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 kernel_map_pages(page, 1 << order, 0);
1148 kasan_free_nondeferred_pages(page, order);
1153 #ifdef CONFIG_DEBUG_VM
1154 static inline bool free_pcp_prepare(struct page *page)
1156 return free_pages_prepare(page, 0, true);
1159 static inline bool bulkfree_pcp_prepare(struct page *page)
1164 static bool free_pcp_prepare(struct page *page)
1166 return free_pages_prepare(page, 0, false);
1169 static bool bulkfree_pcp_prepare(struct page *page)
1171 return free_pages_check(page);
1173 #endif /* CONFIG_DEBUG_VM */
1175 static inline void prefetch_buddy(struct page *page)
1177 unsigned long pfn = page_to_pfn(page);
1178 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1179 struct page *buddy = page + (buddy_pfn - pfn);
1185 * Frees a number of pages from the PCP lists
1186 * Assumes all pages on list are in same zone, and of same order.
1187 * count is the number of pages to free.
1189 * If the zone was previously in an "all pages pinned" state then look to
1190 * see if this freeing clears that state.
1192 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1193 * pinned" detection logic.
1195 static void free_pcppages_bulk(struct zone *zone, int count,
1196 struct per_cpu_pages *pcp)
1198 int migratetype = 0;
1200 int prefetch_nr = 0;
1201 bool isolated_pageblocks;
1202 struct page *page, *tmp;
1206 struct list_head *list;
1209 * Remove pages from lists in a round-robin fashion. A
1210 * batch_free count is maintained that is incremented when an
1211 * empty list is encountered. This is so more pages are freed
1212 * off fuller lists instead of spinning excessively around empty
1217 if (++migratetype == MIGRATE_PCPTYPES)
1219 list = &pcp->lists[migratetype];
1220 } while (list_empty(list));
1222 /* This is the only non-empty list. Free them all. */
1223 if (batch_free == MIGRATE_PCPTYPES)
1227 page = list_last_entry(list, struct page, lru);
1228 /* must delete to avoid corrupting pcp list */
1229 list_del(&page->lru);
1232 if (bulkfree_pcp_prepare(page))
1235 list_add_tail(&page->lru, &head);
1238 * We are going to put the page back to the global
1239 * pool, prefetch its buddy to speed up later access
1240 * under zone->lock. It is believed the overhead of
1241 * an additional test and calculating buddy_pfn here
1242 * can be offset by reduced memory latency later. To
1243 * avoid excessive prefetching due to large count, only
1244 * prefetch buddy for the first pcp->batch nr of pages.
1246 if (prefetch_nr++ < pcp->batch)
1247 prefetch_buddy(page);
1248 } while (--count && --batch_free && !list_empty(list));
1251 spin_lock(&zone->lock);
1252 isolated_pageblocks = has_isolate_pageblock(zone);
1255 * Use safe version since after __free_one_page(),
1256 * page->lru.next will not point to original list.
1258 list_for_each_entry_safe(page, tmp, &head, lru) {
1259 int mt = get_pcppage_migratetype(page);
1260 /* MIGRATE_ISOLATE page should not go to pcplists */
1261 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1262 /* Pageblock could have been isolated meanwhile */
1263 if (unlikely(isolated_pageblocks))
1264 mt = get_pageblock_migratetype(page);
1266 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1267 trace_mm_page_pcpu_drain(page, 0, mt);
1269 spin_unlock(&zone->lock);
1272 static void free_one_page(struct zone *zone,
1273 struct page *page, unsigned long pfn,
1277 spin_lock(&zone->lock);
1278 if (unlikely(has_isolate_pageblock(zone) ||
1279 is_migrate_isolate(migratetype))) {
1280 migratetype = get_pfnblock_migratetype(page, pfn);
1282 __free_one_page(page, pfn, zone, order, migratetype);
1283 spin_unlock(&zone->lock);
1286 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1287 unsigned long zone, int nid)
1289 mm_zero_struct_page(page);
1290 set_page_links(page, zone, nid, pfn);
1291 init_page_count(page);
1292 page_mapcount_reset(page);
1293 page_cpupid_reset_last(page);
1294 page_kasan_tag_reset(page);
1296 INIT_LIST_HEAD(&page->lru);
1297 #ifdef WANT_PAGE_VIRTUAL
1298 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1299 if (!is_highmem_idx(zone))
1300 set_page_address(page, __va(pfn << PAGE_SHIFT));
1304 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1305 static void __meminit init_reserved_page(unsigned long pfn)
1310 if (!early_page_uninitialised(pfn))
1313 nid = early_pfn_to_nid(pfn);
1314 pgdat = NODE_DATA(nid);
1316 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1317 struct zone *zone = &pgdat->node_zones[zid];
1319 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1322 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1325 static inline void init_reserved_page(unsigned long pfn)
1328 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1331 * Initialised pages do not have PageReserved set. This function is
1332 * called for each range allocated by the bootmem allocator and
1333 * marks the pages PageReserved. The remaining valid pages are later
1334 * sent to the buddy page allocator.
1336 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1338 unsigned long start_pfn = PFN_DOWN(start);
1339 unsigned long end_pfn = PFN_UP(end);
1341 for (; start_pfn < end_pfn; start_pfn++) {
1342 if (pfn_valid(start_pfn)) {
1343 struct page *page = pfn_to_page(start_pfn);
1345 init_reserved_page(start_pfn);
1347 /* Avoid false-positive PageTail() */
1348 INIT_LIST_HEAD(&page->lru);
1351 * no need for atomic set_bit because the struct
1352 * page is not visible yet so nobody should
1355 __SetPageReserved(page);
1360 static void __free_pages_ok(struct page *page, unsigned int order)
1362 unsigned long flags;
1364 unsigned long pfn = page_to_pfn(page);
1366 if (!free_pages_prepare(page, order, true))
1369 migratetype = get_pfnblock_migratetype(page, pfn);
1370 local_irq_save(flags);
1371 __count_vm_events(PGFREE, 1 << order);
1372 free_one_page(page_zone(page), page, pfn, order, migratetype);
1373 local_irq_restore(flags);
1376 void __free_pages_core(struct page *page, unsigned int order)
1378 unsigned int nr_pages = 1 << order;
1379 struct page *p = page;
1383 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1385 __ClearPageReserved(p);
1386 set_page_count(p, 0);
1388 __ClearPageReserved(p);
1389 set_page_count(p, 0);
1391 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1392 set_page_refcounted(page);
1393 __free_pages(page, order);
1396 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1397 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1399 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1401 int __meminit early_pfn_to_nid(unsigned long pfn)
1403 static DEFINE_SPINLOCK(early_pfn_lock);
1406 spin_lock(&early_pfn_lock);
1407 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1409 nid = first_online_node;
1410 spin_unlock(&early_pfn_lock);
1416 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1417 static inline bool __meminit __maybe_unused
1418 meminit_pfn_in_nid(unsigned long pfn, int node,
1419 struct mminit_pfnnid_cache *state)
1423 nid = __early_pfn_to_nid(pfn, state);
1424 if (nid >= 0 && nid != node)
1429 /* Only safe to use early in boot when initialisation is single-threaded */
1430 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1432 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1437 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1441 static inline bool __meminit __maybe_unused
1442 meminit_pfn_in_nid(unsigned long pfn, int node,
1443 struct mminit_pfnnid_cache *state)
1450 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1453 if (early_page_uninitialised(pfn))
1455 __free_pages_core(page, order);
1459 * Check that the whole (or subset of) a pageblock given by the interval of
1460 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1461 * with the migration of free compaction scanner. The scanners then need to
1462 * use only pfn_valid_within() check for arches that allow holes within
1465 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1467 * It's possible on some configurations to have a setup like node0 node1 node0
1468 * i.e. it's possible that all pages within a zones range of pages do not
1469 * belong to a single zone. We assume that a border between node0 and node1
1470 * can occur within a single pageblock, but not a node0 node1 node0
1471 * interleaving within a single pageblock. It is therefore sufficient to check
1472 * the first and last page of a pageblock and avoid checking each individual
1473 * page in a pageblock.
1475 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1476 unsigned long end_pfn, struct zone *zone)
1478 struct page *start_page;
1479 struct page *end_page;
1481 /* end_pfn is one past the range we are checking */
1484 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1487 start_page = pfn_to_online_page(start_pfn);
1491 if (page_zone(start_page) != zone)
1494 end_page = pfn_to_page(end_pfn);
1496 /* This gives a shorter code than deriving page_zone(end_page) */
1497 if (page_zone_id(start_page) != page_zone_id(end_page))
1503 void set_zone_contiguous(struct zone *zone)
1505 unsigned long block_start_pfn = zone->zone_start_pfn;
1506 unsigned long block_end_pfn;
1508 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1509 for (; block_start_pfn < zone_end_pfn(zone);
1510 block_start_pfn = block_end_pfn,
1511 block_end_pfn += pageblock_nr_pages) {
1513 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1515 if (!__pageblock_pfn_to_page(block_start_pfn,
1516 block_end_pfn, zone))
1520 /* We confirm that there is no hole */
1521 zone->contiguous = true;
1524 void clear_zone_contiguous(struct zone *zone)
1526 zone->contiguous = false;
1529 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1530 static void __init deferred_free_range(unsigned long pfn,
1531 unsigned long nr_pages)
1539 page = pfn_to_page(pfn);
1541 /* Free a large naturally-aligned chunk if possible */
1542 if (nr_pages == pageblock_nr_pages &&
1543 (pfn & (pageblock_nr_pages - 1)) == 0) {
1544 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1545 __free_pages_core(page, pageblock_order);
1549 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1550 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1551 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1552 __free_pages_core(page, 0);
1556 /* Completion tracking for deferred_init_memmap() threads */
1557 static atomic_t pgdat_init_n_undone __initdata;
1558 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1560 static inline void __init pgdat_init_report_one_done(void)
1562 if (atomic_dec_and_test(&pgdat_init_n_undone))
1563 complete(&pgdat_init_all_done_comp);
1567 * Returns true if page needs to be initialized or freed to buddy allocator.
1569 * First we check if pfn is valid on architectures where it is possible to have
1570 * holes within pageblock_nr_pages. On systems where it is not possible, this
1571 * function is optimized out.
1573 * Then, we check if a current large page is valid by only checking the validity
1576 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1577 * within a node: a pfn is between start and end of a node, but does not belong
1578 * to this memory node.
1580 static inline bool __init
1581 deferred_pfn_valid(int nid, unsigned long pfn,
1582 struct mminit_pfnnid_cache *nid_init_state)
1584 if (!pfn_valid_within(pfn))
1586 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1588 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1594 * Free pages to buddy allocator. Try to free aligned pages in
1595 * pageblock_nr_pages sizes.
1597 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1598 unsigned long end_pfn)
1600 struct mminit_pfnnid_cache nid_init_state = { };
1601 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1602 unsigned long nr_free = 0;
1604 for (; pfn < end_pfn; pfn++) {
1605 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1606 deferred_free_range(pfn - nr_free, nr_free);
1608 } else if (!(pfn & nr_pgmask)) {
1609 deferred_free_range(pfn - nr_free, nr_free);
1611 touch_nmi_watchdog();
1616 /* Free the last block of pages to allocator */
1617 deferred_free_range(pfn - nr_free, nr_free);
1621 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1622 * by performing it only once every pageblock_nr_pages.
1623 * Return number of pages initialized.
1625 static unsigned long __init deferred_init_pages(int nid, int zid,
1627 unsigned long end_pfn)
1629 struct mminit_pfnnid_cache nid_init_state = { };
1630 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1631 unsigned long nr_pages = 0;
1632 struct page *page = NULL;
1634 for (; pfn < end_pfn; pfn++) {
1635 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1638 } else if (!page || !(pfn & nr_pgmask)) {
1639 page = pfn_to_page(pfn);
1640 touch_nmi_watchdog();
1644 __init_single_page(page, pfn, zid, nid);
1650 /* Initialise remaining memory on a node */
1651 static int __init deferred_init_memmap(void *data)
1653 pg_data_t *pgdat = data;
1654 int nid = pgdat->node_id;
1655 unsigned long start = jiffies;
1656 unsigned long nr_pages = 0;
1657 unsigned long spfn, epfn, first_init_pfn, flags;
1658 phys_addr_t spa, epa;
1661 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1664 /* Bind memory initialisation thread to a local node if possible */
1665 if (!cpumask_empty(cpumask))
1666 set_cpus_allowed_ptr(current, cpumask);
1668 pgdat_resize_lock(pgdat, &flags);
1669 first_init_pfn = pgdat->first_deferred_pfn;
1670 if (first_init_pfn == ULONG_MAX) {
1671 pgdat_resize_unlock(pgdat, &flags);
1672 pgdat_init_report_one_done();
1676 /* Sanity check boundaries */
1677 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1678 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1679 pgdat->first_deferred_pfn = ULONG_MAX;
1681 /* Only the highest zone is deferred so find it */
1682 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1683 zone = pgdat->node_zones + zid;
1684 if (first_init_pfn < zone_end_pfn(zone))
1687 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1690 * Initialize and free pages. We do it in two loops: first we initialize
1691 * struct page, than free to buddy allocator, because while we are
1692 * freeing pages we can access pages that are ahead (computing buddy
1693 * page in __free_one_page()).
1695 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1696 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1697 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1698 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1700 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1701 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1702 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1703 deferred_free_pages(nid, zid, spfn, epfn);
1705 pgdat_resize_unlock(pgdat, &flags);
1707 /* Sanity check that the next zone really is unpopulated */
1708 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1710 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1711 jiffies_to_msecs(jiffies - start));
1713 pgdat_init_report_one_done();
1718 * If this zone has deferred pages, try to grow it by initializing enough
1719 * deferred pages to satisfy the allocation specified by order, rounded up to
1720 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1721 * of SECTION_SIZE bytes by initializing struct pages in increments of
1722 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1724 * Return true when zone was grown, otherwise return false. We return true even
1725 * when we grow less than requested, to let the caller decide if there are
1726 * enough pages to satisfy the allocation.
1728 * Note: We use noinline because this function is needed only during boot, and
1729 * it is called from a __ref function _deferred_grow_zone. This way we are
1730 * making sure that it is not inlined into permanent text section.
1732 static noinline bool __init
1733 deferred_grow_zone(struct zone *zone, unsigned int order)
1735 int zid = zone_idx(zone);
1736 int nid = zone_to_nid(zone);
1737 pg_data_t *pgdat = NODE_DATA(nid);
1738 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1739 unsigned long nr_pages = 0;
1740 unsigned long first_init_pfn, spfn, epfn, t, flags;
1741 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1742 phys_addr_t spa, epa;
1745 /* Only the last zone may have deferred pages */
1746 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1749 pgdat_resize_lock(pgdat, &flags);
1752 * If deferred pages have been initialized while we were waiting for
1753 * the lock, return true, as the zone was grown. The caller will retry
1754 * this zone. We won't return to this function since the caller also
1755 * has this static branch.
1757 if (!static_branch_unlikely(&deferred_pages)) {
1758 pgdat_resize_unlock(pgdat, &flags);
1763 * If someone grew this zone while we were waiting for spinlock, return
1764 * true, as there might be enough pages already.
1766 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1767 pgdat_resize_unlock(pgdat, &flags);
1771 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1773 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1774 pgdat_resize_unlock(pgdat, &flags);
1778 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1779 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1780 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1782 while (spfn < epfn && nr_pages < nr_pages_needed) {
1783 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1784 first_deferred_pfn = min(t, epfn);
1785 nr_pages += deferred_init_pages(nid, zid, spfn,
1786 first_deferred_pfn);
1787 spfn = first_deferred_pfn;
1790 if (nr_pages >= nr_pages_needed)
1794 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1795 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1796 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1797 deferred_free_pages(nid, zid, spfn, epfn);
1799 if (first_deferred_pfn == epfn)
1802 pgdat->first_deferred_pfn = first_deferred_pfn;
1803 pgdat_resize_unlock(pgdat, &flags);
1805 return nr_pages > 0;
1809 * deferred_grow_zone() is __init, but it is called from
1810 * get_page_from_freelist() during early boot until deferred_pages permanently
1811 * disables this call. This is why we have refdata wrapper to avoid warning,
1812 * and to ensure that the function body gets unloaded.
1815 _deferred_grow_zone(struct zone *zone, unsigned int order)
1817 return deferred_grow_zone(zone, order);
1820 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1822 void __init page_alloc_init_late(void)
1826 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1829 /* There will be num_node_state(N_MEMORY) threads */
1830 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1831 for_each_node_state(nid, N_MEMORY) {
1832 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1835 /* Block until all are initialised */
1836 wait_for_completion(&pgdat_init_all_done_comp);
1839 * We initialized the rest of the deferred pages. Permanently disable
1840 * on-demand struct page initialization.
1842 static_branch_disable(&deferred_pages);
1844 /* Reinit limits that are based on free pages after the kernel is up */
1845 files_maxfiles_init();
1847 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1848 /* Discard memblock private memory */
1852 for_each_populated_zone(zone)
1853 set_zone_contiguous(zone);
1857 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1858 void __init init_cma_reserved_pageblock(struct page *page)
1860 unsigned i = pageblock_nr_pages;
1861 struct page *p = page;
1864 __ClearPageReserved(p);
1865 set_page_count(p, 0);
1868 set_pageblock_migratetype(page, MIGRATE_CMA);
1870 if (pageblock_order >= MAX_ORDER) {
1871 i = pageblock_nr_pages;
1874 set_page_refcounted(p);
1875 __free_pages(p, MAX_ORDER - 1);
1876 p += MAX_ORDER_NR_PAGES;
1877 } while (i -= MAX_ORDER_NR_PAGES);
1879 set_page_refcounted(page);
1880 __free_pages(page, pageblock_order);
1883 adjust_managed_page_count(page, pageblock_nr_pages);
1888 * The order of subdivision here is critical for the IO subsystem.
1889 * Please do not alter this order without good reasons and regression
1890 * testing. Specifically, as large blocks of memory are subdivided,
1891 * the order in which smaller blocks are delivered depends on the order
1892 * they're subdivided in this function. This is the primary factor
1893 * influencing the order in which pages are delivered to the IO
1894 * subsystem according to empirical testing, and this is also justified
1895 * by considering the behavior of a buddy system containing a single
1896 * large block of memory acted on by a series of small allocations.
1897 * This behavior is a critical factor in sglist merging's success.
1901 static inline void expand(struct zone *zone, struct page *page,
1902 int low, int high, struct free_area *area,
1905 unsigned long size = 1 << high;
1907 while (high > low) {
1911 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1914 * Mark as guard pages (or page), that will allow to
1915 * merge back to allocator when buddy will be freed.
1916 * Corresponding page table entries will not be touched,
1917 * pages will stay not present in virtual address space
1919 if (set_page_guard(zone, &page[size], high, migratetype))
1922 list_add(&page[size].lru, &area->free_list[migratetype]);
1924 set_page_order(&page[size], high);
1928 static void check_new_page_bad(struct page *page)
1930 const char *bad_reason = NULL;
1931 unsigned long bad_flags = 0;
1933 if (unlikely(atomic_read(&page->_mapcount) != -1))
1934 bad_reason = "nonzero mapcount";
1935 if (unlikely(page->mapping != NULL))
1936 bad_reason = "non-NULL mapping";
1937 if (unlikely(page_ref_count(page) != 0))
1938 bad_reason = "nonzero _count";
1939 if (unlikely(page->flags & __PG_HWPOISON)) {
1940 bad_reason = "HWPoisoned (hardware-corrupted)";
1941 bad_flags = __PG_HWPOISON;
1942 /* Don't complain about hwpoisoned pages */
1943 page_mapcount_reset(page); /* remove PageBuddy */
1946 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1947 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1948 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1951 if (unlikely(page->mem_cgroup))
1952 bad_reason = "page still charged to cgroup";
1954 bad_page(page, bad_reason, bad_flags);
1958 * This page is about to be returned from the page allocator
1960 static inline int check_new_page(struct page *page)
1962 if (likely(page_expected_state(page,
1963 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1966 check_new_page_bad(page);
1970 static inline bool free_pages_prezeroed(void)
1972 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1973 page_poisoning_enabled();
1976 #ifdef CONFIG_DEBUG_VM
1977 static bool check_pcp_refill(struct page *page)
1982 static bool check_new_pcp(struct page *page)
1984 return check_new_page(page);
1987 static bool check_pcp_refill(struct page *page)
1989 return check_new_page(page);
1991 static bool check_new_pcp(struct page *page)
1995 #endif /* CONFIG_DEBUG_VM */
1997 static bool check_new_pages(struct page *page, unsigned int order)
2000 for (i = 0; i < (1 << order); i++) {
2001 struct page *p = page + i;
2003 if (unlikely(check_new_page(p)))
2010 inline void post_alloc_hook(struct page *page, unsigned int order,
2013 set_page_private(page, 0);
2014 set_page_refcounted(page);
2016 arch_alloc_page(page, order);
2017 kernel_map_pages(page, 1 << order, 1);
2018 kasan_alloc_pages(page, order);
2019 kernel_poison_pages(page, 1 << order, 1);
2020 set_page_owner(page, order, gfp_flags);
2023 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2024 unsigned int alloc_flags)
2028 post_alloc_hook(page, order, gfp_flags);
2030 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
2031 for (i = 0; i < (1 << order); i++)
2032 clear_highpage(page + i);
2034 if (order && (gfp_flags & __GFP_COMP))
2035 prep_compound_page(page, order);
2038 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2039 * allocate the page. The expectation is that the caller is taking
2040 * steps that will free more memory. The caller should avoid the page
2041 * being used for !PFMEMALLOC purposes.
2043 if (alloc_flags & ALLOC_NO_WATERMARKS)
2044 set_page_pfmemalloc(page);
2046 clear_page_pfmemalloc(page);
2050 * Go through the free lists for the given migratetype and remove
2051 * the smallest available page from the freelists
2053 static __always_inline
2054 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2057 unsigned int current_order;
2058 struct free_area *area;
2061 /* Find a page of the appropriate size in the preferred list */
2062 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2063 area = &(zone->free_area[current_order]);
2064 page = list_first_entry_or_null(&area->free_list[migratetype],
2068 list_del(&page->lru);
2069 rmv_page_order(page);
2071 expand(zone, page, order, current_order, area, migratetype);
2072 set_pcppage_migratetype(page, migratetype);
2081 * This array describes the order lists are fallen back to when
2082 * the free lists for the desirable migrate type are depleted
2084 static int fallbacks[MIGRATE_TYPES][4] = {
2085 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2086 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2087 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2089 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2091 #ifdef CONFIG_MEMORY_ISOLATION
2092 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2097 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2100 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2103 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2104 unsigned int order) { return NULL; }
2108 * Move the free pages in a range to the free lists of the requested type.
2109 * Note that start_page and end_pages are not aligned on a pageblock
2110 * boundary. If alignment is required, use move_freepages_block()
2112 static int move_freepages(struct zone *zone,
2113 struct page *start_page, struct page *end_page,
2114 int migratetype, int *num_movable)
2118 int pages_moved = 0;
2120 #ifndef CONFIG_HOLES_IN_ZONE
2122 * page_zone is not safe to call in this context when
2123 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2124 * anyway as we check zone boundaries in move_freepages_block().
2125 * Remove at a later date when no bug reports exist related to
2126 * grouping pages by mobility
2128 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2129 pfn_valid(page_to_pfn(end_page)) &&
2130 page_zone(start_page) != page_zone(end_page));
2132 for (page = start_page; page <= end_page;) {
2133 if (!pfn_valid_within(page_to_pfn(page))) {
2138 /* Make sure we are not inadvertently changing nodes */
2139 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2141 if (!PageBuddy(page)) {
2143 * We assume that pages that could be isolated for
2144 * migration are movable. But we don't actually try
2145 * isolating, as that would be expensive.
2148 (PageLRU(page) || __PageMovable(page)))
2155 order = page_order(page);
2156 list_move(&page->lru,
2157 &zone->free_area[order].free_list[migratetype]);
2159 pages_moved += 1 << order;
2165 int move_freepages_block(struct zone *zone, struct page *page,
2166 int migratetype, int *num_movable)
2168 unsigned long start_pfn, end_pfn;
2169 struct page *start_page, *end_page;
2174 start_pfn = page_to_pfn(page);
2175 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2176 start_page = pfn_to_page(start_pfn);
2177 end_page = start_page + pageblock_nr_pages - 1;
2178 end_pfn = start_pfn + pageblock_nr_pages - 1;
2180 /* Do not cross zone boundaries */
2181 if (!zone_spans_pfn(zone, start_pfn))
2183 if (!zone_spans_pfn(zone, end_pfn))
2186 return move_freepages(zone, start_page, end_page, migratetype,
2190 static void change_pageblock_range(struct page *pageblock_page,
2191 int start_order, int migratetype)
2193 int nr_pageblocks = 1 << (start_order - pageblock_order);
2195 while (nr_pageblocks--) {
2196 set_pageblock_migratetype(pageblock_page, migratetype);
2197 pageblock_page += pageblock_nr_pages;
2202 * When we are falling back to another migratetype during allocation, try to
2203 * steal extra free pages from the same pageblocks to satisfy further
2204 * allocations, instead of polluting multiple pageblocks.
2206 * If we are stealing a relatively large buddy page, it is likely there will
2207 * be more free pages in the pageblock, so try to steal them all. For
2208 * reclaimable and unmovable allocations, we steal regardless of page size,
2209 * as fragmentation caused by those allocations polluting movable pageblocks
2210 * is worse than movable allocations stealing from unmovable and reclaimable
2213 static bool can_steal_fallback(unsigned int order, int start_mt)
2216 * Leaving this order check is intended, although there is
2217 * relaxed order check in next check. The reason is that
2218 * we can actually steal whole pageblock if this condition met,
2219 * but, below check doesn't guarantee it and that is just heuristic
2220 * so could be changed anytime.
2222 if (order >= pageblock_order)
2225 if (order >= pageblock_order / 2 ||
2226 start_mt == MIGRATE_RECLAIMABLE ||
2227 start_mt == MIGRATE_UNMOVABLE ||
2228 page_group_by_mobility_disabled)
2234 static inline void boost_watermark(struct zone *zone)
2236 unsigned long max_boost;
2238 if (!watermark_boost_factor)
2241 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2242 watermark_boost_factor, 10000);
2245 * high watermark may be uninitialised if fragmentation occurs
2246 * very early in boot so do not boost. We do not fall
2247 * through and boost by pageblock_nr_pages as failing
2248 * allocations that early means that reclaim is not going
2249 * to help and it may even be impossible to reclaim the
2250 * boosted watermark resulting in a hang.
2255 max_boost = max(pageblock_nr_pages, max_boost);
2257 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2262 * This function implements actual steal behaviour. If order is large enough,
2263 * we can steal whole pageblock. If not, we first move freepages in this
2264 * pageblock to our migratetype and determine how many already-allocated pages
2265 * are there in the pageblock with a compatible migratetype. If at least half
2266 * of pages are free or compatible, we can change migratetype of the pageblock
2267 * itself, so pages freed in the future will be put on the correct free list.
2269 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2270 unsigned int alloc_flags, int start_type, bool whole_block)
2272 unsigned int current_order = page_order(page);
2273 struct free_area *area;
2274 int free_pages, movable_pages, alike_pages;
2277 old_block_type = get_pageblock_migratetype(page);
2280 * This can happen due to races and we want to prevent broken
2281 * highatomic accounting.
2283 if (is_migrate_highatomic(old_block_type))
2286 /* Take ownership for orders >= pageblock_order */
2287 if (current_order >= pageblock_order) {
2288 change_pageblock_range(page, current_order, start_type);
2293 * Boost watermarks to increase reclaim pressure to reduce the
2294 * likelihood of future fallbacks. Wake kswapd now as the node
2295 * may be balanced overall and kswapd will not wake naturally.
2297 boost_watermark(zone);
2298 if (alloc_flags & ALLOC_KSWAPD)
2299 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2301 /* We are not allowed to try stealing from the whole block */
2305 free_pages = move_freepages_block(zone, page, start_type,
2308 * Determine how many pages are compatible with our allocation.
2309 * For movable allocation, it's the number of movable pages which
2310 * we just obtained. For other types it's a bit more tricky.
2312 if (start_type == MIGRATE_MOVABLE) {
2313 alike_pages = movable_pages;
2316 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2317 * to MOVABLE pageblock, consider all non-movable pages as
2318 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2319 * vice versa, be conservative since we can't distinguish the
2320 * exact migratetype of non-movable pages.
2322 if (old_block_type == MIGRATE_MOVABLE)
2323 alike_pages = pageblock_nr_pages
2324 - (free_pages + movable_pages);
2329 /* moving whole block can fail due to zone boundary conditions */
2334 * If a sufficient number of pages in the block are either free or of
2335 * comparable migratability as our allocation, claim the whole block.
2337 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2338 page_group_by_mobility_disabled)
2339 set_pageblock_migratetype(page, start_type);
2344 area = &zone->free_area[current_order];
2345 list_move(&page->lru, &area->free_list[start_type]);
2349 * Check whether there is a suitable fallback freepage with requested order.
2350 * If only_stealable is true, this function returns fallback_mt only if
2351 * we can steal other freepages all together. This would help to reduce
2352 * fragmentation due to mixed migratetype pages in one pageblock.
2354 int find_suitable_fallback(struct free_area *area, unsigned int order,
2355 int migratetype, bool only_stealable, bool *can_steal)
2360 if (area->nr_free == 0)
2365 fallback_mt = fallbacks[migratetype][i];
2366 if (fallback_mt == MIGRATE_TYPES)
2369 if (list_empty(&area->free_list[fallback_mt]))
2372 if (can_steal_fallback(order, migratetype))
2375 if (!only_stealable)
2386 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2387 * there are no empty page blocks that contain a page with a suitable order
2389 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2390 unsigned int alloc_order)
2393 unsigned long max_managed, flags;
2396 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2397 * Check is race-prone but harmless.
2399 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2400 if (zone->nr_reserved_highatomic >= max_managed)
2403 spin_lock_irqsave(&zone->lock, flags);
2405 /* Recheck the nr_reserved_highatomic limit under the lock */
2406 if (zone->nr_reserved_highatomic >= max_managed)
2410 mt = get_pageblock_migratetype(page);
2411 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2412 && !is_migrate_cma(mt)) {
2413 zone->nr_reserved_highatomic += pageblock_nr_pages;
2414 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2415 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2419 spin_unlock_irqrestore(&zone->lock, flags);
2423 * Used when an allocation is about to fail under memory pressure. This
2424 * potentially hurts the reliability of high-order allocations when under
2425 * intense memory pressure but failed atomic allocations should be easier
2426 * to recover from than an OOM.
2428 * If @force is true, try to unreserve a pageblock even though highatomic
2429 * pageblock is exhausted.
2431 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2434 struct zonelist *zonelist = ac->zonelist;
2435 unsigned long flags;
2442 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2445 * Preserve at least one pageblock unless memory pressure
2448 if (!force && zone->nr_reserved_highatomic <=
2452 spin_lock_irqsave(&zone->lock, flags);
2453 for (order = 0; order < MAX_ORDER; order++) {
2454 struct free_area *area = &(zone->free_area[order]);
2456 page = list_first_entry_or_null(
2457 &area->free_list[MIGRATE_HIGHATOMIC],
2463 * In page freeing path, migratetype change is racy so
2464 * we can counter several free pages in a pageblock
2465 * in this loop althoug we changed the pageblock type
2466 * from highatomic to ac->migratetype. So we should
2467 * adjust the count once.
2469 if (is_migrate_highatomic_page(page)) {
2471 * It should never happen but changes to
2472 * locking could inadvertently allow a per-cpu
2473 * drain to add pages to MIGRATE_HIGHATOMIC
2474 * while unreserving so be safe and watch for
2477 zone->nr_reserved_highatomic -= min(
2479 zone->nr_reserved_highatomic);
2483 * Convert to ac->migratetype and avoid the normal
2484 * pageblock stealing heuristics. Minimally, the caller
2485 * is doing the work and needs the pages. More
2486 * importantly, if the block was always converted to
2487 * MIGRATE_UNMOVABLE or another type then the number
2488 * of pageblocks that cannot be completely freed
2491 set_pageblock_migratetype(page, ac->migratetype);
2492 ret = move_freepages_block(zone, page, ac->migratetype,
2495 spin_unlock_irqrestore(&zone->lock, flags);
2499 spin_unlock_irqrestore(&zone->lock, flags);
2506 * Try finding a free buddy page on the fallback list and put it on the free
2507 * list of requested migratetype, possibly along with other pages from the same
2508 * block, depending on fragmentation avoidance heuristics. Returns true if
2509 * fallback was found so that __rmqueue_smallest() can grab it.
2511 * The use of signed ints for order and current_order is a deliberate
2512 * deviation from the rest of this file, to make the for loop
2513 * condition simpler.
2515 static __always_inline bool
2516 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2517 unsigned int alloc_flags)
2519 struct free_area *area;
2521 int min_order = order;
2527 * Do not steal pages from freelists belonging to other pageblocks
2528 * i.e. orders < pageblock_order. If there are no local zones free,
2529 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2531 if (alloc_flags & ALLOC_NOFRAGMENT)
2532 min_order = pageblock_order;
2535 * Find the largest available free page in the other list. This roughly
2536 * approximates finding the pageblock with the most free pages, which
2537 * would be too costly to do exactly.
2539 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2541 area = &(zone->free_area[current_order]);
2542 fallback_mt = find_suitable_fallback(area, current_order,
2543 start_migratetype, false, &can_steal);
2544 if (fallback_mt == -1)
2548 * We cannot steal all free pages from the pageblock and the
2549 * requested migratetype is movable. In that case it's better to
2550 * steal and split the smallest available page instead of the
2551 * largest available page, because even if the next movable
2552 * allocation falls back into a different pageblock than this
2553 * one, it won't cause permanent fragmentation.
2555 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2556 && current_order > order)
2565 for (current_order = order; current_order < MAX_ORDER;
2567 area = &(zone->free_area[current_order]);
2568 fallback_mt = find_suitable_fallback(area, current_order,
2569 start_migratetype, false, &can_steal);
2570 if (fallback_mt != -1)
2575 * This should not happen - we already found a suitable fallback
2576 * when looking for the largest page.
2578 VM_BUG_ON(current_order == MAX_ORDER);
2581 page = list_first_entry(&area->free_list[fallback_mt],
2584 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2587 trace_mm_page_alloc_extfrag(page, order, current_order,
2588 start_migratetype, fallback_mt);
2595 * Do the hard work of removing an element from the buddy allocator.
2596 * Call me with the zone->lock already held.
2598 static __always_inline struct page *
2599 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2600 unsigned int alloc_flags)
2605 page = __rmqueue_smallest(zone, order, migratetype);
2606 if (unlikely(!page)) {
2607 if (migratetype == MIGRATE_MOVABLE)
2608 page = __rmqueue_cma_fallback(zone, order);
2610 if (!page && __rmqueue_fallback(zone, order, migratetype,
2615 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2620 * Obtain a specified number of elements from the buddy allocator, all under
2621 * a single hold of the lock, for efficiency. Add them to the supplied list.
2622 * Returns the number of new pages which were placed at *list.
2624 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2625 unsigned long count, struct list_head *list,
2626 int migratetype, unsigned int alloc_flags)
2630 spin_lock(&zone->lock);
2631 for (i = 0; i < count; ++i) {
2632 struct page *page = __rmqueue(zone, order, migratetype,
2634 if (unlikely(page == NULL))
2637 if (unlikely(check_pcp_refill(page)))
2641 * Split buddy pages returned by expand() are received here in
2642 * physical page order. The page is added to the tail of
2643 * caller's list. From the callers perspective, the linked list
2644 * is ordered by page number under some conditions. This is
2645 * useful for IO devices that can forward direction from the
2646 * head, thus also in the physical page order. This is useful
2647 * for IO devices that can merge IO requests if the physical
2648 * pages are ordered properly.
2650 list_add_tail(&page->lru, list);
2652 if (is_migrate_cma(get_pcppage_migratetype(page)))
2653 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2658 * i pages were removed from the buddy list even if some leak due
2659 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2660 * on i. Do not confuse with 'alloced' which is the number of
2661 * pages added to the pcp list.
2663 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2664 spin_unlock(&zone->lock);
2670 * Called from the vmstat counter updater to drain pagesets of this
2671 * currently executing processor on remote nodes after they have
2674 * Note that this function must be called with the thread pinned to
2675 * a single processor.
2677 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2679 unsigned long flags;
2680 int to_drain, batch;
2682 local_irq_save(flags);
2683 batch = READ_ONCE(pcp->batch);
2684 to_drain = min(pcp->count, batch);
2686 free_pcppages_bulk(zone, to_drain, pcp);
2687 local_irq_restore(flags);
2692 * Drain pcplists of the indicated processor and zone.
2694 * The processor must either be the current processor and the
2695 * thread pinned to the current processor or a processor that
2698 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2700 unsigned long flags;
2701 struct per_cpu_pageset *pset;
2702 struct per_cpu_pages *pcp;
2704 local_irq_save(flags);
2705 pset = per_cpu_ptr(zone->pageset, cpu);
2709 free_pcppages_bulk(zone, pcp->count, pcp);
2710 local_irq_restore(flags);
2714 * Drain pcplists of all zones on the indicated processor.
2716 * The processor must either be the current processor and the
2717 * thread pinned to the current processor or a processor that
2720 static void drain_pages(unsigned int cpu)
2724 for_each_populated_zone(zone) {
2725 drain_pages_zone(cpu, zone);
2730 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2732 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2733 * the single zone's pages.
2735 void drain_local_pages(struct zone *zone)
2737 int cpu = smp_processor_id();
2740 drain_pages_zone(cpu, zone);
2745 static void drain_local_pages_wq(struct work_struct *work)
2747 struct pcpu_drain *drain;
2749 drain = container_of(work, struct pcpu_drain, work);
2752 * drain_all_pages doesn't use proper cpu hotplug protection so
2753 * we can race with cpu offline when the WQ can move this from
2754 * a cpu pinned worker to an unbound one. We can operate on a different
2755 * cpu which is allright but we also have to make sure to not move to
2759 drain_local_pages(drain->zone);
2764 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2766 * When zone parameter is non-NULL, spill just the single zone's pages.
2768 * Note that this can be extremely slow as the draining happens in a workqueue.
2770 void drain_all_pages(struct zone *zone)
2775 * Allocate in the BSS so we wont require allocation in
2776 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2778 static cpumask_t cpus_with_pcps;
2781 * Make sure nobody triggers this path before mm_percpu_wq is fully
2784 if (WARN_ON_ONCE(!mm_percpu_wq))
2788 * Do not drain if one is already in progress unless it's specific to
2789 * a zone. Such callers are primarily CMA and memory hotplug and need
2790 * the drain to be complete when the call returns.
2792 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2795 mutex_lock(&pcpu_drain_mutex);
2799 * We don't care about racing with CPU hotplug event
2800 * as offline notification will cause the notified
2801 * cpu to drain that CPU pcps and on_each_cpu_mask
2802 * disables preemption as part of its processing
2804 for_each_online_cpu(cpu) {
2805 struct per_cpu_pageset *pcp;
2807 bool has_pcps = false;
2810 pcp = per_cpu_ptr(zone->pageset, cpu);
2814 for_each_populated_zone(z) {
2815 pcp = per_cpu_ptr(z->pageset, cpu);
2816 if (pcp->pcp.count) {
2824 cpumask_set_cpu(cpu, &cpus_with_pcps);
2826 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2829 for_each_cpu(cpu, &cpus_with_pcps) {
2830 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2833 INIT_WORK(&drain->work, drain_local_pages_wq);
2834 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2836 for_each_cpu(cpu, &cpus_with_pcps)
2837 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2839 mutex_unlock(&pcpu_drain_mutex);
2842 #ifdef CONFIG_HIBERNATION
2845 * Touch the watchdog for every WD_PAGE_COUNT pages.
2847 #define WD_PAGE_COUNT (128*1024)
2849 void mark_free_pages(struct zone *zone)
2851 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2852 unsigned long flags;
2853 unsigned int order, t;
2856 if (zone_is_empty(zone))
2859 spin_lock_irqsave(&zone->lock, flags);
2861 max_zone_pfn = zone_end_pfn(zone);
2862 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2863 if (pfn_valid(pfn)) {
2864 page = pfn_to_page(pfn);
2866 if (!--page_count) {
2867 touch_nmi_watchdog();
2868 page_count = WD_PAGE_COUNT;
2871 if (page_zone(page) != zone)
2874 if (!swsusp_page_is_forbidden(page))
2875 swsusp_unset_page_free(page);
2878 for_each_migratetype_order(order, t) {
2879 list_for_each_entry(page,
2880 &zone->free_area[order].free_list[t], lru) {
2883 pfn = page_to_pfn(page);
2884 for (i = 0; i < (1UL << order); i++) {
2885 if (!--page_count) {
2886 touch_nmi_watchdog();
2887 page_count = WD_PAGE_COUNT;
2889 swsusp_set_page_free(pfn_to_page(pfn + i));
2893 spin_unlock_irqrestore(&zone->lock, flags);
2895 #endif /* CONFIG_PM */
2897 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2901 if (!free_pcp_prepare(page))
2904 migratetype = get_pfnblock_migratetype(page, pfn);
2905 set_pcppage_migratetype(page, migratetype);
2909 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2911 struct zone *zone = page_zone(page);
2912 struct per_cpu_pages *pcp;
2915 migratetype = get_pcppage_migratetype(page);
2916 __count_vm_event(PGFREE);
2919 * We only track unmovable, reclaimable and movable on pcp lists.
2920 * Free ISOLATE pages back to the allocator because they are being
2921 * offlined but treat HIGHATOMIC as movable pages so we can get those
2922 * areas back if necessary. Otherwise, we may have to free
2923 * excessively into the page allocator
2925 if (migratetype >= MIGRATE_PCPTYPES) {
2926 if (unlikely(is_migrate_isolate(migratetype))) {
2927 free_one_page(zone, page, pfn, 0, migratetype);
2930 migratetype = MIGRATE_MOVABLE;
2933 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2934 list_add(&page->lru, &pcp->lists[migratetype]);
2936 if (pcp->count >= pcp->high) {
2937 unsigned long batch = READ_ONCE(pcp->batch);
2938 free_pcppages_bulk(zone, batch, pcp);
2943 * Free a 0-order page
2945 void free_unref_page(struct page *page)
2947 unsigned long flags;
2948 unsigned long pfn = page_to_pfn(page);
2950 if (!free_unref_page_prepare(page, pfn))
2953 local_irq_save(flags);
2954 free_unref_page_commit(page, pfn);
2955 local_irq_restore(flags);
2959 * Free a list of 0-order pages
2961 void free_unref_page_list(struct list_head *list)
2963 struct page *page, *next;
2964 unsigned long flags, pfn;
2965 int batch_count = 0;
2967 /* Prepare pages for freeing */
2968 list_for_each_entry_safe(page, next, list, lru) {
2969 pfn = page_to_pfn(page);
2970 if (!free_unref_page_prepare(page, pfn))
2971 list_del(&page->lru);
2972 set_page_private(page, pfn);
2975 local_irq_save(flags);
2976 list_for_each_entry_safe(page, next, list, lru) {
2977 unsigned long pfn = page_private(page);
2979 set_page_private(page, 0);
2980 trace_mm_page_free_batched(page);
2981 free_unref_page_commit(page, pfn);
2984 * Guard against excessive IRQ disabled times when we get
2985 * a large list of pages to free.
2987 if (++batch_count == SWAP_CLUSTER_MAX) {
2988 local_irq_restore(flags);
2990 local_irq_save(flags);
2993 local_irq_restore(flags);
2997 * split_page takes a non-compound higher-order page, and splits it into
2998 * n (1<<order) sub-pages: page[0..n]
2999 * Each sub-page must be freed individually.
3001 * Note: this is probably too low level an operation for use in drivers.
3002 * Please consult with lkml before using this in your driver.
3004 void split_page(struct page *page, unsigned int order)
3008 VM_BUG_ON_PAGE(PageCompound(page), page);
3009 VM_BUG_ON_PAGE(!page_count(page), page);
3011 for (i = 1; i < (1 << order); i++)
3012 set_page_refcounted(page + i);
3013 split_page_owner(page, order);
3015 EXPORT_SYMBOL_GPL(split_page);
3017 int __isolate_free_page(struct page *page, unsigned int order)
3019 unsigned long watermark;
3023 BUG_ON(!PageBuddy(page));
3025 zone = page_zone(page);
3026 mt = get_pageblock_migratetype(page);
3028 if (!is_migrate_isolate(mt)) {
3030 * Obey watermarks as if the page was being allocated. We can
3031 * emulate a high-order watermark check with a raised order-0
3032 * watermark, because we already know our high-order page
3035 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3036 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3039 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3042 /* Remove page from free list */
3043 list_del(&page->lru);
3044 zone->free_area[order].nr_free--;
3045 rmv_page_order(page);
3048 * Set the pageblock if the isolated page is at least half of a
3051 if (order >= pageblock_order - 1) {
3052 struct page *endpage = page + (1 << order) - 1;
3053 for (; page < endpage; page += pageblock_nr_pages) {
3054 int mt = get_pageblock_migratetype(page);
3055 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3056 && !is_migrate_highatomic(mt))
3057 set_pageblock_migratetype(page,
3063 return 1UL << order;
3067 * Update NUMA hit/miss statistics
3069 * Must be called with interrupts disabled.
3071 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3074 enum numa_stat_item local_stat = NUMA_LOCAL;
3076 /* skip numa counters update if numa stats is disabled */
3077 if (!static_branch_likely(&vm_numa_stat_key))
3080 if (zone_to_nid(z) != numa_node_id())
3081 local_stat = NUMA_OTHER;
3083 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3084 __inc_numa_state(z, NUMA_HIT);
3086 __inc_numa_state(z, NUMA_MISS);
3087 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3089 __inc_numa_state(z, local_stat);
3093 /* Remove page from the per-cpu list, caller must protect the list */
3094 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3095 unsigned int alloc_flags,
3096 struct per_cpu_pages *pcp,
3097 struct list_head *list)
3102 if (list_empty(list)) {
3103 pcp->count += rmqueue_bulk(zone, 0,
3105 migratetype, alloc_flags);
3106 if (unlikely(list_empty(list)))
3110 page = list_first_entry(list, struct page, lru);
3111 list_del(&page->lru);
3113 } while (check_new_pcp(page));
3118 /* Lock and remove page from the per-cpu list */
3119 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3120 struct zone *zone, unsigned int order,
3121 gfp_t gfp_flags, int migratetype,
3122 unsigned int alloc_flags)
3124 struct per_cpu_pages *pcp;
3125 struct list_head *list;
3127 unsigned long flags;
3129 local_irq_save(flags);
3130 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3131 list = &pcp->lists[migratetype];
3132 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3134 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3135 zone_statistics(preferred_zone, zone);
3137 local_irq_restore(flags);
3142 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3145 struct page *rmqueue(struct zone *preferred_zone,
3146 struct zone *zone, unsigned int order,
3147 gfp_t gfp_flags, unsigned int alloc_flags,
3150 unsigned long flags;
3153 if (likely(order == 0)) {
3154 page = rmqueue_pcplist(preferred_zone, zone, order,
3155 gfp_flags, migratetype, alloc_flags);
3160 * We most definitely don't want callers attempting to
3161 * allocate greater than order-1 page units with __GFP_NOFAIL.
3163 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3164 spin_lock_irqsave(&zone->lock, flags);
3168 if (alloc_flags & ALLOC_HARDER) {
3169 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3171 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3174 page = __rmqueue(zone, order, migratetype, alloc_flags);
3175 } while (page && check_new_pages(page, order));
3176 spin_unlock(&zone->lock);
3179 __mod_zone_freepage_state(zone, -(1 << order),
3180 get_pcppage_migratetype(page));
3182 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3183 zone_statistics(preferred_zone, zone);
3184 local_irq_restore(flags);
3187 /* Separate test+clear to avoid unnecessary atomics */
3188 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3189 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3190 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3193 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3197 local_irq_restore(flags);
3201 #ifdef CONFIG_FAIL_PAGE_ALLOC
3204 struct fault_attr attr;
3206 bool ignore_gfp_highmem;
3207 bool ignore_gfp_reclaim;
3209 } fail_page_alloc = {
3210 .attr = FAULT_ATTR_INITIALIZER,
3211 .ignore_gfp_reclaim = true,
3212 .ignore_gfp_highmem = true,
3216 static int __init setup_fail_page_alloc(char *str)
3218 return setup_fault_attr(&fail_page_alloc.attr, str);
3220 __setup("fail_page_alloc=", setup_fail_page_alloc);
3222 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3224 if (order < fail_page_alloc.min_order)
3226 if (gfp_mask & __GFP_NOFAIL)
3228 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3230 if (fail_page_alloc.ignore_gfp_reclaim &&
3231 (gfp_mask & __GFP_DIRECT_RECLAIM))
3234 return should_fail(&fail_page_alloc.attr, 1 << order);
3237 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3239 static int __init fail_page_alloc_debugfs(void)
3241 umode_t mode = S_IFREG | 0600;
3244 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3245 &fail_page_alloc.attr);
3247 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3248 &fail_page_alloc.ignore_gfp_reclaim);
3249 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3250 &fail_page_alloc.ignore_gfp_highmem);
3251 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3256 late_initcall(fail_page_alloc_debugfs);
3258 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3260 #else /* CONFIG_FAIL_PAGE_ALLOC */
3262 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3267 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3269 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3271 return __should_fail_alloc_page(gfp_mask, order);
3273 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3276 * Return true if free base pages are above 'mark'. For high-order checks it
3277 * will return true of the order-0 watermark is reached and there is at least
3278 * one free page of a suitable size. Checking now avoids taking the zone lock
3279 * to check in the allocation paths if no pages are free.
3281 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3282 int classzone_idx, unsigned int alloc_flags,
3287 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3289 /* free_pages may go negative - that's OK */
3290 free_pages -= (1 << order) - 1;
3292 if (alloc_flags & ALLOC_HIGH)
3296 * If the caller does not have rights to ALLOC_HARDER then subtract
3297 * the high-atomic reserves. This will over-estimate the size of the
3298 * atomic reserve but it avoids a search.
3300 if (likely(!alloc_harder)) {
3301 free_pages -= z->nr_reserved_highatomic;
3304 * OOM victims can try even harder than normal ALLOC_HARDER
3305 * users on the grounds that it's definitely going to be in
3306 * the exit path shortly and free memory. Any allocation it
3307 * makes during the free path will be small and short-lived.
3309 if (alloc_flags & ALLOC_OOM)
3317 /* If allocation can't use CMA areas don't use free CMA pages */
3318 if (!(alloc_flags & ALLOC_CMA))
3319 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3323 * Check watermarks for an order-0 allocation request. If these
3324 * are not met, then a high-order request also cannot go ahead
3325 * even if a suitable page happened to be free.
3327 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3330 /* If this is an order-0 request then the watermark is fine */
3334 /* For a high-order request, check at least one suitable page is free */
3335 for (o = order; o < MAX_ORDER; o++) {
3336 struct free_area *area = &z->free_area[o];
3342 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3343 if (!list_empty(&area->free_list[mt]))
3348 if ((alloc_flags & ALLOC_CMA) &&
3349 !list_empty(&area->free_list[MIGRATE_CMA])) {
3354 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3360 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3361 int classzone_idx, unsigned int alloc_flags)
3363 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3364 zone_page_state(z, NR_FREE_PAGES));
3367 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3368 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3370 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3374 /* If allocation can't use CMA areas don't use free CMA pages */
3375 if (!(alloc_flags & ALLOC_CMA))
3376 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3380 * Fast check for order-0 only. If this fails then the reserves
3381 * need to be calculated. There is a corner case where the check
3382 * passes but only the high-order atomic reserve are free. If
3383 * the caller is !atomic then it'll uselessly search the free
3384 * list. That corner case is then slower but it is harmless.
3386 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3389 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3393 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3394 unsigned long mark, int classzone_idx)
3396 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3398 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3399 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3401 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3406 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3408 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3411 #else /* CONFIG_NUMA */
3412 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3416 #endif /* CONFIG_NUMA */
3419 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3420 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3421 * premature use of a lower zone may cause lowmem pressure problems that
3422 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3423 * probably too small. It only makes sense to spread allocations to avoid
3424 * fragmentation between the Normal and DMA32 zones.
3426 static inline unsigned int
3427 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3429 unsigned int alloc_flags = 0;
3431 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3432 alloc_flags |= ALLOC_KSWAPD;
3434 #ifdef CONFIG_ZONE_DMA32
3438 if (zone_idx(zone) != ZONE_NORMAL)
3442 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3443 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3444 * on UMA that if Normal is populated then so is DMA32.
3446 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3447 if (nr_online_nodes > 1 && !populated_zone(--zone))
3450 alloc_flags |= ALLOC_NOFRAGMENT;
3451 #endif /* CONFIG_ZONE_DMA32 */
3456 * get_page_from_freelist goes through the zonelist trying to allocate
3459 static struct page *
3460 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3461 const struct alloc_context *ac)
3465 struct pglist_data *last_pgdat_dirty_limit = NULL;
3470 * Scan zonelist, looking for a zone with enough free.
3471 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3473 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3474 z = ac->preferred_zoneref;
3475 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3480 if (cpusets_enabled() &&
3481 (alloc_flags & ALLOC_CPUSET) &&
3482 !__cpuset_zone_allowed(zone, gfp_mask))
3485 * When allocating a page cache page for writing, we
3486 * want to get it from a node that is within its dirty
3487 * limit, such that no single node holds more than its
3488 * proportional share of globally allowed dirty pages.
3489 * The dirty limits take into account the node's
3490 * lowmem reserves and high watermark so that kswapd
3491 * should be able to balance it without having to
3492 * write pages from its LRU list.
3494 * XXX: For now, allow allocations to potentially
3495 * exceed the per-node dirty limit in the slowpath
3496 * (spread_dirty_pages unset) before going into reclaim,
3497 * which is important when on a NUMA setup the allowed
3498 * nodes are together not big enough to reach the
3499 * global limit. The proper fix for these situations
3500 * will require awareness of nodes in the
3501 * dirty-throttling and the flusher threads.
3503 if (ac->spread_dirty_pages) {
3504 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3507 if (!node_dirty_ok(zone->zone_pgdat)) {
3508 last_pgdat_dirty_limit = zone->zone_pgdat;
3513 if (no_fallback && nr_online_nodes > 1 &&
3514 zone != ac->preferred_zoneref->zone) {
3518 * If moving to a remote node, retry but allow
3519 * fragmenting fallbacks. Locality is more important
3520 * than fragmentation avoidance.
3522 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3523 if (zone_to_nid(zone) != local_nid) {
3524 alloc_flags &= ~ALLOC_NOFRAGMENT;
3529 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3530 if (!zone_watermark_fast(zone, order, mark,
3531 ac_classzone_idx(ac), alloc_flags)) {
3534 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3536 * Watermark failed for this zone, but see if we can
3537 * grow this zone if it contains deferred pages.
3539 if (static_branch_unlikely(&deferred_pages)) {
3540 if (_deferred_grow_zone(zone, order))
3544 /* Checked here to keep the fast path fast */
3545 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3546 if (alloc_flags & ALLOC_NO_WATERMARKS)
3549 if (node_reclaim_mode == 0 ||
3550 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3553 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3555 case NODE_RECLAIM_NOSCAN:
3558 case NODE_RECLAIM_FULL:
3559 /* scanned but unreclaimable */
3562 /* did we reclaim enough */
3563 if (zone_watermark_ok(zone, order, mark,
3564 ac_classzone_idx(ac), alloc_flags))
3572 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3573 gfp_mask, alloc_flags, ac->migratetype);
3575 prep_new_page(page, order, gfp_mask, alloc_flags);
3578 * If this is a high-order atomic allocation then check
3579 * if the pageblock should be reserved for the future
3581 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3582 reserve_highatomic_pageblock(page, zone, order);
3586 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3587 /* Try again if zone has deferred pages */
3588 if (static_branch_unlikely(&deferred_pages)) {
3589 if (_deferred_grow_zone(zone, order))
3597 * It's possible on a UMA machine to get through all zones that are
3598 * fragmented. If avoiding fragmentation, reset and try again.
3601 alloc_flags &= ~ALLOC_NOFRAGMENT;
3608 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3610 unsigned int filter = SHOW_MEM_FILTER_NODES;
3611 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3613 if (!__ratelimit(&show_mem_rs))
3617 * This documents exceptions given to allocations in certain
3618 * contexts that are allowed to allocate outside current's set
3621 if (!(gfp_mask & __GFP_NOMEMALLOC))
3622 if (tsk_is_oom_victim(current) ||
3623 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3624 filter &= ~SHOW_MEM_FILTER_NODES;
3625 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3626 filter &= ~SHOW_MEM_FILTER_NODES;
3628 show_mem(filter, nodemask);
3631 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3633 struct va_format vaf;
3635 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3636 DEFAULT_RATELIMIT_BURST);
3638 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3641 va_start(args, fmt);
3644 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3645 current->comm, &vaf, gfp_mask, &gfp_mask,
3646 nodemask_pr_args(nodemask));
3649 cpuset_print_current_mems_allowed();
3652 warn_alloc_show_mem(gfp_mask, nodemask);
3655 static inline struct page *
3656 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3657 unsigned int alloc_flags,
3658 const struct alloc_context *ac)
3662 page = get_page_from_freelist(gfp_mask, order,
3663 alloc_flags|ALLOC_CPUSET, ac);
3665 * fallback to ignore cpuset restriction if our nodes
3669 page = get_page_from_freelist(gfp_mask, order,
3675 static inline struct page *
3676 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3677 const struct alloc_context *ac, unsigned long *did_some_progress)
3679 struct oom_control oc = {
3680 .zonelist = ac->zonelist,
3681 .nodemask = ac->nodemask,
3683 .gfp_mask = gfp_mask,
3688 *did_some_progress = 0;
3691 * Acquire the oom lock. If that fails, somebody else is
3692 * making progress for us.
3694 if (!mutex_trylock(&oom_lock)) {
3695 *did_some_progress = 1;
3696 schedule_timeout_uninterruptible(1);
3701 * Go through the zonelist yet one more time, keep very high watermark
3702 * here, this is only to catch a parallel oom killing, we must fail if
3703 * we're still under heavy pressure. But make sure that this reclaim
3704 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3705 * allocation which will never fail due to oom_lock already held.
3707 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3708 ~__GFP_DIRECT_RECLAIM, order,
3709 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3713 /* Coredumps can quickly deplete all memory reserves */
3714 if (current->flags & PF_DUMPCORE)
3716 /* The OOM killer will not help higher order allocs */
3717 if (order > PAGE_ALLOC_COSTLY_ORDER)
3720 * We have already exhausted all our reclaim opportunities without any
3721 * success so it is time to admit defeat. We will skip the OOM killer
3722 * because it is very likely that the caller has a more reasonable
3723 * fallback than shooting a random task.
3725 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3727 /* The OOM killer does not needlessly kill tasks for lowmem */
3728 if (ac->high_zoneidx < ZONE_NORMAL)
3730 if (pm_suspended_storage())
3733 * XXX: GFP_NOFS allocations should rather fail than rely on
3734 * other request to make a forward progress.
3735 * We are in an unfortunate situation where out_of_memory cannot
3736 * do much for this context but let's try it to at least get
3737 * access to memory reserved if the current task is killed (see
3738 * out_of_memory). Once filesystems are ready to handle allocation
3739 * failures more gracefully we should just bail out here.
3742 /* The OOM killer may not free memory on a specific node */
3743 if (gfp_mask & __GFP_THISNODE)
3746 /* Exhausted what can be done so it's blame time */
3747 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3748 *did_some_progress = 1;
3751 * Help non-failing allocations by giving them access to memory
3754 if (gfp_mask & __GFP_NOFAIL)
3755 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3756 ALLOC_NO_WATERMARKS, ac);
3759 mutex_unlock(&oom_lock);
3764 * Maximum number of compaction retries wit a progress before OOM
3765 * killer is consider as the only way to move forward.
3767 #define MAX_COMPACT_RETRIES 16
3769 #ifdef CONFIG_COMPACTION
3770 /* Try memory compaction for high-order allocations before reclaim */
3771 static struct page *
3772 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3773 unsigned int alloc_flags, const struct alloc_context *ac,
3774 enum compact_priority prio, enum compact_result *compact_result)
3776 struct page *page = NULL;
3777 unsigned long pflags;
3778 unsigned int noreclaim_flag;
3783 psi_memstall_enter(&pflags);
3784 noreclaim_flag = memalloc_noreclaim_save();
3786 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3789 memalloc_noreclaim_restore(noreclaim_flag);
3790 psi_memstall_leave(&pflags);
3793 * At least in one zone compaction wasn't deferred or skipped, so let's
3794 * count a compaction stall
3796 count_vm_event(COMPACTSTALL);
3798 /* Prep a captured page if available */
3800 prep_new_page(page, order, gfp_mask, alloc_flags);
3802 /* Try get a page from the freelist if available */
3804 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3807 struct zone *zone = page_zone(page);
3809 zone->compact_blockskip_flush = false;
3810 compaction_defer_reset(zone, order, true);
3811 count_vm_event(COMPACTSUCCESS);
3816 * It's bad if compaction run occurs and fails. The most likely reason
3817 * is that pages exist, but not enough to satisfy watermarks.
3819 count_vm_event(COMPACTFAIL);
3827 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3828 enum compact_result compact_result,
3829 enum compact_priority *compact_priority,
3830 int *compaction_retries)
3832 int max_retries = MAX_COMPACT_RETRIES;
3835 int retries = *compaction_retries;
3836 enum compact_priority priority = *compact_priority;
3841 if (compaction_made_progress(compact_result))
3842 (*compaction_retries)++;
3845 * compaction considers all the zone as desperately out of memory
3846 * so it doesn't really make much sense to retry except when the
3847 * failure could be caused by insufficient priority
3849 if (compaction_failed(compact_result))
3850 goto check_priority;
3853 * make sure the compaction wasn't deferred or didn't bail out early
3854 * due to locks contention before we declare that we should give up.
3855 * But do not retry if the given zonelist is not suitable for
3858 if (compaction_withdrawn(compact_result)) {
3859 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3864 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3865 * costly ones because they are de facto nofail and invoke OOM
3866 * killer to move on while costly can fail and users are ready
3867 * to cope with that. 1/4 retries is rather arbitrary but we
3868 * would need much more detailed feedback from compaction to
3869 * make a better decision.
3871 if (order > PAGE_ALLOC_COSTLY_ORDER)
3873 if (*compaction_retries <= max_retries) {
3879 * Make sure there are attempts at the highest priority if we exhausted
3880 * all retries or failed at the lower priorities.
3883 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3884 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3886 if (*compact_priority > min_priority) {
3887 (*compact_priority)--;
3888 *compaction_retries = 0;
3892 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3896 static inline struct page *
3897 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3898 unsigned int alloc_flags, const struct alloc_context *ac,
3899 enum compact_priority prio, enum compact_result *compact_result)
3901 *compact_result = COMPACT_SKIPPED;
3906 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3907 enum compact_result compact_result,
3908 enum compact_priority *compact_priority,
3909 int *compaction_retries)
3914 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3918 * There are setups with compaction disabled which would prefer to loop
3919 * inside the allocator rather than hit the oom killer prematurely.
3920 * Let's give them a good hope and keep retrying while the order-0
3921 * watermarks are OK.
3923 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3925 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3926 ac_classzone_idx(ac), alloc_flags))
3931 #endif /* CONFIG_COMPACTION */
3933 #ifdef CONFIG_LOCKDEP
3934 static struct lockdep_map __fs_reclaim_map =
3935 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3937 static bool __need_fs_reclaim(gfp_t gfp_mask)
3939 gfp_mask = current_gfp_context(gfp_mask);
3941 /* no reclaim without waiting on it */
3942 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3945 /* this guy won't enter reclaim */
3946 if (current->flags & PF_MEMALLOC)
3949 /* We're only interested __GFP_FS allocations for now */
3950 if (!(gfp_mask & __GFP_FS))
3953 if (gfp_mask & __GFP_NOLOCKDEP)
3959 void __fs_reclaim_acquire(void)
3961 lock_map_acquire(&__fs_reclaim_map);
3964 void __fs_reclaim_release(void)
3966 lock_map_release(&__fs_reclaim_map);
3969 void fs_reclaim_acquire(gfp_t gfp_mask)
3971 if (__need_fs_reclaim(gfp_mask))
3972 __fs_reclaim_acquire();
3974 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3976 void fs_reclaim_release(gfp_t gfp_mask)
3978 if (__need_fs_reclaim(gfp_mask))
3979 __fs_reclaim_release();
3981 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3984 /* Perform direct synchronous page reclaim */
3986 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3987 const struct alloc_context *ac)
3989 struct reclaim_state reclaim_state;
3991 unsigned int noreclaim_flag;
3992 unsigned long pflags;
3996 /* We now go into synchronous reclaim */
3997 cpuset_memory_pressure_bump();
3998 psi_memstall_enter(&pflags);
3999 fs_reclaim_acquire(gfp_mask);
4000 noreclaim_flag = memalloc_noreclaim_save();
4001 reclaim_state.reclaimed_slab = 0;
4002 current->reclaim_state = &reclaim_state;
4004 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4007 current->reclaim_state = NULL;
4008 memalloc_noreclaim_restore(noreclaim_flag);
4009 fs_reclaim_release(gfp_mask);
4010 psi_memstall_leave(&pflags);
4017 /* The really slow allocator path where we enter direct reclaim */
4018 static inline struct page *
4019 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4020 unsigned int alloc_flags, const struct alloc_context *ac,
4021 unsigned long *did_some_progress)
4023 struct page *page = NULL;
4024 bool drained = false;
4026 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4027 if (unlikely(!(*did_some_progress)))
4031 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4034 * If an allocation failed after direct reclaim, it could be because
4035 * pages are pinned on the per-cpu lists or in high alloc reserves.
4036 * Shrink them them and try again
4038 if (!page && !drained) {
4039 unreserve_highatomic_pageblock(ac, false);
4040 drain_all_pages(NULL);
4048 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4049 const struct alloc_context *ac)
4053 pg_data_t *last_pgdat = NULL;
4054 enum zone_type high_zoneidx = ac->high_zoneidx;
4056 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4058 if (last_pgdat != zone->zone_pgdat)
4059 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4060 last_pgdat = zone->zone_pgdat;
4064 static inline unsigned int
4065 gfp_to_alloc_flags(gfp_t gfp_mask)
4067 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4069 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4070 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4073 * The caller may dip into page reserves a bit more if the caller
4074 * cannot run direct reclaim, or if the caller has realtime scheduling
4075 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4076 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4078 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4080 if (gfp_mask & __GFP_ATOMIC) {
4082 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4083 * if it can't schedule.
4085 if (!(gfp_mask & __GFP_NOMEMALLOC))
4086 alloc_flags |= ALLOC_HARDER;
4088 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4089 * comment for __cpuset_node_allowed().
4091 alloc_flags &= ~ALLOC_CPUSET;
4092 } else if (unlikely(rt_task(current)) && !in_interrupt())
4093 alloc_flags |= ALLOC_HARDER;
4095 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4096 alloc_flags |= ALLOC_KSWAPD;
4099 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4100 alloc_flags |= ALLOC_CMA;
4105 static bool oom_reserves_allowed(struct task_struct *tsk)
4107 if (!tsk_is_oom_victim(tsk))
4111 * !MMU doesn't have oom reaper so give access to memory reserves
4112 * only to the thread with TIF_MEMDIE set
4114 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4121 * Distinguish requests which really need access to full memory
4122 * reserves from oom victims which can live with a portion of it
4124 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4126 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4128 if (gfp_mask & __GFP_MEMALLOC)
4129 return ALLOC_NO_WATERMARKS;
4130 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4131 return ALLOC_NO_WATERMARKS;
4132 if (!in_interrupt()) {
4133 if (current->flags & PF_MEMALLOC)
4134 return ALLOC_NO_WATERMARKS;
4135 else if (oom_reserves_allowed(current))
4142 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4144 return !!__gfp_pfmemalloc_flags(gfp_mask);
4148 * Checks whether it makes sense to retry the reclaim to make a forward progress
4149 * for the given allocation request.
4151 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4152 * without success, or when we couldn't even meet the watermark if we
4153 * reclaimed all remaining pages on the LRU lists.
4155 * Returns true if a retry is viable or false to enter the oom path.
4158 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4159 struct alloc_context *ac, int alloc_flags,
4160 bool did_some_progress, int *no_progress_loops)
4167 * Costly allocations might have made a progress but this doesn't mean
4168 * their order will become available due to high fragmentation so
4169 * always increment the no progress counter for them
4171 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4172 *no_progress_loops = 0;
4174 (*no_progress_loops)++;
4177 * Make sure we converge to OOM if we cannot make any progress
4178 * several times in the row.
4180 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4181 /* Before OOM, exhaust highatomic_reserve */
4182 return unreserve_highatomic_pageblock(ac, true);
4186 * Keep reclaiming pages while there is a chance this will lead
4187 * somewhere. If none of the target zones can satisfy our allocation
4188 * request even if all reclaimable pages are considered then we are
4189 * screwed and have to go OOM.
4191 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4193 unsigned long available;
4194 unsigned long reclaimable;
4195 unsigned long min_wmark = min_wmark_pages(zone);
4198 available = reclaimable = zone_reclaimable_pages(zone);
4199 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4202 * Would the allocation succeed if we reclaimed all
4203 * reclaimable pages?
4205 wmark = __zone_watermark_ok(zone, order, min_wmark,
4206 ac_classzone_idx(ac), alloc_flags, available);
4207 trace_reclaim_retry_zone(z, order, reclaimable,
4208 available, min_wmark, *no_progress_loops, wmark);
4211 * If we didn't make any progress and have a lot of
4212 * dirty + writeback pages then we should wait for
4213 * an IO to complete to slow down the reclaim and
4214 * prevent from pre mature OOM
4216 if (!did_some_progress) {
4217 unsigned long write_pending;
4219 write_pending = zone_page_state_snapshot(zone,
4220 NR_ZONE_WRITE_PENDING);
4222 if (2 * write_pending > reclaimable) {
4223 congestion_wait(BLK_RW_ASYNC, HZ/10);
4235 * Memory allocation/reclaim might be called from a WQ context and the
4236 * current implementation of the WQ concurrency control doesn't
4237 * recognize that a particular WQ is congested if the worker thread is
4238 * looping without ever sleeping. Therefore we have to do a short sleep
4239 * here rather than calling cond_resched().
4241 if (current->flags & PF_WQ_WORKER)
4242 schedule_timeout_uninterruptible(1);
4249 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4252 * It's possible that cpuset's mems_allowed and the nodemask from
4253 * mempolicy don't intersect. This should be normally dealt with by
4254 * policy_nodemask(), but it's possible to race with cpuset update in
4255 * such a way the check therein was true, and then it became false
4256 * before we got our cpuset_mems_cookie here.
4257 * This assumes that for all allocations, ac->nodemask can come only
4258 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4259 * when it does not intersect with the cpuset restrictions) or the
4260 * caller can deal with a violated nodemask.
4262 if (cpusets_enabled() && ac->nodemask &&
4263 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4264 ac->nodemask = NULL;
4269 * When updating a task's mems_allowed or mempolicy nodemask, it is
4270 * possible to race with parallel threads in such a way that our
4271 * allocation can fail while the mask is being updated. If we are about
4272 * to fail, check if the cpuset changed during allocation and if so,
4275 if (read_mems_allowed_retry(cpuset_mems_cookie))
4281 static inline struct page *
4282 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4283 struct alloc_context *ac)
4285 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4286 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4287 struct page *page = NULL;
4288 unsigned int alloc_flags;
4289 unsigned long did_some_progress;
4290 enum compact_priority compact_priority;
4291 enum compact_result compact_result;
4292 int compaction_retries;
4293 int no_progress_loops;
4294 unsigned int cpuset_mems_cookie;
4298 * We also sanity check to catch abuse of atomic reserves being used by
4299 * callers that are not in atomic context.
4301 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4302 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4303 gfp_mask &= ~__GFP_ATOMIC;
4306 compaction_retries = 0;
4307 no_progress_loops = 0;
4308 compact_priority = DEF_COMPACT_PRIORITY;
4309 cpuset_mems_cookie = read_mems_allowed_begin();
4312 * The fast path uses conservative alloc_flags to succeed only until
4313 * kswapd needs to be woken up, and to avoid the cost of setting up
4314 * alloc_flags precisely. So we do that now.
4316 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4319 * We need to recalculate the starting point for the zonelist iterator
4320 * because we might have used different nodemask in the fast path, or
4321 * there was a cpuset modification and we are retrying - otherwise we
4322 * could end up iterating over non-eligible zones endlessly.
4324 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4325 ac->high_zoneidx, ac->nodemask);
4326 if (!ac->preferred_zoneref->zone)
4329 if (alloc_flags & ALLOC_KSWAPD)
4330 wake_all_kswapds(order, gfp_mask, ac);
4333 * The adjusted alloc_flags might result in immediate success, so try
4336 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4341 * For costly allocations, try direct compaction first, as it's likely
4342 * that we have enough base pages and don't need to reclaim. For non-
4343 * movable high-order allocations, do that as well, as compaction will
4344 * try prevent permanent fragmentation by migrating from blocks of the
4346 * Don't try this for allocations that are allowed to ignore
4347 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4349 if (can_direct_reclaim &&
4351 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4352 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4353 page = __alloc_pages_direct_compact(gfp_mask, order,
4355 INIT_COMPACT_PRIORITY,
4361 * Checks for costly allocations with __GFP_NORETRY, which
4362 * includes THP page fault allocations
4364 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4366 * If compaction is deferred for high-order allocations,
4367 * it is because sync compaction recently failed. If
4368 * this is the case and the caller requested a THP
4369 * allocation, we do not want to heavily disrupt the
4370 * system, so we fail the allocation instead of entering
4373 if (compact_result == COMPACT_DEFERRED)
4377 * Looks like reclaim/compaction is worth trying, but
4378 * sync compaction could be very expensive, so keep
4379 * using async compaction.
4381 compact_priority = INIT_COMPACT_PRIORITY;
4386 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4387 if (alloc_flags & ALLOC_KSWAPD)
4388 wake_all_kswapds(order, gfp_mask, ac);
4390 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4392 alloc_flags = reserve_flags;
4395 * Reset the nodemask and zonelist iterators if memory policies can be
4396 * ignored. These allocations are high priority and system rather than
4399 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4400 ac->nodemask = NULL;
4401 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4402 ac->high_zoneidx, ac->nodemask);
4405 /* Attempt with potentially adjusted zonelist and alloc_flags */
4406 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4410 /* Caller is not willing to reclaim, we can't balance anything */
4411 if (!can_direct_reclaim)
4414 /* Avoid recursion of direct reclaim */
4415 if (current->flags & PF_MEMALLOC)
4418 /* Try direct reclaim and then allocating */
4419 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4420 &did_some_progress);
4424 /* Try direct compaction and then allocating */
4425 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4426 compact_priority, &compact_result);
4430 /* Do not loop if specifically requested */
4431 if (gfp_mask & __GFP_NORETRY)
4435 * Do not retry costly high order allocations unless they are
4436 * __GFP_RETRY_MAYFAIL
4438 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4441 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4442 did_some_progress > 0, &no_progress_loops))
4446 * It doesn't make any sense to retry for the compaction if the order-0
4447 * reclaim is not able to make any progress because the current
4448 * implementation of the compaction depends on the sufficient amount
4449 * of free memory (see __compaction_suitable)
4451 if (did_some_progress > 0 &&
4452 should_compact_retry(ac, order, alloc_flags,
4453 compact_result, &compact_priority,
4454 &compaction_retries))
4458 /* Deal with possible cpuset update races before we start OOM killing */
4459 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4462 /* Reclaim has failed us, start killing things */
4463 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4467 /* Avoid allocations with no watermarks from looping endlessly */
4468 if (tsk_is_oom_victim(current) &&
4469 (alloc_flags == ALLOC_OOM ||
4470 (gfp_mask & __GFP_NOMEMALLOC)))
4473 /* Retry as long as the OOM killer is making progress */
4474 if (did_some_progress) {
4475 no_progress_loops = 0;
4480 /* Deal with possible cpuset update races before we fail */
4481 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4485 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4488 if (gfp_mask & __GFP_NOFAIL) {
4490 * All existing users of the __GFP_NOFAIL are blockable, so warn
4491 * of any new users that actually require GFP_NOWAIT
4493 if (WARN_ON_ONCE(!can_direct_reclaim))
4497 * PF_MEMALLOC request from this context is rather bizarre
4498 * because we cannot reclaim anything and only can loop waiting
4499 * for somebody to do a work for us
4501 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4504 * non failing costly orders are a hard requirement which we
4505 * are not prepared for much so let's warn about these users
4506 * so that we can identify them and convert them to something
4509 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4512 * Help non-failing allocations by giving them access to memory
4513 * reserves but do not use ALLOC_NO_WATERMARKS because this
4514 * could deplete whole memory reserves which would just make
4515 * the situation worse
4517 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4525 warn_alloc(gfp_mask, ac->nodemask,
4526 "page allocation failure: order:%u", order);
4531 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4532 int preferred_nid, nodemask_t *nodemask,
4533 struct alloc_context *ac, gfp_t *alloc_mask,
4534 unsigned int *alloc_flags)
4536 ac->high_zoneidx = gfp_zone(gfp_mask);
4537 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4538 ac->nodemask = nodemask;
4539 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4541 if (cpusets_enabled()) {
4542 *alloc_mask |= __GFP_HARDWALL;
4544 ac->nodemask = &cpuset_current_mems_allowed;
4546 *alloc_flags |= ALLOC_CPUSET;
4549 fs_reclaim_acquire(gfp_mask);
4550 fs_reclaim_release(gfp_mask);
4552 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4554 if (should_fail_alloc_page(gfp_mask, order))
4557 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4558 *alloc_flags |= ALLOC_CMA;
4563 /* Determine whether to spread dirty pages and what the first usable zone */
4564 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4566 /* Dirty zone balancing only done in the fast path */
4567 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4570 * The preferred zone is used for statistics but crucially it is
4571 * also used as the starting point for the zonelist iterator. It
4572 * may get reset for allocations that ignore memory policies.
4574 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4575 ac->high_zoneidx, ac->nodemask);
4579 * This is the 'heart' of the zoned buddy allocator.
4582 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4583 nodemask_t *nodemask)
4586 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4587 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4588 struct alloc_context ac = { };
4591 * There are several places where we assume that the order value is sane
4592 * so bail out early if the request is out of bound.
4594 if (unlikely(order >= MAX_ORDER)) {
4595 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4599 gfp_mask &= gfp_allowed_mask;
4600 alloc_mask = gfp_mask;
4601 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4604 finalise_ac(gfp_mask, &ac);
4607 * Forbid the first pass from falling back to types that fragment
4608 * memory until all local zones are considered.
4610 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4612 /* First allocation attempt */
4613 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4618 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4619 * resp. GFP_NOIO which has to be inherited for all allocation requests
4620 * from a particular context which has been marked by
4621 * memalloc_no{fs,io}_{save,restore}.
4623 alloc_mask = current_gfp_context(gfp_mask);
4624 ac.spread_dirty_pages = false;
4627 * Restore the original nodemask if it was potentially replaced with
4628 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4630 if (unlikely(ac.nodemask != nodemask))
4631 ac.nodemask = nodemask;
4633 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4636 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4637 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4638 __free_pages(page, order);
4642 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4646 EXPORT_SYMBOL(__alloc_pages_nodemask);
4649 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4650 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4651 * you need to access high mem.
4653 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4657 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4660 return (unsigned long) page_address(page);
4662 EXPORT_SYMBOL(__get_free_pages);
4664 unsigned long get_zeroed_page(gfp_t gfp_mask)
4666 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4668 EXPORT_SYMBOL(get_zeroed_page);
4670 static inline void free_the_page(struct page *page, unsigned int order)
4672 if (order == 0) /* Via pcp? */
4673 free_unref_page(page);
4675 __free_pages_ok(page, order);
4678 void __free_pages(struct page *page, unsigned int order)
4680 if (put_page_testzero(page))
4681 free_the_page(page, order);
4683 EXPORT_SYMBOL(__free_pages);
4685 void free_pages(unsigned long addr, unsigned int order)
4688 VM_BUG_ON(!virt_addr_valid((void *)addr));
4689 __free_pages(virt_to_page((void *)addr), order);
4693 EXPORT_SYMBOL(free_pages);
4697 * An arbitrary-length arbitrary-offset area of memory which resides
4698 * within a 0 or higher order page. Multiple fragments within that page
4699 * are individually refcounted, in the page's reference counter.
4701 * The page_frag functions below provide a simple allocation framework for
4702 * page fragments. This is used by the network stack and network device
4703 * drivers to provide a backing region of memory for use as either an
4704 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4706 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4709 struct page *page = NULL;
4710 gfp_t gfp = gfp_mask;
4712 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4713 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4715 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4716 PAGE_FRAG_CACHE_MAX_ORDER);
4717 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4719 if (unlikely(!page))
4720 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4722 nc->va = page ? page_address(page) : NULL;
4727 void __page_frag_cache_drain(struct page *page, unsigned int count)
4729 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4731 if (page_ref_sub_and_test(page, count))
4732 free_the_page(page, compound_order(page));
4734 EXPORT_SYMBOL(__page_frag_cache_drain);
4736 void *page_frag_alloc(struct page_frag_cache *nc,
4737 unsigned int fragsz, gfp_t gfp_mask)
4739 unsigned int size = PAGE_SIZE;
4743 if (unlikely(!nc->va)) {
4745 page = __page_frag_cache_refill(nc, gfp_mask);
4749 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4750 /* if size can vary use size else just use PAGE_SIZE */
4753 /* Even if we own the page, we do not use atomic_set().
4754 * This would break get_page_unless_zero() users.
4756 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4758 /* reset page count bias and offset to start of new frag */
4759 nc->pfmemalloc = page_is_pfmemalloc(page);
4760 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4764 offset = nc->offset - fragsz;
4765 if (unlikely(offset < 0)) {
4766 page = virt_to_page(nc->va);
4768 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4771 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4772 /* if size can vary use size else just use PAGE_SIZE */
4775 /* OK, page count is 0, we can safely set it */
4776 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4778 /* reset page count bias and offset to start of new frag */
4779 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4780 offset = size - fragsz;
4784 nc->offset = offset;
4786 return nc->va + offset;
4788 EXPORT_SYMBOL(page_frag_alloc);
4791 * Frees a page fragment allocated out of either a compound or order 0 page.
4793 void page_frag_free(void *addr)
4795 struct page *page = virt_to_head_page(addr);
4797 if (unlikely(put_page_testzero(page)))
4798 free_the_page(page, compound_order(page));
4800 EXPORT_SYMBOL(page_frag_free);
4802 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4806 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4807 unsigned long used = addr + PAGE_ALIGN(size);
4809 split_page(virt_to_page((void *)addr), order);
4810 while (used < alloc_end) {
4815 return (void *)addr;
4819 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4820 * @size: the number of bytes to allocate
4821 * @gfp_mask: GFP flags for the allocation
4823 * This function is similar to alloc_pages(), except that it allocates the
4824 * minimum number of pages to satisfy the request. alloc_pages() can only
4825 * allocate memory in power-of-two pages.
4827 * This function is also limited by MAX_ORDER.
4829 * Memory allocated by this function must be released by free_pages_exact().
4831 * Return: pointer to the allocated area or %NULL in case of error.
4833 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4835 unsigned int order = get_order(size);
4838 addr = __get_free_pages(gfp_mask, order);
4839 return make_alloc_exact(addr, order, size);
4841 EXPORT_SYMBOL(alloc_pages_exact);
4844 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4846 * @nid: the preferred node ID where memory should be allocated
4847 * @size: the number of bytes to allocate
4848 * @gfp_mask: GFP flags for the allocation
4850 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4853 * Return: pointer to the allocated area or %NULL in case of error.
4855 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4857 unsigned int order = get_order(size);
4858 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4861 return make_alloc_exact((unsigned long)page_address(p), order, size);
4865 * free_pages_exact - release memory allocated via alloc_pages_exact()
4866 * @virt: the value returned by alloc_pages_exact.
4867 * @size: size of allocation, same value as passed to alloc_pages_exact().
4869 * Release the memory allocated by a previous call to alloc_pages_exact.
4871 void free_pages_exact(void *virt, size_t size)
4873 unsigned long addr = (unsigned long)virt;
4874 unsigned long end = addr + PAGE_ALIGN(size);
4876 while (addr < end) {
4881 EXPORT_SYMBOL(free_pages_exact);
4884 * nr_free_zone_pages - count number of pages beyond high watermark
4885 * @offset: The zone index of the highest zone
4887 * nr_free_zone_pages() counts the number of pages which are beyond the
4888 * high watermark within all zones at or below a given zone index. For each
4889 * zone, the number of pages is calculated as:
4891 * nr_free_zone_pages = managed_pages - high_pages
4893 * Return: number of pages beyond high watermark.
4895 static unsigned long nr_free_zone_pages(int offset)
4900 /* Just pick one node, since fallback list is circular */
4901 unsigned long sum = 0;
4903 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4905 for_each_zone_zonelist(zone, z, zonelist, offset) {
4906 unsigned long size = zone_managed_pages(zone);
4907 unsigned long high = high_wmark_pages(zone);
4916 * nr_free_buffer_pages - count number of pages beyond high watermark
4918 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4919 * watermark within ZONE_DMA and ZONE_NORMAL.
4921 * Return: number of pages beyond high watermark within ZONE_DMA and
4924 unsigned long nr_free_buffer_pages(void)
4926 return nr_free_zone_pages(gfp_zone(GFP_USER));
4928 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4931 * nr_free_pagecache_pages - count number of pages beyond high watermark
4933 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4934 * high watermark within all zones.
4936 * Return: number of pages beyond high watermark within all zones.
4938 unsigned long nr_free_pagecache_pages(void)
4940 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4943 static inline void show_node(struct zone *zone)
4945 if (IS_ENABLED(CONFIG_NUMA))
4946 printk("Node %d ", zone_to_nid(zone));
4949 long si_mem_available(void)
4952 unsigned long pagecache;
4953 unsigned long wmark_low = 0;
4954 unsigned long pages[NR_LRU_LISTS];
4955 unsigned long reclaimable;
4959 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4960 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4963 wmark_low += low_wmark_pages(zone);
4966 * Estimate the amount of memory available for userspace allocations,
4967 * without causing swapping.
4969 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4972 * Not all the page cache can be freed, otherwise the system will
4973 * start swapping. Assume at least half of the page cache, or the
4974 * low watermark worth of cache, needs to stay.
4976 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4977 pagecache -= min(pagecache / 2, wmark_low);
4978 available += pagecache;
4981 * Part of the reclaimable slab and other kernel memory consists of
4982 * items that are in use, and cannot be freed. Cap this estimate at the
4985 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
4986 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
4987 available += reclaimable - min(reclaimable / 2, wmark_low);
4993 EXPORT_SYMBOL_GPL(si_mem_available);
4995 void si_meminfo(struct sysinfo *val)
4997 val->totalram = totalram_pages();
4998 val->sharedram = global_node_page_state(NR_SHMEM);
4999 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5000 val->bufferram = nr_blockdev_pages();
5001 val->totalhigh = totalhigh_pages();
5002 val->freehigh = nr_free_highpages();
5003 val->mem_unit = PAGE_SIZE;
5006 EXPORT_SYMBOL(si_meminfo);
5009 void si_meminfo_node(struct sysinfo *val, int nid)
5011 int zone_type; /* needs to be signed */
5012 unsigned long managed_pages = 0;
5013 unsigned long managed_highpages = 0;
5014 unsigned long free_highpages = 0;
5015 pg_data_t *pgdat = NODE_DATA(nid);
5017 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5018 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5019 val->totalram = managed_pages;
5020 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5021 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5022 #ifdef CONFIG_HIGHMEM
5023 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5024 struct zone *zone = &pgdat->node_zones[zone_type];
5026 if (is_highmem(zone)) {
5027 managed_highpages += zone_managed_pages(zone);
5028 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5031 val->totalhigh = managed_highpages;
5032 val->freehigh = free_highpages;
5034 val->totalhigh = managed_highpages;
5035 val->freehigh = free_highpages;
5037 val->mem_unit = PAGE_SIZE;
5042 * Determine whether the node should be displayed or not, depending on whether
5043 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5045 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5047 if (!(flags & SHOW_MEM_FILTER_NODES))
5051 * no node mask - aka implicit memory numa policy. Do not bother with
5052 * the synchronization - read_mems_allowed_begin - because we do not
5053 * have to be precise here.
5056 nodemask = &cpuset_current_mems_allowed;
5058 return !node_isset(nid, *nodemask);
5061 #define K(x) ((x) << (PAGE_SHIFT-10))
5063 static void show_migration_types(unsigned char type)
5065 static const char types[MIGRATE_TYPES] = {
5066 [MIGRATE_UNMOVABLE] = 'U',
5067 [MIGRATE_MOVABLE] = 'M',
5068 [MIGRATE_RECLAIMABLE] = 'E',
5069 [MIGRATE_HIGHATOMIC] = 'H',
5071 [MIGRATE_CMA] = 'C',
5073 #ifdef CONFIG_MEMORY_ISOLATION
5074 [MIGRATE_ISOLATE] = 'I',
5077 char tmp[MIGRATE_TYPES + 1];
5081 for (i = 0; i < MIGRATE_TYPES; i++) {
5082 if (type & (1 << i))
5087 printk(KERN_CONT "(%s) ", tmp);
5091 * Show free area list (used inside shift_scroll-lock stuff)
5092 * We also calculate the percentage fragmentation. We do this by counting the
5093 * memory on each free list with the exception of the first item on the list.
5096 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5099 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5101 unsigned long free_pcp = 0;
5106 for_each_populated_zone(zone) {
5107 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5110 for_each_online_cpu(cpu)
5111 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5114 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5115 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5116 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5117 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5118 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5119 " free:%lu free_pcp:%lu free_cma:%lu\n",
5120 global_node_page_state(NR_ACTIVE_ANON),
5121 global_node_page_state(NR_INACTIVE_ANON),
5122 global_node_page_state(NR_ISOLATED_ANON),
5123 global_node_page_state(NR_ACTIVE_FILE),
5124 global_node_page_state(NR_INACTIVE_FILE),
5125 global_node_page_state(NR_ISOLATED_FILE),
5126 global_node_page_state(NR_UNEVICTABLE),
5127 global_node_page_state(NR_FILE_DIRTY),
5128 global_node_page_state(NR_WRITEBACK),
5129 global_node_page_state(NR_UNSTABLE_NFS),
5130 global_node_page_state(NR_SLAB_RECLAIMABLE),
5131 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5132 global_node_page_state(NR_FILE_MAPPED),
5133 global_node_page_state(NR_SHMEM),
5134 global_zone_page_state(NR_PAGETABLE),
5135 global_zone_page_state(NR_BOUNCE),
5136 global_zone_page_state(NR_FREE_PAGES),
5138 global_zone_page_state(NR_FREE_CMA_PAGES));
5140 for_each_online_pgdat(pgdat) {
5141 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5145 " active_anon:%lukB"
5146 " inactive_anon:%lukB"
5147 " active_file:%lukB"
5148 " inactive_file:%lukB"
5149 " unevictable:%lukB"
5150 " isolated(anon):%lukB"
5151 " isolated(file):%lukB"
5156 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5158 " shmem_pmdmapped: %lukB"
5161 " writeback_tmp:%lukB"
5163 " all_unreclaimable? %s"
5166 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5167 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5168 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5169 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5170 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5171 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5172 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5173 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5174 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5175 K(node_page_state(pgdat, NR_WRITEBACK)),
5176 K(node_page_state(pgdat, NR_SHMEM)),
5177 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5178 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5179 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5181 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5183 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5184 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5185 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5189 for_each_populated_zone(zone) {
5192 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5196 for_each_online_cpu(cpu)
5197 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5206 " active_anon:%lukB"
5207 " inactive_anon:%lukB"
5208 " active_file:%lukB"
5209 " inactive_file:%lukB"
5210 " unevictable:%lukB"
5211 " writepending:%lukB"
5215 " kernel_stack:%lukB"
5223 K(zone_page_state(zone, NR_FREE_PAGES)),
5224 K(min_wmark_pages(zone)),
5225 K(low_wmark_pages(zone)),
5226 K(high_wmark_pages(zone)),
5227 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5228 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5229 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5230 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5231 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5232 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5233 K(zone->present_pages),
5234 K(zone_managed_pages(zone)),
5235 K(zone_page_state(zone, NR_MLOCK)),
5236 zone_page_state(zone, NR_KERNEL_STACK_KB),
5237 K(zone_page_state(zone, NR_PAGETABLE)),
5238 K(zone_page_state(zone, NR_BOUNCE)),
5240 K(this_cpu_read(zone->pageset->pcp.count)),
5241 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5242 printk("lowmem_reserve[]:");
5243 for (i = 0; i < MAX_NR_ZONES; i++)
5244 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5245 printk(KERN_CONT "\n");
5248 for_each_populated_zone(zone) {
5250 unsigned long nr[MAX_ORDER], flags, total = 0;
5251 unsigned char types[MAX_ORDER];
5253 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5256 printk(KERN_CONT "%s: ", zone->name);
5258 spin_lock_irqsave(&zone->lock, flags);
5259 for (order = 0; order < MAX_ORDER; order++) {
5260 struct free_area *area = &zone->free_area[order];
5263 nr[order] = area->nr_free;
5264 total += nr[order] << order;
5267 for (type = 0; type < MIGRATE_TYPES; type++) {
5268 if (!list_empty(&area->free_list[type]))
5269 types[order] |= 1 << type;
5272 spin_unlock_irqrestore(&zone->lock, flags);
5273 for (order = 0; order < MAX_ORDER; order++) {
5274 printk(KERN_CONT "%lu*%lukB ",
5275 nr[order], K(1UL) << order);
5277 show_migration_types(types[order]);
5279 printk(KERN_CONT "= %lukB\n", K(total));
5282 hugetlb_show_meminfo();
5284 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5286 show_swap_cache_info();
5289 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5291 zoneref->zone = zone;
5292 zoneref->zone_idx = zone_idx(zone);
5296 * Builds allocation fallback zone lists.
5298 * Add all populated zones of a node to the zonelist.
5300 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5303 enum zone_type zone_type = MAX_NR_ZONES;
5308 zone = pgdat->node_zones + zone_type;
5309 if (managed_zone(zone)) {
5310 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5311 check_highest_zone(zone_type);
5313 } while (zone_type);
5320 static int __parse_numa_zonelist_order(char *s)
5323 * We used to support different zonlists modes but they turned
5324 * out to be just not useful. Let's keep the warning in place
5325 * if somebody still use the cmd line parameter so that we do
5326 * not fail it silently
5328 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5329 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5335 static __init int setup_numa_zonelist_order(char *s)
5340 return __parse_numa_zonelist_order(s);
5342 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5344 char numa_zonelist_order[] = "Node";
5347 * sysctl handler for numa_zonelist_order
5349 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5350 void __user *buffer, size_t *length,
5357 return proc_dostring(table, write, buffer, length, ppos);
5358 str = memdup_user_nul(buffer, 16);
5360 return PTR_ERR(str);
5362 ret = __parse_numa_zonelist_order(str);
5368 #define MAX_NODE_LOAD (nr_online_nodes)
5369 static int node_load[MAX_NUMNODES];
5372 * find_next_best_node - find the next node that should appear in a given node's fallback list
5373 * @node: node whose fallback list we're appending
5374 * @used_node_mask: nodemask_t of already used nodes
5376 * We use a number of factors to determine which is the next node that should
5377 * appear on a given node's fallback list. The node should not have appeared
5378 * already in @node's fallback list, and it should be the next closest node
5379 * according to the distance array (which contains arbitrary distance values
5380 * from each node to each node in the system), and should also prefer nodes
5381 * with no CPUs, since presumably they'll have very little allocation pressure
5382 * on them otherwise.
5384 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5386 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5389 int min_val = INT_MAX;
5390 int best_node = NUMA_NO_NODE;
5391 const struct cpumask *tmp = cpumask_of_node(0);
5393 /* Use the local node if we haven't already */
5394 if (!node_isset(node, *used_node_mask)) {
5395 node_set(node, *used_node_mask);
5399 for_each_node_state(n, N_MEMORY) {
5401 /* Don't want a node to appear more than once */
5402 if (node_isset(n, *used_node_mask))
5405 /* Use the distance array to find the distance */
5406 val = node_distance(node, n);
5408 /* Penalize nodes under us ("prefer the next node") */
5411 /* Give preference to headless and unused nodes */
5412 tmp = cpumask_of_node(n);
5413 if (!cpumask_empty(tmp))
5414 val += PENALTY_FOR_NODE_WITH_CPUS;
5416 /* Slight preference for less loaded node */
5417 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5418 val += node_load[n];
5420 if (val < min_val) {
5427 node_set(best_node, *used_node_mask);
5434 * Build zonelists ordered by node and zones within node.
5435 * This results in maximum locality--normal zone overflows into local
5436 * DMA zone, if any--but risks exhausting DMA zone.
5438 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5441 struct zoneref *zonerefs;
5444 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5446 for (i = 0; i < nr_nodes; i++) {
5449 pg_data_t *node = NODE_DATA(node_order[i]);
5451 nr_zones = build_zonerefs_node(node, zonerefs);
5452 zonerefs += nr_zones;
5454 zonerefs->zone = NULL;
5455 zonerefs->zone_idx = 0;
5459 * Build gfp_thisnode zonelists
5461 static void build_thisnode_zonelists(pg_data_t *pgdat)
5463 struct zoneref *zonerefs;
5466 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5467 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5468 zonerefs += nr_zones;
5469 zonerefs->zone = NULL;
5470 zonerefs->zone_idx = 0;
5474 * Build zonelists ordered by zone and nodes within zones.
5475 * This results in conserving DMA zone[s] until all Normal memory is
5476 * exhausted, but results in overflowing to remote node while memory
5477 * may still exist in local DMA zone.
5480 static void build_zonelists(pg_data_t *pgdat)
5482 static int node_order[MAX_NUMNODES];
5483 int node, load, nr_nodes = 0;
5484 nodemask_t used_mask;
5485 int local_node, prev_node;
5487 /* NUMA-aware ordering of nodes */
5488 local_node = pgdat->node_id;
5489 load = nr_online_nodes;
5490 prev_node = local_node;
5491 nodes_clear(used_mask);
5493 memset(node_order, 0, sizeof(node_order));
5494 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5496 * We don't want to pressure a particular node.
5497 * So adding penalty to the first node in same
5498 * distance group to make it round-robin.
5500 if (node_distance(local_node, node) !=
5501 node_distance(local_node, prev_node))
5502 node_load[node] = load;
5504 node_order[nr_nodes++] = node;
5509 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5510 build_thisnode_zonelists(pgdat);
5513 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5515 * Return node id of node used for "local" allocations.
5516 * I.e., first node id of first zone in arg node's generic zonelist.
5517 * Used for initializing percpu 'numa_mem', which is used primarily
5518 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5520 int local_memory_node(int node)
5524 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5525 gfp_zone(GFP_KERNEL),
5527 return zone_to_nid(z->zone);
5531 static void setup_min_unmapped_ratio(void);
5532 static void setup_min_slab_ratio(void);
5533 #else /* CONFIG_NUMA */
5535 static void build_zonelists(pg_data_t *pgdat)
5537 int node, local_node;
5538 struct zoneref *zonerefs;
5541 local_node = pgdat->node_id;
5543 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5544 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5545 zonerefs += nr_zones;
5548 * Now we build the zonelist so that it contains the zones
5549 * of all the other nodes.
5550 * We don't want to pressure a particular node, so when
5551 * building the zones for node N, we make sure that the
5552 * zones coming right after the local ones are those from
5553 * node N+1 (modulo N)
5555 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5556 if (!node_online(node))
5558 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5559 zonerefs += nr_zones;
5561 for (node = 0; node < local_node; node++) {
5562 if (!node_online(node))
5564 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5565 zonerefs += nr_zones;
5568 zonerefs->zone = NULL;
5569 zonerefs->zone_idx = 0;
5572 #endif /* CONFIG_NUMA */
5575 * Boot pageset table. One per cpu which is going to be used for all
5576 * zones and all nodes. The parameters will be set in such a way
5577 * that an item put on a list will immediately be handed over to
5578 * the buddy list. This is safe since pageset manipulation is done
5579 * with interrupts disabled.
5581 * The boot_pagesets must be kept even after bootup is complete for
5582 * unused processors and/or zones. They do play a role for bootstrapping
5583 * hotplugged processors.
5585 * zoneinfo_show() and maybe other functions do
5586 * not check if the processor is online before following the pageset pointer.
5587 * Other parts of the kernel may not check if the zone is available.
5589 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5590 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5591 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5593 static void __build_all_zonelists(void *data)
5596 int __maybe_unused cpu;
5597 pg_data_t *self = data;
5598 static DEFINE_SPINLOCK(lock);
5603 memset(node_load, 0, sizeof(node_load));
5607 * This node is hotadded and no memory is yet present. So just
5608 * building zonelists is fine - no need to touch other nodes.
5610 if (self && !node_online(self->node_id)) {
5611 build_zonelists(self);
5613 for_each_online_node(nid) {
5614 pg_data_t *pgdat = NODE_DATA(nid);
5616 build_zonelists(pgdat);
5619 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5621 * We now know the "local memory node" for each node--
5622 * i.e., the node of the first zone in the generic zonelist.
5623 * Set up numa_mem percpu variable for on-line cpus. During
5624 * boot, only the boot cpu should be on-line; we'll init the
5625 * secondary cpus' numa_mem as they come on-line. During
5626 * node/memory hotplug, we'll fixup all on-line cpus.
5628 for_each_online_cpu(cpu)
5629 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5636 static noinline void __init
5637 build_all_zonelists_init(void)
5641 __build_all_zonelists(NULL);
5644 * Initialize the boot_pagesets that are going to be used
5645 * for bootstrapping processors. The real pagesets for
5646 * each zone will be allocated later when the per cpu
5647 * allocator is available.
5649 * boot_pagesets are used also for bootstrapping offline
5650 * cpus if the system is already booted because the pagesets
5651 * are needed to initialize allocators on a specific cpu too.
5652 * F.e. the percpu allocator needs the page allocator which
5653 * needs the percpu allocator in order to allocate its pagesets
5654 * (a chicken-egg dilemma).
5656 for_each_possible_cpu(cpu)
5657 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5659 mminit_verify_zonelist();
5660 cpuset_init_current_mems_allowed();
5664 * unless system_state == SYSTEM_BOOTING.
5666 * __ref due to call of __init annotated helper build_all_zonelists_init
5667 * [protected by SYSTEM_BOOTING].
5669 void __ref build_all_zonelists(pg_data_t *pgdat)
5671 if (system_state == SYSTEM_BOOTING) {
5672 build_all_zonelists_init();
5674 __build_all_zonelists(pgdat);
5675 /* cpuset refresh routine should be here */
5677 vm_total_pages = nr_free_pagecache_pages();
5679 * Disable grouping by mobility if the number of pages in the
5680 * system is too low to allow the mechanism to work. It would be
5681 * more accurate, but expensive to check per-zone. This check is
5682 * made on memory-hotadd so a system can start with mobility
5683 * disabled and enable it later
5685 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5686 page_group_by_mobility_disabled = 1;
5688 page_group_by_mobility_disabled = 0;
5690 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5692 page_group_by_mobility_disabled ? "off" : "on",
5695 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5699 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5700 static bool __meminit
5701 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5703 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5704 static struct memblock_region *r;
5706 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5707 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5708 for_each_memblock(memory, r) {
5709 if (*pfn < memblock_region_memory_end_pfn(r))
5713 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5714 memblock_is_mirror(r)) {
5715 *pfn = memblock_region_memory_end_pfn(r);
5724 * Initially all pages are reserved - free ones are freed
5725 * up by memblock_free_all() once the early boot process is
5726 * done. Non-atomic initialization, single-pass.
5728 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5729 unsigned long start_pfn, enum memmap_context context,
5730 struct vmem_altmap *altmap)
5732 unsigned long pfn, end_pfn = start_pfn + size;
5735 if (highest_memmap_pfn < end_pfn - 1)
5736 highest_memmap_pfn = end_pfn - 1;
5738 #ifdef CONFIG_ZONE_DEVICE
5740 * Honor reservation requested by the driver for this ZONE_DEVICE
5741 * memory. We limit the total number of pages to initialize to just
5742 * those that might contain the memory mapping. We will defer the
5743 * ZONE_DEVICE page initialization until after we have released
5746 if (zone == ZONE_DEVICE) {
5750 if (start_pfn == altmap->base_pfn)
5751 start_pfn += altmap->reserve;
5752 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5756 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5758 * There can be holes in boot-time mem_map[]s handed to this
5759 * function. They do not exist on hotplugged memory.
5761 if (context == MEMMAP_EARLY) {
5762 if (!early_pfn_valid(pfn))
5764 if (!early_pfn_in_nid(pfn, nid))
5766 if (overlap_memmap_init(zone, &pfn))
5768 if (defer_init(nid, pfn, end_pfn))
5772 page = pfn_to_page(pfn);
5773 __init_single_page(page, pfn, zone, nid);
5774 if (context == MEMMAP_HOTPLUG)
5775 __SetPageReserved(page);
5778 * Mark the block movable so that blocks are reserved for
5779 * movable at startup. This will force kernel allocations
5780 * to reserve their blocks rather than leaking throughout
5781 * the address space during boot when many long-lived
5782 * kernel allocations are made.
5784 * bitmap is created for zone's valid pfn range. but memmap
5785 * can be created for invalid pages (for alignment)
5786 * check here not to call set_pageblock_migratetype() against
5789 if (!(pfn & (pageblock_nr_pages - 1))) {
5790 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5796 #ifdef CONFIG_ZONE_DEVICE
5797 void __ref memmap_init_zone_device(struct zone *zone,
5798 unsigned long start_pfn,
5800 struct dev_pagemap *pgmap)
5802 unsigned long pfn, end_pfn = start_pfn + size;
5803 struct pglist_data *pgdat = zone->zone_pgdat;
5804 unsigned long zone_idx = zone_idx(zone);
5805 unsigned long start = jiffies;
5806 int nid = pgdat->node_id;
5808 if (WARN_ON_ONCE(!pgmap || !is_dev_zone(zone)))
5812 * The call to memmap_init_zone should have already taken care
5813 * of the pages reserved for the memmap, so we can just jump to
5814 * the end of that region and start processing the device pages.
5816 if (pgmap->altmap_valid) {
5817 struct vmem_altmap *altmap = &pgmap->altmap;
5819 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5820 size = end_pfn - start_pfn;
5823 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5824 struct page *page = pfn_to_page(pfn);
5826 __init_single_page(page, pfn, zone_idx, nid);
5829 * Mark page reserved as it will need to wait for onlining
5830 * phase for it to be fully associated with a zone.
5832 * We can use the non-atomic __set_bit operation for setting
5833 * the flag as we are still initializing the pages.
5835 __SetPageReserved(page);
5838 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5839 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5840 * page is ever freed or placed on a driver-private list.
5842 page->pgmap = pgmap;
5846 * Mark the block movable so that blocks are reserved for
5847 * movable at startup. This will force kernel allocations
5848 * to reserve their blocks rather than leaking throughout
5849 * the address space during boot when many long-lived
5850 * kernel allocations are made.
5852 * bitmap is created for zone's valid pfn range. but memmap
5853 * can be created for invalid pages (for alignment)
5854 * check here not to call set_pageblock_migratetype() against
5857 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5858 * because this is done early in sparse_add_one_section
5860 if (!(pfn & (pageblock_nr_pages - 1))) {
5861 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5866 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap->dev),
5867 size, jiffies_to_msecs(jiffies - start));
5871 static void __meminit zone_init_free_lists(struct zone *zone)
5873 unsigned int order, t;
5874 for_each_migratetype_order(order, t) {
5875 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5876 zone->free_area[order].nr_free = 0;
5880 void __meminit __weak memmap_init(unsigned long size, int nid,
5881 unsigned long zone, unsigned long start_pfn)
5883 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
5886 static int zone_batchsize(struct zone *zone)
5892 * The per-cpu-pages pools are set to around 1000th of the
5895 batch = zone_managed_pages(zone) / 1024;
5896 /* But no more than a meg. */
5897 if (batch * PAGE_SIZE > 1024 * 1024)
5898 batch = (1024 * 1024) / PAGE_SIZE;
5899 batch /= 4; /* We effectively *= 4 below */
5904 * Clamp the batch to a 2^n - 1 value. Having a power
5905 * of 2 value was found to be more likely to have
5906 * suboptimal cache aliasing properties in some cases.
5908 * For example if 2 tasks are alternately allocating
5909 * batches of pages, one task can end up with a lot
5910 * of pages of one half of the possible page colors
5911 * and the other with pages of the other colors.
5913 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5918 /* The deferral and batching of frees should be suppressed under NOMMU
5921 * The problem is that NOMMU needs to be able to allocate large chunks
5922 * of contiguous memory as there's no hardware page translation to
5923 * assemble apparent contiguous memory from discontiguous pages.
5925 * Queueing large contiguous runs of pages for batching, however,
5926 * causes the pages to actually be freed in smaller chunks. As there
5927 * can be a significant delay between the individual batches being
5928 * recycled, this leads to the once large chunks of space being
5929 * fragmented and becoming unavailable for high-order allocations.
5936 * pcp->high and pcp->batch values are related and dependent on one another:
5937 * ->batch must never be higher then ->high.
5938 * The following function updates them in a safe manner without read side
5941 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5942 * those fields changing asynchronously (acording the the above rule).
5944 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5945 * outside of boot time (or some other assurance that no concurrent updaters
5948 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5949 unsigned long batch)
5951 /* start with a fail safe value for batch */
5955 /* Update high, then batch, in order */
5962 /* a companion to pageset_set_high() */
5963 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5965 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5968 static void pageset_init(struct per_cpu_pageset *p)
5970 struct per_cpu_pages *pcp;
5973 memset(p, 0, sizeof(*p));
5976 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5977 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5980 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5983 pageset_set_batch(p, batch);
5987 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5988 * to the value high for the pageset p.
5990 static void pageset_set_high(struct per_cpu_pageset *p,
5993 unsigned long batch = max(1UL, high / 4);
5994 if ((high / 4) > (PAGE_SHIFT * 8))
5995 batch = PAGE_SHIFT * 8;
5997 pageset_update(&p->pcp, high, batch);
6000 static void pageset_set_high_and_batch(struct zone *zone,
6001 struct per_cpu_pageset *pcp)
6003 if (percpu_pagelist_fraction)
6004 pageset_set_high(pcp,
6005 (zone_managed_pages(zone) /
6006 percpu_pagelist_fraction));
6008 pageset_set_batch(pcp, zone_batchsize(zone));
6011 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6013 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6016 pageset_set_high_and_batch(zone, pcp);
6019 void __meminit setup_zone_pageset(struct zone *zone)
6022 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6023 for_each_possible_cpu(cpu)
6024 zone_pageset_init(zone, cpu);
6028 * Allocate per cpu pagesets and initialize them.
6029 * Before this call only boot pagesets were available.
6031 void __init setup_per_cpu_pageset(void)
6033 struct pglist_data *pgdat;
6036 for_each_populated_zone(zone)
6037 setup_zone_pageset(zone);
6039 for_each_online_pgdat(pgdat)
6040 pgdat->per_cpu_nodestats =
6041 alloc_percpu(struct per_cpu_nodestat);
6044 static __meminit void zone_pcp_init(struct zone *zone)
6047 * per cpu subsystem is not up at this point. The following code
6048 * relies on the ability of the linker to provide the
6049 * offset of a (static) per cpu variable into the per cpu area.
6051 zone->pageset = &boot_pageset;
6053 if (populated_zone(zone))
6054 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6055 zone->name, zone->present_pages,
6056 zone_batchsize(zone));
6059 void __meminit init_currently_empty_zone(struct zone *zone,
6060 unsigned long zone_start_pfn,
6063 struct pglist_data *pgdat = zone->zone_pgdat;
6064 int zone_idx = zone_idx(zone) + 1;
6066 if (zone_idx > pgdat->nr_zones)
6067 pgdat->nr_zones = zone_idx;
6069 zone->zone_start_pfn = zone_start_pfn;
6071 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6072 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6074 (unsigned long)zone_idx(zone),
6075 zone_start_pfn, (zone_start_pfn + size));
6077 zone_init_free_lists(zone);
6078 zone->initialized = 1;
6081 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6082 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6085 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6087 int __meminit __early_pfn_to_nid(unsigned long pfn,
6088 struct mminit_pfnnid_cache *state)
6090 unsigned long start_pfn, end_pfn;
6093 if (state->last_start <= pfn && pfn < state->last_end)
6094 return state->last_nid;
6096 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6097 if (nid != NUMA_NO_NODE) {
6098 state->last_start = start_pfn;
6099 state->last_end = end_pfn;
6100 state->last_nid = nid;
6105 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6108 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6109 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6110 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6112 * If an architecture guarantees that all ranges registered contain no holes
6113 * and may be freed, this this function may be used instead of calling
6114 * memblock_free_early_nid() manually.
6116 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6118 unsigned long start_pfn, end_pfn;
6121 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6122 start_pfn = min(start_pfn, max_low_pfn);
6123 end_pfn = min(end_pfn, max_low_pfn);
6125 if (start_pfn < end_pfn)
6126 memblock_free_early_nid(PFN_PHYS(start_pfn),
6127 (end_pfn - start_pfn) << PAGE_SHIFT,
6133 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6134 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6136 * If an architecture guarantees that all ranges registered contain no holes and may
6137 * be freed, this function may be used instead of calling memory_present() manually.
6139 void __init sparse_memory_present_with_active_regions(int nid)
6141 unsigned long start_pfn, end_pfn;
6144 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6145 memory_present(this_nid, start_pfn, end_pfn);
6149 * get_pfn_range_for_nid - Return the start and end page frames for a node
6150 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6151 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6152 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6154 * It returns the start and end page frame of a node based on information
6155 * provided by memblock_set_node(). If called for a node
6156 * with no available memory, a warning is printed and the start and end
6159 void __init get_pfn_range_for_nid(unsigned int nid,
6160 unsigned long *start_pfn, unsigned long *end_pfn)
6162 unsigned long this_start_pfn, this_end_pfn;
6168 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6169 *start_pfn = min(*start_pfn, this_start_pfn);
6170 *end_pfn = max(*end_pfn, this_end_pfn);
6173 if (*start_pfn == -1UL)
6178 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6179 * assumption is made that zones within a node are ordered in monotonic
6180 * increasing memory addresses so that the "highest" populated zone is used
6182 static void __init find_usable_zone_for_movable(void)
6185 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6186 if (zone_index == ZONE_MOVABLE)
6189 if (arch_zone_highest_possible_pfn[zone_index] >
6190 arch_zone_lowest_possible_pfn[zone_index])
6194 VM_BUG_ON(zone_index == -1);
6195 movable_zone = zone_index;
6199 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6200 * because it is sized independent of architecture. Unlike the other zones,
6201 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6202 * in each node depending on the size of each node and how evenly kernelcore
6203 * is distributed. This helper function adjusts the zone ranges
6204 * provided by the architecture for a given node by using the end of the
6205 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6206 * zones within a node are in order of monotonic increases memory addresses
6208 static void __init adjust_zone_range_for_zone_movable(int nid,
6209 unsigned long zone_type,
6210 unsigned long node_start_pfn,
6211 unsigned long node_end_pfn,
6212 unsigned long *zone_start_pfn,
6213 unsigned long *zone_end_pfn)
6215 /* Only adjust if ZONE_MOVABLE is on this node */
6216 if (zone_movable_pfn[nid]) {
6217 /* Size ZONE_MOVABLE */
6218 if (zone_type == ZONE_MOVABLE) {
6219 *zone_start_pfn = zone_movable_pfn[nid];
6220 *zone_end_pfn = min(node_end_pfn,
6221 arch_zone_highest_possible_pfn[movable_zone]);
6223 /* Adjust for ZONE_MOVABLE starting within this range */
6224 } else if (!mirrored_kernelcore &&
6225 *zone_start_pfn < zone_movable_pfn[nid] &&
6226 *zone_end_pfn > zone_movable_pfn[nid]) {
6227 *zone_end_pfn = zone_movable_pfn[nid];
6229 /* Check if this whole range is within ZONE_MOVABLE */
6230 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6231 *zone_start_pfn = *zone_end_pfn;
6236 * Return the number of pages a zone spans in a node, including holes
6237 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6239 static unsigned long __init zone_spanned_pages_in_node(int nid,
6240 unsigned long zone_type,
6241 unsigned long node_start_pfn,
6242 unsigned long node_end_pfn,
6243 unsigned long *zone_start_pfn,
6244 unsigned long *zone_end_pfn,
6245 unsigned long *ignored)
6247 /* When hotadd a new node from cpu_up(), the node should be empty */
6248 if (!node_start_pfn && !node_end_pfn)
6251 /* Get the start and end of the zone */
6252 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
6253 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
6254 adjust_zone_range_for_zone_movable(nid, zone_type,
6255 node_start_pfn, node_end_pfn,
6256 zone_start_pfn, zone_end_pfn);
6258 /* Check that this node has pages within the zone's required range */
6259 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6262 /* Move the zone boundaries inside the node if necessary */
6263 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6264 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6266 /* Return the spanned pages */
6267 return *zone_end_pfn - *zone_start_pfn;
6271 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6272 * then all holes in the requested range will be accounted for.
6274 unsigned long __init __absent_pages_in_range(int nid,
6275 unsigned long range_start_pfn,
6276 unsigned long range_end_pfn)
6278 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6279 unsigned long start_pfn, end_pfn;
6282 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6283 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6284 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6285 nr_absent -= end_pfn - start_pfn;
6291 * absent_pages_in_range - Return number of page frames in holes within a range
6292 * @start_pfn: The start PFN to start searching for holes
6293 * @end_pfn: The end PFN to stop searching for holes
6295 * Return: the number of pages frames in memory holes within a range.
6297 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6298 unsigned long end_pfn)
6300 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6303 /* Return the number of page frames in holes in a zone on a node */
6304 static unsigned long __init zone_absent_pages_in_node(int nid,
6305 unsigned long zone_type,
6306 unsigned long node_start_pfn,
6307 unsigned long node_end_pfn,
6308 unsigned long *ignored)
6310 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6311 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6312 unsigned long zone_start_pfn, zone_end_pfn;
6313 unsigned long nr_absent;
6315 /* When hotadd a new node from cpu_up(), the node should be empty */
6316 if (!node_start_pfn && !node_end_pfn)
6319 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6320 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6322 adjust_zone_range_for_zone_movable(nid, zone_type,
6323 node_start_pfn, node_end_pfn,
6324 &zone_start_pfn, &zone_end_pfn);
6325 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6328 * ZONE_MOVABLE handling.
6329 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6332 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6333 unsigned long start_pfn, end_pfn;
6334 struct memblock_region *r;
6336 for_each_memblock(memory, r) {
6337 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6338 zone_start_pfn, zone_end_pfn);
6339 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6340 zone_start_pfn, zone_end_pfn);
6342 if (zone_type == ZONE_MOVABLE &&
6343 memblock_is_mirror(r))
6344 nr_absent += end_pfn - start_pfn;
6346 if (zone_type == ZONE_NORMAL &&
6347 !memblock_is_mirror(r))
6348 nr_absent += end_pfn - start_pfn;
6355 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6356 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6357 unsigned long zone_type,
6358 unsigned long node_start_pfn,
6359 unsigned long node_end_pfn,
6360 unsigned long *zone_start_pfn,
6361 unsigned long *zone_end_pfn,
6362 unsigned long *zones_size)
6366 *zone_start_pfn = node_start_pfn;
6367 for (zone = 0; zone < zone_type; zone++)
6368 *zone_start_pfn += zones_size[zone];
6370 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6372 return zones_size[zone_type];
6375 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6376 unsigned long zone_type,
6377 unsigned long node_start_pfn,
6378 unsigned long node_end_pfn,
6379 unsigned long *zholes_size)
6384 return zholes_size[zone_type];
6387 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6389 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6390 unsigned long node_start_pfn,
6391 unsigned long node_end_pfn,
6392 unsigned long *zones_size,
6393 unsigned long *zholes_size)
6395 unsigned long realtotalpages = 0, totalpages = 0;
6398 for (i = 0; i < MAX_NR_ZONES; i++) {
6399 struct zone *zone = pgdat->node_zones + i;
6400 unsigned long zone_start_pfn, zone_end_pfn;
6401 unsigned long size, real_size;
6403 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6409 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6410 node_start_pfn, node_end_pfn,
6413 zone->zone_start_pfn = zone_start_pfn;
6415 zone->zone_start_pfn = 0;
6416 zone->spanned_pages = size;
6417 zone->present_pages = real_size;
6420 realtotalpages += real_size;
6423 pgdat->node_spanned_pages = totalpages;
6424 pgdat->node_present_pages = realtotalpages;
6425 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6429 #ifndef CONFIG_SPARSEMEM
6431 * Calculate the size of the zone->blockflags rounded to an unsigned long
6432 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6433 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6434 * round what is now in bits to nearest long in bits, then return it in
6437 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6439 unsigned long usemapsize;
6441 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6442 usemapsize = roundup(zonesize, pageblock_nr_pages);
6443 usemapsize = usemapsize >> pageblock_order;
6444 usemapsize *= NR_PAGEBLOCK_BITS;
6445 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6447 return usemapsize / 8;
6450 static void __ref setup_usemap(struct pglist_data *pgdat,
6452 unsigned long zone_start_pfn,
6453 unsigned long zonesize)
6455 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6456 zone->pageblock_flags = NULL;
6458 zone->pageblock_flags =
6459 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6461 if (!zone->pageblock_flags)
6462 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6463 usemapsize, zone->name, pgdat->node_id);
6467 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6468 unsigned long zone_start_pfn, unsigned long zonesize) {}
6469 #endif /* CONFIG_SPARSEMEM */
6471 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6473 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6474 void __init set_pageblock_order(void)
6478 /* Check that pageblock_nr_pages has not already been setup */
6479 if (pageblock_order)
6482 if (HPAGE_SHIFT > PAGE_SHIFT)
6483 order = HUGETLB_PAGE_ORDER;
6485 order = MAX_ORDER - 1;
6488 * Assume the largest contiguous order of interest is a huge page.
6489 * This value may be variable depending on boot parameters on IA64 and
6492 pageblock_order = order;
6494 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6497 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6498 * is unused as pageblock_order is set at compile-time. See
6499 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6502 void __init set_pageblock_order(void)
6506 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6508 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6509 unsigned long present_pages)
6511 unsigned long pages = spanned_pages;
6514 * Provide a more accurate estimation if there are holes within
6515 * the zone and SPARSEMEM is in use. If there are holes within the
6516 * zone, each populated memory region may cost us one or two extra
6517 * memmap pages due to alignment because memmap pages for each
6518 * populated regions may not be naturally aligned on page boundary.
6519 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6521 if (spanned_pages > present_pages + (present_pages >> 4) &&
6522 IS_ENABLED(CONFIG_SPARSEMEM))
6523 pages = present_pages;
6525 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6528 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6529 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6531 spin_lock_init(&pgdat->split_queue_lock);
6532 INIT_LIST_HEAD(&pgdat->split_queue);
6533 pgdat->split_queue_len = 0;
6536 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6539 #ifdef CONFIG_COMPACTION
6540 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6542 init_waitqueue_head(&pgdat->kcompactd_wait);
6545 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6548 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6550 pgdat_resize_init(pgdat);
6552 pgdat_init_split_queue(pgdat);
6553 pgdat_init_kcompactd(pgdat);
6555 init_waitqueue_head(&pgdat->kswapd_wait);
6556 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6558 pgdat_page_ext_init(pgdat);
6559 spin_lock_init(&pgdat->lru_lock);
6560 lruvec_init(node_lruvec(pgdat));
6563 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6564 unsigned long remaining_pages)
6566 atomic_long_set(&zone->managed_pages, remaining_pages);
6567 zone_set_nid(zone, nid);
6568 zone->name = zone_names[idx];
6569 zone->zone_pgdat = NODE_DATA(nid);
6570 spin_lock_init(&zone->lock);
6571 zone_seqlock_init(zone);
6572 zone_pcp_init(zone);
6576 * Set up the zone data structures
6577 * - init pgdat internals
6578 * - init all zones belonging to this node
6580 * NOTE: this function is only called during memory hotplug
6582 #ifdef CONFIG_MEMORY_HOTPLUG
6583 void __ref free_area_init_core_hotplug(int nid)
6586 pg_data_t *pgdat = NODE_DATA(nid);
6588 pgdat_init_internals(pgdat);
6589 for (z = 0; z < MAX_NR_ZONES; z++)
6590 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6595 * Set up the zone data structures:
6596 * - mark all pages reserved
6597 * - mark all memory queues empty
6598 * - clear the memory bitmaps
6600 * NOTE: pgdat should get zeroed by caller.
6601 * NOTE: this function is only called during early init.
6603 static void __init free_area_init_core(struct pglist_data *pgdat)
6606 int nid = pgdat->node_id;
6608 pgdat_init_internals(pgdat);
6609 pgdat->per_cpu_nodestats = &boot_nodestats;
6611 for (j = 0; j < MAX_NR_ZONES; j++) {
6612 struct zone *zone = pgdat->node_zones + j;
6613 unsigned long size, freesize, memmap_pages;
6614 unsigned long zone_start_pfn = zone->zone_start_pfn;
6616 size = zone->spanned_pages;
6617 freesize = zone->present_pages;
6620 * Adjust freesize so that it accounts for how much memory
6621 * is used by this zone for memmap. This affects the watermark
6622 * and per-cpu initialisations
6624 memmap_pages = calc_memmap_size(size, freesize);
6625 if (!is_highmem_idx(j)) {
6626 if (freesize >= memmap_pages) {
6627 freesize -= memmap_pages;
6630 " %s zone: %lu pages used for memmap\n",
6631 zone_names[j], memmap_pages);
6633 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6634 zone_names[j], memmap_pages, freesize);
6637 /* Account for reserved pages */
6638 if (j == 0 && freesize > dma_reserve) {
6639 freesize -= dma_reserve;
6640 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6641 zone_names[0], dma_reserve);
6644 if (!is_highmem_idx(j))
6645 nr_kernel_pages += freesize;
6646 /* Charge for highmem memmap if there are enough kernel pages */
6647 else if (nr_kernel_pages > memmap_pages * 2)
6648 nr_kernel_pages -= memmap_pages;
6649 nr_all_pages += freesize;
6652 * Set an approximate value for lowmem here, it will be adjusted
6653 * when the bootmem allocator frees pages into the buddy system.
6654 * And all highmem pages will be managed by the buddy system.
6656 zone_init_internals(zone, j, nid, freesize);
6661 set_pageblock_order();
6662 setup_usemap(pgdat, zone, zone_start_pfn, size);
6663 init_currently_empty_zone(zone, zone_start_pfn, size);
6664 memmap_init(size, nid, j, zone_start_pfn);
6668 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6669 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6671 unsigned long __maybe_unused start = 0;
6672 unsigned long __maybe_unused offset = 0;
6674 /* Skip empty nodes */
6675 if (!pgdat->node_spanned_pages)
6678 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6679 offset = pgdat->node_start_pfn - start;
6680 /* ia64 gets its own node_mem_map, before this, without bootmem */
6681 if (!pgdat->node_mem_map) {
6682 unsigned long size, end;
6686 * The zone's endpoints aren't required to be MAX_ORDER
6687 * aligned but the node_mem_map endpoints must be in order
6688 * for the buddy allocator to function correctly.
6690 end = pgdat_end_pfn(pgdat);
6691 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6692 size = (end - start) * sizeof(struct page);
6693 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6696 panic("Failed to allocate %ld bytes for node %d memory map\n",
6697 size, pgdat->node_id);
6698 pgdat->node_mem_map = map + offset;
6700 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6701 __func__, pgdat->node_id, (unsigned long)pgdat,
6702 (unsigned long)pgdat->node_mem_map);
6703 #ifndef CONFIG_NEED_MULTIPLE_NODES
6705 * With no DISCONTIG, the global mem_map is just set as node 0's
6707 if (pgdat == NODE_DATA(0)) {
6708 mem_map = NODE_DATA(0)->node_mem_map;
6709 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6710 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6712 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6717 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6718 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6720 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6721 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6723 pgdat->first_deferred_pfn = ULONG_MAX;
6726 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6729 void __init free_area_init_node(int nid, unsigned long *zones_size,
6730 unsigned long node_start_pfn,
6731 unsigned long *zholes_size)
6733 pg_data_t *pgdat = NODE_DATA(nid);
6734 unsigned long start_pfn = 0;
6735 unsigned long end_pfn = 0;
6737 /* pg_data_t should be reset to zero when it's allocated */
6738 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6740 pgdat->node_id = nid;
6741 pgdat->node_start_pfn = node_start_pfn;
6742 pgdat->per_cpu_nodestats = NULL;
6743 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6744 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6745 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6746 (u64)start_pfn << PAGE_SHIFT,
6747 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6749 start_pfn = node_start_pfn;
6751 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6752 zones_size, zholes_size);
6754 alloc_node_mem_map(pgdat);
6755 pgdat_set_deferred_range(pgdat);
6757 free_area_init_core(pgdat);
6760 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6762 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6765 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6770 for (pfn = spfn; pfn < epfn; pfn++) {
6771 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6772 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6773 + pageblock_nr_pages - 1;
6776 mm_zero_struct_page(pfn_to_page(pfn));
6784 * Only struct pages that are backed by physical memory are zeroed and
6785 * initialized by going through __init_single_page(). But, there are some
6786 * struct pages which are reserved in memblock allocator and their fields
6787 * may be accessed (for example page_to_pfn() on some configuration accesses
6788 * flags). We must explicitly zero those struct pages.
6790 * This function also addresses a similar issue where struct pages are left
6791 * uninitialized because the physical address range is not covered by
6792 * memblock.memory or memblock.reserved. That could happen when memblock
6793 * layout is manually configured via memmap=.
6795 void __init zero_resv_unavail(void)
6797 phys_addr_t start, end;
6799 phys_addr_t next = 0;
6802 * Loop through unavailable ranges not covered by memblock.memory.
6805 for_each_mem_range(i, &memblock.memory, NULL,
6806 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6808 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6811 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6814 * Struct pages that do not have backing memory. This could be because
6815 * firmware is using some of this memory, or for some other reasons.
6818 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6820 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6822 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6824 #if MAX_NUMNODES > 1
6826 * Figure out the number of possible node ids.
6828 void __init setup_nr_node_ids(void)
6830 unsigned int highest;
6832 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6833 nr_node_ids = highest + 1;
6838 * node_map_pfn_alignment - determine the maximum internode alignment
6840 * This function should be called after node map is populated and sorted.
6841 * It calculates the maximum power of two alignment which can distinguish
6844 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6845 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6846 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6847 * shifted, 1GiB is enough and this function will indicate so.
6849 * This is used to test whether pfn -> nid mapping of the chosen memory
6850 * model has fine enough granularity to avoid incorrect mapping for the
6851 * populated node map.
6853 * Return: the determined alignment in pfn's. 0 if there is no alignment
6854 * requirement (single node).
6856 unsigned long __init node_map_pfn_alignment(void)
6858 unsigned long accl_mask = 0, last_end = 0;
6859 unsigned long start, end, mask;
6860 int last_nid = NUMA_NO_NODE;
6863 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6864 if (!start || last_nid < 0 || last_nid == nid) {
6871 * Start with a mask granular enough to pin-point to the
6872 * start pfn and tick off bits one-by-one until it becomes
6873 * too coarse to separate the current node from the last.
6875 mask = ~((1 << __ffs(start)) - 1);
6876 while (mask && last_end <= (start & (mask << 1)))
6879 /* accumulate all internode masks */
6883 /* convert mask to number of pages */
6884 return ~accl_mask + 1;
6887 /* Find the lowest pfn for a node */
6888 static unsigned long __init find_min_pfn_for_node(int nid)
6890 unsigned long min_pfn = ULONG_MAX;
6891 unsigned long start_pfn;
6894 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6895 min_pfn = min(min_pfn, start_pfn);
6897 if (min_pfn == ULONG_MAX) {
6898 pr_warn("Could not find start_pfn for node %d\n", nid);
6906 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6908 * Return: the minimum PFN based on information provided via
6909 * memblock_set_node().
6911 unsigned long __init find_min_pfn_with_active_regions(void)
6913 return find_min_pfn_for_node(MAX_NUMNODES);
6917 * early_calculate_totalpages()
6918 * Sum pages in active regions for movable zone.
6919 * Populate N_MEMORY for calculating usable_nodes.
6921 static unsigned long __init early_calculate_totalpages(void)
6923 unsigned long totalpages = 0;
6924 unsigned long start_pfn, end_pfn;
6927 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6928 unsigned long pages = end_pfn - start_pfn;
6930 totalpages += pages;
6932 node_set_state(nid, N_MEMORY);
6938 * Find the PFN the Movable zone begins in each node. Kernel memory
6939 * is spread evenly between nodes as long as the nodes have enough
6940 * memory. When they don't, some nodes will have more kernelcore than
6943 static void __init find_zone_movable_pfns_for_nodes(void)
6946 unsigned long usable_startpfn;
6947 unsigned long kernelcore_node, kernelcore_remaining;
6948 /* save the state before borrow the nodemask */
6949 nodemask_t saved_node_state = node_states[N_MEMORY];
6950 unsigned long totalpages = early_calculate_totalpages();
6951 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6952 struct memblock_region *r;
6954 /* Need to find movable_zone earlier when movable_node is specified. */
6955 find_usable_zone_for_movable();
6958 * If movable_node is specified, ignore kernelcore and movablecore
6961 if (movable_node_is_enabled()) {
6962 for_each_memblock(memory, r) {
6963 if (!memblock_is_hotpluggable(r))
6968 usable_startpfn = PFN_DOWN(r->base);
6969 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6970 min(usable_startpfn, zone_movable_pfn[nid]) :
6978 * If kernelcore=mirror is specified, ignore movablecore option
6980 if (mirrored_kernelcore) {
6981 bool mem_below_4gb_not_mirrored = false;
6983 for_each_memblock(memory, r) {
6984 if (memblock_is_mirror(r))
6989 usable_startpfn = memblock_region_memory_base_pfn(r);
6991 if (usable_startpfn < 0x100000) {
6992 mem_below_4gb_not_mirrored = true;
6996 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6997 min(usable_startpfn, zone_movable_pfn[nid]) :
7001 if (mem_below_4gb_not_mirrored)
7002 pr_warn("This configuration results in unmirrored kernel memory.");
7008 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7009 * amount of necessary memory.
7011 if (required_kernelcore_percent)
7012 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7014 if (required_movablecore_percent)
7015 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7019 * If movablecore= was specified, calculate what size of
7020 * kernelcore that corresponds so that memory usable for
7021 * any allocation type is evenly spread. If both kernelcore
7022 * and movablecore are specified, then the value of kernelcore
7023 * will be used for required_kernelcore if it's greater than
7024 * what movablecore would have allowed.
7026 if (required_movablecore) {
7027 unsigned long corepages;
7030 * Round-up so that ZONE_MOVABLE is at least as large as what
7031 * was requested by the user
7033 required_movablecore =
7034 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7035 required_movablecore = min(totalpages, required_movablecore);
7036 corepages = totalpages - required_movablecore;
7038 required_kernelcore = max(required_kernelcore, corepages);
7042 * If kernelcore was not specified or kernelcore size is larger
7043 * than totalpages, there is no ZONE_MOVABLE.
7045 if (!required_kernelcore || required_kernelcore >= totalpages)
7048 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7049 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7052 /* Spread kernelcore memory as evenly as possible throughout nodes */
7053 kernelcore_node = required_kernelcore / usable_nodes;
7054 for_each_node_state(nid, N_MEMORY) {
7055 unsigned long start_pfn, end_pfn;
7058 * Recalculate kernelcore_node if the division per node
7059 * now exceeds what is necessary to satisfy the requested
7060 * amount of memory for the kernel
7062 if (required_kernelcore < kernelcore_node)
7063 kernelcore_node = required_kernelcore / usable_nodes;
7066 * As the map is walked, we track how much memory is usable
7067 * by the kernel using kernelcore_remaining. When it is
7068 * 0, the rest of the node is usable by ZONE_MOVABLE
7070 kernelcore_remaining = kernelcore_node;
7072 /* Go through each range of PFNs within this node */
7073 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7074 unsigned long size_pages;
7076 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7077 if (start_pfn >= end_pfn)
7080 /* Account for what is only usable for kernelcore */
7081 if (start_pfn < usable_startpfn) {
7082 unsigned long kernel_pages;
7083 kernel_pages = min(end_pfn, usable_startpfn)
7086 kernelcore_remaining -= min(kernel_pages,
7087 kernelcore_remaining);
7088 required_kernelcore -= min(kernel_pages,
7089 required_kernelcore);
7091 /* Continue if range is now fully accounted */
7092 if (end_pfn <= usable_startpfn) {
7095 * Push zone_movable_pfn to the end so
7096 * that if we have to rebalance
7097 * kernelcore across nodes, we will
7098 * not double account here
7100 zone_movable_pfn[nid] = end_pfn;
7103 start_pfn = usable_startpfn;
7107 * The usable PFN range for ZONE_MOVABLE is from
7108 * start_pfn->end_pfn. Calculate size_pages as the
7109 * number of pages used as kernelcore
7111 size_pages = end_pfn - start_pfn;
7112 if (size_pages > kernelcore_remaining)
7113 size_pages = kernelcore_remaining;
7114 zone_movable_pfn[nid] = start_pfn + size_pages;
7117 * Some kernelcore has been met, update counts and
7118 * break if the kernelcore for this node has been
7121 required_kernelcore -= min(required_kernelcore,
7123 kernelcore_remaining -= size_pages;
7124 if (!kernelcore_remaining)
7130 * If there is still required_kernelcore, we do another pass with one
7131 * less node in the count. This will push zone_movable_pfn[nid] further
7132 * along on the nodes that still have memory until kernelcore is
7136 if (usable_nodes && required_kernelcore > usable_nodes)
7140 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7141 for (nid = 0; nid < MAX_NUMNODES; nid++)
7142 zone_movable_pfn[nid] =
7143 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7146 /* restore the node_state */
7147 node_states[N_MEMORY] = saved_node_state;
7150 /* Any regular or high memory on that node ? */
7151 static void check_for_memory(pg_data_t *pgdat, int nid)
7153 enum zone_type zone_type;
7155 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7156 struct zone *zone = &pgdat->node_zones[zone_type];
7157 if (populated_zone(zone)) {
7158 if (IS_ENABLED(CONFIG_HIGHMEM))
7159 node_set_state(nid, N_HIGH_MEMORY);
7160 if (zone_type <= ZONE_NORMAL)
7161 node_set_state(nid, N_NORMAL_MEMORY);
7168 * free_area_init_nodes - Initialise all pg_data_t and zone data
7169 * @max_zone_pfn: an array of max PFNs for each zone
7171 * This will call free_area_init_node() for each active node in the system.
7172 * Using the page ranges provided by memblock_set_node(), the size of each
7173 * zone in each node and their holes is calculated. If the maximum PFN
7174 * between two adjacent zones match, it is assumed that the zone is empty.
7175 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7176 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7177 * starts where the previous one ended. For example, ZONE_DMA32 starts
7178 * at arch_max_dma_pfn.
7180 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7182 unsigned long start_pfn, end_pfn;
7185 /* Record where the zone boundaries are */
7186 memset(arch_zone_lowest_possible_pfn, 0,
7187 sizeof(arch_zone_lowest_possible_pfn));
7188 memset(arch_zone_highest_possible_pfn, 0,
7189 sizeof(arch_zone_highest_possible_pfn));
7191 start_pfn = find_min_pfn_with_active_regions();
7193 for (i = 0; i < MAX_NR_ZONES; i++) {
7194 if (i == ZONE_MOVABLE)
7197 end_pfn = max(max_zone_pfn[i], start_pfn);
7198 arch_zone_lowest_possible_pfn[i] = start_pfn;
7199 arch_zone_highest_possible_pfn[i] = end_pfn;
7201 start_pfn = end_pfn;
7204 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7205 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7206 find_zone_movable_pfns_for_nodes();
7208 /* Print out the zone ranges */
7209 pr_info("Zone ranges:\n");
7210 for (i = 0; i < MAX_NR_ZONES; i++) {
7211 if (i == ZONE_MOVABLE)
7213 pr_info(" %-8s ", zone_names[i]);
7214 if (arch_zone_lowest_possible_pfn[i] ==
7215 arch_zone_highest_possible_pfn[i])
7218 pr_cont("[mem %#018Lx-%#018Lx]\n",
7219 (u64)arch_zone_lowest_possible_pfn[i]
7221 ((u64)arch_zone_highest_possible_pfn[i]
7222 << PAGE_SHIFT) - 1);
7225 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7226 pr_info("Movable zone start for each node\n");
7227 for (i = 0; i < MAX_NUMNODES; i++) {
7228 if (zone_movable_pfn[i])
7229 pr_info(" Node %d: %#018Lx\n", i,
7230 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7233 /* Print out the early node map */
7234 pr_info("Early memory node ranges\n");
7235 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
7236 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7237 (u64)start_pfn << PAGE_SHIFT,
7238 ((u64)end_pfn << PAGE_SHIFT) - 1);
7240 /* Initialise every node */
7241 mminit_verify_pageflags_layout();
7242 setup_nr_node_ids();
7243 zero_resv_unavail();
7244 for_each_online_node(nid) {
7245 pg_data_t *pgdat = NODE_DATA(nid);
7246 free_area_init_node(nid, NULL,
7247 find_min_pfn_for_node(nid), NULL);
7249 /* Any memory on that node */
7250 if (pgdat->node_present_pages)
7251 node_set_state(nid, N_MEMORY);
7252 check_for_memory(pgdat, nid);
7256 static int __init cmdline_parse_core(char *p, unsigned long *core,
7257 unsigned long *percent)
7259 unsigned long long coremem;
7265 /* Value may be a percentage of total memory, otherwise bytes */
7266 coremem = simple_strtoull(p, &endptr, 0);
7267 if (*endptr == '%') {
7268 /* Paranoid check for percent values greater than 100 */
7269 WARN_ON(coremem > 100);
7273 coremem = memparse(p, &p);
7274 /* Paranoid check that UL is enough for the coremem value */
7275 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7277 *core = coremem >> PAGE_SHIFT;
7284 * kernelcore=size sets the amount of memory for use for allocations that
7285 * cannot be reclaimed or migrated.
7287 static int __init cmdline_parse_kernelcore(char *p)
7289 /* parse kernelcore=mirror */
7290 if (parse_option_str(p, "mirror")) {
7291 mirrored_kernelcore = true;
7295 return cmdline_parse_core(p, &required_kernelcore,
7296 &required_kernelcore_percent);
7300 * movablecore=size sets the amount of memory for use for allocations that
7301 * can be reclaimed or migrated.
7303 static int __init cmdline_parse_movablecore(char *p)
7305 return cmdline_parse_core(p, &required_movablecore,
7306 &required_movablecore_percent);
7309 early_param("kernelcore", cmdline_parse_kernelcore);
7310 early_param("movablecore", cmdline_parse_movablecore);
7312 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7314 void adjust_managed_page_count(struct page *page, long count)
7316 atomic_long_add(count, &page_zone(page)->managed_pages);
7317 totalram_pages_add(count);
7318 #ifdef CONFIG_HIGHMEM
7319 if (PageHighMem(page))
7320 totalhigh_pages_add(count);
7323 EXPORT_SYMBOL(adjust_managed_page_count);
7325 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7328 unsigned long pages = 0;
7330 start = (void *)PAGE_ALIGN((unsigned long)start);
7331 end = (void *)((unsigned long)end & PAGE_MASK);
7332 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7333 struct page *page = virt_to_page(pos);
7334 void *direct_map_addr;
7337 * 'direct_map_addr' might be different from 'pos'
7338 * because some architectures' virt_to_page()
7339 * work with aliases. Getting the direct map
7340 * address ensures that we get a _writeable_
7341 * alias for the memset().
7343 direct_map_addr = page_address(page);
7344 if ((unsigned int)poison <= 0xFF)
7345 memset(direct_map_addr, poison, PAGE_SIZE);
7347 free_reserved_page(page);
7351 pr_info("Freeing %s memory: %ldK\n",
7352 s, pages << (PAGE_SHIFT - 10));
7357 #ifdef CONFIG_HIGHMEM
7358 void free_highmem_page(struct page *page)
7360 __free_reserved_page(page);
7361 totalram_pages_inc();
7362 atomic_long_inc(&page_zone(page)->managed_pages);
7363 totalhigh_pages_inc();
7368 void __init mem_init_print_info(const char *str)
7370 unsigned long physpages, codesize, datasize, rosize, bss_size;
7371 unsigned long init_code_size, init_data_size;
7373 physpages = get_num_physpages();
7374 codesize = _etext - _stext;
7375 datasize = _edata - _sdata;
7376 rosize = __end_rodata - __start_rodata;
7377 bss_size = __bss_stop - __bss_start;
7378 init_data_size = __init_end - __init_begin;
7379 init_code_size = _einittext - _sinittext;
7382 * Detect special cases and adjust section sizes accordingly:
7383 * 1) .init.* may be embedded into .data sections
7384 * 2) .init.text.* may be out of [__init_begin, __init_end],
7385 * please refer to arch/tile/kernel/vmlinux.lds.S.
7386 * 3) .rodata.* may be embedded into .text or .data sections.
7388 #define adj_init_size(start, end, size, pos, adj) \
7390 if (start <= pos && pos < end && size > adj) \
7394 adj_init_size(__init_begin, __init_end, init_data_size,
7395 _sinittext, init_code_size);
7396 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7397 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7398 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7399 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7401 #undef adj_init_size
7403 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7404 #ifdef CONFIG_HIGHMEM
7408 nr_free_pages() << (PAGE_SHIFT - 10),
7409 physpages << (PAGE_SHIFT - 10),
7410 codesize >> 10, datasize >> 10, rosize >> 10,
7411 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7412 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7413 totalcma_pages << (PAGE_SHIFT - 10),
7414 #ifdef CONFIG_HIGHMEM
7415 totalhigh_pages() << (PAGE_SHIFT - 10),
7417 str ? ", " : "", str ? str : "");
7421 * set_dma_reserve - set the specified number of pages reserved in the first zone
7422 * @new_dma_reserve: The number of pages to mark reserved
7424 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7425 * In the DMA zone, a significant percentage may be consumed by kernel image
7426 * and other unfreeable allocations which can skew the watermarks badly. This
7427 * function may optionally be used to account for unfreeable pages in the
7428 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7429 * smaller per-cpu batchsize.
7431 void __init set_dma_reserve(unsigned long new_dma_reserve)
7433 dma_reserve = new_dma_reserve;
7436 void __init free_area_init(unsigned long *zones_size)
7438 zero_resv_unavail();
7439 free_area_init_node(0, zones_size,
7440 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7443 static int page_alloc_cpu_dead(unsigned int cpu)
7446 lru_add_drain_cpu(cpu);
7450 * Spill the event counters of the dead processor
7451 * into the current processors event counters.
7452 * This artificially elevates the count of the current
7455 vm_events_fold_cpu(cpu);
7458 * Zero the differential counters of the dead processor
7459 * so that the vm statistics are consistent.
7461 * This is only okay since the processor is dead and cannot
7462 * race with what we are doing.
7464 cpu_vm_stats_fold(cpu);
7468 void __init page_alloc_init(void)
7472 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7473 "mm/page_alloc:dead", NULL,
7474 page_alloc_cpu_dead);
7479 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7480 * or min_free_kbytes changes.
7482 static void calculate_totalreserve_pages(void)
7484 struct pglist_data *pgdat;
7485 unsigned long reserve_pages = 0;
7486 enum zone_type i, j;
7488 for_each_online_pgdat(pgdat) {
7490 pgdat->totalreserve_pages = 0;
7492 for (i = 0; i < MAX_NR_ZONES; i++) {
7493 struct zone *zone = pgdat->node_zones + i;
7495 unsigned long managed_pages = zone_managed_pages(zone);
7497 /* Find valid and maximum lowmem_reserve in the zone */
7498 for (j = i; j < MAX_NR_ZONES; j++) {
7499 if (zone->lowmem_reserve[j] > max)
7500 max = zone->lowmem_reserve[j];
7503 /* we treat the high watermark as reserved pages. */
7504 max += high_wmark_pages(zone);
7506 if (max > managed_pages)
7507 max = managed_pages;
7509 pgdat->totalreserve_pages += max;
7511 reserve_pages += max;
7514 totalreserve_pages = reserve_pages;
7518 * setup_per_zone_lowmem_reserve - called whenever
7519 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7520 * has a correct pages reserved value, so an adequate number of
7521 * pages are left in the zone after a successful __alloc_pages().
7523 static void setup_per_zone_lowmem_reserve(void)
7525 struct pglist_data *pgdat;
7526 enum zone_type j, idx;
7528 for_each_online_pgdat(pgdat) {
7529 for (j = 0; j < MAX_NR_ZONES; j++) {
7530 struct zone *zone = pgdat->node_zones + j;
7531 unsigned long managed_pages = zone_managed_pages(zone);
7533 zone->lowmem_reserve[j] = 0;
7537 struct zone *lower_zone;
7540 lower_zone = pgdat->node_zones + idx;
7542 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7543 sysctl_lowmem_reserve_ratio[idx] = 0;
7544 lower_zone->lowmem_reserve[j] = 0;
7546 lower_zone->lowmem_reserve[j] =
7547 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7549 managed_pages += zone_managed_pages(lower_zone);
7554 /* update totalreserve_pages */
7555 calculate_totalreserve_pages();
7558 static void __setup_per_zone_wmarks(void)
7560 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7561 unsigned long lowmem_pages = 0;
7563 unsigned long flags;
7565 /* Calculate total number of !ZONE_HIGHMEM pages */
7566 for_each_zone(zone) {
7567 if (!is_highmem(zone))
7568 lowmem_pages += zone_managed_pages(zone);
7571 for_each_zone(zone) {
7574 spin_lock_irqsave(&zone->lock, flags);
7575 tmp = (u64)pages_min * zone_managed_pages(zone);
7576 do_div(tmp, lowmem_pages);
7577 if (is_highmem(zone)) {
7579 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7580 * need highmem pages, so cap pages_min to a small
7583 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7584 * deltas control async page reclaim, and so should
7585 * not be capped for highmem.
7587 unsigned long min_pages;
7589 min_pages = zone_managed_pages(zone) / 1024;
7590 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7591 zone->_watermark[WMARK_MIN] = min_pages;
7594 * If it's a lowmem zone, reserve a number of pages
7595 * proportionate to the zone's size.
7597 zone->_watermark[WMARK_MIN] = tmp;
7601 * Set the kswapd watermarks distance according to the
7602 * scale factor in proportion to available memory, but
7603 * ensure a minimum size on small systems.
7605 tmp = max_t(u64, tmp >> 2,
7606 mult_frac(zone_managed_pages(zone),
7607 watermark_scale_factor, 10000));
7609 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7610 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7611 zone->watermark_boost = 0;
7613 spin_unlock_irqrestore(&zone->lock, flags);
7616 /* update totalreserve_pages */
7617 calculate_totalreserve_pages();
7621 * setup_per_zone_wmarks - called when min_free_kbytes changes
7622 * or when memory is hot-{added|removed}
7624 * Ensures that the watermark[min,low,high] values for each zone are set
7625 * correctly with respect to min_free_kbytes.
7627 void setup_per_zone_wmarks(void)
7629 static DEFINE_SPINLOCK(lock);
7632 __setup_per_zone_wmarks();
7637 * Initialise min_free_kbytes.
7639 * For small machines we want it small (128k min). For large machines
7640 * we want it large (64MB max). But it is not linear, because network
7641 * bandwidth does not increase linearly with machine size. We use
7643 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7644 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7660 int __meminit init_per_zone_wmark_min(void)
7662 unsigned long lowmem_kbytes;
7663 int new_min_free_kbytes;
7665 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7666 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7668 if (new_min_free_kbytes > user_min_free_kbytes) {
7669 min_free_kbytes = new_min_free_kbytes;
7670 if (min_free_kbytes < 128)
7671 min_free_kbytes = 128;
7672 if (min_free_kbytes > 65536)
7673 min_free_kbytes = 65536;
7675 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7676 new_min_free_kbytes, user_min_free_kbytes);
7678 setup_per_zone_wmarks();
7679 refresh_zone_stat_thresholds();
7680 setup_per_zone_lowmem_reserve();
7683 setup_min_unmapped_ratio();
7684 setup_min_slab_ratio();
7689 core_initcall(init_per_zone_wmark_min)
7692 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7693 * that we can call two helper functions whenever min_free_kbytes
7696 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7697 void __user *buffer, size_t *length, loff_t *ppos)
7701 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7706 user_min_free_kbytes = min_free_kbytes;
7707 setup_per_zone_wmarks();
7712 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7713 void __user *buffer, size_t *length, loff_t *ppos)
7717 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7724 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7725 void __user *buffer, size_t *length, loff_t *ppos)
7729 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7734 setup_per_zone_wmarks();
7740 static void setup_min_unmapped_ratio(void)
7745 for_each_online_pgdat(pgdat)
7746 pgdat->min_unmapped_pages = 0;
7749 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7750 sysctl_min_unmapped_ratio) / 100;
7754 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7755 void __user *buffer, size_t *length, loff_t *ppos)
7759 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7763 setup_min_unmapped_ratio();
7768 static void setup_min_slab_ratio(void)
7773 for_each_online_pgdat(pgdat)
7774 pgdat->min_slab_pages = 0;
7777 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7778 sysctl_min_slab_ratio) / 100;
7781 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7782 void __user *buffer, size_t *length, loff_t *ppos)
7786 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7790 setup_min_slab_ratio();
7797 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7798 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7799 * whenever sysctl_lowmem_reserve_ratio changes.
7801 * The reserve ratio obviously has absolutely no relation with the
7802 * minimum watermarks. The lowmem reserve ratio can only make sense
7803 * if in function of the boot time zone sizes.
7805 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7806 void __user *buffer, size_t *length, loff_t *ppos)
7808 proc_dointvec_minmax(table, write, buffer, length, ppos);
7809 setup_per_zone_lowmem_reserve();
7814 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7815 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7816 * pagelist can have before it gets flushed back to buddy allocator.
7818 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7819 void __user *buffer, size_t *length, loff_t *ppos)
7822 int old_percpu_pagelist_fraction;
7825 mutex_lock(&pcp_batch_high_lock);
7826 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7828 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7829 if (!write || ret < 0)
7832 /* Sanity checking to avoid pcp imbalance */
7833 if (percpu_pagelist_fraction &&
7834 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7835 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7841 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7844 for_each_populated_zone(zone) {
7847 for_each_possible_cpu(cpu)
7848 pageset_set_high_and_batch(zone,
7849 per_cpu_ptr(zone->pageset, cpu));
7852 mutex_unlock(&pcp_batch_high_lock);
7857 int hashdist = HASHDIST_DEFAULT;
7859 static int __init set_hashdist(char *str)
7863 hashdist = simple_strtoul(str, &str, 0);
7866 __setup("hashdist=", set_hashdist);
7869 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7871 * Returns the number of pages that arch has reserved but
7872 * is not known to alloc_large_system_hash().
7874 static unsigned long __init arch_reserved_kernel_pages(void)
7881 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7882 * machines. As memory size is increased the scale is also increased but at
7883 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7884 * quadruples the scale is increased by one, which means the size of hash table
7885 * only doubles, instead of quadrupling as well.
7886 * Because 32-bit systems cannot have large physical memory, where this scaling
7887 * makes sense, it is disabled on such platforms.
7889 #if __BITS_PER_LONG > 32
7890 #define ADAPT_SCALE_BASE (64ul << 30)
7891 #define ADAPT_SCALE_SHIFT 2
7892 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7896 * allocate a large system hash table from bootmem
7897 * - it is assumed that the hash table must contain an exact power-of-2
7898 * quantity of entries
7899 * - limit is the number of hash buckets, not the total allocation size
7901 void *__init alloc_large_system_hash(const char *tablename,
7902 unsigned long bucketsize,
7903 unsigned long numentries,
7906 unsigned int *_hash_shift,
7907 unsigned int *_hash_mask,
7908 unsigned long low_limit,
7909 unsigned long high_limit)
7911 unsigned long long max = high_limit;
7912 unsigned long log2qty, size;
7916 /* allow the kernel cmdline to have a say */
7918 /* round applicable memory size up to nearest megabyte */
7919 numentries = nr_kernel_pages;
7920 numentries -= arch_reserved_kernel_pages();
7922 /* It isn't necessary when PAGE_SIZE >= 1MB */
7923 if (PAGE_SHIFT < 20)
7924 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7926 #if __BITS_PER_LONG > 32
7928 unsigned long adapt;
7930 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7931 adapt <<= ADAPT_SCALE_SHIFT)
7936 /* limit to 1 bucket per 2^scale bytes of low memory */
7937 if (scale > PAGE_SHIFT)
7938 numentries >>= (scale - PAGE_SHIFT);
7940 numentries <<= (PAGE_SHIFT - scale);
7942 /* Make sure we've got at least a 0-order allocation.. */
7943 if (unlikely(flags & HASH_SMALL)) {
7944 /* Makes no sense without HASH_EARLY */
7945 WARN_ON(!(flags & HASH_EARLY));
7946 if (!(numentries >> *_hash_shift)) {
7947 numentries = 1UL << *_hash_shift;
7948 BUG_ON(!numentries);
7950 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7951 numentries = PAGE_SIZE / bucketsize;
7953 numentries = roundup_pow_of_two(numentries);
7955 /* limit allocation size to 1/16 total memory by default */
7957 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7958 do_div(max, bucketsize);
7960 max = min(max, 0x80000000ULL);
7962 if (numentries < low_limit)
7963 numentries = low_limit;
7964 if (numentries > max)
7967 log2qty = ilog2(numentries);
7969 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7971 size = bucketsize << log2qty;
7972 if (flags & HASH_EARLY) {
7973 if (flags & HASH_ZERO)
7974 table = memblock_alloc(size, SMP_CACHE_BYTES);
7976 table = memblock_alloc_raw(size,
7978 } else if (hashdist) {
7979 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7982 * If bucketsize is not a power-of-two, we may free
7983 * some pages at the end of hash table which
7984 * alloc_pages_exact() automatically does
7986 if (get_order(size) < MAX_ORDER) {
7987 table = alloc_pages_exact(size, gfp_flags);
7988 kmemleak_alloc(table, size, 1, gfp_flags);
7991 } while (!table && size > PAGE_SIZE && --log2qty);
7994 panic("Failed to allocate %s hash table\n", tablename);
7996 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7997 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
8000 *_hash_shift = log2qty;
8002 *_hash_mask = (1 << log2qty) - 1;
8008 * This function checks whether pageblock includes unmovable pages or not.
8009 * If @count is not zero, it is okay to include less @count unmovable pages
8011 * PageLRU check without isolation or lru_lock could race so that
8012 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8013 * check without lock_page also may miss some movable non-lru pages at
8014 * race condition. So you can't expect this function should be exact.
8016 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
8017 int migratetype, int flags)
8019 unsigned long found;
8020 unsigned long iter = 0;
8021 unsigned long pfn = page_to_pfn(page);
8022 const char *reason = "unmovable page";
8025 * TODO we could make this much more efficient by not checking every
8026 * page in the range if we know all of them are in MOVABLE_ZONE and
8027 * that the movable zone guarantees that pages are migratable but
8028 * the later is not the case right now unfortunatelly. E.g. movablecore
8029 * can still lead to having bootmem allocations in zone_movable.
8032 if (is_migrate_cma_page(page)) {
8034 * CMA allocations (alloc_contig_range) really need to mark
8035 * isolate CMA pageblocks even when they are not movable in fact
8036 * so consider them movable here.
8038 if (is_migrate_cma(migratetype))
8041 reason = "CMA page";
8045 for (found = 0; iter < pageblock_nr_pages; iter++) {
8046 unsigned long check = pfn + iter;
8048 if (!pfn_valid_within(check))
8051 page = pfn_to_page(check);
8053 if (PageReserved(page))
8057 * If the zone is movable and we have ruled out all reserved
8058 * pages then it should be reasonably safe to assume the rest
8061 if (zone_idx(zone) == ZONE_MOVABLE)
8065 * Hugepages are not in LRU lists, but they're movable.
8066 * We need not scan over tail pages because we don't
8067 * handle each tail page individually in migration.
8069 if (PageHuge(page)) {
8070 struct page *head = compound_head(page);
8071 unsigned int skip_pages;
8073 if (!hugepage_migration_supported(page_hstate(head)))
8076 skip_pages = (1 << compound_order(head)) - (page - head);
8077 iter += skip_pages - 1;
8082 * We can't use page_count without pin a page
8083 * because another CPU can free compound page.
8084 * This check already skips compound tails of THP
8085 * because their page->_refcount is zero at all time.
8087 if (!page_ref_count(page)) {
8088 if (PageBuddy(page))
8089 iter += (1 << page_order(page)) - 1;
8094 * The HWPoisoned page may be not in buddy system, and
8095 * page_count() is not 0.
8097 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
8100 if (__PageMovable(page))
8106 * If there are RECLAIMABLE pages, we need to check
8107 * it. But now, memory offline itself doesn't call
8108 * shrink_node_slabs() and it still to be fixed.
8111 * If the page is not RAM, page_count()should be 0.
8112 * we don't need more check. This is an _used_ not-movable page.
8114 * The problematic thing here is PG_reserved pages. PG_reserved
8115 * is set to both of a memory hole page and a _used_ kernel
8123 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8124 if (flags & REPORT_FAILURE)
8125 dump_page(pfn_to_page(pfn + iter), reason);
8129 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
8131 static unsigned long pfn_max_align_down(unsigned long pfn)
8133 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8134 pageblock_nr_pages) - 1);
8137 static unsigned long pfn_max_align_up(unsigned long pfn)
8139 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8140 pageblock_nr_pages));
8143 /* [start, end) must belong to a single zone. */
8144 static int __alloc_contig_migrate_range(struct compact_control *cc,
8145 unsigned long start, unsigned long end)
8147 /* This function is based on compact_zone() from compaction.c. */
8148 unsigned long nr_reclaimed;
8149 unsigned long pfn = start;
8150 unsigned int tries = 0;
8155 while (pfn < end || !list_empty(&cc->migratepages)) {
8156 if (fatal_signal_pending(current)) {
8161 if (list_empty(&cc->migratepages)) {
8162 cc->nr_migratepages = 0;
8163 pfn = isolate_migratepages_range(cc, pfn, end);
8169 } else if (++tries == 5) {
8170 ret = ret < 0 ? ret : -EBUSY;
8174 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8176 cc->nr_migratepages -= nr_reclaimed;
8178 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8179 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8182 putback_movable_pages(&cc->migratepages);
8189 * alloc_contig_range() -- tries to allocate given range of pages
8190 * @start: start PFN to allocate
8191 * @end: one-past-the-last PFN to allocate
8192 * @migratetype: migratetype of the underlaying pageblocks (either
8193 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8194 * in range must have the same migratetype and it must
8195 * be either of the two.
8196 * @gfp_mask: GFP mask to use during compaction
8198 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8199 * aligned. The PFN range must belong to a single zone.
8201 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8202 * pageblocks in the range. Once isolated, the pageblocks should not
8203 * be modified by others.
8205 * Return: zero on success or negative error code. On success all
8206 * pages which PFN is in [start, end) are allocated for the caller and
8207 * need to be freed with free_contig_range().
8209 int alloc_contig_range(unsigned long start, unsigned long end,
8210 unsigned migratetype, gfp_t gfp_mask)
8212 unsigned long outer_start, outer_end;
8216 struct compact_control cc = {
8217 .nr_migratepages = 0,
8219 .zone = page_zone(pfn_to_page(start)),
8220 .mode = MIGRATE_SYNC,
8221 .ignore_skip_hint = true,
8222 .no_set_skip_hint = true,
8223 .gfp_mask = current_gfp_context(gfp_mask),
8225 INIT_LIST_HEAD(&cc.migratepages);
8228 * What we do here is we mark all pageblocks in range as
8229 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8230 * have different sizes, and due to the way page allocator
8231 * work, we align the range to biggest of the two pages so
8232 * that page allocator won't try to merge buddies from
8233 * different pageblocks and change MIGRATE_ISOLATE to some
8234 * other migration type.
8236 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8237 * migrate the pages from an unaligned range (ie. pages that
8238 * we are interested in). This will put all the pages in
8239 * range back to page allocator as MIGRATE_ISOLATE.
8241 * When this is done, we take the pages in range from page
8242 * allocator removing them from the buddy system. This way
8243 * page allocator will never consider using them.
8245 * This lets us mark the pageblocks back as
8246 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8247 * aligned range but not in the unaligned, original range are
8248 * put back to page allocator so that buddy can use them.
8251 ret = start_isolate_page_range(pfn_max_align_down(start),
8252 pfn_max_align_up(end), migratetype, 0);
8257 * In case of -EBUSY, we'd like to know which page causes problem.
8258 * So, just fall through. test_pages_isolated() has a tracepoint
8259 * which will report the busy page.
8261 * It is possible that busy pages could become available before
8262 * the call to test_pages_isolated, and the range will actually be
8263 * allocated. So, if we fall through be sure to clear ret so that
8264 * -EBUSY is not accidentally used or returned to caller.
8266 ret = __alloc_contig_migrate_range(&cc, start, end);
8267 if (ret && ret != -EBUSY)
8272 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8273 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8274 * more, all pages in [start, end) are free in page allocator.
8275 * What we are going to do is to allocate all pages from
8276 * [start, end) (that is remove them from page allocator).
8278 * The only problem is that pages at the beginning and at the
8279 * end of interesting range may be not aligned with pages that
8280 * page allocator holds, ie. they can be part of higher order
8281 * pages. Because of this, we reserve the bigger range and
8282 * once this is done free the pages we are not interested in.
8284 * We don't have to hold zone->lock here because the pages are
8285 * isolated thus they won't get removed from buddy.
8288 lru_add_drain_all();
8291 outer_start = start;
8292 while (!PageBuddy(pfn_to_page(outer_start))) {
8293 if (++order >= MAX_ORDER) {
8294 outer_start = start;
8297 outer_start &= ~0UL << order;
8300 if (outer_start != start) {
8301 order = page_order(pfn_to_page(outer_start));
8304 * outer_start page could be small order buddy page and
8305 * it doesn't include start page. Adjust outer_start
8306 * in this case to report failed page properly
8307 * on tracepoint in test_pages_isolated()
8309 if (outer_start + (1UL << order) <= start)
8310 outer_start = start;
8313 /* Make sure the range is really isolated. */
8314 if (test_pages_isolated(outer_start, end, false)) {
8315 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8316 __func__, outer_start, end);
8321 /* Grab isolated pages from freelists. */
8322 outer_end = isolate_freepages_range(&cc, outer_start, end);
8328 /* Free head and tail (if any) */
8329 if (start != outer_start)
8330 free_contig_range(outer_start, start - outer_start);
8331 if (end != outer_end)
8332 free_contig_range(end, outer_end - end);
8335 undo_isolate_page_range(pfn_max_align_down(start),
8336 pfn_max_align_up(end), migratetype);
8340 void free_contig_range(unsigned long pfn, unsigned nr_pages)
8342 unsigned int count = 0;
8344 for (; nr_pages--; pfn++) {
8345 struct page *page = pfn_to_page(pfn);
8347 count += page_count(page) != 1;
8350 WARN(count != 0, "%d pages are still in use!\n", count);
8354 #ifdef CONFIG_MEMORY_HOTPLUG
8356 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8357 * page high values need to be recalulated.
8359 void __meminit zone_pcp_update(struct zone *zone)
8362 mutex_lock(&pcp_batch_high_lock);
8363 for_each_possible_cpu(cpu)
8364 pageset_set_high_and_batch(zone,
8365 per_cpu_ptr(zone->pageset, cpu));
8366 mutex_unlock(&pcp_batch_high_lock);
8370 void zone_pcp_reset(struct zone *zone)
8372 unsigned long flags;
8374 struct per_cpu_pageset *pset;
8376 /* avoid races with drain_pages() */
8377 local_irq_save(flags);
8378 if (zone->pageset != &boot_pageset) {
8379 for_each_online_cpu(cpu) {
8380 pset = per_cpu_ptr(zone->pageset, cpu);
8381 drain_zonestat(zone, pset);
8383 free_percpu(zone->pageset);
8384 zone->pageset = &boot_pageset;
8386 local_irq_restore(flags);
8389 #ifdef CONFIG_MEMORY_HOTREMOVE
8391 * All pages in the range must be in a single zone and isolated
8392 * before calling this.
8395 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8399 unsigned int order, i;
8401 unsigned long flags;
8402 /* find the first valid pfn */
8403 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8408 offline_mem_sections(pfn, end_pfn);
8409 zone = page_zone(pfn_to_page(pfn));
8410 spin_lock_irqsave(&zone->lock, flags);
8412 while (pfn < end_pfn) {
8413 if (!pfn_valid(pfn)) {
8417 page = pfn_to_page(pfn);
8419 * The HWPoisoned page may be not in buddy system, and
8420 * page_count() is not 0.
8422 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8424 SetPageReserved(page);
8428 BUG_ON(page_count(page));
8429 BUG_ON(!PageBuddy(page));
8430 order = page_order(page);
8431 #ifdef CONFIG_DEBUG_VM
8432 pr_info("remove from free list %lx %d %lx\n",
8433 pfn, 1 << order, end_pfn);
8435 list_del(&page->lru);
8436 rmv_page_order(page);
8437 zone->free_area[order].nr_free--;
8438 for (i = 0; i < (1 << order); i++)
8439 SetPageReserved((page+i));
8440 pfn += (1 << order);
8442 spin_unlock_irqrestore(&zone->lock, flags);
8446 bool is_free_buddy_page(struct page *page)
8448 struct zone *zone = page_zone(page);
8449 unsigned long pfn = page_to_pfn(page);
8450 unsigned long flags;
8453 spin_lock_irqsave(&zone->lock, flags);
8454 for (order = 0; order < MAX_ORDER; order++) {
8455 struct page *page_head = page - (pfn & ((1 << order) - 1));
8457 if (PageBuddy(page_head) && page_order(page_head) >= order)
8460 spin_unlock_irqrestore(&zone->lock, flags);
8462 return order < MAX_ORDER;
8465 #ifdef CONFIG_MEMORY_FAILURE
8467 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8468 * test is performed under the zone lock to prevent a race against page
8471 bool set_hwpoison_free_buddy_page(struct page *page)
8473 struct zone *zone = page_zone(page);
8474 unsigned long pfn = page_to_pfn(page);
8475 unsigned long flags;
8477 bool hwpoisoned = false;
8479 spin_lock_irqsave(&zone->lock, flags);
8480 for (order = 0; order < MAX_ORDER; order++) {
8481 struct page *page_head = page - (pfn & ((1 << order) - 1));
8483 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8484 if (!TestSetPageHWPoison(page))
8489 spin_unlock_irqrestore(&zone->lock, flags);