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 */
102 struct work_struct work;
104 DEFINE_MUTEX(pcpu_drain_mutex);
105 DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
107 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
108 volatile unsigned long latent_entropy __latent_entropy;
109 EXPORT_SYMBOL(latent_entropy);
113 * Array of node states.
115 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
116 [N_POSSIBLE] = NODE_MASK_ALL,
117 [N_ONLINE] = { { [0] = 1UL } },
119 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
120 #ifdef CONFIG_HIGHMEM
121 [N_HIGH_MEMORY] = { { [0] = 1UL } },
123 [N_MEMORY] = { { [0] = 1UL } },
124 [N_CPU] = { { [0] = 1UL } },
127 EXPORT_SYMBOL(node_states);
129 atomic_long_t _totalram_pages __read_mostly;
130 EXPORT_SYMBOL(_totalram_pages);
131 unsigned long totalreserve_pages __read_mostly;
132 unsigned long totalcma_pages __read_mostly;
134 int percpu_pagelist_fraction;
135 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
138 * A cached value of the page's pageblock's migratetype, used when the page is
139 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
140 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
141 * Also the migratetype set in the page does not necessarily match the pcplist
142 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
143 * other index - this ensures that it will be put on the correct CMA freelist.
145 static inline int get_pcppage_migratetype(struct page *page)
150 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
152 page->index = migratetype;
155 #ifdef CONFIG_PM_SLEEP
157 * The following functions are used by the suspend/hibernate code to temporarily
158 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
159 * while devices are suspended. To avoid races with the suspend/hibernate code,
160 * they should always be called with system_transition_mutex held
161 * (gfp_allowed_mask also should only be modified with system_transition_mutex
162 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
163 * with that modification).
166 static gfp_t saved_gfp_mask;
168 void pm_restore_gfp_mask(void)
170 WARN_ON(!mutex_is_locked(&system_transition_mutex));
171 if (saved_gfp_mask) {
172 gfp_allowed_mask = saved_gfp_mask;
177 void pm_restrict_gfp_mask(void)
179 WARN_ON(!mutex_is_locked(&system_transition_mutex));
180 WARN_ON(saved_gfp_mask);
181 saved_gfp_mask = gfp_allowed_mask;
182 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
185 bool pm_suspended_storage(void)
187 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
191 #endif /* CONFIG_PM_SLEEP */
193 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
194 unsigned int pageblock_order __read_mostly;
197 static void __free_pages_ok(struct page *page, unsigned int order);
200 * results with 256, 32 in the lowmem_reserve sysctl:
201 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
202 * 1G machine -> (16M dma, 784M normal, 224M high)
203 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
204 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
205 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
207 * TBD: should special case ZONE_DMA32 machines here - in those we normally
208 * don't need any ZONE_NORMAL reservation
210 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
211 #ifdef CONFIG_ZONE_DMA
214 #ifdef CONFIG_ZONE_DMA32
218 #ifdef CONFIG_HIGHMEM
224 EXPORT_SYMBOL(totalram_pages);
226 static char * const zone_names[MAX_NR_ZONES] = {
227 #ifdef CONFIG_ZONE_DMA
230 #ifdef CONFIG_ZONE_DMA32
234 #ifdef CONFIG_HIGHMEM
238 #ifdef CONFIG_ZONE_DEVICE
243 const char * const migratetype_names[MIGRATE_TYPES] = {
251 #ifdef CONFIG_MEMORY_ISOLATION
256 compound_page_dtor * const compound_page_dtors[] = {
259 #ifdef CONFIG_HUGETLB_PAGE
262 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
267 int min_free_kbytes = 1024;
268 int user_min_free_kbytes = -1;
269 #ifdef CONFIG_DISCONTIGMEM
271 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
272 * are not on separate NUMA nodes. Functionally this works but with
273 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
274 * quite small. By default, do not boost watermarks on discontigmem as in
275 * many cases very high-order allocations like THP are likely to be
276 * unsupported and the premature reclaim offsets the advantage of long-term
277 * fragmentation avoidance.
279 int watermark_boost_factor __read_mostly;
281 int watermark_boost_factor __read_mostly = 15000;
283 int watermark_scale_factor = 10;
285 static unsigned long nr_kernel_pages __initdata;
286 static unsigned long nr_all_pages __initdata;
287 static unsigned long dma_reserve __initdata;
289 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
290 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
291 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
292 static unsigned long required_kernelcore __initdata;
293 static unsigned long required_kernelcore_percent __initdata;
294 static unsigned long required_movablecore __initdata;
295 static unsigned long required_movablecore_percent __initdata;
296 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
297 static bool mirrored_kernelcore __meminitdata;
299 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
301 EXPORT_SYMBOL(movable_zone);
302 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
305 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
306 unsigned int nr_online_nodes __read_mostly = 1;
307 EXPORT_SYMBOL(nr_node_ids);
308 EXPORT_SYMBOL(nr_online_nodes);
311 int page_group_by_mobility_disabled __read_mostly;
313 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
315 * During boot we initialize deferred pages on-demand, as needed, but once
316 * page_alloc_init_late() has finished, the deferred pages are all initialized,
317 * and we can permanently disable that path.
319 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
322 * Calling kasan_free_pages() only after deferred memory initialization
323 * has completed. Poisoning pages during deferred memory init will greatly
324 * lengthen the process and cause problem in large memory systems as the
325 * deferred pages initialization is done with interrupt disabled.
327 * Assuming that there will be no reference to those newly initialized
328 * pages before they are ever allocated, this should have no effect on
329 * KASAN memory tracking as the poison will be properly inserted at page
330 * allocation time. The only corner case is when pages are allocated by
331 * on-demand allocation and then freed again before the deferred pages
332 * initialization is done, but this is not likely to happen.
334 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
336 if (!static_branch_unlikely(&deferred_pages))
337 kasan_free_pages(page, order);
340 /* Returns true if the struct page for the pfn is uninitialised */
341 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
343 int nid = early_pfn_to_nid(pfn);
345 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
352 * Returns true when the remaining initialisation should be deferred until
353 * later in the boot cycle when it can be parallelised.
355 static bool __meminit
356 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
358 static unsigned long prev_end_pfn, nr_initialised;
361 * prev_end_pfn static that contains the end of previous zone
362 * No need to protect because called very early in boot before smp_init.
364 if (prev_end_pfn != end_pfn) {
365 prev_end_pfn = end_pfn;
369 /* Always populate low zones for address-constrained allocations */
370 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
374 * We start only with one section of pages, more pages are added as
375 * needed until the rest of deferred pages are initialized.
378 if ((nr_initialised > PAGES_PER_SECTION) &&
379 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
380 NODE_DATA(nid)->first_deferred_pfn = pfn;
386 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
388 static inline bool early_page_uninitialised(unsigned long pfn)
393 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
399 /* Return a pointer to the bitmap storing bits affecting a block of pages */
400 static inline unsigned long *get_pageblock_bitmap(struct page *page,
403 #ifdef CONFIG_SPARSEMEM
404 return __pfn_to_section(pfn)->pageblock_flags;
406 return page_zone(page)->pageblock_flags;
407 #endif /* CONFIG_SPARSEMEM */
410 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
412 #ifdef CONFIG_SPARSEMEM
413 pfn &= (PAGES_PER_SECTION-1);
414 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
416 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
417 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
418 #endif /* CONFIG_SPARSEMEM */
422 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
423 * @page: The page within the block of interest
424 * @pfn: The target page frame number
425 * @end_bitidx: The last bit of interest to retrieve
426 * @mask: mask of bits that the caller is interested in
428 * Return: pageblock_bits flags
430 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
432 unsigned long end_bitidx,
435 unsigned long *bitmap;
436 unsigned long bitidx, word_bitidx;
439 bitmap = get_pageblock_bitmap(page, pfn);
440 bitidx = pfn_to_bitidx(page, pfn);
441 word_bitidx = bitidx / BITS_PER_LONG;
442 bitidx &= (BITS_PER_LONG-1);
444 word = bitmap[word_bitidx];
445 bitidx += end_bitidx;
446 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
449 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
450 unsigned long end_bitidx,
453 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
456 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
458 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
462 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
463 * @page: The page within the block of interest
464 * @flags: The flags to set
465 * @pfn: The target page frame number
466 * @end_bitidx: The last bit of interest
467 * @mask: mask of bits that the caller is interested in
469 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
471 unsigned long end_bitidx,
474 unsigned long *bitmap;
475 unsigned long bitidx, word_bitidx;
476 unsigned long old_word, word;
478 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
479 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
481 bitmap = get_pageblock_bitmap(page, pfn);
482 bitidx = pfn_to_bitidx(page, pfn);
483 word_bitidx = bitidx / BITS_PER_LONG;
484 bitidx &= (BITS_PER_LONG-1);
486 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
488 bitidx += end_bitidx;
489 mask <<= (BITS_PER_LONG - bitidx - 1);
490 flags <<= (BITS_PER_LONG - bitidx - 1);
492 word = READ_ONCE(bitmap[word_bitidx]);
494 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
495 if (word == old_word)
501 void set_pageblock_migratetype(struct page *page, int migratetype)
503 if (unlikely(page_group_by_mobility_disabled &&
504 migratetype < MIGRATE_PCPTYPES))
505 migratetype = MIGRATE_UNMOVABLE;
507 set_pageblock_flags_group(page, (unsigned long)migratetype,
508 PB_migrate, PB_migrate_end);
511 #ifdef CONFIG_DEBUG_VM
512 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
516 unsigned long pfn = page_to_pfn(page);
517 unsigned long sp, start_pfn;
520 seq = zone_span_seqbegin(zone);
521 start_pfn = zone->zone_start_pfn;
522 sp = zone->spanned_pages;
523 if (!zone_spans_pfn(zone, pfn))
525 } while (zone_span_seqretry(zone, seq));
528 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
529 pfn, zone_to_nid(zone), zone->name,
530 start_pfn, start_pfn + sp);
535 static int page_is_consistent(struct zone *zone, struct page *page)
537 if (!pfn_valid_within(page_to_pfn(page)))
539 if (zone != page_zone(page))
545 * Temporary debugging check for pages not lying within a given zone.
547 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
549 if (page_outside_zone_boundaries(zone, page))
551 if (!page_is_consistent(zone, page))
557 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
563 static void bad_page(struct page *page, const char *reason,
564 unsigned long bad_flags)
566 static unsigned long resume;
567 static unsigned long nr_shown;
568 static unsigned long nr_unshown;
571 * Allow a burst of 60 reports, then keep quiet for that minute;
572 * or allow a steady drip of one report per second.
574 if (nr_shown == 60) {
575 if (time_before(jiffies, resume)) {
581 "BUG: Bad page state: %lu messages suppressed\n",
588 resume = jiffies + 60 * HZ;
590 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
591 current->comm, page_to_pfn(page));
592 __dump_page(page, reason);
593 bad_flags &= page->flags;
595 pr_alert("bad because of flags: %#lx(%pGp)\n",
596 bad_flags, &bad_flags);
597 dump_page_owner(page);
602 /* Leave bad fields for debug, except PageBuddy could make trouble */
603 page_mapcount_reset(page); /* remove PageBuddy */
604 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
608 * Higher-order pages are called "compound pages". They are structured thusly:
610 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
612 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
613 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
615 * The first tail page's ->compound_dtor holds the offset in array of compound
616 * page destructors. See compound_page_dtors.
618 * The first tail page's ->compound_order holds the order of allocation.
619 * This usage means that zero-order pages may not be compound.
622 void free_compound_page(struct page *page)
624 __free_pages_ok(page, compound_order(page));
627 void prep_compound_page(struct page *page, unsigned int order)
630 int nr_pages = 1 << order;
632 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
633 set_compound_order(page, order);
635 for (i = 1; i < nr_pages; i++) {
636 struct page *p = page + i;
637 set_page_count(p, 0);
638 p->mapping = TAIL_MAPPING;
639 set_compound_head(p, page);
641 atomic_set(compound_mapcount_ptr(page), -1);
644 #ifdef CONFIG_DEBUG_PAGEALLOC
645 unsigned int _debug_guardpage_minorder;
646 bool _debug_pagealloc_enabled __read_mostly
647 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
648 EXPORT_SYMBOL(_debug_pagealloc_enabled);
649 bool _debug_guardpage_enabled __read_mostly;
651 static int __init early_debug_pagealloc(char *buf)
655 return kstrtobool(buf, &_debug_pagealloc_enabled);
657 early_param("debug_pagealloc", early_debug_pagealloc);
659 static bool need_debug_guardpage(void)
661 /* If we don't use debug_pagealloc, we don't need guard page */
662 if (!debug_pagealloc_enabled())
665 if (!debug_guardpage_minorder())
671 static void init_debug_guardpage(void)
673 if (!debug_pagealloc_enabled())
676 if (!debug_guardpage_minorder())
679 _debug_guardpage_enabled = true;
682 struct page_ext_operations debug_guardpage_ops = {
683 .need = need_debug_guardpage,
684 .init = init_debug_guardpage,
687 static int __init debug_guardpage_minorder_setup(char *buf)
691 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
692 pr_err("Bad debug_guardpage_minorder value\n");
695 _debug_guardpage_minorder = res;
696 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
699 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
701 static inline bool set_page_guard(struct zone *zone, struct page *page,
702 unsigned int order, int migratetype)
704 struct page_ext *page_ext;
706 if (!debug_guardpage_enabled())
709 if (order >= debug_guardpage_minorder())
712 page_ext = lookup_page_ext(page);
713 if (unlikely(!page_ext))
716 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
718 INIT_LIST_HEAD(&page->lru);
719 set_page_private(page, order);
720 /* Guard pages are not available for any usage */
721 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
726 static inline void clear_page_guard(struct zone *zone, struct page *page,
727 unsigned int order, int migratetype)
729 struct page_ext *page_ext;
731 if (!debug_guardpage_enabled())
734 page_ext = lookup_page_ext(page);
735 if (unlikely(!page_ext))
738 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
740 set_page_private(page, 0);
741 if (!is_migrate_isolate(migratetype))
742 __mod_zone_freepage_state(zone, (1 << order), migratetype);
745 struct page_ext_operations debug_guardpage_ops;
746 static inline bool set_page_guard(struct zone *zone, struct page *page,
747 unsigned int order, int migratetype) { return false; }
748 static inline void clear_page_guard(struct zone *zone, struct page *page,
749 unsigned int order, int migratetype) {}
752 static inline void set_page_order(struct page *page, unsigned int order)
754 set_page_private(page, order);
755 __SetPageBuddy(page);
758 static inline void rmv_page_order(struct page *page)
760 __ClearPageBuddy(page);
761 set_page_private(page, 0);
765 * This function checks whether a page is free && is the buddy
766 * we can coalesce a page and its buddy if
767 * (a) the buddy is not in a hole (check before calling!) &&
768 * (b) the buddy is in the buddy system &&
769 * (c) a page and its buddy have the same order &&
770 * (d) a page and its buddy are in the same zone.
772 * For recording whether a page is in the buddy system, we set PageBuddy.
773 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
775 * For recording page's order, we use page_private(page).
777 static inline int page_is_buddy(struct page *page, struct page *buddy,
780 if (page_is_guard(buddy) && page_order(buddy) == order) {
781 if (page_zone_id(page) != page_zone_id(buddy))
784 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
789 if (PageBuddy(buddy) && page_order(buddy) == order) {
791 * zone check is done late to avoid uselessly
792 * calculating zone/node ids for pages that could
795 if (page_zone_id(page) != page_zone_id(buddy))
798 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
805 #ifdef CONFIG_COMPACTION
806 static inline struct capture_control *task_capc(struct zone *zone)
808 struct capture_control *capc = current->capture_control;
811 !(current->flags & PF_KTHREAD) &&
813 capc->cc->zone == zone &&
814 capc->cc->direct_compaction ? capc : NULL;
818 compaction_capture(struct capture_control *capc, struct page *page,
819 int order, int migratetype)
821 if (!capc || order != capc->cc->order)
824 /* Do not accidentally pollute CMA or isolated regions*/
825 if (is_migrate_cma(migratetype) ||
826 is_migrate_isolate(migratetype))
830 * Do not let lower order allocations polluate a movable pageblock.
831 * This might let an unmovable request use a reclaimable pageblock
832 * and vice-versa but no more than normal fallback logic which can
833 * have trouble finding a high-order free page.
835 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
843 static inline struct capture_control *task_capc(struct zone *zone)
849 compaction_capture(struct capture_control *capc, struct page *page,
850 int order, int migratetype)
854 #endif /* CONFIG_COMPACTION */
857 * Freeing function for a buddy system allocator.
859 * The concept of a buddy system is to maintain direct-mapped table
860 * (containing bit values) for memory blocks of various "orders".
861 * The bottom level table contains the map for the smallest allocatable
862 * units of memory (here, pages), and each level above it describes
863 * pairs of units from the levels below, hence, "buddies".
864 * At a high level, all that happens here is marking the table entry
865 * at the bottom level available, and propagating the changes upward
866 * as necessary, plus some accounting needed to play nicely with other
867 * parts of the VM system.
868 * At each level, we keep a list of pages, which are heads of continuous
869 * free pages of length of (1 << order) and marked with PageBuddy.
870 * Page's order is recorded in page_private(page) field.
871 * So when we are allocating or freeing one, we can derive the state of the
872 * other. That is, if we allocate a small block, and both were
873 * free, the remainder of the region must be split into blocks.
874 * If a block is freed, and its buddy is also free, then this
875 * triggers coalescing into a block of larger size.
880 static inline void __free_one_page(struct page *page,
882 struct zone *zone, unsigned int order,
885 unsigned long combined_pfn;
886 unsigned long uninitialized_var(buddy_pfn);
888 unsigned int max_order;
889 struct capture_control *capc = task_capc(zone);
891 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
893 VM_BUG_ON(!zone_is_initialized(zone));
894 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
896 VM_BUG_ON(migratetype == -1);
897 if (likely(!is_migrate_isolate(migratetype)))
898 __mod_zone_freepage_state(zone, 1 << order, migratetype);
900 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
901 VM_BUG_ON_PAGE(bad_range(zone, page), page);
904 while (order < max_order - 1) {
905 if (compaction_capture(capc, page, order, migratetype)) {
906 __mod_zone_freepage_state(zone, -(1 << order),
910 buddy_pfn = __find_buddy_pfn(pfn, order);
911 buddy = page + (buddy_pfn - pfn);
913 if (!pfn_valid_within(buddy_pfn))
915 if (!page_is_buddy(page, buddy, order))
918 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
919 * merge with it and move up one order.
921 if (page_is_guard(buddy)) {
922 clear_page_guard(zone, buddy, order, migratetype);
924 list_del(&buddy->lru);
925 zone->free_area[order].nr_free--;
926 rmv_page_order(buddy);
928 combined_pfn = buddy_pfn & pfn;
929 page = page + (combined_pfn - pfn);
933 if (max_order < MAX_ORDER) {
934 /* If we are here, it means order is >= pageblock_order.
935 * We want to prevent merge between freepages on isolate
936 * pageblock and normal pageblock. Without this, pageblock
937 * isolation could cause incorrect freepage or CMA accounting.
939 * We don't want to hit this code for the more frequent
942 if (unlikely(has_isolate_pageblock(zone))) {
945 buddy_pfn = __find_buddy_pfn(pfn, order);
946 buddy = page + (buddy_pfn - pfn);
947 buddy_mt = get_pageblock_migratetype(buddy);
949 if (migratetype != buddy_mt
950 && (is_migrate_isolate(migratetype) ||
951 is_migrate_isolate(buddy_mt)))
955 goto continue_merging;
959 set_page_order(page, order);
962 * If this is not the largest possible page, check if the buddy
963 * of the next-highest order is free. If it is, it's possible
964 * that pages are being freed that will coalesce soon. In case,
965 * that is happening, add the free page to the tail of the list
966 * so it's less likely to be used soon and more likely to be merged
967 * as a higher order page
969 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
970 struct page *higher_page, *higher_buddy;
971 combined_pfn = buddy_pfn & pfn;
972 higher_page = page + (combined_pfn - pfn);
973 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
974 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
975 if (pfn_valid_within(buddy_pfn) &&
976 page_is_buddy(higher_page, higher_buddy, order + 1)) {
977 list_add_tail(&page->lru,
978 &zone->free_area[order].free_list[migratetype]);
983 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
985 zone->free_area[order].nr_free++;
989 * A bad page could be due to a number of fields. Instead of multiple branches,
990 * try and check multiple fields with one check. The caller must do a detailed
991 * check if necessary.
993 static inline bool page_expected_state(struct page *page,
994 unsigned long check_flags)
996 if (unlikely(atomic_read(&page->_mapcount) != -1))
999 if (unlikely((unsigned long)page->mapping |
1000 page_ref_count(page) |
1002 (unsigned long)page->mem_cgroup |
1004 (page->flags & check_flags)))
1010 static void free_pages_check_bad(struct page *page)
1012 const char *bad_reason;
1013 unsigned long bad_flags;
1018 if (unlikely(atomic_read(&page->_mapcount) != -1))
1019 bad_reason = "nonzero mapcount";
1020 if (unlikely(page->mapping != NULL))
1021 bad_reason = "non-NULL mapping";
1022 if (unlikely(page_ref_count(page) != 0))
1023 bad_reason = "nonzero _refcount";
1024 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1025 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1026 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1029 if (unlikely(page->mem_cgroup))
1030 bad_reason = "page still charged to cgroup";
1032 bad_page(page, bad_reason, bad_flags);
1035 static inline int free_pages_check(struct page *page)
1037 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1040 /* Something has gone sideways, find it */
1041 free_pages_check_bad(page);
1045 static int free_tail_pages_check(struct page *head_page, struct page *page)
1050 * We rely page->lru.next never has bit 0 set, unless the page
1051 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1053 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1055 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1059 switch (page - head_page) {
1061 /* the first tail page: ->mapping may be compound_mapcount() */
1062 if (unlikely(compound_mapcount(page))) {
1063 bad_page(page, "nonzero compound_mapcount", 0);
1069 * the second tail page: ->mapping is
1070 * deferred_list.next -- ignore value.
1074 if (page->mapping != TAIL_MAPPING) {
1075 bad_page(page, "corrupted mapping in tail page", 0);
1080 if (unlikely(!PageTail(page))) {
1081 bad_page(page, "PageTail not set", 0);
1084 if (unlikely(compound_head(page) != head_page)) {
1085 bad_page(page, "compound_head not consistent", 0);
1090 page->mapping = NULL;
1091 clear_compound_head(page);
1095 static __always_inline bool free_pages_prepare(struct page *page,
1096 unsigned int order, bool check_free)
1100 VM_BUG_ON_PAGE(PageTail(page), page);
1102 trace_mm_page_free(page, order);
1105 * Check tail pages before head page information is cleared to
1106 * avoid checking PageCompound for order-0 pages.
1108 if (unlikely(order)) {
1109 bool compound = PageCompound(page);
1112 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1115 ClearPageDoubleMap(page);
1116 for (i = 1; i < (1 << order); i++) {
1118 bad += free_tail_pages_check(page, page + i);
1119 if (unlikely(free_pages_check(page + i))) {
1123 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1126 if (PageMappingFlags(page))
1127 page->mapping = NULL;
1128 if (memcg_kmem_enabled() && PageKmemcg(page))
1129 __memcg_kmem_uncharge(page, order);
1131 bad += free_pages_check(page);
1135 page_cpupid_reset_last(page);
1136 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1137 reset_page_owner(page, order);
1139 if (!PageHighMem(page)) {
1140 debug_check_no_locks_freed(page_address(page),
1141 PAGE_SIZE << order);
1142 debug_check_no_obj_freed(page_address(page),
1143 PAGE_SIZE << order);
1145 arch_free_page(page, order);
1146 kernel_poison_pages(page, 1 << order, 0);
1147 if (debug_pagealloc_enabled())
1148 kernel_map_pages(page, 1 << order, 0);
1150 kasan_free_nondeferred_pages(page, order);
1155 #ifdef CONFIG_DEBUG_VM
1156 static inline bool free_pcp_prepare(struct page *page)
1158 return free_pages_prepare(page, 0, true);
1161 static inline bool bulkfree_pcp_prepare(struct page *page)
1166 static bool free_pcp_prepare(struct page *page)
1168 return free_pages_prepare(page, 0, false);
1171 static bool bulkfree_pcp_prepare(struct page *page)
1173 return free_pages_check(page);
1175 #endif /* CONFIG_DEBUG_VM */
1177 static inline void prefetch_buddy(struct page *page)
1179 unsigned long pfn = page_to_pfn(page);
1180 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1181 struct page *buddy = page + (buddy_pfn - pfn);
1187 * Frees a number of pages from the PCP lists
1188 * Assumes all pages on list are in same zone, and of same order.
1189 * count is the number of pages to free.
1191 * If the zone was previously in an "all pages pinned" state then look to
1192 * see if this freeing clears that state.
1194 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1195 * pinned" detection logic.
1197 static void free_pcppages_bulk(struct zone *zone, int count,
1198 struct per_cpu_pages *pcp)
1200 int migratetype = 0;
1202 int prefetch_nr = 0;
1203 bool isolated_pageblocks;
1204 struct page *page, *tmp;
1208 struct list_head *list;
1211 * Remove pages from lists in a round-robin fashion. A
1212 * batch_free count is maintained that is incremented when an
1213 * empty list is encountered. This is so more pages are freed
1214 * off fuller lists instead of spinning excessively around empty
1219 if (++migratetype == MIGRATE_PCPTYPES)
1221 list = &pcp->lists[migratetype];
1222 } while (list_empty(list));
1224 /* This is the only non-empty list. Free them all. */
1225 if (batch_free == MIGRATE_PCPTYPES)
1229 page = list_last_entry(list, struct page, lru);
1230 /* must delete to avoid corrupting pcp list */
1231 list_del(&page->lru);
1234 if (bulkfree_pcp_prepare(page))
1237 list_add_tail(&page->lru, &head);
1240 * We are going to put the page back to the global
1241 * pool, prefetch its buddy to speed up later access
1242 * under zone->lock. It is believed the overhead of
1243 * an additional test and calculating buddy_pfn here
1244 * can be offset by reduced memory latency later. To
1245 * avoid excessive prefetching due to large count, only
1246 * prefetch buddy for the first pcp->batch nr of pages.
1248 if (prefetch_nr++ < pcp->batch)
1249 prefetch_buddy(page);
1250 } while (--count && --batch_free && !list_empty(list));
1253 spin_lock(&zone->lock);
1254 isolated_pageblocks = has_isolate_pageblock(zone);
1257 * Use safe version since after __free_one_page(),
1258 * page->lru.next will not point to original list.
1260 list_for_each_entry_safe(page, tmp, &head, lru) {
1261 int mt = get_pcppage_migratetype(page);
1262 /* MIGRATE_ISOLATE page should not go to pcplists */
1263 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1264 /* Pageblock could have been isolated meanwhile */
1265 if (unlikely(isolated_pageblocks))
1266 mt = get_pageblock_migratetype(page);
1268 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1269 trace_mm_page_pcpu_drain(page, 0, mt);
1271 spin_unlock(&zone->lock);
1274 static void free_one_page(struct zone *zone,
1275 struct page *page, unsigned long pfn,
1279 spin_lock(&zone->lock);
1280 if (unlikely(has_isolate_pageblock(zone) ||
1281 is_migrate_isolate(migratetype))) {
1282 migratetype = get_pfnblock_migratetype(page, pfn);
1284 __free_one_page(page, pfn, zone, order, migratetype);
1285 spin_unlock(&zone->lock);
1288 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1289 unsigned long zone, int nid)
1291 mm_zero_struct_page(page);
1292 set_page_links(page, zone, nid, pfn);
1293 init_page_count(page);
1294 page_mapcount_reset(page);
1295 page_cpupid_reset_last(page);
1296 page_kasan_tag_reset(page);
1298 INIT_LIST_HEAD(&page->lru);
1299 #ifdef WANT_PAGE_VIRTUAL
1300 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1301 if (!is_highmem_idx(zone))
1302 set_page_address(page, __va(pfn << PAGE_SHIFT));
1306 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1307 static void __meminit init_reserved_page(unsigned long pfn)
1312 if (!early_page_uninitialised(pfn))
1315 nid = early_pfn_to_nid(pfn);
1316 pgdat = NODE_DATA(nid);
1318 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1319 struct zone *zone = &pgdat->node_zones[zid];
1321 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1324 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1327 static inline void init_reserved_page(unsigned long pfn)
1330 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1333 * Initialised pages do not have PageReserved set. This function is
1334 * called for each range allocated by the bootmem allocator and
1335 * marks the pages PageReserved. The remaining valid pages are later
1336 * sent to the buddy page allocator.
1338 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1340 unsigned long start_pfn = PFN_DOWN(start);
1341 unsigned long end_pfn = PFN_UP(end);
1343 for (; start_pfn < end_pfn; start_pfn++) {
1344 if (pfn_valid(start_pfn)) {
1345 struct page *page = pfn_to_page(start_pfn);
1347 init_reserved_page(start_pfn);
1349 /* Avoid false-positive PageTail() */
1350 INIT_LIST_HEAD(&page->lru);
1353 * no need for atomic set_bit because the struct
1354 * page is not visible yet so nobody should
1357 __SetPageReserved(page);
1362 static void __free_pages_ok(struct page *page, unsigned int order)
1364 unsigned long flags;
1366 unsigned long pfn = page_to_pfn(page);
1368 if (!free_pages_prepare(page, order, true))
1371 migratetype = get_pfnblock_migratetype(page, pfn);
1372 local_irq_save(flags);
1373 __count_vm_events(PGFREE, 1 << order);
1374 free_one_page(page_zone(page), page, pfn, order, migratetype);
1375 local_irq_restore(flags);
1378 void __free_pages_core(struct page *page, unsigned int order)
1380 unsigned int nr_pages = 1 << order;
1381 struct page *p = page;
1385 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1387 __ClearPageReserved(p);
1388 set_page_count(p, 0);
1390 __ClearPageReserved(p);
1391 set_page_count(p, 0);
1393 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1394 set_page_refcounted(page);
1395 __free_pages(page, order);
1398 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1399 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1401 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1403 int __meminit early_pfn_to_nid(unsigned long pfn)
1405 static DEFINE_SPINLOCK(early_pfn_lock);
1408 spin_lock(&early_pfn_lock);
1409 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1411 nid = first_online_node;
1412 spin_unlock(&early_pfn_lock);
1418 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1419 static inline bool __meminit __maybe_unused
1420 meminit_pfn_in_nid(unsigned long pfn, int node,
1421 struct mminit_pfnnid_cache *state)
1425 nid = __early_pfn_to_nid(pfn, state);
1426 if (nid >= 0 && nid != node)
1431 /* Only safe to use early in boot when initialisation is single-threaded */
1432 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1434 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1439 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1443 static inline bool __meminit __maybe_unused
1444 meminit_pfn_in_nid(unsigned long pfn, int node,
1445 struct mminit_pfnnid_cache *state)
1452 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1455 if (early_page_uninitialised(pfn))
1457 __free_pages_core(page, order);
1461 * Check that the whole (or subset of) a pageblock given by the interval of
1462 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1463 * with the migration of free compaction scanner. The scanners then need to
1464 * use only pfn_valid_within() check for arches that allow holes within
1467 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1469 * It's possible on some configurations to have a setup like node0 node1 node0
1470 * i.e. it's possible that all pages within a zones range of pages do not
1471 * belong to a single zone. We assume that a border between node0 and node1
1472 * can occur within a single pageblock, but not a node0 node1 node0
1473 * interleaving within a single pageblock. It is therefore sufficient to check
1474 * the first and last page of a pageblock and avoid checking each individual
1475 * page in a pageblock.
1477 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1478 unsigned long end_pfn, struct zone *zone)
1480 struct page *start_page;
1481 struct page *end_page;
1483 /* end_pfn is one past the range we are checking */
1486 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1489 start_page = pfn_to_online_page(start_pfn);
1493 if (page_zone(start_page) != zone)
1496 end_page = pfn_to_page(end_pfn);
1498 /* This gives a shorter code than deriving page_zone(end_page) */
1499 if (page_zone_id(start_page) != page_zone_id(end_page))
1505 void set_zone_contiguous(struct zone *zone)
1507 unsigned long block_start_pfn = zone->zone_start_pfn;
1508 unsigned long block_end_pfn;
1510 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1511 for (; block_start_pfn < zone_end_pfn(zone);
1512 block_start_pfn = block_end_pfn,
1513 block_end_pfn += pageblock_nr_pages) {
1515 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1517 if (!__pageblock_pfn_to_page(block_start_pfn,
1518 block_end_pfn, zone))
1522 /* We confirm that there is no hole */
1523 zone->contiguous = true;
1526 void clear_zone_contiguous(struct zone *zone)
1528 zone->contiguous = false;
1531 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1532 static void __init deferred_free_range(unsigned long pfn,
1533 unsigned long nr_pages)
1541 page = pfn_to_page(pfn);
1543 /* Free a large naturally-aligned chunk if possible */
1544 if (nr_pages == pageblock_nr_pages &&
1545 (pfn & (pageblock_nr_pages - 1)) == 0) {
1546 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1547 __free_pages_core(page, pageblock_order);
1551 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1552 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1553 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1554 __free_pages_core(page, 0);
1558 /* Completion tracking for deferred_init_memmap() threads */
1559 static atomic_t pgdat_init_n_undone __initdata;
1560 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1562 static inline void __init pgdat_init_report_one_done(void)
1564 if (atomic_dec_and_test(&pgdat_init_n_undone))
1565 complete(&pgdat_init_all_done_comp);
1569 * Returns true if page needs to be initialized or freed to buddy allocator.
1571 * First we check if pfn is valid on architectures where it is possible to have
1572 * holes within pageblock_nr_pages. On systems where it is not possible, this
1573 * function is optimized out.
1575 * Then, we check if a current large page is valid by only checking the validity
1578 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1579 * within a node: a pfn is between start and end of a node, but does not belong
1580 * to this memory node.
1582 static inline bool __init
1583 deferred_pfn_valid(int nid, unsigned long pfn,
1584 struct mminit_pfnnid_cache *nid_init_state)
1586 if (!pfn_valid_within(pfn))
1588 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1590 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1596 * Free pages to buddy allocator. Try to free aligned pages in
1597 * pageblock_nr_pages sizes.
1599 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1600 unsigned long end_pfn)
1602 struct mminit_pfnnid_cache nid_init_state = { };
1603 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1604 unsigned long nr_free = 0;
1606 for (; pfn < end_pfn; pfn++) {
1607 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1608 deferred_free_range(pfn - nr_free, nr_free);
1610 } else if (!(pfn & nr_pgmask)) {
1611 deferred_free_range(pfn - nr_free, nr_free);
1613 touch_nmi_watchdog();
1618 /* Free the last block of pages to allocator */
1619 deferred_free_range(pfn - nr_free, nr_free);
1623 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1624 * by performing it only once every pageblock_nr_pages.
1625 * Return number of pages initialized.
1627 static unsigned long __init deferred_init_pages(int nid, int zid,
1629 unsigned long end_pfn)
1631 struct mminit_pfnnid_cache nid_init_state = { };
1632 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1633 unsigned long nr_pages = 0;
1634 struct page *page = NULL;
1636 for (; pfn < end_pfn; pfn++) {
1637 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1640 } else if (!page || !(pfn & nr_pgmask)) {
1641 page = pfn_to_page(pfn);
1642 touch_nmi_watchdog();
1646 __init_single_page(page, pfn, zid, nid);
1652 /* Initialise remaining memory on a node */
1653 static int __init deferred_init_memmap(void *data)
1655 pg_data_t *pgdat = data;
1656 int nid = pgdat->node_id;
1657 unsigned long start = jiffies;
1658 unsigned long nr_pages = 0;
1659 unsigned long spfn, epfn, first_init_pfn, flags;
1660 phys_addr_t spa, epa;
1663 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1666 /* Bind memory initialisation thread to a local node if possible */
1667 if (!cpumask_empty(cpumask))
1668 set_cpus_allowed_ptr(current, cpumask);
1670 pgdat_resize_lock(pgdat, &flags);
1671 first_init_pfn = pgdat->first_deferred_pfn;
1672 if (first_init_pfn == ULONG_MAX) {
1673 pgdat_resize_unlock(pgdat, &flags);
1674 pgdat_init_report_one_done();
1678 /* Sanity check boundaries */
1679 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1680 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1681 pgdat->first_deferred_pfn = ULONG_MAX;
1683 /* Only the highest zone is deferred so find it */
1684 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1685 zone = pgdat->node_zones + zid;
1686 if (first_init_pfn < zone_end_pfn(zone))
1689 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1692 * Initialize and free pages. We do it in two loops: first we initialize
1693 * struct page, than free to buddy allocator, because while we are
1694 * freeing pages we can access pages that are ahead (computing buddy
1695 * page in __free_one_page()).
1697 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1698 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1699 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1700 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1702 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1703 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1704 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1705 deferred_free_pages(nid, zid, spfn, epfn);
1707 pgdat_resize_unlock(pgdat, &flags);
1709 /* Sanity check that the next zone really is unpopulated */
1710 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1712 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1713 jiffies_to_msecs(jiffies - start));
1715 pgdat_init_report_one_done();
1720 * If this zone has deferred pages, try to grow it by initializing enough
1721 * deferred pages to satisfy the allocation specified by order, rounded up to
1722 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1723 * of SECTION_SIZE bytes by initializing struct pages in increments of
1724 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1726 * Return true when zone was grown, otherwise return false. We return true even
1727 * when we grow less than requested, to let the caller decide if there are
1728 * enough pages to satisfy the allocation.
1730 * Note: We use noinline because this function is needed only during boot, and
1731 * it is called from a __ref function _deferred_grow_zone. This way we are
1732 * making sure that it is not inlined into permanent text section.
1734 static noinline bool __init
1735 deferred_grow_zone(struct zone *zone, unsigned int order)
1737 int zid = zone_idx(zone);
1738 int nid = zone_to_nid(zone);
1739 pg_data_t *pgdat = NODE_DATA(nid);
1740 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1741 unsigned long nr_pages = 0;
1742 unsigned long first_init_pfn, spfn, epfn, t, flags;
1743 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1744 phys_addr_t spa, epa;
1747 /* Only the last zone may have deferred pages */
1748 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1751 pgdat_resize_lock(pgdat, &flags);
1754 * If deferred pages have been initialized while we were waiting for
1755 * the lock, return true, as the zone was grown. The caller will retry
1756 * this zone. We won't return to this function since the caller also
1757 * has this static branch.
1759 if (!static_branch_unlikely(&deferred_pages)) {
1760 pgdat_resize_unlock(pgdat, &flags);
1765 * If someone grew this zone while we were waiting for spinlock, return
1766 * true, as there might be enough pages already.
1768 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1769 pgdat_resize_unlock(pgdat, &flags);
1773 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1775 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1776 pgdat_resize_unlock(pgdat, &flags);
1780 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1781 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1782 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1784 while (spfn < epfn && nr_pages < nr_pages_needed) {
1785 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1786 first_deferred_pfn = min(t, epfn);
1787 nr_pages += deferred_init_pages(nid, zid, spfn,
1788 first_deferred_pfn);
1789 spfn = first_deferred_pfn;
1792 if (nr_pages >= nr_pages_needed)
1796 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1797 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1798 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1799 deferred_free_pages(nid, zid, spfn, epfn);
1801 if (first_deferred_pfn == epfn)
1804 pgdat->first_deferred_pfn = first_deferred_pfn;
1805 pgdat_resize_unlock(pgdat, &flags);
1807 return nr_pages > 0;
1811 * deferred_grow_zone() is __init, but it is called from
1812 * get_page_from_freelist() during early boot until deferred_pages permanently
1813 * disables this call. This is why we have refdata wrapper to avoid warning,
1814 * and to ensure that the function body gets unloaded.
1817 _deferred_grow_zone(struct zone *zone, unsigned int order)
1819 return deferred_grow_zone(zone, order);
1822 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1824 void __init page_alloc_init_late(void)
1828 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1831 /* There will be num_node_state(N_MEMORY) threads */
1832 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1833 for_each_node_state(nid, N_MEMORY) {
1834 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1837 /* Block until all are initialised */
1838 wait_for_completion(&pgdat_init_all_done_comp);
1841 * We initialized the rest of the deferred pages. Permanently disable
1842 * on-demand struct page initialization.
1844 static_branch_disable(&deferred_pages);
1846 /* Reinit limits that are based on free pages after the kernel is up */
1847 files_maxfiles_init();
1849 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1850 /* Discard memblock private memory */
1854 for_each_populated_zone(zone)
1855 set_zone_contiguous(zone);
1859 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1860 void __init init_cma_reserved_pageblock(struct page *page)
1862 unsigned i = pageblock_nr_pages;
1863 struct page *p = page;
1866 __ClearPageReserved(p);
1867 set_page_count(p, 0);
1870 set_pageblock_migratetype(page, MIGRATE_CMA);
1872 if (pageblock_order >= MAX_ORDER) {
1873 i = pageblock_nr_pages;
1876 set_page_refcounted(p);
1877 __free_pages(p, MAX_ORDER - 1);
1878 p += MAX_ORDER_NR_PAGES;
1879 } while (i -= MAX_ORDER_NR_PAGES);
1881 set_page_refcounted(page);
1882 __free_pages(page, pageblock_order);
1885 adjust_managed_page_count(page, pageblock_nr_pages);
1890 * The order of subdivision here is critical for the IO subsystem.
1891 * Please do not alter this order without good reasons and regression
1892 * testing. Specifically, as large blocks of memory are subdivided,
1893 * the order in which smaller blocks are delivered depends on the order
1894 * they're subdivided in this function. This is the primary factor
1895 * influencing the order in which pages are delivered to the IO
1896 * subsystem according to empirical testing, and this is also justified
1897 * by considering the behavior of a buddy system containing a single
1898 * large block of memory acted on by a series of small allocations.
1899 * This behavior is a critical factor in sglist merging's success.
1903 static inline void expand(struct zone *zone, struct page *page,
1904 int low, int high, struct free_area *area,
1907 unsigned long size = 1 << high;
1909 while (high > low) {
1913 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1916 * Mark as guard pages (or page), that will allow to
1917 * merge back to allocator when buddy will be freed.
1918 * Corresponding page table entries will not be touched,
1919 * pages will stay not present in virtual address space
1921 if (set_page_guard(zone, &page[size], high, migratetype))
1924 list_add(&page[size].lru, &area->free_list[migratetype]);
1926 set_page_order(&page[size], high);
1930 static void check_new_page_bad(struct page *page)
1932 const char *bad_reason = NULL;
1933 unsigned long bad_flags = 0;
1935 if (unlikely(atomic_read(&page->_mapcount) != -1))
1936 bad_reason = "nonzero mapcount";
1937 if (unlikely(page->mapping != NULL))
1938 bad_reason = "non-NULL mapping";
1939 if (unlikely(page_ref_count(page) != 0))
1940 bad_reason = "nonzero _count";
1941 if (unlikely(page->flags & __PG_HWPOISON)) {
1942 bad_reason = "HWPoisoned (hardware-corrupted)";
1943 bad_flags = __PG_HWPOISON;
1944 /* Don't complain about hwpoisoned pages */
1945 page_mapcount_reset(page); /* remove PageBuddy */
1948 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1949 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1950 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1953 if (unlikely(page->mem_cgroup))
1954 bad_reason = "page still charged to cgroup";
1956 bad_page(page, bad_reason, bad_flags);
1960 * This page is about to be returned from the page allocator
1962 static inline int check_new_page(struct page *page)
1964 if (likely(page_expected_state(page,
1965 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1968 check_new_page_bad(page);
1972 static inline bool free_pages_prezeroed(void)
1974 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1975 page_poisoning_enabled();
1978 #ifdef CONFIG_DEBUG_VM
1979 static bool check_pcp_refill(struct page *page)
1984 static bool check_new_pcp(struct page *page)
1986 return check_new_page(page);
1989 static bool check_pcp_refill(struct page *page)
1991 return check_new_page(page);
1993 static bool check_new_pcp(struct page *page)
1997 #endif /* CONFIG_DEBUG_VM */
1999 static bool check_new_pages(struct page *page, unsigned int order)
2002 for (i = 0; i < (1 << order); i++) {
2003 struct page *p = page + i;
2005 if (unlikely(check_new_page(p)))
2012 inline void post_alloc_hook(struct page *page, unsigned int order,
2015 set_page_private(page, 0);
2016 set_page_refcounted(page);
2018 arch_alloc_page(page, order);
2019 if (debug_pagealloc_enabled())
2020 kernel_map_pages(page, 1 << order, 1);
2021 kasan_alloc_pages(page, order);
2022 kernel_poison_pages(page, 1 << order, 1);
2023 set_page_owner(page, order, gfp_flags);
2026 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2027 unsigned int alloc_flags)
2031 post_alloc_hook(page, order, gfp_flags);
2033 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
2034 for (i = 0; i < (1 << order); i++)
2035 clear_highpage(page + i);
2037 if (order && (gfp_flags & __GFP_COMP))
2038 prep_compound_page(page, order);
2041 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2042 * allocate the page. The expectation is that the caller is taking
2043 * steps that will free more memory. The caller should avoid the page
2044 * being used for !PFMEMALLOC purposes.
2046 if (alloc_flags & ALLOC_NO_WATERMARKS)
2047 set_page_pfmemalloc(page);
2049 clear_page_pfmemalloc(page);
2053 * Go through the free lists for the given migratetype and remove
2054 * the smallest available page from the freelists
2056 static __always_inline
2057 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2060 unsigned int current_order;
2061 struct free_area *area;
2064 /* Find a page of the appropriate size in the preferred list */
2065 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2066 area = &(zone->free_area[current_order]);
2067 page = list_first_entry_or_null(&area->free_list[migratetype],
2071 list_del(&page->lru);
2072 rmv_page_order(page);
2074 expand(zone, page, order, current_order, area, migratetype);
2075 set_pcppage_migratetype(page, migratetype);
2084 * This array describes the order lists are fallen back to when
2085 * the free lists for the desirable migrate type are depleted
2087 static int fallbacks[MIGRATE_TYPES][4] = {
2088 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2089 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2090 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2092 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2094 #ifdef CONFIG_MEMORY_ISOLATION
2095 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2100 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2103 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2106 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2107 unsigned int order) { return NULL; }
2111 * Move the free pages in a range to the free lists of the requested type.
2112 * Note that start_page and end_pages are not aligned on a pageblock
2113 * boundary. If alignment is required, use move_freepages_block()
2115 static int move_freepages(struct zone *zone,
2116 struct page *start_page, struct page *end_page,
2117 int migratetype, int *num_movable)
2121 int pages_moved = 0;
2123 #ifndef CONFIG_HOLES_IN_ZONE
2125 * page_zone is not safe to call in this context when
2126 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2127 * anyway as we check zone boundaries in move_freepages_block().
2128 * Remove at a later date when no bug reports exist related to
2129 * grouping pages by mobility
2131 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2132 pfn_valid(page_to_pfn(end_page)) &&
2133 page_zone(start_page) != page_zone(end_page));
2135 for (page = start_page; page <= end_page;) {
2136 if (!pfn_valid_within(page_to_pfn(page))) {
2141 /* Make sure we are not inadvertently changing nodes */
2142 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2144 if (!PageBuddy(page)) {
2146 * We assume that pages that could be isolated for
2147 * migration are movable. But we don't actually try
2148 * isolating, as that would be expensive.
2151 (PageLRU(page) || __PageMovable(page)))
2158 order = page_order(page);
2159 list_move(&page->lru,
2160 &zone->free_area[order].free_list[migratetype]);
2162 pages_moved += 1 << order;
2168 int move_freepages_block(struct zone *zone, struct page *page,
2169 int migratetype, int *num_movable)
2171 unsigned long start_pfn, end_pfn;
2172 struct page *start_page, *end_page;
2177 start_pfn = page_to_pfn(page);
2178 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2179 start_page = pfn_to_page(start_pfn);
2180 end_page = start_page + pageblock_nr_pages - 1;
2181 end_pfn = start_pfn + pageblock_nr_pages - 1;
2183 /* Do not cross zone boundaries */
2184 if (!zone_spans_pfn(zone, start_pfn))
2186 if (!zone_spans_pfn(zone, end_pfn))
2189 return move_freepages(zone, start_page, end_page, migratetype,
2193 static void change_pageblock_range(struct page *pageblock_page,
2194 int start_order, int migratetype)
2196 int nr_pageblocks = 1 << (start_order - pageblock_order);
2198 while (nr_pageblocks--) {
2199 set_pageblock_migratetype(pageblock_page, migratetype);
2200 pageblock_page += pageblock_nr_pages;
2205 * When we are falling back to another migratetype during allocation, try to
2206 * steal extra free pages from the same pageblocks to satisfy further
2207 * allocations, instead of polluting multiple pageblocks.
2209 * If we are stealing a relatively large buddy page, it is likely there will
2210 * be more free pages in the pageblock, so try to steal them all. For
2211 * reclaimable and unmovable allocations, we steal regardless of page size,
2212 * as fragmentation caused by those allocations polluting movable pageblocks
2213 * is worse than movable allocations stealing from unmovable and reclaimable
2216 static bool can_steal_fallback(unsigned int order, int start_mt)
2219 * Leaving this order check is intended, although there is
2220 * relaxed order check in next check. The reason is that
2221 * we can actually steal whole pageblock if this condition met,
2222 * but, below check doesn't guarantee it and that is just heuristic
2223 * so could be changed anytime.
2225 if (order >= pageblock_order)
2228 if (order >= pageblock_order / 2 ||
2229 start_mt == MIGRATE_RECLAIMABLE ||
2230 start_mt == MIGRATE_UNMOVABLE ||
2231 page_group_by_mobility_disabled)
2237 static inline void boost_watermark(struct zone *zone)
2239 unsigned long max_boost;
2241 if (!watermark_boost_factor)
2244 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2245 watermark_boost_factor, 10000);
2248 * high watermark may be uninitialised if fragmentation occurs
2249 * very early in boot so do not boost. We do not fall
2250 * through and boost by pageblock_nr_pages as failing
2251 * allocations that early means that reclaim is not going
2252 * to help and it may even be impossible to reclaim the
2253 * boosted watermark resulting in a hang.
2258 max_boost = max(pageblock_nr_pages, max_boost);
2260 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2265 * This function implements actual steal behaviour. If order is large enough,
2266 * we can steal whole pageblock. If not, we first move freepages in this
2267 * pageblock to our migratetype and determine how many already-allocated pages
2268 * are there in the pageblock with a compatible migratetype. If at least half
2269 * of pages are free or compatible, we can change migratetype of the pageblock
2270 * itself, so pages freed in the future will be put on the correct free list.
2272 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2273 unsigned int alloc_flags, int start_type, bool whole_block)
2275 unsigned int current_order = page_order(page);
2276 struct free_area *area;
2277 int free_pages, movable_pages, alike_pages;
2280 old_block_type = get_pageblock_migratetype(page);
2283 * This can happen due to races and we want to prevent broken
2284 * highatomic accounting.
2286 if (is_migrate_highatomic(old_block_type))
2289 /* Take ownership for orders >= pageblock_order */
2290 if (current_order >= pageblock_order) {
2291 change_pageblock_range(page, current_order, start_type);
2296 * Boost watermarks to increase reclaim pressure to reduce the
2297 * likelihood of future fallbacks. Wake kswapd now as the node
2298 * may be balanced overall and kswapd will not wake naturally.
2300 boost_watermark(zone);
2301 if (alloc_flags & ALLOC_KSWAPD)
2302 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2304 /* We are not allowed to try stealing from the whole block */
2308 free_pages = move_freepages_block(zone, page, start_type,
2311 * Determine how many pages are compatible with our allocation.
2312 * For movable allocation, it's the number of movable pages which
2313 * we just obtained. For other types it's a bit more tricky.
2315 if (start_type == MIGRATE_MOVABLE) {
2316 alike_pages = movable_pages;
2319 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2320 * to MOVABLE pageblock, consider all non-movable pages as
2321 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2322 * vice versa, be conservative since we can't distinguish the
2323 * exact migratetype of non-movable pages.
2325 if (old_block_type == MIGRATE_MOVABLE)
2326 alike_pages = pageblock_nr_pages
2327 - (free_pages + movable_pages);
2332 /* moving whole block can fail due to zone boundary conditions */
2337 * If a sufficient number of pages in the block are either free or of
2338 * comparable migratability as our allocation, claim the whole block.
2340 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2341 page_group_by_mobility_disabled)
2342 set_pageblock_migratetype(page, start_type);
2347 area = &zone->free_area[current_order];
2348 list_move(&page->lru, &area->free_list[start_type]);
2352 * Check whether there is a suitable fallback freepage with requested order.
2353 * If only_stealable is true, this function returns fallback_mt only if
2354 * we can steal other freepages all together. This would help to reduce
2355 * fragmentation due to mixed migratetype pages in one pageblock.
2357 int find_suitable_fallback(struct free_area *area, unsigned int order,
2358 int migratetype, bool only_stealable, bool *can_steal)
2363 if (area->nr_free == 0)
2368 fallback_mt = fallbacks[migratetype][i];
2369 if (fallback_mt == MIGRATE_TYPES)
2372 if (list_empty(&area->free_list[fallback_mt]))
2375 if (can_steal_fallback(order, migratetype))
2378 if (!only_stealable)
2389 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2390 * there are no empty page blocks that contain a page with a suitable order
2392 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2393 unsigned int alloc_order)
2396 unsigned long max_managed, flags;
2399 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2400 * Check is race-prone but harmless.
2402 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2403 if (zone->nr_reserved_highatomic >= max_managed)
2406 spin_lock_irqsave(&zone->lock, flags);
2408 /* Recheck the nr_reserved_highatomic limit under the lock */
2409 if (zone->nr_reserved_highatomic >= max_managed)
2413 mt = get_pageblock_migratetype(page);
2414 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2415 && !is_migrate_cma(mt)) {
2416 zone->nr_reserved_highatomic += pageblock_nr_pages;
2417 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2418 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2422 spin_unlock_irqrestore(&zone->lock, flags);
2426 * Used when an allocation is about to fail under memory pressure. This
2427 * potentially hurts the reliability of high-order allocations when under
2428 * intense memory pressure but failed atomic allocations should be easier
2429 * to recover from than an OOM.
2431 * If @force is true, try to unreserve a pageblock even though highatomic
2432 * pageblock is exhausted.
2434 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2437 struct zonelist *zonelist = ac->zonelist;
2438 unsigned long flags;
2445 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2448 * Preserve at least one pageblock unless memory pressure
2451 if (!force && zone->nr_reserved_highatomic <=
2455 spin_lock_irqsave(&zone->lock, flags);
2456 for (order = 0; order < MAX_ORDER; order++) {
2457 struct free_area *area = &(zone->free_area[order]);
2459 page = list_first_entry_or_null(
2460 &area->free_list[MIGRATE_HIGHATOMIC],
2466 * In page freeing path, migratetype change is racy so
2467 * we can counter several free pages in a pageblock
2468 * in this loop althoug we changed the pageblock type
2469 * from highatomic to ac->migratetype. So we should
2470 * adjust the count once.
2472 if (is_migrate_highatomic_page(page)) {
2474 * It should never happen but changes to
2475 * locking could inadvertently allow a per-cpu
2476 * drain to add pages to MIGRATE_HIGHATOMIC
2477 * while unreserving so be safe and watch for
2480 zone->nr_reserved_highatomic -= min(
2482 zone->nr_reserved_highatomic);
2486 * Convert to ac->migratetype and avoid the normal
2487 * pageblock stealing heuristics. Minimally, the caller
2488 * is doing the work and needs the pages. More
2489 * importantly, if the block was always converted to
2490 * MIGRATE_UNMOVABLE or another type then the number
2491 * of pageblocks that cannot be completely freed
2494 set_pageblock_migratetype(page, ac->migratetype);
2495 ret = move_freepages_block(zone, page, ac->migratetype,
2498 spin_unlock_irqrestore(&zone->lock, flags);
2502 spin_unlock_irqrestore(&zone->lock, flags);
2509 * Try finding a free buddy page on the fallback list and put it on the free
2510 * list of requested migratetype, possibly along with other pages from the same
2511 * block, depending on fragmentation avoidance heuristics. Returns true if
2512 * fallback was found so that __rmqueue_smallest() can grab it.
2514 * The use of signed ints for order and current_order is a deliberate
2515 * deviation from the rest of this file, to make the for loop
2516 * condition simpler.
2518 static __always_inline bool
2519 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2520 unsigned int alloc_flags)
2522 struct free_area *area;
2524 int min_order = order;
2530 * Do not steal pages from freelists belonging to other pageblocks
2531 * i.e. orders < pageblock_order. If there are no local zones free,
2532 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2534 if (alloc_flags & ALLOC_NOFRAGMENT)
2535 min_order = pageblock_order;
2538 * Find the largest available free page in the other list. This roughly
2539 * approximates finding the pageblock with the most free pages, which
2540 * would be too costly to do exactly.
2542 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2544 area = &(zone->free_area[current_order]);
2545 fallback_mt = find_suitable_fallback(area, current_order,
2546 start_migratetype, false, &can_steal);
2547 if (fallback_mt == -1)
2551 * We cannot steal all free pages from the pageblock and the
2552 * requested migratetype is movable. In that case it's better to
2553 * steal and split the smallest available page instead of the
2554 * largest available page, because even if the next movable
2555 * allocation falls back into a different pageblock than this
2556 * one, it won't cause permanent fragmentation.
2558 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2559 && current_order > order)
2568 for (current_order = order; current_order < MAX_ORDER;
2570 area = &(zone->free_area[current_order]);
2571 fallback_mt = find_suitable_fallback(area, current_order,
2572 start_migratetype, false, &can_steal);
2573 if (fallback_mt != -1)
2578 * This should not happen - we already found a suitable fallback
2579 * when looking for the largest page.
2581 VM_BUG_ON(current_order == MAX_ORDER);
2584 page = list_first_entry(&area->free_list[fallback_mt],
2587 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2590 trace_mm_page_alloc_extfrag(page, order, current_order,
2591 start_migratetype, fallback_mt);
2598 * Do the hard work of removing an element from the buddy allocator.
2599 * Call me with the zone->lock already held.
2601 static __always_inline struct page *
2602 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2603 unsigned int alloc_flags)
2608 page = __rmqueue_smallest(zone, order, migratetype);
2609 if (unlikely(!page)) {
2610 if (migratetype == MIGRATE_MOVABLE)
2611 page = __rmqueue_cma_fallback(zone, order);
2613 if (!page && __rmqueue_fallback(zone, order, migratetype,
2618 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2623 * Obtain a specified number of elements from the buddy allocator, all under
2624 * a single hold of the lock, for efficiency. Add them to the supplied list.
2625 * Returns the number of new pages which were placed at *list.
2627 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2628 unsigned long count, struct list_head *list,
2629 int migratetype, unsigned int alloc_flags)
2633 spin_lock(&zone->lock);
2634 for (i = 0; i < count; ++i) {
2635 struct page *page = __rmqueue(zone, order, migratetype,
2637 if (unlikely(page == NULL))
2640 if (unlikely(check_pcp_refill(page)))
2644 * Split buddy pages returned by expand() are received here in
2645 * physical page order. The page is added to the tail of
2646 * caller's list. From the callers perspective, the linked list
2647 * is ordered by page number under some conditions. This is
2648 * useful for IO devices that can forward direction from the
2649 * head, thus also in the physical page order. This is useful
2650 * for IO devices that can merge IO requests if the physical
2651 * pages are ordered properly.
2653 list_add_tail(&page->lru, list);
2655 if (is_migrate_cma(get_pcppage_migratetype(page)))
2656 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2661 * i pages were removed from the buddy list even if some leak due
2662 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2663 * on i. Do not confuse with 'alloced' which is the number of
2664 * pages added to the pcp list.
2666 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2667 spin_unlock(&zone->lock);
2673 * Called from the vmstat counter updater to drain pagesets of this
2674 * currently executing processor on remote nodes after they have
2677 * Note that this function must be called with the thread pinned to
2678 * a single processor.
2680 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2682 unsigned long flags;
2683 int to_drain, batch;
2685 local_irq_save(flags);
2686 batch = READ_ONCE(pcp->batch);
2687 to_drain = min(pcp->count, batch);
2689 free_pcppages_bulk(zone, to_drain, pcp);
2690 local_irq_restore(flags);
2695 * Drain pcplists of the indicated processor and zone.
2697 * The processor must either be the current processor and the
2698 * thread pinned to the current processor or a processor that
2701 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2703 unsigned long flags;
2704 struct per_cpu_pageset *pset;
2705 struct per_cpu_pages *pcp;
2707 local_irq_save(flags);
2708 pset = per_cpu_ptr(zone->pageset, cpu);
2712 free_pcppages_bulk(zone, pcp->count, pcp);
2713 local_irq_restore(flags);
2717 * Drain pcplists of all zones on the indicated processor.
2719 * The processor must either be the current processor and the
2720 * thread pinned to the current processor or a processor that
2723 static void drain_pages(unsigned int cpu)
2727 for_each_populated_zone(zone) {
2728 drain_pages_zone(cpu, zone);
2733 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2735 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2736 * the single zone's pages.
2738 void drain_local_pages(struct zone *zone)
2740 int cpu = smp_processor_id();
2743 drain_pages_zone(cpu, zone);
2748 static void drain_local_pages_wq(struct work_struct *work)
2750 struct pcpu_drain *drain;
2752 drain = container_of(work, struct pcpu_drain, work);
2755 * drain_all_pages doesn't use proper cpu hotplug protection so
2756 * we can race with cpu offline when the WQ can move this from
2757 * a cpu pinned worker to an unbound one. We can operate on a different
2758 * cpu which is allright but we also have to make sure to not move to
2762 drain_local_pages(drain->zone);
2767 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2769 * When zone parameter is non-NULL, spill just the single zone's pages.
2771 * Note that this can be extremely slow as the draining happens in a workqueue.
2773 void drain_all_pages(struct zone *zone)
2778 * Allocate in the BSS so we wont require allocation in
2779 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2781 static cpumask_t cpus_with_pcps;
2784 * Make sure nobody triggers this path before mm_percpu_wq is fully
2787 if (WARN_ON_ONCE(!mm_percpu_wq))
2791 * Do not drain if one is already in progress unless it's specific to
2792 * a zone. Such callers are primarily CMA and memory hotplug and need
2793 * the drain to be complete when the call returns.
2795 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2798 mutex_lock(&pcpu_drain_mutex);
2802 * We don't care about racing with CPU hotplug event
2803 * as offline notification will cause the notified
2804 * cpu to drain that CPU pcps and on_each_cpu_mask
2805 * disables preemption as part of its processing
2807 for_each_online_cpu(cpu) {
2808 struct per_cpu_pageset *pcp;
2810 bool has_pcps = false;
2813 pcp = per_cpu_ptr(zone->pageset, cpu);
2817 for_each_populated_zone(z) {
2818 pcp = per_cpu_ptr(z->pageset, cpu);
2819 if (pcp->pcp.count) {
2827 cpumask_set_cpu(cpu, &cpus_with_pcps);
2829 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2832 for_each_cpu(cpu, &cpus_with_pcps) {
2833 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2836 INIT_WORK(&drain->work, drain_local_pages_wq);
2837 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2839 for_each_cpu(cpu, &cpus_with_pcps)
2840 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2842 mutex_unlock(&pcpu_drain_mutex);
2845 #ifdef CONFIG_HIBERNATION
2848 * Touch the watchdog for every WD_PAGE_COUNT pages.
2850 #define WD_PAGE_COUNT (128*1024)
2852 void mark_free_pages(struct zone *zone)
2854 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2855 unsigned long flags;
2856 unsigned int order, t;
2859 if (zone_is_empty(zone))
2862 spin_lock_irqsave(&zone->lock, flags);
2864 max_zone_pfn = zone_end_pfn(zone);
2865 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2866 if (pfn_valid(pfn)) {
2867 page = pfn_to_page(pfn);
2869 if (!--page_count) {
2870 touch_nmi_watchdog();
2871 page_count = WD_PAGE_COUNT;
2874 if (page_zone(page) != zone)
2877 if (!swsusp_page_is_forbidden(page))
2878 swsusp_unset_page_free(page);
2881 for_each_migratetype_order(order, t) {
2882 list_for_each_entry(page,
2883 &zone->free_area[order].free_list[t], lru) {
2886 pfn = page_to_pfn(page);
2887 for (i = 0; i < (1UL << order); i++) {
2888 if (!--page_count) {
2889 touch_nmi_watchdog();
2890 page_count = WD_PAGE_COUNT;
2892 swsusp_set_page_free(pfn_to_page(pfn + i));
2896 spin_unlock_irqrestore(&zone->lock, flags);
2898 #endif /* CONFIG_PM */
2900 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2904 if (!free_pcp_prepare(page))
2907 migratetype = get_pfnblock_migratetype(page, pfn);
2908 set_pcppage_migratetype(page, migratetype);
2912 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2914 struct zone *zone = page_zone(page);
2915 struct per_cpu_pages *pcp;
2918 migratetype = get_pcppage_migratetype(page);
2919 __count_vm_event(PGFREE);
2922 * We only track unmovable, reclaimable and movable on pcp lists.
2923 * Free ISOLATE pages back to the allocator because they are being
2924 * offlined but treat HIGHATOMIC as movable pages so we can get those
2925 * areas back if necessary. Otherwise, we may have to free
2926 * excessively into the page allocator
2928 if (migratetype >= MIGRATE_PCPTYPES) {
2929 if (unlikely(is_migrate_isolate(migratetype))) {
2930 free_one_page(zone, page, pfn, 0, migratetype);
2933 migratetype = MIGRATE_MOVABLE;
2936 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2937 list_add(&page->lru, &pcp->lists[migratetype]);
2939 if (pcp->count >= pcp->high) {
2940 unsigned long batch = READ_ONCE(pcp->batch);
2941 free_pcppages_bulk(zone, batch, pcp);
2946 * Free a 0-order page
2948 void free_unref_page(struct page *page)
2950 unsigned long flags;
2951 unsigned long pfn = page_to_pfn(page);
2953 if (!free_unref_page_prepare(page, pfn))
2956 local_irq_save(flags);
2957 free_unref_page_commit(page, pfn);
2958 local_irq_restore(flags);
2962 * Free a list of 0-order pages
2964 void free_unref_page_list(struct list_head *list)
2966 struct page *page, *next;
2967 unsigned long flags, pfn;
2968 int batch_count = 0;
2970 /* Prepare pages for freeing */
2971 list_for_each_entry_safe(page, next, list, lru) {
2972 pfn = page_to_pfn(page);
2973 if (!free_unref_page_prepare(page, pfn))
2974 list_del(&page->lru);
2975 set_page_private(page, pfn);
2978 local_irq_save(flags);
2979 list_for_each_entry_safe(page, next, list, lru) {
2980 unsigned long pfn = page_private(page);
2982 set_page_private(page, 0);
2983 trace_mm_page_free_batched(page);
2984 free_unref_page_commit(page, pfn);
2987 * Guard against excessive IRQ disabled times when we get
2988 * a large list of pages to free.
2990 if (++batch_count == SWAP_CLUSTER_MAX) {
2991 local_irq_restore(flags);
2993 local_irq_save(flags);
2996 local_irq_restore(flags);
3000 * split_page takes a non-compound higher-order page, and splits it into
3001 * n (1<<order) sub-pages: page[0..n]
3002 * Each sub-page must be freed individually.
3004 * Note: this is probably too low level an operation for use in drivers.
3005 * Please consult with lkml before using this in your driver.
3007 void split_page(struct page *page, unsigned int order)
3011 VM_BUG_ON_PAGE(PageCompound(page), page);
3012 VM_BUG_ON_PAGE(!page_count(page), page);
3014 for (i = 1; i < (1 << order); i++)
3015 set_page_refcounted(page + i);
3016 split_page_owner(page, order);
3018 EXPORT_SYMBOL_GPL(split_page);
3020 int __isolate_free_page(struct page *page, unsigned int order)
3022 unsigned long watermark;
3026 BUG_ON(!PageBuddy(page));
3028 zone = page_zone(page);
3029 mt = get_pageblock_migratetype(page);
3031 if (!is_migrate_isolate(mt)) {
3033 * Obey watermarks as if the page was being allocated. We can
3034 * emulate a high-order watermark check with a raised order-0
3035 * watermark, because we already know our high-order page
3038 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3039 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3042 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3045 /* Remove page from free list */
3046 list_del(&page->lru);
3047 zone->free_area[order].nr_free--;
3048 rmv_page_order(page);
3051 * Set the pageblock if the isolated page is at least half of a
3054 if (order >= pageblock_order - 1) {
3055 struct page *endpage = page + (1 << order) - 1;
3056 for (; page < endpage; page += pageblock_nr_pages) {
3057 int mt = get_pageblock_migratetype(page);
3058 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3059 && !is_migrate_highatomic(mt))
3060 set_pageblock_migratetype(page,
3066 return 1UL << order;
3070 * Update NUMA hit/miss statistics
3072 * Must be called with interrupts disabled.
3074 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3077 enum numa_stat_item local_stat = NUMA_LOCAL;
3079 /* skip numa counters update if numa stats is disabled */
3080 if (!static_branch_likely(&vm_numa_stat_key))
3083 if (zone_to_nid(z) != numa_node_id())
3084 local_stat = NUMA_OTHER;
3086 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3087 __inc_numa_state(z, NUMA_HIT);
3089 __inc_numa_state(z, NUMA_MISS);
3090 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3092 __inc_numa_state(z, local_stat);
3096 /* Remove page from the per-cpu list, caller must protect the list */
3097 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3098 unsigned int alloc_flags,
3099 struct per_cpu_pages *pcp,
3100 struct list_head *list)
3105 if (list_empty(list)) {
3106 pcp->count += rmqueue_bulk(zone, 0,
3108 migratetype, alloc_flags);
3109 if (unlikely(list_empty(list)))
3113 page = list_first_entry(list, struct page, lru);
3114 list_del(&page->lru);
3116 } while (check_new_pcp(page));
3121 /* Lock and remove page from the per-cpu list */
3122 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3123 struct zone *zone, unsigned int order,
3124 gfp_t gfp_flags, int migratetype,
3125 unsigned int alloc_flags)
3127 struct per_cpu_pages *pcp;
3128 struct list_head *list;
3130 unsigned long flags;
3132 local_irq_save(flags);
3133 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3134 list = &pcp->lists[migratetype];
3135 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3137 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3138 zone_statistics(preferred_zone, zone);
3140 local_irq_restore(flags);
3145 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3148 struct page *rmqueue(struct zone *preferred_zone,
3149 struct zone *zone, unsigned int order,
3150 gfp_t gfp_flags, unsigned int alloc_flags,
3153 unsigned long flags;
3156 if (likely(order == 0)) {
3157 page = rmqueue_pcplist(preferred_zone, zone, order,
3158 gfp_flags, migratetype, alloc_flags);
3163 * We most definitely don't want callers attempting to
3164 * allocate greater than order-1 page units with __GFP_NOFAIL.
3166 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3167 spin_lock_irqsave(&zone->lock, flags);
3171 if (alloc_flags & ALLOC_HARDER) {
3172 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3174 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3177 page = __rmqueue(zone, order, migratetype, alloc_flags);
3178 } while (page && check_new_pages(page, order));
3179 spin_unlock(&zone->lock);
3182 __mod_zone_freepage_state(zone, -(1 << order),
3183 get_pcppage_migratetype(page));
3185 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3186 zone_statistics(preferred_zone, zone);
3187 local_irq_restore(flags);
3190 /* Separate test+clear to avoid unnecessary atomics */
3191 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3192 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3193 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3196 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3200 local_irq_restore(flags);
3204 #ifdef CONFIG_FAIL_PAGE_ALLOC
3207 struct fault_attr attr;
3209 bool ignore_gfp_highmem;
3210 bool ignore_gfp_reclaim;
3212 } fail_page_alloc = {
3213 .attr = FAULT_ATTR_INITIALIZER,
3214 .ignore_gfp_reclaim = true,
3215 .ignore_gfp_highmem = true,
3219 static int __init setup_fail_page_alloc(char *str)
3221 return setup_fault_attr(&fail_page_alloc.attr, str);
3223 __setup("fail_page_alloc=", setup_fail_page_alloc);
3225 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3227 if (order < fail_page_alloc.min_order)
3229 if (gfp_mask & __GFP_NOFAIL)
3231 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3233 if (fail_page_alloc.ignore_gfp_reclaim &&
3234 (gfp_mask & __GFP_DIRECT_RECLAIM))
3237 return should_fail(&fail_page_alloc.attr, 1 << order);
3240 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3242 static int __init fail_page_alloc_debugfs(void)
3244 umode_t mode = S_IFREG | 0600;
3247 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3248 &fail_page_alloc.attr);
3250 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3251 &fail_page_alloc.ignore_gfp_reclaim);
3252 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3253 &fail_page_alloc.ignore_gfp_highmem);
3254 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3259 late_initcall(fail_page_alloc_debugfs);
3261 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3263 #else /* CONFIG_FAIL_PAGE_ALLOC */
3265 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3270 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3272 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3274 return __should_fail_alloc_page(gfp_mask, order);
3276 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3279 * Return true if free base pages are above 'mark'. For high-order checks it
3280 * will return true of the order-0 watermark is reached and there is at least
3281 * one free page of a suitable size. Checking now avoids taking the zone lock
3282 * to check in the allocation paths if no pages are free.
3284 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3285 int classzone_idx, unsigned int alloc_flags,
3290 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3292 /* free_pages may go negative - that's OK */
3293 free_pages -= (1 << order) - 1;
3295 if (alloc_flags & ALLOC_HIGH)
3299 * If the caller does not have rights to ALLOC_HARDER then subtract
3300 * the high-atomic reserves. This will over-estimate the size of the
3301 * atomic reserve but it avoids a search.
3303 if (likely(!alloc_harder)) {
3304 free_pages -= z->nr_reserved_highatomic;
3307 * OOM victims can try even harder than normal ALLOC_HARDER
3308 * users on the grounds that it's definitely going to be in
3309 * the exit path shortly and free memory. Any allocation it
3310 * makes during the free path will be small and short-lived.
3312 if (alloc_flags & ALLOC_OOM)
3320 /* If allocation can't use CMA areas don't use free CMA pages */
3321 if (!(alloc_flags & ALLOC_CMA))
3322 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3326 * Check watermarks for an order-0 allocation request. If these
3327 * are not met, then a high-order request also cannot go ahead
3328 * even if a suitable page happened to be free.
3330 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3333 /* If this is an order-0 request then the watermark is fine */
3337 /* For a high-order request, check at least one suitable page is free */
3338 for (o = order; o < MAX_ORDER; o++) {
3339 struct free_area *area = &z->free_area[o];
3345 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3346 if (!list_empty(&area->free_list[mt]))
3351 if ((alloc_flags & ALLOC_CMA) &&
3352 !list_empty(&area->free_list[MIGRATE_CMA])) {
3357 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3363 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3364 int classzone_idx, unsigned int alloc_flags)
3366 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3367 zone_page_state(z, NR_FREE_PAGES));
3370 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3371 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3373 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3377 /* If allocation can't use CMA areas don't use free CMA pages */
3378 if (!(alloc_flags & ALLOC_CMA))
3379 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3383 * Fast check for order-0 only. If this fails then the reserves
3384 * need to be calculated. There is a corner case where the check
3385 * passes but only the high-order atomic reserve are free. If
3386 * the caller is !atomic then it'll uselessly search the free
3387 * list. That corner case is then slower but it is harmless.
3389 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3392 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3396 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3397 unsigned long mark, int classzone_idx)
3399 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3401 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3402 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3404 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3409 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3411 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3414 #else /* CONFIG_NUMA */
3415 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3419 #endif /* CONFIG_NUMA */
3422 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3423 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3424 * premature use of a lower zone may cause lowmem pressure problems that
3425 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3426 * probably too small. It only makes sense to spread allocations to avoid
3427 * fragmentation between the Normal and DMA32 zones.
3429 static inline unsigned int
3430 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3432 unsigned int alloc_flags = 0;
3434 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3435 alloc_flags |= ALLOC_KSWAPD;
3437 #ifdef CONFIG_ZONE_DMA32
3441 if (zone_idx(zone) != ZONE_NORMAL)
3445 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3446 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3447 * on UMA that if Normal is populated then so is DMA32.
3449 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3450 if (nr_online_nodes > 1 && !populated_zone(--zone))
3453 alloc_flags |= ALLOC_NOFRAGMENT;
3454 #endif /* CONFIG_ZONE_DMA32 */
3459 * get_page_from_freelist goes through the zonelist trying to allocate
3462 static struct page *
3463 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3464 const struct alloc_context *ac)
3468 struct pglist_data *last_pgdat_dirty_limit = NULL;
3473 * Scan zonelist, looking for a zone with enough free.
3474 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3476 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3477 z = ac->preferred_zoneref;
3478 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3483 if (cpusets_enabled() &&
3484 (alloc_flags & ALLOC_CPUSET) &&
3485 !__cpuset_zone_allowed(zone, gfp_mask))
3488 * When allocating a page cache page for writing, we
3489 * want to get it from a node that is within its dirty
3490 * limit, such that no single node holds more than its
3491 * proportional share of globally allowed dirty pages.
3492 * The dirty limits take into account the node's
3493 * lowmem reserves and high watermark so that kswapd
3494 * should be able to balance it without having to
3495 * write pages from its LRU list.
3497 * XXX: For now, allow allocations to potentially
3498 * exceed the per-node dirty limit in the slowpath
3499 * (spread_dirty_pages unset) before going into reclaim,
3500 * which is important when on a NUMA setup the allowed
3501 * nodes are together not big enough to reach the
3502 * global limit. The proper fix for these situations
3503 * will require awareness of nodes in the
3504 * dirty-throttling and the flusher threads.
3506 if (ac->spread_dirty_pages) {
3507 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3510 if (!node_dirty_ok(zone->zone_pgdat)) {
3511 last_pgdat_dirty_limit = zone->zone_pgdat;
3516 if (no_fallback && nr_online_nodes > 1 &&
3517 zone != ac->preferred_zoneref->zone) {
3521 * If moving to a remote node, retry but allow
3522 * fragmenting fallbacks. Locality is more important
3523 * than fragmentation avoidance.
3525 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3526 if (zone_to_nid(zone) != local_nid) {
3527 alloc_flags &= ~ALLOC_NOFRAGMENT;
3532 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3533 if (!zone_watermark_fast(zone, order, mark,
3534 ac_classzone_idx(ac), alloc_flags)) {
3537 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3539 * Watermark failed for this zone, but see if we can
3540 * grow this zone if it contains deferred pages.
3542 if (static_branch_unlikely(&deferred_pages)) {
3543 if (_deferred_grow_zone(zone, order))
3547 /* Checked here to keep the fast path fast */
3548 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3549 if (alloc_flags & ALLOC_NO_WATERMARKS)
3552 if (node_reclaim_mode == 0 ||
3553 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3556 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3558 case NODE_RECLAIM_NOSCAN:
3561 case NODE_RECLAIM_FULL:
3562 /* scanned but unreclaimable */
3565 /* did we reclaim enough */
3566 if (zone_watermark_ok(zone, order, mark,
3567 ac_classzone_idx(ac), alloc_flags))
3575 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3576 gfp_mask, alloc_flags, ac->migratetype);
3578 prep_new_page(page, order, gfp_mask, alloc_flags);
3581 * If this is a high-order atomic allocation then check
3582 * if the pageblock should be reserved for the future
3584 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3585 reserve_highatomic_pageblock(page, zone, order);
3589 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3590 /* Try again if zone has deferred pages */
3591 if (static_branch_unlikely(&deferred_pages)) {
3592 if (_deferred_grow_zone(zone, order))
3600 * It's possible on a UMA machine to get through all zones that are
3601 * fragmented. If avoiding fragmentation, reset and try again.
3604 alloc_flags &= ~ALLOC_NOFRAGMENT;
3611 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3613 unsigned int filter = SHOW_MEM_FILTER_NODES;
3614 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3616 if (!__ratelimit(&show_mem_rs))
3620 * This documents exceptions given to allocations in certain
3621 * contexts that are allowed to allocate outside current's set
3624 if (!(gfp_mask & __GFP_NOMEMALLOC))
3625 if (tsk_is_oom_victim(current) ||
3626 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3627 filter &= ~SHOW_MEM_FILTER_NODES;
3628 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3629 filter &= ~SHOW_MEM_FILTER_NODES;
3631 show_mem(filter, nodemask);
3634 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3636 struct va_format vaf;
3638 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3639 DEFAULT_RATELIMIT_BURST);
3641 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3644 va_start(args, fmt);
3647 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3648 current->comm, &vaf, gfp_mask, &gfp_mask,
3649 nodemask_pr_args(nodemask));
3652 cpuset_print_current_mems_allowed();
3655 warn_alloc_show_mem(gfp_mask, nodemask);
3658 static inline struct page *
3659 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3660 unsigned int alloc_flags,
3661 const struct alloc_context *ac)
3665 page = get_page_from_freelist(gfp_mask, order,
3666 alloc_flags|ALLOC_CPUSET, ac);
3668 * fallback to ignore cpuset restriction if our nodes
3672 page = get_page_from_freelist(gfp_mask, order,
3678 static inline struct page *
3679 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3680 const struct alloc_context *ac, unsigned long *did_some_progress)
3682 struct oom_control oc = {
3683 .zonelist = ac->zonelist,
3684 .nodemask = ac->nodemask,
3686 .gfp_mask = gfp_mask,
3691 *did_some_progress = 0;
3694 * Acquire the oom lock. If that fails, somebody else is
3695 * making progress for us.
3697 if (!mutex_trylock(&oom_lock)) {
3698 *did_some_progress = 1;
3699 schedule_timeout_uninterruptible(1);
3704 * Go through the zonelist yet one more time, keep very high watermark
3705 * here, this is only to catch a parallel oom killing, we must fail if
3706 * we're still under heavy pressure. But make sure that this reclaim
3707 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3708 * allocation which will never fail due to oom_lock already held.
3710 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3711 ~__GFP_DIRECT_RECLAIM, order,
3712 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3716 /* Coredumps can quickly deplete all memory reserves */
3717 if (current->flags & PF_DUMPCORE)
3719 /* The OOM killer will not help higher order allocs */
3720 if (order > PAGE_ALLOC_COSTLY_ORDER)
3723 * We have already exhausted all our reclaim opportunities without any
3724 * success so it is time to admit defeat. We will skip the OOM killer
3725 * because it is very likely that the caller has a more reasonable
3726 * fallback than shooting a random task.
3728 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3730 /* The OOM killer does not needlessly kill tasks for lowmem */
3731 if (ac->high_zoneidx < ZONE_NORMAL)
3733 if (pm_suspended_storage())
3736 * XXX: GFP_NOFS allocations should rather fail than rely on
3737 * other request to make a forward progress.
3738 * We are in an unfortunate situation where out_of_memory cannot
3739 * do much for this context but let's try it to at least get
3740 * access to memory reserved if the current task is killed (see
3741 * out_of_memory). Once filesystems are ready to handle allocation
3742 * failures more gracefully we should just bail out here.
3745 /* The OOM killer may not free memory on a specific node */
3746 if (gfp_mask & __GFP_THISNODE)
3749 /* Exhausted what can be done so it's blame time */
3750 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3751 *did_some_progress = 1;
3754 * Help non-failing allocations by giving them access to memory
3757 if (gfp_mask & __GFP_NOFAIL)
3758 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3759 ALLOC_NO_WATERMARKS, ac);
3762 mutex_unlock(&oom_lock);
3767 * Maximum number of compaction retries wit a progress before OOM
3768 * killer is consider as the only way to move forward.
3770 #define MAX_COMPACT_RETRIES 16
3772 #ifdef CONFIG_COMPACTION
3773 /* Try memory compaction for high-order allocations before reclaim */
3774 static 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 struct page *page = NULL;
3780 unsigned long pflags;
3781 unsigned int noreclaim_flag;
3786 psi_memstall_enter(&pflags);
3787 noreclaim_flag = memalloc_noreclaim_save();
3789 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3792 memalloc_noreclaim_restore(noreclaim_flag);
3793 psi_memstall_leave(&pflags);
3796 * At least in one zone compaction wasn't deferred or skipped, so let's
3797 * count a compaction stall
3799 count_vm_event(COMPACTSTALL);
3801 /* Prep a captured page if available */
3803 prep_new_page(page, order, gfp_mask, alloc_flags);
3805 /* Try get a page from the freelist if available */
3807 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3810 struct zone *zone = page_zone(page);
3812 zone->compact_blockskip_flush = false;
3813 compaction_defer_reset(zone, order, true);
3814 count_vm_event(COMPACTSUCCESS);
3819 * It's bad if compaction run occurs and fails. The most likely reason
3820 * is that pages exist, but not enough to satisfy watermarks.
3822 count_vm_event(COMPACTFAIL);
3830 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3831 enum compact_result compact_result,
3832 enum compact_priority *compact_priority,
3833 int *compaction_retries)
3835 int max_retries = MAX_COMPACT_RETRIES;
3838 int retries = *compaction_retries;
3839 enum compact_priority priority = *compact_priority;
3844 if (compaction_made_progress(compact_result))
3845 (*compaction_retries)++;
3848 * compaction considers all the zone as desperately out of memory
3849 * so it doesn't really make much sense to retry except when the
3850 * failure could be caused by insufficient priority
3852 if (compaction_failed(compact_result))
3853 goto check_priority;
3856 * make sure the compaction wasn't deferred or didn't bail out early
3857 * due to locks contention before we declare that we should give up.
3858 * But do not retry if the given zonelist is not suitable for
3861 if (compaction_withdrawn(compact_result)) {
3862 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3867 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3868 * costly ones because they are de facto nofail and invoke OOM
3869 * killer to move on while costly can fail and users are ready
3870 * to cope with that. 1/4 retries is rather arbitrary but we
3871 * would need much more detailed feedback from compaction to
3872 * make a better decision.
3874 if (order > PAGE_ALLOC_COSTLY_ORDER)
3876 if (*compaction_retries <= max_retries) {
3882 * Make sure there are attempts at the highest priority if we exhausted
3883 * all retries or failed at the lower priorities.
3886 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3887 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3889 if (*compact_priority > min_priority) {
3890 (*compact_priority)--;
3891 *compaction_retries = 0;
3895 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3899 static inline struct page *
3900 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3901 unsigned int alloc_flags, const struct alloc_context *ac,
3902 enum compact_priority prio, enum compact_result *compact_result)
3904 *compact_result = COMPACT_SKIPPED;
3909 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3910 enum compact_result compact_result,
3911 enum compact_priority *compact_priority,
3912 int *compaction_retries)
3917 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3921 * There are setups with compaction disabled which would prefer to loop
3922 * inside the allocator rather than hit the oom killer prematurely.
3923 * Let's give them a good hope and keep retrying while the order-0
3924 * watermarks are OK.
3926 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3928 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3929 ac_classzone_idx(ac), alloc_flags))
3934 #endif /* CONFIG_COMPACTION */
3936 #ifdef CONFIG_LOCKDEP
3937 static struct lockdep_map __fs_reclaim_map =
3938 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3940 static bool __need_fs_reclaim(gfp_t gfp_mask)
3942 gfp_mask = current_gfp_context(gfp_mask);
3944 /* no reclaim without waiting on it */
3945 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3948 /* this guy won't enter reclaim */
3949 if (current->flags & PF_MEMALLOC)
3952 /* We're only interested __GFP_FS allocations for now */
3953 if (!(gfp_mask & __GFP_FS))
3956 if (gfp_mask & __GFP_NOLOCKDEP)
3962 void __fs_reclaim_acquire(void)
3964 lock_map_acquire(&__fs_reclaim_map);
3967 void __fs_reclaim_release(void)
3969 lock_map_release(&__fs_reclaim_map);
3972 void fs_reclaim_acquire(gfp_t gfp_mask)
3974 if (__need_fs_reclaim(gfp_mask))
3975 __fs_reclaim_acquire();
3977 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3979 void fs_reclaim_release(gfp_t gfp_mask)
3981 if (__need_fs_reclaim(gfp_mask))
3982 __fs_reclaim_release();
3984 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3987 /* Perform direct synchronous page reclaim */
3989 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3990 const struct alloc_context *ac)
3992 struct reclaim_state reclaim_state;
3994 unsigned int noreclaim_flag;
3995 unsigned long pflags;
3999 /* We now go into synchronous reclaim */
4000 cpuset_memory_pressure_bump();
4001 psi_memstall_enter(&pflags);
4002 fs_reclaim_acquire(gfp_mask);
4003 noreclaim_flag = memalloc_noreclaim_save();
4004 reclaim_state.reclaimed_slab = 0;
4005 current->reclaim_state = &reclaim_state;
4007 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4010 current->reclaim_state = NULL;
4011 memalloc_noreclaim_restore(noreclaim_flag);
4012 fs_reclaim_release(gfp_mask);
4013 psi_memstall_leave(&pflags);
4020 /* The really slow allocator path where we enter direct reclaim */
4021 static inline struct page *
4022 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4023 unsigned int alloc_flags, const struct alloc_context *ac,
4024 unsigned long *did_some_progress)
4026 struct page *page = NULL;
4027 bool drained = false;
4029 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4030 if (unlikely(!(*did_some_progress)))
4034 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4037 * If an allocation failed after direct reclaim, it could be because
4038 * pages are pinned on the per-cpu lists or in high alloc reserves.
4039 * Shrink them them and try again
4041 if (!page && !drained) {
4042 unreserve_highatomic_pageblock(ac, false);
4043 drain_all_pages(NULL);
4051 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4052 const struct alloc_context *ac)
4056 pg_data_t *last_pgdat = NULL;
4057 enum zone_type high_zoneidx = ac->high_zoneidx;
4059 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4061 if (last_pgdat != zone->zone_pgdat)
4062 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4063 last_pgdat = zone->zone_pgdat;
4067 static inline unsigned int
4068 gfp_to_alloc_flags(gfp_t gfp_mask)
4070 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4072 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4073 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4076 * The caller may dip into page reserves a bit more if the caller
4077 * cannot run direct reclaim, or if the caller has realtime scheduling
4078 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4079 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4081 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4083 if (gfp_mask & __GFP_ATOMIC) {
4085 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4086 * if it can't schedule.
4088 if (!(gfp_mask & __GFP_NOMEMALLOC))
4089 alloc_flags |= ALLOC_HARDER;
4091 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4092 * comment for __cpuset_node_allowed().
4094 alloc_flags &= ~ALLOC_CPUSET;
4095 } else if (unlikely(rt_task(current)) && !in_interrupt())
4096 alloc_flags |= ALLOC_HARDER;
4098 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4099 alloc_flags |= ALLOC_KSWAPD;
4102 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4103 alloc_flags |= ALLOC_CMA;
4108 static bool oom_reserves_allowed(struct task_struct *tsk)
4110 if (!tsk_is_oom_victim(tsk))
4114 * !MMU doesn't have oom reaper so give access to memory reserves
4115 * only to the thread with TIF_MEMDIE set
4117 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4124 * Distinguish requests which really need access to full memory
4125 * reserves from oom victims which can live with a portion of it
4127 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4129 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4131 if (gfp_mask & __GFP_MEMALLOC)
4132 return ALLOC_NO_WATERMARKS;
4133 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4134 return ALLOC_NO_WATERMARKS;
4135 if (!in_interrupt()) {
4136 if (current->flags & PF_MEMALLOC)
4137 return ALLOC_NO_WATERMARKS;
4138 else if (oom_reserves_allowed(current))
4145 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4147 return !!__gfp_pfmemalloc_flags(gfp_mask);
4151 * Checks whether it makes sense to retry the reclaim to make a forward progress
4152 * for the given allocation request.
4154 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4155 * without success, or when we couldn't even meet the watermark if we
4156 * reclaimed all remaining pages on the LRU lists.
4158 * Returns true if a retry is viable or false to enter the oom path.
4161 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4162 struct alloc_context *ac, int alloc_flags,
4163 bool did_some_progress, int *no_progress_loops)
4170 * Costly allocations might have made a progress but this doesn't mean
4171 * their order will become available due to high fragmentation so
4172 * always increment the no progress counter for them
4174 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4175 *no_progress_loops = 0;
4177 (*no_progress_loops)++;
4180 * Make sure we converge to OOM if we cannot make any progress
4181 * several times in the row.
4183 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4184 /* Before OOM, exhaust highatomic_reserve */
4185 return unreserve_highatomic_pageblock(ac, true);
4189 * Keep reclaiming pages while there is a chance this will lead
4190 * somewhere. If none of the target zones can satisfy our allocation
4191 * request even if all reclaimable pages are considered then we are
4192 * screwed and have to go OOM.
4194 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4196 unsigned long available;
4197 unsigned long reclaimable;
4198 unsigned long min_wmark = min_wmark_pages(zone);
4201 available = reclaimable = zone_reclaimable_pages(zone);
4202 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4205 * Would the allocation succeed if we reclaimed all
4206 * reclaimable pages?
4208 wmark = __zone_watermark_ok(zone, order, min_wmark,
4209 ac_classzone_idx(ac), alloc_flags, available);
4210 trace_reclaim_retry_zone(z, order, reclaimable,
4211 available, min_wmark, *no_progress_loops, wmark);
4214 * If we didn't make any progress and have a lot of
4215 * dirty + writeback pages then we should wait for
4216 * an IO to complete to slow down the reclaim and
4217 * prevent from pre mature OOM
4219 if (!did_some_progress) {
4220 unsigned long write_pending;
4222 write_pending = zone_page_state_snapshot(zone,
4223 NR_ZONE_WRITE_PENDING);
4225 if (2 * write_pending > reclaimable) {
4226 congestion_wait(BLK_RW_ASYNC, HZ/10);
4238 * Memory allocation/reclaim might be called from a WQ context and the
4239 * current implementation of the WQ concurrency control doesn't
4240 * recognize that a particular WQ is congested if the worker thread is
4241 * looping without ever sleeping. Therefore we have to do a short sleep
4242 * here rather than calling cond_resched().
4244 if (current->flags & PF_WQ_WORKER)
4245 schedule_timeout_uninterruptible(1);
4252 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4255 * It's possible that cpuset's mems_allowed and the nodemask from
4256 * mempolicy don't intersect. This should be normally dealt with by
4257 * policy_nodemask(), but it's possible to race with cpuset update in
4258 * such a way the check therein was true, and then it became false
4259 * before we got our cpuset_mems_cookie here.
4260 * This assumes that for all allocations, ac->nodemask can come only
4261 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4262 * when it does not intersect with the cpuset restrictions) or the
4263 * caller can deal with a violated nodemask.
4265 if (cpusets_enabled() && ac->nodemask &&
4266 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4267 ac->nodemask = NULL;
4272 * When updating a task's mems_allowed or mempolicy nodemask, it is
4273 * possible to race with parallel threads in such a way that our
4274 * allocation can fail while the mask is being updated. If we are about
4275 * to fail, check if the cpuset changed during allocation and if so,
4278 if (read_mems_allowed_retry(cpuset_mems_cookie))
4284 static inline struct page *
4285 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4286 struct alloc_context *ac)
4288 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4289 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4290 struct page *page = NULL;
4291 unsigned int alloc_flags;
4292 unsigned long did_some_progress;
4293 enum compact_priority compact_priority;
4294 enum compact_result compact_result;
4295 int compaction_retries;
4296 int no_progress_loops;
4297 unsigned int cpuset_mems_cookie;
4301 * We also sanity check to catch abuse of atomic reserves being used by
4302 * callers that are not in atomic context.
4304 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4305 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4306 gfp_mask &= ~__GFP_ATOMIC;
4309 compaction_retries = 0;
4310 no_progress_loops = 0;
4311 compact_priority = DEF_COMPACT_PRIORITY;
4312 cpuset_mems_cookie = read_mems_allowed_begin();
4315 * The fast path uses conservative alloc_flags to succeed only until
4316 * kswapd needs to be woken up, and to avoid the cost of setting up
4317 * alloc_flags precisely. So we do that now.
4319 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4322 * We need to recalculate the starting point for the zonelist iterator
4323 * because we might have used different nodemask in the fast path, or
4324 * there was a cpuset modification and we are retrying - otherwise we
4325 * could end up iterating over non-eligible zones endlessly.
4327 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4328 ac->high_zoneidx, ac->nodemask);
4329 if (!ac->preferred_zoneref->zone)
4332 if (alloc_flags & ALLOC_KSWAPD)
4333 wake_all_kswapds(order, gfp_mask, ac);
4336 * The adjusted alloc_flags might result in immediate success, so try
4339 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4344 * For costly allocations, try direct compaction first, as it's likely
4345 * that we have enough base pages and don't need to reclaim. For non-
4346 * movable high-order allocations, do that as well, as compaction will
4347 * try prevent permanent fragmentation by migrating from blocks of the
4349 * Don't try this for allocations that are allowed to ignore
4350 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4352 if (can_direct_reclaim &&
4354 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4355 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4356 page = __alloc_pages_direct_compact(gfp_mask, order,
4358 INIT_COMPACT_PRIORITY,
4364 * Checks for costly allocations with __GFP_NORETRY, which
4365 * includes THP page fault allocations
4367 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4369 * If compaction is deferred for high-order allocations,
4370 * it is because sync compaction recently failed. If
4371 * this is the case and the caller requested a THP
4372 * allocation, we do not want to heavily disrupt the
4373 * system, so we fail the allocation instead of entering
4376 if (compact_result == COMPACT_DEFERRED)
4380 * Looks like reclaim/compaction is worth trying, but
4381 * sync compaction could be very expensive, so keep
4382 * using async compaction.
4384 compact_priority = INIT_COMPACT_PRIORITY;
4389 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4390 if (alloc_flags & ALLOC_KSWAPD)
4391 wake_all_kswapds(order, gfp_mask, ac);
4393 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4395 alloc_flags = reserve_flags;
4398 * Reset the nodemask and zonelist iterators if memory policies can be
4399 * ignored. These allocations are high priority and system rather than
4402 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4403 ac->nodemask = NULL;
4404 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4405 ac->high_zoneidx, ac->nodemask);
4408 /* Attempt with potentially adjusted zonelist and alloc_flags */
4409 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4413 /* Caller is not willing to reclaim, we can't balance anything */
4414 if (!can_direct_reclaim)
4417 /* Avoid recursion of direct reclaim */
4418 if (current->flags & PF_MEMALLOC)
4421 /* Try direct reclaim and then allocating */
4422 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4423 &did_some_progress);
4427 /* Try direct compaction and then allocating */
4428 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4429 compact_priority, &compact_result);
4433 /* Do not loop if specifically requested */
4434 if (gfp_mask & __GFP_NORETRY)
4438 * Do not retry costly high order allocations unless they are
4439 * __GFP_RETRY_MAYFAIL
4441 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4444 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4445 did_some_progress > 0, &no_progress_loops))
4449 * It doesn't make any sense to retry for the compaction if the order-0
4450 * reclaim is not able to make any progress because the current
4451 * implementation of the compaction depends on the sufficient amount
4452 * of free memory (see __compaction_suitable)
4454 if (did_some_progress > 0 &&
4455 should_compact_retry(ac, order, alloc_flags,
4456 compact_result, &compact_priority,
4457 &compaction_retries))
4461 /* Deal with possible cpuset update races before we start OOM killing */
4462 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4465 /* Reclaim has failed us, start killing things */
4466 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4470 /* Avoid allocations with no watermarks from looping endlessly */
4471 if (tsk_is_oom_victim(current) &&
4472 (alloc_flags == ALLOC_OOM ||
4473 (gfp_mask & __GFP_NOMEMALLOC)))
4476 /* Retry as long as the OOM killer is making progress */
4477 if (did_some_progress) {
4478 no_progress_loops = 0;
4483 /* Deal with possible cpuset update races before we fail */
4484 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4488 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4491 if (gfp_mask & __GFP_NOFAIL) {
4493 * All existing users of the __GFP_NOFAIL are blockable, so warn
4494 * of any new users that actually require GFP_NOWAIT
4496 if (WARN_ON_ONCE(!can_direct_reclaim))
4500 * PF_MEMALLOC request from this context is rather bizarre
4501 * because we cannot reclaim anything and only can loop waiting
4502 * for somebody to do a work for us
4504 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4507 * non failing costly orders are a hard requirement which we
4508 * are not prepared for much so let's warn about these users
4509 * so that we can identify them and convert them to something
4512 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4515 * Help non-failing allocations by giving them access to memory
4516 * reserves but do not use ALLOC_NO_WATERMARKS because this
4517 * could deplete whole memory reserves which would just make
4518 * the situation worse
4520 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4528 warn_alloc(gfp_mask, ac->nodemask,
4529 "page allocation failure: order:%u", order);
4534 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4535 int preferred_nid, nodemask_t *nodemask,
4536 struct alloc_context *ac, gfp_t *alloc_mask,
4537 unsigned int *alloc_flags)
4539 ac->high_zoneidx = gfp_zone(gfp_mask);
4540 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4541 ac->nodemask = nodemask;
4542 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4544 if (cpusets_enabled()) {
4545 *alloc_mask |= __GFP_HARDWALL;
4547 ac->nodemask = &cpuset_current_mems_allowed;
4549 *alloc_flags |= ALLOC_CPUSET;
4552 fs_reclaim_acquire(gfp_mask);
4553 fs_reclaim_release(gfp_mask);
4555 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4557 if (should_fail_alloc_page(gfp_mask, order))
4560 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4561 *alloc_flags |= ALLOC_CMA;
4566 /* Determine whether to spread dirty pages and what the first usable zone */
4567 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4569 /* Dirty zone balancing only done in the fast path */
4570 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4573 * The preferred zone is used for statistics but crucially it is
4574 * also used as the starting point for the zonelist iterator. It
4575 * may get reset for allocations that ignore memory policies.
4577 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4578 ac->high_zoneidx, ac->nodemask);
4582 * This is the 'heart' of the zoned buddy allocator.
4585 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4586 nodemask_t *nodemask)
4589 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4590 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4591 struct alloc_context ac = { };
4594 * There are several places where we assume that the order value is sane
4595 * so bail out early if the request is out of bound.
4597 if (unlikely(order >= MAX_ORDER)) {
4598 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4602 gfp_mask &= gfp_allowed_mask;
4603 alloc_mask = gfp_mask;
4604 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4607 finalise_ac(gfp_mask, &ac);
4610 * Forbid the first pass from falling back to types that fragment
4611 * memory until all local zones are considered.
4613 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4615 /* First allocation attempt */
4616 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4621 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4622 * resp. GFP_NOIO which has to be inherited for all allocation requests
4623 * from a particular context which has been marked by
4624 * memalloc_no{fs,io}_{save,restore}.
4626 alloc_mask = current_gfp_context(gfp_mask);
4627 ac.spread_dirty_pages = false;
4630 * Restore the original nodemask if it was potentially replaced with
4631 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4633 if (unlikely(ac.nodemask != nodemask))
4634 ac.nodemask = nodemask;
4636 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4639 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4640 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4641 __free_pages(page, order);
4645 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4649 EXPORT_SYMBOL(__alloc_pages_nodemask);
4652 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4653 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4654 * you need to access high mem.
4656 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4660 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4663 return (unsigned long) page_address(page);
4665 EXPORT_SYMBOL(__get_free_pages);
4667 unsigned long get_zeroed_page(gfp_t gfp_mask)
4669 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4671 EXPORT_SYMBOL(get_zeroed_page);
4673 static inline void free_the_page(struct page *page, unsigned int order)
4675 if (order == 0) /* Via pcp? */
4676 free_unref_page(page);
4678 __free_pages_ok(page, order);
4681 void __free_pages(struct page *page, unsigned int order)
4683 if (put_page_testzero(page))
4684 free_the_page(page, order);
4686 EXPORT_SYMBOL(__free_pages);
4688 void free_pages(unsigned long addr, unsigned int order)
4691 VM_BUG_ON(!virt_addr_valid((void *)addr));
4692 __free_pages(virt_to_page((void *)addr), order);
4696 EXPORT_SYMBOL(free_pages);
4700 * An arbitrary-length arbitrary-offset area of memory which resides
4701 * within a 0 or higher order page. Multiple fragments within that page
4702 * are individually refcounted, in the page's reference counter.
4704 * The page_frag functions below provide a simple allocation framework for
4705 * page fragments. This is used by the network stack and network device
4706 * drivers to provide a backing region of memory for use as either an
4707 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4709 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4712 struct page *page = NULL;
4713 gfp_t gfp = gfp_mask;
4715 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4716 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4718 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4719 PAGE_FRAG_CACHE_MAX_ORDER);
4720 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4722 if (unlikely(!page))
4723 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4725 nc->va = page ? page_address(page) : NULL;
4730 void __page_frag_cache_drain(struct page *page, unsigned int count)
4732 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4734 if (page_ref_sub_and_test(page, count))
4735 free_the_page(page, compound_order(page));
4737 EXPORT_SYMBOL(__page_frag_cache_drain);
4739 void *page_frag_alloc(struct page_frag_cache *nc,
4740 unsigned int fragsz, gfp_t gfp_mask)
4742 unsigned int size = PAGE_SIZE;
4746 if (unlikely(!nc->va)) {
4748 page = __page_frag_cache_refill(nc, gfp_mask);
4752 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4753 /* if size can vary use size else just use PAGE_SIZE */
4756 /* Even if we own the page, we do not use atomic_set().
4757 * This would break get_page_unless_zero() users.
4759 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4761 /* reset page count bias and offset to start of new frag */
4762 nc->pfmemalloc = page_is_pfmemalloc(page);
4763 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4767 offset = nc->offset - fragsz;
4768 if (unlikely(offset < 0)) {
4769 page = virt_to_page(nc->va);
4771 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4774 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4775 /* if size can vary use size else just use PAGE_SIZE */
4778 /* OK, page count is 0, we can safely set it */
4779 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4781 /* reset page count bias and offset to start of new frag */
4782 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4783 offset = size - fragsz;
4787 nc->offset = offset;
4789 return nc->va + offset;
4791 EXPORT_SYMBOL(page_frag_alloc);
4794 * Frees a page fragment allocated out of either a compound or order 0 page.
4796 void page_frag_free(void *addr)
4798 struct page *page = virt_to_head_page(addr);
4800 if (unlikely(put_page_testzero(page)))
4801 free_the_page(page, compound_order(page));
4803 EXPORT_SYMBOL(page_frag_free);
4805 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4809 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4810 unsigned long used = addr + PAGE_ALIGN(size);
4812 split_page(virt_to_page((void *)addr), order);
4813 while (used < alloc_end) {
4818 return (void *)addr;
4822 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4823 * @size: the number of bytes to allocate
4824 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4826 * This function is similar to alloc_pages(), except that it allocates the
4827 * minimum number of pages to satisfy the request. alloc_pages() can only
4828 * allocate memory in power-of-two pages.
4830 * This function is also limited by MAX_ORDER.
4832 * Memory allocated by this function must be released by free_pages_exact().
4834 * Return: pointer to the allocated area or %NULL in case of error.
4836 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4838 unsigned int order = get_order(size);
4841 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4842 gfp_mask &= ~__GFP_COMP;
4844 addr = __get_free_pages(gfp_mask, order);
4845 return make_alloc_exact(addr, order, size);
4847 EXPORT_SYMBOL(alloc_pages_exact);
4850 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4852 * @nid: the preferred node ID where memory should be allocated
4853 * @size: the number of bytes to allocate
4854 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4856 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4859 * Return: pointer to the allocated area or %NULL in case of error.
4861 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4863 unsigned int order = get_order(size);
4866 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4867 gfp_mask &= ~__GFP_COMP;
4869 p = alloc_pages_node(nid, gfp_mask, order);
4872 return make_alloc_exact((unsigned long)page_address(p), order, size);
4876 * free_pages_exact - release memory allocated via alloc_pages_exact()
4877 * @virt: the value returned by alloc_pages_exact.
4878 * @size: size of allocation, same value as passed to alloc_pages_exact().
4880 * Release the memory allocated by a previous call to alloc_pages_exact.
4882 void free_pages_exact(void *virt, size_t size)
4884 unsigned long addr = (unsigned long)virt;
4885 unsigned long end = addr + PAGE_ALIGN(size);
4887 while (addr < end) {
4892 EXPORT_SYMBOL(free_pages_exact);
4895 * nr_free_zone_pages - count number of pages beyond high watermark
4896 * @offset: The zone index of the highest zone
4898 * nr_free_zone_pages() counts the number of pages which are beyond the
4899 * high watermark within all zones at or below a given zone index. For each
4900 * zone, the number of pages is calculated as:
4902 * nr_free_zone_pages = managed_pages - high_pages
4904 * Return: number of pages beyond high watermark.
4906 static unsigned long nr_free_zone_pages(int offset)
4911 /* Just pick one node, since fallback list is circular */
4912 unsigned long sum = 0;
4914 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4916 for_each_zone_zonelist(zone, z, zonelist, offset) {
4917 unsigned long size = zone_managed_pages(zone);
4918 unsigned long high = high_wmark_pages(zone);
4927 * nr_free_buffer_pages - count number of pages beyond high watermark
4929 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4930 * watermark within ZONE_DMA and ZONE_NORMAL.
4932 * Return: number of pages beyond high watermark within ZONE_DMA and
4935 unsigned long nr_free_buffer_pages(void)
4937 return nr_free_zone_pages(gfp_zone(GFP_USER));
4939 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4942 * nr_free_pagecache_pages - count number of pages beyond high watermark
4944 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4945 * high watermark within all zones.
4947 * Return: number of pages beyond high watermark within all zones.
4949 unsigned long nr_free_pagecache_pages(void)
4951 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4954 static inline void show_node(struct zone *zone)
4956 if (IS_ENABLED(CONFIG_NUMA))
4957 printk("Node %d ", zone_to_nid(zone));
4960 long si_mem_available(void)
4963 unsigned long pagecache;
4964 unsigned long wmark_low = 0;
4965 unsigned long pages[NR_LRU_LISTS];
4966 unsigned long reclaimable;
4970 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4971 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4974 wmark_low += low_wmark_pages(zone);
4977 * Estimate the amount of memory available for userspace allocations,
4978 * without causing swapping.
4980 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4983 * Not all the page cache can be freed, otherwise the system will
4984 * start swapping. Assume at least half of the page cache, or the
4985 * low watermark worth of cache, needs to stay.
4987 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4988 pagecache -= min(pagecache / 2, wmark_low);
4989 available += pagecache;
4992 * Part of the reclaimable slab and other kernel memory consists of
4993 * items that are in use, and cannot be freed. Cap this estimate at the
4996 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
4997 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
4998 available += reclaimable - min(reclaimable / 2, wmark_low);
5004 EXPORT_SYMBOL_GPL(si_mem_available);
5006 void si_meminfo(struct sysinfo *val)
5008 val->totalram = totalram_pages();
5009 val->sharedram = global_node_page_state(NR_SHMEM);
5010 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5011 val->bufferram = nr_blockdev_pages();
5012 val->totalhigh = totalhigh_pages();
5013 val->freehigh = nr_free_highpages();
5014 val->mem_unit = PAGE_SIZE;
5017 EXPORT_SYMBOL(si_meminfo);
5020 void si_meminfo_node(struct sysinfo *val, int nid)
5022 int zone_type; /* needs to be signed */
5023 unsigned long managed_pages = 0;
5024 unsigned long managed_highpages = 0;
5025 unsigned long free_highpages = 0;
5026 pg_data_t *pgdat = NODE_DATA(nid);
5028 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5029 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5030 val->totalram = managed_pages;
5031 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5032 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5033 #ifdef CONFIG_HIGHMEM
5034 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5035 struct zone *zone = &pgdat->node_zones[zone_type];
5037 if (is_highmem(zone)) {
5038 managed_highpages += zone_managed_pages(zone);
5039 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5042 val->totalhigh = managed_highpages;
5043 val->freehigh = free_highpages;
5045 val->totalhigh = managed_highpages;
5046 val->freehigh = free_highpages;
5048 val->mem_unit = PAGE_SIZE;
5053 * Determine whether the node should be displayed or not, depending on whether
5054 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5056 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5058 if (!(flags & SHOW_MEM_FILTER_NODES))
5062 * no node mask - aka implicit memory numa policy. Do not bother with
5063 * the synchronization - read_mems_allowed_begin - because we do not
5064 * have to be precise here.
5067 nodemask = &cpuset_current_mems_allowed;
5069 return !node_isset(nid, *nodemask);
5072 #define K(x) ((x) << (PAGE_SHIFT-10))
5074 static void show_migration_types(unsigned char type)
5076 static const char types[MIGRATE_TYPES] = {
5077 [MIGRATE_UNMOVABLE] = 'U',
5078 [MIGRATE_MOVABLE] = 'M',
5079 [MIGRATE_RECLAIMABLE] = 'E',
5080 [MIGRATE_HIGHATOMIC] = 'H',
5082 [MIGRATE_CMA] = 'C',
5084 #ifdef CONFIG_MEMORY_ISOLATION
5085 [MIGRATE_ISOLATE] = 'I',
5088 char tmp[MIGRATE_TYPES + 1];
5092 for (i = 0; i < MIGRATE_TYPES; i++) {
5093 if (type & (1 << i))
5098 printk(KERN_CONT "(%s) ", tmp);
5102 * Show free area list (used inside shift_scroll-lock stuff)
5103 * We also calculate the percentage fragmentation. We do this by counting the
5104 * memory on each free list with the exception of the first item on the list.
5107 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5110 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5112 unsigned long free_pcp = 0;
5117 for_each_populated_zone(zone) {
5118 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5121 for_each_online_cpu(cpu)
5122 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5125 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5126 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5127 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5128 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5129 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5130 " free:%lu free_pcp:%lu free_cma:%lu\n",
5131 global_node_page_state(NR_ACTIVE_ANON),
5132 global_node_page_state(NR_INACTIVE_ANON),
5133 global_node_page_state(NR_ISOLATED_ANON),
5134 global_node_page_state(NR_ACTIVE_FILE),
5135 global_node_page_state(NR_INACTIVE_FILE),
5136 global_node_page_state(NR_ISOLATED_FILE),
5137 global_node_page_state(NR_UNEVICTABLE),
5138 global_node_page_state(NR_FILE_DIRTY),
5139 global_node_page_state(NR_WRITEBACK),
5140 global_node_page_state(NR_UNSTABLE_NFS),
5141 global_node_page_state(NR_SLAB_RECLAIMABLE),
5142 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5143 global_node_page_state(NR_FILE_MAPPED),
5144 global_node_page_state(NR_SHMEM),
5145 global_zone_page_state(NR_PAGETABLE),
5146 global_zone_page_state(NR_BOUNCE),
5147 global_zone_page_state(NR_FREE_PAGES),
5149 global_zone_page_state(NR_FREE_CMA_PAGES));
5151 for_each_online_pgdat(pgdat) {
5152 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5156 " active_anon:%lukB"
5157 " inactive_anon:%lukB"
5158 " active_file:%lukB"
5159 " inactive_file:%lukB"
5160 " unevictable:%lukB"
5161 " isolated(anon):%lukB"
5162 " isolated(file):%lukB"
5167 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5169 " shmem_pmdmapped: %lukB"
5172 " writeback_tmp:%lukB"
5174 " all_unreclaimable? %s"
5177 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5178 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5179 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5180 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5181 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5182 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5183 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5184 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5185 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5186 K(node_page_state(pgdat, NR_WRITEBACK)),
5187 K(node_page_state(pgdat, NR_SHMEM)),
5188 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5189 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5190 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5192 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5194 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5195 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5196 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5200 for_each_populated_zone(zone) {
5203 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5207 for_each_online_cpu(cpu)
5208 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5217 " active_anon:%lukB"
5218 " inactive_anon:%lukB"
5219 " active_file:%lukB"
5220 " inactive_file:%lukB"
5221 " unevictable:%lukB"
5222 " writepending:%lukB"
5226 " kernel_stack:%lukB"
5234 K(zone_page_state(zone, NR_FREE_PAGES)),
5235 K(min_wmark_pages(zone)),
5236 K(low_wmark_pages(zone)),
5237 K(high_wmark_pages(zone)),
5238 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5239 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5240 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5241 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5242 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5243 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5244 K(zone->present_pages),
5245 K(zone_managed_pages(zone)),
5246 K(zone_page_state(zone, NR_MLOCK)),
5247 zone_page_state(zone, NR_KERNEL_STACK_KB),
5248 K(zone_page_state(zone, NR_PAGETABLE)),
5249 K(zone_page_state(zone, NR_BOUNCE)),
5251 K(this_cpu_read(zone->pageset->pcp.count)),
5252 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5253 printk("lowmem_reserve[]:");
5254 for (i = 0; i < MAX_NR_ZONES; i++)
5255 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5256 printk(KERN_CONT "\n");
5259 for_each_populated_zone(zone) {
5261 unsigned long nr[MAX_ORDER], flags, total = 0;
5262 unsigned char types[MAX_ORDER];
5264 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5267 printk(KERN_CONT "%s: ", zone->name);
5269 spin_lock_irqsave(&zone->lock, flags);
5270 for (order = 0; order < MAX_ORDER; order++) {
5271 struct free_area *area = &zone->free_area[order];
5274 nr[order] = area->nr_free;
5275 total += nr[order] << order;
5278 for (type = 0; type < MIGRATE_TYPES; type++) {
5279 if (!list_empty(&area->free_list[type]))
5280 types[order] |= 1 << type;
5283 spin_unlock_irqrestore(&zone->lock, flags);
5284 for (order = 0; order < MAX_ORDER; order++) {
5285 printk(KERN_CONT "%lu*%lukB ",
5286 nr[order], K(1UL) << order);
5288 show_migration_types(types[order]);
5290 printk(KERN_CONT "= %lukB\n", K(total));
5293 hugetlb_show_meminfo();
5295 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5297 show_swap_cache_info();
5300 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5302 zoneref->zone = zone;
5303 zoneref->zone_idx = zone_idx(zone);
5307 * Builds allocation fallback zone lists.
5309 * Add all populated zones of a node to the zonelist.
5311 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5314 enum zone_type zone_type = MAX_NR_ZONES;
5319 zone = pgdat->node_zones + zone_type;
5320 if (managed_zone(zone)) {
5321 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5322 check_highest_zone(zone_type);
5324 } while (zone_type);
5331 static int __parse_numa_zonelist_order(char *s)
5334 * We used to support different zonlists modes but they turned
5335 * out to be just not useful. Let's keep the warning in place
5336 * if somebody still use the cmd line parameter so that we do
5337 * not fail it silently
5339 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5340 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5346 static __init int setup_numa_zonelist_order(char *s)
5351 return __parse_numa_zonelist_order(s);
5353 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5355 char numa_zonelist_order[] = "Node";
5358 * sysctl handler for numa_zonelist_order
5360 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5361 void __user *buffer, size_t *length,
5368 return proc_dostring(table, write, buffer, length, ppos);
5369 str = memdup_user_nul(buffer, 16);
5371 return PTR_ERR(str);
5373 ret = __parse_numa_zonelist_order(str);
5379 #define MAX_NODE_LOAD (nr_online_nodes)
5380 static int node_load[MAX_NUMNODES];
5383 * find_next_best_node - find the next node that should appear in a given node's fallback list
5384 * @node: node whose fallback list we're appending
5385 * @used_node_mask: nodemask_t of already used nodes
5387 * We use a number of factors to determine which is the next node that should
5388 * appear on a given node's fallback list. The node should not have appeared
5389 * already in @node's fallback list, and it should be the next closest node
5390 * according to the distance array (which contains arbitrary distance values
5391 * from each node to each node in the system), and should also prefer nodes
5392 * with no CPUs, since presumably they'll have very little allocation pressure
5393 * on them otherwise.
5395 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5397 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5400 int min_val = INT_MAX;
5401 int best_node = NUMA_NO_NODE;
5402 const struct cpumask *tmp = cpumask_of_node(0);
5404 /* Use the local node if we haven't already */
5405 if (!node_isset(node, *used_node_mask)) {
5406 node_set(node, *used_node_mask);
5410 for_each_node_state(n, N_MEMORY) {
5412 /* Don't want a node to appear more than once */
5413 if (node_isset(n, *used_node_mask))
5416 /* Use the distance array to find the distance */
5417 val = node_distance(node, n);
5419 /* Penalize nodes under us ("prefer the next node") */
5422 /* Give preference to headless and unused nodes */
5423 tmp = cpumask_of_node(n);
5424 if (!cpumask_empty(tmp))
5425 val += PENALTY_FOR_NODE_WITH_CPUS;
5427 /* Slight preference for less loaded node */
5428 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5429 val += node_load[n];
5431 if (val < min_val) {
5438 node_set(best_node, *used_node_mask);
5445 * Build zonelists ordered by node and zones within node.
5446 * This results in maximum locality--normal zone overflows into local
5447 * DMA zone, if any--but risks exhausting DMA zone.
5449 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5452 struct zoneref *zonerefs;
5455 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5457 for (i = 0; i < nr_nodes; i++) {
5460 pg_data_t *node = NODE_DATA(node_order[i]);
5462 nr_zones = build_zonerefs_node(node, zonerefs);
5463 zonerefs += nr_zones;
5465 zonerefs->zone = NULL;
5466 zonerefs->zone_idx = 0;
5470 * Build gfp_thisnode zonelists
5472 static void build_thisnode_zonelists(pg_data_t *pgdat)
5474 struct zoneref *zonerefs;
5477 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5478 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5479 zonerefs += nr_zones;
5480 zonerefs->zone = NULL;
5481 zonerefs->zone_idx = 0;
5485 * Build zonelists ordered by zone and nodes within zones.
5486 * This results in conserving DMA zone[s] until all Normal memory is
5487 * exhausted, but results in overflowing to remote node while memory
5488 * may still exist in local DMA zone.
5491 static void build_zonelists(pg_data_t *pgdat)
5493 static int node_order[MAX_NUMNODES];
5494 int node, load, nr_nodes = 0;
5495 nodemask_t used_mask;
5496 int local_node, prev_node;
5498 /* NUMA-aware ordering of nodes */
5499 local_node = pgdat->node_id;
5500 load = nr_online_nodes;
5501 prev_node = local_node;
5502 nodes_clear(used_mask);
5504 memset(node_order, 0, sizeof(node_order));
5505 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5507 * We don't want to pressure a particular node.
5508 * So adding penalty to the first node in same
5509 * distance group to make it round-robin.
5511 if (node_distance(local_node, node) !=
5512 node_distance(local_node, prev_node))
5513 node_load[node] = load;
5515 node_order[nr_nodes++] = node;
5520 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5521 build_thisnode_zonelists(pgdat);
5524 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5526 * Return node id of node used for "local" allocations.
5527 * I.e., first node id of first zone in arg node's generic zonelist.
5528 * Used for initializing percpu 'numa_mem', which is used primarily
5529 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5531 int local_memory_node(int node)
5535 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5536 gfp_zone(GFP_KERNEL),
5538 return zone_to_nid(z->zone);
5542 static void setup_min_unmapped_ratio(void);
5543 static void setup_min_slab_ratio(void);
5544 #else /* CONFIG_NUMA */
5546 static void build_zonelists(pg_data_t *pgdat)
5548 int node, local_node;
5549 struct zoneref *zonerefs;
5552 local_node = pgdat->node_id;
5554 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5555 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5556 zonerefs += nr_zones;
5559 * Now we build the zonelist so that it contains the zones
5560 * of all the other nodes.
5561 * We don't want to pressure a particular node, so when
5562 * building the zones for node N, we make sure that the
5563 * zones coming right after the local ones are those from
5564 * node N+1 (modulo N)
5566 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5567 if (!node_online(node))
5569 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5570 zonerefs += nr_zones;
5572 for (node = 0; node < local_node; node++) {
5573 if (!node_online(node))
5575 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5576 zonerefs += nr_zones;
5579 zonerefs->zone = NULL;
5580 zonerefs->zone_idx = 0;
5583 #endif /* CONFIG_NUMA */
5586 * Boot pageset table. One per cpu which is going to be used for all
5587 * zones and all nodes. The parameters will be set in such a way
5588 * that an item put on a list will immediately be handed over to
5589 * the buddy list. This is safe since pageset manipulation is done
5590 * with interrupts disabled.
5592 * The boot_pagesets must be kept even after bootup is complete for
5593 * unused processors and/or zones. They do play a role for bootstrapping
5594 * hotplugged processors.
5596 * zoneinfo_show() and maybe other functions do
5597 * not check if the processor is online before following the pageset pointer.
5598 * Other parts of the kernel may not check if the zone is available.
5600 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5601 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5602 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5604 static void __build_all_zonelists(void *data)
5607 int __maybe_unused cpu;
5608 pg_data_t *self = data;
5609 static DEFINE_SPINLOCK(lock);
5614 memset(node_load, 0, sizeof(node_load));
5618 * This node is hotadded and no memory is yet present. So just
5619 * building zonelists is fine - no need to touch other nodes.
5621 if (self && !node_online(self->node_id)) {
5622 build_zonelists(self);
5624 for_each_online_node(nid) {
5625 pg_data_t *pgdat = NODE_DATA(nid);
5627 build_zonelists(pgdat);
5630 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5632 * We now know the "local memory node" for each node--
5633 * i.e., the node of the first zone in the generic zonelist.
5634 * Set up numa_mem percpu variable for on-line cpus. During
5635 * boot, only the boot cpu should be on-line; we'll init the
5636 * secondary cpus' numa_mem as they come on-line. During
5637 * node/memory hotplug, we'll fixup all on-line cpus.
5639 for_each_online_cpu(cpu)
5640 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5647 static noinline void __init
5648 build_all_zonelists_init(void)
5652 __build_all_zonelists(NULL);
5655 * Initialize the boot_pagesets that are going to be used
5656 * for bootstrapping processors. The real pagesets for
5657 * each zone will be allocated later when the per cpu
5658 * allocator is available.
5660 * boot_pagesets are used also for bootstrapping offline
5661 * cpus if the system is already booted because the pagesets
5662 * are needed to initialize allocators on a specific cpu too.
5663 * F.e. the percpu allocator needs the page allocator which
5664 * needs the percpu allocator in order to allocate its pagesets
5665 * (a chicken-egg dilemma).
5667 for_each_possible_cpu(cpu)
5668 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5670 mminit_verify_zonelist();
5671 cpuset_init_current_mems_allowed();
5675 * unless system_state == SYSTEM_BOOTING.
5677 * __ref due to call of __init annotated helper build_all_zonelists_init
5678 * [protected by SYSTEM_BOOTING].
5680 void __ref build_all_zonelists(pg_data_t *pgdat)
5682 if (system_state == SYSTEM_BOOTING) {
5683 build_all_zonelists_init();
5685 __build_all_zonelists(pgdat);
5686 /* cpuset refresh routine should be here */
5688 vm_total_pages = nr_free_pagecache_pages();
5690 * Disable grouping by mobility if the number of pages in the
5691 * system is too low to allow the mechanism to work. It would be
5692 * more accurate, but expensive to check per-zone. This check is
5693 * made on memory-hotadd so a system can start with mobility
5694 * disabled and enable it later
5696 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5697 page_group_by_mobility_disabled = 1;
5699 page_group_by_mobility_disabled = 0;
5701 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5703 page_group_by_mobility_disabled ? "off" : "on",
5706 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5710 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5711 static bool __meminit
5712 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5714 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5715 static struct memblock_region *r;
5717 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5718 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5719 for_each_memblock(memory, r) {
5720 if (*pfn < memblock_region_memory_end_pfn(r))
5724 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5725 memblock_is_mirror(r)) {
5726 *pfn = memblock_region_memory_end_pfn(r);
5735 * Initially all pages are reserved - free ones are freed
5736 * up by memblock_free_all() once the early boot process is
5737 * done. Non-atomic initialization, single-pass.
5739 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5740 unsigned long start_pfn, enum memmap_context context,
5741 struct vmem_altmap *altmap)
5743 unsigned long pfn, end_pfn = start_pfn + size;
5746 if (highest_memmap_pfn < end_pfn - 1)
5747 highest_memmap_pfn = end_pfn - 1;
5749 #ifdef CONFIG_ZONE_DEVICE
5751 * Honor reservation requested by the driver for this ZONE_DEVICE
5752 * memory. We limit the total number of pages to initialize to just
5753 * those that might contain the memory mapping. We will defer the
5754 * ZONE_DEVICE page initialization until after we have released
5757 if (zone == ZONE_DEVICE) {
5761 if (start_pfn == altmap->base_pfn)
5762 start_pfn += altmap->reserve;
5763 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5767 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5769 * There can be holes in boot-time mem_map[]s handed to this
5770 * function. They do not exist on hotplugged memory.
5772 if (context == MEMMAP_EARLY) {
5773 if (!early_pfn_valid(pfn))
5775 if (!early_pfn_in_nid(pfn, nid))
5777 if (overlap_memmap_init(zone, &pfn))
5779 if (defer_init(nid, pfn, end_pfn))
5783 page = pfn_to_page(pfn);
5784 __init_single_page(page, pfn, zone, nid);
5785 if (context == MEMMAP_HOTPLUG)
5786 __SetPageReserved(page);
5789 * Mark the block movable so that blocks are reserved for
5790 * movable at startup. This will force kernel allocations
5791 * to reserve their blocks rather than leaking throughout
5792 * the address space during boot when many long-lived
5793 * kernel allocations are made.
5795 * bitmap is created for zone's valid pfn range. but memmap
5796 * can be created for invalid pages (for alignment)
5797 * check here not to call set_pageblock_migratetype() against
5800 if (!(pfn & (pageblock_nr_pages - 1))) {
5801 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5807 #ifdef CONFIG_ZONE_DEVICE
5808 void __ref memmap_init_zone_device(struct zone *zone,
5809 unsigned long start_pfn,
5811 struct dev_pagemap *pgmap)
5813 unsigned long pfn, end_pfn = start_pfn + size;
5814 struct pglist_data *pgdat = zone->zone_pgdat;
5815 unsigned long zone_idx = zone_idx(zone);
5816 unsigned long start = jiffies;
5817 int nid = pgdat->node_id;
5819 if (WARN_ON_ONCE(!pgmap || !is_dev_zone(zone)))
5823 * The call to memmap_init_zone should have already taken care
5824 * of the pages reserved for the memmap, so we can just jump to
5825 * the end of that region and start processing the device pages.
5827 if (pgmap->altmap_valid) {
5828 struct vmem_altmap *altmap = &pgmap->altmap;
5830 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5831 size = end_pfn - start_pfn;
5834 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5835 struct page *page = pfn_to_page(pfn);
5837 __init_single_page(page, pfn, zone_idx, nid);
5840 * Mark page reserved as it will need to wait for onlining
5841 * phase for it to be fully associated with a zone.
5843 * We can use the non-atomic __set_bit operation for setting
5844 * the flag as we are still initializing the pages.
5846 __SetPageReserved(page);
5849 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5850 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5851 * page is ever freed or placed on a driver-private list.
5853 page->pgmap = pgmap;
5857 * Mark the block movable so that blocks are reserved for
5858 * movable at startup. This will force kernel allocations
5859 * to reserve their blocks rather than leaking throughout
5860 * the address space during boot when many long-lived
5861 * kernel allocations are made.
5863 * bitmap is created for zone's valid pfn range. but memmap
5864 * can be created for invalid pages (for alignment)
5865 * check here not to call set_pageblock_migratetype() against
5868 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5869 * because this is done early in sparse_add_one_section
5871 if (!(pfn & (pageblock_nr_pages - 1))) {
5872 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5877 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap->dev),
5878 size, jiffies_to_msecs(jiffies - start));
5882 static void __meminit zone_init_free_lists(struct zone *zone)
5884 unsigned int order, t;
5885 for_each_migratetype_order(order, t) {
5886 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5887 zone->free_area[order].nr_free = 0;
5891 void __meminit __weak memmap_init(unsigned long size, int nid,
5892 unsigned long zone, unsigned long start_pfn)
5894 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
5897 static int zone_batchsize(struct zone *zone)
5903 * The per-cpu-pages pools are set to around 1000th of the
5906 batch = zone_managed_pages(zone) / 1024;
5907 /* But no more than a meg. */
5908 if (batch * PAGE_SIZE > 1024 * 1024)
5909 batch = (1024 * 1024) / PAGE_SIZE;
5910 batch /= 4; /* We effectively *= 4 below */
5915 * Clamp the batch to a 2^n - 1 value. Having a power
5916 * of 2 value was found to be more likely to have
5917 * suboptimal cache aliasing properties in some cases.
5919 * For example if 2 tasks are alternately allocating
5920 * batches of pages, one task can end up with a lot
5921 * of pages of one half of the possible page colors
5922 * and the other with pages of the other colors.
5924 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5929 /* The deferral and batching of frees should be suppressed under NOMMU
5932 * The problem is that NOMMU needs to be able to allocate large chunks
5933 * of contiguous memory as there's no hardware page translation to
5934 * assemble apparent contiguous memory from discontiguous pages.
5936 * Queueing large contiguous runs of pages for batching, however,
5937 * causes the pages to actually be freed in smaller chunks. As there
5938 * can be a significant delay between the individual batches being
5939 * recycled, this leads to the once large chunks of space being
5940 * fragmented and becoming unavailable for high-order allocations.
5947 * pcp->high and pcp->batch values are related and dependent on one another:
5948 * ->batch must never be higher then ->high.
5949 * The following function updates them in a safe manner without read side
5952 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5953 * those fields changing asynchronously (acording the the above rule).
5955 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5956 * outside of boot time (or some other assurance that no concurrent updaters
5959 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5960 unsigned long batch)
5962 /* start with a fail safe value for batch */
5966 /* Update high, then batch, in order */
5973 /* a companion to pageset_set_high() */
5974 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5976 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5979 static void pageset_init(struct per_cpu_pageset *p)
5981 struct per_cpu_pages *pcp;
5984 memset(p, 0, sizeof(*p));
5987 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5988 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5991 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5994 pageset_set_batch(p, batch);
5998 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5999 * to the value high for the pageset p.
6001 static void pageset_set_high(struct per_cpu_pageset *p,
6004 unsigned long batch = max(1UL, high / 4);
6005 if ((high / 4) > (PAGE_SHIFT * 8))
6006 batch = PAGE_SHIFT * 8;
6008 pageset_update(&p->pcp, high, batch);
6011 static void pageset_set_high_and_batch(struct zone *zone,
6012 struct per_cpu_pageset *pcp)
6014 if (percpu_pagelist_fraction)
6015 pageset_set_high(pcp,
6016 (zone_managed_pages(zone) /
6017 percpu_pagelist_fraction));
6019 pageset_set_batch(pcp, zone_batchsize(zone));
6022 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6024 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6027 pageset_set_high_and_batch(zone, pcp);
6030 void __meminit setup_zone_pageset(struct zone *zone)
6033 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6034 for_each_possible_cpu(cpu)
6035 zone_pageset_init(zone, cpu);
6039 * Allocate per cpu pagesets and initialize them.
6040 * Before this call only boot pagesets were available.
6042 void __init setup_per_cpu_pageset(void)
6044 struct pglist_data *pgdat;
6047 for_each_populated_zone(zone)
6048 setup_zone_pageset(zone);
6050 for_each_online_pgdat(pgdat)
6051 pgdat->per_cpu_nodestats =
6052 alloc_percpu(struct per_cpu_nodestat);
6055 static __meminit void zone_pcp_init(struct zone *zone)
6058 * per cpu subsystem is not up at this point. The following code
6059 * relies on the ability of the linker to provide the
6060 * offset of a (static) per cpu variable into the per cpu area.
6062 zone->pageset = &boot_pageset;
6064 if (populated_zone(zone))
6065 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6066 zone->name, zone->present_pages,
6067 zone_batchsize(zone));
6070 void __meminit init_currently_empty_zone(struct zone *zone,
6071 unsigned long zone_start_pfn,
6074 struct pglist_data *pgdat = zone->zone_pgdat;
6075 int zone_idx = zone_idx(zone) + 1;
6077 if (zone_idx > pgdat->nr_zones)
6078 pgdat->nr_zones = zone_idx;
6080 zone->zone_start_pfn = zone_start_pfn;
6082 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6083 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6085 (unsigned long)zone_idx(zone),
6086 zone_start_pfn, (zone_start_pfn + size));
6088 zone_init_free_lists(zone);
6089 zone->initialized = 1;
6092 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6093 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6096 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6098 int __meminit __early_pfn_to_nid(unsigned long pfn,
6099 struct mminit_pfnnid_cache *state)
6101 unsigned long start_pfn, end_pfn;
6104 if (state->last_start <= pfn && pfn < state->last_end)
6105 return state->last_nid;
6107 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6108 if (nid != NUMA_NO_NODE) {
6109 state->last_start = start_pfn;
6110 state->last_end = end_pfn;
6111 state->last_nid = nid;
6116 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6119 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6120 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6121 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6123 * If an architecture guarantees that all ranges registered contain no holes
6124 * and may be freed, this this function may be used instead of calling
6125 * memblock_free_early_nid() manually.
6127 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6129 unsigned long start_pfn, end_pfn;
6132 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6133 start_pfn = min(start_pfn, max_low_pfn);
6134 end_pfn = min(end_pfn, max_low_pfn);
6136 if (start_pfn < end_pfn)
6137 memblock_free_early_nid(PFN_PHYS(start_pfn),
6138 (end_pfn - start_pfn) << PAGE_SHIFT,
6144 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6145 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6147 * If an architecture guarantees that all ranges registered contain no holes and may
6148 * be freed, this function may be used instead of calling memory_present() manually.
6150 void __init sparse_memory_present_with_active_regions(int nid)
6152 unsigned long start_pfn, end_pfn;
6155 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6156 memory_present(this_nid, start_pfn, end_pfn);
6160 * get_pfn_range_for_nid - Return the start and end page frames for a node
6161 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6162 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6163 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6165 * It returns the start and end page frame of a node based on information
6166 * provided by memblock_set_node(). If called for a node
6167 * with no available memory, a warning is printed and the start and end
6170 void __init get_pfn_range_for_nid(unsigned int nid,
6171 unsigned long *start_pfn, unsigned long *end_pfn)
6173 unsigned long this_start_pfn, this_end_pfn;
6179 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6180 *start_pfn = min(*start_pfn, this_start_pfn);
6181 *end_pfn = max(*end_pfn, this_end_pfn);
6184 if (*start_pfn == -1UL)
6189 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6190 * assumption is made that zones within a node are ordered in monotonic
6191 * increasing memory addresses so that the "highest" populated zone is used
6193 static void __init find_usable_zone_for_movable(void)
6196 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6197 if (zone_index == ZONE_MOVABLE)
6200 if (arch_zone_highest_possible_pfn[zone_index] >
6201 arch_zone_lowest_possible_pfn[zone_index])
6205 VM_BUG_ON(zone_index == -1);
6206 movable_zone = zone_index;
6210 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6211 * because it is sized independent of architecture. Unlike the other zones,
6212 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6213 * in each node depending on the size of each node and how evenly kernelcore
6214 * is distributed. This helper function adjusts the zone ranges
6215 * provided by the architecture for a given node by using the end of the
6216 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6217 * zones within a node are in order of monotonic increases memory addresses
6219 static void __init adjust_zone_range_for_zone_movable(int nid,
6220 unsigned long zone_type,
6221 unsigned long node_start_pfn,
6222 unsigned long node_end_pfn,
6223 unsigned long *zone_start_pfn,
6224 unsigned long *zone_end_pfn)
6226 /* Only adjust if ZONE_MOVABLE is on this node */
6227 if (zone_movable_pfn[nid]) {
6228 /* Size ZONE_MOVABLE */
6229 if (zone_type == ZONE_MOVABLE) {
6230 *zone_start_pfn = zone_movable_pfn[nid];
6231 *zone_end_pfn = min(node_end_pfn,
6232 arch_zone_highest_possible_pfn[movable_zone]);
6234 /* Adjust for ZONE_MOVABLE starting within this range */
6235 } else if (!mirrored_kernelcore &&
6236 *zone_start_pfn < zone_movable_pfn[nid] &&
6237 *zone_end_pfn > zone_movable_pfn[nid]) {
6238 *zone_end_pfn = zone_movable_pfn[nid];
6240 /* Check if this whole range is within ZONE_MOVABLE */
6241 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6242 *zone_start_pfn = *zone_end_pfn;
6247 * Return the number of pages a zone spans in a node, including holes
6248 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6250 static unsigned long __init zone_spanned_pages_in_node(int nid,
6251 unsigned long zone_type,
6252 unsigned long node_start_pfn,
6253 unsigned long node_end_pfn,
6254 unsigned long *zone_start_pfn,
6255 unsigned long *zone_end_pfn,
6256 unsigned long *ignored)
6258 /* When hotadd a new node from cpu_up(), the node should be empty */
6259 if (!node_start_pfn && !node_end_pfn)
6262 /* Get the start and end of the zone */
6263 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
6264 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
6265 adjust_zone_range_for_zone_movable(nid, zone_type,
6266 node_start_pfn, node_end_pfn,
6267 zone_start_pfn, zone_end_pfn);
6269 /* Check that this node has pages within the zone's required range */
6270 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6273 /* Move the zone boundaries inside the node if necessary */
6274 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6275 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6277 /* Return the spanned pages */
6278 return *zone_end_pfn - *zone_start_pfn;
6282 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6283 * then all holes in the requested range will be accounted for.
6285 unsigned long __init __absent_pages_in_range(int nid,
6286 unsigned long range_start_pfn,
6287 unsigned long range_end_pfn)
6289 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6290 unsigned long start_pfn, end_pfn;
6293 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6294 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6295 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6296 nr_absent -= end_pfn - start_pfn;
6302 * absent_pages_in_range - Return number of page frames in holes within a range
6303 * @start_pfn: The start PFN to start searching for holes
6304 * @end_pfn: The end PFN to stop searching for holes
6306 * Return: the number of pages frames in memory holes within a range.
6308 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6309 unsigned long end_pfn)
6311 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6314 /* Return the number of page frames in holes in a zone on a node */
6315 static unsigned long __init zone_absent_pages_in_node(int nid,
6316 unsigned long zone_type,
6317 unsigned long node_start_pfn,
6318 unsigned long node_end_pfn,
6319 unsigned long *ignored)
6321 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6322 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6323 unsigned long zone_start_pfn, zone_end_pfn;
6324 unsigned long nr_absent;
6326 /* When hotadd a new node from cpu_up(), the node should be empty */
6327 if (!node_start_pfn && !node_end_pfn)
6330 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6331 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6333 adjust_zone_range_for_zone_movable(nid, zone_type,
6334 node_start_pfn, node_end_pfn,
6335 &zone_start_pfn, &zone_end_pfn);
6336 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6339 * ZONE_MOVABLE handling.
6340 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6343 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6344 unsigned long start_pfn, end_pfn;
6345 struct memblock_region *r;
6347 for_each_memblock(memory, r) {
6348 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6349 zone_start_pfn, zone_end_pfn);
6350 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6351 zone_start_pfn, zone_end_pfn);
6353 if (zone_type == ZONE_MOVABLE &&
6354 memblock_is_mirror(r))
6355 nr_absent += end_pfn - start_pfn;
6357 if (zone_type == ZONE_NORMAL &&
6358 !memblock_is_mirror(r))
6359 nr_absent += end_pfn - start_pfn;
6366 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6367 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6368 unsigned long zone_type,
6369 unsigned long node_start_pfn,
6370 unsigned long node_end_pfn,
6371 unsigned long *zone_start_pfn,
6372 unsigned long *zone_end_pfn,
6373 unsigned long *zones_size)
6377 *zone_start_pfn = node_start_pfn;
6378 for (zone = 0; zone < zone_type; zone++)
6379 *zone_start_pfn += zones_size[zone];
6381 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6383 return zones_size[zone_type];
6386 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6387 unsigned long zone_type,
6388 unsigned long node_start_pfn,
6389 unsigned long node_end_pfn,
6390 unsigned long *zholes_size)
6395 return zholes_size[zone_type];
6398 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6400 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6401 unsigned long node_start_pfn,
6402 unsigned long node_end_pfn,
6403 unsigned long *zones_size,
6404 unsigned long *zholes_size)
6406 unsigned long realtotalpages = 0, totalpages = 0;
6409 for (i = 0; i < MAX_NR_ZONES; i++) {
6410 struct zone *zone = pgdat->node_zones + i;
6411 unsigned long zone_start_pfn, zone_end_pfn;
6412 unsigned long size, real_size;
6414 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6420 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6421 node_start_pfn, node_end_pfn,
6424 zone->zone_start_pfn = zone_start_pfn;
6426 zone->zone_start_pfn = 0;
6427 zone->spanned_pages = size;
6428 zone->present_pages = real_size;
6431 realtotalpages += real_size;
6434 pgdat->node_spanned_pages = totalpages;
6435 pgdat->node_present_pages = realtotalpages;
6436 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6440 #ifndef CONFIG_SPARSEMEM
6442 * Calculate the size of the zone->blockflags rounded to an unsigned long
6443 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6444 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6445 * round what is now in bits to nearest long in bits, then return it in
6448 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6450 unsigned long usemapsize;
6452 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6453 usemapsize = roundup(zonesize, pageblock_nr_pages);
6454 usemapsize = usemapsize >> pageblock_order;
6455 usemapsize *= NR_PAGEBLOCK_BITS;
6456 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6458 return usemapsize / 8;
6461 static void __ref setup_usemap(struct pglist_data *pgdat,
6463 unsigned long zone_start_pfn,
6464 unsigned long zonesize)
6466 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6467 zone->pageblock_flags = NULL;
6469 zone->pageblock_flags =
6470 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6472 if (!zone->pageblock_flags)
6473 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6474 usemapsize, zone->name, pgdat->node_id);
6478 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6479 unsigned long zone_start_pfn, unsigned long zonesize) {}
6480 #endif /* CONFIG_SPARSEMEM */
6482 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6484 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6485 void __init set_pageblock_order(void)
6489 /* Check that pageblock_nr_pages has not already been setup */
6490 if (pageblock_order)
6493 if (HPAGE_SHIFT > PAGE_SHIFT)
6494 order = HUGETLB_PAGE_ORDER;
6496 order = MAX_ORDER - 1;
6499 * Assume the largest contiguous order of interest is a huge page.
6500 * This value may be variable depending on boot parameters on IA64 and
6503 pageblock_order = order;
6505 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6508 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6509 * is unused as pageblock_order is set at compile-time. See
6510 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6513 void __init set_pageblock_order(void)
6517 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6519 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6520 unsigned long present_pages)
6522 unsigned long pages = spanned_pages;
6525 * Provide a more accurate estimation if there are holes within
6526 * the zone and SPARSEMEM is in use. If there are holes within the
6527 * zone, each populated memory region may cost us one or two extra
6528 * memmap pages due to alignment because memmap pages for each
6529 * populated regions may not be naturally aligned on page boundary.
6530 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6532 if (spanned_pages > present_pages + (present_pages >> 4) &&
6533 IS_ENABLED(CONFIG_SPARSEMEM))
6534 pages = present_pages;
6536 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6539 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6540 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6542 spin_lock_init(&pgdat->split_queue_lock);
6543 INIT_LIST_HEAD(&pgdat->split_queue);
6544 pgdat->split_queue_len = 0;
6547 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6550 #ifdef CONFIG_COMPACTION
6551 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6553 init_waitqueue_head(&pgdat->kcompactd_wait);
6556 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6559 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6561 pgdat_resize_init(pgdat);
6563 pgdat_init_split_queue(pgdat);
6564 pgdat_init_kcompactd(pgdat);
6566 init_waitqueue_head(&pgdat->kswapd_wait);
6567 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6569 pgdat_page_ext_init(pgdat);
6570 spin_lock_init(&pgdat->lru_lock);
6571 lruvec_init(node_lruvec(pgdat));
6574 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6575 unsigned long remaining_pages)
6577 atomic_long_set(&zone->managed_pages, remaining_pages);
6578 zone_set_nid(zone, nid);
6579 zone->name = zone_names[idx];
6580 zone->zone_pgdat = NODE_DATA(nid);
6581 spin_lock_init(&zone->lock);
6582 zone_seqlock_init(zone);
6583 zone_pcp_init(zone);
6587 * Set up the zone data structures
6588 * - init pgdat internals
6589 * - init all zones belonging to this node
6591 * NOTE: this function is only called during memory hotplug
6593 #ifdef CONFIG_MEMORY_HOTPLUG
6594 void __ref free_area_init_core_hotplug(int nid)
6597 pg_data_t *pgdat = NODE_DATA(nid);
6599 pgdat_init_internals(pgdat);
6600 for (z = 0; z < MAX_NR_ZONES; z++)
6601 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6606 * Set up the zone data structures:
6607 * - mark all pages reserved
6608 * - mark all memory queues empty
6609 * - clear the memory bitmaps
6611 * NOTE: pgdat should get zeroed by caller.
6612 * NOTE: this function is only called during early init.
6614 static void __init free_area_init_core(struct pglist_data *pgdat)
6617 int nid = pgdat->node_id;
6619 pgdat_init_internals(pgdat);
6620 pgdat->per_cpu_nodestats = &boot_nodestats;
6622 for (j = 0; j < MAX_NR_ZONES; j++) {
6623 struct zone *zone = pgdat->node_zones + j;
6624 unsigned long size, freesize, memmap_pages;
6625 unsigned long zone_start_pfn = zone->zone_start_pfn;
6627 size = zone->spanned_pages;
6628 freesize = zone->present_pages;
6631 * Adjust freesize so that it accounts for how much memory
6632 * is used by this zone for memmap. This affects the watermark
6633 * and per-cpu initialisations
6635 memmap_pages = calc_memmap_size(size, freesize);
6636 if (!is_highmem_idx(j)) {
6637 if (freesize >= memmap_pages) {
6638 freesize -= memmap_pages;
6641 " %s zone: %lu pages used for memmap\n",
6642 zone_names[j], memmap_pages);
6644 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6645 zone_names[j], memmap_pages, freesize);
6648 /* Account for reserved pages */
6649 if (j == 0 && freesize > dma_reserve) {
6650 freesize -= dma_reserve;
6651 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6652 zone_names[0], dma_reserve);
6655 if (!is_highmem_idx(j))
6656 nr_kernel_pages += freesize;
6657 /* Charge for highmem memmap if there are enough kernel pages */
6658 else if (nr_kernel_pages > memmap_pages * 2)
6659 nr_kernel_pages -= memmap_pages;
6660 nr_all_pages += freesize;
6663 * Set an approximate value for lowmem here, it will be adjusted
6664 * when the bootmem allocator frees pages into the buddy system.
6665 * And all highmem pages will be managed by the buddy system.
6667 zone_init_internals(zone, j, nid, freesize);
6672 set_pageblock_order();
6673 setup_usemap(pgdat, zone, zone_start_pfn, size);
6674 init_currently_empty_zone(zone, zone_start_pfn, size);
6675 memmap_init(size, nid, j, zone_start_pfn);
6679 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6680 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6682 unsigned long __maybe_unused start = 0;
6683 unsigned long __maybe_unused offset = 0;
6685 /* Skip empty nodes */
6686 if (!pgdat->node_spanned_pages)
6689 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6690 offset = pgdat->node_start_pfn - start;
6691 /* ia64 gets its own node_mem_map, before this, without bootmem */
6692 if (!pgdat->node_mem_map) {
6693 unsigned long size, end;
6697 * The zone's endpoints aren't required to be MAX_ORDER
6698 * aligned but the node_mem_map endpoints must be in order
6699 * for the buddy allocator to function correctly.
6701 end = pgdat_end_pfn(pgdat);
6702 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6703 size = (end - start) * sizeof(struct page);
6704 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6707 panic("Failed to allocate %ld bytes for node %d memory map\n",
6708 size, pgdat->node_id);
6709 pgdat->node_mem_map = map + offset;
6711 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6712 __func__, pgdat->node_id, (unsigned long)pgdat,
6713 (unsigned long)pgdat->node_mem_map);
6714 #ifndef CONFIG_NEED_MULTIPLE_NODES
6716 * With no DISCONTIG, the global mem_map is just set as node 0's
6718 if (pgdat == NODE_DATA(0)) {
6719 mem_map = NODE_DATA(0)->node_mem_map;
6720 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6721 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6723 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6728 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6729 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6731 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6732 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6734 pgdat->first_deferred_pfn = ULONG_MAX;
6737 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6740 void __init free_area_init_node(int nid, unsigned long *zones_size,
6741 unsigned long node_start_pfn,
6742 unsigned long *zholes_size)
6744 pg_data_t *pgdat = NODE_DATA(nid);
6745 unsigned long start_pfn = 0;
6746 unsigned long end_pfn = 0;
6748 /* pg_data_t should be reset to zero when it's allocated */
6749 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6751 pgdat->node_id = nid;
6752 pgdat->node_start_pfn = node_start_pfn;
6753 pgdat->per_cpu_nodestats = NULL;
6754 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6755 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6756 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6757 (u64)start_pfn << PAGE_SHIFT,
6758 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6760 start_pfn = node_start_pfn;
6762 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6763 zones_size, zholes_size);
6765 alloc_node_mem_map(pgdat);
6766 pgdat_set_deferred_range(pgdat);
6768 free_area_init_core(pgdat);
6771 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6773 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6776 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6781 for (pfn = spfn; pfn < epfn; pfn++) {
6782 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6783 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6784 + pageblock_nr_pages - 1;
6787 mm_zero_struct_page(pfn_to_page(pfn));
6795 * Only struct pages that are backed by physical memory are zeroed and
6796 * initialized by going through __init_single_page(). But, there are some
6797 * struct pages which are reserved in memblock allocator and their fields
6798 * may be accessed (for example page_to_pfn() on some configuration accesses
6799 * flags). We must explicitly zero those struct pages.
6801 * This function also addresses a similar issue where struct pages are left
6802 * uninitialized because the physical address range is not covered by
6803 * memblock.memory or memblock.reserved. That could happen when memblock
6804 * layout is manually configured via memmap=.
6806 void __init zero_resv_unavail(void)
6808 phys_addr_t start, end;
6810 phys_addr_t next = 0;
6813 * Loop through unavailable ranges not covered by memblock.memory.
6816 for_each_mem_range(i, &memblock.memory, NULL,
6817 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6819 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6822 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6825 * Struct pages that do not have backing memory. This could be because
6826 * firmware is using some of this memory, or for some other reasons.
6829 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6831 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6833 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6835 #if MAX_NUMNODES > 1
6837 * Figure out the number of possible node ids.
6839 void __init setup_nr_node_ids(void)
6841 unsigned int highest;
6843 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6844 nr_node_ids = highest + 1;
6849 * node_map_pfn_alignment - determine the maximum internode alignment
6851 * This function should be called after node map is populated and sorted.
6852 * It calculates the maximum power of two alignment which can distinguish
6855 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6856 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6857 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6858 * shifted, 1GiB is enough and this function will indicate so.
6860 * This is used to test whether pfn -> nid mapping of the chosen memory
6861 * model has fine enough granularity to avoid incorrect mapping for the
6862 * populated node map.
6864 * Return: the determined alignment in pfn's. 0 if there is no alignment
6865 * requirement (single node).
6867 unsigned long __init node_map_pfn_alignment(void)
6869 unsigned long accl_mask = 0, last_end = 0;
6870 unsigned long start, end, mask;
6871 int last_nid = NUMA_NO_NODE;
6874 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6875 if (!start || last_nid < 0 || last_nid == nid) {
6882 * Start with a mask granular enough to pin-point to the
6883 * start pfn and tick off bits one-by-one until it becomes
6884 * too coarse to separate the current node from the last.
6886 mask = ~((1 << __ffs(start)) - 1);
6887 while (mask && last_end <= (start & (mask << 1)))
6890 /* accumulate all internode masks */
6894 /* convert mask to number of pages */
6895 return ~accl_mask + 1;
6898 /* Find the lowest pfn for a node */
6899 static unsigned long __init find_min_pfn_for_node(int nid)
6901 unsigned long min_pfn = ULONG_MAX;
6902 unsigned long start_pfn;
6905 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6906 min_pfn = min(min_pfn, start_pfn);
6908 if (min_pfn == ULONG_MAX) {
6909 pr_warn("Could not find start_pfn for node %d\n", nid);
6917 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6919 * Return: the minimum PFN based on information provided via
6920 * memblock_set_node().
6922 unsigned long __init find_min_pfn_with_active_regions(void)
6924 return find_min_pfn_for_node(MAX_NUMNODES);
6928 * early_calculate_totalpages()
6929 * Sum pages in active regions for movable zone.
6930 * Populate N_MEMORY for calculating usable_nodes.
6932 static unsigned long __init early_calculate_totalpages(void)
6934 unsigned long totalpages = 0;
6935 unsigned long start_pfn, end_pfn;
6938 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6939 unsigned long pages = end_pfn - start_pfn;
6941 totalpages += pages;
6943 node_set_state(nid, N_MEMORY);
6949 * Find the PFN the Movable zone begins in each node. Kernel memory
6950 * is spread evenly between nodes as long as the nodes have enough
6951 * memory. When they don't, some nodes will have more kernelcore than
6954 static void __init find_zone_movable_pfns_for_nodes(void)
6957 unsigned long usable_startpfn;
6958 unsigned long kernelcore_node, kernelcore_remaining;
6959 /* save the state before borrow the nodemask */
6960 nodemask_t saved_node_state = node_states[N_MEMORY];
6961 unsigned long totalpages = early_calculate_totalpages();
6962 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6963 struct memblock_region *r;
6965 /* Need to find movable_zone earlier when movable_node is specified. */
6966 find_usable_zone_for_movable();
6969 * If movable_node is specified, ignore kernelcore and movablecore
6972 if (movable_node_is_enabled()) {
6973 for_each_memblock(memory, r) {
6974 if (!memblock_is_hotpluggable(r))
6979 usable_startpfn = PFN_DOWN(r->base);
6980 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6981 min(usable_startpfn, zone_movable_pfn[nid]) :
6989 * If kernelcore=mirror is specified, ignore movablecore option
6991 if (mirrored_kernelcore) {
6992 bool mem_below_4gb_not_mirrored = false;
6994 for_each_memblock(memory, r) {
6995 if (memblock_is_mirror(r))
7000 usable_startpfn = memblock_region_memory_base_pfn(r);
7002 if (usable_startpfn < 0x100000) {
7003 mem_below_4gb_not_mirrored = true;
7007 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7008 min(usable_startpfn, zone_movable_pfn[nid]) :
7012 if (mem_below_4gb_not_mirrored)
7013 pr_warn("This configuration results in unmirrored kernel memory.");
7019 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7020 * amount of necessary memory.
7022 if (required_kernelcore_percent)
7023 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7025 if (required_movablecore_percent)
7026 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7030 * If movablecore= was specified, calculate what size of
7031 * kernelcore that corresponds so that memory usable for
7032 * any allocation type is evenly spread. If both kernelcore
7033 * and movablecore are specified, then the value of kernelcore
7034 * will be used for required_kernelcore if it's greater than
7035 * what movablecore would have allowed.
7037 if (required_movablecore) {
7038 unsigned long corepages;
7041 * Round-up so that ZONE_MOVABLE is at least as large as what
7042 * was requested by the user
7044 required_movablecore =
7045 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7046 required_movablecore = min(totalpages, required_movablecore);
7047 corepages = totalpages - required_movablecore;
7049 required_kernelcore = max(required_kernelcore, corepages);
7053 * If kernelcore was not specified or kernelcore size is larger
7054 * than totalpages, there is no ZONE_MOVABLE.
7056 if (!required_kernelcore || required_kernelcore >= totalpages)
7059 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7060 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7063 /* Spread kernelcore memory as evenly as possible throughout nodes */
7064 kernelcore_node = required_kernelcore / usable_nodes;
7065 for_each_node_state(nid, N_MEMORY) {
7066 unsigned long start_pfn, end_pfn;
7069 * Recalculate kernelcore_node if the division per node
7070 * now exceeds what is necessary to satisfy the requested
7071 * amount of memory for the kernel
7073 if (required_kernelcore < kernelcore_node)
7074 kernelcore_node = required_kernelcore / usable_nodes;
7077 * As the map is walked, we track how much memory is usable
7078 * by the kernel using kernelcore_remaining. When it is
7079 * 0, the rest of the node is usable by ZONE_MOVABLE
7081 kernelcore_remaining = kernelcore_node;
7083 /* Go through each range of PFNs within this node */
7084 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7085 unsigned long size_pages;
7087 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7088 if (start_pfn >= end_pfn)
7091 /* Account for what is only usable for kernelcore */
7092 if (start_pfn < usable_startpfn) {
7093 unsigned long kernel_pages;
7094 kernel_pages = min(end_pfn, usable_startpfn)
7097 kernelcore_remaining -= min(kernel_pages,
7098 kernelcore_remaining);
7099 required_kernelcore -= min(kernel_pages,
7100 required_kernelcore);
7102 /* Continue if range is now fully accounted */
7103 if (end_pfn <= usable_startpfn) {
7106 * Push zone_movable_pfn to the end so
7107 * that if we have to rebalance
7108 * kernelcore across nodes, we will
7109 * not double account here
7111 zone_movable_pfn[nid] = end_pfn;
7114 start_pfn = usable_startpfn;
7118 * The usable PFN range for ZONE_MOVABLE is from
7119 * start_pfn->end_pfn. Calculate size_pages as the
7120 * number of pages used as kernelcore
7122 size_pages = end_pfn - start_pfn;
7123 if (size_pages > kernelcore_remaining)
7124 size_pages = kernelcore_remaining;
7125 zone_movable_pfn[nid] = start_pfn + size_pages;
7128 * Some kernelcore has been met, update counts and
7129 * break if the kernelcore for this node has been
7132 required_kernelcore -= min(required_kernelcore,
7134 kernelcore_remaining -= size_pages;
7135 if (!kernelcore_remaining)
7141 * If there is still required_kernelcore, we do another pass with one
7142 * less node in the count. This will push zone_movable_pfn[nid] further
7143 * along on the nodes that still have memory until kernelcore is
7147 if (usable_nodes && required_kernelcore > usable_nodes)
7151 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7152 for (nid = 0; nid < MAX_NUMNODES; nid++)
7153 zone_movable_pfn[nid] =
7154 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7157 /* restore the node_state */
7158 node_states[N_MEMORY] = saved_node_state;
7161 /* Any regular or high memory on that node ? */
7162 static void check_for_memory(pg_data_t *pgdat, int nid)
7164 enum zone_type zone_type;
7166 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7167 struct zone *zone = &pgdat->node_zones[zone_type];
7168 if (populated_zone(zone)) {
7169 if (IS_ENABLED(CONFIG_HIGHMEM))
7170 node_set_state(nid, N_HIGH_MEMORY);
7171 if (zone_type <= ZONE_NORMAL)
7172 node_set_state(nid, N_NORMAL_MEMORY);
7179 * free_area_init_nodes - Initialise all pg_data_t and zone data
7180 * @max_zone_pfn: an array of max PFNs for each zone
7182 * This will call free_area_init_node() for each active node in the system.
7183 * Using the page ranges provided by memblock_set_node(), the size of each
7184 * zone in each node and their holes is calculated. If the maximum PFN
7185 * between two adjacent zones match, it is assumed that the zone is empty.
7186 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7187 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7188 * starts where the previous one ended. For example, ZONE_DMA32 starts
7189 * at arch_max_dma_pfn.
7191 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7193 unsigned long start_pfn, end_pfn;
7196 /* Record where the zone boundaries are */
7197 memset(arch_zone_lowest_possible_pfn, 0,
7198 sizeof(arch_zone_lowest_possible_pfn));
7199 memset(arch_zone_highest_possible_pfn, 0,
7200 sizeof(arch_zone_highest_possible_pfn));
7202 start_pfn = find_min_pfn_with_active_regions();
7204 for (i = 0; i < MAX_NR_ZONES; i++) {
7205 if (i == ZONE_MOVABLE)
7208 end_pfn = max(max_zone_pfn[i], start_pfn);
7209 arch_zone_lowest_possible_pfn[i] = start_pfn;
7210 arch_zone_highest_possible_pfn[i] = end_pfn;
7212 start_pfn = end_pfn;
7215 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7216 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7217 find_zone_movable_pfns_for_nodes();
7219 /* Print out the zone ranges */
7220 pr_info("Zone ranges:\n");
7221 for (i = 0; i < MAX_NR_ZONES; i++) {
7222 if (i == ZONE_MOVABLE)
7224 pr_info(" %-8s ", zone_names[i]);
7225 if (arch_zone_lowest_possible_pfn[i] ==
7226 arch_zone_highest_possible_pfn[i])
7229 pr_cont("[mem %#018Lx-%#018Lx]\n",
7230 (u64)arch_zone_lowest_possible_pfn[i]
7232 ((u64)arch_zone_highest_possible_pfn[i]
7233 << PAGE_SHIFT) - 1);
7236 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7237 pr_info("Movable zone start for each node\n");
7238 for (i = 0; i < MAX_NUMNODES; i++) {
7239 if (zone_movable_pfn[i])
7240 pr_info(" Node %d: %#018Lx\n", i,
7241 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7244 /* Print out the early node map */
7245 pr_info("Early memory node ranges\n");
7246 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
7247 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7248 (u64)start_pfn << PAGE_SHIFT,
7249 ((u64)end_pfn << PAGE_SHIFT) - 1);
7251 /* Initialise every node */
7252 mminit_verify_pageflags_layout();
7253 setup_nr_node_ids();
7254 zero_resv_unavail();
7255 for_each_online_node(nid) {
7256 pg_data_t *pgdat = NODE_DATA(nid);
7257 free_area_init_node(nid, NULL,
7258 find_min_pfn_for_node(nid), NULL);
7260 /* Any memory on that node */
7261 if (pgdat->node_present_pages)
7262 node_set_state(nid, N_MEMORY);
7263 check_for_memory(pgdat, nid);
7267 static int __init cmdline_parse_core(char *p, unsigned long *core,
7268 unsigned long *percent)
7270 unsigned long long coremem;
7276 /* Value may be a percentage of total memory, otherwise bytes */
7277 coremem = simple_strtoull(p, &endptr, 0);
7278 if (*endptr == '%') {
7279 /* Paranoid check for percent values greater than 100 */
7280 WARN_ON(coremem > 100);
7284 coremem = memparse(p, &p);
7285 /* Paranoid check that UL is enough for the coremem value */
7286 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7288 *core = coremem >> PAGE_SHIFT;
7295 * kernelcore=size sets the amount of memory for use for allocations that
7296 * cannot be reclaimed or migrated.
7298 static int __init cmdline_parse_kernelcore(char *p)
7300 /* parse kernelcore=mirror */
7301 if (parse_option_str(p, "mirror")) {
7302 mirrored_kernelcore = true;
7306 return cmdline_parse_core(p, &required_kernelcore,
7307 &required_kernelcore_percent);
7311 * movablecore=size sets the amount of memory for use for allocations that
7312 * can be reclaimed or migrated.
7314 static int __init cmdline_parse_movablecore(char *p)
7316 return cmdline_parse_core(p, &required_movablecore,
7317 &required_movablecore_percent);
7320 early_param("kernelcore", cmdline_parse_kernelcore);
7321 early_param("movablecore", cmdline_parse_movablecore);
7323 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7325 void adjust_managed_page_count(struct page *page, long count)
7327 atomic_long_add(count, &page_zone(page)->managed_pages);
7328 totalram_pages_add(count);
7329 #ifdef CONFIG_HIGHMEM
7330 if (PageHighMem(page))
7331 totalhigh_pages_add(count);
7334 EXPORT_SYMBOL(adjust_managed_page_count);
7336 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7339 unsigned long pages = 0;
7341 start = (void *)PAGE_ALIGN((unsigned long)start);
7342 end = (void *)((unsigned long)end & PAGE_MASK);
7343 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7344 struct page *page = virt_to_page(pos);
7345 void *direct_map_addr;
7348 * 'direct_map_addr' might be different from 'pos'
7349 * because some architectures' virt_to_page()
7350 * work with aliases. Getting the direct map
7351 * address ensures that we get a _writeable_
7352 * alias for the memset().
7354 direct_map_addr = page_address(page);
7355 if ((unsigned int)poison <= 0xFF)
7356 memset(direct_map_addr, poison, PAGE_SIZE);
7358 free_reserved_page(page);
7362 pr_info("Freeing %s memory: %ldK\n",
7363 s, pages << (PAGE_SHIFT - 10));
7368 #ifdef CONFIG_HIGHMEM
7369 void free_highmem_page(struct page *page)
7371 __free_reserved_page(page);
7372 totalram_pages_inc();
7373 atomic_long_inc(&page_zone(page)->managed_pages);
7374 totalhigh_pages_inc();
7379 void __init mem_init_print_info(const char *str)
7381 unsigned long physpages, codesize, datasize, rosize, bss_size;
7382 unsigned long init_code_size, init_data_size;
7384 physpages = get_num_physpages();
7385 codesize = _etext - _stext;
7386 datasize = _edata - _sdata;
7387 rosize = __end_rodata - __start_rodata;
7388 bss_size = __bss_stop - __bss_start;
7389 init_data_size = __init_end - __init_begin;
7390 init_code_size = _einittext - _sinittext;
7393 * Detect special cases and adjust section sizes accordingly:
7394 * 1) .init.* may be embedded into .data sections
7395 * 2) .init.text.* may be out of [__init_begin, __init_end],
7396 * please refer to arch/tile/kernel/vmlinux.lds.S.
7397 * 3) .rodata.* may be embedded into .text or .data sections.
7399 #define adj_init_size(start, end, size, pos, adj) \
7401 if (start <= pos && pos < end && size > adj) \
7405 adj_init_size(__init_begin, __init_end, init_data_size,
7406 _sinittext, init_code_size);
7407 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7408 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7409 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7410 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7412 #undef adj_init_size
7414 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7415 #ifdef CONFIG_HIGHMEM
7419 nr_free_pages() << (PAGE_SHIFT - 10),
7420 physpages << (PAGE_SHIFT - 10),
7421 codesize >> 10, datasize >> 10, rosize >> 10,
7422 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7423 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7424 totalcma_pages << (PAGE_SHIFT - 10),
7425 #ifdef CONFIG_HIGHMEM
7426 totalhigh_pages() << (PAGE_SHIFT - 10),
7428 str ? ", " : "", str ? str : "");
7432 * set_dma_reserve - set the specified number of pages reserved in the first zone
7433 * @new_dma_reserve: The number of pages to mark reserved
7435 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7436 * In the DMA zone, a significant percentage may be consumed by kernel image
7437 * and other unfreeable allocations which can skew the watermarks badly. This
7438 * function may optionally be used to account for unfreeable pages in the
7439 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7440 * smaller per-cpu batchsize.
7442 void __init set_dma_reserve(unsigned long new_dma_reserve)
7444 dma_reserve = new_dma_reserve;
7447 void __init free_area_init(unsigned long *zones_size)
7449 zero_resv_unavail();
7450 free_area_init_node(0, zones_size,
7451 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7454 static int page_alloc_cpu_dead(unsigned int cpu)
7457 lru_add_drain_cpu(cpu);
7461 * Spill the event counters of the dead processor
7462 * into the current processors event counters.
7463 * This artificially elevates the count of the current
7466 vm_events_fold_cpu(cpu);
7469 * Zero the differential counters of the dead processor
7470 * so that the vm statistics are consistent.
7472 * This is only okay since the processor is dead and cannot
7473 * race with what we are doing.
7475 cpu_vm_stats_fold(cpu);
7479 void __init page_alloc_init(void)
7483 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7484 "mm/page_alloc:dead", NULL,
7485 page_alloc_cpu_dead);
7490 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7491 * or min_free_kbytes changes.
7493 static void calculate_totalreserve_pages(void)
7495 struct pglist_data *pgdat;
7496 unsigned long reserve_pages = 0;
7497 enum zone_type i, j;
7499 for_each_online_pgdat(pgdat) {
7501 pgdat->totalreserve_pages = 0;
7503 for (i = 0; i < MAX_NR_ZONES; i++) {
7504 struct zone *zone = pgdat->node_zones + i;
7506 unsigned long managed_pages = zone_managed_pages(zone);
7508 /* Find valid and maximum lowmem_reserve in the zone */
7509 for (j = i; j < MAX_NR_ZONES; j++) {
7510 if (zone->lowmem_reserve[j] > max)
7511 max = zone->lowmem_reserve[j];
7514 /* we treat the high watermark as reserved pages. */
7515 max += high_wmark_pages(zone);
7517 if (max > managed_pages)
7518 max = managed_pages;
7520 pgdat->totalreserve_pages += max;
7522 reserve_pages += max;
7525 totalreserve_pages = reserve_pages;
7529 * setup_per_zone_lowmem_reserve - called whenever
7530 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7531 * has a correct pages reserved value, so an adequate number of
7532 * pages are left in the zone after a successful __alloc_pages().
7534 static void setup_per_zone_lowmem_reserve(void)
7536 struct pglist_data *pgdat;
7537 enum zone_type j, idx;
7539 for_each_online_pgdat(pgdat) {
7540 for (j = 0; j < MAX_NR_ZONES; j++) {
7541 struct zone *zone = pgdat->node_zones + j;
7542 unsigned long managed_pages = zone_managed_pages(zone);
7544 zone->lowmem_reserve[j] = 0;
7548 struct zone *lower_zone;
7551 lower_zone = pgdat->node_zones + idx;
7553 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7554 sysctl_lowmem_reserve_ratio[idx] = 0;
7555 lower_zone->lowmem_reserve[j] = 0;
7557 lower_zone->lowmem_reserve[j] =
7558 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7560 managed_pages += zone_managed_pages(lower_zone);
7565 /* update totalreserve_pages */
7566 calculate_totalreserve_pages();
7569 static void __setup_per_zone_wmarks(void)
7571 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7572 unsigned long lowmem_pages = 0;
7574 unsigned long flags;
7576 /* Calculate total number of !ZONE_HIGHMEM pages */
7577 for_each_zone(zone) {
7578 if (!is_highmem(zone))
7579 lowmem_pages += zone_managed_pages(zone);
7582 for_each_zone(zone) {
7585 spin_lock_irqsave(&zone->lock, flags);
7586 tmp = (u64)pages_min * zone_managed_pages(zone);
7587 do_div(tmp, lowmem_pages);
7588 if (is_highmem(zone)) {
7590 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7591 * need highmem pages, so cap pages_min to a small
7594 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7595 * deltas control async page reclaim, and so should
7596 * not be capped for highmem.
7598 unsigned long min_pages;
7600 min_pages = zone_managed_pages(zone) / 1024;
7601 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7602 zone->_watermark[WMARK_MIN] = min_pages;
7605 * If it's a lowmem zone, reserve a number of pages
7606 * proportionate to the zone's size.
7608 zone->_watermark[WMARK_MIN] = tmp;
7612 * Set the kswapd watermarks distance according to the
7613 * scale factor in proportion to available memory, but
7614 * ensure a minimum size on small systems.
7616 tmp = max_t(u64, tmp >> 2,
7617 mult_frac(zone_managed_pages(zone),
7618 watermark_scale_factor, 10000));
7620 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7621 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7622 zone->watermark_boost = 0;
7624 spin_unlock_irqrestore(&zone->lock, flags);
7627 /* update totalreserve_pages */
7628 calculate_totalreserve_pages();
7632 * setup_per_zone_wmarks - called when min_free_kbytes changes
7633 * or when memory is hot-{added|removed}
7635 * Ensures that the watermark[min,low,high] values for each zone are set
7636 * correctly with respect to min_free_kbytes.
7638 void setup_per_zone_wmarks(void)
7640 static DEFINE_SPINLOCK(lock);
7643 __setup_per_zone_wmarks();
7648 * Initialise min_free_kbytes.
7650 * For small machines we want it small (128k min). For large machines
7651 * we want it large (64MB max). But it is not linear, because network
7652 * bandwidth does not increase linearly with machine size. We use
7654 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7655 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7671 int __meminit init_per_zone_wmark_min(void)
7673 unsigned long lowmem_kbytes;
7674 int new_min_free_kbytes;
7676 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7677 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7679 if (new_min_free_kbytes > user_min_free_kbytes) {
7680 min_free_kbytes = new_min_free_kbytes;
7681 if (min_free_kbytes < 128)
7682 min_free_kbytes = 128;
7683 if (min_free_kbytes > 65536)
7684 min_free_kbytes = 65536;
7686 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7687 new_min_free_kbytes, user_min_free_kbytes);
7689 setup_per_zone_wmarks();
7690 refresh_zone_stat_thresholds();
7691 setup_per_zone_lowmem_reserve();
7694 setup_min_unmapped_ratio();
7695 setup_min_slab_ratio();
7700 core_initcall(init_per_zone_wmark_min)
7703 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7704 * that we can call two helper functions whenever min_free_kbytes
7707 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7708 void __user *buffer, size_t *length, loff_t *ppos)
7712 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7717 user_min_free_kbytes = min_free_kbytes;
7718 setup_per_zone_wmarks();
7723 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7724 void __user *buffer, size_t *length, loff_t *ppos)
7728 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7735 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7736 void __user *buffer, size_t *length, loff_t *ppos)
7740 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7745 setup_per_zone_wmarks();
7751 static void setup_min_unmapped_ratio(void)
7756 for_each_online_pgdat(pgdat)
7757 pgdat->min_unmapped_pages = 0;
7760 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7761 sysctl_min_unmapped_ratio) / 100;
7765 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7766 void __user *buffer, size_t *length, loff_t *ppos)
7770 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7774 setup_min_unmapped_ratio();
7779 static void setup_min_slab_ratio(void)
7784 for_each_online_pgdat(pgdat)
7785 pgdat->min_slab_pages = 0;
7788 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7789 sysctl_min_slab_ratio) / 100;
7792 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7793 void __user *buffer, size_t *length, loff_t *ppos)
7797 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7801 setup_min_slab_ratio();
7808 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7809 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7810 * whenever sysctl_lowmem_reserve_ratio changes.
7812 * The reserve ratio obviously has absolutely no relation with the
7813 * minimum watermarks. The lowmem reserve ratio can only make sense
7814 * if in function of the boot time zone sizes.
7816 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7817 void __user *buffer, size_t *length, loff_t *ppos)
7819 proc_dointvec_minmax(table, write, buffer, length, ppos);
7820 setup_per_zone_lowmem_reserve();
7825 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7826 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7827 * pagelist can have before it gets flushed back to buddy allocator.
7829 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7830 void __user *buffer, size_t *length, loff_t *ppos)
7833 int old_percpu_pagelist_fraction;
7836 mutex_lock(&pcp_batch_high_lock);
7837 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7839 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7840 if (!write || ret < 0)
7843 /* Sanity checking to avoid pcp imbalance */
7844 if (percpu_pagelist_fraction &&
7845 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7846 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7852 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7855 for_each_populated_zone(zone) {
7858 for_each_possible_cpu(cpu)
7859 pageset_set_high_and_batch(zone,
7860 per_cpu_ptr(zone->pageset, cpu));
7863 mutex_unlock(&pcp_batch_high_lock);
7868 int hashdist = HASHDIST_DEFAULT;
7870 static int __init set_hashdist(char *str)
7874 hashdist = simple_strtoul(str, &str, 0);
7877 __setup("hashdist=", set_hashdist);
7880 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7882 * Returns the number of pages that arch has reserved but
7883 * is not known to alloc_large_system_hash().
7885 static unsigned long __init arch_reserved_kernel_pages(void)
7892 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7893 * machines. As memory size is increased the scale is also increased but at
7894 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7895 * quadruples the scale is increased by one, which means the size of hash table
7896 * only doubles, instead of quadrupling as well.
7897 * Because 32-bit systems cannot have large physical memory, where this scaling
7898 * makes sense, it is disabled on such platforms.
7900 #if __BITS_PER_LONG > 32
7901 #define ADAPT_SCALE_BASE (64ul << 30)
7902 #define ADAPT_SCALE_SHIFT 2
7903 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7907 * allocate a large system hash table from bootmem
7908 * - it is assumed that the hash table must contain an exact power-of-2
7909 * quantity of entries
7910 * - limit is the number of hash buckets, not the total allocation size
7912 void *__init alloc_large_system_hash(const char *tablename,
7913 unsigned long bucketsize,
7914 unsigned long numentries,
7917 unsigned int *_hash_shift,
7918 unsigned int *_hash_mask,
7919 unsigned long low_limit,
7920 unsigned long high_limit)
7922 unsigned long long max = high_limit;
7923 unsigned long log2qty, size;
7927 /* allow the kernel cmdline to have a say */
7929 /* round applicable memory size up to nearest megabyte */
7930 numentries = nr_kernel_pages;
7931 numentries -= arch_reserved_kernel_pages();
7933 /* It isn't necessary when PAGE_SIZE >= 1MB */
7934 if (PAGE_SHIFT < 20)
7935 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7937 #if __BITS_PER_LONG > 32
7939 unsigned long adapt;
7941 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7942 adapt <<= ADAPT_SCALE_SHIFT)
7947 /* limit to 1 bucket per 2^scale bytes of low memory */
7948 if (scale > PAGE_SHIFT)
7949 numentries >>= (scale - PAGE_SHIFT);
7951 numentries <<= (PAGE_SHIFT - scale);
7953 /* Make sure we've got at least a 0-order allocation.. */
7954 if (unlikely(flags & HASH_SMALL)) {
7955 /* Makes no sense without HASH_EARLY */
7956 WARN_ON(!(flags & HASH_EARLY));
7957 if (!(numentries >> *_hash_shift)) {
7958 numentries = 1UL << *_hash_shift;
7959 BUG_ON(!numentries);
7961 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7962 numentries = PAGE_SIZE / bucketsize;
7964 numentries = roundup_pow_of_two(numentries);
7966 /* limit allocation size to 1/16 total memory by default */
7968 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7969 do_div(max, bucketsize);
7971 max = min(max, 0x80000000ULL);
7973 if (numentries < low_limit)
7974 numentries = low_limit;
7975 if (numentries > max)
7978 log2qty = ilog2(numentries);
7980 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7982 size = bucketsize << log2qty;
7983 if (flags & HASH_EARLY) {
7984 if (flags & HASH_ZERO)
7985 table = memblock_alloc(size, SMP_CACHE_BYTES);
7987 table = memblock_alloc_raw(size,
7989 } else if (hashdist) {
7990 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7993 * If bucketsize is not a power-of-two, we may free
7994 * some pages at the end of hash table which
7995 * alloc_pages_exact() automatically does
7997 if (get_order(size) < MAX_ORDER) {
7998 table = alloc_pages_exact(size, gfp_flags);
7999 kmemleak_alloc(table, size, 1, gfp_flags);
8002 } while (!table && size > PAGE_SIZE && --log2qty);
8005 panic("Failed to allocate %s hash table\n", tablename);
8007 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
8008 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
8011 *_hash_shift = log2qty;
8013 *_hash_mask = (1 << log2qty) - 1;
8019 * This function checks whether pageblock includes unmovable pages or not.
8020 * If @count is not zero, it is okay to include less @count unmovable pages
8022 * PageLRU check without isolation or lru_lock could race so that
8023 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8024 * check without lock_page also may miss some movable non-lru pages at
8025 * race condition. So you can't expect this function should be exact.
8027 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
8028 int migratetype, int flags)
8030 unsigned long found;
8031 unsigned long iter = 0;
8032 unsigned long pfn = page_to_pfn(page);
8033 const char *reason = "unmovable page";
8036 * TODO we could make this much more efficient by not checking every
8037 * page in the range if we know all of them are in MOVABLE_ZONE and
8038 * that the movable zone guarantees that pages are migratable but
8039 * the later is not the case right now unfortunatelly. E.g. movablecore
8040 * can still lead to having bootmem allocations in zone_movable.
8043 if (is_migrate_cma_page(page)) {
8045 * CMA allocations (alloc_contig_range) really need to mark
8046 * isolate CMA pageblocks even when they are not movable in fact
8047 * so consider them movable here.
8049 if (is_migrate_cma(migratetype))
8052 reason = "CMA page";
8056 for (found = 0; iter < pageblock_nr_pages; iter++) {
8057 unsigned long check = pfn + iter;
8059 if (!pfn_valid_within(check))
8062 page = pfn_to_page(check);
8064 if (PageReserved(page))
8068 * If the zone is movable and we have ruled out all reserved
8069 * pages then it should be reasonably safe to assume the rest
8072 if (zone_idx(zone) == ZONE_MOVABLE)
8076 * Hugepages are not in LRU lists, but they're movable.
8077 * We need not scan over tail pages because we don't
8078 * handle each tail page individually in migration.
8080 if (PageHuge(page)) {
8081 struct page *head = compound_head(page);
8082 unsigned int skip_pages;
8084 if (!hugepage_migration_supported(page_hstate(head)))
8087 skip_pages = (1 << compound_order(head)) - (page - head);
8088 iter += skip_pages - 1;
8093 * We can't use page_count without pin a page
8094 * because another CPU can free compound page.
8095 * This check already skips compound tails of THP
8096 * because their page->_refcount is zero at all time.
8098 if (!page_ref_count(page)) {
8099 if (PageBuddy(page))
8100 iter += (1 << page_order(page)) - 1;
8105 * The HWPoisoned page may be not in buddy system, and
8106 * page_count() is not 0.
8108 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
8111 if (__PageMovable(page))
8117 * If there are RECLAIMABLE pages, we need to check
8118 * it. But now, memory offline itself doesn't call
8119 * shrink_node_slabs() and it still to be fixed.
8122 * If the page is not RAM, page_count()should be 0.
8123 * we don't need more check. This is an _used_ not-movable page.
8125 * The problematic thing here is PG_reserved pages. PG_reserved
8126 * is set to both of a memory hole page and a _used_ kernel
8134 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8135 if (flags & REPORT_FAILURE)
8136 dump_page(pfn_to_page(pfn + iter), reason);
8140 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
8142 static unsigned long pfn_max_align_down(unsigned long pfn)
8144 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8145 pageblock_nr_pages) - 1);
8148 static unsigned long pfn_max_align_up(unsigned long pfn)
8150 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8151 pageblock_nr_pages));
8154 /* [start, end) must belong to a single zone. */
8155 static int __alloc_contig_migrate_range(struct compact_control *cc,
8156 unsigned long start, unsigned long end)
8158 /* This function is based on compact_zone() from compaction.c. */
8159 unsigned long nr_reclaimed;
8160 unsigned long pfn = start;
8161 unsigned int tries = 0;
8166 while (pfn < end || !list_empty(&cc->migratepages)) {
8167 if (fatal_signal_pending(current)) {
8172 if (list_empty(&cc->migratepages)) {
8173 cc->nr_migratepages = 0;
8174 pfn = isolate_migratepages_range(cc, pfn, end);
8180 } else if (++tries == 5) {
8181 ret = ret < 0 ? ret : -EBUSY;
8185 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8187 cc->nr_migratepages -= nr_reclaimed;
8189 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8190 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8193 putback_movable_pages(&cc->migratepages);
8200 * alloc_contig_range() -- tries to allocate given range of pages
8201 * @start: start PFN to allocate
8202 * @end: one-past-the-last PFN to allocate
8203 * @migratetype: migratetype of the underlaying pageblocks (either
8204 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8205 * in range must have the same migratetype and it must
8206 * be either of the two.
8207 * @gfp_mask: GFP mask to use during compaction
8209 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8210 * aligned. The PFN range must belong to a single zone.
8212 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8213 * pageblocks in the range. Once isolated, the pageblocks should not
8214 * be modified by others.
8216 * Return: zero on success or negative error code. On success all
8217 * pages which PFN is in [start, end) are allocated for the caller and
8218 * need to be freed with free_contig_range().
8220 int alloc_contig_range(unsigned long start, unsigned long end,
8221 unsigned migratetype, gfp_t gfp_mask)
8223 unsigned long outer_start, outer_end;
8227 struct compact_control cc = {
8228 .nr_migratepages = 0,
8230 .zone = page_zone(pfn_to_page(start)),
8231 .mode = MIGRATE_SYNC,
8232 .ignore_skip_hint = true,
8233 .no_set_skip_hint = true,
8234 .gfp_mask = current_gfp_context(gfp_mask),
8236 INIT_LIST_HEAD(&cc.migratepages);
8239 * What we do here is we mark all pageblocks in range as
8240 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8241 * have different sizes, and due to the way page allocator
8242 * work, we align the range to biggest of the two pages so
8243 * that page allocator won't try to merge buddies from
8244 * different pageblocks and change MIGRATE_ISOLATE to some
8245 * other migration type.
8247 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8248 * migrate the pages from an unaligned range (ie. pages that
8249 * we are interested in). This will put all the pages in
8250 * range back to page allocator as MIGRATE_ISOLATE.
8252 * When this is done, we take the pages in range from page
8253 * allocator removing them from the buddy system. This way
8254 * page allocator will never consider using them.
8256 * This lets us mark the pageblocks back as
8257 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8258 * aligned range but not in the unaligned, original range are
8259 * put back to page allocator so that buddy can use them.
8262 ret = start_isolate_page_range(pfn_max_align_down(start),
8263 pfn_max_align_up(end), migratetype, 0);
8268 * In case of -EBUSY, we'd like to know which page causes problem.
8269 * So, just fall through. test_pages_isolated() has a tracepoint
8270 * which will report the busy page.
8272 * It is possible that busy pages could become available before
8273 * the call to test_pages_isolated, and the range will actually be
8274 * allocated. So, if we fall through be sure to clear ret so that
8275 * -EBUSY is not accidentally used or returned to caller.
8277 ret = __alloc_contig_migrate_range(&cc, start, end);
8278 if (ret && ret != -EBUSY)
8283 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8284 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8285 * more, all pages in [start, end) are free in page allocator.
8286 * What we are going to do is to allocate all pages from
8287 * [start, end) (that is remove them from page allocator).
8289 * The only problem is that pages at the beginning and at the
8290 * end of interesting range may be not aligned with pages that
8291 * page allocator holds, ie. they can be part of higher order
8292 * pages. Because of this, we reserve the bigger range and
8293 * once this is done free the pages we are not interested in.
8295 * We don't have to hold zone->lock here because the pages are
8296 * isolated thus they won't get removed from buddy.
8299 lru_add_drain_all();
8302 outer_start = start;
8303 while (!PageBuddy(pfn_to_page(outer_start))) {
8304 if (++order >= MAX_ORDER) {
8305 outer_start = start;
8308 outer_start &= ~0UL << order;
8311 if (outer_start != start) {
8312 order = page_order(pfn_to_page(outer_start));
8315 * outer_start page could be small order buddy page and
8316 * it doesn't include start page. Adjust outer_start
8317 * in this case to report failed page properly
8318 * on tracepoint in test_pages_isolated()
8320 if (outer_start + (1UL << order) <= start)
8321 outer_start = start;
8324 /* Make sure the range is really isolated. */
8325 if (test_pages_isolated(outer_start, end, false)) {
8326 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8327 __func__, outer_start, end);
8332 /* Grab isolated pages from freelists. */
8333 outer_end = isolate_freepages_range(&cc, outer_start, end);
8339 /* Free head and tail (if any) */
8340 if (start != outer_start)
8341 free_contig_range(outer_start, start - outer_start);
8342 if (end != outer_end)
8343 free_contig_range(end, outer_end - end);
8346 undo_isolate_page_range(pfn_max_align_down(start),
8347 pfn_max_align_up(end), migratetype);
8351 void free_contig_range(unsigned long pfn, unsigned nr_pages)
8353 unsigned int count = 0;
8355 for (; nr_pages--; pfn++) {
8356 struct page *page = pfn_to_page(pfn);
8358 count += page_count(page) != 1;
8361 WARN(count != 0, "%d pages are still in use!\n", count);
8365 #ifdef CONFIG_MEMORY_HOTPLUG
8367 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8368 * page high values need to be recalulated.
8370 void __meminit zone_pcp_update(struct zone *zone)
8373 mutex_lock(&pcp_batch_high_lock);
8374 for_each_possible_cpu(cpu)
8375 pageset_set_high_and_batch(zone,
8376 per_cpu_ptr(zone->pageset, cpu));
8377 mutex_unlock(&pcp_batch_high_lock);
8381 void zone_pcp_reset(struct zone *zone)
8383 unsigned long flags;
8385 struct per_cpu_pageset *pset;
8387 /* avoid races with drain_pages() */
8388 local_irq_save(flags);
8389 if (zone->pageset != &boot_pageset) {
8390 for_each_online_cpu(cpu) {
8391 pset = per_cpu_ptr(zone->pageset, cpu);
8392 drain_zonestat(zone, pset);
8394 free_percpu(zone->pageset);
8395 zone->pageset = &boot_pageset;
8397 local_irq_restore(flags);
8400 #ifdef CONFIG_MEMORY_HOTREMOVE
8402 * All pages in the range must be in a single zone and isolated
8403 * before calling this.
8406 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8410 unsigned int order, i;
8412 unsigned long flags;
8413 /* find the first valid pfn */
8414 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8419 offline_mem_sections(pfn, end_pfn);
8420 zone = page_zone(pfn_to_page(pfn));
8421 spin_lock_irqsave(&zone->lock, flags);
8423 while (pfn < end_pfn) {
8424 if (!pfn_valid(pfn)) {
8428 page = pfn_to_page(pfn);
8430 * The HWPoisoned page may be not in buddy system, and
8431 * page_count() is not 0.
8433 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8435 SetPageReserved(page);
8439 BUG_ON(page_count(page));
8440 BUG_ON(!PageBuddy(page));
8441 order = page_order(page);
8442 #ifdef CONFIG_DEBUG_VM
8443 pr_info("remove from free list %lx %d %lx\n",
8444 pfn, 1 << order, end_pfn);
8446 list_del(&page->lru);
8447 rmv_page_order(page);
8448 zone->free_area[order].nr_free--;
8449 for (i = 0; i < (1 << order); i++)
8450 SetPageReserved((page+i));
8451 pfn += (1 << order);
8453 spin_unlock_irqrestore(&zone->lock, flags);
8457 bool is_free_buddy_page(struct page *page)
8459 struct zone *zone = page_zone(page);
8460 unsigned long pfn = page_to_pfn(page);
8461 unsigned long flags;
8464 spin_lock_irqsave(&zone->lock, flags);
8465 for (order = 0; order < MAX_ORDER; order++) {
8466 struct page *page_head = page - (pfn & ((1 << order) - 1));
8468 if (PageBuddy(page_head) && page_order(page_head) >= order)
8471 spin_unlock_irqrestore(&zone->lock, flags);
8473 return order < MAX_ORDER;
8476 #ifdef CONFIG_MEMORY_FAILURE
8478 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8479 * test is performed under the zone lock to prevent a race against page
8482 bool set_hwpoison_free_buddy_page(struct page *page)
8484 struct zone *zone = page_zone(page);
8485 unsigned long pfn = page_to_pfn(page);
8486 unsigned long flags;
8488 bool hwpoisoned = false;
8490 spin_lock_irqsave(&zone->lock, flags);
8491 for (order = 0; order < MAX_ORDER; order++) {
8492 struct page *page_head = page - (pfn & ((1 << order) - 1));
8494 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8495 if (!TestSetPageHWPoison(page))
8500 spin_unlock_irqrestore(&zone->lock, flags);