2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/highmem.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/jiffies.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/vmstat.h>
43 #include <linux/mempolicy.h>
44 #include <linux/memremap.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_ext.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <trace/events/oom.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
62 #include <linux/sched/mm.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
65 #include <linux/memcontrol.h>
66 #include <linux/ftrace.h>
67 #include <linux/lockdep.h>
68 #include <linux/nmi.h>
69 #include <linux/psi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node);
82 EXPORT_PER_CPU_SYMBOL(numa_node);
85 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
87 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
94 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96 int _node_numa_mem_[MAX_NUMNODES];
99 /* work_structs for global per-cpu drains */
100 DEFINE_MUTEX(pcpu_drain_mutex);
101 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
103 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104 volatile unsigned long latent_entropy __latent_entropy;
105 EXPORT_SYMBOL(latent_entropy);
109 * Array of node states.
111 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
112 [N_POSSIBLE] = NODE_MASK_ALL,
113 [N_ONLINE] = { { [0] = 1UL } },
115 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
116 #ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY] = { { [0] = 1UL } },
119 [N_MEMORY] = { { [0] = 1UL } },
120 [N_CPU] = { { [0] = 1UL } },
123 EXPORT_SYMBOL(node_states);
125 atomic_long_t _totalram_pages __read_mostly;
126 EXPORT_SYMBOL(_totalram_pages);
127 unsigned long totalreserve_pages __read_mostly;
128 unsigned long totalcma_pages __read_mostly;
130 int percpu_pagelist_fraction;
131 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
134 * A cached value of the page's pageblock's migratetype, used when the page is
135 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
136 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
137 * Also the migratetype set in the page does not necessarily match the pcplist
138 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
139 * other index - this ensures that it will be put on the correct CMA freelist.
141 static inline int get_pcppage_migratetype(struct page *page)
146 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
148 page->index = migratetype;
151 #ifdef CONFIG_PM_SLEEP
153 * The following functions are used by the suspend/hibernate code to temporarily
154 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
155 * while devices are suspended. To avoid races with the suspend/hibernate code,
156 * they should always be called with system_transition_mutex held
157 * (gfp_allowed_mask also should only be modified with system_transition_mutex
158 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
159 * with that modification).
162 static gfp_t saved_gfp_mask;
164 void pm_restore_gfp_mask(void)
166 WARN_ON(!mutex_is_locked(&system_transition_mutex));
167 if (saved_gfp_mask) {
168 gfp_allowed_mask = saved_gfp_mask;
173 void pm_restrict_gfp_mask(void)
175 WARN_ON(!mutex_is_locked(&system_transition_mutex));
176 WARN_ON(saved_gfp_mask);
177 saved_gfp_mask = gfp_allowed_mask;
178 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
181 bool pm_suspended_storage(void)
183 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
187 #endif /* CONFIG_PM_SLEEP */
189 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
190 unsigned int pageblock_order __read_mostly;
193 static void __free_pages_ok(struct page *page, unsigned int order);
196 * results with 256, 32 in the lowmem_reserve sysctl:
197 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
198 * 1G machine -> (16M dma, 784M normal, 224M high)
199 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
200 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
201 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 * TBD: should special case ZONE_DMA32 machines here - in those we normally
204 * don't need any ZONE_NORMAL reservation
206 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
207 #ifdef CONFIG_ZONE_DMA
210 #ifdef CONFIG_ZONE_DMA32
214 #ifdef CONFIG_HIGHMEM
220 EXPORT_SYMBOL(totalram_pages);
222 static char * const zone_names[MAX_NR_ZONES] = {
223 #ifdef CONFIG_ZONE_DMA
226 #ifdef CONFIG_ZONE_DMA32
230 #ifdef CONFIG_HIGHMEM
234 #ifdef CONFIG_ZONE_DEVICE
239 const 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_boost_factor __read_mostly = 15000;
266 int watermark_scale_factor = 10;
268 static unsigned long nr_kernel_pages __initdata;
269 static unsigned long nr_all_pages __initdata;
270 static unsigned long dma_reserve __initdata;
272 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
274 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
275 static unsigned long required_kernelcore __initdata;
276 static unsigned long required_kernelcore_percent __initdata;
277 static unsigned long required_movablecore __initdata;
278 static unsigned long required_movablecore_percent __initdata;
279 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
280 static bool mirrored_kernelcore __meminitdata;
282 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
284 EXPORT_SYMBOL(movable_zone);
285 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
288 int nr_node_ids __read_mostly = MAX_NUMNODES;
289 int nr_online_nodes __read_mostly = 1;
290 EXPORT_SYMBOL(nr_node_ids);
291 EXPORT_SYMBOL(nr_online_nodes);
294 int page_group_by_mobility_disabled __read_mostly;
296 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
297 /* Returns true if the struct page for the pfn is uninitialised */
298 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
300 int nid = early_pfn_to_nid(pfn);
302 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
309 * Returns true when the remaining initialisation should be deferred until
310 * later in the boot cycle when it can be parallelised.
312 static bool __meminit
313 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
315 static unsigned long prev_end_pfn, nr_initialised;
318 * prev_end_pfn static that contains the end of previous zone
319 * No need to protect because called very early in boot before smp_init.
321 if (prev_end_pfn != end_pfn) {
322 prev_end_pfn = end_pfn;
326 /* Always populate low zones for address-constrained allocations */
327 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
331 * We start only with one section of pages, more pages are added as
332 * needed until the rest of deferred pages are initialized.
335 if ((nr_initialised > PAGES_PER_SECTION) &&
336 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
337 NODE_DATA(nid)->first_deferred_pfn = pfn;
343 static inline bool early_page_uninitialised(unsigned long pfn)
348 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
354 /* Return a pointer to the bitmap storing bits affecting a block of pages */
355 static inline unsigned long *get_pageblock_bitmap(struct page *page,
358 #ifdef CONFIG_SPARSEMEM
359 return __pfn_to_section(pfn)->pageblock_flags;
361 return page_zone(page)->pageblock_flags;
362 #endif /* CONFIG_SPARSEMEM */
365 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
367 #ifdef CONFIG_SPARSEMEM
368 pfn &= (PAGES_PER_SECTION-1);
369 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
371 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
372 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
373 #endif /* CONFIG_SPARSEMEM */
377 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
378 * @page: The page within the block of interest
379 * @pfn: The target page frame number
380 * @end_bitidx: The last bit of interest to retrieve
381 * @mask: mask of bits that the caller is interested in
383 * Return: pageblock_bits flags
385 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
387 unsigned long end_bitidx,
390 unsigned long *bitmap;
391 unsigned long bitidx, word_bitidx;
394 bitmap = get_pageblock_bitmap(page, pfn);
395 bitidx = pfn_to_bitidx(page, pfn);
396 word_bitidx = bitidx / BITS_PER_LONG;
397 bitidx &= (BITS_PER_LONG-1);
399 word = bitmap[word_bitidx];
400 bitidx += end_bitidx;
401 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
404 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
405 unsigned long end_bitidx,
408 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
411 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
413 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
417 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
418 * @page: The page within the block of interest
419 * @flags: The flags to set
420 * @pfn: The target page frame number
421 * @end_bitidx: The last bit of interest
422 * @mask: mask of bits that the caller is interested in
424 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
426 unsigned long end_bitidx,
429 unsigned long *bitmap;
430 unsigned long bitidx, word_bitidx;
431 unsigned long old_word, word;
433 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
435 bitmap = get_pageblock_bitmap(page, pfn);
436 bitidx = pfn_to_bitidx(page, pfn);
437 word_bitidx = bitidx / BITS_PER_LONG;
438 bitidx &= (BITS_PER_LONG-1);
440 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
442 bitidx += end_bitidx;
443 mask <<= (BITS_PER_LONG - bitidx - 1);
444 flags <<= (BITS_PER_LONG - bitidx - 1);
446 word = READ_ONCE(bitmap[word_bitidx]);
448 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
449 if (word == old_word)
455 void set_pageblock_migratetype(struct page *page, int migratetype)
457 if (unlikely(page_group_by_mobility_disabled &&
458 migratetype < MIGRATE_PCPTYPES))
459 migratetype = MIGRATE_UNMOVABLE;
461 set_pageblock_flags_group(page, (unsigned long)migratetype,
462 PB_migrate, PB_migrate_end);
465 #ifdef CONFIG_DEBUG_VM
466 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
470 unsigned long pfn = page_to_pfn(page);
471 unsigned long sp, start_pfn;
474 seq = zone_span_seqbegin(zone);
475 start_pfn = zone->zone_start_pfn;
476 sp = zone->spanned_pages;
477 if (!zone_spans_pfn(zone, pfn))
479 } while (zone_span_seqretry(zone, seq));
482 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
483 pfn, zone_to_nid(zone), zone->name,
484 start_pfn, start_pfn + sp);
489 static int page_is_consistent(struct zone *zone, struct page *page)
491 if (!pfn_valid_within(page_to_pfn(page)))
493 if (zone != page_zone(page))
499 * Temporary debugging check for pages not lying within a given zone.
501 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
503 if (page_outside_zone_boundaries(zone, page))
505 if (!page_is_consistent(zone, page))
511 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
517 static void bad_page(struct page *page, const char *reason,
518 unsigned long bad_flags)
520 static unsigned long resume;
521 static unsigned long nr_shown;
522 static unsigned long nr_unshown;
525 * Allow a burst of 60 reports, then keep quiet for that minute;
526 * or allow a steady drip of one report per second.
528 if (nr_shown == 60) {
529 if (time_before(jiffies, resume)) {
535 "BUG: Bad page state: %lu messages suppressed\n",
542 resume = jiffies + 60 * HZ;
544 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
545 current->comm, page_to_pfn(page));
546 __dump_page(page, reason);
547 bad_flags &= page->flags;
549 pr_alert("bad because of flags: %#lx(%pGp)\n",
550 bad_flags, &bad_flags);
551 dump_page_owner(page);
556 /* Leave bad fields for debug, except PageBuddy could make trouble */
557 page_mapcount_reset(page); /* remove PageBuddy */
558 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
562 * Higher-order pages are called "compound pages". They are structured thusly:
564 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
566 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
567 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
569 * The first tail page's ->compound_dtor holds the offset in array of compound
570 * page destructors. See compound_page_dtors.
572 * The first tail page's ->compound_order holds the order of allocation.
573 * This usage means that zero-order pages may not be compound.
576 void free_compound_page(struct page *page)
578 __free_pages_ok(page, compound_order(page));
581 void prep_compound_page(struct page *page, unsigned int order)
584 int nr_pages = 1 << order;
586 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
587 set_compound_order(page, order);
589 for (i = 1; i < nr_pages; i++) {
590 struct page *p = page + i;
591 set_page_count(p, 0);
592 p->mapping = TAIL_MAPPING;
593 set_compound_head(p, page);
595 atomic_set(compound_mapcount_ptr(page), -1);
598 #ifdef CONFIG_DEBUG_PAGEALLOC
599 unsigned int _debug_guardpage_minorder;
600 bool _debug_pagealloc_enabled __read_mostly
601 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
602 EXPORT_SYMBOL(_debug_pagealloc_enabled);
603 bool _debug_guardpage_enabled __read_mostly;
605 static int __init early_debug_pagealloc(char *buf)
609 return kstrtobool(buf, &_debug_pagealloc_enabled);
611 early_param("debug_pagealloc", early_debug_pagealloc);
613 static bool need_debug_guardpage(void)
615 /* If we don't use debug_pagealloc, we don't need guard page */
616 if (!debug_pagealloc_enabled())
619 if (!debug_guardpage_minorder())
625 static void init_debug_guardpage(void)
627 if (!debug_pagealloc_enabled())
630 if (!debug_guardpage_minorder())
633 _debug_guardpage_enabled = true;
636 struct page_ext_operations debug_guardpage_ops = {
637 .need = need_debug_guardpage,
638 .init = init_debug_guardpage,
641 static int __init debug_guardpage_minorder_setup(char *buf)
645 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
646 pr_err("Bad debug_guardpage_minorder value\n");
649 _debug_guardpage_minorder = res;
650 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
653 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
655 static inline bool set_page_guard(struct zone *zone, struct page *page,
656 unsigned int order, int migratetype)
658 struct page_ext *page_ext;
660 if (!debug_guardpage_enabled())
663 if (order >= debug_guardpage_minorder())
666 page_ext = lookup_page_ext(page);
667 if (unlikely(!page_ext))
670 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
672 INIT_LIST_HEAD(&page->lru);
673 set_page_private(page, order);
674 /* Guard pages are not available for any usage */
675 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
680 static inline void clear_page_guard(struct zone *zone, struct page *page,
681 unsigned int order, int migratetype)
683 struct page_ext *page_ext;
685 if (!debug_guardpage_enabled())
688 page_ext = lookup_page_ext(page);
689 if (unlikely(!page_ext))
692 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
694 set_page_private(page, 0);
695 if (!is_migrate_isolate(migratetype))
696 __mod_zone_freepage_state(zone, (1 << order), migratetype);
699 struct page_ext_operations debug_guardpage_ops;
700 static inline bool set_page_guard(struct zone *zone, struct page *page,
701 unsigned int order, int migratetype) { return false; }
702 static inline void clear_page_guard(struct zone *zone, struct page *page,
703 unsigned int order, int migratetype) {}
706 static inline void set_page_order(struct page *page, unsigned int order)
708 set_page_private(page, order);
709 __SetPageBuddy(page);
712 static inline void rmv_page_order(struct page *page)
714 __ClearPageBuddy(page);
715 set_page_private(page, 0);
719 * This function checks whether a page is free && is the buddy
720 * we can coalesce a page and its buddy if
721 * (a) the buddy is not in a hole (check before calling!) &&
722 * (b) the buddy is in the buddy system &&
723 * (c) a page and its buddy have the same order &&
724 * (d) a page and its buddy are in the same zone.
726 * For recording whether a page is in the buddy system, we set PageBuddy.
727 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
729 * For recording page's order, we use page_private(page).
731 static inline int page_is_buddy(struct page *page, struct page *buddy,
734 if (page_is_guard(buddy) && page_order(buddy) == order) {
735 if (page_zone_id(page) != page_zone_id(buddy))
738 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
743 if (PageBuddy(buddy) && page_order(buddy) == order) {
745 * zone check is done late to avoid uselessly
746 * calculating zone/node ids for pages that could
749 if (page_zone_id(page) != page_zone_id(buddy))
752 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
760 * Freeing function for a buddy system allocator.
762 * The concept of a buddy system is to maintain direct-mapped table
763 * (containing bit values) for memory blocks of various "orders".
764 * The bottom level table contains the map for the smallest allocatable
765 * units of memory (here, pages), and each level above it describes
766 * pairs of units from the levels below, hence, "buddies".
767 * At a high level, all that happens here is marking the table entry
768 * at the bottom level available, and propagating the changes upward
769 * as necessary, plus some accounting needed to play nicely with other
770 * parts of the VM system.
771 * At each level, we keep a list of pages, which are heads of continuous
772 * free pages of length of (1 << order) and marked with PageBuddy.
773 * Page's order is recorded in page_private(page) field.
774 * So when we are allocating or freeing one, we can derive the state of the
775 * other. That is, if we allocate a small block, and both were
776 * free, the remainder of the region must be split into blocks.
777 * If a block is freed, and its buddy is also free, then this
778 * triggers coalescing into a block of larger size.
783 static inline void __free_one_page(struct page *page,
785 struct zone *zone, unsigned int order,
788 unsigned long combined_pfn;
789 unsigned long uninitialized_var(buddy_pfn);
791 unsigned int max_order;
793 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
795 VM_BUG_ON(!zone_is_initialized(zone));
796 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
798 VM_BUG_ON(migratetype == -1);
799 if (likely(!is_migrate_isolate(migratetype)))
800 __mod_zone_freepage_state(zone, 1 << order, migratetype);
802 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
803 VM_BUG_ON_PAGE(bad_range(zone, page), page);
806 while (order < max_order - 1) {
807 buddy_pfn = __find_buddy_pfn(pfn, order);
808 buddy = page + (buddy_pfn - pfn);
810 if (!pfn_valid_within(buddy_pfn))
812 if (!page_is_buddy(page, buddy, order))
815 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
816 * merge with it and move up one order.
818 if (page_is_guard(buddy)) {
819 clear_page_guard(zone, buddy, order, migratetype);
821 list_del(&buddy->lru);
822 zone->free_area[order].nr_free--;
823 rmv_page_order(buddy);
825 combined_pfn = buddy_pfn & pfn;
826 page = page + (combined_pfn - pfn);
830 if (max_order < MAX_ORDER) {
831 /* If we are here, it means order is >= pageblock_order.
832 * We want to prevent merge between freepages on isolate
833 * pageblock and normal pageblock. Without this, pageblock
834 * isolation could cause incorrect freepage or CMA accounting.
836 * We don't want to hit this code for the more frequent
839 if (unlikely(has_isolate_pageblock(zone))) {
842 buddy_pfn = __find_buddy_pfn(pfn, order);
843 buddy = page + (buddy_pfn - pfn);
844 buddy_mt = get_pageblock_migratetype(buddy);
846 if (migratetype != buddy_mt
847 && (is_migrate_isolate(migratetype) ||
848 is_migrate_isolate(buddy_mt)))
852 goto continue_merging;
856 set_page_order(page, order);
859 * If this is not the largest possible page, check if the buddy
860 * of the next-highest order is free. If it is, it's possible
861 * that pages are being freed that will coalesce soon. In case,
862 * that is happening, add the free page to the tail of the list
863 * so it's less likely to be used soon and more likely to be merged
864 * as a higher order page
866 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
867 struct page *higher_page, *higher_buddy;
868 combined_pfn = buddy_pfn & pfn;
869 higher_page = page + (combined_pfn - pfn);
870 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
871 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
872 if (pfn_valid_within(buddy_pfn) &&
873 page_is_buddy(higher_page, higher_buddy, order + 1)) {
874 list_add_tail(&page->lru,
875 &zone->free_area[order].free_list[migratetype]);
880 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
882 zone->free_area[order].nr_free++;
886 * A bad page could be due to a number of fields. Instead of multiple branches,
887 * try and check multiple fields with one check. The caller must do a detailed
888 * check if necessary.
890 static inline bool page_expected_state(struct page *page,
891 unsigned long check_flags)
893 if (unlikely(atomic_read(&page->_mapcount) != -1))
896 if (unlikely((unsigned long)page->mapping |
897 page_ref_count(page) |
899 (unsigned long)page->mem_cgroup |
901 (page->flags & check_flags)))
907 static void free_pages_check_bad(struct page *page)
909 const char *bad_reason;
910 unsigned long bad_flags;
915 if (unlikely(atomic_read(&page->_mapcount) != -1))
916 bad_reason = "nonzero mapcount";
917 if (unlikely(page->mapping != NULL))
918 bad_reason = "non-NULL mapping";
919 if (unlikely(page_ref_count(page) != 0))
920 bad_reason = "nonzero _refcount";
921 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
922 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
923 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
926 if (unlikely(page->mem_cgroup))
927 bad_reason = "page still charged to cgroup";
929 bad_page(page, bad_reason, bad_flags);
932 static inline int free_pages_check(struct page *page)
934 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
937 /* Something has gone sideways, find it */
938 free_pages_check_bad(page);
942 static int free_tail_pages_check(struct page *head_page, struct page *page)
947 * We rely page->lru.next never has bit 0 set, unless the page
948 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
950 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
952 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
956 switch (page - head_page) {
958 /* the first tail page: ->mapping may be compound_mapcount() */
959 if (unlikely(compound_mapcount(page))) {
960 bad_page(page, "nonzero compound_mapcount", 0);
966 * the second tail page: ->mapping is
967 * deferred_list.next -- ignore value.
971 if (page->mapping != TAIL_MAPPING) {
972 bad_page(page, "corrupted mapping in tail page", 0);
977 if (unlikely(!PageTail(page))) {
978 bad_page(page, "PageTail not set", 0);
981 if (unlikely(compound_head(page) != head_page)) {
982 bad_page(page, "compound_head not consistent", 0);
987 page->mapping = NULL;
988 clear_compound_head(page);
992 static __always_inline bool free_pages_prepare(struct page *page,
993 unsigned int order, bool check_free)
997 VM_BUG_ON_PAGE(PageTail(page), page);
999 trace_mm_page_free(page, order);
1002 * Check tail pages before head page information is cleared to
1003 * avoid checking PageCompound for order-0 pages.
1005 if (unlikely(order)) {
1006 bool compound = PageCompound(page);
1009 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1012 ClearPageDoubleMap(page);
1013 for (i = 1; i < (1 << order); i++) {
1015 bad += free_tail_pages_check(page, page + i);
1016 if (unlikely(free_pages_check(page + i))) {
1020 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1023 if (PageMappingFlags(page))
1024 page->mapping = NULL;
1025 if (memcg_kmem_enabled() && PageKmemcg(page))
1026 memcg_kmem_uncharge(page, order);
1028 bad += free_pages_check(page);
1032 page_cpupid_reset_last(page);
1033 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1034 reset_page_owner(page, order);
1036 if (!PageHighMem(page)) {
1037 debug_check_no_locks_freed(page_address(page),
1038 PAGE_SIZE << order);
1039 debug_check_no_obj_freed(page_address(page),
1040 PAGE_SIZE << order);
1042 arch_free_page(page, order);
1043 kernel_poison_pages(page, 1 << order, 0);
1044 kernel_map_pages(page, 1 << order, 0);
1045 kasan_free_pages(page, order);
1050 #ifdef CONFIG_DEBUG_VM
1051 static inline bool free_pcp_prepare(struct page *page)
1053 return free_pages_prepare(page, 0, true);
1056 static inline bool bulkfree_pcp_prepare(struct page *page)
1061 static bool free_pcp_prepare(struct page *page)
1063 return free_pages_prepare(page, 0, false);
1066 static bool bulkfree_pcp_prepare(struct page *page)
1068 return free_pages_check(page);
1070 #endif /* CONFIG_DEBUG_VM */
1072 static inline void prefetch_buddy(struct page *page)
1074 unsigned long pfn = page_to_pfn(page);
1075 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1076 struct page *buddy = page + (buddy_pfn - pfn);
1082 * Frees a number of pages from the PCP lists
1083 * Assumes all pages on list are in same zone, and of same order.
1084 * count is the number of pages to free.
1086 * If the zone was previously in an "all pages pinned" state then look to
1087 * see if this freeing clears that state.
1089 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1090 * pinned" detection logic.
1092 static void free_pcppages_bulk(struct zone *zone, int count,
1093 struct per_cpu_pages *pcp)
1095 int migratetype = 0;
1097 int prefetch_nr = 0;
1098 bool isolated_pageblocks;
1099 struct page *page, *tmp;
1103 struct list_head *list;
1106 * Remove pages from lists in a round-robin fashion. A
1107 * batch_free count is maintained that is incremented when an
1108 * empty list is encountered. This is so more pages are freed
1109 * off fuller lists instead of spinning excessively around empty
1114 if (++migratetype == MIGRATE_PCPTYPES)
1116 list = &pcp->lists[migratetype];
1117 } while (list_empty(list));
1119 /* This is the only non-empty list. Free them all. */
1120 if (batch_free == MIGRATE_PCPTYPES)
1124 page = list_last_entry(list, struct page, lru);
1125 /* must delete to avoid corrupting pcp list */
1126 list_del(&page->lru);
1129 if (bulkfree_pcp_prepare(page))
1132 list_add_tail(&page->lru, &head);
1135 * We are going to put the page back to the global
1136 * pool, prefetch its buddy to speed up later access
1137 * under zone->lock. It is believed the overhead of
1138 * an additional test and calculating buddy_pfn here
1139 * can be offset by reduced memory latency later. To
1140 * avoid excessive prefetching due to large count, only
1141 * prefetch buddy for the first pcp->batch nr of pages.
1143 if (prefetch_nr++ < pcp->batch)
1144 prefetch_buddy(page);
1145 } while (--count && --batch_free && !list_empty(list));
1148 spin_lock(&zone->lock);
1149 isolated_pageblocks = has_isolate_pageblock(zone);
1152 * Use safe version since after __free_one_page(),
1153 * page->lru.next will not point to original list.
1155 list_for_each_entry_safe(page, tmp, &head, lru) {
1156 int mt = get_pcppage_migratetype(page);
1157 /* MIGRATE_ISOLATE page should not go to pcplists */
1158 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1159 /* Pageblock could have been isolated meanwhile */
1160 if (unlikely(isolated_pageblocks))
1161 mt = get_pageblock_migratetype(page);
1163 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1164 trace_mm_page_pcpu_drain(page, 0, mt);
1166 spin_unlock(&zone->lock);
1169 static void free_one_page(struct zone *zone,
1170 struct page *page, unsigned long pfn,
1174 spin_lock(&zone->lock);
1175 if (unlikely(has_isolate_pageblock(zone) ||
1176 is_migrate_isolate(migratetype))) {
1177 migratetype = get_pfnblock_migratetype(page, pfn);
1179 __free_one_page(page, pfn, zone, order, migratetype);
1180 spin_unlock(&zone->lock);
1183 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1184 unsigned long zone, int nid)
1186 mm_zero_struct_page(page);
1187 set_page_links(page, zone, nid, pfn);
1188 init_page_count(page);
1189 page_mapcount_reset(page);
1190 page_cpupid_reset_last(page);
1191 page_kasan_tag_reset(page);
1193 INIT_LIST_HEAD(&page->lru);
1194 #ifdef WANT_PAGE_VIRTUAL
1195 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1196 if (!is_highmem_idx(zone))
1197 set_page_address(page, __va(pfn << PAGE_SHIFT));
1201 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1202 static void __meminit init_reserved_page(unsigned long pfn)
1207 if (!early_page_uninitialised(pfn))
1210 nid = early_pfn_to_nid(pfn);
1211 pgdat = NODE_DATA(nid);
1213 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1214 struct zone *zone = &pgdat->node_zones[zid];
1216 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1219 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1222 static inline void init_reserved_page(unsigned long pfn)
1225 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1228 * Initialised pages do not have PageReserved set. This function is
1229 * called for each range allocated by the bootmem allocator and
1230 * marks the pages PageReserved. The remaining valid pages are later
1231 * sent to the buddy page allocator.
1233 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1235 unsigned long start_pfn = PFN_DOWN(start);
1236 unsigned long end_pfn = PFN_UP(end);
1238 for (; start_pfn < end_pfn; start_pfn++) {
1239 if (pfn_valid(start_pfn)) {
1240 struct page *page = pfn_to_page(start_pfn);
1242 init_reserved_page(start_pfn);
1244 /* Avoid false-positive PageTail() */
1245 INIT_LIST_HEAD(&page->lru);
1248 * no need for atomic set_bit because the struct
1249 * page is not visible yet so nobody should
1252 __SetPageReserved(page);
1257 static void __free_pages_ok(struct page *page, unsigned int order)
1259 unsigned long flags;
1261 unsigned long pfn = page_to_pfn(page);
1263 if (!free_pages_prepare(page, order, true))
1266 migratetype = get_pfnblock_migratetype(page, pfn);
1267 local_irq_save(flags);
1268 __count_vm_events(PGFREE, 1 << order);
1269 free_one_page(page_zone(page), page, pfn, order, migratetype);
1270 local_irq_restore(flags);
1273 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1275 unsigned int nr_pages = 1 << order;
1276 struct page *p = page;
1280 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1282 __ClearPageReserved(p);
1283 set_page_count(p, 0);
1285 __ClearPageReserved(p);
1286 set_page_count(p, 0);
1288 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1289 set_page_refcounted(page);
1290 __free_pages(page, order);
1293 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1294 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1296 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1298 int __meminit early_pfn_to_nid(unsigned long pfn)
1300 static DEFINE_SPINLOCK(early_pfn_lock);
1303 spin_lock(&early_pfn_lock);
1304 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1306 nid = first_online_node;
1307 spin_unlock(&early_pfn_lock);
1313 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1314 static inline bool __meminit __maybe_unused
1315 meminit_pfn_in_nid(unsigned long pfn, int node,
1316 struct mminit_pfnnid_cache *state)
1320 nid = __early_pfn_to_nid(pfn, state);
1321 if (nid >= 0 && nid != node)
1326 /* Only safe to use early in boot when initialisation is single-threaded */
1327 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1329 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1334 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1338 static inline bool __meminit __maybe_unused
1339 meminit_pfn_in_nid(unsigned long pfn, int node,
1340 struct mminit_pfnnid_cache *state)
1347 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1350 if (early_page_uninitialised(pfn))
1352 return __free_pages_boot_core(page, order);
1356 * Check that the whole (or subset of) a pageblock given by the interval of
1357 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1358 * with the migration of free compaction scanner. The scanners then need to
1359 * use only pfn_valid_within() check for arches that allow holes within
1362 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1364 * It's possible on some configurations to have a setup like node0 node1 node0
1365 * i.e. it's possible that all pages within a zones range of pages do not
1366 * belong to a single zone. We assume that a border between node0 and node1
1367 * can occur within a single pageblock, but not a node0 node1 node0
1368 * interleaving within a single pageblock. It is therefore sufficient to check
1369 * the first and last page of a pageblock and avoid checking each individual
1370 * page in a pageblock.
1372 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1373 unsigned long end_pfn, struct zone *zone)
1375 struct page *start_page;
1376 struct page *end_page;
1378 /* end_pfn is one past the range we are checking */
1381 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1384 start_page = pfn_to_online_page(start_pfn);
1388 if (page_zone(start_page) != zone)
1391 end_page = pfn_to_page(end_pfn);
1393 /* This gives a shorter code than deriving page_zone(end_page) */
1394 if (page_zone_id(start_page) != page_zone_id(end_page))
1400 void set_zone_contiguous(struct zone *zone)
1402 unsigned long block_start_pfn = zone->zone_start_pfn;
1403 unsigned long block_end_pfn;
1405 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1406 for (; block_start_pfn < zone_end_pfn(zone);
1407 block_start_pfn = block_end_pfn,
1408 block_end_pfn += pageblock_nr_pages) {
1410 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1412 if (!__pageblock_pfn_to_page(block_start_pfn,
1413 block_end_pfn, zone))
1417 /* We confirm that there is no hole */
1418 zone->contiguous = true;
1421 void clear_zone_contiguous(struct zone *zone)
1423 zone->contiguous = false;
1426 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1427 static void __init deferred_free_range(unsigned long pfn,
1428 unsigned long nr_pages)
1436 page = pfn_to_page(pfn);
1438 /* Free a large naturally-aligned chunk if possible */
1439 if (nr_pages == pageblock_nr_pages &&
1440 (pfn & (pageblock_nr_pages - 1)) == 0) {
1441 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1442 __free_pages_boot_core(page, pageblock_order);
1446 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1447 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1448 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1449 __free_pages_boot_core(page, 0);
1453 /* Completion tracking for deferred_init_memmap() threads */
1454 static atomic_t pgdat_init_n_undone __initdata;
1455 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1457 static inline void __init pgdat_init_report_one_done(void)
1459 if (atomic_dec_and_test(&pgdat_init_n_undone))
1460 complete(&pgdat_init_all_done_comp);
1464 * Returns true if page needs to be initialized or freed to buddy allocator.
1466 * First we check if pfn is valid on architectures where it is possible to have
1467 * holes within pageblock_nr_pages. On systems where it is not possible, this
1468 * function is optimized out.
1470 * Then, we check if a current large page is valid by only checking the validity
1473 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1474 * within a node: a pfn is between start and end of a node, but does not belong
1475 * to this memory node.
1477 static inline bool __init
1478 deferred_pfn_valid(int nid, unsigned long pfn,
1479 struct mminit_pfnnid_cache *nid_init_state)
1481 if (!pfn_valid_within(pfn))
1483 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1485 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1491 * Free pages to buddy allocator. Try to free aligned pages in
1492 * pageblock_nr_pages sizes.
1494 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1495 unsigned long end_pfn)
1497 struct mminit_pfnnid_cache nid_init_state = { };
1498 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1499 unsigned long nr_free = 0;
1501 for (; pfn < end_pfn; pfn++) {
1502 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1503 deferred_free_range(pfn - nr_free, nr_free);
1505 } else if (!(pfn & nr_pgmask)) {
1506 deferred_free_range(pfn - nr_free, nr_free);
1508 touch_nmi_watchdog();
1513 /* Free the last block of pages to allocator */
1514 deferred_free_range(pfn - nr_free, nr_free);
1518 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1519 * by performing it only once every pageblock_nr_pages.
1520 * Return number of pages initialized.
1522 static unsigned long __init deferred_init_pages(int nid, int zid,
1524 unsigned long end_pfn)
1526 struct mminit_pfnnid_cache nid_init_state = { };
1527 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1528 unsigned long nr_pages = 0;
1529 struct page *page = NULL;
1531 for (; pfn < end_pfn; pfn++) {
1532 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1535 } else if (!page || !(pfn & nr_pgmask)) {
1536 page = pfn_to_page(pfn);
1537 touch_nmi_watchdog();
1541 __init_single_page(page, pfn, zid, nid);
1547 /* Initialise remaining memory on a node */
1548 static int __init deferred_init_memmap(void *data)
1550 pg_data_t *pgdat = data;
1551 int nid = pgdat->node_id;
1552 unsigned long start = jiffies;
1553 unsigned long nr_pages = 0;
1554 unsigned long spfn, epfn, first_init_pfn, flags;
1555 phys_addr_t spa, epa;
1558 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1561 /* Bind memory initialisation thread to a local node if possible */
1562 if (!cpumask_empty(cpumask))
1563 set_cpus_allowed_ptr(current, cpumask);
1565 pgdat_resize_lock(pgdat, &flags);
1566 first_init_pfn = pgdat->first_deferred_pfn;
1567 if (first_init_pfn == ULONG_MAX) {
1568 pgdat_resize_unlock(pgdat, &flags);
1569 pgdat_init_report_one_done();
1573 /* Sanity check boundaries */
1574 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1575 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1576 pgdat->first_deferred_pfn = ULONG_MAX;
1578 /* Only the highest zone is deferred so find it */
1579 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1580 zone = pgdat->node_zones + zid;
1581 if (first_init_pfn < zone_end_pfn(zone))
1584 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1587 * Initialize and free pages. We do it in two loops: first we initialize
1588 * struct page, than free to buddy allocator, because while we are
1589 * freeing pages we can access pages that are ahead (computing buddy
1590 * page in __free_one_page()).
1592 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1593 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1594 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1595 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1597 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1598 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1599 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1600 deferred_free_pages(nid, zid, spfn, epfn);
1602 pgdat_resize_unlock(pgdat, &flags);
1604 /* Sanity check that the next zone really is unpopulated */
1605 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1607 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1608 jiffies_to_msecs(jiffies - start));
1610 pgdat_init_report_one_done();
1615 * During boot we initialize deferred pages on-demand, as needed, but once
1616 * page_alloc_init_late() has finished, the deferred pages are all initialized,
1617 * and we can permanently disable that path.
1619 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
1622 * If this zone has deferred pages, try to grow it by initializing enough
1623 * deferred pages to satisfy the allocation specified by order, rounded up to
1624 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1625 * of SECTION_SIZE bytes by initializing struct pages in increments of
1626 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1628 * Return true when zone was grown, otherwise return false. We return true even
1629 * when we grow less than requested, to let the caller decide if there are
1630 * enough pages to satisfy the allocation.
1632 * Note: We use noinline because this function is needed only during boot, and
1633 * it is called from a __ref function _deferred_grow_zone. This way we are
1634 * making sure that it is not inlined into permanent text section.
1636 static noinline bool __init
1637 deferred_grow_zone(struct zone *zone, unsigned int order)
1639 int zid = zone_idx(zone);
1640 int nid = zone_to_nid(zone);
1641 pg_data_t *pgdat = NODE_DATA(nid);
1642 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1643 unsigned long nr_pages = 0;
1644 unsigned long first_init_pfn, spfn, epfn, t, flags;
1645 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1646 phys_addr_t spa, epa;
1649 /* Only the last zone may have deferred pages */
1650 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1653 pgdat_resize_lock(pgdat, &flags);
1656 * If deferred pages have been initialized while we were waiting for
1657 * the lock, return true, as the zone was grown. The caller will retry
1658 * this zone. We won't return to this function since the caller also
1659 * has this static branch.
1661 if (!static_branch_unlikely(&deferred_pages)) {
1662 pgdat_resize_unlock(pgdat, &flags);
1667 * If someone grew this zone while we were waiting for spinlock, return
1668 * true, as there might be enough pages already.
1670 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1671 pgdat_resize_unlock(pgdat, &flags);
1675 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1677 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1678 pgdat_resize_unlock(pgdat, &flags);
1682 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1683 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1684 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1686 while (spfn < epfn && nr_pages < nr_pages_needed) {
1687 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1688 first_deferred_pfn = min(t, epfn);
1689 nr_pages += deferred_init_pages(nid, zid, spfn,
1690 first_deferred_pfn);
1691 spfn = first_deferred_pfn;
1694 if (nr_pages >= nr_pages_needed)
1698 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1699 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1700 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1701 deferred_free_pages(nid, zid, spfn, epfn);
1703 if (first_deferred_pfn == epfn)
1706 pgdat->first_deferred_pfn = first_deferred_pfn;
1707 pgdat_resize_unlock(pgdat, &flags);
1709 return nr_pages > 0;
1713 * deferred_grow_zone() is __init, but it is called from
1714 * get_page_from_freelist() during early boot until deferred_pages permanently
1715 * disables this call. This is why we have refdata wrapper to avoid warning,
1716 * and to ensure that the function body gets unloaded.
1719 _deferred_grow_zone(struct zone *zone, unsigned int order)
1721 return deferred_grow_zone(zone, order);
1724 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1726 void __init page_alloc_init_late(void)
1730 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1733 /* There will be num_node_state(N_MEMORY) threads */
1734 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1735 for_each_node_state(nid, N_MEMORY) {
1736 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1739 /* Block until all are initialised */
1740 wait_for_completion(&pgdat_init_all_done_comp);
1743 * We initialized the rest of the deferred pages. Permanently disable
1744 * on-demand struct page initialization.
1746 static_branch_disable(&deferred_pages);
1748 /* Reinit limits that are based on free pages after the kernel is up */
1749 files_maxfiles_init();
1751 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1752 /* Discard memblock private memory */
1756 for_each_populated_zone(zone)
1757 set_zone_contiguous(zone);
1761 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1762 void __init init_cma_reserved_pageblock(struct page *page)
1764 unsigned i = pageblock_nr_pages;
1765 struct page *p = page;
1768 __ClearPageReserved(p);
1769 set_page_count(p, 0);
1772 set_pageblock_migratetype(page, MIGRATE_CMA);
1774 if (pageblock_order >= MAX_ORDER) {
1775 i = pageblock_nr_pages;
1778 set_page_refcounted(p);
1779 __free_pages(p, MAX_ORDER - 1);
1780 p += MAX_ORDER_NR_PAGES;
1781 } while (i -= MAX_ORDER_NR_PAGES);
1783 set_page_refcounted(page);
1784 __free_pages(page, pageblock_order);
1787 adjust_managed_page_count(page, pageblock_nr_pages);
1792 * The order of subdivision here is critical for the IO subsystem.
1793 * Please do not alter this order without good reasons and regression
1794 * testing. Specifically, as large blocks of memory are subdivided,
1795 * the order in which smaller blocks are delivered depends on the order
1796 * they're subdivided in this function. This is the primary factor
1797 * influencing the order in which pages are delivered to the IO
1798 * subsystem according to empirical testing, and this is also justified
1799 * by considering the behavior of a buddy system containing a single
1800 * large block of memory acted on by a series of small allocations.
1801 * This behavior is a critical factor in sglist merging's success.
1805 static inline void expand(struct zone *zone, struct page *page,
1806 int low, int high, struct free_area *area,
1809 unsigned long size = 1 << high;
1811 while (high > low) {
1815 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1818 * Mark as guard pages (or page), that will allow to
1819 * merge back to allocator when buddy will be freed.
1820 * Corresponding page table entries will not be touched,
1821 * pages will stay not present in virtual address space
1823 if (set_page_guard(zone, &page[size], high, migratetype))
1826 list_add(&page[size].lru, &area->free_list[migratetype]);
1828 set_page_order(&page[size], high);
1832 static void check_new_page_bad(struct page *page)
1834 const char *bad_reason = NULL;
1835 unsigned long bad_flags = 0;
1837 if (unlikely(atomic_read(&page->_mapcount) != -1))
1838 bad_reason = "nonzero mapcount";
1839 if (unlikely(page->mapping != NULL))
1840 bad_reason = "non-NULL mapping";
1841 if (unlikely(page_ref_count(page) != 0))
1842 bad_reason = "nonzero _count";
1843 if (unlikely(page->flags & __PG_HWPOISON)) {
1844 bad_reason = "HWPoisoned (hardware-corrupted)";
1845 bad_flags = __PG_HWPOISON;
1846 /* Don't complain about hwpoisoned pages */
1847 page_mapcount_reset(page); /* remove PageBuddy */
1850 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1851 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1852 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1855 if (unlikely(page->mem_cgroup))
1856 bad_reason = "page still charged to cgroup";
1858 bad_page(page, bad_reason, bad_flags);
1862 * This page is about to be returned from the page allocator
1864 static inline int check_new_page(struct page *page)
1866 if (likely(page_expected_state(page,
1867 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1870 check_new_page_bad(page);
1874 static inline bool free_pages_prezeroed(void)
1876 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1877 page_poisoning_enabled();
1880 #ifdef CONFIG_DEBUG_VM
1881 static bool check_pcp_refill(struct page *page)
1886 static bool check_new_pcp(struct page *page)
1888 return check_new_page(page);
1891 static bool check_pcp_refill(struct page *page)
1893 return check_new_page(page);
1895 static bool check_new_pcp(struct page *page)
1899 #endif /* CONFIG_DEBUG_VM */
1901 static bool check_new_pages(struct page *page, unsigned int order)
1904 for (i = 0; i < (1 << order); i++) {
1905 struct page *p = page + i;
1907 if (unlikely(check_new_page(p)))
1914 inline void post_alloc_hook(struct page *page, unsigned int order,
1917 set_page_private(page, 0);
1918 set_page_refcounted(page);
1920 arch_alloc_page(page, order);
1921 kernel_map_pages(page, 1 << order, 1);
1922 kernel_poison_pages(page, 1 << order, 1);
1923 kasan_alloc_pages(page, order);
1924 set_page_owner(page, order, gfp_flags);
1927 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1928 unsigned int alloc_flags)
1932 post_alloc_hook(page, order, gfp_flags);
1934 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1935 for (i = 0; i < (1 << order); i++)
1936 clear_highpage(page + i);
1938 if (order && (gfp_flags & __GFP_COMP))
1939 prep_compound_page(page, order);
1942 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1943 * allocate the page. The expectation is that the caller is taking
1944 * steps that will free more memory. The caller should avoid the page
1945 * being used for !PFMEMALLOC purposes.
1947 if (alloc_flags & ALLOC_NO_WATERMARKS)
1948 set_page_pfmemalloc(page);
1950 clear_page_pfmemalloc(page);
1954 * Go through the free lists for the given migratetype and remove
1955 * the smallest available page from the freelists
1957 static __always_inline
1958 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1961 unsigned int current_order;
1962 struct free_area *area;
1965 /* Find a page of the appropriate size in the preferred list */
1966 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1967 area = &(zone->free_area[current_order]);
1968 page = list_first_entry_or_null(&area->free_list[migratetype],
1972 list_del(&page->lru);
1973 rmv_page_order(page);
1975 expand(zone, page, order, current_order, area, migratetype);
1976 set_pcppage_migratetype(page, migratetype);
1985 * This array describes the order lists are fallen back to when
1986 * the free lists for the desirable migrate type are depleted
1988 static int fallbacks[MIGRATE_TYPES][4] = {
1989 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1990 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1991 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1993 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1995 #ifdef CONFIG_MEMORY_ISOLATION
1996 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2001 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2004 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2007 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2008 unsigned int order) { return NULL; }
2012 * Move the free pages in a range to the free lists of the requested type.
2013 * Note that start_page and end_pages are not aligned on a pageblock
2014 * boundary. If alignment is required, use move_freepages_block()
2016 static int move_freepages(struct zone *zone,
2017 struct page *start_page, struct page *end_page,
2018 int migratetype, int *num_movable)
2022 int pages_moved = 0;
2024 #ifndef CONFIG_HOLES_IN_ZONE
2026 * page_zone is not safe to call in this context when
2027 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2028 * anyway as we check zone boundaries in move_freepages_block().
2029 * Remove at a later date when no bug reports exist related to
2030 * grouping pages by mobility
2032 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2033 pfn_valid(page_to_pfn(end_page)) &&
2034 page_zone(start_page) != page_zone(end_page));
2036 for (page = start_page; page <= end_page;) {
2037 if (!pfn_valid_within(page_to_pfn(page))) {
2042 /* Make sure we are not inadvertently changing nodes */
2043 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2045 if (!PageBuddy(page)) {
2047 * We assume that pages that could be isolated for
2048 * migration are movable. But we don't actually try
2049 * isolating, as that would be expensive.
2052 (PageLRU(page) || __PageMovable(page)))
2059 order = page_order(page);
2060 list_move(&page->lru,
2061 &zone->free_area[order].free_list[migratetype]);
2063 pages_moved += 1 << order;
2069 int move_freepages_block(struct zone *zone, struct page *page,
2070 int migratetype, int *num_movable)
2072 unsigned long start_pfn, end_pfn;
2073 struct page *start_page, *end_page;
2078 start_pfn = page_to_pfn(page);
2079 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2080 start_page = pfn_to_page(start_pfn);
2081 end_page = start_page + pageblock_nr_pages - 1;
2082 end_pfn = start_pfn + pageblock_nr_pages - 1;
2084 /* Do not cross zone boundaries */
2085 if (!zone_spans_pfn(zone, start_pfn))
2087 if (!zone_spans_pfn(zone, end_pfn))
2090 return move_freepages(zone, start_page, end_page, migratetype,
2094 static void change_pageblock_range(struct page *pageblock_page,
2095 int start_order, int migratetype)
2097 int nr_pageblocks = 1 << (start_order - pageblock_order);
2099 while (nr_pageblocks--) {
2100 set_pageblock_migratetype(pageblock_page, migratetype);
2101 pageblock_page += pageblock_nr_pages;
2106 * When we are falling back to another migratetype during allocation, try to
2107 * steal extra free pages from the same pageblocks to satisfy further
2108 * allocations, instead of polluting multiple pageblocks.
2110 * If we are stealing a relatively large buddy page, it is likely there will
2111 * be more free pages in the pageblock, so try to steal them all. For
2112 * reclaimable and unmovable allocations, we steal regardless of page size,
2113 * as fragmentation caused by those allocations polluting movable pageblocks
2114 * is worse than movable allocations stealing from unmovable and reclaimable
2117 static bool can_steal_fallback(unsigned int order, int start_mt)
2120 * Leaving this order check is intended, although there is
2121 * relaxed order check in next check. The reason is that
2122 * we can actually steal whole pageblock if this condition met,
2123 * but, below check doesn't guarantee it and that is just heuristic
2124 * so could be changed anytime.
2126 if (order >= pageblock_order)
2129 if (order >= pageblock_order / 2 ||
2130 start_mt == MIGRATE_RECLAIMABLE ||
2131 start_mt == MIGRATE_UNMOVABLE ||
2132 page_group_by_mobility_disabled)
2138 static inline void boost_watermark(struct zone *zone)
2140 unsigned long max_boost;
2142 if (!watermark_boost_factor)
2145 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2146 watermark_boost_factor, 10000);
2147 max_boost = max(pageblock_nr_pages, max_boost);
2149 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2154 * This function implements actual steal behaviour. If order is large enough,
2155 * we can steal whole pageblock. If not, we first move freepages in this
2156 * pageblock to our migratetype and determine how many already-allocated pages
2157 * are there in the pageblock with a compatible migratetype. If at least half
2158 * of pages are free or compatible, we can change migratetype of the pageblock
2159 * itself, so pages freed in the future will be put on the correct free list.
2161 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2162 unsigned int alloc_flags, int start_type, bool whole_block)
2164 unsigned int current_order = page_order(page);
2165 struct free_area *area;
2166 int free_pages, movable_pages, alike_pages;
2169 old_block_type = get_pageblock_migratetype(page);
2172 * This can happen due to races and we want to prevent broken
2173 * highatomic accounting.
2175 if (is_migrate_highatomic(old_block_type))
2178 /* Take ownership for orders >= pageblock_order */
2179 if (current_order >= pageblock_order) {
2180 change_pageblock_range(page, current_order, start_type);
2185 * Boost watermarks to increase reclaim pressure to reduce the
2186 * likelihood of future fallbacks. Wake kswapd now as the node
2187 * may be balanced overall and kswapd will not wake naturally.
2189 boost_watermark(zone);
2190 if (alloc_flags & ALLOC_KSWAPD)
2191 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
2193 /* We are not allowed to try stealing from the whole block */
2197 free_pages = move_freepages_block(zone, page, start_type,
2200 * Determine how many pages are compatible with our allocation.
2201 * For movable allocation, it's the number of movable pages which
2202 * we just obtained. For other types it's a bit more tricky.
2204 if (start_type == MIGRATE_MOVABLE) {
2205 alike_pages = movable_pages;
2208 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2209 * to MOVABLE pageblock, consider all non-movable pages as
2210 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2211 * vice versa, be conservative since we can't distinguish the
2212 * exact migratetype of non-movable pages.
2214 if (old_block_type == MIGRATE_MOVABLE)
2215 alike_pages = pageblock_nr_pages
2216 - (free_pages + movable_pages);
2221 /* moving whole block can fail due to zone boundary conditions */
2226 * If a sufficient number of pages in the block are either free or of
2227 * comparable migratability as our allocation, claim the whole block.
2229 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2230 page_group_by_mobility_disabled)
2231 set_pageblock_migratetype(page, start_type);
2236 area = &zone->free_area[current_order];
2237 list_move(&page->lru, &area->free_list[start_type]);
2241 * Check whether there is a suitable fallback freepage with requested order.
2242 * If only_stealable is true, this function returns fallback_mt only if
2243 * we can steal other freepages all together. This would help to reduce
2244 * fragmentation due to mixed migratetype pages in one pageblock.
2246 int find_suitable_fallback(struct free_area *area, unsigned int order,
2247 int migratetype, bool only_stealable, bool *can_steal)
2252 if (area->nr_free == 0)
2257 fallback_mt = fallbacks[migratetype][i];
2258 if (fallback_mt == MIGRATE_TYPES)
2261 if (list_empty(&area->free_list[fallback_mt]))
2264 if (can_steal_fallback(order, migratetype))
2267 if (!only_stealable)
2278 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2279 * there are no empty page blocks that contain a page with a suitable order
2281 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2282 unsigned int alloc_order)
2285 unsigned long max_managed, flags;
2288 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2289 * Check is race-prone but harmless.
2291 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2292 if (zone->nr_reserved_highatomic >= max_managed)
2295 spin_lock_irqsave(&zone->lock, flags);
2297 /* Recheck the nr_reserved_highatomic limit under the lock */
2298 if (zone->nr_reserved_highatomic >= max_managed)
2302 mt = get_pageblock_migratetype(page);
2303 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2304 && !is_migrate_cma(mt)) {
2305 zone->nr_reserved_highatomic += pageblock_nr_pages;
2306 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2307 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2311 spin_unlock_irqrestore(&zone->lock, flags);
2315 * Used when an allocation is about to fail under memory pressure. This
2316 * potentially hurts the reliability of high-order allocations when under
2317 * intense memory pressure but failed atomic allocations should be easier
2318 * to recover from than an OOM.
2320 * If @force is true, try to unreserve a pageblock even though highatomic
2321 * pageblock is exhausted.
2323 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2326 struct zonelist *zonelist = ac->zonelist;
2327 unsigned long flags;
2334 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2337 * Preserve at least one pageblock unless memory pressure
2340 if (!force && zone->nr_reserved_highatomic <=
2344 spin_lock_irqsave(&zone->lock, flags);
2345 for (order = 0; order < MAX_ORDER; order++) {
2346 struct free_area *area = &(zone->free_area[order]);
2348 page = list_first_entry_or_null(
2349 &area->free_list[MIGRATE_HIGHATOMIC],
2355 * In page freeing path, migratetype change is racy so
2356 * we can counter several free pages in a pageblock
2357 * in this loop althoug we changed the pageblock type
2358 * from highatomic to ac->migratetype. So we should
2359 * adjust the count once.
2361 if (is_migrate_highatomic_page(page)) {
2363 * It should never happen but changes to
2364 * locking could inadvertently allow a per-cpu
2365 * drain to add pages to MIGRATE_HIGHATOMIC
2366 * while unreserving so be safe and watch for
2369 zone->nr_reserved_highatomic -= min(
2371 zone->nr_reserved_highatomic);
2375 * Convert to ac->migratetype and avoid the normal
2376 * pageblock stealing heuristics. Minimally, the caller
2377 * is doing the work and needs the pages. More
2378 * importantly, if the block was always converted to
2379 * MIGRATE_UNMOVABLE or another type then the number
2380 * of pageblocks that cannot be completely freed
2383 set_pageblock_migratetype(page, ac->migratetype);
2384 ret = move_freepages_block(zone, page, ac->migratetype,
2387 spin_unlock_irqrestore(&zone->lock, flags);
2391 spin_unlock_irqrestore(&zone->lock, flags);
2398 * Try finding a free buddy page on the fallback list and put it on the free
2399 * list of requested migratetype, possibly along with other pages from the same
2400 * block, depending on fragmentation avoidance heuristics. Returns true if
2401 * fallback was found so that __rmqueue_smallest() can grab it.
2403 * The use of signed ints for order and current_order is a deliberate
2404 * deviation from the rest of this file, to make the for loop
2405 * condition simpler.
2407 static __always_inline bool
2408 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2409 unsigned int alloc_flags)
2411 struct free_area *area;
2413 int min_order = order;
2419 * Do not steal pages from freelists belonging to other pageblocks
2420 * i.e. orders < pageblock_order. If there are no local zones free,
2421 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2423 if (alloc_flags & ALLOC_NOFRAGMENT)
2424 min_order = pageblock_order;
2427 * Find the largest available free page in the other list. This roughly
2428 * approximates finding the pageblock with the most free pages, which
2429 * would be too costly to do exactly.
2431 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2433 area = &(zone->free_area[current_order]);
2434 fallback_mt = find_suitable_fallback(area, current_order,
2435 start_migratetype, false, &can_steal);
2436 if (fallback_mt == -1)
2440 * We cannot steal all free pages from the pageblock and the
2441 * requested migratetype is movable. In that case it's better to
2442 * steal and split the smallest available page instead of the
2443 * largest available page, because even if the next movable
2444 * allocation falls back into a different pageblock than this
2445 * one, it won't cause permanent fragmentation.
2447 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2448 && current_order > order)
2457 for (current_order = order; current_order < MAX_ORDER;
2459 area = &(zone->free_area[current_order]);
2460 fallback_mt = find_suitable_fallback(area, current_order,
2461 start_migratetype, false, &can_steal);
2462 if (fallback_mt != -1)
2467 * This should not happen - we already found a suitable fallback
2468 * when looking for the largest page.
2470 VM_BUG_ON(current_order == MAX_ORDER);
2473 page = list_first_entry(&area->free_list[fallback_mt],
2476 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2479 trace_mm_page_alloc_extfrag(page, order, current_order,
2480 start_migratetype, fallback_mt);
2487 * Do the hard work of removing an element from the buddy allocator.
2488 * Call me with the zone->lock already held.
2490 static __always_inline struct page *
2491 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2492 unsigned int alloc_flags)
2497 page = __rmqueue_smallest(zone, order, migratetype);
2498 if (unlikely(!page)) {
2499 if (migratetype == MIGRATE_MOVABLE)
2500 page = __rmqueue_cma_fallback(zone, order);
2502 if (!page && __rmqueue_fallback(zone, order, migratetype,
2507 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2512 * Obtain a specified number of elements from the buddy allocator, all under
2513 * a single hold of the lock, for efficiency. Add them to the supplied list.
2514 * Returns the number of new pages which were placed at *list.
2516 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2517 unsigned long count, struct list_head *list,
2518 int migratetype, unsigned int alloc_flags)
2522 spin_lock(&zone->lock);
2523 for (i = 0; i < count; ++i) {
2524 struct page *page = __rmqueue(zone, order, migratetype,
2526 if (unlikely(page == NULL))
2529 if (unlikely(check_pcp_refill(page)))
2533 * Split buddy pages returned by expand() are received here in
2534 * physical page order. The page is added to the tail of
2535 * caller's list. From the callers perspective, the linked list
2536 * is ordered by page number under some conditions. This is
2537 * useful for IO devices that can forward direction from the
2538 * head, thus also in the physical page order. This is useful
2539 * for IO devices that can merge IO requests if the physical
2540 * pages are ordered properly.
2542 list_add_tail(&page->lru, list);
2544 if (is_migrate_cma(get_pcppage_migratetype(page)))
2545 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2550 * i pages were removed from the buddy list even if some leak due
2551 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2552 * on i. Do not confuse with 'alloced' which is the number of
2553 * pages added to the pcp list.
2555 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2556 spin_unlock(&zone->lock);
2562 * Called from the vmstat counter updater to drain pagesets of this
2563 * currently executing processor on remote nodes after they have
2566 * Note that this function must be called with the thread pinned to
2567 * a single processor.
2569 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2571 unsigned long flags;
2572 int to_drain, batch;
2574 local_irq_save(flags);
2575 batch = READ_ONCE(pcp->batch);
2576 to_drain = min(pcp->count, batch);
2578 free_pcppages_bulk(zone, to_drain, pcp);
2579 local_irq_restore(flags);
2584 * Drain pcplists of the indicated processor and zone.
2586 * The processor must either be the current processor and the
2587 * thread pinned to the current processor or a processor that
2590 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2592 unsigned long flags;
2593 struct per_cpu_pageset *pset;
2594 struct per_cpu_pages *pcp;
2596 local_irq_save(flags);
2597 pset = per_cpu_ptr(zone->pageset, cpu);
2601 free_pcppages_bulk(zone, pcp->count, pcp);
2602 local_irq_restore(flags);
2606 * Drain pcplists of all zones on the indicated processor.
2608 * The processor must either be the current processor and the
2609 * thread pinned to the current processor or a processor that
2612 static void drain_pages(unsigned int cpu)
2616 for_each_populated_zone(zone) {
2617 drain_pages_zone(cpu, zone);
2622 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2624 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2625 * the single zone's pages.
2627 void drain_local_pages(struct zone *zone)
2629 int cpu = smp_processor_id();
2632 drain_pages_zone(cpu, zone);
2637 static void drain_local_pages_wq(struct work_struct *work)
2640 * drain_all_pages doesn't use proper cpu hotplug protection so
2641 * we can race with cpu offline when the WQ can move this from
2642 * a cpu pinned worker to an unbound one. We can operate on a different
2643 * cpu which is allright but we also have to make sure to not move to
2647 drain_local_pages(NULL);
2652 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2654 * When zone parameter is non-NULL, spill just the single zone's pages.
2656 * Note that this can be extremely slow as the draining happens in a workqueue.
2658 void drain_all_pages(struct zone *zone)
2663 * Allocate in the BSS so we wont require allocation in
2664 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2666 static cpumask_t cpus_with_pcps;
2669 * Make sure nobody triggers this path before mm_percpu_wq is fully
2672 if (WARN_ON_ONCE(!mm_percpu_wq))
2676 * Do not drain if one is already in progress unless it's specific to
2677 * a zone. Such callers are primarily CMA and memory hotplug and need
2678 * the drain to be complete when the call returns.
2680 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2683 mutex_lock(&pcpu_drain_mutex);
2687 * We don't care about racing with CPU hotplug event
2688 * as offline notification will cause the notified
2689 * cpu to drain that CPU pcps and on_each_cpu_mask
2690 * disables preemption as part of its processing
2692 for_each_online_cpu(cpu) {
2693 struct per_cpu_pageset *pcp;
2695 bool has_pcps = false;
2698 pcp = per_cpu_ptr(zone->pageset, cpu);
2702 for_each_populated_zone(z) {
2703 pcp = per_cpu_ptr(z->pageset, cpu);
2704 if (pcp->pcp.count) {
2712 cpumask_set_cpu(cpu, &cpus_with_pcps);
2714 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2717 for_each_cpu(cpu, &cpus_with_pcps) {
2718 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2719 INIT_WORK(work, drain_local_pages_wq);
2720 queue_work_on(cpu, mm_percpu_wq, work);
2722 for_each_cpu(cpu, &cpus_with_pcps)
2723 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2725 mutex_unlock(&pcpu_drain_mutex);
2728 #ifdef CONFIG_HIBERNATION
2731 * Touch the watchdog for every WD_PAGE_COUNT pages.
2733 #define WD_PAGE_COUNT (128*1024)
2735 void mark_free_pages(struct zone *zone)
2737 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2738 unsigned long flags;
2739 unsigned int order, t;
2742 if (zone_is_empty(zone))
2745 spin_lock_irqsave(&zone->lock, flags);
2747 max_zone_pfn = zone_end_pfn(zone);
2748 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2749 if (pfn_valid(pfn)) {
2750 page = pfn_to_page(pfn);
2752 if (!--page_count) {
2753 touch_nmi_watchdog();
2754 page_count = WD_PAGE_COUNT;
2757 if (page_zone(page) != zone)
2760 if (!swsusp_page_is_forbidden(page))
2761 swsusp_unset_page_free(page);
2764 for_each_migratetype_order(order, t) {
2765 list_for_each_entry(page,
2766 &zone->free_area[order].free_list[t], lru) {
2769 pfn = page_to_pfn(page);
2770 for (i = 0; i < (1UL << order); i++) {
2771 if (!--page_count) {
2772 touch_nmi_watchdog();
2773 page_count = WD_PAGE_COUNT;
2775 swsusp_set_page_free(pfn_to_page(pfn + i));
2779 spin_unlock_irqrestore(&zone->lock, flags);
2781 #endif /* CONFIG_PM */
2783 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2787 if (!free_pcp_prepare(page))
2790 migratetype = get_pfnblock_migratetype(page, pfn);
2791 set_pcppage_migratetype(page, migratetype);
2795 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2797 struct zone *zone = page_zone(page);
2798 struct per_cpu_pages *pcp;
2801 migratetype = get_pcppage_migratetype(page);
2802 __count_vm_event(PGFREE);
2805 * We only track unmovable, reclaimable and movable on pcp lists.
2806 * Free ISOLATE pages back to the allocator because they are being
2807 * offlined but treat HIGHATOMIC as movable pages so we can get those
2808 * areas back if necessary. Otherwise, we may have to free
2809 * excessively into the page allocator
2811 if (migratetype >= MIGRATE_PCPTYPES) {
2812 if (unlikely(is_migrate_isolate(migratetype))) {
2813 free_one_page(zone, page, pfn, 0, migratetype);
2816 migratetype = MIGRATE_MOVABLE;
2819 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2820 list_add(&page->lru, &pcp->lists[migratetype]);
2822 if (pcp->count >= pcp->high) {
2823 unsigned long batch = READ_ONCE(pcp->batch);
2824 free_pcppages_bulk(zone, batch, pcp);
2829 * Free a 0-order page
2831 void free_unref_page(struct page *page)
2833 unsigned long flags;
2834 unsigned long pfn = page_to_pfn(page);
2836 if (!free_unref_page_prepare(page, pfn))
2839 local_irq_save(flags);
2840 free_unref_page_commit(page, pfn);
2841 local_irq_restore(flags);
2845 * Free a list of 0-order pages
2847 void free_unref_page_list(struct list_head *list)
2849 struct page *page, *next;
2850 unsigned long flags, pfn;
2851 int batch_count = 0;
2853 /* Prepare pages for freeing */
2854 list_for_each_entry_safe(page, next, list, lru) {
2855 pfn = page_to_pfn(page);
2856 if (!free_unref_page_prepare(page, pfn))
2857 list_del(&page->lru);
2858 set_page_private(page, pfn);
2861 local_irq_save(flags);
2862 list_for_each_entry_safe(page, next, list, lru) {
2863 unsigned long pfn = page_private(page);
2865 set_page_private(page, 0);
2866 trace_mm_page_free_batched(page);
2867 free_unref_page_commit(page, pfn);
2870 * Guard against excessive IRQ disabled times when we get
2871 * a large list of pages to free.
2873 if (++batch_count == SWAP_CLUSTER_MAX) {
2874 local_irq_restore(flags);
2876 local_irq_save(flags);
2879 local_irq_restore(flags);
2883 * split_page takes a non-compound higher-order page, and splits it into
2884 * n (1<<order) sub-pages: page[0..n]
2885 * Each sub-page must be freed individually.
2887 * Note: this is probably too low level an operation for use in drivers.
2888 * Please consult with lkml before using this in your driver.
2890 void split_page(struct page *page, unsigned int order)
2894 VM_BUG_ON_PAGE(PageCompound(page), page);
2895 VM_BUG_ON_PAGE(!page_count(page), page);
2897 for (i = 1; i < (1 << order); i++)
2898 set_page_refcounted(page + i);
2899 split_page_owner(page, order);
2901 EXPORT_SYMBOL_GPL(split_page);
2903 int __isolate_free_page(struct page *page, unsigned int order)
2905 unsigned long watermark;
2909 BUG_ON(!PageBuddy(page));
2911 zone = page_zone(page);
2912 mt = get_pageblock_migratetype(page);
2914 if (!is_migrate_isolate(mt)) {
2916 * Obey watermarks as if the page was being allocated. We can
2917 * emulate a high-order watermark check with a raised order-0
2918 * watermark, because we already know our high-order page
2921 watermark = min_wmark_pages(zone) + (1UL << order);
2922 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2925 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2928 /* Remove page from free list */
2929 list_del(&page->lru);
2930 zone->free_area[order].nr_free--;
2931 rmv_page_order(page);
2934 * Set the pageblock if the isolated page is at least half of a
2937 if (order >= pageblock_order - 1) {
2938 struct page *endpage = page + (1 << order) - 1;
2939 for (; page < endpage; page += pageblock_nr_pages) {
2940 int mt = get_pageblock_migratetype(page);
2941 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2942 && !is_migrate_highatomic(mt))
2943 set_pageblock_migratetype(page,
2949 return 1UL << order;
2953 * Update NUMA hit/miss statistics
2955 * Must be called with interrupts disabled.
2957 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2960 enum numa_stat_item local_stat = NUMA_LOCAL;
2962 /* skip numa counters update if numa stats is disabled */
2963 if (!static_branch_likely(&vm_numa_stat_key))
2966 if (zone_to_nid(z) != numa_node_id())
2967 local_stat = NUMA_OTHER;
2969 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2970 __inc_numa_state(z, NUMA_HIT);
2972 __inc_numa_state(z, NUMA_MISS);
2973 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2975 __inc_numa_state(z, local_stat);
2979 /* Remove page from the per-cpu list, caller must protect the list */
2980 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2981 unsigned int alloc_flags,
2982 struct per_cpu_pages *pcp,
2983 struct list_head *list)
2988 if (list_empty(list)) {
2989 pcp->count += rmqueue_bulk(zone, 0,
2991 migratetype, alloc_flags);
2992 if (unlikely(list_empty(list)))
2996 page = list_first_entry(list, struct page, lru);
2997 list_del(&page->lru);
2999 } while (check_new_pcp(page));
3004 /* Lock and remove page from the per-cpu list */
3005 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3006 struct zone *zone, unsigned int order,
3007 gfp_t gfp_flags, int migratetype,
3008 unsigned int alloc_flags)
3010 struct per_cpu_pages *pcp;
3011 struct list_head *list;
3013 unsigned long flags;
3015 local_irq_save(flags);
3016 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3017 list = &pcp->lists[migratetype];
3018 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3020 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3021 zone_statistics(preferred_zone, zone);
3023 local_irq_restore(flags);
3028 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3031 struct page *rmqueue(struct zone *preferred_zone,
3032 struct zone *zone, unsigned int order,
3033 gfp_t gfp_flags, unsigned int alloc_flags,
3036 unsigned long flags;
3039 if (likely(order == 0)) {
3040 page = rmqueue_pcplist(preferred_zone, zone, order,
3041 gfp_flags, migratetype, alloc_flags);
3046 * We most definitely don't want callers attempting to
3047 * allocate greater than order-1 page units with __GFP_NOFAIL.
3049 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3050 spin_lock_irqsave(&zone->lock, flags);
3054 if (alloc_flags & ALLOC_HARDER) {
3055 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3057 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3060 page = __rmqueue(zone, order, migratetype, alloc_flags);
3061 } while (page && check_new_pages(page, order));
3062 spin_unlock(&zone->lock);
3065 __mod_zone_freepage_state(zone, -(1 << order),
3066 get_pcppage_migratetype(page));
3068 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3069 zone_statistics(preferred_zone, zone);
3070 local_irq_restore(flags);
3073 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3077 local_irq_restore(flags);
3081 #ifdef CONFIG_FAIL_PAGE_ALLOC
3084 struct fault_attr attr;
3086 bool ignore_gfp_highmem;
3087 bool ignore_gfp_reclaim;
3089 } fail_page_alloc = {
3090 .attr = FAULT_ATTR_INITIALIZER,
3091 .ignore_gfp_reclaim = true,
3092 .ignore_gfp_highmem = true,
3096 static int __init setup_fail_page_alloc(char *str)
3098 return setup_fault_attr(&fail_page_alloc.attr, str);
3100 __setup("fail_page_alloc=", setup_fail_page_alloc);
3102 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3104 if (order < fail_page_alloc.min_order)
3106 if (gfp_mask & __GFP_NOFAIL)
3108 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3110 if (fail_page_alloc.ignore_gfp_reclaim &&
3111 (gfp_mask & __GFP_DIRECT_RECLAIM))
3114 return should_fail(&fail_page_alloc.attr, 1 << order);
3117 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3119 static int __init fail_page_alloc_debugfs(void)
3121 umode_t mode = S_IFREG | 0600;
3124 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3125 &fail_page_alloc.attr);
3127 return PTR_ERR(dir);
3129 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3130 &fail_page_alloc.ignore_gfp_reclaim))
3132 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3133 &fail_page_alloc.ignore_gfp_highmem))
3135 if (!debugfs_create_u32("min-order", mode, dir,
3136 &fail_page_alloc.min_order))
3141 debugfs_remove_recursive(dir);
3146 late_initcall(fail_page_alloc_debugfs);
3148 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3150 #else /* CONFIG_FAIL_PAGE_ALLOC */
3152 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3157 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3160 * Return true if free base pages are above 'mark'. For high-order checks it
3161 * will return true of the order-0 watermark is reached and there is at least
3162 * one free page of a suitable size. Checking now avoids taking the zone lock
3163 * to check in the allocation paths if no pages are free.
3165 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3166 int classzone_idx, unsigned int alloc_flags,
3171 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3173 /* free_pages may go negative - that's OK */
3174 free_pages -= (1 << order) - 1;
3176 if (alloc_flags & ALLOC_HIGH)
3180 * If the caller does not have rights to ALLOC_HARDER then subtract
3181 * the high-atomic reserves. This will over-estimate the size of the
3182 * atomic reserve but it avoids a search.
3184 if (likely(!alloc_harder)) {
3185 free_pages -= z->nr_reserved_highatomic;
3188 * OOM victims can try even harder than normal ALLOC_HARDER
3189 * users on the grounds that it's definitely going to be in
3190 * the exit path shortly and free memory. Any allocation it
3191 * makes during the free path will be small and short-lived.
3193 if (alloc_flags & ALLOC_OOM)
3201 /* If allocation can't use CMA areas don't use free CMA pages */
3202 if (!(alloc_flags & ALLOC_CMA))
3203 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3207 * Check watermarks for an order-0 allocation request. If these
3208 * are not met, then a high-order request also cannot go ahead
3209 * even if a suitable page happened to be free.
3211 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3214 /* If this is an order-0 request then the watermark is fine */
3218 /* For a high-order request, check at least one suitable page is free */
3219 for (o = order; o < MAX_ORDER; o++) {
3220 struct free_area *area = &z->free_area[o];
3226 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3227 if (!list_empty(&area->free_list[mt]))
3232 if ((alloc_flags & ALLOC_CMA) &&
3233 !list_empty(&area->free_list[MIGRATE_CMA])) {
3238 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3244 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3245 int classzone_idx, unsigned int alloc_flags)
3247 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3248 zone_page_state(z, NR_FREE_PAGES));
3251 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3252 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3254 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3258 /* If allocation can't use CMA areas don't use free CMA pages */
3259 if (!(alloc_flags & ALLOC_CMA))
3260 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3264 * Fast check for order-0 only. If this fails then the reserves
3265 * need to be calculated. There is a corner case where the check
3266 * passes but only the high-order atomic reserve are free. If
3267 * the caller is !atomic then it'll uselessly search the free
3268 * list. That corner case is then slower but it is harmless.
3270 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3273 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3277 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3278 unsigned long mark, int classzone_idx)
3280 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3282 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3283 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3285 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3290 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3292 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3295 #else /* CONFIG_NUMA */
3296 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3300 #endif /* CONFIG_NUMA */
3303 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3304 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3305 * premature use of a lower zone may cause lowmem pressure problems that
3306 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3307 * probably too small. It only makes sense to spread allocations to avoid
3308 * fragmentation between the Normal and DMA32 zones.
3310 static inline unsigned int
3311 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3313 unsigned int alloc_flags = 0;
3315 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3316 alloc_flags |= ALLOC_KSWAPD;
3318 #ifdef CONFIG_ZONE_DMA32
3319 if (zone_idx(zone) != ZONE_NORMAL)
3323 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3324 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3325 * on UMA that if Normal is populated then so is DMA32.
3327 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3328 if (nr_online_nodes > 1 && !populated_zone(--zone))
3332 #endif /* CONFIG_ZONE_DMA32 */
3337 * get_page_from_freelist goes through the zonelist trying to allocate
3340 static struct page *
3341 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3342 const struct alloc_context *ac)
3346 struct pglist_data *last_pgdat_dirty_limit = NULL;
3351 * Scan zonelist, looking for a zone with enough free.
3352 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3354 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3355 z = ac->preferred_zoneref;
3356 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3361 if (cpusets_enabled() &&
3362 (alloc_flags & ALLOC_CPUSET) &&
3363 !__cpuset_zone_allowed(zone, gfp_mask))
3366 * When allocating a page cache page for writing, we
3367 * want to get it from a node that is within its dirty
3368 * limit, such that no single node holds more than its
3369 * proportional share of globally allowed dirty pages.
3370 * The dirty limits take into account the node's
3371 * lowmem reserves and high watermark so that kswapd
3372 * should be able to balance it without having to
3373 * write pages from its LRU list.
3375 * XXX: For now, allow allocations to potentially
3376 * exceed the per-node dirty limit in the slowpath
3377 * (spread_dirty_pages unset) before going into reclaim,
3378 * which is important when on a NUMA setup the allowed
3379 * nodes are together not big enough to reach the
3380 * global limit. The proper fix for these situations
3381 * will require awareness of nodes in the
3382 * dirty-throttling and the flusher threads.
3384 if (ac->spread_dirty_pages) {
3385 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3388 if (!node_dirty_ok(zone->zone_pgdat)) {
3389 last_pgdat_dirty_limit = zone->zone_pgdat;
3394 if (no_fallback && nr_online_nodes > 1 &&
3395 zone != ac->preferred_zoneref->zone) {
3399 * If moving to a remote node, retry but allow
3400 * fragmenting fallbacks. Locality is more important
3401 * than fragmentation avoidance.
3403 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3404 if (zone_to_nid(zone) != local_nid) {
3405 alloc_flags &= ~ALLOC_NOFRAGMENT;
3410 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3411 if (!zone_watermark_fast(zone, order, mark,
3412 ac_classzone_idx(ac), alloc_flags)) {
3415 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3417 * Watermark failed for this zone, but see if we can
3418 * grow this zone if it contains deferred pages.
3420 if (static_branch_unlikely(&deferred_pages)) {
3421 if (_deferred_grow_zone(zone, order))
3425 /* Checked here to keep the fast path fast */
3426 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3427 if (alloc_flags & ALLOC_NO_WATERMARKS)
3430 if (node_reclaim_mode == 0 ||
3431 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3434 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3436 case NODE_RECLAIM_NOSCAN:
3439 case NODE_RECLAIM_FULL:
3440 /* scanned but unreclaimable */
3443 /* did we reclaim enough */
3444 if (zone_watermark_ok(zone, order, mark,
3445 ac_classzone_idx(ac), alloc_flags))
3453 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3454 gfp_mask, alloc_flags, ac->migratetype);
3456 prep_new_page(page, order, gfp_mask, alloc_flags);
3459 * If this is a high-order atomic allocation then check
3460 * if the pageblock should be reserved for the future
3462 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3463 reserve_highatomic_pageblock(page, zone, order);
3467 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3468 /* Try again if zone has deferred pages */
3469 if (static_branch_unlikely(&deferred_pages)) {
3470 if (_deferred_grow_zone(zone, order))
3478 * It's possible on a UMA machine to get through all zones that are
3479 * fragmented. If avoiding fragmentation, reset and try again.
3482 alloc_flags &= ~ALLOC_NOFRAGMENT;
3489 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3491 unsigned int filter = SHOW_MEM_FILTER_NODES;
3492 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3494 if (!__ratelimit(&show_mem_rs))
3498 * This documents exceptions given to allocations in certain
3499 * contexts that are allowed to allocate outside current's set
3502 if (!(gfp_mask & __GFP_NOMEMALLOC))
3503 if (tsk_is_oom_victim(current) ||
3504 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3505 filter &= ~SHOW_MEM_FILTER_NODES;
3506 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3507 filter &= ~SHOW_MEM_FILTER_NODES;
3509 show_mem(filter, nodemask);
3512 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3514 struct va_format vaf;
3516 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3517 DEFAULT_RATELIMIT_BURST);
3519 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3522 va_start(args, fmt);
3525 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3526 current->comm, &vaf, gfp_mask, &gfp_mask,
3527 nodemask_pr_args(nodemask));
3530 cpuset_print_current_mems_allowed();
3533 warn_alloc_show_mem(gfp_mask, nodemask);
3536 static inline struct page *
3537 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3538 unsigned int alloc_flags,
3539 const struct alloc_context *ac)
3543 page = get_page_from_freelist(gfp_mask, order,
3544 alloc_flags|ALLOC_CPUSET, ac);
3546 * fallback to ignore cpuset restriction if our nodes
3550 page = get_page_from_freelist(gfp_mask, order,
3556 static inline struct page *
3557 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3558 const struct alloc_context *ac, unsigned long *did_some_progress)
3560 struct oom_control oc = {
3561 .zonelist = ac->zonelist,
3562 .nodemask = ac->nodemask,
3564 .gfp_mask = gfp_mask,
3569 *did_some_progress = 0;
3572 * Acquire the oom lock. If that fails, somebody else is
3573 * making progress for us.
3575 if (!mutex_trylock(&oom_lock)) {
3576 *did_some_progress = 1;
3577 schedule_timeout_uninterruptible(1);
3582 * Go through the zonelist yet one more time, keep very high watermark
3583 * here, this is only to catch a parallel oom killing, we must fail if
3584 * we're still under heavy pressure. But make sure that this reclaim
3585 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3586 * allocation which will never fail due to oom_lock already held.
3588 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3589 ~__GFP_DIRECT_RECLAIM, order,
3590 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3594 /* Coredumps can quickly deplete all memory reserves */
3595 if (current->flags & PF_DUMPCORE)
3597 /* The OOM killer will not help higher order allocs */
3598 if (order > PAGE_ALLOC_COSTLY_ORDER)
3601 * We have already exhausted all our reclaim opportunities without any
3602 * success so it is time to admit defeat. We will skip the OOM killer
3603 * because it is very likely that the caller has a more reasonable
3604 * fallback than shooting a random task.
3606 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3608 /* The OOM killer does not needlessly kill tasks for lowmem */
3609 if (ac->high_zoneidx < ZONE_NORMAL)
3611 if (pm_suspended_storage())
3614 * XXX: GFP_NOFS allocations should rather fail than rely on
3615 * other request to make a forward progress.
3616 * We are in an unfortunate situation where out_of_memory cannot
3617 * do much for this context but let's try it to at least get
3618 * access to memory reserved if the current task is killed (see
3619 * out_of_memory). Once filesystems are ready to handle allocation
3620 * failures more gracefully we should just bail out here.
3623 /* The OOM killer may not free memory on a specific node */
3624 if (gfp_mask & __GFP_THISNODE)
3627 /* Exhausted what can be done so it's blame time */
3628 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3629 *did_some_progress = 1;
3632 * Help non-failing allocations by giving them access to memory
3635 if (gfp_mask & __GFP_NOFAIL)
3636 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3637 ALLOC_NO_WATERMARKS, ac);
3640 mutex_unlock(&oom_lock);
3645 * Maximum number of compaction retries wit a progress before OOM
3646 * killer is consider as the only way to move forward.
3648 #define MAX_COMPACT_RETRIES 16
3650 #ifdef CONFIG_COMPACTION
3651 /* Try memory compaction for high-order allocations before reclaim */
3652 static struct page *
3653 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3654 unsigned int alloc_flags, const struct alloc_context *ac,
3655 enum compact_priority prio, enum compact_result *compact_result)
3658 unsigned long pflags;
3659 unsigned int noreclaim_flag;
3664 psi_memstall_enter(&pflags);
3665 noreclaim_flag = memalloc_noreclaim_save();
3667 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3670 memalloc_noreclaim_restore(noreclaim_flag);
3671 psi_memstall_leave(&pflags);
3673 if (*compact_result <= COMPACT_INACTIVE)
3677 * At least in one zone compaction wasn't deferred or skipped, so let's
3678 * count a compaction stall
3680 count_vm_event(COMPACTSTALL);
3682 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3685 struct zone *zone = page_zone(page);
3687 zone->compact_blockskip_flush = false;
3688 compaction_defer_reset(zone, order, true);
3689 count_vm_event(COMPACTSUCCESS);
3694 * It's bad if compaction run occurs and fails. The most likely reason
3695 * is that pages exist, but not enough to satisfy watermarks.
3697 count_vm_event(COMPACTFAIL);
3705 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3706 enum compact_result compact_result,
3707 enum compact_priority *compact_priority,
3708 int *compaction_retries)
3710 int max_retries = MAX_COMPACT_RETRIES;
3713 int retries = *compaction_retries;
3714 enum compact_priority priority = *compact_priority;
3719 if (compaction_made_progress(compact_result))
3720 (*compaction_retries)++;
3723 * compaction considers all the zone as desperately out of memory
3724 * so it doesn't really make much sense to retry except when the
3725 * failure could be caused by insufficient priority
3727 if (compaction_failed(compact_result))
3728 goto check_priority;
3731 * make sure the compaction wasn't deferred or didn't bail out early
3732 * due to locks contention before we declare that we should give up.
3733 * But do not retry if the given zonelist is not suitable for
3736 if (compaction_withdrawn(compact_result)) {
3737 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3742 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3743 * costly ones because they are de facto nofail and invoke OOM
3744 * killer to move on while costly can fail and users are ready
3745 * to cope with that. 1/4 retries is rather arbitrary but we
3746 * would need much more detailed feedback from compaction to
3747 * make a better decision.
3749 if (order > PAGE_ALLOC_COSTLY_ORDER)
3751 if (*compaction_retries <= max_retries) {
3757 * Make sure there are attempts at the highest priority if we exhausted
3758 * all retries or failed at the lower priorities.
3761 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3762 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3764 if (*compact_priority > min_priority) {
3765 (*compact_priority)--;
3766 *compaction_retries = 0;
3770 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3774 static inline struct page *
3775 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3776 unsigned int alloc_flags, const struct alloc_context *ac,
3777 enum compact_priority prio, enum compact_result *compact_result)
3779 *compact_result = COMPACT_SKIPPED;
3784 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3785 enum compact_result compact_result,
3786 enum compact_priority *compact_priority,
3787 int *compaction_retries)
3792 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3796 * There are setups with compaction disabled which would prefer to loop
3797 * inside the allocator rather than hit the oom killer prematurely.
3798 * Let's give them a good hope and keep retrying while the order-0
3799 * watermarks are OK.
3801 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3803 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3804 ac_classzone_idx(ac), alloc_flags))
3809 #endif /* CONFIG_COMPACTION */
3811 #ifdef CONFIG_LOCKDEP
3812 static struct lockdep_map __fs_reclaim_map =
3813 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3815 static bool __need_fs_reclaim(gfp_t gfp_mask)
3817 gfp_mask = current_gfp_context(gfp_mask);
3819 /* no reclaim without waiting on it */
3820 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3823 /* this guy won't enter reclaim */
3824 if (current->flags & PF_MEMALLOC)
3827 /* We're only interested __GFP_FS allocations for now */
3828 if (!(gfp_mask & __GFP_FS))
3831 if (gfp_mask & __GFP_NOLOCKDEP)
3837 void __fs_reclaim_acquire(void)
3839 lock_map_acquire(&__fs_reclaim_map);
3842 void __fs_reclaim_release(void)
3844 lock_map_release(&__fs_reclaim_map);
3847 void fs_reclaim_acquire(gfp_t gfp_mask)
3849 if (__need_fs_reclaim(gfp_mask))
3850 __fs_reclaim_acquire();
3852 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3854 void fs_reclaim_release(gfp_t gfp_mask)
3856 if (__need_fs_reclaim(gfp_mask))
3857 __fs_reclaim_release();
3859 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3862 /* Perform direct synchronous page reclaim */
3864 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3865 const struct alloc_context *ac)
3867 struct reclaim_state reclaim_state;
3869 unsigned int noreclaim_flag;
3870 unsigned long pflags;
3874 /* We now go into synchronous reclaim */
3875 cpuset_memory_pressure_bump();
3876 psi_memstall_enter(&pflags);
3877 fs_reclaim_acquire(gfp_mask);
3878 noreclaim_flag = memalloc_noreclaim_save();
3879 reclaim_state.reclaimed_slab = 0;
3880 current->reclaim_state = &reclaim_state;
3882 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3885 current->reclaim_state = NULL;
3886 memalloc_noreclaim_restore(noreclaim_flag);
3887 fs_reclaim_release(gfp_mask);
3888 psi_memstall_leave(&pflags);
3895 /* The really slow allocator path where we enter direct reclaim */
3896 static inline struct page *
3897 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3898 unsigned int alloc_flags, const struct alloc_context *ac,
3899 unsigned long *did_some_progress)
3901 struct page *page = NULL;
3902 bool drained = false;
3904 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3905 if (unlikely(!(*did_some_progress)))
3909 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3912 * If an allocation failed after direct reclaim, it could be because
3913 * pages are pinned on the per-cpu lists or in high alloc reserves.
3914 * Shrink them them and try again
3916 if (!page && !drained) {
3917 unreserve_highatomic_pageblock(ac, false);
3918 drain_all_pages(NULL);
3926 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3927 const struct alloc_context *ac)
3931 pg_data_t *last_pgdat = NULL;
3932 enum zone_type high_zoneidx = ac->high_zoneidx;
3934 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3936 if (last_pgdat != zone->zone_pgdat)
3937 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3938 last_pgdat = zone->zone_pgdat;
3942 static inline unsigned int
3943 gfp_to_alloc_flags(gfp_t gfp_mask)
3945 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3947 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3948 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3951 * The caller may dip into page reserves a bit more if the caller
3952 * cannot run direct reclaim, or if the caller has realtime scheduling
3953 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3954 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3956 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3958 if (gfp_mask & __GFP_ATOMIC) {
3960 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3961 * if it can't schedule.
3963 if (!(gfp_mask & __GFP_NOMEMALLOC))
3964 alloc_flags |= ALLOC_HARDER;
3966 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3967 * comment for __cpuset_node_allowed().
3969 alloc_flags &= ~ALLOC_CPUSET;
3970 } else if (unlikely(rt_task(current)) && !in_interrupt())
3971 alloc_flags |= ALLOC_HARDER;
3973 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3974 alloc_flags |= ALLOC_KSWAPD;
3977 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3978 alloc_flags |= ALLOC_CMA;
3983 static bool oom_reserves_allowed(struct task_struct *tsk)
3985 if (!tsk_is_oom_victim(tsk))
3989 * !MMU doesn't have oom reaper so give access to memory reserves
3990 * only to the thread with TIF_MEMDIE set
3992 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3999 * Distinguish requests which really need access to full memory
4000 * reserves from oom victims which can live with a portion of it
4002 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4004 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4006 if (gfp_mask & __GFP_MEMALLOC)
4007 return ALLOC_NO_WATERMARKS;
4008 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4009 return ALLOC_NO_WATERMARKS;
4010 if (!in_interrupt()) {
4011 if (current->flags & PF_MEMALLOC)
4012 return ALLOC_NO_WATERMARKS;
4013 else if (oom_reserves_allowed(current))
4020 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4022 return !!__gfp_pfmemalloc_flags(gfp_mask);
4026 * Checks whether it makes sense to retry the reclaim to make a forward progress
4027 * for the given allocation request.
4029 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4030 * without success, or when we couldn't even meet the watermark if we
4031 * reclaimed all remaining pages on the LRU lists.
4033 * Returns true if a retry is viable or false to enter the oom path.
4036 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4037 struct alloc_context *ac, int alloc_flags,
4038 bool did_some_progress, int *no_progress_loops)
4045 * Costly allocations might have made a progress but this doesn't mean
4046 * their order will become available due to high fragmentation so
4047 * always increment the no progress counter for them
4049 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4050 *no_progress_loops = 0;
4052 (*no_progress_loops)++;
4055 * Make sure we converge to OOM if we cannot make any progress
4056 * several times in the row.
4058 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4059 /* Before OOM, exhaust highatomic_reserve */
4060 return unreserve_highatomic_pageblock(ac, true);
4064 * Keep reclaiming pages while there is a chance this will lead
4065 * somewhere. If none of the target zones can satisfy our allocation
4066 * request even if all reclaimable pages are considered then we are
4067 * screwed and have to go OOM.
4069 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4071 unsigned long available;
4072 unsigned long reclaimable;
4073 unsigned long min_wmark = min_wmark_pages(zone);
4076 available = reclaimable = zone_reclaimable_pages(zone);
4077 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4080 * Would the allocation succeed if we reclaimed all
4081 * reclaimable pages?
4083 wmark = __zone_watermark_ok(zone, order, min_wmark,
4084 ac_classzone_idx(ac), alloc_flags, available);
4085 trace_reclaim_retry_zone(z, order, reclaimable,
4086 available, min_wmark, *no_progress_loops, wmark);
4089 * If we didn't make any progress and have a lot of
4090 * dirty + writeback pages then we should wait for
4091 * an IO to complete to slow down the reclaim and
4092 * prevent from pre mature OOM
4094 if (!did_some_progress) {
4095 unsigned long write_pending;
4097 write_pending = zone_page_state_snapshot(zone,
4098 NR_ZONE_WRITE_PENDING);
4100 if (2 * write_pending > reclaimable) {
4101 congestion_wait(BLK_RW_ASYNC, HZ/10);
4113 * Memory allocation/reclaim might be called from a WQ context and the
4114 * current implementation of the WQ concurrency control doesn't
4115 * recognize that a particular WQ is congested if the worker thread is
4116 * looping without ever sleeping. Therefore we have to do a short sleep
4117 * here rather than calling cond_resched().
4119 if (current->flags & PF_WQ_WORKER)
4120 schedule_timeout_uninterruptible(1);
4127 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4130 * It's possible that cpuset's mems_allowed and the nodemask from
4131 * mempolicy don't intersect. This should be normally dealt with by
4132 * policy_nodemask(), but it's possible to race with cpuset update in
4133 * such a way the check therein was true, and then it became false
4134 * before we got our cpuset_mems_cookie here.
4135 * This assumes that for all allocations, ac->nodemask can come only
4136 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4137 * when it does not intersect with the cpuset restrictions) or the
4138 * caller can deal with a violated nodemask.
4140 if (cpusets_enabled() && ac->nodemask &&
4141 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4142 ac->nodemask = NULL;
4147 * When updating a task's mems_allowed or mempolicy nodemask, it is
4148 * possible to race with parallel threads in such a way that our
4149 * allocation can fail while the mask is being updated. If we are about
4150 * to fail, check if the cpuset changed during allocation and if so,
4153 if (read_mems_allowed_retry(cpuset_mems_cookie))
4159 static inline struct page *
4160 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4161 struct alloc_context *ac)
4163 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4164 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4165 struct page *page = NULL;
4166 unsigned int alloc_flags;
4167 unsigned long did_some_progress;
4168 enum compact_priority compact_priority;
4169 enum compact_result compact_result;
4170 int compaction_retries;
4171 int no_progress_loops;
4172 unsigned int cpuset_mems_cookie;
4176 * We also sanity check to catch abuse of atomic reserves being used by
4177 * callers that are not in atomic context.
4179 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4180 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4181 gfp_mask &= ~__GFP_ATOMIC;
4184 compaction_retries = 0;
4185 no_progress_loops = 0;
4186 compact_priority = DEF_COMPACT_PRIORITY;
4187 cpuset_mems_cookie = read_mems_allowed_begin();
4190 * The fast path uses conservative alloc_flags to succeed only until
4191 * kswapd needs to be woken up, and to avoid the cost of setting up
4192 * alloc_flags precisely. So we do that now.
4194 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4197 * We need to recalculate the starting point for the zonelist iterator
4198 * because we might have used different nodemask in the fast path, or
4199 * there was a cpuset modification and we are retrying - otherwise we
4200 * could end up iterating over non-eligible zones endlessly.
4202 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4203 ac->high_zoneidx, ac->nodemask);
4204 if (!ac->preferred_zoneref->zone)
4207 if (alloc_flags & ALLOC_KSWAPD)
4208 wake_all_kswapds(order, gfp_mask, ac);
4211 * The adjusted alloc_flags might result in immediate success, so try
4214 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4219 * For costly allocations, try direct compaction first, as it's likely
4220 * that we have enough base pages and don't need to reclaim. For non-
4221 * movable high-order allocations, do that as well, as compaction will
4222 * try prevent permanent fragmentation by migrating from blocks of the
4224 * Don't try this for allocations that are allowed to ignore
4225 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4227 if (can_direct_reclaim &&
4229 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4230 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4231 page = __alloc_pages_direct_compact(gfp_mask, order,
4233 INIT_COMPACT_PRIORITY,
4239 * Checks for costly allocations with __GFP_NORETRY, which
4240 * includes THP page fault allocations
4242 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4244 * If compaction is deferred for high-order allocations,
4245 * it is because sync compaction recently failed. If
4246 * this is the case and the caller requested a THP
4247 * allocation, we do not want to heavily disrupt the
4248 * system, so we fail the allocation instead of entering
4251 if (compact_result == COMPACT_DEFERRED)
4255 * Looks like reclaim/compaction is worth trying, but
4256 * sync compaction could be very expensive, so keep
4257 * using async compaction.
4259 compact_priority = INIT_COMPACT_PRIORITY;
4264 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4265 if (alloc_flags & ALLOC_KSWAPD)
4266 wake_all_kswapds(order, gfp_mask, ac);
4268 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4270 alloc_flags = reserve_flags;
4273 * Reset the nodemask and zonelist iterators if memory policies can be
4274 * ignored. These allocations are high priority and system rather than
4277 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4278 ac->nodemask = NULL;
4279 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4280 ac->high_zoneidx, ac->nodemask);
4283 /* Attempt with potentially adjusted zonelist and alloc_flags */
4284 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4288 /* Caller is not willing to reclaim, we can't balance anything */
4289 if (!can_direct_reclaim)
4292 /* Avoid recursion of direct reclaim */
4293 if (current->flags & PF_MEMALLOC)
4296 /* Try direct reclaim and then allocating */
4297 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4298 &did_some_progress);
4302 /* Try direct compaction and then allocating */
4303 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4304 compact_priority, &compact_result);
4308 /* Do not loop if specifically requested */
4309 if (gfp_mask & __GFP_NORETRY)
4313 * Do not retry costly high order allocations unless they are
4314 * __GFP_RETRY_MAYFAIL
4316 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4319 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4320 did_some_progress > 0, &no_progress_loops))
4324 * It doesn't make any sense to retry for the compaction if the order-0
4325 * reclaim is not able to make any progress because the current
4326 * implementation of the compaction depends on the sufficient amount
4327 * of free memory (see __compaction_suitable)
4329 if (did_some_progress > 0 &&
4330 should_compact_retry(ac, order, alloc_flags,
4331 compact_result, &compact_priority,
4332 &compaction_retries))
4336 /* Deal with possible cpuset update races before we start OOM killing */
4337 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4340 /* Reclaim has failed us, start killing things */
4341 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4345 /* Avoid allocations with no watermarks from looping endlessly */
4346 if (tsk_is_oom_victim(current) &&
4347 (alloc_flags == ALLOC_OOM ||
4348 (gfp_mask & __GFP_NOMEMALLOC)))
4351 /* Retry as long as the OOM killer is making progress */
4352 if (did_some_progress) {
4353 no_progress_loops = 0;
4358 /* Deal with possible cpuset update races before we fail */
4359 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4363 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4366 if (gfp_mask & __GFP_NOFAIL) {
4368 * All existing users of the __GFP_NOFAIL are blockable, so warn
4369 * of any new users that actually require GFP_NOWAIT
4371 if (WARN_ON_ONCE(!can_direct_reclaim))
4375 * PF_MEMALLOC request from this context is rather bizarre
4376 * because we cannot reclaim anything and only can loop waiting
4377 * for somebody to do a work for us
4379 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4382 * non failing costly orders are a hard requirement which we
4383 * are not prepared for much so let's warn about these users
4384 * so that we can identify them and convert them to something
4387 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4390 * Help non-failing allocations by giving them access to memory
4391 * reserves but do not use ALLOC_NO_WATERMARKS because this
4392 * could deplete whole memory reserves which would just make
4393 * the situation worse
4395 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4403 warn_alloc(gfp_mask, ac->nodemask,
4404 "page allocation failure: order:%u", order);
4409 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4410 int preferred_nid, nodemask_t *nodemask,
4411 struct alloc_context *ac, gfp_t *alloc_mask,
4412 unsigned int *alloc_flags)
4414 ac->high_zoneidx = gfp_zone(gfp_mask);
4415 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4416 ac->nodemask = nodemask;
4417 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4419 if (cpusets_enabled()) {
4420 *alloc_mask |= __GFP_HARDWALL;
4422 ac->nodemask = &cpuset_current_mems_allowed;
4424 *alloc_flags |= ALLOC_CPUSET;
4427 fs_reclaim_acquire(gfp_mask);
4428 fs_reclaim_release(gfp_mask);
4430 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4432 if (should_fail_alloc_page(gfp_mask, order))
4435 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4436 *alloc_flags |= ALLOC_CMA;
4441 /* Determine whether to spread dirty pages and what the first usable zone */
4442 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4444 /* Dirty zone balancing only done in the fast path */
4445 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4448 * The preferred zone is used for statistics but crucially it is
4449 * also used as the starting point for the zonelist iterator. It
4450 * may get reset for allocations that ignore memory policies.
4452 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4453 ac->high_zoneidx, ac->nodemask);
4457 * This is the 'heart' of the zoned buddy allocator.
4460 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4461 nodemask_t *nodemask)
4464 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4465 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4466 struct alloc_context ac = { };
4469 * There are several places where we assume that the order value is sane
4470 * so bail out early if the request is out of bound.
4472 if (unlikely(order >= MAX_ORDER)) {
4473 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4477 gfp_mask &= gfp_allowed_mask;
4478 alloc_mask = gfp_mask;
4479 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4482 finalise_ac(gfp_mask, &ac);
4485 * Forbid the first pass from falling back to types that fragment
4486 * memory until all local zones are considered.
4488 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4490 /* First allocation attempt */
4491 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4496 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4497 * resp. GFP_NOIO which has to be inherited for all allocation requests
4498 * from a particular context which has been marked by
4499 * memalloc_no{fs,io}_{save,restore}.
4501 alloc_mask = current_gfp_context(gfp_mask);
4502 ac.spread_dirty_pages = false;
4505 * Restore the original nodemask if it was potentially replaced with
4506 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4508 if (unlikely(ac.nodemask != nodemask))
4509 ac.nodemask = nodemask;
4511 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4514 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4515 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4516 __free_pages(page, order);
4520 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4524 EXPORT_SYMBOL(__alloc_pages_nodemask);
4527 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4528 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4529 * you need to access high mem.
4531 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4535 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4538 return (unsigned long) page_address(page);
4540 EXPORT_SYMBOL(__get_free_pages);
4542 unsigned long get_zeroed_page(gfp_t gfp_mask)
4544 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4546 EXPORT_SYMBOL(get_zeroed_page);
4548 static inline void free_the_page(struct page *page, unsigned int order)
4550 if (order == 0) /* Via pcp? */
4551 free_unref_page(page);
4553 __free_pages_ok(page, order);
4556 void __free_pages(struct page *page, unsigned int order)
4558 if (put_page_testzero(page))
4559 free_the_page(page, order);
4561 EXPORT_SYMBOL(__free_pages);
4563 void free_pages(unsigned long addr, unsigned int order)
4566 VM_BUG_ON(!virt_addr_valid((void *)addr));
4567 __free_pages(virt_to_page((void *)addr), order);
4571 EXPORT_SYMBOL(free_pages);
4575 * An arbitrary-length arbitrary-offset area of memory which resides
4576 * within a 0 or higher order page. Multiple fragments within that page
4577 * are individually refcounted, in the page's reference counter.
4579 * The page_frag functions below provide a simple allocation framework for
4580 * page fragments. This is used by the network stack and network device
4581 * drivers to provide a backing region of memory for use as either an
4582 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4584 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4587 struct page *page = NULL;
4588 gfp_t gfp = gfp_mask;
4590 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4591 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4593 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4594 PAGE_FRAG_CACHE_MAX_ORDER);
4595 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4597 if (unlikely(!page))
4598 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4600 nc->va = page ? page_address(page) : NULL;
4605 void __page_frag_cache_drain(struct page *page, unsigned int count)
4607 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4609 if (page_ref_sub_and_test(page, count))
4610 free_the_page(page, compound_order(page));
4612 EXPORT_SYMBOL(__page_frag_cache_drain);
4614 void *page_frag_alloc(struct page_frag_cache *nc,
4615 unsigned int fragsz, gfp_t gfp_mask)
4617 unsigned int size = PAGE_SIZE;
4621 if (unlikely(!nc->va)) {
4623 page = __page_frag_cache_refill(nc, gfp_mask);
4627 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4628 /* if size can vary use size else just use PAGE_SIZE */
4631 /* Even if we own the page, we do not use atomic_set().
4632 * This would break get_page_unless_zero() users.
4634 page_ref_add(page, size - 1);
4636 /* reset page count bias and offset to start of new frag */
4637 nc->pfmemalloc = page_is_pfmemalloc(page);
4638 nc->pagecnt_bias = size;
4642 offset = nc->offset - fragsz;
4643 if (unlikely(offset < 0)) {
4644 page = virt_to_page(nc->va);
4646 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4649 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4650 /* if size can vary use size else just use PAGE_SIZE */
4653 /* OK, page count is 0, we can safely set it */
4654 set_page_count(page, size);
4656 /* reset page count bias and offset to start of new frag */
4657 nc->pagecnt_bias = size;
4658 offset = size - fragsz;
4662 nc->offset = offset;
4664 return nc->va + offset;
4666 EXPORT_SYMBOL(page_frag_alloc);
4669 * Frees a page fragment allocated out of either a compound or order 0 page.
4671 void page_frag_free(void *addr)
4673 struct page *page = virt_to_head_page(addr);
4675 if (unlikely(put_page_testzero(page)))
4676 free_the_page(page, compound_order(page));
4678 EXPORT_SYMBOL(page_frag_free);
4680 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4684 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4685 unsigned long used = addr + PAGE_ALIGN(size);
4687 split_page(virt_to_page((void *)addr), order);
4688 while (used < alloc_end) {
4693 return (void *)addr;
4697 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4698 * @size: the number of bytes to allocate
4699 * @gfp_mask: GFP flags for the allocation
4701 * This function is similar to alloc_pages(), except that it allocates the
4702 * minimum number of pages to satisfy the request. alloc_pages() can only
4703 * allocate memory in power-of-two pages.
4705 * This function is also limited by MAX_ORDER.
4707 * Memory allocated by this function must be released by free_pages_exact().
4709 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4711 unsigned int order = get_order(size);
4714 addr = __get_free_pages(gfp_mask, order);
4715 return make_alloc_exact(addr, order, size);
4717 EXPORT_SYMBOL(alloc_pages_exact);
4720 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4722 * @nid: the preferred node ID where memory should be allocated
4723 * @size: the number of bytes to allocate
4724 * @gfp_mask: GFP flags for the allocation
4726 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4729 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4731 unsigned int order = get_order(size);
4732 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4735 return make_alloc_exact((unsigned long)page_address(p), order, size);
4739 * free_pages_exact - release memory allocated via alloc_pages_exact()
4740 * @virt: the value returned by alloc_pages_exact.
4741 * @size: size of allocation, same value as passed to alloc_pages_exact().
4743 * Release the memory allocated by a previous call to alloc_pages_exact.
4745 void free_pages_exact(void *virt, size_t size)
4747 unsigned long addr = (unsigned long)virt;
4748 unsigned long end = addr + PAGE_ALIGN(size);
4750 while (addr < end) {
4755 EXPORT_SYMBOL(free_pages_exact);
4758 * nr_free_zone_pages - count number of pages beyond high watermark
4759 * @offset: The zone index of the highest zone
4761 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4762 * high watermark within all zones at or below a given zone index. For each
4763 * zone, the number of pages is calculated as:
4765 * nr_free_zone_pages = managed_pages - high_pages
4767 static unsigned long nr_free_zone_pages(int offset)
4772 /* Just pick one node, since fallback list is circular */
4773 unsigned long sum = 0;
4775 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4777 for_each_zone_zonelist(zone, z, zonelist, offset) {
4778 unsigned long size = zone_managed_pages(zone);
4779 unsigned long high = high_wmark_pages(zone);
4788 * nr_free_buffer_pages - count number of pages beyond high watermark
4790 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4791 * watermark within ZONE_DMA and ZONE_NORMAL.
4793 unsigned long nr_free_buffer_pages(void)
4795 return nr_free_zone_pages(gfp_zone(GFP_USER));
4797 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4800 * nr_free_pagecache_pages - count number of pages beyond high watermark
4802 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4803 * high watermark within all zones.
4805 unsigned long nr_free_pagecache_pages(void)
4807 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4810 static inline void show_node(struct zone *zone)
4812 if (IS_ENABLED(CONFIG_NUMA))
4813 printk("Node %d ", zone_to_nid(zone));
4816 long si_mem_available(void)
4819 unsigned long pagecache;
4820 unsigned long wmark_low = 0;
4821 unsigned long pages[NR_LRU_LISTS];
4822 unsigned long reclaimable;
4826 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4827 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4830 wmark_low += low_wmark_pages(zone);
4833 * Estimate the amount of memory available for userspace allocations,
4834 * without causing swapping.
4836 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4839 * Not all the page cache can be freed, otherwise the system will
4840 * start swapping. Assume at least half of the page cache, or the
4841 * low watermark worth of cache, needs to stay.
4843 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4844 pagecache -= min(pagecache / 2, wmark_low);
4845 available += pagecache;
4848 * Part of the reclaimable slab and other kernel memory consists of
4849 * items that are in use, and cannot be freed. Cap this estimate at the
4852 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
4853 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
4854 available += reclaimable - min(reclaimable / 2, wmark_low);
4860 EXPORT_SYMBOL_GPL(si_mem_available);
4862 void si_meminfo(struct sysinfo *val)
4864 val->totalram = totalram_pages();
4865 val->sharedram = global_node_page_state(NR_SHMEM);
4866 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4867 val->bufferram = nr_blockdev_pages();
4868 val->totalhigh = totalhigh_pages();
4869 val->freehigh = nr_free_highpages();
4870 val->mem_unit = PAGE_SIZE;
4873 EXPORT_SYMBOL(si_meminfo);
4876 void si_meminfo_node(struct sysinfo *val, int nid)
4878 int zone_type; /* needs to be signed */
4879 unsigned long managed_pages = 0;
4880 unsigned long managed_highpages = 0;
4881 unsigned long free_highpages = 0;
4882 pg_data_t *pgdat = NODE_DATA(nid);
4884 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4885 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
4886 val->totalram = managed_pages;
4887 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4888 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4889 #ifdef CONFIG_HIGHMEM
4890 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4891 struct zone *zone = &pgdat->node_zones[zone_type];
4893 if (is_highmem(zone)) {
4894 managed_highpages += zone_managed_pages(zone);
4895 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4898 val->totalhigh = managed_highpages;
4899 val->freehigh = free_highpages;
4901 val->totalhigh = managed_highpages;
4902 val->freehigh = free_highpages;
4904 val->mem_unit = PAGE_SIZE;
4909 * Determine whether the node should be displayed or not, depending on whether
4910 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4912 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4914 if (!(flags & SHOW_MEM_FILTER_NODES))
4918 * no node mask - aka implicit memory numa policy. Do not bother with
4919 * the synchronization - read_mems_allowed_begin - because we do not
4920 * have to be precise here.
4923 nodemask = &cpuset_current_mems_allowed;
4925 return !node_isset(nid, *nodemask);
4928 #define K(x) ((x) << (PAGE_SHIFT-10))
4930 static void show_migration_types(unsigned char type)
4932 static const char types[MIGRATE_TYPES] = {
4933 [MIGRATE_UNMOVABLE] = 'U',
4934 [MIGRATE_MOVABLE] = 'M',
4935 [MIGRATE_RECLAIMABLE] = 'E',
4936 [MIGRATE_HIGHATOMIC] = 'H',
4938 [MIGRATE_CMA] = 'C',
4940 #ifdef CONFIG_MEMORY_ISOLATION
4941 [MIGRATE_ISOLATE] = 'I',
4944 char tmp[MIGRATE_TYPES + 1];
4948 for (i = 0; i < MIGRATE_TYPES; i++) {
4949 if (type & (1 << i))
4954 printk(KERN_CONT "(%s) ", tmp);
4958 * Show free area list (used inside shift_scroll-lock stuff)
4959 * We also calculate the percentage fragmentation. We do this by counting the
4960 * memory on each free list with the exception of the first item on the list.
4963 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4966 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4968 unsigned long free_pcp = 0;
4973 for_each_populated_zone(zone) {
4974 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4977 for_each_online_cpu(cpu)
4978 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4981 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4982 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4983 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4984 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4985 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4986 " free:%lu free_pcp:%lu free_cma:%lu\n",
4987 global_node_page_state(NR_ACTIVE_ANON),
4988 global_node_page_state(NR_INACTIVE_ANON),
4989 global_node_page_state(NR_ISOLATED_ANON),
4990 global_node_page_state(NR_ACTIVE_FILE),
4991 global_node_page_state(NR_INACTIVE_FILE),
4992 global_node_page_state(NR_ISOLATED_FILE),
4993 global_node_page_state(NR_UNEVICTABLE),
4994 global_node_page_state(NR_FILE_DIRTY),
4995 global_node_page_state(NR_WRITEBACK),
4996 global_node_page_state(NR_UNSTABLE_NFS),
4997 global_node_page_state(NR_SLAB_RECLAIMABLE),
4998 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4999 global_node_page_state(NR_FILE_MAPPED),
5000 global_node_page_state(NR_SHMEM),
5001 global_zone_page_state(NR_PAGETABLE),
5002 global_zone_page_state(NR_BOUNCE),
5003 global_zone_page_state(NR_FREE_PAGES),
5005 global_zone_page_state(NR_FREE_CMA_PAGES));
5007 for_each_online_pgdat(pgdat) {
5008 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5012 " active_anon:%lukB"
5013 " inactive_anon:%lukB"
5014 " active_file:%lukB"
5015 " inactive_file:%lukB"
5016 " unevictable:%lukB"
5017 " isolated(anon):%lukB"
5018 " isolated(file):%lukB"
5023 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5025 " shmem_pmdmapped: %lukB"
5028 " writeback_tmp:%lukB"
5030 " all_unreclaimable? %s"
5033 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5034 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5035 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5036 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5037 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5038 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5039 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5040 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5041 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5042 K(node_page_state(pgdat, NR_WRITEBACK)),
5043 K(node_page_state(pgdat, NR_SHMEM)),
5044 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5045 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5046 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5048 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5050 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5051 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5052 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5056 for_each_populated_zone(zone) {
5059 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5063 for_each_online_cpu(cpu)
5064 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5073 " active_anon:%lukB"
5074 " inactive_anon:%lukB"
5075 " active_file:%lukB"
5076 " inactive_file:%lukB"
5077 " unevictable:%lukB"
5078 " writepending:%lukB"
5082 " kernel_stack:%lukB"
5090 K(zone_page_state(zone, NR_FREE_PAGES)),
5091 K(min_wmark_pages(zone)),
5092 K(low_wmark_pages(zone)),
5093 K(high_wmark_pages(zone)),
5094 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5095 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5096 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5097 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5098 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5099 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5100 K(zone->present_pages),
5101 K(zone_managed_pages(zone)),
5102 K(zone_page_state(zone, NR_MLOCK)),
5103 zone_page_state(zone, NR_KERNEL_STACK_KB),
5104 K(zone_page_state(zone, NR_PAGETABLE)),
5105 K(zone_page_state(zone, NR_BOUNCE)),
5107 K(this_cpu_read(zone->pageset->pcp.count)),
5108 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5109 printk("lowmem_reserve[]:");
5110 for (i = 0; i < MAX_NR_ZONES; i++)
5111 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5112 printk(KERN_CONT "\n");
5115 for_each_populated_zone(zone) {
5117 unsigned long nr[MAX_ORDER], flags, total = 0;
5118 unsigned char types[MAX_ORDER];
5120 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5123 printk(KERN_CONT "%s: ", zone->name);
5125 spin_lock_irqsave(&zone->lock, flags);
5126 for (order = 0; order < MAX_ORDER; order++) {
5127 struct free_area *area = &zone->free_area[order];
5130 nr[order] = area->nr_free;
5131 total += nr[order] << order;
5134 for (type = 0; type < MIGRATE_TYPES; type++) {
5135 if (!list_empty(&area->free_list[type]))
5136 types[order] |= 1 << type;
5139 spin_unlock_irqrestore(&zone->lock, flags);
5140 for (order = 0; order < MAX_ORDER; order++) {
5141 printk(KERN_CONT "%lu*%lukB ",
5142 nr[order], K(1UL) << order);
5144 show_migration_types(types[order]);
5146 printk(KERN_CONT "= %lukB\n", K(total));
5149 hugetlb_show_meminfo();
5151 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5153 show_swap_cache_info();
5156 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5158 zoneref->zone = zone;
5159 zoneref->zone_idx = zone_idx(zone);
5163 * Builds allocation fallback zone lists.
5165 * Add all populated zones of a node to the zonelist.
5167 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5170 enum zone_type zone_type = MAX_NR_ZONES;
5175 zone = pgdat->node_zones + zone_type;
5176 if (managed_zone(zone)) {
5177 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5178 check_highest_zone(zone_type);
5180 } while (zone_type);
5187 static int __parse_numa_zonelist_order(char *s)
5190 * We used to support different zonlists modes but they turned
5191 * out to be just not useful. Let's keep the warning in place
5192 * if somebody still use the cmd line parameter so that we do
5193 * not fail it silently
5195 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5196 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5202 static __init int setup_numa_zonelist_order(char *s)
5207 return __parse_numa_zonelist_order(s);
5209 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5211 char numa_zonelist_order[] = "Node";
5214 * sysctl handler for numa_zonelist_order
5216 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5217 void __user *buffer, size_t *length,
5224 return proc_dostring(table, write, buffer, length, ppos);
5225 str = memdup_user_nul(buffer, 16);
5227 return PTR_ERR(str);
5229 ret = __parse_numa_zonelist_order(str);
5235 #define MAX_NODE_LOAD (nr_online_nodes)
5236 static int node_load[MAX_NUMNODES];
5239 * find_next_best_node - find the next node that should appear in a given node's fallback list
5240 * @node: node whose fallback list we're appending
5241 * @used_node_mask: nodemask_t of already used nodes
5243 * We use a number of factors to determine which is the next node that should
5244 * appear on a given node's fallback list. The node should not have appeared
5245 * already in @node's fallback list, and it should be the next closest node
5246 * according to the distance array (which contains arbitrary distance values
5247 * from each node to each node in the system), and should also prefer nodes
5248 * with no CPUs, since presumably they'll have very little allocation pressure
5249 * on them otherwise.
5250 * It returns -1 if no node is found.
5252 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5255 int min_val = INT_MAX;
5256 int best_node = NUMA_NO_NODE;
5257 const struct cpumask *tmp = cpumask_of_node(0);
5259 /* Use the local node if we haven't already */
5260 if (!node_isset(node, *used_node_mask)) {
5261 node_set(node, *used_node_mask);
5265 for_each_node_state(n, N_MEMORY) {
5267 /* Don't want a node to appear more than once */
5268 if (node_isset(n, *used_node_mask))
5271 /* Use the distance array to find the distance */
5272 val = node_distance(node, n);
5274 /* Penalize nodes under us ("prefer the next node") */
5277 /* Give preference to headless and unused nodes */
5278 tmp = cpumask_of_node(n);
5279 if (!cpumask_empty(tmp))
5280 val += PENALTY_FOR_NODE_WITH_CPUS;
5282 /* Slight preference for less loaded node */
5283 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5284 val += node_load[n];
5286 if (val < min_val) {
5293 node_set(best_node, *used_node_mask);
5300 * Build zonelists ordered by node and zones within node.
5301 * This results in maximum locality--normal zone overflows into local
5302 * DMA zone, if any--but risks exhausting DMA zone.
5304 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5307 struct zoneref *zonerefs;
5310 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5312 for (i = 0; i < nr_nodes; i++) {
5315 pg_data_t *node = NODE_DATA(node_order[i]);
5317 nr_zones = build_zonerefs_node(node, zonerefs);
5318 zonerefs += nr_zones;
5320 zonerefs->zone = NULL;
5321 zonerefs->zone_idx = 0;
5325 * Build gfp_thisnode zonelists
5327 static void build_thisnode_zonelists(pg_data_t *pgdat)
5329 struct zoneref *zonerefs;
5332 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5333 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5334 zonerefs += nr_zones;
5335 zonerefs->zone = NULL;
5336 zonerefs->zone_idx = 0;
5340 * Build zonelists ordered by zone and nodes within zones.
5341 * This results in conserving DMA zone[s] until all Normal memory is
5342 * exhausted, but results in overflowing to remote node while memory
5343 * may still exist in local DMA zone.
5346 static void build_zonelists(pg_data_t *pgdat)
5348 static int node_order[MAX_NUMNODES];
5349 int node, load, nr_nodes = 0;
5350 nodemask_t used_mask;
5351 int local_node, prev_node;
5353 /* NUMA-aware ordering of nodes */
5354 local_node = pgdat->node_id;
5355 load = nr_online_nodes;
5356 prev_node = local_node;
5357 nodes_clear(used_mask);
5359 memset(node_order, 0, sizeof(node_order));
5360 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5362 * We don't want to pressure a particular node.
5363 * So adding penalty to the first node in same
5364 * distance group to make it round-robin.
5366 if (node_distance(local_node, node) !=
5367 node_distance(local_node, prev_node))
5368 node_load[node] = load;
5370 node_order[nr_nodes++] = node;
5375 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5376 build_thisnode_zonelists(pgdat);
5379 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5381 * Return node id of node used for "local" allocations.
5382 * I.e., first node id of first zone in arg node's generic zonelist.
5383 * Used for initializing percpu 'numa_mem', which is used primarily
5384 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5386 int local_memory_node(int node)
5390 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5391 gfp_zone(GFP_KERNEL),
5393 return zone_to_nid(z->zone);
5397 static void setup_min_unmapped_ratio(void);
5398 static void setup_min_slab_ratio(void);
5399 #else /* CONFIG_NUMA */
5401 static void build_zonelists(pg_data_t *pgdat)
5403 int node, local_node;
5404 struct zoneref *zonerefs;
5407 local_node = pgdat->node_id;
5409 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5410 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5411 zonerefs += nr_zones;
5414 * Now we build the zonelist so that it contains the zones
5415 * of all the other nodes.
5416 * We don't want to pressure a particular node, so when
5417 * building the zones for node N, we make sure that the
5418 * zones coming right after the local ones are those from
5419 * node N+1 (modulo N)
5421 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5422 if (!node_online(node))
5424 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5425 zonerefs += nr_zones;
5427 for (node = 0; node < local_node; node++) {
5428 if (!node_online(node))
5430 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5431 zonerefs += nr_zones;
5434 zonerefs->zone = NULL;
5435 zonerefs->zone_idx = 0;
5438 #endif /* CONFIG_NUMA */
5441 * Boot pageset table. One per cpu which is going to be used for all
5442 * zones and all nodes. The parameters will be set in such a way
5443 * that an item put on a list will immediately be handed over to
5444 * the buddy list. This is safe since pageset manipulation is done
5445 * with interrupts disabled.
5447 * The boot_pagesets must be kept even after bootup is complete for
5448 * unused processors and/or zones. They do play a role for bootstrapping
5449 * hotplugged processors.
5451 * zoneinfo_show() and maybe other functions do
5452 * not check if the processor is online before following the pageset pointer.
5453 * Other parts of the kernel may not check if the zone is available.
5455 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5456 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5457 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5459 static void __build_all_zonelists(void *data)
5462 int __maybe_unused cpu;
5463 pg_data_t *self = data;
5464 static DEFINE_SPINLOCK(lock);
5469 memset(node_load, 0, sizeof(node_load));
5473 * This node is hotadded and no memory is yet present. So just
5474 * building zonelists is fine - no need to touch other nodes.
5476 if (self && !node_online(self->node_id)) {
5477 build_zonelists(self);
5479 for_each_online_node(nid) {
5480 pg_data_t *pgdat = NODE_DATA(nid);
5482 build_zonelists(pgdat);
5485 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5487 * We now know the "local memory node" for each node--
5488 * i.e., the node of the first zone in the generic zonelist.
5489 * Set up numa_mem percpu variable for on-line cpus. During
5490 * boot, only the boot cpu should be on-line; we'll init the
5491 * secondary cpus' numa_mem as they come on-line. During
5492 * node/memory hotplug, we'll fixup all on-line cpus.
5494 for_each_online_cpu(cpu)
5495 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5502 static noinline void __init
5503 build_all_zonelists_init(void)
5507 __build_all_zonelists(NULL);
5510 * Initialize the boot_pagesets that are going to be used
5511 * for bootstrapping processors. The real pagesets for
5512 * each zone will be allocated later when the per cpu
5513 * allocator is available.
5515 * boot_pagesets are used also for bootstrapping offline
5516 * cpus if the system is already booted because the pagesets
5517 * are needed to initialize allocators on a specific cpu too.
5518 * F.e. the percpu allocator needs the page allocator which
5519 * needs the percpu allocator in order to allocate its pagesets
5520 * (a chicken-egg dilemma).
5522 for_each_possible_cpu(cpu)
5523 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5525 mminit_verify_zonelist();
5526 cpuset_init_current_mems_allowed();
5530 * unless system_state == SYSTEM_BOOTING.
5532 * __ref due to call of __init annotated helper build_all_zonelists_init
5533 * [protected by SYSTEM_BOOTING].
5535 void __ref build_all_zonelists(pg_data_t *pgdat)
5537 if (system_state == SYSTEM_BOOTING) {
5538 build_all_zonelists_init();
5540 __build_all_zonelists(pgdat);
5541 /* cpuset refresh routine should be here */
5543 vm_total_pages = nr_free_pagecache_pages();
5545 * Disable grouping by mobility if the number of pages in the
5546 * system is too low to allow the mechanism to work. It would be
5547 * more accurate, but expensive to check per-zone. This check is
5548 * made on memory-hotadd so a system can start with mobility
5549 * disabled and enable it later
5551 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5552 page_group_by_mobility_disabled = 1;
5554 page_group_by_mobility_disabled = 0;
5556 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5558 page_group_by_mobility_disabled ? "off" : "on",
5561 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5565 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5566 static bool __meminit
5567 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5569 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5570 static struct memblock_region *r;
5572 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5573 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5574 for_each_memblock(memory, r) {
5575 if (*pfn < memblock_region_memory_end_pfn(r))
5579 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5580 memblock_is_mirror(r)) {
5581 *pfn = memblock_region_memory_end_pfn(r);
5590 * Initially all pages are reserved - free ones are freed
5591 * up by memblock_free_all() once the early boot process is
5592 * done. Non-atomic initialization, single-pass.
5594 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5595 unsigned long start_pfn, enum memmap_context context,
5596 struct vmem_altmap *altmap)
5598 unsigned long pfn, end_pfn = start_pfn + size;
5601 if (highest_memmap_pfn < end_pfn - 1)
5602 highest_memmap_pfn = end_pfn - 1;
5604 #ifdef CONFIG_ZONE_DEVICE
5606 * Honor reservation requested by the driver for this ZONE_DEVICE
5607 * memory. We limit the total number of pages to initialize to just
5608 * those that might contain the memory mapping. We will defer the
5609 * ZONE_DEVICE page initialization until after we have released
5612 if (zone == ZONE_DEVICE) {
5616 if (start_pfn == altmap->base_pfn)
5617 start_pfn += altmap->reserve;
5618 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5622 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5624 * There can be holes in boot-time mem_map[]s handed to this
5625 * function. They do not exist on hotplugged memory.
5627 if (context == MEMMAP_EARLY) {
5628 if (!early_pfn_valid(pfn))
5630 if (!early_pfn_in_nid(pfn, nid))
5632 if (overlap_memmap_init(zone, &pfn))
5634 if (defer_init(nid, pfn, end_pfn))
5638 page = pfn_to_page(pfn);
5639 __init_single_page(page, pfn, zone, nid);
5640 if (context == MEMMAP_HOTPLUG)
5641 __SetPageReserved(page);
5644 * Mark the block movable so that blocks are reserved for
5645 * movable at startup. This will force kernel allocations
5646 * to reserve their blocks rather than leaking throughout
5647 * the address space during boot when many long-lived
5648 * kernel allocations are made.
5650 * bitmap is created for zone's valid pfn range. but memmap
5651 * can be created for invalid pages (for alignment)
5652 * check here not to call set_pageblock_migratetype() against
5655 if (!(pfn & (pageblock_nr_pages - 1))) {
5656 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5660 #ifdef CONFIG_SPARSEMEM
5662 * If the zone does not span the rest of the section then
5663 * we should at least initialize those pages. Otherwise we
5664 * could blow up on a poisoned page in some paths which depend
5665 * on full sections being initialized (e.g. memory hotplug).
5667 while (end_pfn % PAGES_PER_SECTION) {
5668 __init_single_page(pfn_to_page(end_pfn), end_pfn, zone, nid);
5674 #ifdef CONFIG_ZONE_DEVICE
5675 void __ref memmap_init_zone_device(struct zone *zone,
5676 unsigned long start_pfn,
5678 struct dev_pagemap *pgmap)
5680 unsigned long pfn, end_pfn = start_pfn + size;
5681 struct pglist_data *pgdat = zone->zone_pgdat;
5682 unsigned long zone_idx = zone_idx(zone);
5683 unsigned long start = jiffies;
5684 int nid = pgdat->node_id;
5686 if (WARN_ON_ONCE(!pgmap || !is_dev_zone(zone)))
5690 * The call to memmap_init_zone should have already taken care
5691 * of the pages reserved for the memmap, so we can just jump to
5692 * the end of that region and start processing the device pages.
5694 if (pgmap->altmap_valid) {
5695 struct vmem_altmap *altmap = &pgmap->altmap;
5697 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5698 size = end_pfn - start_pfn;
5701 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5702 struct page *page = pfn_to_page(pfn);
5704 __init_single_page(page, pfn, zone_idx, nid);
5707 * Mark page reserved as it will need to wait for onlining
5708 * phase for it to be fully associated with a zone.
5710 * We can use the non-atomic __set_bit operation for setting
5711 * the flag as we are still initializing the pages.
5713 __SetPageReserved(page);
5716 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5717 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5718 * page is ever freed or placed on a driver-private list.
5720 page->pgmap = pgmap;
5724 * Mark the block movable so that blocks are reserved for
5725 * movable at startup. This will force kernel allocations
5726 * to reserve their blocks rather than leaking throughout
5727 * the address space during boot when many long-lived
5728 * kernel allocations are made.
5730 * bitmap is created for zone's valid pfn range. but memmap
5731 * can be created for invalid pages (for alignment)
5732 * check here not to call set_pageblock_migratetype() against
5735 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5736 * because this is done early in sparse_add_one_section
5738 if (!(pfn & (pageblock_nr_pages - 1))) {
5739 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5744 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap->dev),
5745 size, jiffies_to_msecs(jiffies - start));
5749 static void __meminit zone_init_free_lists(struct zone *zone)
5751 unsigned int order, t;
5752 for_each_migratetype_order(order, t) {
5753 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5754 zone->free_area[order].nr_free = 0;
5758 void __meminit __weak memmap_init(unsigned long size, int nid,
5759 unsigned long zone, unsigned long start_pfn)
5761 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
5764 static int zone_batchsize(struct zone *zone)
5770 * The per-cpu-pages pools are set to around 1000th of the
5773 batch = zone_managed_pages(zone) / 1024;
5774 /* But no more than a meg. */
5775 if (batch * PAGE_SIZE > 1024 * 1024)
5776 batch = (1024 * 1024) / PAGE_SIZE;
5777 batch /= 4; /* We effectively *= 4 below */
5782 * Clamp the batch to a 2^n - 1 value. Having a power
5783 * of 2 value was found to be more likely to have
5784 * suboptimal cache aliasing properties in some cases.
5786 * For example if 2 tasks are alternately allocating
5787 * batches of pages, one task can end up with a lot
5788 * of pages of one half of the possible page colors
5789 * and the other with pages of the other colors.
5791 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5796 /* The deferral and batching of frees should be suppressed under NOMMU
5799 * The problem is that NOMMU needs to be able to allocate large chunks
5800 * of contiguous memory as there's no hardware page translation to
5801 * assemble apparent contiguous memory from discontiguous pages.
5803 * Queueing large contiguous runs of pages for batching, however,
5804 * causes the pages to actually be freed in smaller chunks. As there
5805 * can be a significant delay between the individual batches being
5806 * recycled, this leads to the once large chunks of space being
5807 * fragmented and becoming unavailable for high-order allocations.
5814 * pcp->high and pcp->batch values are related and dependent on one another:
5815 * ->batch must never be higher then ->high.
5816 * The following function updates them in a safe manner without read side
5819 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5820 * those fields changing asynchronously (acording the the above rule).
5822 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5823 * outside of boot time (or some other assurance that no concurrent updaters
5826 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5827 unsigned long batch)
5829 /* start with a fail safe value for batch */
5833 /* Update high, then batch, in order */
5840 /* a companion to pageset_set_high() */
5841 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5843 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5846 static void pageset_init(struct per_cpu_pageset *p)
5848 struct per_cpu_pages *pcp;
5851 memset(p, 0, sizeof(*p));
5854 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5855 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5858 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5861 pageset_set_batch(p, batch);
5865 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5866 * to the value high for the pageset p.
5868 static void pageset_set_high(struct per_cpu_pageset *p,
5871 unsigned long batch = max(1UL, high / 4);
5872 if ((high / 4) > (PAGE_SHIFT * 8))
5873 batch = PAGE_SHIFT * 8;
5875 pageset_update(&p->pcp, high, batch);
5878 static void pageset_set_high_and_batch(struct zone *zone,
5879 struct per_cpu_pageset *pcp)
5881 if (percpu_pagelist_fraction)
5882 pageset_set_high(pcp,
5883 (zone_managed_pages(zone) /
5884 percpu_pagelist_fraction));
5886 pageset_set_batch(pcp, zone_batchsize(zone));
5889 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5891 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5894 pageset_set_high_and_batch(zone, pcp);
5897 void __meminit setup_zone_pageset(struct zone *zone)
5900 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5901 for_each_possible_cpu(cpu)
5902 zone_pageset_init(zone, cpu);
5906 * Allocate per cpu pagesets and initialize them.
5907 * Before this call only boot pagesets were available.
5909 void __init setup_per_cpu_pageset(void)
5911 struct pglist_data *pgdat;
5914 for_each_populated_zone(zone)
5915 setup_zone_pageset(zone);
5917 for_each_online_pgdat(pgdat)
5918 pgdat->per_cpu_nodestats =
5919 alloc_percpu(struct per_cpu_nodestat);
5922 static __meminit void zone_pcp_init(struct zone *zone)
5925 * per cpu subsystem is not up at this point. The following code
5926 * relies on the ability of the linker to provide the
5927 * offset of a (static) per cpu variable into the per cpu area.
5929 zone->pageset = &boot_pageset;
5931 if (populated_zone(zone))
5932 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5933 zone->name, zone->present_pages,
5934 zone_batchsize(zone));
5937 void __meminit init_currently_empty_zone(struct zone *zone,
5938 unsigned long zone_start_pfn,
5941 struct pglist_data *pgdat = zone->zone_pgdat;
5942 int zone_idx = zone_idx(zone) + 1;
5944 if (zone_idx > pgdat->nr_zones)
5945 pgdat->nr_zones = zone_idx;
5947 zone->zone_start_pfn = zone_start_pfn;
5949 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5950 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5952 (unsigned long)zone_idx(zone),
5953 zone_start_pfn, (zone_start_pfn + size));
5955 zone_init_free_lists(zone);
5956 zone->initialized = 1;
5959 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5960 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5963 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5965 int __meminit __early_pfn_to_nid(unsigned long pfn,
5966 struct mminit_pfnnid_cache *state)
5968 unsigned long start_pfn, end_pfn;
5971 if (state->last_start <= pfn && pfn < state->last_end)
5972 return state->last_nid;
5974 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5976 state->last_start = start_pfn;
5977 state->last_end = end_pfn;
5978 state->last_nid = nid;
5983 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5986 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5987 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5988 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5990 * If an architecture guarantees that all ranges registered contain no holes
5991 * and may be freed, this this function may be used instead of calling
5992 * memblock_free_early_nid() manually.
5994 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5996 unsigned long start_pfn, end_pfn;
5999 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6000 start_pfn = min(start_pfn, max_low_pfn);
6001 end_pfn = min(end_pfn, max_low_pfn);
6003 if (start_pfn < end_pfn)
6004 memblock_free_early_nid(PFN_PHYS(start_pfn),
6005 (end_pfn - start_pfn) << PAGE_SHIFT,
6011 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6012 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6014 * If an architecture guarantees that all ranges registered contain no holes and may
6015 * be freed, this function may be used instead of calling memory_present() manually.
6017 void __init sparse_memory_present_with_active_regions(int nid)
6019 unsigned long start_pfn, end_pfn;
6022 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6023 memory_present(this_nid, start_pfn, end_pfn);
6027 * get_pfn_range_for_nid - Return the start and end page frames for a node
6028 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6029 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6030 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6032 * It returns the start and end page frame of a node based on information
6033 * provided by memblock_set_node(). If called for a node
6034 * with no available memory, a warning is printed and the start and end
6037 void __init get_pfn_range_for_nid(unsigned int nid,
6038 unsigned long *start_pfn, unsigned long *end_pfn)
6040 unsigned long this_start_pfn, this_end_pfn;
6046 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6047 *start_pfn = min(*start_pfn, this_start_pfn);
6048 *end_pfn = max(*end_pfn, this_end_pfn);
6051 if (*start_pfn == -1UL)
6056 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6057 * assumption is made that zones within a node are ordered in monotonic
6058 * increasing memory addresses so that the "highest" populated zone is used
6060 static void __init find_usable_zone_for_movable(void)
6063 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6064 if (zone_index == ZONE_MOVABLE)
6067 if (arch_zone_highest_possible_pfn[zone_index] >
6068 arch_zone_lowest_possible_pfn[zone_index])
6072 VM_BUG_ON(zone_index == -1);
6073 movable_zone = zone_index;
6077 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6078 * because it is sized independent of architecture. Unlike the other zones,
6079 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6080 * in each node depending on the size of each node and how evenly kernelcore
6081 * is distributed. This helper function adjusts the zone ranges
6082 * provided by the architecture for a given node by using the end of the
6083 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6084 * zones within a node are in order of monotonic increases memory addresses
6086 static void __init adjust_zone_range_for_zone_movable(int nid,
6087 unsigned long zone_type,
6088 unsigned long node_start_pfn,
6089 unsigned long node_end_pfn,
6090 unsigned long *zone_start_pfn,
6091 unsigned long *zone_end_pfn)
6093 /* Only adjust if ZONE_MOVABLE is on this node */
6094 if (zone_movable_pfn[nid]) {
6095 /* Size ZONE_MOVABLE */
6096 if (zone_type == ZONE_MOVABLE) {
6097 *zone_start_pfn = zone_movable_pfn[nid];
6098 *zone_end_pfn = min(node_end_pfn,
6099 arch_zone_highest_possible_pfn[movable_zone]);
6101 /* Adjust for ZONE_MOVABLE starting within this range */
6102 } else if (!mirrored_kernelcore &&
6103 *zone_start_pfn < zone_movable_pfn[nid] &&
6104 *zone_end_pfn > zone_movable_pfn[nid]) {
6105 *zone_end_pfn = zone_movable_pfn[nid];
6107 /* Check if this whole range is within ZONE_MOVABLE */
6108 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6109 *zone_start_pfn = *zone_end_pfn;
6114 * Return the number of pages a zone spans in a node, including holes
6115 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6117 static unsigned long __init zone_spanned_pages_in_node(int nid,
6118 unsigned long zone_type,
6119 unsigned long node_start_pfn,
6120 unsigned long node_end_pfn,
6121 unsigned long *zone_start_pfn,
6122 unsigned long *zone_end_pfn,
6123 unsigned long *ignored)
6125 /* When hotadd a new node from cpu_up(), the node should be empty */
6126 if (!node_start_pfn && !node_end_pfn)
6129 /* Get the start and end of the zone */
6130 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
6131 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
6132 adjust_zone_range_for_zone_movable(nid, zone_type,
6133 node_start_pfn, node_end_pfn,
6134 zone_start_pfn, zone_end_pfn);
6136 /* Check that this node has pages within the zone's required range */
6137 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6140 /* Move the zone boundaries inside the node if necessary */
6141 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6142 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6144 /* Return the spanned pages */
6145 return *zone_end_pfn - *zone_start_pfn;
6149 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6150 * then all holes in the requested range will be accounted for.
6152 unsigned long __init __absent_pages_in_range(int nid,
6153 unsigned long range_start_pfn,
6154 unsigned long range_end_pfn)
6156 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6157 unsigned long start_pfn, end_pfn;
6160 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6161 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6162 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6163 nr_absent -= end_pfn - start_pfn;
6169 * absent_pages_in_range - Return number of page frames in holes within a range
6170 * @start_pfn: The start PFN to start searching for holes
6171 * @end_pfn: The end PFN to stop searching for holes
6173 * It returns the number of pages frames in memory holes within a range.
6175 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6176 unsigned long end_pfn)
6178 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6181 /* Return the number of page frames in holes in a zone on a node */
6182 static unsigned long __init zone_absent_pages_in_node(int nid,
6183 unsigned long zone_type,
6184 unsigned long node_start_pfn,
6185 unsigned long node_end_pfn,
6186 unsigned long *ignored)
6188 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6189 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6190 unsigned long zone_start_pfn, zone_end_pfn;
6191 unsigned long nr_absent;
6193 /* When hotadd a new node from cpu_up(), the node should be empty */
6194 if (!node_start_pfn && !node_end_pfn)
6197 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6198 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6200 adjust_zone_range_for_zone_movable(nid, zone_type,
6201 node_start_pfn, node_end_pfn,
6202 &zone_start_pfn, &zone_end_pfn);
6203 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6206 * ZONE_MOVABLE handling.
6207 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6210 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6211 unsigned long start_pfn, end_pfn;
6212 struct memblock_region *r;
6214 for_each_memblock(memory, r) {
6215 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6216 zone_start_pfn, zone_end_pfn);
6217 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6218 zone_start_pfn, zone_end_pfn);
6220 if (zone_type == ZONE_MOVABLE &&
6221 memblock_is_mirror(r))
6222 nr_absent += end_pfn - start_pfn;
6224 if (zone_type == ZONE_NORMAL &&
6225 !memblock_is_mirror(r))
6226 nr_absent += end_pfn - start_pfn;
6233 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6234 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6235 unsigned long zone_type,
6236 unsigned long node_start_pfn,
6237 unsigned long node_end_pfn,
6238 unsigned long *zone_start_pfn,
6239 unsigned long *zone_end_pfn,
6240 unsigned long *zones_size)
6244 *zone_start_pfn = node_start_pfn;
6245 for (zone = 0; zone < zone_type; zone++)
6246 *zone_start_pfn += zones_size[zone];
6248 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6250 return zones_size[zone_type];
6253 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6254 unsigned long zone_type,
6255 unsigned long node_start_pfn,
6256 unsigned long node_end_pfn,
6257 unsigned long *zholes_size)
6262 return zholes_size[zone_type];
6265 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6267 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6268 unsigned long node_start_pfn,
6269 unsigned long node_end_pfn,
6270 unsigned long *zones_size,
6271 unsigned long *zholes_size)
6273 unsigned long realtotalpages = 0, totalpages = 0;
6276 for (i = 0; i < MAX_NR_ZONES; i++) {
6277 struct zone *zone = pgdat->node_zones + i;
6278 unsigned long zone_start_pfn, zone_end_pfn;
6279 unsigned long size, real_size;
6281 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6287 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6288 node_start_pfn, node_end_pfn,
6291 zone->zone_start_pfn = zone_start_pfn;
6293 zone->zone_start_pfn = 0;
6294 zone->spanned_pages = size;
6295 zone->present_pages = real_size;
6298 realtotalpages += real_size;
6301 pgdat->node_spanned_pages = totalpages;
6302 pgdat->node_present_pages = realtotalpages;
6303 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6307 #ifndef CONFIG_SPARSEMEM
6309 * Calculate the size of the zone->blockflags rounded to an unsigned long
6310 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6311 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6312 * round what is now in bits to nearest long in bits, then return it in
6315 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6317 unsigned long usemapsize;
6319 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6320 usemapsize = roundup(zonesize, pageblock_nr_pages);
6321 usemapsize = usemapsize >> pageblock_order;
6322 usemapsize *= NR_PAGEBLOCK_BITS;
6323 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6325 return usemapsize / 8;
6328 static void __ref setup_usemap(struct pglist_data *pgdat,
6330 unsigned long zone_start_pfn,
6331 unsigned long zonesize)
6333 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6334 zone->pageblock_flags = NULL;
6336 zone->pageblock_flags =
6337 memblock_alloc_node_nopanic(usemapsize,
6341 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6342 unsigned long zone_start_pfn, unsigned long zonesize) {}
6343 #endif /* CONFIG_SPARSEMEM */
6345 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6347 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6348 void __init set_pageblock_order(void)
6352 /* Check that pageblock_nr_pages has not already been setup */
6353 if (pageblock_order)
6356 if (HPAGE_SHIFT > PAGE_SHIFT)
6357 order = HUGETLB_PAGE_ORDER;
6359 order = MAX_ORDER - 1;
6362 * Assume the largest contiguous order of interest is a huge page.
6363 * This value may be variable depending on boot parameters on IA64 and
6366 pageblock_order = order;
6368 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6371 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6372 * is unused as pageblock_order is set at compile-time. See
6373 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6376 void __init set_pageblock_order(void)
6380 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6382 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6383 unsigned long present_pages)
6385 unsigned long pages = spanned_pages;
6388 * Provide a more accurate estimation if there are holes within
6389 * the zone and SPARSEMEM is in use. If there are holes within the
6390 * zone, each populated memory region may cost us one or two extra
6391 * memmap pages due to alignment because memmap pages for each
6392 * populated regions may not be naturally aligned on page boundary.
6393 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6395 if (spanned_pages > present_pages + (present_pages >> 4) &&
6396 IS_ENABLED(CONFIG_SPARSEMEM))
6397 pages = present_pages;
6399 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6402 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6403 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6405 spin_lock_init(&pgdat->split_queue_lock);
6406 INIT_LIST_HEAD(&pgdat->split_queue);
6407 pgdat->split_queue_len = 0;
6410 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6413 #ifdef CONFIG_COMPACTION
6414 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6416 init_waitqueue_head(&pgdat->kcompactd_wait);
6419 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6422 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6424 pgdat_resize_init(pgdat);
6426 pgdat_init_split_queue(pgdat);
6427 pgdat_init_kcompactd(pgdat);
6429 init_waitqueue_head(&pgdat->kswapd_wait);
6430 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6432 pgdat_page_ext_init(pgdat);
6433 spin_lock_init(&pgdat->lru_lock);
6434 lruvec_init(node_lruvec(pgdat));
6437 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6438 unsigned long remaining_pages)
6440 atomic_long_set(&zone->managed_pages, remaining_pages);
6441 zone_set_nid(zone, nid);
6442 zone->name = zone_names[idx];
6443 zone->zone_pgdat = NODE_DATA(nid);
6444 spin_lock_init(&zone->lock);
6445 zone_seqlock_init(zone);
6446 zone_pcp_init(zone);
6450 * Set up the zone data structures
6451 * - init pgdat internals
6452 * - init all zones belonging to this node
6454 * NOTE: this function is only called during memory hotplug
6456 #ifdef CONFIG_MEMORY_HOTPLUG
6457 void __ref free_area_init_core_hotplug(int nid)
6460 pg_data_t *pgdat = NODE_DATA(nid);
6462 pgdat_init_internals(pgdat);
6463 for (z = 0; z < MAX_NR_ZONES; z++)
6464 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6469 * Set up the zone data structures:
6470 * - mark all pages reserved
6471 * - mark all memory queues empty
6472 * - clear the memory bitmaps
6474 * NOTE: pgdat should get zeroed by caller.
6475 * NOTE: this function is only called during early init.
6477 static void __init free_area_init_core(struct pglist_data *pgdat)
6480 int nid = pgdat->node_id;
6482 pgdat_init_internals(pgdat);
6483 pgdat->per_cpu_nodestats = &boot_nodestats;
6485 for (j = 0; j < MAX_NR_ZONES; j++) {
6486 struct zone *zone = pgdat->node_zones + j;
6487 unsigned long size, freesize, memmap_pages;
6488 unsigned long zone_start_pfn = zone->zone_start_pfn;
6490 size = zone->spanned_pages;
6491 freesize = zone->present_pages;
6494 * Adjust freesize so that it accounts for how much memory
6495 * is used by this zone for memmap. This affects the watermark
6496 * and per-cpu initialisations
6498 memmap_pages = calc_memmap_size(size, freesize);
6499 if (!is_highmem_idx(j)) {
6500 if (freesize >= memmap_pages) {
6501 freesize -= memmap_pages;
6504 " %s zone: %lu pages used for memmap\n",
6505 zone_names[j], memmap_pages);
6507 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6508 zone_names[j], memmap_pages, freesize);
6511 /* Account for reserved pages */
6512 if (j == 0 && freesize > dma_reserve) {
6513 freesize -= dma_reserve;
6514 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6515 zone_names[0], dma_reserve);
6518 if (!is_highmem_idx(j))
6519 nr_kernel_pages += freesize;
6520 /* Charge for highmem memmap if there are enough kernel pages */
6521 else if (nr_kernel_pages > memmap_pages * 2)
6522 nr_kernel_pages -= memmap_pages;
6523 nr_all_pages += freesize;
6526 * Set an approximate value for lowmem here, it will be adjusted
6527 * when the bootmem allocator frees pages into the buddy system.
6528 * And all highmem pages will be managed by the buddy system.
6530 zone_init_internals(zone, j, nid, freesize);
6535 set_pageblock_order();
6536 setup_usemap(pgdat, zone, zone_start_pfn, size);
6537 init_currently_empty_zone(zone, zone_start_pfn, size);
6538 memmap_init(size, nid, j, zone_start_pfn);
6542 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6543 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6545 unsigned long __maybe_unused start = 0;
6546 unsigned long __maybe_unused offset = 0;
6548 /* Skip empty nodes */
6549 if (!pgdat->node_spanned_pages)
6552 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6553 offset = pgdat->node_start_pfn - start;
6554 /* ia64 gets its own node_mem_map, before this, without bootmem */
6555 if (!pgdat->node_mem_map) {
6556 unsigned long size, end;
6560 * The zone's endpoints aren't required to be MAX_ORDER
6561 * aligned but the node_mem_map endpoints must be in order
6562 * for the buddy allocator to function correctly.
6564 end = pgdat_end_pfn(pgdat);
6565 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6566 size = (end - start) * sizeof(struct page);
6567 map = memblock_alloc_node_nopanic(size, pgdat->node_id);
6568 pgdat->node_mem_map = map + offset;
6570 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6571 __func__, pgdat->node_id, (unsigned long)pgdat,
6572 (unsigned long)pgdat->node_mem_map);
6573 #ifndef CONFIG_NEED_MULTIPLE_NODES
6575 * With no DISCONTIG, the global mem_map is just set as node 0's
6577 if (pgdat == NODE_DATA(0)) {
6578 mem_map = NODE_DATA(0)->node_mem_map;
6579 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6580 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6582 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6587 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6588 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6590 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6591 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6593 pgdat->first_deferred_pfn = ULONG_MAX;
6596 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6599 void __init free_area_init_node(int nid, unsigned long *zones_size,
6600 unsigned long node_start_pfn,
6601 unsigned long *zholes_size)
6603 pg_data_t *pgdat = NODE_DATA(nid);
6604 unsigned long start_pfn = 0;
6605 unsigned long end_pfn = 0;
6607 /* pg_data_t should be reset to zero when it's allocated */
6608 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6610 pgdat->node_id = nid;
6611 pgdat->node_start_pfn = node_start_pfn;
6612 pgdat->per_cpu_nodestats = NULL;
6613 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6614 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6615 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6616 (u64)start_pfn << PAGE_SHIFT,
6617 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6619 start_pfn = node_start_pfn;
6621 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6622 zones_size, zholes_size);
6624 alloc_node_mem_map(pgdat);
6625 pgdat_set_deferred_range(pgdat);
6627 free_area_init_core(pgdat);
6630 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6632 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6635 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6640 for (pfn = spfn; pfn < epfn; pfn++) {
6641 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6642 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6643 + pageblock_nr_pages - 1;
6646 mm_zero_struct_page(pfn_to_page(pfn));
6654 * Only struct pages that are backed by physical memory are zeroed and
6655 * initialized by going through __init_single_page(). But, there are some
6656 * struct pages which are reserved in memblock allocator and their fields
6657 * may be accessed (for example page_to_pfn() on some configuration accesses
6658 * flags). We must explicitly zero those struct pages.
6660 * This function also addresses a similar issue where struct pages are left
6661 * uninitialized because the physical address range is not covered by
6662 * memblock.memory or memblock.reserved. That could happen when memblock
6663 * layout is manually configured via memmap=.
6665 void __init zero_resv_unavail(void)
6667 phys_addr_t start, end;
6669 phys_addr_t next = 0;
6672 * Loop through unavailable ranges not covered by memblock.memory.
6675 for_each_mem_range(i, &memblock.memory, NULL,
6676 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6678 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6681 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6684 * Struct pages that do not have backing memory. This could be because
6685 * firmware is using some of this memory, or for some other reasons.
6688 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6690 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6692 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6694 #if MAX_NUMNODES > 1
6696 * Figure out the number of possible node ids.
6698 void __init setup_nr_node_ids(void)
6700 unsigned int highest;
6702 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6703 nr_node_ids = highest + 1;
6708 * node_map_pfn_alignment - determine the maximum internode alignment
6710 * This function should be called after node map is populated and sorted.
6711 * It calculates the maximum power of two alignment which can distinguish
6714 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6715 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6716 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6717 * shifted, 1GiB is enough and this function will indicate so.
6719 * This is used to test whether pfn -> nid mapping of the chosen memory
6720 * model has fine enough granularity to avoid incorrect mapping for the
6721 * populated node map.
6723 * Returns the determined alignment in pfn's. 0 if there is no alignment
6724 * requirement (single node).
6726 unsigned long __init node_map_pfn_alignment(void)
6728 unsigned long accl_mask = 0, last_end = 0;
6729 unsigned long start, end, mask;
6733 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6734 if (!start || last_nid < 0 || last_nid == nid) {
6741 * Start with a mask granular enough to pin-point to the
6742 * start pfn and tick off bits one-by-one until it becomes
6743 * too coarse to separate the current node from the last.
6745 mask = ~((1 << __ffs(start)) - 1);
6746 while (mask && last_end <= (start & (mask << 1)))
6749 /* accumulate all internode masks */
6753 /* convert mask to number of pages */
6754 return ~accl_mask + 1;
6757 /* Find the lowest pfn for a node */
6758 static unsigned long __init find_min_pfn_for_node(int nid)
6760 unsigned long min_pfn = ULONG_MAX;
6761 unsigned long start_pfn;
6764 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6765 min_pfn = min(min_pfn, start_pfn);
6767 if (min_pfn == ULONG_MAX) {
6768 pr_warn("Could not find start_pfn for node %d\n", nid);
6776 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6778 * It returns the minimum PFN based on information provided via
6779 * memblock_set_node().
6781 unsigned long __init find_min_pfn_with_active_regions(void)
6783 return find_min_pfn_for_node(MAX_NUMNODES);
6787 * early_calculate_totalpages()
6788 * Sum pages in active regions for movable zone.
6789 * Populate N_MEMORY for calculating usable_nodes.
6791 static unsigned long __init early_calculate_totalpages(void)
6793 unsigned long totalpages = 0;
6794 unsigned long start_pfn, end_pfn;
6797 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6798 unsigned long pages = end_pfn - start_pfn;
6800 totalpages += pages;
6802 node_set_state(nid, N_MEMORY);
6808 * Find the PFN the Movable zone begins in each node. Kernel memory
6809 * is spread evenly between nodes as long as the nodes have enough
6810 * memory. When they don't, some nodes will have more kernelcore than
6813 static void __init find_zone_movable_pfns_for_nodes(void)
6816 unsigned long usable_startpfn;
6817 unsigned long kernelcore_node, kernelcore_remaining;
6818 /* save the state before borrow the nodemask */
6819 nodemask_t saved_node_state = node_states[N_MEMORY];
6820 unsigned long totalpages = early_calculate_totalpages();
6821 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6822 struct memblock_region *r;
6824 /* Need to find movable_zone earlier when movable_node is specified. */
6825 find_usable_zone_for_movable();
6828 * If movable_node is specified, ignore kernelcore and movablecore
6831 if (movable_node_is_enabled()) {
6832 for_each_memblock(memory, r) {
6833 if (!memblock_is_hotpluggable(r))
6838 usable_startpfn = PFN_DOWN(r->base);
6839 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6840 min(usable_startpfn, zone_movable_pfn[nid]) :
6848 * If kernelcore=mirror is specified, ignore movablecore option
6850 if (mirrored_kernelcore) {
6851 bool mem_below_4gb_not_mirrored = false;
6853 for_each_memblock(memory, r) {
6854 if (memblock_is_mirror(r))
6859 usable_startpfn = memblock_region_memory_base_pfn(r);
6861 if (usable_startpfn < 0x100000) {
6862 mem_below_4gb_not_mirrored = true;
6866 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6867 min(usable_startpfn, zone_movable_pfn[nid]) :
6871 if (mem_below_4gb_not_mirrored)
6872 pr_warn("This configuration results in unmirrored kernel memory.");
6878 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6879 * amount of necessary memory.
6881 if (required_kernelcore_percent)
6882 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6884 if (required_movablecore_percent)
6885 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6889 * If movablecore= was specified, calculate what size of
6890 * kernelcore that corresponds so that memory usable for
6891 * any allocation type is evenly spread. If both kernelcore
6892 * and movablecore are specified, then the value of kernelcore
6893 * will be used for required_kernelcore if it's greater than
6894 * what movablecore would have allowed.
6896 if (required_movablecore) {
6897 unsigned long corepages;
6900 * Round-up so that ZONE_MOVABLE is at least as large as what
6901 * was requested by the user
6903 required_movablecore =
6904 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6905 required_movablecore = min(totalpages, required_movablecore);
6906 corepages = totalpages - required_movablecore;
6908 required_kernelcore = max(required_kernelcore, corepages);
6912 * If kernelcore was not specified or kernelcore size is larger
6913 * than totalpages, there is no ZONE_MOVABLE.
6915 if (!required_kernelcore || required_kernelcore >= totalpages)
6918 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6919 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6922 /* Spread kernelcore memory as evenly as possible throughout nodes */
6923 kernelcore_node = required_kernelcore / usable_nodes;
6924 for_each_node_state(nid, N_MEMORY) {
6925 unsigned long start_pfn, end_pfn;
6928 * Recalculate kernelcore_node if the division per node
6929 * now exceeds what is necessary to satisfy the requested
6930 * amount of memory for the kernel
6932 if (required_kernelcore < kernelcore_node)
6933 kernelcore_node = required_kernelcore / usable_nodes;
6936 * As the map is walked, we track how much memory is usable
6937 * by the kernel using kernelcore_remaining. When it is
6938 * 0, the rest of the node is usable by ZONE_MOVABLE
6940 kernelcore_remaining = kernelcore_node;
6942 /* Go through each range of PFNs within this node */
6943 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6944 unsigned long size_pages;
6946 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6947 if (start_pfn >= end_pfn)
6950 /* Account for what is only usable for kernelcore */
6951 if (start_pfn < usable_startpfn) {
6952 unsigned long kernel_pages;
6953 kernel_pages = min(end_pfn, usable_startpfn)
6956 kernelcore_remaining -= min(kernel_pages,
6957 kernelcore_remaining);
6958 required_kernelcore -= min(kernel_pages,
6959 required_kernelcore);
6961 /* Continue if range is now fully accounted */
6962 if (end_pfn <= usable_startpfn) {
6965 * Push zone_movable_pfn to the end so
6966 * that if we have to rebalance
6967 * kernelcore across nodes, we will
6968 * not double account here
6970 zone_movable_pfn[nid] = end_pfn;
6973 start_pfn = usable_startpfn;
6977 * The usable PFN range for ZONE_MOVABLE is from
6978 * start_pfn->end_pfn. Calculate size_pages as the
6979 * number of pages used as kernelcore
6981 size_pages = end_pfn - start_pfn;
6982 if (size_pages > kernelcore_remaining)
6983 size_pages = kernelcore_remaining;
6984 zone_movable_pfn[nid] = start_pfn + size_pages;
6987 * Some kernelcore has been met, update counts and
6988 * break if the kernelcore for this node has been
6991 required_kernelcore -= min(required_kernelcore,
6993 kernelcore_remaining -= size_pages;
6994 if (!kernelcore_remaining)
7000 * If there is still required_kernelcore, we do another pass with one
7001 * less node in the count. This will push zone_movable_pfn[nid] further
7002 * along on the nodes that still have memory until kernelcore is
7006 if (usable_nodes && required_kernelcore > usable_nodes)
7010 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7011 for (nid = 0; nid < MAX_NUMNODES; nid++)
7012 zone_movable_pfn[nid] =
7013 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7016 /* restore the node_state */
7017 node_states[N_MEMORY] = saved_node_state;
7020 /* Any regular or high memory on that node ? */
7021 static void check_for_memory(pg_data_t *pgdat, int nid)
7023 enum zone_type zone_type;
7025 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7026 struct zone *zone = &pgdat->node_zones[zone_type];
7027 if (populated_zone(zone)) {
7028 if (IS_ENABLED(CONFIG_HIGHMEM))
7029 node_set_state(nid, N_HIGH_MEMORY);
7030 if (zone_type <= ZONE_NORMAL)
7031 node_set_state(nid, N_NORMAL_MEMORY);
7038 * free_area_init_nodes - Initialise all pg_data_t and zone data
7039 * @max_zone_pfn: an array of max PFNs for each zone
7041 * This will call free_area_init_node() for each active node in the system.
7042 * Using the page ranges provided by memblock_set_node(), the size of each
7043 * zone in each node and their holes is calculated. If the maximum PFN
7044 * between two adjacent zones match, it is assumed that the zone is empty.
7045 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7046 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7047 * starts where the previous one ended. For example, ZONE_DMA32 starts
7048 * at arch_max_dma_pfn.
7050 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7052 unsigned long start_pfn, end_pfn;
7055 /* Record where the zone boundaries are */
7056 memset(arch_zone_lowest_possible_pfn, 0,
7057 sizeof(arch_zone_lowest_possible_pfn));
7058 memset(arch_zone_highest_possible_pfn, 0,
7059 sizeof(arch_zone_highest_possible_pfn));
7061 start_pfn = find_min_pfn_with_active_regions();
7063 for (i = 0; i < MAX_NR_ZONES; i++) {
7064 if (i == ZONE_MOVABLE)
7067 end_pfn = max(max_zone_pfn[i], start_pfn);
7068 arch_zone_lowest_possible_pfn[i] = start_pfn;
7069 arch_zone_highest_possible_pfn[i] = end_pfn;
7071 start_pfn = end_pfn;
7074 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7075 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7076 find_zone_movable_pfns_for_nodes();
7078 /* Print out the zone ranges */
7079 pr_info("Zone ranges:\n");
7080 for (i = 0; i < MAX_NR_ZONES; i++) {
7081 if (i == ZONE_MOVABLE)
7083 pr_info(" %-8s ", zone_names[i]);
7084 if (arch_zone_lowest_possible_pfn[i] ==
7085 arch_zone_highest_possible_pfn[i])
7088 pr_cont("[mem %#018Lx-%#018Lx]\n",
7089 (u64)arch_zone_lowest_possible_pfn[i]
7091 ((u64)arch_zone_highest_possible_pfn[i]
7092 << PAGE_SHIFT) - 1);
7095 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7096 pr_info("Movable zone start for each node\n");
7097 for (i = 0; i < MAX_NUMNODES; i++) {
7098 if (zone_movable_pfn[i])
7099 pr_info(" Node %d: %#018Lx\n", i,
7100 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7103 /* Print out the early node map */
7104 pr_info("Early memory node ranges\n");
7105 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
7106 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7107 (u64)start_pfn << PAGE_SHIFT,
7108 ((u64)end_pfn << PAGE_SHIFT) - 1);
7110 /* Initialise every node */
7111 mminit_verify_pageflags_layout();
7112 setup_nr_node_ids();
7113 zero_resv_unavail();
7114 for_each_online_node(nid) {
7115 pg_data_t *pgdat = NODE_DATA(nid);
7116 free_area_init_node(nid, NULL,
7117 find_min_pfn_for_node(nid), NULL);
7119 /* Any memory on that node */
7120 if (pgdat->node_present_pages)
7121 node_set_state(nid, N_MEMORY);
7122 check_for_memory(pgdat, nid);
7126 static int __init cmdline_parse_core(char *p, unsigned long *core,
7127 unsigned long *percent)
7129 unsigned long long coremem;
7135 /* Value may be a percentage of total memory, otherwise bytes */
7136 coremem = simple_strtoull(p, &endptr, 0);
7137 if (*endptr == '%') {
7138 /* Paranoid check for percent values greater than 100 */
7139 WARN_ON(coremem > 100);
7143 coremem = memparse(p, &p);
7144 /* Paranoid check that UL is enough for the coremem value */
7145 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7147 *core = coremem >> PAGE_SHIFT;
7154 * kernelcore=size sets the amount of memory for use for allocations that
7155 * cannot be reclaimed or migrated.
7157 static int __init cmdline_parse_kernelcore(char *p)
7159 /* parse kernelcore=mirror */
7160 if (parse_option_str(p, "mirror")) {
7161 mirrored_kernelcore = true;
7165 return cmdline_parse_core(p, &required_kernelcore,
7166 &required_kernelcore_percent);
7170 * movablecore=size sets the amount of memory for use for allocations that
7171 * can be reclaimed or migrated.
7173 static int __init cmdline_parse_movablecore(char *p)
7175 return cmdline_parse_core(p, &required_movablecore,
7176 &required_movablecore_percent);
7179 early_param("kernelcore", cmdline_parse_kernelcore);
7180 early_param("movablecore", cmdline_parse_movablecore);
7182 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7184 void adjust_managed_page_count(struct page *page, long count)
7186 atomic_long_add(count, &page_zone(page)->managed_pages);
7187 totalram_pages_add(count);
7188 #ifdef CONFIG_HIGHMEM
7189 if (PageHighMem(page))
7190 totalhigh_pages_add(count);
7193 EXPORT_SYMBOL(adjust_managed_page_count);
7195 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7198 unsigned long pages = 0;
7200 start = (void *)PAGE_ALIGN((unsigned long)start);
7201 end = (void *)((unsigned long)end & PAGE_MASK);
7202 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7203 struct page *page = virt_to_page(pos);
7204 void *direct_map_addr;
7207 * 'direct_map_addr' might be different from 'pos'
7208 * because some architectures' virt_to_page()
7209 * work with aliases. Getting the direct map
7210 * address ensures that we get a _writeable_
7211 * alias for the memset().
7213 direct_map_addr = page_address(page);
7214 if ((unsigned int)poison <= 0xFF)
7215 memset(direct_map_addr, poison, PAGE_SIZE);
7217 free_reserved_page(page);
7221 pr_info("Freeing %s memory: %ldK\n",
7222 s, pages << (PAGE_SHIFT - 10));
7226 EXPORT_SYMBOL(free_reserved_area);
7228 #ifdef CONFIG_HIGHMEM
7229 void free_highmem_page(struct page *page)
7231 __free_reserved_page(page);
7232 totalram_pages_inc();
7233 atomic_long_inc(&page_zone(page)->managed_pages);
7234 totalhigh_pages_inc();
7239 void __init mem_init_print_info(const char *str)
7241 unsigned long physpages, codesize, datasize, rosize, bss_size;
7242 unsigned long init_code_size, init_data_size;
7244 physpages = get_num_physpages();
7245 codesize = _etext - _stext;
7246 datasize = _edata - _sdata;
7247 rosize = __end_rodata - __start_rodata;
7248 bss_size = __bss_stop - __bss_start;
7249 init_data_size = __init_end - __init_begin;
7250 init_code_size = _einittext - _sinittext;
7253 * Detect special cases and adjust section sizes accordingly:
7254 * 1) .init.* may be embedded into .data sections
7255 * 2) .init.text.* may be out of [__init_begin, __init_end],
7256 * please refer to arch/tile/kernel/vmlinux.lds.S.
7257 * 3) .rodata.* may be embedded into .text or .data sections.
7259 #define adj_init_size(start, end, size, pos, adj) \
7261 if (start <= pos && pos < end && size > adj) \
7265 adj_init_size(__init_begin, __init_end, init_data_size,
7266 _sinittext, init_code_size);
7267 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7268 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7269 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7270 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7272 #undef adj_init_size
7274 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7275 #ifdef CONFIG_HIGHMEM
7279 nr_free_pages() << (PAGE_SHIFT - 10),
7280 physpages << (PAGE_SHIFT - 10),
7281 codesize >> 10, datasize >> 10, rosize >> 10,
7282 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7283 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7284 totalcma_pages << (PAGE_SHIFT - 10),
7285 #ifdef CONFIG_HIGHMEM
7286 totalhigh_pages() << (PAGE_SHIFT - 10),
7288 str ? ", " : "", str ? str : "");
7292 * set_dma_reserve - set the specified number of pages reserved in the first zone
7293 * @new_dma_reserve: The number of pages to mark reserved
7295 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7296 * In the DMA zone, a significant percentage may be consumed by kernel image
7297 * and other unfreeable allocations which can skew the watermarks badly. This
7298 * function may optionally be used to account for unfreeable pages in the
7299 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7300 * smaller per-cpu batchsize.
7302 void __init set_dma_reserve(unsigned long new_dma_reserve)
7304 dma_reserve = new_dma_reserve;
7307 void __init free_area_init(unsigned long *zones_size)
7309 zero_resv_unavail();
7310 free_area_init_node(0, zones_size,
7311 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7314 static int page_alloc_cpu_dead(unsigned int cpu)
7317 lru_add_drain_cpu(cpu);
7321 * Spill the event counters of the dead processor
7322 * into the current processors event counters.
7323 * This artificially elevates the count of the current
7326 vm_events_fold_cpu(cpu);
7329 * Zero the differential counters of the dead processor
7330 * so that the vm statistics are consistent.
7332 * This is only okay since the processor is dead and cannot
7333 * race with what we are doing.
7335 cpu_vm_stats_fold(cpu);
7339 void __init page_alloc_init(void)
7343 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7344 "mm/page_alloc:dead", NULL,
7345 page_alloc_cpu_dead);
7350 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7351 * or min_free_kbytes changes.
7353 static void calculate_totalreserve_pages(void)
7355 struct pglist_data *pgdat;
7356 unsigned long reserve_pages = 0;
7357 enum zone_type i, j;
7359 for_each_online_pgdat(pgdat) {
7361 pgdat->totalreserve_pages = 0;
7363 for (i = 0; i < MAX_NR_ZONES; i++) {
7364 struct zone *zone = pgdat->node_zones + i;
7366 unsigned long managed_pages = zone_managed_pages(zone);
7368 /* Find valid and maximum lowmem_reserve in the zone */
7369 for (j = i; j < MAX_NR_ZONES; j++) {
7370 if (zone->lowmem_reserve[j] > max)
7371 max = zone->lowmem_reserve[j];
7374 /* we treat the high watermark as reserved pages. */
7375 max += high_wmark_pages(zone);
7377 if (max > managed_pages)
7378 max = managed_pages;
7380 pgdat->totalreserve_pages += max;
7382 reserve_pages += max;
7385 totalreserve_pages = reserve_pages;
7389 * setup_per_zone_lowmem_reserve - called whenever
7390 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7391 * has a correct pages reserved value, so an adequate number of
7392 * pages are left in the zone after a successful __alloc_pages().
7394 static void setup_per_zone_lowmem_reserve(void)
7396 struct pglist_data *pgdat;
7397 enum zone_type j, idx;
7399 for_each_online_pgdat(pgdat) {
7400 for (j = 0; j < MAX_NR_ZONES; j++) {
7401 struct zone *zone = pgdat->node_zones + j;
7402 unsigned long managed_pages = zone_managed_pages(zone);
7404 zone->lowmem_reserve[j] = 0;
7408 struct zone *lower_zone;
7411 lower_zone = pgdat->node_zones + idx;
7413 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7414 sysctl_lowmem_reserve_ratio[idx] = 0;
7415 lower_zone->lowmem_reserve[j] = 0;
7417 lower_zone->lowmem_reserve[j] =
7418 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7420 managed_pages += zone_managed_pages(lower_zone);
7425 /* update totalreserve_pages */
7426 calculate_totalreserve_pages();
7429 static void __setup_per_zone_wmarks(void)
7431 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7432 unsigned long lowmem_pages = 0;
7434 unsigned long flags;
7436 /* Calculate total number of !ZONE_HIGHMEM pages */
7437 for_each_zone(zone) {
7438 if (!is_highmem(zone))
7439 lowmem_pages += zone_managed_pages(zone);
7442 for_each_zone(zone) {
7445 spin_lock_irqsave(&zone->lock, flags);
7446 tmp = (u64)pages_min * zone_managed_pages(zone);
7447 do_div(tmp, lowmem_pages);
7448 if (is_highmem(zone)) {
7450 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7451 * need highmem pages, so cap pages_min to a small
7454 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7455 * deltas control asynch page reclaim, and so should
7456 * not be capped for highmem.
7458 unsigned long min_pages;
7460 min_pages = zone_managed_pages(zone) / 1024;
7461 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7462 zone->_watermark[WMARK_MIN] = min_pages;
7465 * If it's a lowmem zone, reserve a number of pages
7466 * proportionate to the zone's size.
7468 zone->_watermark[WMARK_MIN] = tmp;
7472 * Set the kswapd watermarks distance according to the
7473 * scale factor in proportion to available memory, but
7474 * ensure a minimum size on small systems.
7476 tmp = max_t(u64, tmp >> 2,
7477 mult_frac(zone_managed_pages(zone),
7478 watermark_scale_factor, 10000));
7480 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7481 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7482 zone->watermark_boost = 0;
7484 spin_unlock_irqrestore(&zone->lock, flags);
7487 /* update totalreserve_pages */
7488 calculate_totalreserve_pages();
7492 * setup_per_zone_wmarks - called when min_free_kbytes changes
7493 * or when memory is hot-{added|removed}
7495 * Ensures that the watermark[min,low,high] values for each zone are set
7496 * correctly with respect to min_free_kbytes.
7498 void setup_per_zone_wmarks(void)
7500 static DEFINE_SPINLOCK(lock);
7503 __setup_per_zone_wmarks();
7508 * Initialise min_free_kbytes.
7510 * For small machines we want it small (128k min). For large machines
7511 * we want it large (64MB max). But it is not linear, because network
7512 * bandwidth does not increase linearly with machine size. We use
7514 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7515 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7531 int __meminit init_per_zone_wmark_min(void)
7533 unsigned long lowmem_kbytes;
7534 int new_min_free_kbytes;
7536 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7537 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7539 if (new_min_free_kbytes > user_min_free_kbytes) {
7540 min_free_kbytes = new_min_free_kbytes;
7541 if (min_free_kbytes < 128)
7542 min_free_kbytes = 128;
7543 if (min_free_kbytes > 65536)
7544 min_free_kbytes = 65536;
7546 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7547 new_min_free_kbytes, user_min_free_kbytes);
7549 setup_per_zone_wmarks();
7550 refresh_zone_stat_thresholds();
7551 setup_per_zone_lowmem_reserve();
7554 setup_min_unmapped_ratio();
7555 setup_min_slab_ratio();
7560 core_initcall(init_per_zone_wmark_min)
7563 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7564 * that we can call two helper functions whenever min_free_kbytes
7567 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7568 void __user *buffer, size_t *length, loff_t *ppos)
7572 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7577 user_min_free_kbytes = min_free_kbytes;
7578 setup_per_zone_wmarks();
7583 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7584 void __user *buffer, size_t *length, loff_t *ppos)
7588 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7595 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7596 void __user *buffer, size_t *length, loff_t *ppos)
7600 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7605 setup_per_zone_wmarks();
7611 static void setup_min_unmapped_ratio(void)
7616 for_each_online_pgdat(pgdat)
7617 pgdat->min_unmapped_pages = 0;
7620 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7621 sysctl_min_unmapped_ratio) / 100;
7625 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7626 void __user *buffer, size_t *length, loff_t *ppos)
7630 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7634 setup_min_unmapped_ratio();
7639 static void setup_min_slab_ratio(void)
7644 for_each_online_pgdat(pgdat)
7645 pgdat->min_slab_pages = 0;
7648 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7649 sysctl_min_slab_ratio) / 100;
7652 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7653 void __user *buffer, size_t *length, loff_t *ppos)
7657 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7661 setup_min_slab_ratio();
7668 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7669 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7670 * whenever sysctl_lowmem_reserve_ratio changes.
7672 * The reserve ratio obviously has absolutely no relation with the
7673 * minimum watermarks. The lowmem reserve ratio can only make sense
7674 * if in function of the boot time zone sizes.
7676 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7677 void __user *buffer, size_t *length, loff_t *ppos)
7679 proc_dointvec_minmax(table, write, buffer, length, ppos);
7680 setup_per_zone_lowmem_reserve();
7685 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7686 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7687 * pagelist can have before it gets flushed back to buddy allocator.
7689 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7690 void __user *buffer, size_t *length, loff_t *ppos)
7693 int old_percpu_pagelist_fraction;
7696 mutex_lock(&pcp_batch_high_lock);
7697 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7699 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7700 if (!write || ret < 0)
7703 /* Sanity checking to avoid pcp imbalance */
7704 if (percpu_pagelist_fraction &&
7705 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7706 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7712 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7715 for_each_populated_zone(zone) {
7718 for_each_possible_cpu(cpu)
7719 pageset_set_high_and_batch(zone,
7720 per_cpu_ptr(zone->pageset, cpu));
7723 mutex_unlock(&pcp_batch_high_lock);
7728 int hashdist = HASHDIST_DEFAULT;
7730 static int __init set_hashdist(char *str)
7734 hashdist = simple_strtoul(str, &str, 0);
7737 __setup("hashdist=", set_hashdist);
7740 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7742 * Returns the number of pages that arch has reserved but
7743 * is not known to alloc_large_system_hash().
7745 static unsigned long __init arch_reserved_kernel_pages(void)
7752 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7753 * machines. As memory size is increased the scale is also increased but at
7754 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7755 * quadruples the scale is increased by one, which means the size of hash table
7756 * only doubles, instead of quadrupling as well.
7757 * Because 32-bit systems cannot have large physical memory, where this scaling
7758 * makes sense, it is disabled on such platforms.
7760 #if __BITS_PER_LONG > 32
7761 #define ADAPT_SCALE_BASE (64ul << 30)
7762 #define ADAPT_SCALE_SHIFT 2
7763 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7767 * allocate a large system hash table from bootmem
7768 * - it is assumed that the hash table must contain an exact power-of-2
7769 * quantity of entries
7770 * - limit is the number of hash buckets, not the total allocation size
7772 void *__init alloc_large_system_hash(const char *tablename,
7773 unsigned long bucketsize,
7774 unsigned long numentries,
7777 unsigned int *_hash_shift,
7778 unsigned int *_hash_mask,
7779 unsigned long low_limit,
7780 unsigned long high_limit)
7782 unsigned long long max = high_limit;
7783 unsigned long log2qty, size;
7787 /* allow the kernel cmdline to have a say */
7789 /* round applicable memory size up to nearest megabyte */
7790 numentries = nr_kernel_pages;
7791 numentries -= arch_reserved_kernel_pages();
7793 /* It isn't necessary when PAGE_SIZE >= 1MB */
7794 if (PAGE_SHIFT < 20)
7795 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7797 #if __BITS_PER_LONG > 32
7799 unsigned long adapt;
7801 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7802 adapt <<= ADAPT_SCALE_SHIFT)
7807 /* limit to 1 bucket per 2^scale bytes of low memory */
7808 if (scale > PAGE_SHIFT)
7809 numentries >>= (scale - PAGE_SHIFT);
7811 numentries <<= (PAGE_SHIFT - scale);
7813 /* Make sure we've got at least a 0-order allocation.. */
7814 if (unlikely(flags & HASH_SMALL)) {
7815 /* Makes no sense without HASH_EARLY */
7816 WARN_ON(!(flags & HASH_EARLY));
7817 if (!(numentries >> *_hash_shift)) {
7818 numentries = 1UL << *_hash_shift;
7819 BUG_ON(!numentries);
7821 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7822 numentries = PAGE_SIZE / bucketsize;
7824 numentries = roundup_pow_of_two(numentries);
7826 /* limit allocation size to 1/16 total memory by default */
7828 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7829 do_div(max, bucketsize);
7831 max = min(max, 0x80000000ULL);
7833 if (numentries < low_limit)
7834 numentries = low_limit;
7835 if (numentries > max)
7838 log2qty = ilog2(numentries);
7840 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7842 size = bucketsize << log2qty;
7843 if (flags & HASH_EARLY) {
7844 if (flags & HASH_ZERO)
7845 table = memblock_alloc_nopanic(size,
7848 table = memblock_alloc_raw(size,
7850 } else if (hashdist) {
7851 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7854 * If bucketsize is not a power-of-two, we may free
7855 * some pages at the end of hash table which
7856 * alloc_pages_exact() automatically does
7858 if (get_order(size) < MAX_ORDER) {
7859 table = alloc_pages_exact(size, gfp_flags);
7860 kmemleak_alloc(table, size, 1, gfp_flags);
7863 } while (!table && size > PAGE_SIZE && --log2qty);
7866 panic("Failed to allocate %s hash table\n", tablename);
7868 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7869 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7872 *_hash_shift = log2qty;
7874 *_hash_mask = (1 << log2qty) - 1;
7880 * This function checks whether pageblock includes unmovable pages or not.
7881 * If @count is not zero, it is okay to include less @count unmovable pages
7883 * PageLRU check without isolation or lru_lock could race so that
7884 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7885 * check without lock_page also may miss some movable non-lru pages at
7886 * race condition. So you can't expect this function should be exact.
7888 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7889 int migratetype, int flags)
7891 unsigned long pfn, iter, found;
7894 * TODO we could make this much more efficient by not checking every
7895 * page in the range if we know all of them are in MOVABLE_ZONE and
7896 * that the movable zone guarantees that pages are migratable but
7897 * the later is not the case right now unfortunatelly. E.g. movablecore
7898 * can still lead to having bootmem allocations in zone_movable.
7902 * CMA allocations (alloc_contig_range) really need to mark isolate
7903 * CMA pageblocks even when they are not movable in fact so consider
7904 * them movable here.
7906 if (is_migrate_cma(migratetype) &&
7907 is_migrate_cma(get_pageblock_migratetype(page)))
7910 pfn = page_to_pfn(page);
7911 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7912 unsigned long check = pfn + iter;
7914 if (!pfn_valid_within(check))
7917 page = pfn_to_page(check);
7919 if (PageReserved(page))
7923 * If the zone is movable and we have ruled out all reserved
7924 * pages then it should be reasonably safe to assume the rest
7927 if (zone_idx(zone) == ZONE_MOVABLE)
7931 * Hugepages are not in LRU lists, but they're movable.
7932 * We need not scan over tail pages bacause we don't
7933 * handle each tail page individually in migration.
7935 if (PageHuge(page)) {
7936 struct page *head = compound_head(page);
7937 unsigned int skip_pages;
7939 if (!hugepage_migration_supported(page_hstate(head)))
7942 skip_pages = (1 << compound_order(head)) - (page - head);
7943 iter += skip_pages - 1;
7948 * We can't use page_count without pin a page
7949 * because another CPU can free compound page.
7950 * This check already skips compound tails of THP
7951 * because their page->_refcount is zero at all time.
7953 if (!page_ref_count(page)) {
7954 if (PageBuddy(page))
7955 iter += (1 << page_order(page)) - 1;
7960 * The HWPoisoned page may be not in buddy system, and
7961 * page_count() is not 0.
7963 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
7966 if (__PageMovable(page))
7972 * If there are RECLAIMABLE pages, we need to check
7973 * it. But now, memory offline itself doesn't call
7974 * shrink_node_slabs() and it still to be fixed.
7977 * If the page is not RAM, page_count()should be 0.
7978 * we don't need more check. This is an _used_ not-movable page.
7980 * The problematic thing here is PG_reserved pages. PG_reserved
7981 * is set to both of a memory hole page and a _used_ kernel
7989 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
7990 if (flags & REPORT_FAILURE)
7991 dump_page(pfn_to_page(pfn+iter), "unmovable page");
7995 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7997 static unsigned long pfn_max_align_down(unsigned long pfn)
7999 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8000 pageblock_nr_pages) - 1);
8003 static unsigned long pfn_max_align_up(unsigned long pfn)
8005 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8006 pageblock_nr_pages));
8009 /* [start, end) must belong to a single zone. */
8010 static int __alloc_contig_migrate_range(struct compact_control *cc,
8011 unsigned long start, unsigned long end)
8013 /* This function is based on compact_zone() from compaction.c. */
8014 unsigned long nr_reclaimed;
8015 unsigned long pfn = start;
8016 unsigned int tries = 0;
8021 while (pfn < end || !list_empty(&cc->migratepages)) {
8022 if (fatal_signal_pending(current)) {
8027 if (list_empty(&cc->migratepages)) {
8028 cc->nr_migratepages = 0;
8029 pfn = isolate_migratepages_range(cc, pfn, end);
8035 } else if (++tries == 5) {
8036 ret = ret < 0 ? ret : -EBUSY;
8040 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8042 cc->nr_migratepages -= nr_reclaimed;
8044 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8045 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8048 putback_movable_pages(&cc->migratepages);
8055 * alloc_contig_range() -- tries to allocate given range of pages
8056 * @start: start PFN to allocate
8057 * @end: one-past-the-last PFN to allocate
8058 * @migratetype: migratetype of the underlaying pageblocks (either
8059 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8060 * in range must have the same migratetype and it must
8061 * be either of the two.
8062 * @gfp_mask: GFP mask to use during compaction
8064 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8065 * aligned. The PFN range must belong to a single zone.
8067 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8068 * pageblocks in the range. Once isolated, the pageblocks should not
8069 * be modified by others.
8071 * Returns zero on success or negative error code. On success all
8072 * pages which PFN is in [start, end) are allocated for the caller and
8073 * need to be freed with free_contig_range().
8075 int alloc_contig_range(unsigned long start, unsigned long end,
8076 unsigned migratetype, gfp_t gfp_mask)
8078 unsigned long outer_start, outer_end;
8082 struct compact_control cc = {
8083 .nr_migratepages = 0,
8085 .zone = page_zone(pfn_to_page(start)),
8086 .mode = MIGRATE_SYNC,
8087 .ignore_skip_hint = true,
8088 .no_set_skip_hint = true,
8089 .gfp_mask = current_gfp_context(gfp_mask),
8091 INIT_LIST_HEAD(&cc.migratepages);
8094 * What we do here is we mark all pageblocks in range as
8095 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8096 * have different sizes, and due to the way page allocator
8097 * work, we align the range to biggest of the two pages so
8098 * that page allocator won't try to merge buddies from
8099 * different pageblocks and change MIGRATE_ISOLATE to some
8100 * other migration type.
8102 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8103 * migrate the pages from an unaligned range (ie. pages that
8104 * we are interested in). This will put all the pages in
8105 * range back to page allocator as MIGRATE_ISOLATE.
8107 * When this is done, we take the pages in range from page
8108 * allocator removing them from the buddy system. This way
8109 * page allocator will never consider using them.
8111 * This lets us mark the pageblocks back as
8112 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8113 * aligned range but not in the unaligned, original range are
8114 * put back to page allocator so that buddy can use them.
8117 ret = start_isolate_page_range(pfn_max_align_down(start),
8118 pfn_max_align_up(end), migratetype, 0);
8123 * In case of -EBUSY, we'd like to know which page causes problem.
8124 * So, just fall through. test_pages_isolated() has a tracepoint
8125 * which will report the busy page.
8127 * It is possible that busy pages could become available before
8128 * the call to test_pages_isolated, and the range will actually be
8129 * allocated. So, if we fall through be sure to clear ret so that
8130 * -EBUSY is not accidentally used or returned to caller.
8132 ret = __alloc_contig_migrate_range(&cc, start, end);
8133 if (ret && ret != -EBUSY)
8138 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8139 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8140 * more, all pages in [start, end) are free in page allocator.
8141 * What we are going to do is to allocate all pages from
8142 * [start, end) (that is remove them from page allocator).
8144 * The only problem is that pages at the beginning and at the
8145 * end of interesting range may be not aligned with pages that
8146 * page allocator holds, ie. they can be part of higher order
8147 * pages. Because of this, we reserve the bigger range and
8148 * once this is done free the pages we are not interested in.
8150 * We don't have to hold zone->lock here because the pages are
8151 * isolated thus they won't get removed from buddy.
8154 lru_add_drain_all();
8155 drain_all_pages(cc.zone);
8158 outer_start = start;
8159 while (!PageBuddy(pfn_to_page(outer_start))) {
8160 if (++order >= MAX_ORDER) {
8161 outer_start = start;
8164 outer_start &= ~0UL << order;
8167 if (outer_start != start) {
8168 order = page_order(pfn_to_page(outer_start));
8171 * outer_start page could be small order buddy page and
8172 * it doesn't include start page. Adjust outer_start
8173 * in this case to report failed page properly
8174 * on tracepoint in test_pages_isolated()
8176 if (outer_start + (1UL << order) <= start)
8177 outer_start = start;
8180 /* Make sure the range is really isolated. */
8181 if (test_pages_isolated(outer_start, end, false)) {
8182 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8183 __func__, outer_start, end);
8188 /* Grab isolated pages from freelists. */
8189 outer_end = isolate_freepages_range(&cc, outer_start, end);
8195 /* Free head and tail (if any) */
8196 if (start != outer_start)
8197 free_contig_range(outer_start, start - outer_start);
8198 if (end != outer_end)
8199 free_contig_range(end, outer_end - end);
8202 undo_isolate_page_range(pfn_max_align_down(start),
8203 pfn_max_align_up(end), migratetype);
8207 void free_contig_range(unsigned long pfn, unsigned nr_pages)
8209 unsigned int count = 0;
8211 for (; nr_pages--; pfn++) {
8212 struct page *page = pfn_to_page(pfn);
8214 count += page_count(page) != 1;
8217 WARN(count != 0, "%d pages are still in use!\n", count);
8221 #ifdef CONFIG_MEMORY_HOTPLUG
8223 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8224 * page high values need to be recalulated.
8226 void __meminit zone_pcp_update(struct zone *zone)
8229 mutex_lock(&pcp_batch_high_lock);
8230 for_each_possible_cpu(cpu)
8231 pageset_set_high_and_batch(zone,
8232 per_cpu_ptr(zone->pageset, cpu));
8233 mutex_unlock(&pcp_batch_high_lock);
8237 void zone_pcp_reset(struct zone *zone)
8239 unsigned long flags;
8241 struct per_cpu_pageset *pset;
8243 /* avoid races with drain_pages() */
8244 local_irq_save(flags);
8245 if (zone->pageset != &boot_pageset) {
8246 for_each_online_cpu(cpu) {
8247 pset = per_cpu_ptr(zone->pageset, cpu);
8248 drain_zonestat(zone, pset);
8250 free_percpu(zone->pageset);
8251 zone->pageset = &boot_pageset;
8253 local_irq_restore(flags);
8256 #ifdef CONFIG_MEMORY_HOTREMOVE
8258 * All pages in the range must be in a single zone and isolated
8259 * before calling this.
8262 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8266 unsigned int order, i;
8268 unsigned long flags;
8269 /* find the first valid pfn */
8270 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8275 offline_mem_sections(pfn, end_pfn);
8276 zone = page_zone(pfn_to_page(pfn));
8277 spin_lock_irqsave(&zone->lock, flags);
8279 while (pfn < end_pfn) {
8280 if (!pfn_valid(pfn)) {
8284 page = pfn_to_page(pfn);
8286 * The HWPoisoned page may be not in buddy system, and
8287 * page_count() is not 0.
8289 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8291 SetPageReserved(page);
8295 BUG_ON(page_count(page));
8296 BUG_ON(!PageBuddy(page));
8297 order = page_order(page);
8298 #ifdef CONFIG_DEBUG_VM
8299 pr_info("remove from free list %lx %d %lx\n",
8300 pfn, 1 << order, end_pfn);
8302 list_del(&page->lru);
8303 rmv_page_order(page);
8304 zone->free_area[order].nr_free--;
8305 for (i = 0; i < (1 << order); i++)
8306 SetPageReserved((page+i));
8307 pfn += (1 << order);
8309 spin_unlock_irqrestore(&zone->lock, flags);
8313 bool is_free_buddy_page(struct page *page)
8315 struct zone *zone = page_zone(page);
8316 unsigned long pfn = page_to_pfn(page);
8317 unsigned long flags;
8320 spin_lock_irqsave(&zone->lock, flags);
8321 for (order = 0; order < MAX_ORDER; order++) {
8322 struct page *page_head = page - (pfn & ((1 << order) - 1));
8324 if (PageBuddy(page_head) && page_order(page_head) >= order)
8327 spin_unlock_irqrestore(&zone->lock, flags);
8329 return order < MAX_ORDER;
8332 #ifdef CONFIG_MEMORY_FAILURE
8334 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8335 * test is performed under the zone lock to prevent a race against page
8338 bool set_hwpoison_free_buddy_page(struct page *page)
8340 struct zone *zone = page_zone(page);
8341 unsigned long pfn = page_to_pfn(page);
8342 unsigned long flags;
8344 bool hwpoisoned = false;
8346 spin_lock_irqsave(&zone->lock, flags);
8347 for (order = 0; order < MAX_ORDER; order++) {
8348 struct page *page_head = page - (pfn & ((1 << order) - 1));
8350 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8351 if (!TestSetPageHWPoison(page))
8356 spin_unlock_irqrestore(&zone->lock, flags);