2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/highmem.h>
20 #include <linux/swap.h>
21 #include <linux/interrupt.h>
22 #include <linux/pagemap.h>
23 #include <linux/jiffies.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/vmstat.h>
43 #include <linux/mempolicy.h>
44 #include <linux/memremap.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_ext.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <trace/events/oom.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
62 #include <linux/sched/mm.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
65 #include <linux/memcontrol.h>
66 #include <linux/ftrace.h>
67 #include <linux/lockdep.h>
68 #include <linux/nmi.h>
69 #include <linux/psi.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node);
82 EXPORT_PER_CPU_SYMBOL(numa_node);
85 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
87 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
94 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96 int _node_numa_mem_[MAX_NUMNODES];
99 /* work_structs for global per-cpu drains */
102 struct work_struct work;
104 DEFINE_MUTEX(pcpu_drain_mutex);
105 DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
107 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
108 volatile unsigned long latent_entropy __latent_entropy;
109 EXPORT_SYMBOL(latent_entropy);
113 * Array of node states.
115 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
116 [N_POSSIBLE] = NODE_MASK_ALL,
117 [N_ONLINE] = { { [0] = 1UL } },
119 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
120 #ifdef CONFIG_HIGHMEM
121 [N_HIGH_MEMORY] = { { [0] = 1UL } },
123 [N_MEMORY] = { { [0] = 1UL } },
124 [N_CPU] = { { [0] = 1UL } },
127 EXPORT_SYMBOL(node_states);
129 atomic_long_t _totalram_pages __read_mostly;
130 EXPORT_SYMBOL(_totalram_pages);
131 unsigned long totalreserve_pages __read_mostly;
132 unsigned long totalcma_pages __read_mostly;
134 int percpu_pagelist_fraction;
135 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
138 * A cached value of the page's pageblock's migratetype, used when the page is
139 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
140 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
141 * Also the migratetype set in the page does not necessarily match the pcplist
142 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
143 * other index - this ensures that it will be put on the correct CMA freelist.
145 static inline int get_pcppage_migratetype(struct page *page)
150 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
152 page->index = migratetype;
155 #ifdef CONFIG_PM_SLEEP
157 * The following functions are used by the suspend/hibernate code to temporarily
158 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
159 * while devices are suspended. To avoid races with the suspend/hibernate code,
160 * they should always be called with system_transition_mutex held
161 * (gfp_allowed_mask also should only be modified with system_transition_mutex
162 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
163 * with that modification).
166 static gfp_t saved_gfp_mask;
168 void pm_restore_gfp_mask(void)
170 WARN_ON(!mutex_is_locked(&system_transition_mutex));
171 if (saved_gfp_mask) {
172 gfp_allowed_mask = saved_gfp_mask;
177 void pm_restrict_gfp_mask(void)
179 WARN_ON(!mutex_is_locked(&system_transition_mutex));
180 WARN_ON(saved_gfp_mask);
181 saved_gfp_mask = gfp_allowed_mask;
182 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
185 bool pm_suspended_storage(void)
187 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
191 #endif /* CONFIG_PM_SLEEP */
193 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
194 unsigned int pageblock_order __read_mostly;
197 static void __free_pages_ok(struct page *page, unsigned int order);
200 * results with 256, 32 in the lowmem_reserve sysctl:
201 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
202 * 1G machine -> (16M dma, 784M normal, 224M high)
203 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
204 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
205 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
207 * TBD: should special case ZONE_DMA32 machines here - in those we normally
208 * don't need any ZONE_NORMAL reservation
210 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
211 #ifdef CONFIG_ZONE_DMA
214 #ifdef CONFIG_ZONE_DMA32
218 #ifdef CONFIG_HIGHMEM
224 EXPORT_SYMBOL(totalram_pages);
226 static char * const zone_names[MAX_NR_ZONES] = {
227 #ifdef CONFIG_ZONE_DMA
230 #ifdef CONFIG_ZONE_DMA32
234 #ifdef CONFIG_HIGHMEM
238 #ifdef CONFIG_ZONE_DEVICE
243 const char * const migratetype_names[MIGRATE_TYPES] = {
251 #ifdef CONFIG_MEMORY_ISOLATION
256 compound_page_dtor * const compound_page_dtors[] = {
259 #ifdef CONFIG_HUGETLB_PAGE
262 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
267 int min_free_kbytes = 1024;
268 int user_min_free_kbytes = -1;
269 #ifdef CONFIG_DISCONTIGMEM
271 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
272 * are not on separate NUMA nodes. Functionally this works but with
273 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
274 * quite small. By default, do not boost watermarks on discontigmem as in
275 * many cases very high-order allocations like THP are likely to be
276 * unsupported and the premature reclaim offsets the advantage of long-term
277 * fragmentation avoidance.
279 int watermark_boost_factor __read_mostly;
281 int watermark_boost_factor __read_mostly = 15000;
283 int watermark_scale_factor = 10;
285 static unsigned long nr_kernel_pages __initdata;
286 static unsigned long nr_all_pages __initdata;
287 static unsigned long dma_reserve __initdata;
289 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
290 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
291 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
292 static unsigned long required_kernelcore __initdata;
293 static unsigned long required_kernelcore_percent __initdata;
294 static unsigned long required_movablecore __initdata;
295 static unsigned long required_movablecore_percent __initdata;
296 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
297 static bool mirrored_kernelcore __meminitdata;
299 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
301 EXPORT_SYMBOL(movable_zone);
302 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
305 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
306 unsigned int nr_online_nodes __read_mostly = 1;
307 EXPORT_SYMBOL(nr_node_ids);
308 EXPORT_SYMBOL(nr_online_nodes);
311 int page_group_by_mobility_disabled __read_mostly;
313 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
315 * During boot we initialize deferred pages on-demand, as needed, but once
316 * page_alloc_init_late() has finished, the deferred pages are all initialized,
317 * and we can permanently disable that path.
319 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
322 * Calling kasan_free_pages() only after deferred memory initialization
323 * has completed. Poisoning pages during deferred memory init will greatly
324 * lengthen the process and cause problem in large memory systems as the
325 * deferred pages initialization is done with interrupt disabled.
327 * Assuming that there will be no reference to those newly initialized
328 * pages before they are ever allocated, this should have no effect on
329 * KASAN memory tracking as the poison will be properly inserted at page
330 * allocation time. The only corner case is when pages are allocated by
331 * on-demand allocation and then freed again before the deferred pages
332 * initialization is done, but this is not likely to happen.
334 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
336 if (!static_branch_unlikely(&deferred_pages))
337 kasan_free_pages(page, order);
340 /* Returns true if the struct page for the pfn is uninitialised */
341 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
343 int nid = early_pfn_to_nid(pfn);
345 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
352 * Returns true when the remaining initialisation should be deferred until
353 * later in the boot cycle when it can be parallelised.
355 static bool __meminit
356 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
358 static unsigned long prev_end_pfn, nr_initialised;
361 * prev_end_pfn static that contains the end of previous zone
362 * No need to protect because called very early in boot before smp_init.
364 if (prev_end_pfn != end_pfn) {
365 prev_end_pfn = end_pfn;
369 /* Always populate low zones for address-constrained allocations */
370 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
374 * We start only with one section of pages, more pages are added as
375 * needed until the rest of deferred pages are initialized.
378 if ((nr_initialised > PAGES_PER_SECTION) &&
379 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
380 NODE_DATA(nid)->first_deferred_pfn = pfn;
386 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
388 static inline bool early_page_uninitialised(unsigned long pfn)
393 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
399 /* Return a pointer to the bitmap storing bits affecting a block of pages */
400 static inline unsigned long *get_pageblock_bitmap(struct page *page,
403 #ifdef CONFIG_SPARSEMEM
404 return __pfn_to_section(pfn)->pageblock_flags;
406 return page_zone(page)->pageblock_flags;
407 #endif /* CONFIG_SPARSEMEM */
410 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
412 #ifdef CONFIG_SPARSEMEM
413 pfn &= (PAGES_PER_SECTION-1);
414 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
416 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
417 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
418 #endif /* CONFIG_SPARSEMEM */
422 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
423 * @page: The page within the block of interest
424 * @pfn: The target page frame number
425 * @end_bitidx: The last bit of interest to retrieve
426 * @mask: mask of bits that the caller is interested in
428 * Return: pageblock_bits flags
430 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
432 unsigned long end_bitidx,
435 unsigned long *bitmap;
436 unsigned long bitidx, word_bitidx;
439 bitmap = get_pageblock_bitmap(page, pfn);
440 bitidx = pfn_to_bitidx(page, pfn);
441 word_bitidx = bitidx / BITS_PER_LONG;
442 bitidx &= (BITS_PER_LONG-1);
444 word = bitmap[word_bitidx];
445 bitidx += end_bitidx;
446 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
449 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
450 unsigned long end_bitidx,
453 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
456 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
458 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
462 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
463 * @page: The page within the block of interest
464 * @flags: The flags to set
465 * @pfn: The target page frame number
466 * @end_bitidx: The last bit of interest
467 * @mask: mask of bits that the caller is interested in
469 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
471 unsigned long end_bitidx,
474 unsigned long *bitmap;
475 unsigned long bitidx, word_bitidx;
476 unsigned long old_word, word;
478 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
479 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
481 bitmap = get_pageblock_bitmap(page, pfn);
482 bitidx = pfn_to_bitidx(page, pfn);
483 word_bitidx = bitidx / BITS_PER_LONG;
484 bitidx &= (BITS_PER_LONG-1);
486 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
488 bitidx += end_bitidx;
489 mask <<= (BITS_PER_LONG - bitidx - 1);
490 flags <<= (BITS_PER_LONG - bitidx - 1);
492 word = READ_ONCE(bitmap[word_bitidx]);
494 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
495 if (word == old_word)
501 void set_pageblock_migratetype(struct page *page, int migratetype)
503 if (unlikely(page_group_by_mobility_disabled &&
504 migratetype < MIGRATE_PCPTYPES))
505 migratetype = MIGRATE_UNMOVABLE;
507 set_pageblock_flags_group(page, (unsigned long)migratetype,
508 PB_migrate, PB_migrate_end);
511 #ifdef CONFIG_DEBUG_VM
512 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
516 unsigned long pfn = page_to_pfn(page);
517 unsigned long sp, start_pfn;
520 seq = zone_span_seqbegin(zone);
521 start_pfn = zone->zone_start_pfn;
522 sp = zone->spanned_pages;
523 if (!zone_spans_pfn(zone, pfn))
525 } while (zone_span_seqretry(zone, seq));
528 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
529 pfn, zone_to_nid(zone), zone->name,
530 start_pfn, start_pfn + sp);
535 static int page_is_consistent(struct zone *zone, struct page *page)
537 if (!pfn_valid_within(page_to_pfn(page)))
539 if (zone != page_zone(page))
545 * Temporary debugging check for pages not lying within a given zone.
547 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
549 if (page_outside_zone_boundaries(zone, page))
551 if (!page_is_consistent(zone, page))
557 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
563 static void bad_page(struct page *page, const char *reason,
564 unsigned long bad_flags)
566 static unsigned long resume;
567 static unsigned long nr_shown;
568 static unsigned long nr_unshown;
571 * Allow a burst of 60 reports, then keep quiet for that minute;
572 * or allow a steady drip of one report per second.
574 if (nr_shown == 60) {
575 if (time_before(jiffies, resume)) {
581 "BUG: Bad page state: %lu messages suppressed\n",
588 resume = jiffies + 60 * HZ;
590 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
591 current->comm, page_to_pfn(page));
592 __dump_page(page, reason);
593 bad_flags &= page->flags;
595 pr_alert("bad because of flags: %#lx(%pGp)\n",
596 bad_flags, &bad_flags);
597 dump_page_owner(page);
602 /* Leave bad fields for debug, except PageBuddy could make trouble */
603 page_mapcount_reset(page); /* remove PageBuddy */
604 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
608 * Higher-order pages are called "compound pages". They are structured thusly:
610 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
612 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
613 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
615 * The first tail page's ->compound_dtor holds the offset in array of compound
616 * page destructors. See compound_page_dtors.
618 * The first tail page's ->compound_order holds the order of allocation.
619 * This usage means that zero-order pages may not be compound.
622 void free_compound_page(struct page *page)
624 __free_pages_ok(page, compound_order(page));
627 void prep_compound_page(struct page *page, unsigned int order)
630 int nr_pages = 1 << order;
632 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
633 set_compound_order(page, order);
635 for (i = 1; i < nr_pages; i++) {
636 struct page *p = page + i;
637 set_page_count(p, 0);
638 p->mapping = TAIL_MAPPING;
639 set_compound_head(p, page);
641 atomic_set(compound_mapcount_ptr(page), -1);
644 #ifdef CONFIG_DEBUG_PAGEALLOC
645 unsigned int _debug_guardpage_minorder;
646 bool _debug_pagealloc_enabled __read_mostly
647 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
648 EXPORT_SYMBOL(_debug_pagealloc_enabled);
649 bool _debug_guardpage_enabled __read_mostly;
651 static int __init early_debug_pagealloc(char *buf)
655 return kstrtobool(buf, &_debug_pagealloc_enabled);
657 early_param("debug_pagealloc", early_debug_pagealloc);
659 static bool need_debug_guardpage(void)
661 /* If we don't use debug_pagealloc, we don't need guard page */
662 if (!debug_pagealloc_enabled())
665 if (!debug_guardpage_minorder())
671 static void init_debug_guardpage(void)
673 if (!debug_pagealloc_enabled())
676 if (!debug_guardpage_minorder())
679 _debug_guardpage_enabled = true;
682 struct page_ext_operations debug_guardpage_ops = {
683 .need = need_debug_guardpage,
684 .init = init_debug_guardpage,
687 static int __init debug_guardpage_minorder_setup(char *buf)
691 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
692 pr_err("Bad debug_guardpage_minorder value\n");
695 _debug_guardpage_minorder = res;
696 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
699 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
701 static inline bool set_page_guard(struct zone *zone, struct page *page,
702 unsigned int order, int migratetype)
704 struct page_ext *page_ext;
706 if (!debug_guardpage_enabled())
709 if (order >= debug_guardpage_minorder())
712 page_ext = lookup_page_ext(page);
713 if (unlikely(!page_ext))
716 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
718 INIT_LIST_HEAD(&page->lru);
719 set_page_private(page, order);
720 /* Guard pages are not available for any usage */
721 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
726 static inline void clear_page_guard(struct zone *zone, struct page *page,
727 unsigned int order, int migratetype)
729 struct page_ext *page_ext;
731 if (!debug_guardpage_enabled())
734 page_ext = lookup_page_ext(page);
735 if (unlikely(!page_ext))
738 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
740 set_page_private(page, 0);
741 if (!is_migrate_isolate(migratetype))
742 __mod_zone_freepage_state(zone, (1 << order), migratetype);
745 struct page_ext_operations debug_guardpage_ops;
746 static inline bool set_page_guard(struct zone *zone, struct page *page,
747 unsigned int order, int migratetype) { return false; }
748 static inline void clear_page_guard(struct zone *zone, struct page *page,
749 unsigned int order, int migratetype) {}
752 static inline void set_page_order(struct page *page, unsigned int order)
754 set_page_private(page, order);
755 __SetPageBuddy(page);
758 static inline void rmv_page_order(struct page *page)
760 __ClearPageBuddy(page);
761 set_page_private(page, 0);
765 * This function checks whether a page is free && is the buddy
766 * we can coalesce a page and its buddy if
767 * (a) the buddy is not in a hole (check before calling!) &&
768 * (b) the buddy is in the buddy system &&
769 * (c) a page and its buddy have the same order &&
770 * (d) a page and its buddy are in the same zone.
772 * For recording whether a page is in the buddy system, we set PageBuddy.
773 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
775 * For recording page's order, we use page_private(page).
777 static inline int page_is_buddy(struct page *page, struct page *buddy,
780 if (page_is_guard(buddy) && page_order(buddy) == order) {
781 if (page_zone_id(page) != page_zone_id(buddy))
784 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
789 if (PageBuddy(buddy) && page_order(buddy) == order) {
791 * zone check is done late to avoid uselessly
792 * calculating zone/node ids for pages that could
795 if (page_zone_id(page) != page_zone_id(buddy))
798 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
805 #ifdef CONFIG_COMPACTION
806 static inline struct capture_control *task_capc(struct zone *zone)
808 struct capture_control *capc = current->capture_control;
811 !(current->flags & PF_KTHREAD) &&
813 capc->cc->zone == zone &&
814 capc->cc->direct_compaction ? capc : NULL;
818 compaction_capture(struct capture_control *capc, struct page *page,
819 int order, int migratetype)
821 if (!capc || order != capc->cc->order)
824 /* Do not accidentally pollute CMA or isolated regions*/
825 if (is_migrate_cma(migratetype) ||
826 is_migrate_isolate(migratetype))
830 * Do not let lower order allocations polluate a movable pageblock.
831 * This might let an unmovable request use a reclaimable pageblock
832 * and vice-versa but no more than normal fallback logic which can
833 * have trouble finding a high-order free page.
835 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
843 static inline struct capture_control *task_capc(struct zone *zone)
849 compaction_capture(struct capture_control *capc, struct page *page,
850 int order, int migratetype)
854 #endif /* CONFIG_COMPACTION */
857 * Freeing function for a buddy system allocator.
859 * The concept of a buddy system is to maintain direct-mapped table
860 * (containing bit values) for memory blocks of various "orders".
861 * The bottom level table contains the map for the smallest allocatable
862 * units of memory (here, pages), and each level above it describes
863 * pairs of units from the levels below, hence, "buddies".
864 * At a high level, all that happens here is marking the table entry
865 * at the bottom level available, and propagating the changes upward
866 * as necessary, plus some accounting needed to play nicely with other
867 * parts of the VM system.
868 * At each level, we keep a list of pages, which are heads of continuous
869 * free pages of length of (1 << order) and marked with PageBuddy.
870 * Page's order is recorded in page_private(page) field.
871 * So when we are allocating or freeing one, we can derive the state of the
872 * other. That is, if we allocate a small block, and both were
873 * free, the remainder of the region must be split into blocks.
874 * If a block is freed, and its buddy is also free, then this
875 * triggers coalescing into a block of larger size.
880 static inline void __free_one_page(struct page *page,
882 struct zone *zone, unsigned int order,
885 unsigned long combined_pfn;
886 unsigned long uninitialized_var(buddy_pfn);
888 unsigned int max_order;
889 struct capture_control *capc = task_capc(zone);
891 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
893 VM_BUG_ON(!zone_is_initialized(zone));
894 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
896 VM_BUG_ON(migratetype == -1);
897 if (likely(!is_migrate_isolate(migratetype)))
898 __mod_zone_freepage_state(zone, 1 << order, migratetype);
900 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
901 VM_BUG_ON_PAGE(bad_range(zone, page), page);
904 while (order < max_order - 1) {
905 if (compaction_capture(capc, page, order, migratetype)) {
906 __mod_zone_freepage_state(zone, -(1 << order),
910 buddy_pfn = __find_buddy_pfn(pfn, order);
911 buddy = page + (buddy_pfn - pfn);
913 if (!pfn_valid_within(buddy_pfn))
915 if (!page_is_buddy(page, buddy, order))
918 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
919 * merge with it and move up one order.
921 if (page_is_guard(buddy)) {
922 clear_page_guard(zone, buddy, order, migratetype);
924 list_del(&buddy->lru);
925 zone->free_area[order].nr_free--;
926 rmv_page_order(buddy);
928 combined_pfn = buddy_pfn & pfn;
929 page = page + (combined_pfn - pfn);
933 if (max_order < MAX_ORDER) {
934 /* If we are here, it means order is >= pageblock_order.
935 * We want to prevent merge between freepages on isolate
936 * pageblock and normal pageblock. Without this, pageblock
937 * isolation could cause incorrect freepage or CMA accounting.
939 * We don't want to hit this code for the more frequent
942 if (unlikely(has_isolate_pageblock(zone))) {
945 buddy_pfn = __find_buddy_pfn(pfn, order);
946 buddy = page + (buddy_pfn - pfn);
947 buddy_mt = get_pageblock_migratetype(buddy);
949 if (migratetype != buddy_mt
950 && (is_migrate_isolate(migratetype) ||
951 is_migrate_isolate(buddy_mt)))
955 goto continue_merging;
959 set_page_order(page, order);
962 * If this is not the largest possible page, check if the buddy
963 * of the next-highest order is free. If it is, it's possible
964 * that pages are being freed that will coalesce soon. In case,
965 * that is happening, add the free page to the tail of the list
966 * so it's less likely to be used soon and more likely to be merged
967 * as a higher order page
969 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
970 struct page *higher_page, *higher_buddy;
971 combined_pfn = buddy_pfn & pfn;
972 higher_page = page + (combined_pfn - pfn);
973 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
974 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
975 if (pfn_valid_within(buddy_pfn) &&
976 page_is_buddy(higher_page, higher_buddy, order + 1)) {
977 list_add_tail(&page->lru,
978 &zone->free_area[order].free_list[migratetype]);
983 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
985 zone->free_area[order].nr_free++;
989 * A bad page could be due to a number of fields. Instead of multiple branches,
990 * try and check multiple fields with one check. The caller must do a detailed
991 * check if necessary.
993 static inline bool page_expected_state(struct page *page,
994 unsigned long check_flags)
996 if (unlikely(atomic_read(&page->_mapcount) != -1))
999 if (unlikely((unsigned long)page->mapping |
1000 page_ref_count(page) |
1002 (unsigned long)page->mem_cgroup |
1004 (page->flags & check_flags)))
1010 static void free_pages_check_bad(struct page *page)
1012 const char *bad_reason;
1013 unsigned long bad_flags;
1018 if (unlikely(atomic_read(&page->_mapcount) != -1))
1019 bad_reason = "nonzero mapcount";
1020 if (unlikely(page->mapping != NULL))
1021 bad_reason = "non-NULL mapping";
1022 if (unlikely(page_ref_count(page) != 0))
1023 bad_reason = "nonzero _refcount";
1024 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1025 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1026 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1029 if (unlikely(page->mem_cgroup))
1030 bad_reason = "page still charged to cgroup";
1032 bad_page(page, bad_reason, bad_flags);
1035 static inline int free_pages_check(struct page *page)
1037 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1040 /* Something has gone sideways, find it */
1041 free_pages_check_bad(page);
1045 static int free_tail_pages_check(struct page *head_page, struct page *page)
1050 * We rely page->lru.next never has bit 0 set, unless the page
1051 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1053 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1055 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1059 switch (page - head_page) {
1061 /* the first tail page: ->mapping may be compound_mapcount() */
1062 if (unlikely(compound_mapcount(page))) {
1063 bad_page(page, "nonzero compound_mapcount", 0);
1069 * the second tail page: ->mapping is
1070 * deferred_list.next -- ignore value.
1074 if (page->mapping != TAIL_MAPPING) {
1075 bad_page(page, "corrupted mapping in tail page", 0);
1080 if (unlikely(!PageTail(page))) {
1081 bad_page(page, "PageTail not set", 0);
1084 if (unlikely(compound_head(page) != head_page)) {
1085 bad_page(page, "compound_head not consistent", 0);
1090 page->mapping = NULL;
1091 clear_compound_head(page);
1095 static __always_inline bool free_pages_prepare(struct page *page,
1096 unsigned int order, bool check_free)
1100 VM_BUG_ON_PAGE(PageTail(page), page);
1102 trace_mm_page_free(page, order);
1105 * Check tail pages before head page information is cleared to
1106 * avoid checking PageCompound for order-0 pages.
1108 if (unlikely(order)) {
1109 bool compound = PageCompound(page);
1112 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1115 ClearPageDoubleMap(page);
1116 for (i = 1; i < (1 << order); i++) {
1118 bad += free_tail_pages_check(page, page + i);
1119 if (unlikely(free_pages_check(page + i))) {
1123 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1126 if (PageMappingFlags(page))
1127 page->mapping = NULL;
1128 if (memcg_kmem_enabled() && PageKmemcg(page))
1129 __memcg_kmem_uncharge(page, order);
1131 bad += free_pages_check(page);
1135 page_cpupid_reset_last(page);
1136 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1137 reset_page_owner(page, order);
1139 if (!PageHighMem(page)) {
1140 debug_check_no_locks_freed(page_address(page),
1141 PAGE_SIZE << order);
1142 debug_check_no_obj_freed(page_address(page),
1143 PAGE_SIZE << order);
1145 arch_free_page(page, order);
1146 kernel_poison_pages(page, 1 << order, 0);
1147 if (debug_pagealloc_enabled())
1148 kernel_map_pages(page, 1 << order, 0);
1150 kasan_free_nondeferred_pages(page, order);
1155 #ifdef CONFIG_DEBUG_VM
1156 static inline bool free_pcp_prepare(struct page *page)
1158 return free_pages_prepare(page, 0, true);
1161 static inline bool bulkfree_pcp_prepare(struct page *page)
1166 static bool free_pcp_prepare(struct page *page)
1168 return free_pages_prepare(page, 0, false);
1171 static bool bulkfree_pcp_prepare(struct page *page)
1173 return free_pages_check(page);
1175 #endif /* CONFIG_DEBUG_VM */
1177 static inline void prefetch_buddy(struct page *page)
1179 unsigned long pfn = page_to_pfn(page);
1180 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1181 struct page *buddy = page + (buddy_pfn - pfn);
1187 * Frees a number of pages from the PCP lists
1188 * Assumes all pages on list are in same zone, and of same order.
1189 * count is the number of pages to free.
1191 * If the zone was previously in an "all pages pinned" state then look to
1192 * see if this freeing clears that state.
1194 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1195 * pinned" detection logic.
1197 static void free_pcppages_bulk(struct zone *zone, int count,
1198 struct per_cpu_pages *pcp)
1200 int migratetype = 0;
1202 int prefetch_nr = 0;
1203 bool isolated_pageblocks;
1204 struct page *page, *tmp;
1208 struct list_head *list;
1211 * Remove pages from lists in a round-robin fashion. A
1212 * batch_free count is maintained that is incremented when an
1213 * empty list is encountered. This is so more pages are freed
1214 * off fuller lists instead of spinning excessively around empty
1219 if (++migratetype == MIGRATE_PCPTYPES)
1221 list = &pcp->lists[migratetype];
1222 } while (list_empty(list));
1224 /* This is the only non-empty list. Free them all. */
1225 if (batch_free == MIGRATE_PCPTYPES)
1229 page = list_last_entry(list, struct page, lru);
1230 /* must delete to avoid corrupting pcp list */
1231 list_del(&page->lru);
1234 if (bulkfree_pcp_prepare(page))
1237 list_add_tail(&page->lru, &head);
1240 * We are going to put the page back to the global
1241 * pool, prefetch its buddy to speed up later access
1242 * under zone->lock. It is believed the overhead of
1243 * an additional test and calculating buddy_pfn here
1244 * can be offset by reduced memory latency later. To
1245 * avoid excessive prefetching due to large count, only
1246 * prefetch buddy for the first pcp->batch nr of pages.
1248 if (prefetch_nr++ < pcp->batch)
1249 prefetch_buddy(page);
1250 } while (--count && --batch_free && !list_empty(list));
1253 spin_lock(&zone->lock);
1254 isolated_pageblocks = has_isolate_pageblock(zone);
1257 * Use safe version since after __free_one_page(),
1258 * page->lru.next will not point to original list.
1260 list_for_each_entry_safe(page, tmp, &head, lru) {
1261 int mt = get_pcppage_migratetype(page);
1262 /* MIGRATE_ISOLATE page should not go to pcplists */
1263 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1264 /* Pageblock could have been isolated meanwhile */
1265 if (unlikely(isolated_pageblocks))
1266 mt = get_pageblock_migratetype(page);
1268 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1269 trace_mm_page_pcpu_drain(page, 0, mt);
1271 spin_unlock(&zone->lock);
1274 static void free_one_page(struct zone *zone,
1275 struct page *page, unsigned long pfn,
1279 spin_lock(&zone->lock);
1280 if (unlikely(has_isolate_pageblock(zone) ||
1281 is_migrate_isolate(migratetype))) {
1282 migratetype = get_pfnblock_migratetype(page, pfn);
1284 __free_one_page(page, pfn, zone, order, migratetype);
1285 spin_unlock(&zone->lock);
1288 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1289 unsigned long zone, int nid)
1291 mm_zero_struct_page(page);
1292 set_page_links(page, zone, nid, pfn);
1293 init_page_count(page);
1294 page_mapcount_reset(page);
1295 page_cpupid_reset_last(page);
1296 page_kasan_tag_reset(page);
1298 INIT_LIST_HEAD(&page->lru);
1299 #ifdef WANT_PAGE_VIRTUAL
1300 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1301 if (!is_highmem_idx(zone))
1302 set_page_address(page, __va(pfn << PAGE_SHIFT));
1306 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1307 static void __meminit init_reserved_page(unsigned long pfn)
1312 if (!early_page_uninitialised(pfn))
1315 nid = early_pfn_to_nid(pfn);
1316 pgdat = NODE_DATA(nid);
1318 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1319 struct zone *zone = &pgdat->node_zones[zid];
1321 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1324 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1327 static inline void init_reserved_page(unsigned long pfn)
1330 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1333 * Initialised pages do not have PageReserved set. This function is
1334 * called for each range allocated by the bootmem allocator and
1335 * marks the pages PageReserved. The remaining valid pages are later
1336 * sent to the buddy page allocator.
1338 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1340 unsigned long start_pfn = PFN_DOWN(start);
1341 unsigned long end_pfn = PFN_UP(end);
1343 for (; start_pfn < end_pfn; start_pfn++) {
1344 if (pfn_valid(start_pfn)) {
1345 struct page *page = pfn_to_page(start_pfn);
1347 init_reserved_page(start_pfn);
1349 /* Avoid false-positive PageTail() */
1350 INIT_LIST_HEAD(&page->lru);
1353 * no need for atomic set_bit because the struct
1354 * page is not visible yet so nobody should
1357 __SetPageReserved(page);
1362 static void __free_pages_ok(struct page *page, unsigned int order)
1364 unsigned long flags;
1366 unsigned long pfn = page_to_pfn(page);
1368 if (!free_pages_prepare(page, order, true))
1371 migratetype = get_pfnblock_migratetype(page, pfn);
1372 local_irq_save(flags);
1373 __count_vm_events(PGFREE, 1 << order);
1374 free_one_page(page_zone(page), page, pfn, order, migratetype);
1375 local_irq_restore(flags);
1378 void __free_pages_core(struct page *page, unsigned int order)
1380 unsigned int nr_pages = 1 << order;
1381 struct page *p = page;
1385 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1387 __ClearPageReserved(p);
1388 set_page_count(p, 0);
1390 __ClearPageReserved(p);
1391 set_page_count(p, 0);
1393 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1394 set_page_refcounted(page);
1395 __free_pages(page, order);
1398 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1399 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1401 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1403 int __meminit early_pfn_to_nid(unsigned long pfn)
1405 static DEFINE_SPINLOCK(early_pfn_lock);
1408 spin_lock(&early_pfn_lock);
1409 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1411 nid = first_online_node;
1412 spin_unlock(&early_pfn_lock);
1418 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1419 /* Only safe to use early in boot when initialisation is single-threaded */
1420 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1424 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1425 if (nid >= 0 && nid != node)
1431 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1438 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1441 if (early_page_uninitialised(pfn))
1443 __free_pages_core(page, order);
1447 * Check that the whole (or subset of) a pageblock given by the interval of
1448 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1449 * with the migration of free compaction scanner. The scanners then need to
1450 * use only pfn_valid_within() check for arches that allow holes within
1453 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1455 * It's possible on some configurations to have a setup like node0 node1 node0
1456 * i.e. it's possible that all pages within a zones range of pages do not
1457 * belong to a single zone. We assume that a border between node0 and node1
1458 * can occur within a single pageblock, but not a node0 node1 node0
1459 * interleaving within a single pageblock. It is therefore sufficient to check
1460 * the first and last page of a pageblock and avoid checking each individual
1461 * page in a pageblock.
1463 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1464 unsigned long end_pfn, struct zone *zone)
1466 struct page *start_page;
1467 struct page *end_page;
1469 /* end_pfn is one past the range we are checking */
1472 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1475 start_page = pfn_to_online_page(start_pfn);
1479 if (page_zone(start_page) != zone)
1482 end_page = pfn_to_page(end_pfn);
1484 /* This gives a shorter code than deriving page_zone(end_page) */
1485 if (page_zone_id(start_page) != page_zone_id(end_page))
1491 void set_zone_contiguous(struct zone *zone)
1493 unsigned long block_start_pfn = zone->zone_start_pfn;
1494 unsigned long block_end_pfn;
1496 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1497 for (; block_start_pfn < zone_end_pfn(zone);
1498 block_start_pfn = block_end_pfn,
1499 block_end_pfn += pageblock_nr_pages) {
1501 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1503 if (!__pageblock_pfn_to_page(block_start_pfn,
1504 block_end_pfn, zone))
1508 /* We confirm that there is no hole */
1509 zone->contiguous = true;
1512 void clear_zone_contiguous(struct zone *zone)
1514 zone->contiguous = false;
1517 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1518 static void __init deferred_free_range(unsigned long pfn,
1519 unsigned long nr_pages)
1527 page = pfn_to_page(pfn);
1529 /* Free a large naturally-aligned chunk if possible */
1530 if (nr_pages == pageblock_nr_pages &&
1531 (pfn & (pageblock_nr_pages - 1)) == 0) {
1532 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1533 __free_pages_core(page, pageblock_order);
1537 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1538 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1539 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1540 __free_pages_core(page, 0);
1544 /* Completion tracking for deferred_init_memmap() threads */
1545 static atomic_t pgdat_init_n_undone __initdata;
1546 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1548 static inline void __init pgdat_init_report_one_done(void)
1550 if (atomic_dec_and_test(&pgdat_init_n_undone))
1551 complete(&pgdat_init_all_done_comp);
1555 * Returns true if page needs to be initialized or freed to buddy allocator.
1557 * First we check if pfn is valid on architectures where it is possible to have
1558 * holes within pageblock_nr_pages. On systems where it is not possible, this
1559 * function is optimized out.
1561 * Then, we check if a current large page is valid by only checking the validity
1564 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1566 if (!pfn_valid_within(pfn))
1568 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1574 * Free pages to buddy allocator. Try to free aligned pages in
1575 * pageblock_nr_pages sizes.
1577 static void __init deferred_free_pages(unsigned long pfn,
1578 unsigned long end_pfn)
1580 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1581 unsigned long nr_free = 0;
1583 for (; pfn < end_pfn; pfn++) {
1584 if (!deferred_pfn_valid(pfn)) {
1585 deferred_free_range(pfn - nr_free, nr_free);
1587 } else if (!(pfn & nr_pgmask)) {
1588 deferred_free_range(pfn - nr_free, nr_free);
1590 touch_nmi_watchdog();
1595 /* Free the last block of pages to allocator */
1596 deferred_free_range(pfn - nr_free, nr_free);
1600 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1601 * by performing it only once every pageblock_nr_pages.
1602 * Return number of pages initialized.
1604 static unsigned long __init deferred_init_pages(struct zone *zone,
1606 unsigned long end_pfn)
1608 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1609 int nid = zone_to_nid(zone);
1610 unsigned long nr_pages = 0;
1611 int zid = zone_idx(zone);
1612 struct page *page = NULL;
1614 for (; pfn < end_pfn; pfn++) {
1615 if (!deferred_pfn_valid(pfn)) {
1618 } else if (!page || !(pfn & nr_pgmask)) {
1619 page = pfn_to_page(pfn);
1620 touch_nmi_watchdog();
1624 __init_single_page(page, pfn, zid, nid);
1631 * This function is meant to pre-load the iterator for the zone init.
1632 * Specifically it walks through the ranges until we are caught up to the
1633 * first_init_pfn value and exits there. If we never encounter the value we
1634 * return false indicating there are no valid ranges left.
1637 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1638 unsigned long *spfn, unsigned long *epfn,
1639 unsigned long first_init_pfn)
1644 * Start out by walking through the ranges in this zone that have
1645 * already been initialized. We don't need to do anything with them
1646 * so we just need to flush them out of the system.
1648 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1649 if (*epfn <= first_init_pfn)
1651 if (*spfn < first_init_pfn)
1652 *spfn = first_init_pfn;
1661 * Initialize and free pages. We do it in two loops: first we initialize
1662 * struct page, then free to buddy allocator, because while we are
1663 * freeing pages we can access pages that are ahead (computing buddy
1664 * page in __free_one_page()).
1666 * In order to try and keep some memory in the cache we have the loop
1667 * broken along max page order boundaries. This way we will not cause
1668 * any issues with the buddy page computation.
1670 static unsigned long __init
1671 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1672 unsigned long *end_pfn)
1674 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1675 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1676 unsigned long nr_pages = 0;
1679 /* First we loop through and initialize the page values */
1680 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1683 if (mo_pfn <= *start_pfn)
1686 t = min(mo_pfn, *end_pfn);
1687 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1689 if (mo_pfn < *end_pfn) {
1690 *start_pfn = mo_pfn;
1695 /* Reset values and now loop through freeing pages as needed */
1698 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1704 t = min(mo_pfn, epfn);
1705 deferred_free_pages(spfn, t);
1714 /* Initialise remaining memory on a node */
1715 static int __init deferred_init_memmap(void *data)
1717 pg_data_t *pgdat = data;
1718 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1719 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1720 unsigned long first_init_pfn, flags;
1721 unsigned long start = jiffies;
1726 /* Bind memory initialisation thread to a local node if possible */
1727 if (!cpumask_empty(cpumask))
1728 set_cpus_allowed_ptr(current, cpumask);
1730 pgdat_resize_lock(pgdat, &flags);
1731 first_init_pfn = pgdat->first_deferred_pfn;
1732 if (first_init_pfn == ULONG_MAX) {
1733 pgdat_resize_unlock(pgdat, &flags);
1734 pgdat_init_report_one_done();
1738 /* Sanity check boundaries */
1739 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1740 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1741 pgdat->first_deferred_pfn = ULONG_MAX;
1743 /* Only the highest zone is deferred so find it */
1744 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1745 zone = pgdat->node_zones + zid;
1746 if (first_init_pfn < zone_end_pfn(zone))
1750 /* If the zone is empty somebody else may have cleared out the zone */
1751 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1756 * Initialize and free pages in MAX_ORDER sized increments so
1757 * that we can avoid introducing any issues with the buddy
1761 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1763 pgdat_resize_unlock(pgdat, &flags);
1765 /* Sanity check that the next zone really is unpopulated */
1766 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1768 pr_info("node %d initialised, %lu pages in %ums\n",
1769 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1771 pgdat_init_report_one_done();
1776 * If this zone has deferred pages, try to grow it by initializing enough
1777 * deferred pages to satisfy the allocation specified by order, rounded up to
1778 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1779 * of SECTION_SIZE bytes by initializing struct pages in increments of
1780 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1782 * Return true when zone was grown, otherwise return false. We return true even
1783 * when we grow less than requested, to let the caller decide if there are
1784 * enough pages to satisfy the allocation.
1786 * Note: We use noinline because this function is needed only during boot, and
1787 * it is called from a __ref function _deferred_grow_zone. This way we are
1788 * making sure that it is not inlined into permanent text section.
1790 static noinline bool __init
1791 deferred_grow_zone(struct zone *zone, unsigned int order)
1793 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1794 pg_data_t *pgdat = zone->zone_pgdat;
1795 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1796 unsigned long spfn, epfn, flags;
1797 unsigned long nr_pages = 0;
1800 /* Only the last zone may have deferred pages */
1801 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1804 pgdat_resize_lock(pgdat, &flags);
1807 * If deferred pages have been initialized while we were waiting for
1808 * the lock, return true, as the zone was grown. The caller will retry
1809 * this zone. We won't return to this function since the caller also
1810 * has this static branch.
1812 if (!static_branch_unlikely(&deferred_pages)) {
1813 pgdat_resize_unlock(pgdat, &flags);
1818 * If someone grew this zone while we were waiting for spinlock, return
1819 * true, as there might be enough pages already.
1821 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1822 pgdat_resize_unlock(pgdat, &flags);
1826 /* If the zone is empty somebody else may have cleared out the zone */
1827 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1828 first_deferred_pfn)) {
1829 pgdat->first_deferred_pfn = ULONG_MAX;
1830 pgdat_resize_unlock(pgdat, &flags);
1835 * Initialize and free pages in MAX_ORDER sized increments so
1836 * that we can avoid introducing any issues with the buddy
1839 while (spfn < epfn) {
1840 /* update our first deferred PFN for this section */
1841 first_deferred_pfn = spfn;
1843 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1845 /* We should only stop along section boundaries */
1846 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1849 /* If our quota has been met we can stop here */
1850 if (nr_pages >= nr_pages_needed)
1854 pgdat->first_deferred_pfn = spfn;
1855 pgdat_resize_unlock(pgdat, &flags);
1857 return nr_pages > 0;
1861 * deferred_grow_zone() is __init, but it is called from
1862 * get_page_from_freelist() during early boot until deferred_pages permanently
1863 * disables this call. This is why we have refdata wrapper to avoid warning,
1864 * and to ensure that the function body gets unloaded.
1867 _deferred_grow_zone(struct zone *zone, unsigned int order)
1869 return deferred_grow_zone(zone, order);
1872 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1874 void __init page_alloc_init_late(void)
1878 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1881 /* There will be num_node_state(N_MEMORY) threads */
1882 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1883 for_each_node_state(nid, N_MEMORY) {
1884 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1887 /* Block until all are initialised */
1888 wait_for_completion(&pgdat_init_all_done_comp);
1891 * We initialized the rest of the deferred pages. Permanently disable
1892 * on-demand struct page initialization.
1894 static_branch_disable(&deferred_pages);
1896 /* Reinit limits that are based on free pages after the kernel is up */
1897 files_maxfiles_init();
1900 /* Discard memblock private memory */
1903 for_each_populated_zone(zone)
1904 set_zone_contiguous(zone);
1908 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1909 void __init init_cma_reserved_pageblock(struct page *page)
1911 unsigned i = pageblock_nr_pages;
1912 struct page *p = page;
1915 __ClearPageReserved(p);
1916 set_page_count(p, 0);
1919 set_pageblock_migratetype(page, MIGRATE_CMA);
1921 if (pageblock_order >= MAX_ORDER) {
1922 i = pageblock_nr_pages;
1925 set_page_refcounted(p);
1926 __free_pages(p, MAX_ORDER - 1);
1927 p += MAX_ORDER_NR_PAGES;
1928 } while (i -= MAX_ORDER_NR_PAGES);
1930 set_page_refcounted(page);
1931 __free_pages(page, pageblock_order);
1934 adjust_managed_page_count(page, pageblock_nr_pages);
1939 * The order of subdivision here is critical for the IO subsystem.
1940 * Please do not alter this order without good reasons and regression
1941 * testing. Specifically, as large blocks of memory are subdivided,
1942 * the order in which smaller blocks are delivered depends on the order
1943 * they're subdivided in this function. This is the primary factor
1944 * influencing the order in which pages are delivered to the IO
1945 * subsystem according to empirical testing, and this is also justified
1946 * by considering the behavior of a buddy system containing a single
1947 * large block of memory acted on by a series of small allocations.
1948 * This behavior is a critical factor in sglist merging's success.
1952 static inline void expand(struct zone *zone, struct page *page,
1953 int low, int high, struct free_area *area,
1956 unsigned long size = 1 << high;
1958 while (high > low) {
1962 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1965 * Mark as guard pages (or page), that will allow to
1966 * merge back to allocator when buddy will be freed.
1967 * Corresponding page table entries will not be touched,
1968 * pages will stay not present in virtual address space
1970 if (set_page_guard(zone, &page[size], high, migratetype))
1973 list_add(&page[size].lru, &area->free_list[migratetype]);
1975 set_page_order(&page[size], high);
1979 static void check_new_page_bad(struct page *page)
1981 const char *bad_reason = NULL;
1982 unsigned long bad_flags = 0;
1984 if (unlikely(atomic_read(&page->_mapcount) != -1))
1985 bad_reason = "nonzero mapcount";
1986 if (unlikely(page->mapping != NULL))
1987 bad_reason = "non-NULL mapping";
1988 if (unlikely(page_ref_count(page) != 0))
1989 bad_reason = "nonzero _count";
1990 if (unlikely(page->flags & __PG_HWPOISON)) {
1991 bad_reason = "HWPoisoned (hardware-corrupted)";
1992 bad_flags = __PG_HWPOISON;
1993 /* Don't complain about hwpoisoned pages */
1994 page_mapcount_reset(page); /* remove PageBuddy */
1997 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1998 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1999 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2002 if (unlikely(page->mem_cgroup))
2003 bad_reason = "page still charged to cgroup";
2005 bad_page(page, bad_reason, bad_flags);
2009 * This page is about to be returned from the page allocator
2011 static inline int check_new_page(struct page *page)
2013 if (likely(page_expected_state(page,
2014 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2017 check_new_page_bad(page);
2021 static inline bool free_pages_prezeroed(void)
2023 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2024 page_poisoning_enabled();
2027 #ifdef CONFIG_DEBUG_VM
2028 static bool check_pcp_refill(struct page *page)
2033 static bool check_new_pcp(struct page *page)
2035 return check_new_page(page);
2038 static bool check_pcp_refill(struct page *page)
2040 return check_new_page(page);
2042 static bool check_new_pcp(struct page *page)
2046 #endif /* CONFIG_DEBUG_VM */
2048 static bool check_new_pages(struct page *page, unsigned int order)
2051 for (i = 0; i < (1 << order); i++) {
2052 struct page *p = page + i;
2054 if (unlikely(check_new_page(p)))
2061 inline void post_alloc_hook(struct page *page, unsigned int order,
2064 set_page_private(page, 0);
2065 set_page_refcounted(page);
2067 arch_alloc_page(page, order);
2068 if (debug_pagealloc_enabled())
2069 kernel_map_pages(page, 1 << order, 1);
2070 kasan_alloc_pages(page, order);
2071 kernel_poison_pages(page, 1 << order, 1);
2072 set_page_owner(page, order, gfp_flags);
2075 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2076 unsigned int alloc_flags)
2080 post_alloc_hook(page, order, gfp_flags);
2082 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
2083 for (i = 0; i < (1 << order); i++)
2084 clear_highpage(page + i);
2086 if (order && (gfp_flags & __GFP_COMP))
2087 prep_compound_page(page, order);
2090 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2091 * allocate the page. The expectation is that the caller is taking
2092 * steps that will free more memory. The caller should avoid the page
2093 * being used for !PFMEMALLOC purposes.
2095 if (alloc_flags & ALLOC_NO_WATERMARKS)
2096 set_page_pfmemalloc(page);
2098 clear_page_pfmemalloc(page);
2102 * Go through the free lists for the given migratetype and remove
2103 * the smallest available page from the freelists
2105 static __always_inline
2106 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2109 unsigned int current_order;
2110 struct free_area *area;
2113 /* Find a page of the appropriate size in the preferred list */
2114 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2115 area = &(zone->free_area[current_order]);
2116 page = list_first_entry_or_null(&area->free_list[migratetype],
2120 list_del(&page->lru);
2121 rmv_page_order(page);
2123 expand(zone, page, order, current_order, area, migratetype);
2124 set_pcppage_migratetype(page, migratetype);
2133 * This array describes the order lists are fallen back to when
2134 * the free lists for the desirable migrate type are depleted
2136 static int fallbacks[MIGRATE_TYPES][4] = {
2137 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2138 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2139 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2141 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2143 #ifdef CONFIG_MEMORY_ISOLATION
2144 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2149 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2152 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2155 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2156 unsigned int order) { return NULL; }
2160 * Move the free pages in a range to the free lists of the requested type.
2161 * Note that start_page and end_pages are not aligned on a pageblock
2162 * boundary. If alignment is required, use move_freepages_block()
2164 static int move_freepages(struct zone *zone,
2165 struct page *start_page, struct page *end_page,
2166 int migratetype, int *num_movable)
2170 int pages_moved = 0;
2172 #ifndef CONFIG_HOLES_IN_ZONE
2174 * page_zone is not safe to call in this context when
2175 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2176 * anyway as we check zone boundaries in move_freepages_block().
2177 * Remove at a later date when no bug reports exist related to
2178 * grouping pages by mobility
2180 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2181 pfn_valid(page_to_pfn(end_page)) &&
2182 page_zone(start_page) != page_zone(end_page));
2184 for (page = start_page; page <= end_page;) {
2185 if (!pfn_valid_within(page_to_pfn(page))) {
2190 /* Make sure we are not inadvertently changing nodes */
2191 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2193 if (!PageBuddy(page)) {
2195 * We assume that pages that could be isolated for
2196 * migration are movable. But we don't actually try
2197 * isolating, as that would be expensive.
2200 (PageLRU(page) || __PageMovable(page)))
2207 order = page_order(page);
2208 list_move(&page->lru,
2209 &zone->free_area[order].free_list[migratetype]);
2211 pages_moved += 1 << order;
2217 int move_freepages_block(struct zone *zone, struct page *page,
2218 int migratetype, int *num_movable)
2220 unsigned long start_pfn, end_pfn;
2221 struct page *start_page, *end_page;
2226 start_pfn = page_to_pfn(page);
2227 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2228 start_page = pfn_to_page(start_pfn);
2229 end_page = start_page + pageblock_nr_pages - 1;
2230 end_pfn = start_pfn + pageblock_nr_pages - 1;
2232 /* Do not cross zone boundaries */
2233 if (!zone_spans_pfn(zone, start_pfn))
2235 if (!zone_spans_pfn(zone, end_pfn))
2238 return move_freepages(zone, start_page, end_page, migratetype,
2242 static void change_pageblock_range(struct page *pageblock_page,
2243 int start_order, int migratetype)
2245 int nr_pageblocks = 1 << (start_order - pageblock_order);
2247 while (nr_pageblocks--) {
2248 set_pageblock_migratetype(pageblock_page, migratetype);
2249 pageblock_page += pageblock_nr_pages;
2254 * When we are falling back to another migratetype during allocation, try to
2255 * steal extra free pages from the same pageblocks to satisfy further
2256 * allocations, instead of polluting multiple pageblocks.
2258 * If we are stealing a relatively large buddy page, it is likely there will
2259 * be more free pages in the pageblock, so try to steal them all. For
2260 * reclaimable and unmovable allocations, we steal regardless of page size,
2261 * as fragmentation caused by those allocations polluting movable pageblocks
2262 * is worse than movable allocations stealing from unmovable and reclaimable
2265 static bool can_steal_fallback(unsigned int order, int start_mt)
2268 * Leaving this order check is intended, although there is
2269 * relaxed order check in next check. The reason is that
2270 * we can actually steal whole pageblock if this condition met,
2271 * but, below check doesn't guarantee it and that is just heuristic
2272 * so could be changed anytime.
2274 if (order >= pageblock_order)
2277 if (order >= pageblock_order / 2 ||
2278 start_mt == MIGRATE_RECLAIMABLE ||
2279 start_mt == MIGRATE_UNMOVABLE ||
2280 page_group_by_mobility_disabled)
2286 static inline void boost_watermark(struct zone *zone)
2288 unsigned long max_boost;
2290 if (!watermark_boost_factor)
2293 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2294 watermark_boost_factor, 10000);
2297 * high watermark may be uninitialised if fragmentation occurs
2298 * very early in boot so do not boost. We do not fall
2299 * through and boost by pageblock_nr_pages as failing
2300 * allocations that early means that reclaim is not going
2301 * to help and it may even be impossible to reclaim the
2302 * boosted watermark resulting in a hang.
2307 max_boost = max(pageblock_nr_pages, max_boost);
2309 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2314 * This function implements actual steal behaviour. If order is large enough,
2315 * we can steal whole pageblock. If not, we first move freepages in this
2316 * pageblock to our migratetype and determine how many already-allocated pages
2317 * are there in the pageblock with a compatible migratetype. If at least half
2318 * of pages are free or compatible, we can change migratetype of the pageblock
2319 * itself, so pages freed in the future will be put on the correct free list.
2321 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2322 unsigned int alloc_flags, int start_type, bool whole_block)
2324 unsigned int current_order = page_order(page);
2325 struct free_area *area;
2326 int free_pages, movable_pages, alike_pages;
2329 old_block_type = get_pageblock_migratetype(page);
2332 * This can happen due to races and we want to prevent broken
2333 * highatomic accounting.
2335 if (is_migrate_highatomic(old_block_type))
2338 /* Take ownership for orders >= pageblock_order */
2339 if (current_order >= pageblock_order) {
2340 change_pageblock_range(page, current_order, start_type);
2345 * Boost watermarks to increase reclaim pressure to reduce the
2346 * likelihood of future fallbacks. Wake kswapd now as the node
2347 * may be balanced overall and kswapd will not wake naturally.
2349 boost_watermark(zone);
2350 if (alloc_flags & ALLOC_KSWAPD)
2351 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2353 /* We are not allowed to try stealing from the whole block */
2357 free_pages = move_freepages_block(zone, page, start_type,
2360 * Determine how many pages are compatible with our allocation.
2361 * For movable allocation, it's the number of movable pages which
2362 * we just obtained. For other types it's a bit more tricky.
2364 if (start_type == MIGRATE_MOVABLE) {
2365 alike_pages = movable_pages;
2368 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2369 * to MOVABLE pageblock, consider all non-movable pages as
2370 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2371 * vice versa, be conservative since we can't distinguish the
2372 * exact migratetype of non-movable pages.
2374 if (old_block_type == MIGRATE_MOVABLE)
2375 alike_pages = pageblock_nr_pages
2376 - (free_pages + movable_pages);
2381 /* moving whole block can fail due to zone boundary conditions */
2386 * If a sufficient number of pages in the block are either free or of
2387 * comparable migratability as our allocation, claim the whole block.
2389 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2390 page_group_by_mobility_disabled)
2391 set_pageblock_migratetype(page, start_type);
2396 area = &zone->free_area[current_order];
2397 list_move(&page->lru, &area->free_list[start_type]);
2401 * Check whether there is a suitable fallback freepage with requested order.
2402 * If only_stealable is true, this function returns fallback_mt only if
2403 * we can steal other freepages all together. This would help to reduce
2404 * fragmentation due to mixed migratetype pages in one pageblock.
2406 int find_suitable_fallback(struct free_area *area, unsigned int order,
2407 int migratetype, bool only_stealable, bool *can_steal)
2412 if (area->nr_free == 0)
2417 fallback_mt = fallbacks[migratetype][i];
2418 if (fallback_mt == MIGRATE_TYPES)
2421 if (list_empty(&area->free_list[fallback_mt]))
2424 if (can_steal_fallback(order, migratetype))
2427 if (!only_stealable)
2438 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2439 * there are no empty page blocks that contain a page with a suitable order
2441 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2442 unsigned int alloc_order)
2445 unsigned long max_managed, flags;
2448 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2449 * Check is race-prone but harmless.
2451 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2452 if (zone->nr_reserved_highatomic >= max_managed)
2455 spin_lock_irqsave(&zone->lock, flags);
2457 /* Recheck the nr_reserved_highatomic limit under the lock */
2458 if (zone->nr_reserved_highatomic >= max_managed)
2462 mt = get_pageblock_migratetype(page);
2463 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2464 && !is_migrate_cma(mt)) {
2465 zone->nr_reserved_highatomic += pageblock_nr_pages;
2466 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2467 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2471 spin_unlock_irqrestore(&zone->lock, flags);
2475 * Used when an allocation is about to fail under memory pressure. This
2476 * potentially hurts the reliability of high-order allocations when under
2477 * intense memory pressure but failed atomic allocations should be easier
2478 * to recover from than an OOM.
2480 * If @force is true, try to unreserve a pageblock even though highatomic
2481 * pageblock is exhausted.
2483 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2486 struct zonelist *zonelist = ac->zonelist;
2487 unsigned long flags;
2494 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2497 * Preserve at least one pageblock unless memory pressure
2500 if (!force && zone->nr_reserved_highatomic <=
2504 spin_lock_irqsave(&zone->lock, flags);
2505 for (order = 0; order < MAX_ORDER; order++) {
2506 struct free_area *area = &(zone->free_area[order]);
2508 page = list_first_entry_or_null(
2509 &area->free_list[MIGRATE_HIGHATOMIC],
2515 * In page freeing path, migratetype change is racy so
2516 * we can counter several free pages in a pageblock
2517 * in this loop althoug we changed the pageblock type
2518 * from highatomic to ac->migratetype. So we should
2519 * adjust the count once.
2521 if (is_migrate_highatomic_page(page)) {
2523 * It should never happen but changes to
2524 * locking could inadvertently allow a per-cpu
2525 * drain to add pages to MIGRATE_HIGHATOMIC
2526 * while unreserving so be safe and watch for
2529 zone->nr_reserved_highatomic -= min(
2531 zone->nr_reserved_highatomic);
2535 * Convert to ac->migratetype and avoid the normal
2536 * pageblock stealing heuristics. Minimally, the caller
2537 * is doing the work and needs the pages. More
2538 * importantly, if the block was always converted to
2539 * MIGRATE_UNMOVABLE or another type then the number
2540 * of pageblocks that cannot be completely freed
2543 set_pageblock_migratetype(page, ac->migratetype);
2544 ret = move_freepages_block(zone, page, ac->migratetype,
2547 spin_unlock_irqrestore(&zone->lock, flags);
2551 spin_unlock_irqrestore(&zone->lock, flags);
2558 * Try finding a free buddy page on the fallback list and put it on the free
2559 * list of requested migratetype, possibly along with other pages from the same
2560 * block, depending on fragmentation avoidance heuristics. Returns true if
2561 * fallback was found so that __rmqueue_smallest() can grab it.
2563 * The use of signed ints for order and current_order is a deliberate
2564 * deviation from the rest of this file, to make the for loop
2565 * condition simpler.
2567 static __always_inline bool
2568 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2569 unsigned int alloc_flags)
2571 struct free_area *area;
2573 int min_order = order;
2579 * Do not steal pages from freelists belonging to other pageblocks
2580 * i.e. orders < pageblock_order. If there are no local zones free,
2581 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2583 if (alloc_flags & ALLOC_NOFRAGMENT)
2584 min_order = pageblock_order;
2587 * Find the largest available free page in the other list. This roughly
2588 * approximates finding the pageblock with the most free pages, which
2589 * would be too costly to do exactly.
2591 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2593 area = &(zone->free_area[current_order]);
2594 fallback_mt = find_suitable_fallback(area, current_order,
2595 start_migratetype, false, &can_steal);
2596 if (fallback_mt == -1)
2600 * We cannot steal all free pages from the pageblock and the
2601 * requested migratetype is movable. In that case it's better to
2602 * steal and split the smallest available page instead of the
2603 * largest available page, because even if the next movable
2604 * allocation falls back into a different pageblock than this
2605 * one, it won't cause permanent fragmentation.
2607 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2608 && current_order > order)
2617 for (current_order = order; current_order < MAX_ORDER;
2619 area = &(zone->free_area[current_order]);
2620 fallback_mt = find_suitable_fallback(area, current_order,
2621 start_migratetype, false, &can_steal);
2622 if (fallback_mt != -1)
2627 * This should not happen - we already found a suitable fallback
2628 * when looking for the largest page.
2630 VM_BUG_ON(current_order == MAX_ORDER);
2633 page = list_first_entry(&area->free_list[fallback_mt],
2636 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2639 trace_mm_page_alloc_extfrag(page, order, current_order,
2640 start_migratetype, fallback_mt);
2647 * Do the hard work of removing an element from the buddy allocator.
2648 * Call me with the zone->lock already held.
2650 static __always_inline struct page *
2651 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2652 unsigned int alloc_flags)
2657 page = __rmqueue_smallest(zone, order, migratetype);
2658 if (unlikely(!page)) {
2659 if (migratetype == MIGRATE_MOVABLE)
2660 page = __rmqueue_cma_fallback(zone, order);
2662 if (!page && __rmqueue_fallback(zone, order, migratetype,
2667 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2672 * Obtain a specified number of elements from the buddy allocator, all under
2673 * a single hold of the lock, for efficiency. Add them to the supplied list.
2674 * Returns the number of new pages which were placed at *list.
2676 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2677 unsigned long count, struct list_head *list,
2678 int migratetype, unsigned int alloc_flags)
2682 spin_lock(&zone->lock);
2683 for (i = 0; i < count; ++i) {
2684 struct page *page = __rmqueue(zone, order, migratetype,
2686 if (unlikely(page == NULL))
2689 if (unlikely(check_pcp_refill(page)))
2693 * Split buddy pages returned by expand() are received here in
2694 * physical page order. The page is added to the tail of
2695 * caller's list. From the callers perspective, the linked list
2696 * is ordered by page number under some conditions. This is
2697 * useful for IO devices that can forward direction from the
2698 * head, thus also in the physical page order. This is useful
2699 * for IO devices that can merge IO requests if the physical
2700 * pages are ordered properly.
2702 list_add_tail(&page->lru, list);
2704 if (is_migrate_cma(get_pcppage_migratetype(page)))
2705 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2710 * i pages were removed from the buddy list even if some leak due
2711 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2712 * on i. Do not confuse with 'alloced' which is the number of
2713 * pages added to the pcp list.
2715 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2716 spin_unlock(&zone->lock);
2722 * Called from the vmstat counter updater to drain pagesets of this
2723 * currently executing processor on remote nodes after they have
2726 * Note that this function must be called with the thread pinned to
2727 * a single processor.
2729 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2731 unsigned long flags;
2732 int to_drain, batch;
2734 local_irq_save(flags);
2735 batch = READ_ONCE(pcp->batch);
2736 to_drain = min(pcp->count, batch);
2738 free_pcppages_bulk(zone, to_drain, pcp);
2739 local_irq_restore(flags);
2744 * Drain pcplists of the indicated processor and zone.
2746 * The processor must either be the current processor and the
2747 * thread pinned to the current processor or a processor that
2750 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2752 unsigned long flags;
2753 struct per_cpu_pageset *pset;
2754 struct per_cpu_pages *pcp;
2756 local_irq_save(flags);
2757 pset = per_cpu_ptr(zone->pageset, cpu);
2761 free_pcppages_bulk(zone, pcp->count, pcp);
2762 local_irq_restore(flags);
2766 * Drain pcplists of all zones on the indicated processor.
2768 * The processor must either be the current processor and the
2769 * thread pinned to the current processor or a processor that
2772 static void drain_pages(unsigned int cpu)
2776 for_each_populated_zone(zone) {
2777 drain_pages_zone(cpu, zone);
2782 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2784 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2785 * the single zone's pages.
2787 void drain_local_pages(struct zone *zone)
2789 int cpu = smp_processor_id();
2792 drain_pages_zone(cpu, zone);
2797 static void drain_local_pages_wq(struct work_struct *work)
2799 struct pcpu_drain *drain;
2801 drain = container_of(work, struct pcpu_drain, work);
2804 * drain_all_pages doesn't use proper cpu hotplug protection so
2805 * we can race with cpu offline when the WQ can move this from
2806 * a cpu pinned worker to an unbound one. We can operate on a different
2807 * cpu which is allright but we also have to make sure to not move to
2811 drain_local_pages(drain->zone);
2816 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2818 * When zone parameter is non-NULL, spill just the single zone's pages.
2820 * Note that this can be extremely slow as the draining happens in a workqueue.
2822 void drain_all_pages(struct zone *zone)
2827 * Allocate in the BSS so we wont require allocation in
2828 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2830 static cpumask_t cpus_with_pcps;
2833 * Make sure nobody triggers this path before mm_percpu_wq is fully
2836 if (WARN_ON_ONCE(!mm_percpu_wq))
2840 * Do not drain if one is already in progress unless it's specific to
2841 * a zone. Such callers are primarily CMA and memory hotplug and need
2842 * the drain to be complete when the call returns.
2844 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2847 mutex_lock(&pcpu_drain_mutex);
2851 * We don't care about racing with CPU hotplug event
2852 * as offline notification will cause the notified
2853 * cpu to drain that CPU pcps and on_each_cpu_mask
2854 * disables preemption as part of its processing
2856 for_each_online_cpu(cpu) {
2857 struct per_cpu_pageset *pcp;
2859 bool has_pcps = false;
2862 pcp = per_cpu_ptr(zone->pageset, cpu);
2866 for_each_populated_zone(z) {
2867 pcp = per_cpu_ptr(z->pageset, cpu);
2868 if (pcp->pcp.count) {
2876 cpumask_set_cpu(cpu, &cpus_with_pcps);
2878 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2881 for_each_cpu(cpu, &cpus_with_pcps) {
2882 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2885 INIT_WORK(&drain->work, drain_local_pages_wq);
2886 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2888 for_each_cpu(cpu, &cpus_with_pcps)
2889 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2891 mutex_unlock(&pcpu_drain_mutex);
2894 #ifdef CONFIG_HIBERNATION
2897 * Touch the watchdog for every WD_PAGE_COUNT pages.
2899 #define WD_PAGE_COUNT (128*1024)
2901 void mark_free_pages(struct zone *zone)
2903 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2904 unsigned long flags;
2905 unsigned int order, t;
2908 if (zone_is_empty(zone))
2911 spin_lock_irqsave(&zone->lock, flags);
2913 max_zone_pfn = zone_end_pfn(zone);
2914 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2915 if (pfn_valid(pfn)) {
2916 page = pfn_to_page(pfn);
2918 if (!--page_count) {
2919 touch_nmi_watchdog();
2920 page_count = WD_PAGE_COUNT;
2923 if (page_zone(page) != zone)
2926 if (!swsusp_page_is_forbidden(page))
2927 swsusp_unset_page_free(page);
2930 for_each_migratetype_order(order, t) {
2931 list_for_each_entry(page,
2932 &zone->free_area[order].free_list[t], lru) {
2935 pfn = page_to_pfn(page);
2936 for (i = 0; i < (1UL << order); i++) {
2937 if (!--page_count) {
2938 touch_nmi_watchdog();
2939 page_count = WD_PAGE_COUNT;
2941 swsusp_set_page_free(pfn_to_page(pfn + i));
2945 spin_unlock_irqrestore(&zone->lock, flags);
2947 #endif /* CONFIG_PM */
2949 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2953 if (!free_pcp_prepare(page))
2956 migratetype = get_pfnblock_migratetype(page, pfn);
2957 set_pcppage_migratetype(page, migratetype);
2961 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2963 struct zone *zone = page_zone(page);
2964 struct per_cpu_pages *pcp;
2967 migratetype = get_pcppage_migratetype(page);
2968 __count_vm_event(PGFREE);
2971 * We only track unmovable, reclaimable and movable on pcp lists.
2972 * Free ISOLATE pages back to the allocator because they are being
2973 * offlined but treat HIGHATOMIC as movable pages so we can get those
2974 * areas back if necessary. Otherwise, we may have to free
2975 * excessively into the page allocator
2977 if (migratetype >= MIGRATE_PCPTYPES) {
2978 if (unlikely(is_migrate_isolate(migratetype))) {
2979 free_one_page(zone, page, pfn, 0, migratetype);
2982 migratetype = MIGRATE_MOVABLE;
2985 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2986 list_add(&page->lru, &pcp->lists[migratetype]);
2988 if (pcp->count >= pcp->high) {
2989 unsigned long batch = READ_ONCE(pcp->batch);
2990 free_pcppages_bulk(zone, batch, pcp);
2995 * Free a 0-order page
2997 void free_unref_page(struct page *page)
2999 unsigned long flags;
3000 unsigned long pfn = page_to_pfn(page);
3002 if (!free_unref_page_prepare(page, pfn))
3005 local_irq_save(flags);
3006 free_unref_page_commit(page, pfn);
3007 local_irq_restore(flags);
3011 * Free a list of 0-order pages
3013 void free_unref_page_list(struct list_head *list)
3015 struct page *page, *next;
3016 unsigned long flags, pfn;
3017 int batch_count = 0;
3019 /* Prepare pages for freeing */
3020 list_for_each_entry_safe(page, next, list, lru) {
3021 pfn = page_to_pfn(page);
3022 if (!free_unref_page_prepare(page, pfn))
3023 list_del(&page->lru);
3024 set_page_private(page, pfn);
3027 local_irq_save(flags);
3028 list_for_each_entry_safe(page, next, list, lru) {
3029 unsigned long pfn = page_private(page);
3031 set_page_private(page, 0);
3032 trace_mm_page_free_batched(page);
3033 free_unref_page_commit(page, pfn);
3036 * Guard against excessive IRQ disabled times when we get
3037 * a large list of pages to free.
3039 if (++batch_count == SWAP_CLUSTER_MAX) {
3040 local_irq_restore(flags);
3042 local_irq_save(flags);
3045 local_irq_restore(flags);
3049 * split_page takes a non-compound higher-order page, and splits it into
3050 * n (1<<order) sub-pages: page[0..n]
3051 * Each sub-page must be freed individually.
3053 * Note: this is probably too low level an operation for use in drivers.
3054 * Please consult with lkml before using this in your driver.
3056 void split_page(struct page *page, unsigned int order)
3060 VM_BUG_ON_PAGE(PageCompound(page), page);
3061 VM_BUG_ON_PAGE(!page_count(page), page);
3063 for (i = 1; i < (1 << order); i++)
3064 set_page_refcounted(page + i);
3065 split_page_owner(page, order);
3067 EXPORT_SYMBOL_GPL(split_page);
3069 int __isolate_free_page(struct page *page, unsigned int order)
3071 unsigned long watermark;
3075 BUG_ON(!PageBuddy(page));
3077 zone = page_zone(page);
3078 mt = get_pageblock_migratetype(page);
3080 if (!is_migrate_isolate(mt)) {
3082 * Obey watermarks as if the page was being allocated. We can
3083 * emulate a high-order watermark check with a raised order-0
3084 * watermark, because we already know our high-order page
3087 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3088 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3091 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3094 /* Remove page from free list */
3095 list_del(&page->lru);
3096 zone->free_area[order].nr_free--;
3097 rmv_page_order(page);
3100 * Set the pageblock if the isolated page is at least half of a
3103 if (order >= pageblock_order - 1) {
3104 struct page *endpage = page + (1 << order) - 1;
3105 for (; page < endpage; page += pageblock_nr_pages) {
3106 int mt = get_pageblock_migratetype(page);
3107 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3108 && !is_migrate_highatomic(mt))
3109 set_pageblock_migratetype(page,
3115 return 1UL << order;
3119 * Update NUMA hit/miss statistics
3121 * Must be called with interrupts disabled.
3123 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3126 enum numa_stat_item local_stat = NUMA_LOCAL;
3128 /* skip numa counters update if numa stats is disabled */
3129 if (!static_branch_likely(&vm_numa_stat_key))
3132 if (zone_to_nid(z) != numa_node_id())
3133 local_stat = NUMA_OTHER;
3135 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3136 __inc_numa_state(z, NUMA_HIT);
3138 __inc_numa_state(z, NUMA_MISS);
3139 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3141 __inc_numa_state(z, local_stat);
3145 /* Remove page from the per-cpu list, caller must protect the list */
3146 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3147 unsigned int alloc_flags,
3148 struct per_cpu_pages *pcp,
3149 struct list_head *list)
3154 if (list_empty(list)) {
3155 pcp->count += rmqueue_bulk(zone, 0,
3157 migratetype, alloc_flags);
3158 if (unlikely(list_empty(list)))
3162 page = list_first_entry(list, struct page, lru);
3163 list_del(&page->lru);
3165 } while (check_new_pcp(page));
3170 /* Lock and remove page from the per-cpu list */
3171 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3172 struct zone *zone, gfp_t gfp_flags,
3173 int migratetype, unsigned int alloc_flags)
3175 struct per_cpu_pages *pcp;
3176 struct list_head *list;
3178 unsigned long flags;
3180 local_irq_save(flags);
3181 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3182 list = &pcp->lists[migratetype];
3183 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3185 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3186 zone_statistics(preferred_zone, zone);
3188 local_irq_restore(flags);
3193 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3196 struct page *rmqueue(struct zone *preferred_zone,
3197 struct zone *zone, unsigned int order,
3198 gfp_t gfp_flags, unsigned int alloc_flags,
3201 unsigned long flags;
3204 if (likely(order == 0)) {
3205 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3206 migratetype, alloc_flags);
3211 * We most definitely don't want callers attempting to
3212 * allocate greater than order-1 page units with __GFP_NOFAIL.
3214 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3215 spin_lock_irqsave(&zone->lock, flags);
3219 if (alloc_flags & ALLOC_HARDER) {
3220 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3222 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3225 page = __rmqueue(zone, order, migratetype, alloc_flags);
3226 } while (page && check_new_pages(page, order));
3227 spin_unlock(&zone->lock);
3230 __mod_zone_freepage_state(zone, -(1 << order),
3231 get_pcppage_migratetype(page));
3233 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3234 zone_statistics(preferred_zone, zone);
3235 local_irq_restore(flags);
3238 /* Separate test+clear to avoid unnecessary atomics */
3239 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3240 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3241 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3244 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3248 local_irq_restore(flags);
3252 #ifdef CONFIG_FAIL_PAGE_ALLOC
3255 struct fault_attr attr;
3257 bool ignore_gfp_highmem;
3258 bool ignore_gfp_reclaim;
3260 } fail_page_alloc = {
3261 .attr = FAULT_ATTR_INITIALIZER,
3262 .ignore_gfp_reclaim = true,
3263 .ignore_gfp_highmem = true,
3267 static int __init setup_fail_page_alloc(char *str)
3269 return setup_fault_attr(&fail_page_alloc.attr, str);
3271 __setup("fail_page_alloc=", setup_fail_page_alloc);
3273 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3275 if (order < fail_page_alloc.min_order)
3277 if (gfp_mask & __GFP_NOFAIL)
3279 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3281 if (fail_page_alloc.ignore_gfp_reclaim &&
3282 (gfp_mask & __GFP_DIRECT_RECLAIM))
3285 return should_fail(&fail_page_alloc.attr, 1 << order);
3288 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3290 static int __init fail_page_alloc_debugfs(void)
3292 umode_t mode = S_IFREG | 0600;
3295 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3296 &fail_page_alloc.attr);
3298 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3299 &fail_page_alloc.ignore_gfp_reclaim);
3300 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3301 &fail_page_alloc.ignore_gfp_highmem);
3302 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3307 late_initcall(fail_page_alloc_debugfs);
3309 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3311 #else /* CONFIG_FAIL_PAGE_ALLOC */
3313 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3318 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3320 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3322 return __should_fail_alloc_page(gfp_mask, order);
3324 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3327 * Return true if free base pages are above 'mark'. For high-order checks it
3328 * will return true of the order-0 watermark is reached and there is at least
3329 * one free page of a suitable size. Checking now avoids taking the zone lock
3330 * to check in the allocation paths if no pages are free.
3332 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3333 int classzone_idx, unsigned int alloc_flags,
3338 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3340 /* free_pages may go negative - that's OK */
3341 free_pages -= (1 << order) - 1;
3343 if (alloc_flags & ALLOC_HIGH)
3347 * If the caller does not have rights to ALLOC_HARDER then subtract
3348 * the high-atomic reserves. This will over-estimate the size of the
3349 * atomic reserve but it avoids a search.
3351 if (likely(!alloc_harder)) {
3352 free_pages -= z->nr_reserved_highatomic;
3355 * OOM victims can try even harder than normal ALLOC_HARDER
3356 * users on the grounds that it's definitely going to be in
3357 * the exit path shortly and free memory. Any allocation it
3358 * makes during the free path will be small and short-lived.
3360 if (alloc_flags & ALLOC_OOM)
3368 /* If allocation can't use CMA areas don't use free CMA pages */
3369 if (!(alloc_flags & ALLOC_CMA))
3370 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3374 * Check watermarks for an order-0 allocation request. If these
3375 * are not met, then a high-order request also cannot go ahead
3376 * even if a suitable page happened to be free.
3378 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3381 /* If this is an order-0 request then the watermark is fine */
3385 /* For a high-order request, check at least one suitable page is free */
3386 for (o = order; o < MAX_ORDER; o++) {
3387 struct free_area *area = &z->free_area[o];
3393 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3394 if (!list_empty(&area->free_list[mt]))
3399 if ((alloc_flags & ALLOC_CMA) &&
3400 !list_empty(&area->free_list[MIGRATE_CMA])) {
3405 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3411 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3412 int classzone_idx, unsigned int alloc_flags)
3414 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3415 zone_page_state(z, NR_FREE_PAGES));
3418 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3419 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3421 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3425 /* If allocation can't use CMA areas don't use free CMA pages */
3426 if (!(alloc_flags & ALLOC_CMA))
3427 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3431 * Fast check for order-0 only. If this fails then the reserves
3432 * need to be calculated. There is a corner case where the check
3433 * passes but only the high-order atomic reserve are free. If
3434 * the caller is !atomic then it'll uselessly search the free
3435 * list. That corner case is then slower but it is harmless.
3437 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3440 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3444 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3445 unsigned long mark, int classzone_idx)
3447 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3449 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3450 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3452 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3457 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3459 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3462 #else /* CONFIG_NUMA */
3463 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3467 #endif /* CONFIG_NUMA */
3470 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3471 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3472 * premature use of a lower zone may cause lowmem pressure problems that
3473 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3474 * probably too small. It only makes sense to spread allocations to avoid
3475 * fragmentation between the Normal and DMA32 zones.
3477 static inline unsigned int
3478 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3480 unsigned int alloc_flags = 0;
3482 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3483 alloc_flags |= ALLOC_KSWAPD;
3485 #ifdef CONFIG_ZONE_DMA32
3489 if (zone_idx(zone) != ZONE_NORMAL)
3493 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3494 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3495 * on UMA that if Normal is populated then so is DMA32.
3497 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3498 if (nr_online_nodes > 1 && !populated_zone(--zone))
3501 alloc_flags |= ALLOC_NOFRAGMENT;
3502 #endif /* CONFIG_ZONE_DMA32 */
3507 * get_page_from_freelist goes through the zonelist trying to allocate
3510 static struct page *
3511 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3512 const struct alloc_context *ac)
3516 struct pglist_data *last_pgdat_dirty_limit = NULL;
3521 * Scan zonelist, looking for a zone with enough free.
3522 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3524 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3525 z = ac->preferred_zoneref;
3526 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3531 if (cpusets_enabled() &&
3532 (alloc_flags & ALLOC_CPUSET) &&
3533 !__cpuset_zone_allowed(zone, gfp_mask))
3536 * When allocating a page cache page for writing, we
3537 * want to get it from a node that is within its dirty
3538 * limit, such that no single node holds more than its
3539 * proportional share of globally allowed dirty pages.
3540 * The dirty limits take into account the node's
3541 * lowmem reserves and high watermark so that kswapd
3542 * should be able to balance it without having to
3543 * write pages from its LRU list.
3545 * XXX: For now, allow allocations to potentially
3546 * exceed the per-node dirty limit in the slowpath
3547 * (spread_dirty_pages unset) before going into reclaim,
3548 * which is important when on a NUMA setup the allowed
3549 * nodes are together not big enough to reach the
3550 * global limit. The proper fix for these situations
3551 * will require awareness of nodes in the
3552 * dirty-throttling and the flusher threads.
3554 if (ac->spread_dirty_pages) {
3555 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3558 if (!node_dirty_ok(zone->zone_pgdat)) {
3559 last_pgdat_dirty_limit = zone->zone_pgdat;
3564 if (no_fallback && nr_online_nodes > 1 &&
3565 zone != ac->preferred_zoneref->zone) {
3569 * If moving to a remote node, retry but allow
3570 * fragmenting fallbacks. Locality is more important
3571 * than fragmentation avoidance.
3573 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3574 if (zone_to_nid(zone) != local_nid) {
3575 alloc_flags &= ~ALLOC_NOFRAGMENT;
3580 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3581 if (!zone_watermark_fast(zone, order, mark,
3582 ac_classzone_idx(ac), alloc_flags)) {
3585 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3587 * Watermark failed for this zone, but see if we can
3588 * grow this zone if it contains deferred pages.
3590 if (static_branch_unlikely(&deferred_pages)) {
3591 if (_deferred_grow_zone(zone, order))
3595 /* Checked here to keep the fast path fast */
3596 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3597 if (alloc_flags & ALLOC_NO_WATERMARKS)
3600 if (node_reclaim_mode == 0 ||
3601 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3604 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3606 case NODE_RECLAIM_NOSCAN:
3609 case NODE_RECLAIM_FULL:
3610 /* scanned but unreclaimable */
3613 /* did we reclaim enough */
3614 if (zone_watermark_ok(zone, order, mark,
3615 ac_classzone_idx(ac), alloc_flags))
3623 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3624 gfp_mask, alloc_flags, ac->migratetype);
3626 prep_new_page(page, order, gfp_mask, alloc_flags);
3629 * If this is a high-order atomic allocation then check
3630 * if the pageblock should be reserved for the future
3632 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3633 reserve_highatomic_pageblock(page, zone, order);
3637 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3638 /* Try again if zone has deferred pages */
3639 if (static_branch_unlikely(&deferred_pages)) {
3640 if (_deferred_grow_zone(zone, order))
3648 * It's possible on a UMA machine to get through all zones that are
3649 * fragmented. If avoiding fragmentation, reset and try again.
3652 alloc_flags &= ~ALLOC_NOFRAGMENT;
3659 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3661 unsigned int filter = SHOW_MEM_FILTER_NODES;
3662 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3664 if (!__ratelimit(&show_mem_rs))
3668 * This documents exceptions given to allocations in certain
3669 * contexts that are allowed to allocate outside current's set
3672 if (!(gfp_mask & __GFP_NOMEMALLOC))
3673 if (tsk_is_oom_victim(current) ||
3674 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3675 filter &= ~SHOW_MEM_FILTER_NODES;
3676 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3677 filter &= ~SHOW_MEM_FILTER_NODES;
3679 show_mem(filter, nodemask);
3682 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3684 struct va_format vaf;
3686 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3687 DEFAULT_RATELIMIT_BURST);
3689 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3692 va_start(args, fmt);
3695 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3696 current->comm, &vaf, gfp_mask, &gfp_mask,
3697 nodemask_pr_args(nodemask));
3700 cpuset_print_current_mems_allowed();
3703 warn_alloc_show_mem(gfp_mask, nodemask);
3706 static inline struct page *
3707 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3708 unsigned int alloc_flags,
3709 const struct alloc_context *ac)
3713 page = get_page_from_freelist(gfp_mask, order,
3714 alloc_flags|ALLOC_CPUSET, ac);
3716 * fallback to ignore cpuset restriction if our nodes
3720 page = get_page_from_freelist(gfp_mask, order,
3726 static inline struct page *
3727 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3728 const struct alloc_context *ac, unsigned long *did_some_progress)
3730 struct oom_control oc = {
3731 .zonelist = ac->zonelist,
3732 .nodemask = ac->nodemask,
3734 .gfp_mask = gfp_mask,
3739 *did_some_progress = 0;
3742 * Acquire the oom lock. If that fails, somebody else is
3743 * making progress for us.
3745 if (!mutex_trylock(&oom_lock)) {
3746 *did_some_progress = 1;
3747 schedule_timeout_uninterruptible(1);
3752 * Go through the zonelist yet one more time, keep very high watermark
3753 * here, this is only to catch a parallel oom killing, we must fail if
3754 * we're still under heavy pressure. But make sure that this reclaim
3755 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3756 * allocation which will never fail due to oom_lock already held.
3758 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3759 ~__GFP_DIRECT_RECLAIM, order,
3760 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3764 /* Coredumps can quickly deplete all memory reserves */
3765 if (current->flags & PF_DUMPCORE)
3767 /* The OOM killer will not help higher order allocs */
3768 if (order > PAGE_ALLOC_COSTLY_ORDER)
3771 * We have already exhausted all our reclaim opportunities without any
3772 * success so it is time to admit defeat. We will skip the OOM killer
3773 * because it is very likely that the caller has a more reasonable
3774 * fallback than shooting a random task.
3776 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3778 /* The OOM killer does not needlessly kill tasks for lowmem */
3779 if (ac->high_zoneidx < ZONE_NORMAL)
3781 if (pm_suspended_storage())
3784 * XXX: GFP_NOFS allocations should rather fail than rely on
3785 * other request to make a forward progress.
3786 * We are in an unfortunate situation where out_of_memory cannot
3787 * do much for this context but let's try it to at least get
3788 * access to memory reserved if the current task is killed (see
3789 * out_of_memory). Once filesystems are ready to handle allocation
3790 * failures more gracefully we should just bail out here.
3793 /* The OOM killer may not free memory on a specific node */
3794 if (gfp_mask & __GFP_THISNODE)
3797 /* Exhausted what can be done so it's blame time */
3798 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3799 *did_some_progress = 1;
3802 * Help non-failing allocations by giving them access to memory
3805 if (gfp_mask & __GFP_NOFAIL)
3806 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3807 ALLOC_NO_WATERMARKS, ac);
3810 mutex_unlock(&oom_lock);
3815 * Maximum number of compaction retries wit a progress before OOM
3816 * killer is consider as the only way to move forward.
3818 #define MAX_COMPACT_RETRIES 16
3820 #ifdef CONFIG_COMPACTION
3821 /* Try memory compaction for high-order allocations before reclaim */
3822 static struct page *
3823 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3824 unsigned int alloc_flags, const struct alloc_context *ac,
3825 enum compact_priority prio, enum compact_result *compact_result)
3827 struct page *page = NULL;
3828 unsigned long pflags;
3829 unsigned int noreclaim_flag;
3834 psi_memstall_enter(&pflags);
3835 noreclaim_flag = memalloc_noreclaim_save();
3837 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3840 memalloc_noreclaim_restore(noreclaim_flag);
3841 psi_memstall_leave(&pflags);
3844 * At least in one zone compaction wasn't deferred or skipped, so let's
3845 * count a compaction stall
3847 count_vm_event(COMPACTSTALL);
3849 /* Prep a captured page if available */
3851 prep_new_page(page, order, gfp_mask, alloc_flags);
3853 /* Try get a page from the freelist if available */
3855 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3858 struct zone *zone = page_zone(page);
3860 zone->compact_blockskip_flush = false;
3861 compaction_defer_reset(zone, order, true);
3862 count_vm_event(COMPACTSUCCESS);
3867 * It's bad if compaction run occurs and fails. The most likely reason
3868 * is that pages exist, but not enough to satisfy watermarks.
3870 count_vm_event(COMPACTFAIL);
3878 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3879 enum compact_result compact_result,
3880 enum compact_priority *compact_priority,
3881 int *compaction_retries)
3883 int max_retries = MAX_COMPACT_RETRIES;
3886 int retries = *compaction_retries;
3887 enum compact_priority priority = *compact_priority;
3892 if (compaction_made_progress(compact_result))
3893 (*compaction_retries)++;
3896 * compaction considers all the zone as desperately out of memory
3897 * so it doesn't really make much sense to retry except when the
3898 * failure could be caused by insufficient priority
3900 if (compaction_failed(compact_result))
3901 goto check_priority;
3904 * make sure the compaction wasn't deferred or didn't bail out early
3905 * due to locks contention before we declare that we should give up.
3906 * But do not retry if the given zonelist is not suitable for
3909 if (compaction_withdrawn(compact_result)) {
3910 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3915 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3916 * costly ones because they are de facto nofail and invoke OOM
3917 * killer to move on while costly can fail and users are ready
3918 * to cope with that. 1/4 retries is rather arbitrary but we
3919 * would need much more detailed feedback from compaction to
3920 * make a better decision.
3922 if (order > PAGE_ALLOC_COSTLY_ORDER)
3924 if (*compaction_retries <= max_retries) {
3930 * Make sure there are attempts at the highest priority if we exhausted
3931 * all retries or failed at the lower priorities.
3934 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3935 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3937 if (*compact_priority > min_priority) {
3938 (*compact_priority)--;
3939 *compaction_retries = 0;
3943 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3947 static inline struct page *
3948 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3949 unsigned int alloc_flags, const struct alloc_context *ac,
3950 enum compact_priority prio, enum compact_result *compact_result)
3952 *compact_result = COMPACT_SKIPPED;
3957 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3958 enum compact_result compact_result,
3959 enum compact_priority *compact_priority,
3960 int *compaction_retries)
3965 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3969 * There are setups with compaction disabled which would prefer to loop
3970 * inside the allocator rather than hit the oom killer prematurely.
3971 * Let's give them a good hope and keep retrying while the order-0
3972 * watermarks are OK.
3974 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3976 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3977 ac_classzone_idx(ac), alloc_flags))
3982 #endif /* CONFIG_COMPACTION */
3984 #ifdef CONFIG_LOCKDEP
3985 static struct lockdep_map __fs_reclaim_map =
3986 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3988 static bool __need_fs_reclaim(gfp_t gfp_mask)
3990 gfp_mask = current_gfp_context(gfp_mask);
3992 /* no reclaim without waiting on it */
3993 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3996 /* this guy won't enter reclaim */
3997 if (current->flags & PF_MEMALLOC)
4000 /* We're only interested __GFP_FS allocations for now */
4001 if (!(gfp_mask & __GFP_FS))
4004 if (gfp_mask & __GFP_NOLOCKDEP)
4010 void __fs_reclaim_acquire(void)
4012 lock_map_acquire(&__fs_reclaim_map);
4015 void __fs_reclaim_release(void)
4017 lock_map_release(&__fs_reclaim_map);
4020 void fs_reclaim_acquire(gfp_t gfp_mask)
4022 if (__need_fs_reclaim(gfp_mask))
4023 __fs_reclaim_acquire();
4025 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4027 void fs_reclaim_release(gfp_t gfp_mask)
4029 if (__need_fs_reclaim(gfp_mask))
4030 __fs_reclaim_release();
4032 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4035 /* Perform direct synchronous page reclaim */
4037 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4038 const struct alloc_context *ac)
4040 struct reclaim_state reclaim_state;
4042 unsigned int noreclaim_flag;
4043 unsigned long pflags;
4047 /* We now go into synchronous reclaim */
4048 cpuset_memory_pressure_bump();
4049 psi_memstall_enter(&pflags);
4050 fs_reclaim_acquire(gfp_mask);
4051 noreclaim_flag = memalloc_noreclaim_save();
4052 reclaim_state.reclaimed_slab = 0;
4053 current->reclaim_state = &reclaim_state;
4055 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4058 current->reclaim_state = NULL;
4059 memalloc_noreclaim_restore(noreclaim_flag);
4060 fs_reclaim_release(gfp_mask);
4061 psi_memstall_leave(&pflags);
4068 /* The really slow allocator path where we enter direct reclaim */
4069 static inline struct page *
4070 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4071 unsigned int alloc_flags, const struct alloc_context *ac,
4072 unsigned long *did_some_progress)
4074 struct page *page = NULL;
4075 bool drained = false;
4077 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4078 if (unlikely(!(*did_some_progress)))
4082 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4085 * If an allocation failed after direct reclaim, it could be because
4086 * pages are pinned on the per-cpu lists or in high alloc reserves.
4087 * Shrink them them and try again
4089 if (!page && !drained) {
4090 unreserve_highatomic_pageblock(ac, false);
4091 drain_all_pages(NULL);
4099 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4100 const struct alloc_context *ac)
4104 pg_data_t *last_pgdat = NULL;
4105 enum zone_type high_zoneidx = ac->high_zoneidx;
4107 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4109 if (last_pgdat != zone->zone_pgdat)
4110 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4111 last_pgdat = zone->zone_pgdat;
4115 static inline unsigned int
4116 gfp_to_alloc_flags(gfp_t gfp_mask)
4118 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4120 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4121 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4124 * The caller may dip into page reserves a bit more if the caller
4125 * cannot run direct reclaim, or if the caller has realtime scheduling
4126 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4127 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4129 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4131 if (gfp_mask & __GFP_ATOMIC) {
4133 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4134 * if it can't schedule.
4136 if (!(gfp_mask & __GFP_NOMEMALLOC))
4137 alloc_flags |= ALLOC_HARDER;
4139 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4140 * comment for __cpuset_node_allowed().
4142 alloc_flags &= ~ALLOC_CPUSET;
4143 } else if (unlikely(rt_task(current)) && !in_interrupt())
4144 alloc_flags |= ALLOC_HARDER;
4146 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4147 alloc_flags |= ALLOC_KSWAPD;
4150 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4151 alloc_flags |= ALLOC_CMA;
4156 static bool oom_reserves_allowed(struct task_struct *tsk)
4158 if (!tsk_is_oom_victim(tsk))
4162 * !MMU doesn't have oom reaper so give access to memory reserves
4163 * only to the thread with TIF_MEMDIE set
4165 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4172 * Distinguish requests which really need access to full memory
4173 * reserves from oom victims which can live with a portion of it
4175 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4177 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4179 if (gfp_mask & __GFP_MEMALLOC)
4180 return ALLOC_NO_WATERMARKS;
4181 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4182 return ALLOC_NO_WATERMARKS;
4183 if (!in_interrupt()) {
4184 if (current->flags & PF_MEMALLOC)
4185 return ALLOC_NO_WATERMARKS;
4186 else if (oom_reserves_allowed(current))
4193 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4195 return !!__gfp_pfmemalloc_flags(gfp_mask);
4199 * Checks whether it makes sense to retry the reclaim to make a forward progress
4200 * for the given allocation request.
4202 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4203 * without success, or when we couldn't even meet the watermark if we
4204 * reclaimed all remaining pages on the LRU lists.
4206 * Returns true if a retry is viable or false to enter the oom path.
4209 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4210 struct alloc_context *ac, int alloc_flags,
4211 bool did_some_progress, int *no_progress_loops)
4218 * Costly allocations might have made a progress but this doesn't mean
4219 * their order will become available due to high fragmentation so
4220 * always increment the no progress counter for them
4222 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4223 *no_progress_loops = 0;
4225 (*no_progress_loops)++;
4228 * Make sure we converge to OOM if we cannot make any progress
4229 * several times in the row.
4231 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4232 /* Before OOM, exhaust highatomic_reserve */
4233 return unreserve_highatomic_pageblock(ac, true);
4237 * Keep reclaiming pages while there is a chance this will lead
4238 * somewhere. If none of the target zones can satisfy our allocation
4239 * request even if all reclaimable pages are considered then we are
4240 * screwed and have to go OOM.
4242 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4244 unsigned long available;
4245 unsigned long reclaimable;
4246 unsigned long min_wmark = min_wmark_pages(zone);
4249 available = reclaimable = zone_reclaimable_pages(zone);
4250 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4253 * Would the allocation succeed if we reclaimed all
4254 * reclaimable pages?
4256 wmark = __zone_watermark_ok(zone, order, min_wmark,
4257 ac_classzone_idx(ac), alloc_flags, available);
4258 trace_reclaim_retry_zone(z, order, reclaimable,
4259 available, min_wmark, *no_progress_loops, wmark);
4262 * If we didn't make any progress and have a lot of
4263 * dirty + writeback pages then we should wait for
4264 * an IO to complete to slow down the reclaim and
4265 * prevent from pre mature OOM
4267 if (!did_some_progress) {
4268 unsigned long write_pending;
4270 write_pending = zone_page_state_snapshot(zone,
4271 NR_ZONE_WRITE_PENDING);
4273 if (2 * write_pending > reclaimable) {
4274 congestion_wait(BLK_RW_ASYNC, HZ/10);
4286 * Memory allocation/reclaim might be called from a WQ context and the
4287 * current implementation of the WQ concurrency control doesn't
4288 * recognize that a particular WQ is congested if the worker thread is
4289 * looping without ever sleeping. Therefore we have to do a short sleep
4290 * here rather than calling cond_resched().
4292 if (current->flags & PF_WQ_WORKER)
4293 schedule_timeout_uninterruptible(1);
4300 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4303 * It's possible that cpuset's mems_allowed and the nodemask from
4304 * mempolicy don't intersect. This should be normally dealt with by
4305 * policy_nodemask(), but it's possible to race with cpuset update in
4306 * such a way the check therein was true, and then it became false
4307 * before we got our cpuset_mems_cookie here.
4308 * This assumes that for all allocations, ac->nodemask can come only
4309 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4310 * when it does not intersect with the cpuset restrictions) or the
4311 * caller can deal with a violated nodemask.
4313 if (cpusets_enabled() && ac->nodemask &&
4314 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4315 ac->nodemask = NULL;
4320 * When updating a task's mems_allowed or mempolicy nodemask, it is
4321 * possible to race with parallel threads in such a way that our
4322 * allocation can fail while the mask is being updated. If we are about
4323 * to fail, check if the cpuset changed during allocation and if so,
4326 if (read_mems_allowed_retry(cpuset_mems_cookie))
4332 static inline struct page *
4333 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4334 struct alloc_context *ac)
4336 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4337 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4338 struct page *page = NULL;
4339 unsigned int alloc_flags;
4340 unsigned long did_some_progress;
4341 enum compact_priority compact_priority;
4342 enum compact_result compact_result;
4343 int compaction_retries;
4344 int no_progress_loops;
4345 unsigned int cpuset_mems_cookie;
4349 * We also sanity check to catch abuse of atomic reserves being used by
4350 * callers that are not in atomic context.
4352 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4353 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4354 gfp_mask &= ~__GFP_ATOMIC;
4357 compaction_retries = 0;
4358 no_progress_loops = 0;
4359 compact_priority = DEF_COMPACT_PRIORITY;
4360 cpuset_mems_cookie = read_mems_allowed_begin();
4363 * The fast path uses conservative alloc_flags to succeed only until
4364 * kswapd needs to be woken up, and to avoid the cost of setting up
4365 * alloc_flags precisely. So we do that now.
4367 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4370 * We need to recalculate the starting point for the zonelist iterator
4371 * because we might have used different nodemask in the fast path, or
4372 * there was a cpuset modification and we are retrying - otherwise we
4373 * could end up iterating over non-eligible zones endlessly.
4375 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4376 ac->high_zoneidx, ac->nodemask);
4377 if (!ac->preferred_zoneref->zone)
4380 if (alloc_flags & ALLOC_KSWAPD)
4381 wake_all_kswapds(order, gfp_mask, ac);
4384 * The adjusted alloc_flags might result in immediate success, so try
4387 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4392 * For costly allocations, try direct compaction first, as it's likely
4393 * that we have enough base pages and don't need to reclaim. For non-
4394 * movable high-order allocations, do that as well, as compaction will
4395 * try prevent permanent fragmentation by migrating from blocks of the
4397 * Don't try this for allocations that are allowed to ignore
4398 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4400 if (can_direct_reclaim &&
4402 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4403 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4404 page = __alloc_pages_direct_compact(gfp_mask, order,
4406 INIT_COMPACT_PRIORITY,
4412 * Checks for costly allocations with __GFP_NORETRY, which
4413 * includes THP page fault allocations
4415 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4417 * If compaction is deferred for high-order allocations,
4418 * it is because sync compaction recently failed. If
4419 * this is the case and the caller requested a THP
4420 * allocation, we do not want to heavily disrupt the
4421 * system, so we fail the allocation instead of entering
4424 if (compact_result == COMPACT_DEFERRED)
4428 * Looks like reclaim/compaction is worth trying, but
4429 * sync compaction could be very expensive, so keep
4430 * using async compaction.
4432 compact_priority = INIT_COMPACT_PRIORITY;
4437 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4438 if (alloc_flags & ALLOC_KSWAPD)
4439 wake_all_kswapds(order, gfp_mask, ac);
4441 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4443 alloc_flags = reserve_flags;
4446 * Reset the nodemask and zonelist iterators if memory policies can be
4447 * ignored. These allocations are high priority and system rather than
4450 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4451 ac->nodemask = NULL;
4452 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4453 ac->high_zoneidx, ac->nodemask);
4456 /* Attempt with potentially adjusted zonelist and alloc_flags */
4457 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4461 /* Caller is not willing to reclaim, we can't balance anything */
4462 if (!can_direct_reclaim)
4465 /* Avoid recursion of direct reclaim */
4466 if (current->flags & PF_MEMALLOC)
4469 /* Try direct reclaim and then allocating */
4470 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4471 &did_some_progress);
4475 /* Try direct compaction and then allocating */
4476 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4477 compact_priority, &compact_result);
4481 /* Do not loop if specifically requested */
4482 if (gfp_mask & __GFP_NORETRY)
4486 * Do not retry costly high order allocations unless they are
4487 * __GFP_RETRY_MAYFAIL
4489 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4492 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4493 did_some_progress > 0, &no_progress_loops))
4497 * It doesn't make any sense to retry for the compaction if the order-0
4498 * reclaim is not able to make any progress because the current
4499 * implementation of the compaction depends on the sufficient amount
4500 * of free memory (see __compaction_suitable)
4502 if (did_some_progress > 0 &&
4503 should_compact_retry(ac, order, alloc_flags,
4504 compact_result, &compact_priority,
4505 &compaction_retries))
4509 /* Deal with possible cpuset update races before we start OOM killing */
4510 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4513 /* Reclaim has failed us, start killing things */
4514 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4518 /* Avoid allocations with no watermarks from looping endlessly */
4519 if (tsk_is_oom_victim(current) &&
4520 (alloc_flags == ALLOC_OOM ||
4521 (gfp_mask & __GFP_NOMEMALLOC)))
4524 /* Retry as long as the OOM killer is making progress */
4525 if (did_some_progress) {
4526 no_progress_loops = 0;
4531 /* Deal with possible cpuset update races before we fail */
4532 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4536 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4539 if (gfp_mask & __GFP_NOFAIL) {
4541 * All existing users of the __GFP_NOFAIL are blockable, so warn
4542 * of any new users that actually require GFP_NOWAIT
4544 if (WARN_ON_ONCE(!can_direct_reclaim))
4548 * PF_MEMALLOC request from this context is rather bizarre
4549 * because we cannot reclaim anything and only can loop waiting
4550 * for somebody to do a work for us
4552 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4555 * non failing costly orders are a hard requirement which we
4556 * are not prepared for much so let's warn about these users
4557 * so that we can identify them and convert them to something
4560 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4563 * Help non-failing allocations by giving them access to memory
4564 * reserves but do not use ALLOC_NO_WATERMARKS because this
4565 * could deplete whole memory reserves which would just make
4566 * the situation worse
4568 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4576 warn_alloc(gfp_mask, ac->nodemask,
4577 "page allocation failure: order:%u", order);
4582 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4583 int preferred_nid, nodemask_t *nodemask,
4584 struct alloc_context *ac, gfp_t *alloc_mask,
4585 unsigned int *alloc_flags)
4587 ac->high_zoneidx = gfp_zone(gfp_mask);
4588 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4589 ac->nodemask = nodemask;
4590 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4592 if (cpusets_enabled()) {
4593 *alloc_mask |= __GFP_HARDWALL;
4595 ac->nodemask = &cpuset_current_mems_allowed;
4597 *alloc_flags |= ALLOC_CPUSET;
4600 fs_reclaim_acquire(gfp_mask);
4601 fs_reclaim_release(gfp_mask);
4603 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4605 if (should_fail_alloc_page(gfp_mask, order))
4608 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4609 *alloc_flags |= ALLOC_CMA;
4614 /* Determine whether to spread dirty pages and what the first usable zone */
4615 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4617 /* Dirty zone balancing only done in the fast path */
4618 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4621 * The preferred zone is used for statistics but crucially it is
4622 * also used as the starting point for the zonelist iterator. It
4623 * may get reset for allocations that ignore memory policies.
4625 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4626 ac->high_zoneidx, ac->nodemask);
4630 * This is the 'heart' of the zoned buddy allocator.
4633 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4634 nodemask_t *nodemask)
4637 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4638 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4639 struct alloc_context ac = { };
4642 * There are several places where we assume that the order value is sane
4643 * so bail out early if the request is out of bound.
4645 if (unlikely(order >= MAX_ORDER)) {
4646 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4650 gfp_mask &= gfp_allowed_mask;
4651 alloc_mask = gfp_mask;
4652 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4655 finalise_ac(gfp_mask, &ac);
4658 * Forbid the first pass from falling back to types that fragment
4659 * memory until all local zones are considered.
4661 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4663 /* First allocation attempt */
4664 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4669 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4670 * resp. GFP_NOIO which has to be inherited for all allocation requests
4671 * from a particular context which has been marked by
4672 * memalloc_no{fs,io}_{save,restore}.
4674 alloc_mask = current_gfp_context(gfp_mask);
4675 ac.spread_dirty_pages = false;
4678 * Restore the original nodemask if it was potentially replaced with
4679 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4681 if (unlikely(ac.nodemask != nodemask))
4682 ac.nodemask = nodemask;
4684 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4687 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4688 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4689 __free_pages(page, order);
4693 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4697 EXPORT_SYMBOL(__alloc_pages_nodemask);
4700 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4701 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4702 * you need to access high mem.
4704 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4708 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4711 return (unsigned long) page_address(page);
4713 EXPORT_SYMBOL(__get_free_pages);
4715 unsigned long get_zeroed_page(gfp_t gfp_mask)
4717 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4719 EXPORT_SYMBOL(get_zeroed_page);
4721 static inline void free_the_page(struct page *page, unsigned int order)
4723 if (order == 0) /* Via pcp? */
4724 free_unref_page(page);
4726 __free_pages_ok(page, order);
4729 void __free_pages(struct page *page, unsigned int order)
4731 if (put_page_testzero(page))
4732 free_the_page(page, order);
4734 EXPORT_SYMBOL(__free_pages);
4736 void free_pages(unsigned long addr, unsigned int order)
4739 VM_BUG_ON(!virt_addr_valid((void *)addr));
4740 __free_pages(virt_to_page((void *)addr), order);
4744 EXPORT_SYMBOL(free_pages);
4748 * An arbitrary-length arbitrary-offset area of memory which resides
4749 * within a 0 or higher order page. Multiple fragments within that page
4750 * are individually refcounted, in the page's reference counter.
4752 * The page_frag functions below provide a simple allocation framework for
4753 * page fragments. This is used by the network stack and network device
4754 * drivers to provide a backing region of memory for use as either an
4755 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4757 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4760 struct page *page = NULL;
4761 gfp_t gfp = gfp_mask;
4763 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4764 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4766 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4767 PAGE_FRAG_CACHE_MAX_ORDER);
4768 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4770 if (unlikely(!page))
4771 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4773 nc->va = page ? page_address(page) : NULL;
4778 void __page_frag_cache_drain(struct page *page, unsigned int count)
4780 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4782 if (page_ref_sub_and_test(page, count))
4783 free_the_page(page, compound_order(page));
4785 EXPORT_SYMBOL(__page_frag_cache_drain);
4787 void *page_frag_alloc(struct page_frag_cache *nc,
4788 unsigned int fragsz, gfp_t gfp_mask)
4790 unsigned int size = PAGE_SIZE;
4794 if (unlikely(!nc->va)) {
4796 page = __page_frag_cache_refill(nc, gfp_mask);
4800 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4801 /* if size can vary use size else just use PAGE_SIZE */
4804 /* Even if we own the page, we do not use atomic_set().
4805 * This would break get_page_unless_zero() users.
4807 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4809 /* reset page count bias and offset to start of new frag */
4810 nc->pfmemalloc = page_is_pfmemalloc(page);
4811 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4815 offset = nc->offset - fragsz;
4816 if (unlikely(offset < 0)) {
4817 page = virt_to_page(nc->va);
4819 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4822 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4823 /* if size can vary use size else just use PAGE_SIZE */
4826 /* OK, page count is 0, we can safely set it */
4827 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4829 /* reset page count bias and offset to start of new frag */
4830 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4831 offset = size - fragsz;
4835 nc->offset = offset;
4837 return nc->va + offset;
4839 EXPORT_SYMBOL(page_frag_alloc);
4842 * Frees a page fragment allocated out of either a compound or order 0 page.
4844 void page_frag_free(void *addr)
4846 struct page *page = virt_to_head_page(addr);
4848 if (unlikely(put_page_testzero(page)))
4849 free_the_page(page, compound_order(page));
4851 EXPORT_SYMBOL(page_frag_free);
4853 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4857 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4858 unsigned long used = addr + PAGE_ALIGN(size);
4860 split_page(virt_to_page((void *)addr), order);
4861 while (used < alloc_end) {
4866 return (void *)addr;
4870 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4871 * @size: the number of bytes to allocate
4872 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4874 * This function is similar to alloc_pages(), except that it allocates the
4875 * minimum number of pages to satisfy the request. alloc_pages() can only
4876 * allocate memory in power-of-two pages.
4878 * This function is also limited by MAX_ORDER.
4880 * Memory allocated by this function must be released by free_pages_exact().
4882 * Return: pointer to the allocated area or %NULL in case of error.
4884 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4886 unsigned int order = get_order(size);
4889 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4890 gfp_mask &= ~__GFP_COMP;
4892 addr = __get_free_pages(gfp_mask, order);
4893 return make_alloc_exact(addr, order, size);
4895 EXPORT_SYMBOL(alloc_pages_exact);
4898 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4900 * @nid: the preferred node ID where memory should be allocated
4901 * @size: the number of bytes to allocate
4902 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4904 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4907 * Return: pointer to the allocated area or %NULL in case of error.
4909 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4911 unsigned int order = get_order(size);
4914 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4915 gfp_mask &= ~__GFP_COMP;
4917 p = alloc_pages_node(nid, gfp_mask, order);
4920 return make_alloc_exact((unsigned long)page_address(p), order, size);
4924 * free_pages_exact - release memory allocated via alloc_pages_exact()
4925 * @virt: the value returned by alloc_pages_exact.
4926 * @size: size of allocation, same value as passed to alloc_pages_exact().
4928 * Release the memory allocated by a previous call to alloc_pages_exact.
4930 void free_pages_exact(void *virt, size_t size)
4932 unsigned long addr = (unsigned long)virt;
4933 unsigned long end = addr + PAGE_ALIGN(size);
4935 while (addr < end) {
4940 EXPORT_SYMBOL(free_pages_exact);
4943 * nr_free_zone_pages - count number of pages beyond high watermark
4944 * @offset: The zone index of the highest zone
4946 * nr_free_zone_pages() counts the number of pages which are beyond the
4947 * high watermark within all zones at or below a given zone index. For each
4948 * zone, the number of pages is calculated as:
4950 * nr_free_zone_pages = managed_pages - high_pages
4952 * Return: number of pages beyond high watermark.
4954 static unsigned long nr_free_zone_pages(int offset)
4959 /* Just pick one node, since fallback list is circular */
4960 unsigned long sum = 0;
4962 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4964 for_each_zone_zonelist(zone, z, zonelist, offset) {
4965 unsigned long size = zone_managed_pages(zone);
4966 unsigned long high = high_wmark_pages(zone);
4975 * nr_free_buffer_pages - count number of pages beyond high watermark
4977 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4978 * watermark within ZONE_DMA and ZONE_NORMAL.
4980 * Return: number of pages beyond high watermark within ZONE_DMA and
4983 unsigned long nr_free_buffer_pages(void)
4985 return nr_free_zone_pages(gfp_zone(GFP_USER));
4987 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4990 * nr_free_pagecache_pages - count number of pages beyond high watermark
4992 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4993 * high watermark within all zones.
4995 * Return: number of pages beyond high watermark within all zones.
4997 unsigned long nr_free_pagecache_pages(void)
4999 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5002 static inline void show_node(struct zone *zone)
5004 if (IS_ENABLED(CONFIG_NUMA))
5005 printk("Node %d ", zone_to_nid(zone));
5008 long si_mem_available(void)
5011 unsigned long pagecache;
5012 unsigned long wmark_low = 0;
5013 unsigned long pages[NR_LRU_LISTS];
5014 unsigned long reclaimable;
5018 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5019 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5022 wmark_low += low_wmark_pages(zone);
5025 * Estimate the amount of memory available for userspace allocations,
5026 * without causing swapping.
5028 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5031 * Not all the page cache can be freed, otherwise the system will
5032 * start swapping. Assume at least half of the page cache, or the
5033 * low watermark worth of cache, needs to stay.
5035 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5036 pagecache -= min(pagecache / 2, wmark_low);
5037 available += pagecache;
5040 * Part of the reclaimable slab and other kernel memory consists of
5041 * items that are in use, and cannot be freed. Cap this estimate at the
5044 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5045 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5046 available += reclaimable - min(reclaimable / 2, wmark_low);
5052 EXPORT_SYMBOL_GPL(si_mem_available);
5054 void si_meminfo(struct sysinfo *val)
5056 val->totalram = totalram_pages();
5057 val->sharedram = global_node_page_state(NR_SHMEM);
5058 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5059 val->bufferram = nr_blockdev_pages();
5060 val->totalhigh = totalhigh_pages();
5061 val->freehigh = nr_free_highpages();
5062 val->mem_unit = PAGE_SIZE;
5065 EXPORT_SYMBOL(si_meminfo);
5068 void si_meminfo_node(struct sysinfo *val, int nid)
5070 int zone_type; /* needs to be signed */
5071 unsigned long managed_pages = 0;
5072 unsigned long managed_highpages = 0;
5073 unsigned long free_highpages = 0;
5074 pg_data_t *pgdat = NODE_DATA(nid);
5076 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5077 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5078 val->totalram = managed_pages;
5079 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5080 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5081 #ifdef CONFIG_HIGHMEM
5082 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5083 struct zone *zone = &pgdat->node_zones[zone_type];
5085 if (is_highmem(zone)) {
5086 managed_highpages += zone_managed_pages(zone);
5087 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5090 val->totalhigh = managed_highpages;
5091 val->freehigh = free_highpages;
5093 val->totalhigh = managed_highpages;
5094 val->freehigh = free_highpages;
5096 val->mem_unit = PAGE_SIZE;
5101 * Determine whether the node should be displayed or not, depending on whether
5102 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5104 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5106 if (!(flags & SHOW_MEM_FILTER_NODES))
5110 * no node mask - aka implicit memory numa policy. Do not bother with
5111 * the synchronization - read_mems_allowed_begin - because we do not
5112 * have to be precise here.
5115 nodemask = &cpuset_current_mems_allowed;
5117 return !node_isset(nid, *nodemask);
5120 #define K(x) ((x) << (PAGE_SHIFT-10))
5122 static void show_migration_types(unsigned char type)
5124 static const char types[MIGRATE_TYPES] = {
5125 [MIGRATE_UNMOVABLE] = 'U',
5126 [MIGRATE_MOVABLE] = 'M',
5127 [MIGRATE_RECLAIMABLE] = 'E',
5128 [MIGRATE_HIGHATOMIC] = 'H',
5130 [MIGRATE_CMA] = 'C',
5132 #ifdef CONFIG_MEMORY_ISOLATION
5133 [MIGRATE_ISOLATE] = 'I',
5136 char tmp[MIGRATE_TYPES + 1];
5140 for (i = 0; i < MIGRATE_TYPES; i++) {
5141 if (type & (1 << i))
5146 printk(KERN_CONT "(%s) ", tmp);
5150 * Show free area list (used inside shift_scroll-lock stuff)
5151 * We also calculate the percentage fragmentation. We do this by counting the
5152 * memory on each free list with the exception of the first item on the list.
5155 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5158 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5160 unsigned long free_pcp = 0;
5165 for_each_populated_zone(zone) {
5166 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5169 for_each_online_cpu(cpu)
5170 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5173 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5174 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5175 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5176 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5177 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5178 " free:%lu free_pcp:%lu free_cma:%lu\n",
5179 global_node_page_state(NR_ACTIVE_ANON),
5180 global_node_page_state(NR_INACTIVE_ANON),
5181 global_node_page_state(NR_ISOLATED_ANON),
5182 global_node_page_state(NR_ACTIVE_FILE),
5183 global_node_page_state(NR_INACTIVE_FILE),
5184 global_node_page_state(NR_ISOLATED_FILE),
5185 global_node_page_state(NR_UNEVICTABLE),
5186 global_node_page_state(NR_FILE_DIRTY),
5187 global_node_page_state(NR_WRITEBACK),
5188 global_node_page_state(NR_UNSTABLE_NFS),
5189 global_node_page_state(NR_SLAB_RECLAIMABLE),
5190 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5191 global_node_page_state(NR_FILE_MAPPED),
5192 global_node_page_state(NR_SHMEM),
5193 global_zone_page_state(NR_PAGETABLE),
5194 global_zone_page_state(NR_BOUNCE),
5195 global_zone_page_state(NR_FREE_PAGES),
5197 global_zone_page_state(NR_FREE_CMA_PAGES));
5199 for_each_online_pgdat(pgdat) {
5200 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5204 " active_anon:%lukB"
5205 " inactive_anon:%lukB"
5206 " active_file:%lukB"
5207 " inactive_file:%lukB"
5208 " unevictable:%lukB"
5209 " isolated(anon):%lukB"
5210 " isolated(file):%lukB"
5215 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5217 " shmem_pmdmapped: %lukB"
5220 " writeback_tmp:%lukB"
5222 " all_unreclaimable? %s"
5225 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5226 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5227 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5228 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5229 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5230 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5231 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5232 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5233 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5234 K(node_page_state(pgdat, NR_WRITEBACK)),
5235 K(node_page_state(pgdat, NR_SHMEM)),
5236 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5237 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5238 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5240 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5242 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5243 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5244 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5248 for_each_populated_zone(zone) {
5251 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5255 for_each_online_cpu(cpu)
5256 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5265 " active_anon:%lukB"
5266 " inactive_anon:%lukB"
5267 " active_file:%lukB"
5268 " inactive_file:%lukB"
5269 " unevictable:%lukB"
5270 " writepending:%lukB"
5274 " kernel_stack:%lukB"
5282 K(zone_page_state(zone, NR_FREE_PAGES)),
5283 K(min_wmark_pages(zone)),
5284 K(low_wmark_pages(zone)),
5285 K(high_wmark_pages(zone)),
5286 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5287 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5288 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5289 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5290 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5291 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5292 K(zone->present_pages),
5293 K(zone_managed_pages(zone)),
5294 K(zone_page_state(zone, NR_MLOCK)),
5295 zone_page_state(zone, NR_KERNEL_STACK_KB),
5296 K(zone_page_state(zone, NR_PAGETABLE)),
5297 K(zone_page_state(zone, NR_BOUNCE)),
5299 K(this_cpu_read(zone->pageset->pcp.count)),
5300 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5301 printk("lowmem_reserve[]:");
5302 for (i = 0; i < MAX_NR_ZONES; i++)
5303 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5304 printk(KERN_CONT "\n");
5307 for_each_populated_zone(zone) {
5309 unsigned long nr[MAX_ORDER], flags, total = 0;
5310 unsigned char types[MAX_ORDER];
5312 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5315 printk(KERN_CONT "%s: ", zone->name);
5317 spin_lock_irqsave(&zone->lock, flags);
5318 for (order = 0; order < MAX_ORDER; order++) {
5319 struct free_area *area = &zone->free_area[order];
5322 nr[order] = area->nr_free;
5323 total += nr[order] << order;
5326 for (type = 0; type < MIGRATE_TYPES; type++) {
5327 if (!list_empty(&area->free_list[type]))
5328 types[order] |= 1 << type;
5331 spin_unlock_irqrestore(&zone->lock, flags);
5332 for (order = 0; order < MAX_ORDER; order++) {
5333 printk(KERN_CONT "%lu*%lukB ",
5334 nr[order], K(1UL) << order);
5336 show_migration_types(types[order]);
5338 printk(KERN_CONT "= %lukB\n", K(total));
5341 hugetlb_show_meminfo();
5343 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5345 show_swap_cache_info();
5348 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5350 zoneref->zone = zone;
5351 zoneref->zone_idx = zone_idx(zone);
5355 * Builds allocation fallback zone lists.
5357 * Add all populated zones of a node to the zonelist.
5359 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5362 enum zone_type zone_type = MAX_NR_ZONES;
5367 zone = pgdat->node_zones + zone_type;
5368 if (managed_zone(zone)) {
5369 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5370 check_highest_zone(zone_type);
5372 } while (zone_type);
5379 static int __parse_numa_zonelist_order(char *s)
5382 * We used to support different zonlists modes but they turned
5383 * out to be just not useful. Let's keep the warning in place
5384 * if somebody still use the cmd line parameter so that we do
5385 * not fail it silently
5387 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5388 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5394 static __init int setup_numa_zonelist_order(char *s)
5399 return __parse_numa_zonelist_order(s);
5401 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5403 char numa_zonelist_order[] = "Node";
5406 * sysctl handler for numa_zonelist_order
5408 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5409 void __user *buffer, size_t *length,
5416 return proc_dostring(table, write, buffer, length, ppos);
5417 str = memdup_user_nul(buffer, 16);
5419 return PTR_ERR(str);
5421 ret = __parse_numa_zonelist_order(str);
5427 #define MAX_NODE_LOAD (nr_online_nodes)
5428 static int node_load[MAX_NUMNODES];
5431 * find_next_best_node - find the next node that should appear in a given node's fallback list
5432 * @node: node whose fallback list we're appending
5433 * @used_node_mask: nodemask_t of already used nodes
5435 * We use a number of factors to determine which is the next node that should
5436 * appear on a given node's fallback list. The node should not have appeared
5437 * already in @node's fallback list, and it should be the next closest node
5438 * according to the distance array (which contains arbitrary distance values
5439 * from each node to each node in the system), and should also prefer nodes
5440 * with no CPUs, since presumably they'll have very little allocation pressure
5441 * on them otherwise.
5443 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5445 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5448 int min_val = INT_MAX;
5449 int best_node = NUMA_NO_NODE;
5450 const struct cpumask *tmp = cpumask_of_node(0);
5452 /* Use the local node if we haven't already */
5453 if (!node_isset(node, *used_node_mask)) {
5454 node_set(node, *used_node_mask);
5458 for_each_node_state(n, N_MEMORY) {
5460 /* Don't want a node to appear more than once */
5461 if (node_isset(n, *used_node_mask))
5464 /* Use the distance array to find the distance */
5465 val = node_distance(node, n);
5467 /* Penalize nodes under us ("prefer the next node") */
5470 /* Give preference to headless and unused nodes */
5471 tmp = cpumask_of_node(n);
5472 if (!cpumask_empty(tmp))
5473 val += PENALTY_FOR_NODE_WITH_CPUS;
5475 /* Slight preference for less loaded node */
5476 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5477 val += node_load[n];
5479 if (val < min_val) {
5486 node_set(best_node, *used_node_mask);
5493 * Build zonelists ordered by node and zones within node.
5494 * This results in maximum locality--normal zone overflows into local
5495 * DMA zone, if any--but risks exhausting DMA zone.
5497 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5500 struct zoneref *zonerefs;
5503 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5505 for (i = 0; i < nr_nodes; i++) {
5508 pg_data_t *node = NODE_DATA(node_order[i]);
5510 nr_zones = build_zonerefs_node(node, zonerefs);
5511 zonerefs += nr_zones;
5513 zonerefs->zone = NULL;
5514 zonerefs->zone_idx = 0;
5518 * Build gfp_thisnode zonelists
5520 static void build_thisnode_zonelists(pg_data_t *pgdat)
5522 struct zoneref *zonerefs;
5525 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5526 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5527 zonerefs += nr_zones;
5528 zonerefs->zone = NULL;
5529 zonerefs->zone_idx = 0;
5533 * Build zonelists ordered by zone and nodes within zones.
5534 * This results in conserving DMA zone[s] until all Normal memory is
5535 * exhausted, but results in overflowing to remote node while memory
5536 * may still exist in local DMA zone.
5539 static void build_zonelists(pg_data_t *pgdat)
5541 static int node_order[MAX_NUMNODES];
5542 int node, load, nr_nodes = 0;
5543 nodemask_t used_mask;
5544 int local_node, prev_node;
5546 /* NUMA-aware ordering of nodes */
5547 local_node = pgdat->node_id;
5548 load = nr_online_nodes;
5549 prev_node = local_node;
5550 nodes_clear(used_mask);
5552 memset(node_order, 0, sizeof(node_order));
5553 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5555 * We don't want to pressure a particular node.
5556 * So adding penalty to the first node in same
5557 * distance group to make it round-robin.
5559 if (node_distance(local_node, node) !=
5560 node_distance(local_node, prev_node))
5561 node_load[node] = load;
5563 node_order[nr_nodes++] = node;
5568 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5569 build_thisnode_zonelists(pgdat);
5572 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5574 * Return node id of node used for "local" allocations.
5575 * I.e., first node id of first zone in arg node's generic zonelist.
5576 * Used for initializing percpu 'numa_mem', which is used primarily
5577 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5579 int local_memory_node(int node)
5583 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5584 gfp_zone(GFP_KERNEL),
5586 return zone_to_nid(z->zone);
5590 static void setup_min_unmapped_ratio(void);
5591 static void setup_min_slab_ratio(void);
5592 #else /* CONFIG_NUMA */
5594 static void build_zonelists(pg_data_t *pgdat)
5596 int node, local_node;
5597 struct zoneref *zonerefs;
5600 local_node = pgdat->node_id;
5602 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5603 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5604 zonerefs += nr_zones;
5607 * Now we build the zonelist so that it contains the zones
5608 * of all the other nodes.
5609 * We don't want to pressure a particular node, so when
5610 * building the zones for node N, we make sure that the
5611 * zones coming right after the local ones are those from
5612 * node N+1 (modulo N)
5614 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5615 if (!node_online(node))
5617 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5618 zonerefs += nr_zones;
5620 for (node = 0; node < local_node; node++) {
5621 if (!node_online(node))
5623 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5624 zonerefs += nr_zones;
5627 zonerefs->zone = NULL;
5628 zonerefs->zone_idx = 0;
5631 #endif /* CONFIG_NUMA */
5634 * Boot pageset table. One per cpu which is going to be used for all
5635 * zones and all nodes. The parameters will be set in such a way
5636 * that an item put on a list will immediately be handed over to
5637 * the buddy list. This is safe since pageset manipulation is done
5638 * with interrupts disabled.
5640 * The boot_pagesets must be kept even after bootup is complete for
5641 * unused processors and/or zones. They do play a role for bootstrapping
5642 * hotplugged processors.
5644 * zoneinfo_show() and maybe other functions do
5645 * not check if the processor is online before following the pageset pointer.
5646 * Other parts of the kernel may not check if the zone is available.
5648 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5649 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5650 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5652 static void __build_all_zonelists(void *data)
5655 int __maybe_unused cpu;
5656 pg_data_t *self = data;
5657 static DEFINE_SPINLOCK(lock);
5662 memset(node_load, 0, sizeof(node_load));
5666 * This node is hotadded and no memory is yet present. So just
5667 * building zonelists is fine - no need to touch other nodes.
5669 if (self && !node_online(self->node_id)) {
5670 build_zonelists(self);
5672 for_each_online_node(nid) {
5673 pg_data_t *pgdat = NODE_DATA(nid);
5675 build_zonelists(pgdat);
5678 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5680 * We now know the "local memory node" for each node--
5681 * i.e., the node of the first zone in the generic zonelist.
5682 * Set up numa_mem percpu variable for on-line cpus. During
5683 * boot, only the boot cpu should be on-line; we'll init the
5684 * secondary cpus' numa_mem as they come on-line. During
5685 * node/memory hotplug, we'll fixup all on-line cpus.
5687 for_each_online_cpu(cpu)
5688 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5695 static noinline void __init
5696 build_all_zonelists_init(void)
5700 __build_all_zonelists(NULL);
5703 * Initialize the boot_pagesets that are going to be used
5704 * for bootstrapping processors. The real pagesets for
5705 * each zone will be allocated later when the per cpu
5706 * allocator is available.
5708 * boot_pagesets are used also for bootstrapping offline
5709 * cpus if the system is already booted because the pagesets
5710 * are needed to initialize allocators on a specific cpu too.
5711 * F.e. the percpu allocator needs the page allocator which
5712 * needs the percpu allocator in order to allocate its pagesets
5713 * (a chicken-egg dilemma).
5715 for_each_possible_cpu(cpu)
5716 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5718 mminit_verify_zonelist();
5719 cpuset_init_current_mems_allowed();
5723 * unless system_state == SYSTEM_BOOTING.
5725 * __ref due to call of __init annotated helper build_all_zonelists_init
5726 * [protected by SYSTEM_BOOTING].
5728 void __ref build_all_zonelists(pg_data_t *pgdat)
5730 if (system_state == SYSTEM_BOOTING) {
5731 build_all_zonelists_init();
5733 __build_all_zonelists(pgdat);
5734 /* cpuset refresh routine should be here */
5736 vm_total_pages = nr_free_pagecache_pages();
5738 * Disable grouping by mobility if the number of pages in the
5739 * system is too low to allow the mechanism to work. It would be
5740 * more accurate, but expensive to check per-zone. This check is
5741 * made on memory-hotadd so a system can start with mobility
5742 * disabled and enable it later
5744 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5745 page_group_by_mobility_disabled = 1;
5747 page_group_by_mobility_disabled = 0;
5749 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5751 page_group_by_mobility_disabled ? "off" : "on",
5754 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5758 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5759 static bool __meminit
5760 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5762 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5763 static struct memblock_region *r;
5765 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5766 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5767 for_each_memblock(memory, r) {
5768 if (*pfn < memblock_region_memory_end_pfn(r))
5772 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5773 memblock_is_mirror(r)) {
5774 *pfn = memblock_region_memory_end_pfn(r);
5783 * Initially all pages are reserved - free ones are freed
5784 * up by memblock_free_all() once the early boot process is
5785 * done. Non-atomic initialization, single-pass.
5787 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5788 unsigned long start_pfn, enum memmap_context context,
5789 struct vmem_altmap *altmap)
5791 unsigned long pfn, end_pfn = start_pfn + size;
5794 if (highest_memmap_pfn < end_pfn - 1)
5795 highest_memmap_pfn = end_pfn - 1;
5797 #ifdef CONFIG_ZONE_DEVICE
5799 * Honor reservation requested by the driver for this ZONE_DEVICE
5800 * memory. We limit the total number of pages to initialize to just
5801 * those that might contain the memory mapping. We will defer the
5802 * ZONE_DEVICE page initialization until after we have released
5805 if (zone == ZONE_DEVICE) {
5809 if (start_pfn == altmap->base_pfn)
5810 start_pfn += altmap->reserve;
5811 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5815 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5817 * There can be holes in boot-time mem_map[]s handed to this
5818 * function. They do not exist on hotplugged memory.
5820 if (context == MEMMAP_EARLY) {
5821 if (!early_pfn_valid(pfn))
5823 if (!early_pfn_in_nid(pfn, nid))
5825 if (overlap_memmap_init(zone, &pfn))
5827 if (defer_init(nid, pfn, end_pfn))
5831 page = pfn_to_page(pfn);
5832 __init_single_page(page, pfn, zone, nid);
5833 if (context == MEMMAP_HOTPLUG)
5834 __SetPageReserved(page);
5837 * Mark the block movable so that blocks are reserved for
5838 * movable at startup. This will force kernel allocations
5839 * to reserve their blocks rather than leaking throughout
5840 * the address space during boot when many long-lived
5841 * kernel allocations are made.
5843 * bitmap is created for zone's valid pfn range. but memmap
5844 * can be created for invalid pages (for alignment)
5845 * check here not to call set_pageblock_migratetype() against
5848 if (!(pfn & (pageblock_nr_pages - 1))) {
5849 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5855 #ifdef CONFIG_ZONE_DEVICE
5856 void __ref memmap_init_zone_device(struct zone *zone,
5857 unsigned long start_pfn,
5859 struct dev_pagemap *pgmap)
5861 unsigned long pfn, end_pfn = start_pfn + size;
5862 struct pglist_data *pgdat = zone->zone_pgdat;
5863 unsigned long zone_idx = zone_idx(zone);
5864 unsigned long start = jiffies;
5865 int nid = pgdat->node_id;
5867 if (WARN_ON_ONCE(!pgmap || !is_dev_zone(zone)))
5871 * The call to memmap_init_zone should have already taken care
5872 * of the pages reserved for the memmap, so we can just jump to
5873 * the end of that region and start processing the device pages.
5875 if (pgmap->altmap_valid) {
5876 struct vmem_altmap *altmap = &pgmap->altmap;
5878 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5879 size = end_pfn - start_pfn;
5882 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5883 struct page *page = pfn_to_page(pfn);
5885 __init_single_page(page, pfn, zone_idx, nid);
5888 * Mark page reserved as it will need to wait for onlining
5889 * phase for it to be fully associated with a zone.
5891 * We can use the non-atomic __set_bit operation for setting
5892 * the flag as we are still initializing the pages.
5894 __SetPageReserved(page);
5897 * ZONE_DEVICE pages union ->lru with a ->pgmap back
5898 * pointer and hmm_data. It is a bug if a ZONE_DEVICE
5899 * page is ever freed or placed on a driver-private list.
5901 page->pgmap = pgmap;
5905 * Mark the block movable so that blocks are reserved for
5906 * movable at startup. This will force kernel allocations
5907 * to reserve their blocks rather than leaking throughout
5908 * the address space during boot when many long-lived
5909 * kernel allocations are made.
5911 * bitmap is created for zone's valid pfn range. but memmap
5912 * can be created for invalid pages (for alignment)
5913 * check here not to call set_pageblock_migratetype() against
5916 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5917 * because this is done early in sparse_add_one_section
5919 if (!(pfn & (pageblock_nr_pages - 1))) {
5920 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5925 pr_info("%s initialised, %lu pages in %ums\n", dev_name(pgmap->dev),
5926 size, jiffies_to_msecs(jiffies - start));
5930 static void __meminit zone_init_free_lists(struct zone *zone)
5932 unsigned int order, t;
5933 for_each_migratetype_order(order, t) {
5934 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5935 zone->free_area[order].nr_free = 0;
5939 void __meminit __weak memmap_init(unsigned long size, int nid,
5940 unsigned long zone, unsigned long start_pfn)
5942 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
5945 static int zone_batchsize(struct zone *zone)
5951 * The per-cpu-pages pools are set to around 1000th of the
5954 batch = zone_managed_pages(zone) / 1024;
5955 /* But no more than a meg. */
5956 if (batch * PAGE_SIZE > 1024 * 1024)
5957 batch = (1024 * 1024) / PAGE_SIZE;
5958 batch /= 4; /* We effectively *= 4 below */
5963 * Clamp the batch to a 2^n - 1 value. Having a power
5964 * of 2 value was found to be more likely to have
5965 * suboptimal cache aliasing properties in some cases.
5967 * For example if 2 tasks are alternately allocating
5968 * batches of pages, one task can end up with a lot
5969 * of pages of one half of the possible page colors
5970 * and the other with pages of the other colors.
5972 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5977 /* The deferral and batching of frees should be suppressed under NOMMU
5980 * The problem is that NOMMU needs to be able to allocate large chunks
5981 * of contiguous memory as there's no hardware page translation to
5982 * assemble apparent contiguous memory from discontiguous pages.
5984 * Queueing large contiguous runs of pages for batching, however,
5985 * causes the pages to actually be freed in smaller chunks. As there
5986 * can be a significant delay between the individual batches being
5987 * recycled, this leads to the once large chunks of space being
5988 * fragmented and becoming unavailable for high-order allocations.
5995 * pcp->high and pcp->batch values are related and dependent on one another:
5996 * ->batch must never be higher then ->high.
5997 * The following function updates them in a safe manner without read side
6000 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6001 * those fields changing asynchronously (acording the the above rule).
6003 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6004 * outside of boot time (or some other assurance that no concurrent updaters
6007 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6008 unsigned long batch)
6010 /* start with a fail safe value for batch */
6014 /* Update high, then batch, in order */
6021 /* a companion to pageset_set_high() */
6022 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6024 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6027 static void pageset_init(struct per_cpu_pageset *p)
6029 struct per_cpu_pages *pcp;
6032 memset(p, 0, sizeof(*p));
6035 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6036 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6039 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6042 pageset_set_batch(p, batch);
6046 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6047 * to the value high for the pageset p.
6049 static void pageset_set_high(struct per_cpu_pageset *p,
6052 unsigned long batch = max(1UL, high / 4);
6053 if ((high / 4) > (PAGE_SHIFT * 8))
6054 batch = PAGE_SHIFT * 8;
6056 pageset_update(&p->pcp, high, batch);
6059 static void pageset_set_high_and_batch(struct zone *zone,
6060 struct per_cpu_pageset *pcp)
6062 if (percpu_pagelist_fraction)
6063 pageset_set_high(pcp,
6064 (zone_managed_pages(zone) /
6065 percpu_pagelist_fraction));
6067 pageset_set_batch(pcp, zone_batchsize(zone));
6070 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6072 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6075 pageset_set_high_and_batch(zone, pcp);
6078 void __meminit setup_zone_pageset(struct zone *zone)
6081 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6082 for_each_possible_cpu(cpu)
6083 zone_pageset_init(zone, cpu);
6087 * Allocate per cpu pagesets and initialize them.
6088 * Before this call only boot pagesets were available.
6090 void __init setup_per_cpu_pageset(void)
6092 struct pglist_data *pgdat;
6095 for_each_populated_zone(zone)
6096 setup_zone_pageset(zone);
6098 for_each_online_pgdat(pgdat)
6099 pgdat->per_cpu_nodestats =
6100 alloc_percpu(struct per_cpu_nodestat);
6103 static __meminit void zone_pcp_init(struct zone *zone)
6106 * per cpu subsystem is not up at this point. The following code
6107 * relies on the ability of the linker to provide the
6108 * offset of a (static) per cpu variable into the per cpu area.
6110 zone->pageset = &boot_pageset;
6112 if (populated_zone(zone))
6113 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6114 zone->name, zone->present_pages,
6115 zone_batchsize(zone));
6118 void __meminit init_currently_empty_zone(struct zone *zone,
6119 unsigned long zone_start_pfn,
6122 struct pglist_data *pgdat = zone->zone_pgdat;
6123 int zone_idx = zone_idx(zone) + 1;
6125 if (zone_idx > pgdat->nr_zones)
6126 pgdat->nr_zones = zone_idx;
6128 zone->zone_start_pfn = zone_start_pfn;
6130 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6131 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6133 (unsigned long)zone_idx(zone),
6134 zone_start_pfn, (zone_start_pfn + size));
6136 zone_init_free_lists(zone);
6137 zone->initialized = 1;
6140 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6141 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6144 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6146 int __meminit __early_pfn_to_nid(unsigned long pfn,
6147 struct mminit_pfnnid_cache *state)
6149 unsigned long start_pfn, end_pfn;
6152 if (state->last_start <= pfn && pfn < state->last_end)
6153 return state->last_nid;
6155 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6156 if (nid != NUMA_NO_NODE) {
6157 state->last_start = start_pfn;
6158 state->last_end = end_pfn;
6159 state->last_nid = nid;
6164 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6167 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6168 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6169 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6171 * If an architecture guarantees that all ranges registered contain no holes
6172 * and may be freed, this this function may be used instead of calling
6173 * memblock_free_early_nid() manually.
6175 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6177 unsigned long start_pfn, end_pfn;
6180 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6181 start_pfn = min(start_pfn, max_low_pfn);
6182 end_pfn = min(end_pfn, max_low_pfn);
6184 if (start_pfn < end_pfn)
6185 memblock_free_early_nid(PFN_PHYS(start_pfn),
6186 (end_pfn - start_pfn) << PAGE_SHIFT,
6192 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6193 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6195 * If an architecture guarantees that all ranges registered contain no holes and may
6196 * be freed, this function may be used instead of calling memory_present() manually.
6198 void __init sparse_memory_present_with_active_regions(int nid)
6200 unsigned long start_pfn, end_pfn;
6203 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6204 memory_present(this_nid, start_pfn, end_pfn);
6208 * get_pfn_range_for_nid - Return the start and end page frames for a node
6209 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6210 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6211 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6213 * It returns the start and end page frame of a node based on information
6214 * provided by memblock_set_node(). If called for a node
6215 * with no available memory, a warning is printed and the start and end
6218 void __init get_pfn_range_for_nid(unsigned int nid,
6219 unsigned long *start_pfn, unsigned long *end_pfn)
6221 unsigned long this_start_pfn, this_end_pfn;
6227 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6228 *start_pfn = min(*start_pfn, this_start_pfn);
6229 *end_pfn = max(*end_pfn, this_end_pfn);
6232 if (*start_pfn == -1UL)
6237 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6238 * assumption is made that zones within a node are ordered in monotonic
6239 * increasing memory addresses so that the "highest" populated zone is used
6241 static void __init find_usable_zone_for_movable(void)
6244 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6245 if (zone_index == ZONE_MOVABLE)
6248 if (arch_zone_highest_possible_pfn[zone_index] >
6249 arch_zone_lowest_possible_pfn[zone_index])
6253 VM_BUG_ON(zone_index == -1);
6254 movable_zone = zone_index;
6258 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6259 * because it is sized independent of architecture. Unlike the other zones,
6260 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6261 * in each node depending on the size of each node and how evenly kernelcore
6262 * is distributed. This helper function adjusts the zone ranges
6263 * provided by the architecture for a given node by using the end of the
6264 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6265 * zones within a node are in order of monotonic increases memory addresses
6267 static void __init adjust_zone_range_for_zone_movable(int nid,
6268 unsigned long zone_type,
6269 unsigned long node_start_pfn,
6270 unsigned long node_end_pfn,
6271 unsigned long *zone_start_pfn,
6272 unsigned long *zone_end_pfn)
6274 /* Only adjust if ZONE_MOVABLE is on this node */
6275 if (zone_movable_pfn[nid]) {
6276 /* Size ZONE_MOVABLE */
6277 if (zone_type == ZONE_MOVABLE) {
6278 *zone_start_pfn = zone_movable_pfn[nid];
6279 *zone_end_pfn = min(node_end_pfn,
6280 arch_zone_highest_possible_pfn[movable_zone]);
6282 /* Adjust for ZONE_MOVABLE starting within this range */
6283 } else if (!mirrored_kernelcore &&
6284 *zone_start_pfn < zone_movable_pfn[nid] &&
6285 *zone_end_pfn > zone_movable_pfn[nid]) {
6286 *zone_end_pfn = zone_movable_pfn[nid];
6288 /* Check if this whole range is within ZONE_MOVABLE */
6289 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6290 *zone_start_pfn = *zone_end_pfn;
6295 * Return the number of pages a zone spans in a node, including holes
6296 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6298 static unsigned long __init zone_spanned_pages_in_node(int nid,
6299 unsigned long zone_type,
6300 unsigned long node_start_pfn,
6301 unsigned long node_end_pfn,
6302 unsigned long *zone_start_pfn,
6303 unsigned long *zone_end_pfn,
6304 unsigned long *ignored)
6306 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6307 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6308 /* When hotadd a new node from cpu_up(), the node should be empty */
6309 if (!node_start_pfn && !node_end_pfn)
6312 /* Get the start and end of the zone */
6313 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6314 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6315 adjust_zone_range_for_zone_movable(nid, zone_type,
6316 node_start_pfn, node_end_pfn,
6317 zone_start_pfn, zone_end_pfn);
6319 /* Check that this node has pages within the zone's required range */
6320 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6323 /* Move the zone boundaries inside the node if necessary */
6324 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6325 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6327 /* Return the spanned pages */
6328 return *zone_end_pfn - *zone_start_pfn;
6332 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6333 * then all holes in the requested range will be accounted for.
6335 unsigned long __init __absent_pages_in_range(int nid,
6336 unsigned long range_start_pfn,
6337 unsigned long range_end_pfn)
6339 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6340 unsigned long start_pfn, end_pfn;
6343 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6344 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6345 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6346 nr_absent -= end_pfn - start_pfn;
6352 * absent_pages_in_range - Return number of page frames in holes within a range
6353 * @start_pfn: The start PFN to start searching for holes
6354 * @end_pfn: The end PFN to stop searching for holes
6356 * Return: the number of pages frames in memory holes within a range.
6358 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6359 unsigned long end_pfn)
6361 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6364 /* Return the number of page frames in holes in a zone on a node */
6365 static unsigned long __init zone_absent_pages_in_node(int nid,
6366 unsigned long zone_type,
6367 unsigned long node_start_pfn,
6368 unsigned long node_end_pfn,
6369 unsigned long *ignored)
6371 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6372 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6373 unsigned long zone_start_pfn, zone_end_pfn;
6374 unsigned long nr_absent;
6376 /* When hotadd a new node from cpu_up(), the node should be empty */
6377 if (!node_start_pfn && !node_end_pfn)
6380 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6381 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6383 adjust_zone_range_for_zone_movable(nid, zone_type,
6384 node_start_pfn, node_end_pfn,
6385 &zone_start_pfn, &zone_end_pfn);
6386 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6389 * ZONE_MOVABLE handling.
6390 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6393 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6394 unsigned long start_pfn, end_pfn;
6395 struct memblock_region *r;
6397 for_each_memblock(memory, r) {
6398 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6399 zone_start_pfn, zone_end_pfn);
6400 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6401 zone_start_pfn, zone_end_pfn);
6403 if (zone_type == ZONE_MOVABLE &&
6404 memblock_is_mirror(r))
6405 nr_absent += end_pfn - start_pfn;
6407 if (zone_type == ZONE_NORMAL &&
6408 !memblock_is_mirror(r))
6409 nr_absent += end_pfn - start_pfn;
6416 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6417 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6418 unsigned long zone_type,
6419 unsigned long node_start_pfn,
6420 unsigned long node_end_pfn,
6421 unsigned long *zone_start_pfn,
6422 unsigned long *zone_end_pfn,
6423 unsigned long *zones_size)
6427 *zone_start_pfn = node_start_pfn;
6428 for (zone = 0; zone < zone_type; zone++)
6429 *zone_start_pfn += zones_size[zone];
6431 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6433 return zones_size[zone_type];
6436 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6437 unsigned long zone_type,
6438 unsigned long node_start_pfn,
6439 unsigned long node_end_pfn,
6440 unsigned long *zholes_size)
6445 return zholes_size[zone_type];
6448 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6450 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6451 unsigned long node_start_pfn,
6452 unsigned long node_end_pfn,
6453 unsigned long *zones_size,
6454 unsigned long *zholes_size)
6456 unsigned long realtotalpages = 0, totalpages = 0;
6459 for (i = 0; i < MAX_NR_ZONES; i++) {
6460 struct zone *zone = pgdat->node_zones + i;
6461 unsigned long zone_start_pfn, zone_end_pfn;
6462 unsigned long size, real_size;
6464 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6470 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6471 node_start_pfn, node_end_pfn,
6474 zone->zone_start_pfn = zone_start_pfn;
6476 zone->zone_start_pfn = 0;
6477 zone->spanned_pages = size;
6478 zone->present_pages = real_size;
6481 realtotalpages += real_size;
6484 pgdat->node_spanned_pages = totalpages;
6485 pgdat->node_present_pages = realtotalpages;
6486 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6490 #ifndef CONFIG_SPARSEMEM
6492 * Calculate the size of the zone->blockflags rounded to an unsigned long
6493 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6494 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6495 * round what is now in bits to nearest long in bits, then return it in
6498 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6500 unsigned long usemapsize;
6502 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6503 usemapsize = roundup(zonesize, pageblock_nr_pages);
6504 usemapsize = usemapsize >> pageblock_order;
6505 usemapsize *= NR_PAGEBLOCK_BITS;
6506 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6508 return usemapsize / 8;
6511 static void __ref setup_usemap(struct pglist_data *pgdat,
6513 unsigned long zone_start_pfn,
6514 unsigned long zonesize)
6516 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6517 zone->pageblock_flags = NULL;
6519 zone->pageblock_flags =
6520 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6522 if (!zone->pageblock_flags)
6523 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6524 usemapsize, zone->name, pgdat->node_id);
6528 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6529 unsigned long zone_start_pfn, unsigned long zonesize) {}
6530 #endif /* CONFIG_SPARSEMEM */
6532 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6534 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6535 void __init set_pageblock_order(void)
6539 /* Check that pageblock_nr_pages has not already been setup */
6540 if (pageblock_order)
6543 if (HPAGE_SHIFT > PAGE_SHIFT)
6544 order = HUGETLB_PAGE_ORDER;
6546 order = MAX_ORDER - 1;
6549 * Assume the largest contiguous order of interest is a huge page.
6550 * This value may be variable depending on boot parameters on IA64 and
6553 pageblock_order = order;
6555 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6558 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6559 * is unused as pageblock_order is set at compile-time. See
6560 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6563 void __init set_pageblock_order(void)
6567 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6569 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6570 unsigned long present_pages)
6572 unsigned long pages = spanned_pages;
6575 * Provide a more accurate estimation if there are holes within
6576 * the zone and SPARSEMEM is in use. If there are holes within the
6577 * zone, each populated memory region may cost us one or two extra
6578 * memmap pages due to alignment because memmap pages for each
6579 * populated regions may not be naturally aligned on page boundary.
6580 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6582 if (spanned_pages > present_pages + (present_pages >> 4) &&
6583 IS_ENABLED(CONFIG_SPARSEMEM))
6584 pages = present_pages;
6586 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6589 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6590 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6592 spin_lock_init(&pgdat->split_queue_lock);
6593 INIT_LIST_HEAD(&pgdat->split_queue);
6594 pgdat->split_queue_len = 0;
6597 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6600 #ifdef CONFIG_COMPACTION
6601 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6603 init_waitqueue_head(&pgdat->kcompactd_wait);
6606 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6609 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6611 pgdat_resize_init(pgdat);
6613 pgdat_init_split_queue(pgdat);
6614 pgdat_init_kcompactd(pgdat);
6616 init_waitqueue_head(&pgdat->kswapd_wait);
6617 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6619 pgdat_page_ext_init(pgdat);
6620 spin_lock_init(&pgdat->lru_lock);
6621 lruvec_init(node_lruvec(pgdat));
6624 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6625 unsigned long remaining_pages)
6627 atomic_long_set(&zone->managed_pages, remaining_pages);
6628 zone_set_nid(zone, nid);
6629 zone->name = zone_names[idx];
6630 zone->zone_pgdat = NODE_DATA(nid);
6631 spin_lock_init(&zone->lock);
6632 zone_seqlock_init(zone);
6633 zone_pcp_init(zone);
6637 * Set up the zone data structures
6638 * - init pgdat internals
6639 * - init all zones belonging to this node
6641 * NOTE: this function is only called during memory hotplug
6643 #ifdef CONFIG_MEMORY_HOTPLUG
6644 void __ref free_area_init_core_hotplug(int nid)
6647 pg_data_t *pgdat = NODE_DATA(nid);
6649 pgdat_init_internals(pgdat);
6650 for (z = 0; z < MAX_NR_ZONES; z++)
6651 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6656 * Set up the zone data structures:
6657 * - mark all pages reserved
6658 * - mark all memory queues empty
6659 * - clear the memory bitmaps
6661 * NOTE: pgdat should get zeroed by caller.
6662 * NOTE: this function is only called during early init.
6664 static void __init free_area_init_core(struct pglist_data *pgdat)
6667 int nid = pgdat->node_id;
6669 pgdat_init_internals(pgdat);
6670 pgdat->per_cpu_nodestats = &boot_nodestats;
6672 for (j = 0; j < MAX_NR_ZONES; j++) {
6673 struct zone *zone = pgdat->node_zones + j;
6674 unsigned long size, freesize, memmap_pages;
6675 unsigned long zone_start_pfn = zone->zone_start_pfn;
6677 size = zone->spanned_pages;
6678 freesize = zone->present_pages;
6681 * Adjust freesize so that it accounts for how much memory
6682 * is used by this zone for memmap. This affects the watermark
6683 * and per-cpu initialisations
6685 memmap_pages = calc_memmap_size(size, freesize);
6686 if (!is_highmem_idx(j)) {
6687 if (freesize >= memmap_pages) {
6688 freesize -= memmap_pages;
6691 " %s zone: %lu pages used for memmap\n",
6692 zone_names[j], memmap_pages);
6694 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6695 zone_names[j], memmap_pages, freesize);
6698 /* Account for reserved pages */
6699 if (j == 0 && freesize > dma_reserve) {
6700 freesize -= dma_reserve;
6701 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6702 zone_names[0], dma_reserve);
6705 if (!is_highmem_idx(j))
6706 nr_kernel_pages += freesize;
6707 /* Charge for highmem memmap if there are enough kernel pages */
6708 else if (nr_kernel_pages > memmap_pages * 2)
6709 nr_kernel_pages -= memmap_pages;
6710 nr_all_pages += freesize;
6713 * Set an approximate value for lowmem here, it will be adjusted
6714 * when the bootmem allocator frees pages into the buddy system.
6715 * And all highmem pages will be managed by the buddy system.
6717 zone_init_internals(zone, j, nid, freesize);
6722 set_pageblock_order();
6723 setup_usemap(pgdat, zone, zone_start_pfn, size);
6724 init_currently_empty_zone(zone, zone_start_pfn, size);
6725 memmap_init(size, nid, j, zone_start_pfn);
6729 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6730 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6732 unsigned long __maybe_unused start = 0;
6733 unsigned long __maybe_unused offset = 0;
6735 /* Skip empty nodes */
6736 if (!pgdat->node_spanned_pages)
6739 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6740 offset = pgdat->node_start_pfn - start;
6741 /* ia64 gets its own node_mem_map, before this, without bootmem */
6742 if (!pgdat->node_mem_map) {
6743 unsigned long size, end;
6747 * The zone's endpoints aren't required to be MAX_ORDER
6748 * aligned but the node_mem_map endpoints must be in order
6749 * for the buddy allocator to function correctly.
6751 end = pgdat_end_pfn(pgdat);
6752 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6753 size = (end - start) * sizeof(struct page);
6754 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6757 panic("Failed to allocate %ld bytes for node %d memory map\n",
6758 size, pgdat->node_id);
6759 pgdat->node_mem_map = map + offset;
6761 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6762 __func__, pgdat->node_id, (unsigned long)pgdat,
6763 (unsigned long)pgdat->node_mem_map);
6764 #ifndef CONFIG_NEED_MULTIPLE_NODES
6766 * With no DISCONTIG, the global mem_map is just set as node 0's
6768 if (pgdat == NODE_DATA(0)) {
6769 mem_map = NODE_DATA(0)->node_mem_map;
6770 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6771 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6773 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6778 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6779 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6781 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6782 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6784 pgdat->first_deferred_pfn = ULONG_MAX;
6787 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6790 void __init free_area_init_node(int nid, unsigned long *zones_size,
6791 unsigned long node_start_pfn,
6792 unsigned long *zholes_size)
6794 pg_data_t *pgdat = NODE_DATA(nid);
6795 unsigned long start_pfn = 0;
6796 unsigned long end_pfn = 0;
6798 /* pg_data_t should be reset to zero when it's allocated */
6799 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6801 pgdat->node_id = nid;
6802 pgdat->node_start_pfn = node_start_pfn;
6803 pgdat->per_cpu_nodestats = NULL;
6804 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6805 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6806 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6807 (u64)start_pfn << PAGE_SHIFT,
6808 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6810 start_pfn = node_start_pfn;
6812 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6813 zones_size, zholes_size);
6815 alloc_node_mem_map(pgdat);
6816 pgdat_set_deferred_range(pgdat);
6818 free_area_init_core(pgdat);
6821 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6823 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6826 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6831 for (pfn = spfn; pfn < epfn; pfn++) {
6832 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6833 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6834 + pageblock_nr_pages - 1;
6837 mm_zero_struct_page(pfn_to_page(pfn));
6845 * Only struct pages that are backed by physical memory are zeroed and
6846 * initialized by going through __init_single_page(). But, there are some
6847 * struct pages which are reserved in memblock allocator and their fields
6848 * may be accessed (for example page_to_pfn() on some configuration accesses
6849 * flags). We must explicitly zero those struct pages.
6851 * This function also addresses a similar issue where struct pages are left
6852 * uninitialized because the physical address range is not covered by
6853 * memblock.memory or memblock.reserved. That could happen when memblock
6854 * layout is manually configured via memmap=.
6856 void __init zero_resv_unavail(void)
6858 phys_addr_t start, end;
6860 phys_addr_t next = 0;
6863 * Loop through unavailable ranges not covered by memblock.memory.
6866 for_each_mem_range(i, &memblock.memory, NULL,
6867 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6869 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6872 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6875 * Struct pages that do not have backing memory. This could be because
6876 * firmware is using some of this memory, or for some other reasons.
6879 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6881 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6883 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6885 #if MAX_NUMNODES > 1
6887 * Figure out the number of possible node ids.
6889 void __init setup_nr_node_ids(void)
6891 unsigned int highest;
6893 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6894 nr_node_ids = highest + 1;
6899 * node_map_pfn_alignment - determine the maximum internode alignment
6901 * This function should be called after node map is populated and sorted.
6902 * It calculates the maximum power of two alignment which can distinguish
6905 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6906 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6907 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6908 * shifted, 1GiB is enough and this function will indicate so.
6910 * This is used to test whether pfn -> nid mapping of the chosen memory
6911 * model has fine enough granularity to avoid incorrect mapping for the
6912 * populated node map.
6914 * Return: the determined alignment in pfn's. 0 if there is no alignment
6915 * requirement (single node).
6917 unsigned long __init node_map_pfn_alignment(void)
6919 unsigned long accl_mask = 0, last_end = 0;
6920 unsigned long start, end, mask;
6921 int last_nid = NUMA_NO_NODE;
6924 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6925 if (!start || last_nid < 0 || last_nid == nid) {
6932 * Start with a mask granular enough to pin-point to the
6933 * start pfn and tick off bits one-by-one until it becomes
6934 * too coarse to separate the current node from the last.
6936 mask = ~((1 << __ffs(start)) - 1);
6937 while (mask && last_end <= (start & (mask << 1)))
6940 /* accumulate all internode masks */
6944 /* convert mask to number of pages */
6945 return ~accl_mask + 1;
6948 /* Find the lowest pfn for a node */
6949 static unsigned long __init find_min_pfn_for_node(int nid)
6951 unsigned long min_pfn = ULONG_MAX;
6952 unsigned long start_pfn;
6955 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6956 min_pfn = min(min_pfn, start_pfn);
6958 if (min_pfn == ULONG_MAX) {
6959 pr_warn("Could not find start_pfn for node %d\n", nid);
6967 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6969 * Return: the minimum PFN based on information provided via
6970 * memblock_set_node().
6972 unsigned long __init find_min_pfn_with_active_regions(void)
6974 return find_min_pfn_for_node(MAX_NUMNODES);
6978 * early_calculate_totalpages()
6979 * Sum pages in active regions for movable zone.
6980 * Populate N_MEMORY for calculating usable_nodes.
6982 static unsigned long __init early_calculate_totalpages(void)
6984 unsigned long totalpages = 0;
6985 unsigned long start_pfn, end_pfn;
6988 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6989 unsigned long pages = end_pfn - start_pfn;
6991 totalpages += pages;
6993 node_set_state(nid, N_MEMORY);
6999 * Find the PFN the Movable zone begins in each node. Kernel memory
7000 * is spread evenly between nodes as long as the nodes have enough
7001 * memory. When they don't, some nodes will have more kernelcore than
7004 static void __init find_zone_movable_pfns_for_nodes(void)
7007 unsigned long usable_startpfn;
7008 unsigned long kernelcore_node, kernelcore_remaining;
7009 /* save the state before borrow the nodemask */
7010 nodemask_t saved_node_state = node_states[N_MEMORY];
7011 unsigned long totalpages = early_calculate_totalpages();
7012 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7013 struct memblock_region *r;
7015 /* Need to find movable_zone earlier when movable_node is specified. */
7016 find_usable_zone_for_movable();
7019 * If movable_node is specified, ignore kernelcore and movablecore
7022 if (movable_node_is_enabled()) {
7023 for_each_memblock(memory, r) {
7024 if (!memblock_is_hotpluggable(r))
7029 usable_startpfn = PFN_DOWN(r->base);
7030 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7031 min(usable_startpfn, zone_movable_pfn[nid]) :
7039 * If kernelcore=mirror is specified, ignore movablecore option
7041 if (mirrored_kernelcore) {
7042 bool mem_below_4gb_not_mirrored = false;
7044 for_each_memblock(memory, r) {
7045 if (memblock_is_mirror(r))
7050 usable_startpfn = memblock_region_memory_base_pfn(r);
7052 if (usable_startpfn < 0x100000) {
7053 mem_below_4gb_not_mirrored = true;
7057 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7058 min(usable_startpfn, zone_movable_pfn[nid]) :
7062 if (mem_below_4gb_not_mirrored)
7063 pr_warn("This configuration results in unmirrored kernel memory.");
7069 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7070 * amount of necessary memory.
7072 if (required_kernelcore_percent)
7073 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7075 if (required_movablecore_percent)
7076 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7080 * If movablecore= was specified, calculate what size of
7081 * kernelcore that corresponds so that memory usable for
7082 * any allocation type is evenly spread. If both kernelcore
7083 * and movablecore are specified, then the value of kernelcore
7084 * will be used for required_kernelcore if it's greater than
7085 * what movablecore would have allowed.
7087 if (required_movablecore) {
7088 unsigned long corepages;
7091 * Round-up so that ZONE_MOVABLE is at least as large as what
7092 * was requested by the user
7094 required_movablecore =
7095 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7096 required_movablecore = min(totalpages, required_movablecore);
7097 corepages = totalpages - required_movablecore;
7099 required_kernelcore = max(required_kernelcore, corepages);
7103 * If kernelcore was not specified or kernelcore size is larger
7104 * than totalpages, there is no ZONE_MOVABLE.
7106 if (!required_kernelcore || required_kernelcore >= totalpages)
7109 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7110 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7113 /* Spread kernelcore memory as evenly as possible throughout nodes */
7114 kernelcore_node = required_kernelcore / usable_nodes;
7115 for_each_node_state(nid, N_MEMORY) {
7116 unsigned long start_pfn, end_pfn;
7119 * Recalculate kernelcore_node if the division per node
7120 * now exceeds what is necessary to satisfy the requested
7121 * amount of memory for the kernel
7123 if (required_kernelcore < kernelcore_node)
7124 kernelcore_node = required_kernelcore / usable_nodes;
7127 * As the map is walked, we track how much memory is usable
7128 * by the kernel using kernelcore_remaining. When it is
7129 * 0, the rest of the node is usable by ZONE_MOVABLE
7131 kernelcore_remaining = kernelcore_node;
7133 /* Go through each range of PFNs within this node */
7134 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7135 unsigned long size_pages;
7137 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7138 if (start_pfn >= end_pfn)
7141 /* Account for what is only usable for kernelcore */
7142 if (start_pfn < usable_startpfn) {
7143 unsigned long kernel_pages;
7144 kernel_pages = min(end_pfn, usable_startpfn)
7147 kernelcore_remaining -= min(kernel_pages,
7148 kernelcore_remaining);
7149 required_kernelcore -= min(kernel_pages,
7150 required_kernelcore);
7152 /* Continue if range is now fully accounted */
7153 if (end_pfn <= usable_startpfn) {
7156 * Push zone_movable_pfn to the end so
7157 * that if we have to rebalance
7158 * kernelcore across nodes, we will
7159 * not double account here
7161 zone_movable_pfn[nid] = end_pfn;
7164 start_pfn = usable_startpfn;
7168 * The usable PFN range for ZONE_MOVABLE is from
7169 * start_pfn->end_pfn. Calculate size_pages as the
7170 * number of pages used as kernelcore
7172 size_pages = end_pfn - start_pfn;
7173 if (size_pages > kernelcore_remaining)
7174 size_pages = kernelcore_remaining;
7175 zone_movable_pfn[nid] = start_pfn + size_pages;
7178 * Some kernelcore has been met, update counts and
7179 * break if the kernelcore for this node has been
7182 required_kernelcore -= min(required_kernelcore,
7184 kernelcore_remaining -= size_pages;
7185 if (!kernelcore_remaining)
7191 * If there is still required_kernelcore, we do another pass with one
7192 * less node in the count. This will push zone_movable_pfn[nid] further
7193 * along on the nodes that still have memory until kernelcore is
7197 if (usable_nodes && required_kernelcore > usable_nodes)
7201 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7202 for (nid = 0; nid < MAX_NUMNODES; nid++)
7203 zone_movable_pfn[nid] =
7204 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7207 /* restore the node_state */
7208 node_states[N_MEMORY] = saved_node_state;
7211 /* Any regular or high memory on that node ? */
7212 static void check_for_memory(pg_data_t *pgdat, int nid)
7214 enum zone_type zone_type;
7216 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7217 struct zone *zone = &pgdat->node_zones[zone_type];
7218 if (populated_zone(zone)) {
7219 if (IS_ENABLED(CONFIG_HIGHMEM))
7220 node_set_state(nid, N_HIGH_MEMORY);
7221 if (zone_type <= ZONE_NORMAL)
7222 node_set_state(nid, N_NORMAL_MEMORY);
7229 * free_area_init_nodes - Initialise all pg_data_t and zone data
7230 * @max_zone_pfn: an array of max PFNs for each zone
7232 * This will call free_area_init_node() for each active node in the system.
7233 * Using the page ranges provided by memblock_set_node(), the size of each
7234 * zone in each node and their holes is calculated. If the maximum PFN
7235 * between two adjacent zones match, it is assumed that the zone is empty.
7236 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7237 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7238 * starts where the previous one ended. For example, ZONE_DMA32 starts
7239 * at arch_max_dma_pfn.
7241 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7243 unsigned long start_pfn, end_pfn;
7246 /* Record where the zone boundaries are */
7247 memset(arch_zone_lowest_possible_pfn, 0,
7248 sizeof(arch_zone_lowest_possible_pfn));
7249 memset(arch_zone_highest_possible_pfn, 0,
7250 sizeof(arch_zone_highest_possible_pfn));
7252 start_pfn = find_min_pfn_with_active_regions();
7254 for (i = 0; i < MAX_NR_ZONES; i++) {
7255 if (i == ZONE_MOVABLE)
7258 end_pfn = max(max_zone_pfn[i], start_pfn);
7259 arch_zone_lowest_possible_pfn[i] = start_pfn;
7260 arch_zone_highest_possible_pfn[i] = end_pfn;
7262 start_pfn = end_pfn;
7265 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7266 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7267 find_zone_movable_pfns_for_nodes();
7269 /* Print out the zone ranges */
7270 pr_info("Zone ranges:\n");
7271 for (i = 0; i < MAX_NR_ZONES; i++) {
7272 if (i == ZONE_MOVABLE)
7274 pr_info(" %-8s ", zone_names[i]);
7275 if (arch_zone_lowest_possible_pfn[i] ==
7276 arch_zone_highest_possible_pfn[i])
7279 pr_cont("[mem %#018Lx-%#018Lx]\n",
7280 (u64)arch_zone_lowest_possible_pfn[i]
7282 ((u64)arch_zone_highest_possible_pfn[i]
7283 << PAGE_SHIFT) - 1);
7286 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7287 pr_info("Movable zone start for each node\n");
7288 for (i = 0; i < MAX_NUMNODES; i++) {
7289 if (zone_movable_pfn[i])
7290 pr_info(" Node %d: %#018Lx\n", i,
7291 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7294 /* Print out the early node map */
7295 pr_info("Early memory node ranges\n");
7296 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
7297 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7298 (u64)start_pfn << PAGE_SHIFT,
7299 ((u64)end_pfn << PAGE_SHIFT) - 1);
7301 /* Initialise every node */
7302 mminit_verify_pageflags_layout();
7303 setup_nr_node_ids();
7304 zero_resv_unavail();
7305 for_each_online_node(nid) {
7306 pg_data_t *pgdat = NODE_DATA(nid);
7307 free_area_init_node(nid, NULL,
7308 find_min_pfn_for_node(nid), NULL);
7310 /* Any memory on that node */
7311 if (pgdat->node_present_pages)
7312 node_set_state(nid, N_MEMORY);
7313 check_for_memory(pgdat, nid);
7317 static int __init cmdline_parse_core(char *p, unsigned long *core,
7318 unsigned long *percent)
7320 unsigned long long coremem;
7326 /* Value may be a percentage of total memory, otherwise bytes */
7327 coremem = simple_strtoull(p, &endptr, 0);
7328 if (*endptr == '%') {
7329 /* Paranoid check for percent values greater than 100 */
7330 WARN_ON(coremem > 100);
7334 coremem = memparse(p, &p);
7335 /* Paranoid check that UL is enough for the coremem value */
7336 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7338 *core = coremem >> PAGE_SHIFT;
7345 * kernelcore=size sets the amount of memory for use for allocations that
7346 * cannot be reclaimed or migrated.
7348 static int __init cmdline_parse_kernelcore(char *p)
7350 /* parse kernelcore=mirror */
7351 if (parse_option_str(p, "mirror")) {
7352 mirrored_kernelcore = true;
7356 return cmdline_parse_core(p, &required_kernelcore,
7357 &required_kernelcore_percent);
7361 * movablecore=size sets the amount of memory for use for allocations that
7362 * can be reclaimed or migrated.
7364 static int __init cmdline_parse_movablecore(char *p)
7366 return cmdline_parse_core(p, &required_movablecore,
7367 &required_movablecore_percent);
7370 early_param("kernelcore", cmdline_parse_kernelcore);
7371 early_param("movablecore", cmdline_parse_movablecore);
7373 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7375 void adjust_managed_page_count(struct page *page, long count)
7377 atomic_long_add(count, &page_zone(page)->managed_pages);
7378 totalram_pages_add(count);
7379 #ifdef CONFIG_HIGHMEM
7380 if (PageHighMem(page))
7381 totalhigh_pages_add(count);
7384 EXPORT_SYMBOL(adjust_managed_page_count);
7386 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7389 unsigned long pages = 0;
7391 start = (void *)PAGE_ALIGN((unsigned long)start);
7392 end = (void *)((unsigned long)end & PAGE_MASK);
7393 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7394 struct page *page = virt_to_page(pos);
7395 void *direct_map_addr;
7398 * 'direct_map_addr' might be different from 'pos'
7399 * because some architectures' virt_to_page()
7400 * work with aliases. Getting the direct map
7401 * address ensures that we get a _writeable_
7402 * alias for the memset().
7404 direct_map_addr = page_address(page);
7405 if ((unsigned int)poison <= 0xFF)
7406 memset(direct_map_addr, poison, PAGE_SIZE);
7408 free_reserved_page(page);
7412 pr_info("Freeing %s memory: %ldK\n",
7413 s, pages << (PAGE_SHIFT - 10));
7418 #ifdef CONFIG_HIGHMEM
7419 void free_highmem_page(struct page *page)
7421 __free_reserved_page(page);
7422 totalram_pages_inc();
7423 atomic_long_inc(&page_zone(page)->managed_pages);
7424 totalhigh_pages_inc();
7429 void __init mem_init_print_info(const char *str)
7431 unsigned long physpages, codesize, datasize, rosize, bss_size;
7432 unsigned long init_code_size, init_data_size;
7434 physpages = get_num_physpages();
7435 codesize = _etext - _stext;
7436 datasize = _edata - _sdata;
7437 rosize = __end_rodata - __start_rodata;
7438 bss_size = __bss_stop - __bss_start;
7439 init_data_size = __init_end - __init_begin;
7440 init_code_size = _einittext - _sinittext;
7443 * Detect special cases and adjust section sizes accordingly:
7444 * 1) .init.* may be embedded into .data sections
7445 * 2) .init.text.* may be out of [__init_begin, __init_end],
7446 * please refer to arch/tile/kernel/vmlinux.lds.S.
7447 * 3) .rodata.* may be embedded into .text or .data sections.
7449 #define adj_init_size(start, end, size, pos, adj) \
7451 if (start <= pos && pos < end && size > adj) \
7455 adj_init_size(__init_begin, __init_end, init_data_size,
7456 _sinittext, init_code_size);
7457 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7458 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7459 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7460 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7462 #undef adj_init_size
7464 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7465 #ifdef CONFIG_HIGHMEM
7469 nr_free_pages() << (PAGE_SHIFT - 10),
7470 physpages << (PAGE_SHIFT - 10),
7471 codesize >> 10, datasize >> 10, rosize >> 10,
7472 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7473 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7474 totalcma_pages << (PAGE_SHIFT - 10),
7475 #ifdef CONFIG_HIGHMEM
7476 totalhigh_pages() << (PAGE_SHIFT - 10),
7478 str ? ", " : "", str ? str : "");
7482 * set_dma_reserve - set the specified number of pages reserved in the first zone
7483 * @new_dma_reserve: The number of pages to mark reserved
7485 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7486 * In the DMA zone, a significant percentage may be consumed by kernel image
7487 * and other unfreeable allocations which can skew the watermarks badly. This
7488 * function may optionally be used to account for unfreeable pages in the
7489 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7490 * smaller per-cpu batchsize.
7492 void __init set_dma_reserve(unsigned long new_dma_reserve)
7494 dma_reserve = new_dma_reserve;
7497 void __init free_area_init(unsigned long *zones_size)
7499 zero_resv_unavail();
7500 free_area_init_node(0, zones_size,
7501 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7504 static int page_alloc_cpu_dead(unsigned int cpu)
7507 lru_add_drain_cpu(cpu);
7511 * Spill the event counters of the dead processor
7512 * into the current processors event counters.
7513 * This artificially elevates the count of the current
7516 vm_events_fold_cpu(cpu);
7519 * Zero the differential counters of the dead processor
7520 * so that the vm statistics are consistent.
7522 * This is only okay since the processor is dead and cannot
7523 * race with what we are doing.
7525 cpu_vm_stats_fold(cpu);
7529 void __init page_alloc_init(void)
7533 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7534 "mm/page_alloc:dead", NULL,
7535 page_alloc_cpu_dead);
7540 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7541 * or min_free_kbytes changes.
7543 static void calculate_totalreserve_pages(void)
7545 struct pglist_data *pgdat;
7546 unsigned long reserve_pages = 0;
7547 enum zone_type i, j;
7549 for_each_online_pgdat(pgdat) {
7551 pgdat->totalreserve_pages = 0;
7553 for (i = 0; i < MAX_NR_ZONES; i++) {
7554 struct zone *zone = pgdat->node_zones + i;
7556 unsigned long managed_pages = zone_managed_pages(zone);
7558 /* Find valid and maximum lowmem_reserve in the zone */
7559 for (j = i; j < MAX_NR_ZONES; j++) {
7560 if (zone->lowmem_reserve[j] > max)
7561 max = zone->lowmem_reserve[j];
7564 /* we treat the high watermark as reserved pages. */
7565 max += high_wmark_pages(zone);
7567 if (max > managed_pages)
7568 max = managed_pages;
7570 pgdat->totalreserve_pages += max;
7572 reserve_pages += max;
7575 totalreserve_pages = reserve_pages;
7579 * setup_per_zone_lowmem_reserve - called whenever
7580 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7581 * has a correct pages reserved value, so an adequate number of
7582 * pages are left in the zone after a successful __alloc_pages().
7584 static void setup_per_zone_lowmem_reserve(void)
7586 struct pglist_data *pgdat;
7587 enum zone_type j, idx;
7589 for_each_online_pgdat(pgdat) {
7590 for (j = 0; j < MAX_NR_ZONES; j++) {
7591 struct zone *zone = pgdat->node_zones + j;
7592 unsigned long managed_pages = zone_managed_pages(zone);
7594 zone->lowmem_reserve[j] = 0;
7598 struct zone *lower_zone;
7601 lower_zone = pgdat->node_zones + idx;
7603 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7604 sysctl_lowmem_reserve_ratio[idx] = 0;
7605 lower_zone->lowmem_reserve[j] = 0;
7607 lower_zone->lowmem_reserve[j] =
7608 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7610 managed_pages += zone_managed_pages(lower_zone);
7615 /* update totalreserve_pages */
7616 calculate_totalreserve_pages();
7619 static void __setup_per_zone_wmarks(void)
7621 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7622 unsigned long lowmem_pages = 0;
7624 unsigned long flags;
7626 /* Calculate total number of !ZONE_HIGHMEM pages */
7627 for_each_zone(zone) {
7628 if (!is_highmem(zone))
7629 lowmem_pages += zone_managed_pages(zone);
7632 for_each_zone(zone) {
7635 spin_lock_irqsave(&zone->lock, flags);
7636 tmp = (u64)pages_min * zone_managed_pages(zone);
7637 do_div(tmp, lowmem_pages);
7638 if (is_highmem(zone)) {
7640 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7641 * need highmem pages, so cap pages_min to a small
7644 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7645 * deltas control async page reclaim, and so should
7646 * not be capped for highmem.
7648 unsigned long min_pages;
7650 min_pages = zone_managed_pages(zone) / 1024;
7651 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7652 zone->_watermark[WMARK_MIN] = min_pages;
7655 * If it's a lowmem zone, reserve a number of pages
7656 * proportionate to the zone's size.
7658 zone->_watermark[WMARK_MIN] = tmp;
7662 * Set the kswapd watermarks distance according to the
7663 * scale factor in proportion to available memory, but
7664 * ensure a minimum size on small systems.
7666 tmp = max_t(u64, tmp >> 2,
7667 mult_frac(zone_managed_pages(zone),
7668 watermark_scale_factor, 10000));
7670 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7671 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7672 zone->watermark_boost = 0;
7674 spin_unlock_irqrestore(&zone->lock, flags);
7677 /* update totalreserve_pages */
7678 calculate_totalreserve_pages();
7682 * setup_per_zone_wmarks - called when min_free_kbytes changes
7683 * or when memory is hot-{added|removed}
7685 * Ensures that the watermark[min,low,high] values for each zone are set
7686 * correctly with respect to min_free_kbytes.
7688 void setup_per_zone_wmarks(void)
7690 static DEFINE_SPINLOCK(lock);
7693 __setup_per_zone_wmarks();
7698 * Initialise min_free_kbytes.
7700 * For small machines we want it small (128k min). For large machines
7701 * we want it large (64MB max). But it is not linear, because network
7702 * bandwidth does not increase linearly with machine size. We use
7704 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7705 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7721 int __meminit init_per_zone_wmark_min(void)
7723 unsigned long lowmem_kbytes;
7724 int new_min_free_kbytes;
7726 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7727 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7729 if (new_min_free_kbytes > user_min_free_kbytes) {
7730 min_free_kbytes = new_min_free_kbytes;
7731 if (min_free_kbytes < 128)
7732 min_free_kbytes = 128;
7733 if (min_free_kbytes > 65536)
7734 min_free_kbytes = 65536;
7736 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7737 new_min_free_kbytes, user_min_free_kbytes);
7739 setup_per_zone_wmarks();
7740 refresh_zone_stat_thresholds();
7741 setup_per_zone_lowmem_reserve();
7744 setup_min_unmapped_ratio();
7745 setup_min_slab_ratio();
7750 core_initcall(init_per_zone_wmark_min)
7753 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7754 * that we can call two helper functions whenever min_free_kbytes
7757 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7758 void __user *buffer, size_t *length, loff_t *ppos)
7762 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7767 user_min_free_kbytes = min_free_kbytes;
7768 setup_per_zone_wmarks();
7773 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7774 void __user *buffer, size_t *length, loff_t *ppos)
7778 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7785 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7786 void __user *buffer, size_t *length, loff_t *ppos)
7790 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7795 setup_per_zone_wmarks();
7801 static void setup_min_unmapped_ratio(void)
7806 for_each_online_pgdat(pgdat)
7807 pgdat->min_unmapped_pages = 0;
7810 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7811 sysctl_min_unmapped_ratio) / 100;
7815 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7816 void __user *buffer, size_t *length, loff_t *ppos)
7820 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7824 setup_min_unmapped_ratio();
7829 static void setup_min_slab_ratio(void)
7834 for_each_online_pgdat(pgdat)
7835 pgdat->min_slab_pages = 0;
7838 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7839 sysctl_min_slab_ratio) / 100;
7842 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7843 void __user *buffer, size_t *length, loff_t *ppos)
7847 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7851 setup_min_slab_ratio();
7858 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7859 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7860 * whenever sysctl_lowmem_reserve_ratio changes.
7862 * The reserve ratio obviously has absolutely no relation with the
7863 * minimum watermarks. The lowmem reserve ratio can only make sense
7864 * if in function of the boot time zone sizes.
7866 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7867 void __user *buffer, size_t *length, loff_t *ppos)
7869 proc_dointvec_minmax(table, write, buffer, length, ppos);
7870 setup_per_zone_lowmem_reserve();
7875 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7876 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7877 * pagelist can have before it gets flushed back to buddy allocator.
7879 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7880 void __user *buffer, size_t *length, loff_t *ppos)
7883 int old_percpu_pagelist_fraction;
7886 mutex_lock(&pcp_batch_high_lock);
7887 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7889 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7890 if (!write || ret < 0)
7893 /* Sanity checking to avoid pcp imbalance */
7894 if (percpu_pagelist_fraction &&
7895 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7896 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7902 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7905 for_each_populated_zone(zone) {
7908 for_each_possible_cpu(cpu)
7909 pageset_set_high_and_batch(zone,
7910 per_cpu_ptr(zone->pageset, cpu));
7913 mutex_unlock(&pcp_batch_high_lock);
7918 int hashdist = HASHDIST_DEFAULT;
7920 static int __init set_hashdist(char *str)
7924 hashdist = simple_strtoul(str, &str, 0);
7927 __setup("hashdist=", set_hashdist);
7930 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7932 * Returns the number of pages that arch has reserved but
7933 * is not known to alloc_large_system_hash().
7935 static unsigned long __init arch_reserved_kernel_pages(void)
7942 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7943 * machines. As memory size is increased the scale is also increased but at
7944 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7945 * quadruples the scale is increased by one, which means the size of hash table
7946 * only doubles, instead of quadrupling as well.
7947 * Because 32-bit systems cannot have large physical memory, where this scaling
7948 * makes sense, it is disabled on such platforms.
7950 #if __BITS_PER_LONG > 32
7951 #define ADAPT_SCALE_BASE (64ul << 30)
7952 #define ADAPT_SCALE_SHIFT 2
7953 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7957 * allocate a large system hash table from bootmem
7958 * - it is assumed that the hash table must contain an exact power-of-2
7959 * quantity of entries
7960 * - limit is the number of hash buckets, not the total allocation size
7962 void *__init alloc_large_system_hash(const char *tablename,
7963 unsigned long bucketsize,
7964 unsigned long numentries,
7967 unsigned int *_hash_shift,
7968 unsigned int *_hash_mask,
7969 unsigned long low_limit,
7970 unsigned long high_limit)
7972 unsigned long long max = high_limit;
7973 unsigned long log2qty, size;
7977 /* allow the kernel cmdline to have a say */
7979 /* round applicable memory size up to nearest megabyte */
7980 numentries = nr_kernel_pages;
7981 numentries -= arch_reserved_kernel_pages();
7983 /* It isn't necessary when PAGE_SIZE >= 1MB */
7984 if (PAGE_SHIFT < 20)
7985 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7987 #if __BITS_PER_LONG > 32
7989 unsigned long adapt;
7991 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7992 adapt <<= ADAPT_SCALE_SHIFT)
7997 /* limit to 1 bucket per 2^scale bytes of low memory */
7998 if (scale > PAGE_SHIFT)
7999 numentries >>= (scale - PAGE_SHIFT);
8001 numentries <<= (PAGE_SHIFT - scale);
8003 /* Make sure we've got at least a 0-order allocation.. */
8004 if (unlikely(flags & HASH_SMALL)) {
8005 /* Makes no sense without HASH_EARLY */
8006 WARN_ON(!(flags & HASH_EARLY));
8007 if (!(numentries >> *_hash_shift)) {
8008 numentries = 1UL << *_hash_shift;
8009 BUG_ON(!numentries);
8011 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8012 numentries = PAGE_SIZE / bucketsize;
8014 numentries = roundup_pow_of_two(numentries);
8016 /* limit allocation size to 1/16 total memory by default */
8018 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8019 do_div(max, bucketsize);
8021 max = min(max, 0x80000000ULL);
8023 if (numentries < low_limit)
8024 numentries = low_limit;
8025 if (numentries > max)
8028 log2qty = ilog2(numentries);
8030 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8032 size = bucketsize << log2qty;
8033 if (flags & HASH_EARLY) {
8034 if (flags & HASH_ZERO)
8035 table = memblock_alloc(size, SMP_CACHE_BYTES);
8037 table = memblock_alloc_raw(size,
8039 } else if (hashdist) {
8040 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8043 * If bucketsize is not a power-of-two, we may free
8044 * some pages at the end of hash table which
8045 * alloc_pages_exact() automatically does
8047 if (get_order(size) < MAX_ORDER) {
8048 table = alloc_pages_exact(size, gfp_flags);
8049 kmemleak_alloc(table, size, 1, gfp_flags);
8052 } while (!table && size > PAGE_SIZE && --log2qty);
8055 panic("Failed to allocate %s hash table\n", tablename);
8057 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
8058 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
8061 *_hash_shift = log2qty;
8063 *_hash_mask = (1 << log2qty) - 1;
8069 * This function checks whether pageblock includes unmovable pages or not.
8070 * If @count is not zero, it is okay to include less @count unmovable pages
8072 * PageLRU check without isolation or lru_lock could race so that
8073 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8074 * check without lock_page also may miss some movable non-lru pages at
8075 * race condition. So you can't expect this function should be exact.
8077 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
8078 int migratetype, int flags)
8080 unsigned long found;
8081 unsigned long iter = 0;
8082 unsigned long pfn = page_to_pfn(page);
8083 const char *reason = "unmovable page";
8086 * TODO we could make this much more efficient by not checking every
8087 * page in the range if we know all of them are in MOVABLE_ZONE and
8088 * that the movable zone guarantees that pages are migratable but
8089 * the later is not the case right now unfortunatelly. E.g. movablecore
8090 * can still lead to having bootmem allocations in zone_movable.
8093 if (is_migrate_cma_page(page)) {
8095 * CMA allocations (alloc_contig_range) really need to mark
8096 * isolate CMA pageblocks even when they are not movable in fact
8097 * so consider them movable here.
8099 if (is_migrate_cma(migratetype))
8102 reason = "CMA page";
8106 for (found = 0; iter < pageblock_nr_pages; iter++) {
8107 unsigned long check = pfn + iter;
8109 if (!pfn_valid_within(check))
8112 page = pfn_to_page(check);
8114 if (PageReserved(page))
8118 * If the zone is movable and we have ruled out all reserved
8119 * pages then it should be reasonably safe to assume the rest
8122 if (zone_idx(zone) == ZONE_MOVABLE)
8126 * Hugepages are not in LRU lists, but they're movable.
8127 * We need not scan over tail pages because we don't
8128 * handle each tail page individually in migration.
8130 if (PageHuge(page)) {
8131 struct page *head = compound_head(page);
8132 unsigned int skip_pages;
8134 if (!hugepage_migration_supported(page_hstate(head)))
8137 skip_pages = (1 << compound_order(head)) - (page - head);
8138 iter += skip_pages - 1;
8143 * We can't use page_count without pin a page
8144 * because another CPU can free compound page.
8145 * This check already skips compound tails of THP
8146 * because their page->_refcount is zero at all time.
8148 if (!page_ref_count(page)) {
8149 if (PageBuddy(page))
8150 iter += (1 << page_order(page)) - 1;
8155 * The HWPoisoned page may be not in buddy system, and
8156 * page_count() is not 0.
8158 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
8161 if (__PageMovable(page))
8167 * If there are RECLAIMABLE pages, we need to check
8168 * it. But now, memory offline itself doesn't call
8169 * shrink_node_slabs() and it still to be fixed.
8172 * If the page is not RAM, page_count()should be 0.
8173 * we don't need more check. This is an _used_ not-movable page.
8175 * The problematic thing here is PG_reserved pages. PG_reserved
8176 * is set to both of a memory hole page and a _used_ kernel
8184 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8185 if (flags & REPORT_FAILURE)
8186 dump_page(pfn_to_page(pfn + iter), reason);
8190 #ifdef CONFIG_CONTIG_ALLOC
8191 static unsigned long pfn_max_align_down(unsigned long pfn)
8193 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8194 pageblock_nr_pages) - 1);
8197 static unsigned long pfn_max_align_up(unsigned long pfn)
8199 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8200 pageblock_nr_pages));
8203 /* [start, end) must belong to a single zone. */
8204 static int __alloc_contig_migrate_range(struct compact_control *cc,
8205 unsigned long start, unsigned long end)
8207 /* This function is based on compact_zone() from compaction.c. */
8208 unsigned long nr_reclaimed;
8209 unsigned long pfn = start;
8210 unsigned int tries = 0;
8215 while (pfn < end || !list_empty(&cc->migratepages)) {
8216 if (fatal_signal_pending(current)) {
8221 if (list_empty(&cc->migratepages)) {
8222 cc->nr_migratepages = 0;
8223 pfn = isolate_migratepages_range(cc, pfn, end);
8229 } else if (++tries == 5) {
8230 ret = ret < 0 ? ret : -EBUSY;
8234 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8236 cc->nr_migratepages -= nr_reclaimed;
8238 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8239 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8242 putback_movable_pages(&cc->migratepages);
8249 * alloc_contig_range() -- tries to allocate given range of pages
8250 * @start: start PFN to allocate
8251 * @end: one-past-the-last PFN to allocate
8252 * @migratetype: migratetype of the underlaying pageblocks (either
8253 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8254 * in range must have the same migratetype and it must
8255 * be either of the two.
8256 * @gfp_mask: GFP mask to use during compaction
8258 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8259 * aligned. The PFN range must belong to a single zone.
8261 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8262 * pageblocks in the range. Once isolated, the pageblocks should not
8263 * be modified by others.
8265 * Return: zero on success or negative error code. On success all
8266 * pages which PFN is in [start, end) are allocated for the caller and
8267 * need to be freed with free_contig_range().
8269 int alloc_contig_range(unsigned long start, unsigned long end,
8270 unsigned migratetype, gfp_t gfp_mask)
8272 unsigned long outer_start, outer_end;
8276 struct compact_control cc = {
8277 .nr_migratepages = 0,
8279 .zone = page_zone(pfn_to_page(start)),
8280 .mode = MIGRATE_SYNC,
8281 .ignore_skip_hint = true,
8282 .no_set_skip_hint = true,
8283 .gfp_mask = current_gfp_context(gfp_mask),
8285 INIT_LIST_HEAD(&cc.migratepages);
8288 * What we do here is we mark all pageblocks in range as
8289 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8290 * have different sizes, and due to the way page allocator
8291 * work, we align the range to biggest of the two pages so
8292 * that page allocator won't try to merge buddies from
8293 * different pageblocks and change MIGRATE_ISOLATE to some
8294 * other migration type.
8296 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8297 * migrate the pages from an unaligned range (ie. pages that
8298 * we are interested in). This will put all the pages in
8299 * range back to page allocator as MIGRATE_ISOLATE.
8301 * When this is done, we take the pages in range from page
8302 * allocator removing them from the buddy system. This way
8303 * page allocator will never consider using them.
8305 * This lets us mark the pageblocks back as
8306 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8307 * aligned range but not in the unaligned, original range are
8308 * put back to page allocator so that buddy can use them.
8311 ret = start_isolate_page_range(pfn_max_align_down(start),
8312 pfn_max_align_up(end), migratetype, 0);
8317 * In case of -EBUSY, we'd like to know which page causes problem.
8318 * So, just fall through. test_pages_isolated() has a tracepoint
8319 * which will report the busy page.
8321 * It is possible that busy pages could become available before
8322 * the call to test_pages_isolated, and the range will actually be
8323 * allocated. So, if we fall through be sure to clear ret so that
8324 * -EBUSY is not accidentally used or returned to caller.
8326 ret = __alloc_contig_migrate_range(&cc, start, end);
8327 if (ret && ret != -EBUSY)
8332 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8333 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8334 * more, all pages in [start, end) are free in page allocator.
8335 * What we are going to do is to allocate all pages from
8336 * [start, end) (that is remove them from page allocator).
8338 * The only problem is that pages at the beginning and at the
8339 * end of interesting range may be not aligned with pages that
8340 * page allocator holds, ie. they can be part of higher order
8341 * pages. Because of this, we reserve the bigger range and
8342 * once this is done free the pages we are not interested in.
8344 * We don't have to hold zone->lock here because the pages are
8345 * isolated thus they won't get removed from buddy.
8348 lru_add_drain_all();
8351 outer_start = start;
8352 while (!PageBuddy(pfn_to_page(outer_start))) {
8353 if (++order >= MAX_ORDER) {
8354 outer_start = start;
8357 outer_start &= ~0UL << order;
8360 if (outer_start != start) {
8361 order = page_order(pfn_to_page(outer_start));
8364 * outer_start page could be small order buddy page and
8365 * it doesn't include start page. Adjust outer_start
8366 * in this case to report failed page properly
8367 * on tracepoint in test_pages_isolated()
8369 if (outer_start + (1UL << order) <= start)
8370 outer_start = start;
8373 /* Make sure the range is really isolated. */
8374 if (test_pages_isolated(outer_start, end, false)) {
8375 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8376 __func__, outer_start, end);
8381 /* Grab isolated pages from freelists. */
8382 outer_end = isolate_freepages_range(&cc, outer_start, end);
8388 /* Free head and tail (if any) */
8389 if (start != outer_start)
8390 free_contig_range(outer_start, start - outer_start);
8391 if (end != outer_end)
8392 free_contig_range(end, outer_end - end);
8395 undo_isolate_page_range(pfn_max_align_down(start),
8396 pfn_max_align_up(end), migratetype);
8399 #endif /* CONFIG_CONTIG_ALLOC */
8401 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8403 unsigned int count = 0;
8405 for (; nr_pages--; pfn++) {
8406 struct page *page = pfn_to_page(pfn);
8408 count += page_count(page) != 1;
8411 WARN(count != 0, "%d pages are still in use!\n", count);
8414 #ifdef CONFIG_MEMORY_HOTPLUG
8416 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8417 * page high values need to be recalulated.
8419 void __meminit zone_pcp_update(struct zone *zone)
8422 mutex_lock(&pcp_batch_high_lock);
8423 for_each_possible_cpu(cpu)
8424 pageset_set_high_and_batch(zone,
8425 per_cpu_ptr(zone->pageset, cpu));
8426 mutex_unlock(&pcp_batch_high_lock);
8430 void zone_pcp_reset(struct zone *zone)
8432 unsigned long flags;
8434 struct per_cpu_pageset *pset;
8436 /* avoid races with drain_pages() */
8437 local_irq_save(flags);
8438 if (zone->pageset != &boot_pageset) {
8439 for_each_online_cpu(cpu) {
8440 pset = per_cpu_ptr(zone->pageset, cpu);
8441 drain_zonestat(zone, pset);
8443 free_percpu(zone->pageset);
8444 zone->pageset = &boot_pageset;
8446 local_irq_restore(flags);
8449 #ifdef CONFIG_MEMORY_HOTREMOVE
8451 * All pages in the range must be in a single zone and isolated
8452 * before calling this.
8455 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8459 unsigned int order, i;
8461 unsigned long flags;
8462 unsigned long offlined_pages = 0;
8464 /* find the first valid pfn */
8465 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8469 return offlined_pages;
8471 offline_mem_sections(pfn, end_pfn);
8472 zone = page_zone(pfn_to_page(pfn));
8473 spin_lock_irqsave(&zone->lock, flags);
8475 while (pfn < end_pfn) {
8476 if (!pfn_valid(pfn)) {
8480 page = pfn_to_page(pfn);
8482 * The HWPoisoned page may be not in buddy system, and
8483 * page_count() is not 0.
8485 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8487 SetPageReserved(page);
8492 BUG_ON(page_count(page));
8493 BUG_ON(!PageBuddy(page));
8494 order = page_order(page);
8495 offlined_pages += 1 << order;
8496 #ifdef CONFIG_DEBUG_VM
8497 pr_info("remove from free list %lx %d %lx\n",
8498 pfn, 1 << order, end_pfn);
8500 list_del(&page->lru);
8501 rmv_page_order(page);
8502 zone->free_area[order].nr_free--;
8503 for (i = 0; i < (1 << order); i++)
8504 SetPageReserved((page+i));
8505 pfn += (1 << order);
8507 spin_unlock_irqrestore(&zone->lock, flags);
8509 return offlined_pages;
8513 bool is_free_buddy_page(struct page *page)
8515 struct zone *zone = page_zone(page);
8516 unsigned long pfn = page_to_pfn(page);
8517 unsigned long flags;
8520 spin_lock_irqsave(&zone->lock, flags);
8521 for (order = 0; order < MAX_ORDER; order++) {
8522 struct page *page_head = page - (pfn & ((1 << order) - 1));
8524 if (PageBuddy(page_head) && page_order(page_head) >= order)
8527 spin_unlock_irqrestore(&zone->lock, flags);
8529 return order < MAX_ORDER;
8532 #ifdef CONFIG_MEMORY_FAILURE
8534 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8535 * test is performed under the zone lock to prevent a race against page
8538 bool set_hwpoison_free_buddy_page(struct page *page)
8540 struct zone *zone = page_zone(page);
8541 unsigned long pfn = page_to_pfn(page);
8542 unsigned long flags;
8544 bool hwpoisoned = false;
8546 spin_lock_irqsave(&zone->lock, flags);
8547 for (order = 0; order < MAX_ORDER; order++) {
8548 struct page *page_head = page - (pfn & ((1 << order) - 1));
8550 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8551 if (!TestSetPageHWPoison(page))
8556 spin_unlock_irqrestore(&zone->lock, flags);