1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
78 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
79 static DEFINE_MUTEX(pcp_batch_high_lock);
80 #define MIN_PERCPU_PAGELIST_FRACTION (8)
82 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
83 DEFINE_PER_CPU(int, numa_node);
84 EXPORT_PER_CPU_SYMBOL(numa_node);
87 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
89 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
91 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
92 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
93 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
94 * defined in <linux/topology.h>.
96 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
97 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
100 /* work_structs for global per-cpu drains */
103 struct work_struct work;
105 DEFINE_MUTEX(pcpu_drain_mutex);
106 DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
108 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
109 volatile unsigned long latent_entropy __latent_entropy;
110 EXPORT_SYMBOL(latent_entropy);
114 * Array of node states.
116 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
117 [N_POSSIBLE] = NODE_MASK_ALL,
118 [N_ONLINE] = { { [0] = 1UL } },
120 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
121 #ifdef CONFIG_HIGHMEM
122 [N_HIGH_MEMORY] = { { [0] = 1UL } },
124 [N_MEMORY] = { { [0] = 1UL } },
125 [N_CPU] = { { [0] = 1UL } },
128 EXPORT_SYMBOL(node_states);
130 atomic_long_t _totalram_pages __read_mostly;
131 EXPORT_SYMBOL(_totalram_pages);
132 unsigned long totalreserve_pages __read_mostly;
133 unsigned long totalcma_pages __read_mostly;
135 int percpu_pagelist_fraction;
136 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
137 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
138 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
140 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
142 EXPORT_SYMBOL(init_on_alloc);
144 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
145 DEFINE_STATIC_KEY_TRUE(init_on_free);
147 DEFINE_STATIC_KEY_FALSE(init_on_free);
149 EXPORT_SYMBOL(init_on_free);
151 static int __init early_init_on_alloc(char *buf)
158 ret = kstrtobool(buf, &bool_result);
159 if (bool_result && page_poisoning_enabled())
160 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
162 static_branch_enable(&init_on_alloc);
164 static_branch_disable(&init_on_alloc);
167 early_param("init_on_alloc", early_init_on_alloc);
169 static int __init early_init_on_free(char *buf)
176 ret = kstrtobool(buf, &bool_result);
177 if (bool_result && page_poisoning_enabled())
178 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
180 static_branch_enable(&init_on_free);
182 static_branch_disable(&init_on_free);
185 early_param("init_on_free", early_init_on_free);
188 * A cached value of the page's pageblock's migratetype, used when the page is
189 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
190 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
191 * Also the migratetype set in the page does not necessarily match the pcplist
192 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
193 * other index - this ensures that it will be put on the correct CMA freelist.
195 static inline int get_pcppage_migratetype(struct page *page)
200 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
202 page->index = migratetype;
205 #ifdef CONFIG_PM_SLEEP
207 * The following functions are used by the suspend/hibernate code to temporarily
208 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
209 * while devices are suspended. To avoid races with the suspend/hibernate code,
210 * they should always be called with system_transition_mutex held
211 * (gfp_allowed_mask also should only be modified with system_transition_mutex
212 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
213 * with that modification).
216 static gfp_t saved_gfp_mask;
218 void pm_restore_gfp_mask(void)
220 WARN_ON(!mutex_is_locked(&system_transition_mutex));
221 if (saved_gfp_mask) {
222 gfp_allowed_mask = saved_gfp_mask;
227 void pm_restrict_gfp_mask(void)
229 WARN_ON(!mutex_is_locked(&system_transition_mutex));
230 WARN_ON(saved_gfp_mask);
231 saved_gfp_mask = gfp_allowed_mask;
232 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
235 bool pm_suspended_storage(void)
237 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
241 #endif /* CONFIG_PM_SLEEP */
243 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
244 unsigned int pageblock_order __read_mostly;
247 static void __free_pages_ok(struct page *page, unsigned int order);
250 * results with 256, 32 in the lowmem_reserve sysctl:
251 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
252 * 1G machine -> (16M dma, 784M normal, 224M high)
253 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
254 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
255 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
257 * TBD: should special case ZONE_DMA32 machines here - in those we normally
258 * don't need any ZONE_NORMAL reservation
260 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
261 #ifdef CONFIG_ZONE_DMA
264 #ifdef CONFIG_ZONE_DMA32
268 #ifdef CONFIG_HIGHMEM
274 static char * const zone_names[MAX_NR_ZONES] = {
275 #ifdef CONFIG_ZONE_DMA
278 #ifdef CONFIG_ZONE_DMA32
282 #ifdef CONFIG_HIGHMEM
286 #ifdef CONFIG_ZONE_DEVICE
291 const char * const migratetype_names[MIGRATE_TYPES] = {
299 #ifdef CONFIG_MEMORY_ISOLATION
304 compound_page_dtor * const compound_page_dtors[] = {
307 #ifdef CONFIG_HUGETLB_PAGE
310 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
315 int min_free_kbytes = 1024;
316 int user_min_free_kbytes = -1;
317 #ifdef CONFIG_DISCONTIGMEM
319 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
320 * are not on separate NUMA nodes. Functionally this works but with
321 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
322 * quite small. By default, do not boost watermarks on discontigmem as in
323 * many cases very high-order allocations like THP are likely to be
324 * unsupported and the premature reclaim offsets the advantage of long-term
325 * fragmentation avoidance.
327 int watermark_boost_factor __read_mostly;
329 int watermark_boost_factor __read_mostly = 15000;
331 int watermark_scale_factor = 10;
333 static unsigned long nr_kernel_pages __initdata;
334 static unsigned long nr_all_pages __initdata;
335 static unsigned long dma_reserve __initdata;
337 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
338 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
339 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
340 static unsigned long required_kernelcore __initdata;
341 static unsigned long required_kernelcore_percent __initdata;
342 static unsigned long required_movablecore __initdata;
343 static unsigned long required_movablecore_percent __initdata;
344 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
345 static bool mirrored_kernelcore __meminitdata;
347 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
349 EXPORT_SYMBOL(movable_zone);
350 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
353 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
354 unsigned int nr_online_nodes __read_mostly = 1;
355 EXPORT_SYMBOL(nr_node_ids);
356 EXPORT_SYMBOL(nr_online_nodes);
359 int page_group_by_mobility_disabled __read_mostly;
361 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
363 * During boot we initialize deferred pages on-demand, as needed, but once
364 * page_alloc_init_late() has finished, the deferred pages are all initialized,
365 * and we can permanently disable that path.
367 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
370 * Calling kasan_free_pages() only after deferred memory initialization
371 * has completed. Poisoning pages during deferred memory init will greatly
372 * lengthen the process and cause problem in large memory systems as the
373 * deferred pages initialization is done with interrupt disabled.
375 * Assuming that there will be no reference to those newly initialized
376 * pages before they are ever allocated, this should have no effect on
377 * KASAN memory tracking as the poison will be properly inserted at page
378 * allocation time. The only corner case is when pages are allocated by
379 * on-demand allocation and then freed again before the deferred pages
380 * initialization is done, but this is not likely to happen.
382 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
384 if (!static_branch_unlikely(&deferred_pages))
385 kasan_free_pages(page, order);
388 /* Returns true if the struct page for the pfn is uninitialised */
389 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
391 int nid = early_pfn_to_nid(pfn);
393 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
400 * Returns true when the remaining initialisation should be deferred until
401 * later in the boot cycle when it can be parallelised.
403 static bool __meminit
404 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
406 static unsigned long prev_end_pfn, nr_initialised;
409 * prev_end_pfn static that contains the end of previous zone
410 * No need to protect because called very early in boot before smp_init.
412 if (prev_end_pfn != end_pfn) {
413 prev_end_pfn = end_pfn;
417 /* Always populate low zones for address-constrained allocations */
418 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
422 * We start only with one section of pages, more pages are added as
423 * needed until the rest of deferred pages are initialized.
426 if ((nr_initialised > PAGES_PER_SECTION) &&
427 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
428 NODE_DATA(nid)->first_deferred_pfn = pfn;
434 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
436 static inline bool early_page_uninitialised(unsigned long pfn)
441 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
447 /* Return a pointer to the bitmap storing bits affecting a block of pages */
448 static inline unsigned long *get_pageblock_bitmap(struct page *page,
451 #ifdef CONFIG_SPARSEMEM
452 return section_to_usemap(__pfn_to_section(pfn));
454 return page_zone(page)->pageblock_flags;
455 #endif /* CONFIG_SPARSEMEM */
458 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
460 #ifdef CONFIG_SPARSEMEM
461 pfn &= (PAGES_PER_SECTION-1);
462 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
464 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
465 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
466 #endif /* CONFIG_SPARSEMEM */
470 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
471 * @page: The page within the block of interest
472 * @pfn: The target page frame number
473 * @end_bitidx: The last bit of interest to retrieve
474 * @mask: mask of bits that the caller is interested in
476 * Return: pageblock_bits flags
478 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
480 unsigned long end_bitidx,
483 unsigned long *bitmap;
484 unsigned long bitidx, word_bitidx;
487 bitmap = get_pageblock_bitmap(page, pfn);
488 bitidx = pfn_to_bitidx(page, pfn);
489 word_bitidx = bitidx / BITS_PER_LONG;
490 bitidx &= (BITS_PER_LONG-1);
492 word = bitmap[word_bitidx];
493 bitidx += end_bitidx;
494 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
497 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
498 unsigned long end_bitidx,
501 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
504 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
506 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
510 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
511 * @page: The page within the block of interest
512 * @flags: The flags to set
513 * @pfn: The target page frame number
514 * @end_bitidx: The last bit of interest
515 * @mask: mask of bits that the caller is interested in
517 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
519 unsigned long end_bitidx,
522 unsigned long *bitmap;
523 unsigned long bitidx, word_bitidx;
524 unsigned long old_word, word;
526 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
527 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
529 bitmap = get_pageblock_bitmap(page, pfn);
530 bitidx = pfn_to_bitidx(page, pfn);
531 word_bitidx = bitidx / BITS_PER_LONG;
532 bitidx &= (BITS_PER_LONG-1);
534 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
536 bitidx += end_bitidx;
537 mask <<= (BITS_PER_LONG - bitidx - 1);
538 flags <<= (BITS_PER_LONG - bitidx - 1);
540 word = READ_ONCE(bitmap[word_bitidx]);
542 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
543 if (word == old_word)
549 void set_pageblock_migratetype(struct page *page, int migratetype)
551 if (unlikely(page_group_by_mobility_disabled &&
552 migratetype < MIGRATE_PCPTYPES))
553 migratetype = MIGRATE_UNMOVABLE;
555 set_pageblock_flags_group(page, (unsigned long)migratetype,
556 PB_migrate, PB_migrate_end);
559 #ifdef CONFIG_DEBUG_VM
560 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
564 unsigned long pfn = page_to_pfn(page);
565 unsigned long sp, start_pfn;
568 seq = zone_span_seqbegin(zone);
569 start_pfn = zone->zone_start_pfn;
570 sp = zone->spanned_pages;
571 if (!zone_spans_pfn(zone, pfn))
573 } while (zone_span_seqretry(zone, seq));
576 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
577 pfn, zone_to_nid(zone), zone->name,
578 start_pfn, start_pfn + sp);
583 static int page_is_consistent(struct zone *zone, struct page *page)
585 if (!pfn_valid_within(page_to_pfn(page)))
587 if (zone != page_zone(page))
593 * Temporary debugging check for pages not lying within a given zone.
595 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
597 if (page_outside_zone_boundaries(zone, page))
599 if (!page_is_consistent(zone, page))
605 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
611 static void bad_page(struct page *page, const char *reason,
612 unsigned long bad_flags)
614 static unsigned long resume;
615 static unsigned long nr_shown;
616 static unsigned long nr_unshown;
619 * Allow a burst of 60 reports, then keep quiet for that minute;
620 * or allow a steady drip of one report per second.
622 if (nr_shown == 60) {
623 if (time_before(jiffies, resume)) {
629 "BUG: Bad page state: %lu messages suppressed\n",
636 resume = jiffies + 60 * HZ;
638 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
639 current->comm, page_to_pfn(page));
640 __dump_page(page, reason);
641 bad_flags &= page->flags;
643 pr_alert("bad because of flags: %#lx(%pGp)\n",
644 bad_flags, &bad_flags);
645 dump_page_owner(page);
650 /* Leave bad fields for debug, except PageBuddy could make trouble */
651 page_mapcount_reset(page); /* remove PageBuddy */
652 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
656 * Higher-order pages are called "compound pages". They are structured thusly:
658 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
660 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
661 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
663 * The first tail page's ->compound_dtor holds the offset in array of compound
664 * page destructors. See compound_page_dtors.
666 * The first tail page's ->compound_order holds the order of allocation.
667 * This usage means that zero-order pages may not be compound.
670 void free_compound_page(struct page *page)
672 mem_cgroup_uncharge(page);
673 __free_pages_ok(page, compound_order(page));
676 void prep_compound_page(struct page *page, unsigned int order)
679 int nr_pages = 1 << order;
681 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
682 set_compound_order(page, order);
684 for (i = 1; i < nr_pages; i++) {
685 struct page *p = page + i;
686 set_page_count(p, 0);
687 p->mapping = TAIL_MAPPING;
688 set_compound_head(p, page);
690 atomic_set(compound_mapcount_ptr(page), -1);
693 #ifdef CONFIG_DEBUG_PAGEALLOC
694 unsigned int _debug_guardpage_minorder;
696 bool _debug_pagealloc_enabled_early __read_mostly
697 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
698 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
699 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
700 EXPORT_SYMBOL(_debug_pagealloc_enabled);
702 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
704 static int __init early_debug_pagealloc(char *buf)
706 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
708 early_param("debug_pagealloc", early_debug_pagealloc);
710 void init_debug_pagealloc(void)
712 if (!debug_pagealloc_enabled())
715 static_branch_enable(&_debug_pagealloc_enabled);
717 if (!debug_guardpage_minorder())
720 static_branch_enable(&_debug_guardpage_enabled);
723 static int __init debug_guardpage_minorder_setup(char *buf)
727 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
728 pr_err("Bad debug_guardpage_minorder value\n");
731 _debug_guardpage_minorder = res;
732 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
735 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
737 static inline bool set_page_guard(struct zone *zone, struct page *page,
738 unsigned int order, int migratetype)
740 if (!debug_guardpage_enabled())
743 if (order >= debug_guardpage_minorder())
746 __SetPageGuard(page);
747 INIT_LIST_HEAD(&page->lru);
748 set_page_private(page, order);
749 /* Guard pages are not available for any usage */
750 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
755 static inline void clear_page_guard(struct zone *zone, struct page *page,
756 unsigned int order, int migratetype)
758 if (!debug_guardpage_enabled())
761 __ClearPageGuard(page);
763 set_page_private(page, 0);
764 if (!is_migrate_isolate(migratetype))
765 __mod_zone_freepage_state(zone, (1 << order), migratetype);
768 static inline bool set_page_guard(struct zone *zone, struct page *page,
769 unsigned int order, int migratetype) { return false; }
770 static inline void clear_page_guard(struct zone *zone, struct page *page,
771 unsigned int order, int migratetype) {}
774 static inline void set_page_order(struct page *page, unsigned int order)
776 set_page_private(page, order);
777 __SetPageBuddy(page);
781 * This function checks whether a page is free && is the buddy
782 * we can coalesce a page and its buddy if
783 * (a) the buddy is not in a hole (check before calling!) &&
784 * (b) the buddy is in the buddy system &&
785 * (c) a page and its buddy have the same order &&
786 * (d) a page and its buddy are in the same zone.
788 * For recording whether a page is in the buddy system, we set PageBuddy.
789 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
791 * For recording page's order, we use page_private(page).
793 static inline int page_is_buddy(struct page *page, struct page *buddy,
796 if (page_is_guard(buddy) && page_order(buddy) == order) {
797 if (page_zone_id(page) != page_zone_id(buddy))
800 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
805 if (PageBuddy(buddy) && page_order(buddy) == order) {
807 * zone check is done late to avoid uselessly
808 * calculating zone/node ids for pages that could
811 if (page_zone_id(page) != page_zone_id(buddy))
814 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
821 #ifdef CONFIG_COMPACTION
822 static inline struct capture_control *task_capc(struct zone *zone)
824 struct capture_control *capc = current->capture_control;
827 !(current->flags & PF_KTHREAD) &&
829 capc->cc->zone == zone &&
830 capc->cc->direct_compaction ? capc : NULL;
834 compaction_capture(struct capture_control *capc, struct page *page,
835 int order, int migratetype)
837 if (!capc || order != capc->cc->order)
840 /* Do not accidentally pollute CMA or isolated regions*/
841 if (is_migrate_cma(migratetype) ||
842 is_migrate_isolate(migratetype))
846 * Do not let lower order allocations polluate a movable pageblock.
847 * This might let an unmovable request use a reclaimable pageblock
848 * and vice-versa but no more than normal fallback logic which can
849 * have trouble finding a high-order free page.
851 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
859 static inline struct capture_control *task_capc(struct zone *zone)
865 compaction_capture(struct capture_control *capc, struct page *page,
866 int order, int migratetype)
870 #endif /* CONFIG_COMPACTION */
873 * Freeing function for a buddy system allocator.
875 * The concept of a buddy system is to maintain direct-mapped table
876 * (containing bit values) for memory blocks of various "orders".
877 * The bottom level table contains the map for the smallest allocatable
878 * units of memory (here, pages), and each level above it describes
879 * pairs of units from the levels below, hence, "buddies".
880 * At a high level, all that happens here is marking the table entry
881 * at the bottom level available, and propagating the changes upward
882 * as necessary, plus some accounting needed to play nicely with other
883 * parts of the VM system.
884 * At each level, we keep a list of pages, which are heads of continuous
885 * free pages of length of (1 << order) and marked with PageBuddy.
886 * Page's order is recorded in page_private(page) field.
887 * So when we are allocating or freeing one, we can derive the state of the
888 * other. That is, if we allocate a small block, and both were
889 * free, the remainder of the region must be split into blocks.
890 * If a block is freed, and its buddy is also free, then this
891 * triggers coalescing into a block of larger size.
896 static inline void __free_one_page(struct page *page,
898 struct zone *zone, unsigned int order,
901 unsigned long combined_pfn;
902 unsigned long uninitialized_var(buddy_pfn);
904 unsigned int max_order;
905 struct capture_control *capc = task_capc(zone);
907 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
909 VM_BUG_ON(!zone_is_initialized(zone));
910 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
912 VM_BUG_ON(migratetype == -1);
913 if (likely(!is_migrate_isolate(migratetype)))
914 __mod_zone_freepage_state(zone, 1 << order, migratetype);
916 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
917 VM_BUG_ON_PAGE(bad_range(zone, page), page);
920 while (order < max_order - 1) {
921 if (compaction_capture(capc, page, order, migratetype)) {
922 __mod_zone_freepage_state(zone, -(1 << order),
926 buddy_pfn = __find_buddy_pfn(pfn, order);
927 buddy = page + (buddy_pfn - pfn);
929 if (!pfn_valid_within(buddy_pfn))
931 if (!page_is_buddy(page, buddy, order))
934 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
935 * merge with it and move up one order.
937 if (page_is_guard(buddy))
938 clear_page_guard(zone, buddy, order, migratetype);
940 del_page_from_free_area(buddy, &zone->free_area[order]);
941 combined_pfn = buddy_pfn & pfn;
942 page = page + (combined_pfn - pfn);
946 if (max_order < MAX_ORDER) {
947 /* If we are here, it means order is >= pageblock_order.
948 * We want to prevent merge between freepages on isolate
949 * pageblock and normal pageblock. Without this, pageblock
950 * isolation could cause incorrect freepage or CMA accounting.
952 * We don't want to hit this code for the more frequent
955 if (unlikely(has_isolate_pageblock(zone))) {
958 buddy_pfn = __find_buddy_pfn(pfn, order);
959 buddy = page + (buddy_pfn - pfn);
960 buddy_mt = get_pageblock_migratetype(buddy);
962 if (migratetype != buddy_mt
963 && (is_migrate_isolate(migratetype) ||
964 is_migrate_isolate(buddy_mt)))
968 goto continue_merging;
972 set_page_order(page, order);
975 * If this is not the largest possible page, check if the buddy
976 * of the next-highest order is free. If it is, it's possible
977 * that pages are being freed that will coalesce soon. In case,
978 * that is happening, add the free page to the tail of the list
979 * so it's less likely to be used soon and more likely to be merged
980 * as a higher order page
982 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)
983 && !is_shuffle_order(order)) {
984 struct page *higher_page, *higher_buddy;
985 combined_pfn = buddy_pfn & pfn;
986 higher_page = page + (combined_pfn - pfn);
987 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
988 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
989 if (pfn_valid_within(buddy_pfn) &&
990 page_is_buddy(higher_page, higher_buddy, order + 1)) {
991 add_to_free_area_tail(page, &zone->free_area[order],
997 if (is_shuffle_order(order))
998 add_to_free_area_random(page, &zone->free_area[order],
1001 add_to_free_area(page, &zone->free_area[order], migratetype);
1006 * A bad page could be due to a number of fields. Instead of multiple branches,
1007 * try and check multiple fields with one check. The caller must do a detailed
1008 * check if necessary.
1010 static inline bool page_expected_state(struct page *page,
1011 unsigned long check_flags)
1013 if (unlikely(atomic_read(&page->_mapcount) != -1))
1016 if (unlikely((unsigned long)page->mapping |
1017 page_ref_count(page) |
1019 (unsigned long)page->mem_cgroup |
1021 (page->flags & check_flags)))
1027 static void free_pages_check_bad(struct page *page)
1029 const char *bad_reason;
1030 unsigned long bad_flags;
1035 if (unlikely(atomic_read(&page->_mapcount) != -1))
1036 bad_reason = "nonzero mapcount";
1037 if (unlikely(page->mapping != NULL))
1038 bad_reason = "non-NULL mapping";
1039 if (unlikely(page_ref_count(page) != 0))
1040 bad_reason = "nonzero _refcount";
1041 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1042 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1043 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1046 if (unlikely(page->mem_cgroup))
1047 bad_reason = "page still charged to cgroup";
1049 bad_page(page, bad_reason, bad_flags);
1052 static inline int free_pages_check(struct page *page)
1054 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1057 /* Something has gone sideways, find it */
1058 free_pages_check_bad(page);
1062 static int free_tail_pages_check(struct page *head_page, struct page *page)
1067 * We rely page->lru.next never has bit 0 set, unless the page
1068 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1070 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1072 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1076 switch (page - head_page) {
1078 /* the first tail page: ->mapping may be compound_mapcount() */
1079 if (unlikely(compound_mapcount(page))) {
1080 bad_page(page, "nonzero compound_mapcount", 0);
1086 * the second tail page: ->mapping is
1087 * deferred_list.next -- ignore value.
1091 if (page->mapping != TAIL_MAPPING) {
1092 bad_page(page, "corrupted mapping in tail page", 0);
1097 if (unlikely(!PageTail(page))) {
1098 bad_page(page, "PageTail not set", 0);
1101 if (unlikely(compound_head(page) != head_page)) {
1102 bad_page(page, "compound_head not consistent", 0);
1107 page->mapping = NULL;
1108 clear_compound_head(page);
1112 static void kernel_init_free_pages(struct page *page, int numpages)
1116 for (i = 0; i < numpages; i++)
1117 clear_highpage(page + i);
1120 static __always_inline bool free_pages_prepare(struct page *page,
1121 unsigned int order, bool check_free)
1125 VM_BUG_ON_PAGE(PageTail(page), page);
1127 trace_mm_page_free(page, order);
1130 * Check tail pages before head page information is cleared to
1131 * avoid checking PageCompound for order-0 pages.
1133 if (unlikely(order)) {
1134 bool compound = PageCompound(page);
1137 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1140 ClearPageDoubleMap(page);
1141 for (i = 1; i < (1 << order); i++) {
1143 bad += free_tail_pages_check(page, page + i);
1144 if (unlikely(free_pages_check(page + i))) {
1148 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1151 if (PageMappingFlags(page))
1152 page->mapping = NULL;
1153 if (memcg_kmem_enabled() && PageKmemcg(page))
1154 __memcg_kmem_uncharge(page, order);
1156 bad += free_pages_check(page);
1160 page_cpupid_reset_last(page);
1161 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1162 reset_page_owner(page, order);
1164 if (!PageHighMem(page)) {
1165 debug_check_no_locks_freed(page_address(page),
1166 PAGE_SIZE << order);
1167 debug_check_no_obj_freed(page_address(page),
1168 PAGE_SIZE << order);
1170 if (want_init_on_free())
1171 kernel_init_free_pages(page, 1 << order);
1173 kernel_poison_pages(page, 1 << order, 0);
1175 * arch_free_page() can make the page's contents inaccessible. s390
1176 * does this. So nothing which can access the page's contents should
1177 * happen after this.
1179 arch_free_page(page, order);
1181 if (debug_pagealloc_enabled_static())
1182 kernel_map_pages(page, 1 << order, 0);
1184 kasan_free_nondeferred_pages(page, order);
1189 #ifdef CONFIG_DEBUG_VM
1191 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1192 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1193 * moved from pcp lists to free lists.
1195 static bool free_pcp_prepare(struct page *page)
1197 return free_pages_prepare(page, 0, true);
1200 static bool bulkfree_pcp_prepare(struct page *page)
1202 if (debug_pagealloc_enabled_static())
1203 return free_pages_check(page);
1209 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1210 * moving from pcp lists to free list in order to reduce overhead. With
1211 * debug_pagealloc enabled, they are checked also immediately when being freed
1214 static bool free_pcp_prepare(struct page *page)
1216 if (debug_pagealloc_enabled_static())
1217 return free_pages_prepare(page, 0, true);
1219 return free_pages_prepare(page, 0, false);
1222 static bool bulkfree_pcp_prepare(struct page *page)
1224 return free_pages_check(page);
1226 #endif /* CONFIG_DEBUG_VM */
1228 static inline void prefetch_buddy(struct page *page)
1230 unsigned long pfn = page_to_pfn(page);
1231 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1232 struct page *buddy = page + (buddy_pfn - pfn);
1238 * Frees a number of pages from the PCP lists
1239 * Assumes all pages on list are in same zone, and of same order.
1240 * count is the number of pages to free.
1242 * If the zone was previously in an "all pages pinned" state then look to
1243 * see if this freeing clears that state.
1245 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1246 * pinned" detection logic.
1248 static void free_pcppages_bulk(struct zone *zone, int count,
1249 struct per_cpu_pages *pcp)
1251 int migratetype = 0;
1253 int prefetch_nr = 0;
1254 bool isolated_pageblocks;
1255 struct page *page, *tmp;
1259 struct list_head *list;
1262 * Remove pages from lists in a round-robin fashion. A
1263 * batch_free count is maintained that is incremented when an
1264 * empty list is encountered. This is so more pages are freed
1265 * off fuller lists instead of spinning excessively around empty
1270 if (++migratetype == MIGRATE_PCPTYPES)
1272 list = &pcp->lists[migratetype];
1273 } while (list_empty(list));
1275 /* This is the only non-empty list. Free them all. */
1276 if (batch_free == MIGRATE_PCPTYPES)
1280 page = list_last_entry(list, struct page, lru);
1281 /* must delete to avoid corrupting pcp list */
1282 list_del(&page->lru);
1285 if (bulkfree_pcp_prepare(page))
1288 list_add_tail(&page->lru, &head);
1291 * We are going to put the page back to the global
1292 * pool, prefetch its buddy to speed up later access
1293 * under zone->lock. It is believed the overhead of
1294 * an additional test and calculating buddy_pfn here
1295 * can be offset by reduced memory latency later. To
1296 * avoid excessive prefetching due to large count, only
1297 * prefetch buddy for the first pcp->batch nr of pages.
1299 if (prefetch_nr++ < pcp->batch)
1300 prefetch_buddy(page);
1301 } while (--count && --batch_free && !list_empty(list));
1304 spin_lock(&zone->lock);
1305 isolated_pageblocks = has_isolate_pageblock(zone);
1308 * Use safe version since after __free_one_page(),
1309 * page->lru.next will not point to original list.
1311 list_for_each_entry_safe(page, tmp, &head, lru) {
1312 int mt = get_pcppage_migratetype(page);
1313 /* MIGRATE_ISOLATE page should not go to pcplists */
1314 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1315 /* Pageblock could have been isolated meanwhile */
1316 if (unlikely(isolated_pageblocks))
1317 mt = get_pageblock_migratetype(page);
1319 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1320 trace_mm_page_pcpu_drain(page, 0, mt);
1322 spin_unlock(&zone->lock);
1325 static void free_one_page(struct zone *zone,
1326 struct page *page, unsigned long pfn,
1330 spin_lock(&zone->lock);
1331 if (unlikely(has_isolate_pageblock(zone) ||
1332 is_migrate_isolate(migratetype))) {
1333 migratetype = get_pfnblock_migratetype(page, pfn);
1335 __free_one_page(page, pfn, zone, order, migratetype);
1336 spin_unlock(&zone->lock);
1339 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1340 unsigned long zone, int nid)
1342 mm_zero_struct_page(page);
1343 set_page_links(page, zone, nid, pfn);
1344 init_page_count(page);
1345 page_mapcount_reset(page);
1346 page_cpupid_reset_last(page);
1347 page_kasan_tag_reset(page);
1349 INIT_LIST_HEAD(&page->lru);
1350 #ifdef WANT_PAGE_VIRTUAL
1351 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1352 if (!is_highmem_idx(zone))
1353 set_page_address(page, __va(pfn << PAGE_SHIFT));
1357 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1358 static void __meminit init_reserved_page(unsigned long pfn)
1363 if (!early_page_uninitialised(pfn))
1366 nid = early_pfn_to_nid(pfn);
1367 pgdat = NODE_DATA(nid);
1369 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1370 struct zone *zone = &pgdat->node_zones[zid];
1372 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1375 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1378 static inline void init_reserved_page(unsigned long pfn)
1381 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1384 * Initialised pages do not have PageReserved set. This function is
1385 * called for each range allocated by the bootmem allocator and
1386 * marks the pages PageReserved. The remaining valid pages are later
1387 * sent to the buddy page allocator.
1389 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1391 unsigned long start_pfn = PFN_DOWN(start);
1392 unsigned long end_pfn = PFN_UP(end);
1394 for (; start_pfn < end_pfn; start_pfn++) {
1395 if (pfn_valid(start_pfn)) {
1396 struct page *page = pfn_to_page(start_pfn);
1398 init_reserved_page(start_pfn);
1400 /* Avoid false-positive PageTail() */
1401 INIT_LIST_HEAD(&page->lru);
1404 * no need for atomic set_bit because the struct
1405 * page is not visible yet so nobody should
1408 __SetPageReserved(page);
1413 static void __free_pages_ok(struct page *page, unsigned int order)
1415 unsigned long flags;
1417 unsigned long pfn = page_to_pfn(page);
1419 if (!free_pages_prepare(page, order, true))
1422 migratetype = get_pfnblock_migratetype(page, pfn);
1423 local_irq_save(flags);
1424 __count_vm_events(PGFREE, 1 << order);
1425 free_one_page(page_zone(page), page, pfn, order, migratetype);
1426 local_irq_restore(flags);
1429 void __free_pages_core(struct page *page, unsigned int order)
1431 unsigned int nr_pages = 1 << order;
1432 struct page *p = page;
1436 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1438 __ClearPageReserved(p);
1439 set_page_count(p, 0);
1441 __ClearPageReserved(p);
1442 set_page_count(p, 0);
1444 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1445 set_page_refcounted(page);
1446 __free_pages(page, order);
1449 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1450 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1452 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1454 int __meminit early_pfn_to_nid(unsigned long pfn)
1456 static DEFINE_SPINLOCK(early_pfn_lock);
1459 spin_lock(&early_pfn_lock);
1460 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1462 nid = first_online_node;
1463 spin_unlock(&early_pfn_lock);
1469 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1470 /* Only safe to use early in boot when initialisation is single-threaded */
1471 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1475 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1476 if (nid >= 0 && nid != node)
1482 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1489 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1492 if (early_page_uninitialised(pfn))
1494 __free_pages_core(page, order);
1498 * Check that the whole (or subset of) a pageblock given by the interval of
1499 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1500 * with the migration of free compaction scanner. The scanners then need to
1501 * use only pfn_valid_within() check for arches that allow holes within
1504 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1506 * It's possible on some configurations to have a setup like node0 node1 node0
1507 * i.e. it's possible that all pages within a zones range of pages do not
1508 * belong to a single zone. We assume that a border between node0 and node1
1509 * can occur within a single pageblock, but not a node0 node1 node0
1510 * interleaving within a single pageblock. It is therefore sufficient to check
1511 * the first and last page of a pageblock and avoid checking each individual
1512 * page in a pageblock.
1514 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1515 unsigned long end_pfn, struct zone *zone)
1517 struct page *start_page;
1518 struct page *end_page;
1520 /* end_pfn is one past the range we are checking */
1523 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1526 start_page = pfn_to_online_page(start_pfn);
1530 if (page_zone(start_page) != zone)
1533 end_page = pfn_to_page(end_pfn);
1535 /* This gives a shorter code than deriving page_zone(end_page) */
1536 if (page_zone_id(start_page) != page_zone_id(end_page))
1542 void set_zone_contiguous(struct zone *zone)
1544 unsigned long block_start_pfn = zone->zone_start_pfn;
1545 unsigned long block_end_pfn;
1547 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1548 for (; block_start_pfn < zone_end_pfn(zone);
1549 block_start_pfn = block_end_pfn,
1550 block_end_pfn += pageblock_nr_pages) {
1552 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1554 if (!__pageblock_pfn_to_page(block_start_pfn,
1555 block_end_pfn, zone))
1559 /* We confirm that there is no hole */
1560 zone->contiguous = true;
1563 void clear_zone_contiguous(struct zone *zone)
1565 zone->contiguous = false;
1568 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1569 static void __init deferred_free_range(unsigned long pfn,
1570 unsigned long nr_pages)
1578 page = pfn_to_page(pfn);
1580 /* Free a large naturally-aligned chunk if possible */
1581 if (nr_pages == pageblock_nr_pages &&
1582 (pfn & (pageblock_nr_pages - 1)) == 0) {
1583 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1584 __free_pages_core(page, pageblock_order);
1588 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1589 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1590 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1591 __free_pages_core(page, 0);
1595 /* Completion tracking for deferred_init_memmap() threads */
1596 static atomic_t pgdat_init_n_undone __initdata;
1597 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1599 static inline void __init pgdat_init_report_one_done(void)
1601 if (atomic_dec_and_test(&pgdat_init_n_undone))
1602 complete(&pgdat_init_all_done_comp);
1606 * Returns true if page needs to be initialized or freed to buddy allocator.
1608 * First we check if pfn is valid on architectures where it is possible to have
1609 * holes within pageblock_nr_pages. On systems where it is not possible, this
1610 * function is optimized out.
1612 * Then, we check if a current large page is valid by only checking the validity
1615 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1617 if (!pfn_valid_within(pfn))
1619 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1625 * Free pages to buddy allocator. Try to free aligned pages in
1626 * pageblock_nr_pages sizes.
1628 static void __init deferred_free_pages(unsigned long pfn,
1629 unsigned long end_pfn)
1631 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1632 unsigned long nr_free = 0;
1634 for (; pfn < end_pfn; pfn++) {
1635 if (!deferred_pfn_valid(pfn)) {
1636 deferred_free_range(pfn - nr_free, nr_free);
1638 } else if (!(pfn & nr_pgmask)) {
1639 deferred_free_range(pfn - nr_free, nr_free);
1641 touch_nmi_watchdog();
1646 /* Free the last block of pages to allocator */
1647 deferred_free_range(pfn - nr_free, nr_free);
1651 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1652 * by performing it only once every pageblock_nr_pages.
1653 * Return number of pages initialized.
1655 static unsigned long __init deferred_init_pages(struct zone *zone,
1657 unsigned long end_pfn)
1659 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1660 int nid = zone_to_nid(zone);
1661 unsigned long nr_pages = 0;
1662 int zid = zone_idx(zone);
1663 struct page *page = NULL;
1665 for (; pfn < end_pfn; pfn++) {
1666 if (!deferred_pfn_valid(pfn)) {
1669 } else if (!page || !(pfn & nr_pgmask)) {
1670 page = pfn_to_page(pfn);
1671 touch_nmi_watchdog();
1675 __init_single_page(page, pfn, zid, nid);
1682 * This function is meant to pre-load the iterator for the zone init.
1683 * Specifically it walks through the ranges until we are caught up to the
1684 * first_init_pfn value and exits there. If we never encounter the value we
1685 * return false indicating there are no valid ranges left.
1688 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1689 unsigned long *spfn, unsigned long *epfn,
1690 unsigned long first_init_pfn)
1695 * Start out by walking through the ranges in this zone that have
1696 * already been initialized. We don't need to do anything with them
1697 * so we just need to flush them out of the system.
1699 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1700 if (*epfn <= first_init_pfn)
1702 if (*spfn < first_init_pfn)
1703 *spfn = first_init_pfn;
1712 * Initialize and free pages. We do it in two loops: first we initialize
1713 * struct page, then free to buddy allocator, because while we are
1714 * freeing pages we can access pages that are ahead (computing buddy
1715 * page in __free_one_page()).
1717 * In order to try and keep some memory in the cache we have the loop
1718 * broken along max page order boundaries. This way we will not cause
1719 * any issues with the buddy page computation.
1721 static unsigned long __init
1722 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1723 unsigned long *end_pfn)
1725 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1726 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1727 unsigned long nr_pages = 0;
1730 /* First we loop through and initialize the page values */
1731 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1734 if (mo_pfn <= *start_pfn)
1737 t = min(mo_pfn, *end_pfn);
1738 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1740 if (mo_pfn < *end_pfn) {
1741 *start_pfn = mo_pfn;
1746 /* Reset values and now loop through freeing pages as needed */
1749 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1755 t = min(mo_pfn, epfn);
1756 deferred_free_pages(spfn, t);
1765 /* Initialise remaining memory on a node */
1766 static int __init deferred_init_memmap(void *data)
1768 pg_data_t *pgdat = data;
1769 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1770 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1771 unsigned long first_init_pfn, flags;
1772 unsigned long start = jiffies;
1777 /* Bind memory initialisation thread to a local node if possible */
1778 if (!cpumask_empty(cpumask))
1779 set_cpus_allowed_ptr(current, cpumask);
1781 pgdat_resize_lock(pgdat, &flags);
1782 first_init_pfn = pgdat->first_deferred_pfn;
1783 if (first_init_pfn == ULONG_MAX) {
1784 pgdat_resize_unlock(pgdat, &flags);
1785 pgdat_init_report_one_done();
1789 /* Sanity check boundaries */
1790 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1791 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1792 pgdat->first_deferred_pfn = ULONG_MAX;
1794 /* Only the highest zone is deferred so find it */
1795 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1796 zone = pgdat->node_zones + zid;
1797 if (first_init_pfn < zone_end_pfn(zone))
1801 /* If the zone is empty somebody else may have cleared out the zone */
1802 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1807 * Initialize and free pages in MAX_ORDER sized increments so
1808 * that we can avoid introducing any issues with the buddy
1812 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1814 pgdat_resize_unlock(pgdat, &flags);
1816 /* Sanity check that the next zone really is unpopulated */
1817 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1819 pr_info("node %d initialised, %lu pages in %ums\n",
1820 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1822 pgdat_init_report_one_done();
1827 * If this zone has deferred pages, try to grow it by initializing enough
1828 * deferred pages to satisfy the allocation specified by order, rounded up to
1829 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1830 * of SECTION_SIZE bytes by initializing struct pages in increments of
1831 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1833 * Return true when zone was grown, otherwise return false. We return true even
1834 * when we grow less than requested, to let the caller decide if there are
1835 * enough pages to satisfy the allocation.
1837 * Note: We use noinline because this function is needed only during boot, and
1838 * it is called from a __ref function _deferred_grow_zone. This way we are
1839 * making sure that it is not inlined into permanent text section.
1841 static noinline bool __init
1842 deferred_grow_zone(struct zone *zone, unsigned int order)
1844 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1845 pg_data_t *pgdat = zone->zone_pgdat;
1846 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1847 unsigned long spfn, epfn, flags;
1848 unsigned long nr_pages = 0;
1851 /* Only the last zone may have deferred pages */
1852 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1855 pgdat_resize_lock(pgdat, &flags);
1858 * If deferred pages have been initialized while we were waiting for
1859 * the lock, return true, as the zone was grown. The caller will retry
1860 * this zone. We won't return to this function since the caller also
1861 * has this static branch.
1863 if (!static_branch_unlikely(&deferred_pages)) {
1864 pgdat_resize_unlock(pgdat, &flags);
1869 * If someone grew this zone while we were waiting for spinlock, return
1870 * true, as there might be enough pages already.
1872 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1873 pgdat_resize_unlock(pgdat, &flags);
1877 /* If the zone is empty somebody else may have cleared out the zone */
1878 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1879 first_deferred_pfn)) {
1880 pgdat->first_deferred_pfn = ULONG_MAX;
1881 pgdat_resize_unlock(pgdat, &flags);
1882 /* Retry only once. */
1883 return first_deferred_pfn != ULONG_MAX;
1887 * Initialize and free pages in MAX_ORDER sized increments so
1888 * that we can avoid introducing any issues with the buddy
1891 while (spfn < epfn) {
1892 /* update our first deferred PFN for this section */
1893 first_deferred_pfn = spfn;
1895 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1897 /* We should only stop along section boundaries */
1898 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1901 /* If our quota has been met we can stop here */
1902 if (nr_pages >= nr_pages_needed)
1906 pgdat->first_deferred_pfn = spfn;
1907 pgdat_resize_unlock(pgdat, &flags);
1909 return nr_pages > 0;
1913 * deferred_grow_zone() is __init, but it is called from
1914 * get_page_from_freelist() during early boot until deferred_pages permanently
1915 * disables this call. This is why we have refdata wrapper to avoid warning,
1916 * and to ensure that the function body gets unloaded.
1919 _deferred_grow_zone(struct zone *zone, unsigned int order)
1921 return deferred_grow_zone(zone, order);
1924 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1926 void __init page_alloc_init_late(void)
1931 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1933 /* There will be num_node_state(N_MEMORY) threads */
1934 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1935 for_each_node_state(nid, N_MEMORY) {
1936 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1939 /* Block until all are initialised */
1940 wait_for_completion(&pgdat_init_all_done_comp);
1943 * The number of managed pages has changed due to the initialisation
1944 * so the pcpu batch and high limits needs to be updated or the limits
1945 * will be artificially small.
1947 for_each_populated_zone(zone)
1948 zone_pcp_update(zone);
1951 * We initialized the rest of the deferred pages. Permanently disable
1952 * on-demand struct page initialization.
1954 static_branch_disable(&deferred_pages);
1956 /* Reinit limits that are based on free pages after the kernel is up */
1957 files_maxfiles_init();
1960 /* Discard memblock private memory */
1963 for_each_node_state(nid, N_MEMORY)
1964 shuffle_free_memory(NODE_DATA(nid));
1966 for_each_populated_zone(zone)
1967 set_zone_contiguous(zone);
1971 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1972 void __init init_cma_reserved_pageblock(struct page *page)
1974 unsigned i = pageblock_nr_pages;
1975 struct page *p = page;
1978 __ClearPageReserved(p);
1979 set_page_count(p, 0);
1982 set_pageblock_migratetype(page, MIGRATE_CMA);
1984 if (pageblock_order >= MAX_ORDER) {
1985 i = pageblock_nr_pages;
1988 set_page_refcounted(p);
1989 __free_pages(p, MAX_ORDER - 1);
1990 p += MAX_ORDER_NR_PAGES;
1991 } while (i -= MAX_ORDER_NR_PAGES);
1993 set_page_refcounted(page);
1994 __free_pages(page, pageblock_order);
1997 adjust_managed_page_count(page, pageblock_nr_pages);
2002 * The order of subdivision here is critical for the IO subsystem.
2003 * Please do not alter this order without good reasons and regression
2004 * testing. Specifically, as large blocks of memory are subdivided,
2005 * the order in which smaller blocks are delivered depends on the order
2006 * they're subdivided in this function. This is the primary factor
2007 * influencing the order in which pages are delivered to the IO
2008 * subsystem according to empirical testing, and this is also justified
2009 * by considering the behavior of a buddy system containing a single
2010 * large block of memory acted on by a series of small allocations.
2011 * This behavior is a critical factor in sglist merging's success.
2015 static inline void expand(struct zone *zone, struct page *page,
2016 int low, int high, struct free_area *area,
2019 unsigned long size = 1 << high;
2021 while (high > low) {
2025 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2028 * Mark as guard pages (or page), that will allow to
2029 * merge back to allocator when buddy will be freed.
2030 * Corresponding page table entries will not be touched,
2031 * pages will stay not present in virtual address space
2033 if (set_page_guard(zone, &page[size], high, migratetype))
2036 add_to_free_area(&page[size], area, migratetype);
2037 set_page_order(&page[size], high);
2041 static void check_new_page_bad(struct page *page)
2043 const char *bad_reason = NULL;
2044 unsigned long bad_flags = 0;
2046 if (unlikely(atomic_read(&page->_mapcount) != -1))
2047 bad_reason = "nonzero mapcount";
2048 if (unlikely(page->mapping != NULL))
2049 bad_reason = "non-NULL mapping";
2050 if (unlikely(page_ref_count(page) != 0))
2051 bad_reason = "nonzero _refcount";
2052 if (unlikely(page->flags & __PG_HWPOISON)) {
2053 bad_reason = "HWPoisoned (hardware-corrupted)";
2054 bad_flags = __PG_HWPOISON;
2055 /* Don't complain about hwpoisoned pages */
2056 page_mapcount_reset(page); /* remove PageBuddy */
2059 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2060 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2061 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2064 if (unlikely(page->mem_cgroup))
2065 bad_reason = "page still charged to cgroup";
2067 bad_page(page, bad_reason, bad_flags);
2071 * This page is about to be returned from the page allocator
2073 static inline int check_new_page(struct page *page)
2075 if (likely(page_expected_state(page,
2076 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2079 check_new_page_bad(page);
2083 static inline bool free_pages_prezeroed(void)
2085 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2086 page_poisoning_enabled()) || want_init_on_free();
2089 #ifdef CONFIG_DEBUG_VM
2091 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2092 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2093 * also checked when pcp lists are refilled from the free lists.
2095 static inline bool check_pcp_refill(struct page *page)
2097 if (debug_pagealloc_enabled_static())
2098 return check_new_page(page);
2103 static inline bool check_new_pcp(struct page *page)
2105 return check_new_page(page);
2109 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2110 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2111 * enabled, they are also checked when being allocated from the pcp lists.
2113 static inline bool check_pcp_refill(struct page *page)
2115 return check_new_page(page);
2117 static inline bool check_new_pcp(struct page *page)
2119 if (debug_pagealloc_enabled_static())
2120 return check_new_page(page);
2124 #endif /* CONFIG_DEBUG_VM */
2126 static bool check_new_pages(struct page *page, unsigned int order)
2129 for (i = 0; i < (1 << order); i++) {
2130 struct page *p = page + i;
2132 if (unlikely(check_new_page(p)))
2139 inline void post_alloc_hook(struct page *page, unsigned int order,
2142 set_page_private(page, 0);
2143 set_page_refcounted(page);
2145 arch_alloc_page(page, order);
2146 if (debug_pagealloc_enabled_static())
2147 kernel_map_pages(page, 1 << order, 1);
2148 kasan_alloc_pages(page, order);
2149 kernel_poison_pages(page, 1 << order, 1);
2150 set_page_owner(page, order, gfp_flags);
2153 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2154 unsigned int alloc_flags)
2156 post_alloc_hook(page, order, gfp_flags);
2158 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2159 kernel_init_free_pages(page, 1 << order);
2161 if (order && (gfp_flags & __GFP_COMP))
2162 prep_compound_page(page, order);
2165 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2166 * allocate the page. The expectation is that the caller is taking
2167 * steps that will free more memory. The caller should avoid the page
2168 * being used for !PFMEMALLOC purposes.
2170 if (alloc_flags & ALLOC_NO_WATERMARKS)
2171 set_page_pfmemalloc(page);
2173 clear_page_pfmemalloc(page);
2177 * Go through the free lists for the given migratetype and remove
2178 * the smallest available page from the freelists
2180 static __always_inline
2181 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2184 unsigned int current_order;
2185 struct free_area *area;
2188 /* Find a page of the appropriate size in the preferred list */
2189 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2190 area = &(zone->free_area[current_order]);
2191 page = get_page_from_free_area(area, migratetype);
2194 del_page_from_free_area(page, area);
2195 expand(zone, page, order, current_order, area, migratetype);
2196 set_pcppage_migratetype(page, migratetype);
2205 * This array describes the order lists are fallen back to when
2206 * the free lists for the desirable migrate type are depleted
2208 static int fallbacks[MIGRATE_TYPES][4] = {
2209 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2210 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2211 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2213 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2215 #ifdef CONFIG_MEMORY_ISOLATION
2216 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2221 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2224 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2227 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2228 unsigned int order) { return NULL; }
2232 * Move the free pages in a range to the free lists of the requested type.
2233 * Note that start_page and end_pages are not aligned on a pageblock
2234 * boundary. If alignment is required, use move_freepages_block()
2236 static int move_freepages(struct zone *zone,
2237 struct page *start_page, struct page *end_page,
2238 int migratetype, int *num_movable)
2242 int pages_moved = 0;
2244 for (page = start_page; page <= end_page;) {
2245 if (!pfn_valid_within(page_to_pfn(page))) {
2250 if (!PageBuddy(page)) {
2252 * We assume that pages that could be isolated for
2253 * migration are movable. But we don't actually try
2254 * isolating, as that would be expensive.
2257 (PageLRU(page) || __PageMovable(page)))
2264 /* Make sure we are not inadvertently changing nodes */
2265 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2266 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2268 order = page_order(page);
2269 move_to_free_area(page, &zone->free_area[order], migratetype);
2271 pages_moved += 1 << order;
2277 int move_freepages_block(struct zone *zone, struct page *page,
2278 int migratetype, int *num_movable)
2280 unsigned long start_pfn, end_pfn;
2281 struct page *start_page, *end_page;
2286 start_pfn = page_to_pfn(page);
2287 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2288 start_page = pfn_to_page(start_pfn);
2289 end_page = start_page + pageblock_nr_pages - 1;
2290 end_pfn = start_pfn + pageblock_nr_pages - 1;
2292 /* Do not cross zone boundaries */
2293 if (!zone_spans_pfn(zone, start_pfn))
2295 if (!zone_spans_pfn(zone, end_pfn))
2298 return move_freepages(zone, start_page, end_page, migratetype,
2302 static void change_pageblock_range(struct page *pageblock_page,
2303 int start_order, int migratetype)
2305 int nr_pageblocks = 1 << (start_order - pageblock_order);
2307 while (nr_pageblocks--) {
2308 set_pageblock_migratetype(pageblock_page, migratetype);
2309 pageblock_page += pageblock_nr_pages;
2314 * When we are falling back to another migratetype during allocation, try to
2315 * steal extra free pages from the same pageblocks to satisfy further
2316 * allocations, instead of polluting multiple pageblocks.
2318 * If we are stealing a relatively large buddy page, it is likely there will
2319 * be more free pages in the pageblock, so try to steal them all. For
2320 * reclaimable and unmovable allocations, we steal regardless of page size,
2321 * as fragmentation caused by those allocations polluting movable pageblocks
2322 * is worse than movable allocations stealing from unmovable and reclaimable
2325 static bool can_steal_fallback(unsigned int order, int start_mt)
2328 * Leaving this order check is intended, although there is
2329 * relaxed order check in next check. The reason is that
2330 * we can actually steal whole pageblock if this condition met,
2331 * but, below check doesn't guarantee it and that is just heuristic
2332 * so could be changed anytime.
2334 if (order >= pageblock_order)
2337 if (order >= pageblock_order / 2 ||
2338 start_mt == MIGRATE_RECLAIMABLE ||
2339 start_mt == MIGRATE_UNMOVABLE ||
2340 page_group_by_mobility_disabled)
2346 static inline void boost_watermark(struct zone *zone)
2348 unsigned long max_boost;
2350 if (!watermark_boost_factor)
2353 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2354 watermark_boost_factor, 10000);
2357 * high watermark may be uninitialised if fragmentation occurs
2358 * very early in boot so do not boost. We do not fall
2359 * through and boost by pageblock_nr_pages as failing
2360 * allocations that early means that reclaim is not going
2361 * to help and it may even be impossible to reclaim the
2362 * boosted watermark resulting in a hang.
2367 max_boost = max(pageblock_nr_pages, max_boost);
2369 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2374 * This function implements actual steal behaviour. If order is large enough,
2375 * we can steal whole pageblock. If not, we first move freepages in this
2376 * pageblock to our migratetype and determine how many already-allocated pages
2377 * are there in the pageblock with a compatible migratetype. If at least half
2378 * of pages are free or compatible, we can change migratetype of the pageblock
2379 * itself, so pages freed in the future will be put on the correct free list.
2381 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2382 unsigned int alloc_flags, int start_type, bool whole_block)
2384 unsigned int current_order = page_order(page);
2385 struct free_area *area;
2386 int free_pages, movable_pages, alike_pages;
2389 old_block_type = get_pageblock_migratetype(page);
2392 * This can happen due to races and we want to prevent broken
2393 * highatomic accounting.
2395 if (is_migrate_highatomic(old_block_type))
2398 /* Take ownership for orders >= pageblock_order */
2399 if (current_order >= pageblock_order) {
2400 change_pageblock_range(page, current_order, start_type);
2405 * Boost watermarks to increase reclaim pressure to reduce the
2406 * likelihood of future fallbacks. Wake kswapd now as the node
2407 * may be balanced overall and kswapd will not wake naturally.
2409 boost_watermark(zone);
2410 if (alloc_flags & ALLOC_KSWAPD)
2411 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2413 /* We are not allowed to try stealing from the whole block */
2417 free_pages = move_freepages_block(zone, page, start_type,
2420 * Determine how many pages are compatible with our allocation.
2421 * For movable allocation, it's the number of movable pages which
2422 * we just obtained. For other types it's a bit more tricky.
2424 if (start_type == MIGRATE_MOVABLE) {
2425 alike_pages = movable_pages;
2428 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2429 * to MOVABLE pageblock, consider all non-movable pages as
2430 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2431 * vice versa, be conservative since we can't distinguish the
2432 * exact migratetype of non-movable pages.
2434 if (old_block_type == MIGRATE_MOVABLE)
2435 alike_pages = pageblock_nr_pages
2436 - (free_pages + movable_pages);
2441 /* moving whole block can fail due to zone boundary conditions */
2446 * If a sufficient number of pages in the block are either free or of
2447 * comparable migratability as our allocation, claim the whole block.
2449 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2450 page_group_by_mobility_disabled)
2451 set_pageblock_migratetype(page, start_type);
2456 area = &zone->free_area[current_order];
2457 move_to_free_area(page, area, start_type);
2461 * Check whether there is a suitable fallback freepage with requested order.
2462 * If only_stealable is true, this function returns fallback_mt only if
2463 * we can steal other freepages all together. This would help to reduce
2464 * fragmentation due to mixed migratetype pages in one pageblock.
2466 int find_suitable_fallback(struct free_area *area, unsigned int order,
2467 int migratetype, bool only_stealable, bool *can_steal)
2472 if (area->nr_free == 0)
2477 fallback_mt = fallbacks[migratetype][i];
2478 if (fallback_mt == MIGRATE_TYPES)
2481 if (free_area_empty(area, fallback_mt))
2484 if (can_steal_fallback(order, migratetype))
2487 if (!only_stealable)
2498 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2499 * there are no empty page blocks that contain a page with a suitable order
2501 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2502 unsigned int alloc_order)
2505 unsigned long max_managed, flags;
2508 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2509 * Check is race-prone but harmless.
2511 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2512 if (zone->nr_reserved_highatomic >= max_managed)
2515 spin_lock_irqsave(&zone->lock, flags);
2517 /* Recheck the nr_reserved_highatomic limit under the lock */
2518 if (zone->nr_reserved_highatomic >= max_managed)
2522 mt = get_pageblock_migratetype(page);
2523 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2524 && !is_migrate_cma(mt)) {
2525 zone->nr_reserved_highatomic += pageblock_nr_pages;
2526 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2527 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2531 spin_unlock_irqrestore(&zone->lock, flags);
2535 * Used when an allocation is about to fail under memory pressure. This
2536 * potentially hurts the reliability of high-order allocations when under
2537 * intense memory pressure but failed atomic allocations should be easier
2538 * to recover from than an OOM.
2540 * If @force is true, try to unreserve a pageblock even though highatomic
2541 * pageblock is exhausted.
2543 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2546 struct zonelist *zonelist = ac->zonelist;
2547 unsigned long flags;
2554 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2557 * Preserve at least one pageblock unless memory pressure
2560 if (!force && zone->nr_reserved_highatomic <=
2564 spin_lock_irqsave(&zone->lock, flags);
2565 for (order = 0; order < MAX_ORDER; order++) {
2566 struct free_area *area = &(zone->free_area[order]);
2568 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2573 * In page freeing path, migratetype change is racy so
2574 * we can counter several free pages in a pageblock
2575 * in this loop althoug we changed the pageblock type
2576 * from highatomic to ac->migratetype. So we should
2577 * adjust the count once.
2579 if (is_migrate_highatomic_page(page)) {
2581 * It should never happen but changes to
2582 * locking could inadvertently allow a per-cpu
2583 * drain to add pages to MIGRATE_HIGHATOMIC
2584 * while unreserving so be safe and watch for
2587 zone->nr_reserved_highatomic -= min(
2589 zone->nr_reserved_highatomic);
2593 * Convert to ac->migratetype and avoid the normal
2594 * pageblock stealing heuristics. Minimally, the caller
2595 * is doing the work and needs the pages. More
2596 * importantly, if the block was always converted to
2597 * MIGRATE_UNMOVABLE or another type then the number
2598 * of pageblocks that cannot be completely freed
2601 set_pageblock_migratetype(page, ac->migratetype);
2602 ret = move_freepages_block(zone, page, ac->migratetype,
2605 spin_unlock_irqrestore(&zone->lock, flags);
2609 spin_unlock_irqrestore(&zone->lock, flags);
2616 * Try finding a free buddy page on the fallback list and put it on the free
2617 * list of requested migratetype, possibly along with other pages from the same
2618 * block, depending on fragmentation avoidance heuristics. Returns true if
2619 * fallback was found so that __rmqueue_smallest() can grab it.
2621 * The use of signed ints for order and current_order is a deliberate
2622 * deviation from the rest of this file, to make the for loop
2623 * condition simpler.
2625 static __always_inline bool
2626 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2627 unsigned int alloc_flags)
2629 struct free_area *area;
2631 int min_order = order;
2637 * Do not steal pages from freelists belonging to other pageblocks
2638 * i.e. orders < pageblock_order. If there are no local zones free,
2639 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2641 if (alloc_flags & ALLOC_NOFRAGMENT)
2642 min_order = pageblock_order;
2645 * Find the largest available free page in the other list. This roughly
2646 * approximates finding the pageblock with the most free pages, which
2647 * would be too costly to do exactly.
2649 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2651 area = &(zone->free_area[current_order]);
2652 fallback_mt = find_suitable_fallback(area, current_order,
2653 start_migratetype, false, &can_steal);
2654 if (fallback_mt == -1)
2658 * We cannot steal all free pages from the pageblock and the
2659 * requested migratetype is movable. In that case it's better to
2660 * steal and split the smallest available page instead of the
2661 * largest available page, because even if the next movable
2662 * allocation falls back into a different pageblock than this
2663 * one, it won't cause permanent fragmentation.
2665 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2666 && current_order > order)
2675 for (current_order = order; current_order < MAX_ORDER;
2677 area = &(zone->free_area[current_order]);
2678 fallback_mt = find_suitable_fallback(area, current_order,
2679 start_migratetype, false, &can_steal);
2680 if (fallback_mt != -1)
2685 * This should not happen - we already found a suitable fallback
2686 * when looking for the largest page.
2688 VM_BUG_ON(current_order == MAX_ORDER);
2691 page = get_page_from_free_area(area, fallback_mt);
2693 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2696 trace_mm_page_alloc_extfrag(page, order, current_order,
2697 start_migratetype, fallback_mt);
2704 * Do the hard work of removing an element from the buddy allocator.
2705 * Call me with the zone->lock already held.
2707 static __always_inline struct page *
2708 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2709 unsigned int alloc_flags)
2714 page = __rmqueue_smallest(zone, order, migratetype);
2715 if (unlikely(!page)) {
2716 if (migratetype == MIGRATE_MOVABLE)
2717 page = __rmqueue_cma_fallback(zone, order);
2719 if (!page && __rmqueue_fallback(zone, order, migratetype,
2724 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2729 * Obtain a specified number of elements from the buddy allocator, all under
2730 * a single hold of the lock, for efficiency. Add them to the supplied list.
2731 * Returns the number of new pages which were placed at *list.
2733 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2734 unsigned long count, struct list_head *list,
2735 int migratetype, unsigned int alloc_flags)
2739 spin_lock(&zone->lock);
2740 for (i = 0; i < count; ++i) {
2741 struct page *page = __rmqueue(zone, order, migratetype,
2743 if (unlikely(page == NULL))
2746 if (unlikely(check_pcp_refill(page)))
2750 * Split buddy pages returned by expand() are received here in
2751 * physical page order. The page is added to the tail of
2752 * caller's list. From the callers perspective, the linked list
2753 * is ordered by page number under some conditions. This is
2754 * useful for IO devices that can forward direction from the
2755 * head, thus also in the physical page order. This is useful
2756 * for IO devices that can merge IO requests if the physical
2757 * pages are ordered properly.
2759 list_add_tail(&page->lru, list);
2761 if (is_migrate_cma(get_pcppage_migratetype(page)))
2762 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2767 * i pages were removed from the buddy list even if some leak due
2768 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2769 * on i. Do not confuse with 'alloced' which is the number of
2770 * pages added to the pcp list.
2772 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2773 spin_unlock(&zone->lock);
2779 * Called from the vmstat counter updater to drain pagesets of this
2780 * currently executing processor on remote nodes after they have
2783 * Note that this function must be called with the thread pinned to
2784 * a single processor.
2786 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2788 unsigned long flags;
2789 int to_drain, batch;
2791 local_irq_save(flags);
2792 batch = READ_ONCE(pcp->batch);
2793 to_drain = min(pcp->count, batch);
2795 free_pcppages_bulk(zone, to_drain, pcp);
2796 local_irq_restore(flags);
2801 * Drain pcplists of the indicated processor and zone.
2803 * The processor must either be the current processor and the
2804 * thread pinned to the current processor or a processor that
2807 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2809 unsigned long flags;
2810 struct per_cpu_pageset *pset;
2811 struct per_cpu_pages *pcp;
2813 local_irq_save(flags);
2814 pset = per_cpu_ptr(zone->pageset, cpu);
2818 free_pcppages_bulk(zone, pcp->count, pcp);
2819 local_irq_restore(flags);
2823 * Drain pcplists of all zones on the indicated processor.
2825 * The processor must either be the current processor and the
2826 * thread pinned to the current processor or a processor that
2829 static void drain_pages(unsigned int cpu)
2833 for_each_populated_zone(zone) {
2834 drain_pages_zone(cpu, zone);
2839 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2841 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2842 * the single zone's pages.
2844 void drain_local_pages(struct zone *zone)
2846 int cpu = smp_processor_id();
2849 drain_pages_zone(cpu, zone);
2854 static void drain_local_pages_wq(struct work_struct *work)
2856 struct pcpu_drain *drain;
2858 drain = container_of(work, struct pcpu_drain, work);
2861 * drain_all_pages doesn't use proper cpu hotplug protection so
2862 * we can race with cpu offline when the WQ can move this from
2863 * a cpu pinned worker to an unbound one. We can operate on a different
2864 * cpu which is allright but we also have to make sure to not move to
2868 drain_local_pages(drain->zone);
2873 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2875 * When zone parameter is non-NULL, spill just the single zone's pages.
2877 * Note that this can be extremely slow as the draining happens in a workqueue.
2879 void drain_all_pages(struct zone *zone)
2884 * Allocate in the BSS so we wont require allocation in
2885 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2887 static cpumask_t cpus_with_pcps;
2890 * Make sure nobody triggers this path before mm_percpu_wq is fully
2893 if (WARN_ON_ONCE(!mm_percpu_wq))
2897 * Do not drain if one is already in progress unless it's specific to
2898 * a zone. Such callers are primarily CMA and memory hotplug and need
2899 * the drain to be complete when the call returns.
2901 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2904 mutex_lock(&pcpu_drain_mutex);
2908 * We don't care about racing with CPU hotplug event
2909 * as offline notification will cause the notified
2910 * cpu to drain that CPU pcps and on_each_cpu_mask
2911 * disables preemption as part of its processing
2913 for_each_online_cpu(cpu) {
2914 struct per_cpu_pageset *pcp;
2916 bool has_pcps = false;
2919 pcp = per_cpu_ptr(zone->pageset, cpu);
2923 for_each_populated_zone(z) {
2924 pcp = per_cpu_ptr(z->pageset, cpu);
2925 if (pcp->pcp.count) {
2933 cpumask_set_cpu(cpu, &cpus_with_pcps);
2935 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2938 for_each_cpu(cpu, &cpus_with_pcps) {
2939 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2942 INIT_WORK(&drain->work, drain_local_pages_wq);
2943 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2945 for_each_cpu(cpu, &cpus_with_pcps)
2946 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2948 mutex_unlock(&pcpu_drain_mutex);
2951 #ifdef CONFIG_HIBERNATION
2954 * Touch the watchdog for every WD_PAGE_COUNT pages.
2956 #define WD_PAGE_COUNT (128*1024)
2958 void mark_free_pages(struct zone *zone)
2960 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2961 unsigned long flags;
2962 unsigned int order, t;
2965 if (zone_is_empty(zone))
2968 spin_lock_irqsave(&zone->lock, flags);
2970 max_zone_pfn = zone_end_pfn(zone);
2971 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2972 if (pfn_valid(pfn)) {
2973 page = pfn_to_page(pfn);
2975 if (!--page_count) {
2976 touch_nmi_watchdog();
2977 page_count = WD_PAGE_COUNT;
2980 if (page_zone(page) != zone)
2983 if (!swsusp_page_is_forbidden(page))
2984 swsusp_unset_page_free(page);
2987 for_each_migratetype_order(order, t) {
2988 list_for_each_entry(page,
2989 &zone->free_area[order].free_list[t], lru) {
2992 pfn = page_to_pfn(page);
2993 for (i = 0; i < (1UL << order); i++) {
2994 if (!--page_count) {
2995 touch_nmi_watchdog();
2996 page_count = WD_PAGE_COUNT;
2998 swsusp_set_page_free(pfn_to_page(pfn + i));
3002 spin_unlock_irqrestore(&zone->lock, flags);
3004 #endif /* CONFIG_PM */
3006 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3010 if (!free_pcp_prepare(page))
3013 migratetype = get_pfnblock_migratetype(page, pfn);
3014 set_pcppage_migratetype(page, migratetype);
3018 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3020 struct zone *zone = page_zone(page);
3021 struct per_cpu_pages *pcp;
3024 migratetype = get_pcppage_migratetype(page);
3025 __count_vm_event(PGFREE);
3028 * We only track unmovable, reclaimable and movable on pcp lists.
3029 * Free ISOLATE pages back to the allocator because they are being
3030 * offlined but treat HIGHATOMIC as movable pages so we can get those
3031 * areas back if necessary. Otherwise, we may have to free
3032 * excessively into the page allocator
3034 if (migratetype >= MIGRATE_PCPTYPES) {
3035 if (unlikely(is_migrate_isolate(migratetype))) {
3036 free_one_page(zone, page, pfn, 0, migratetype);
3039 migratetype = MIGRATE_MOVABLE;
3042 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3043 list_add(&page->lru, &pcp->lists[migratetype]);
3045 if (pcp->count >= pcp->high) {
3046 unsigned long batch = READ_ONCE(pcp->batch);
3047 free_pcppages_bulk(zone, batch, pcp);
3052 * Free a 0-order page
3054 void free_unref_page(struct page *page)
3056 unsigned long flags;
3057 unsigned long pfn = page_to_pfn(page);
3059 if (!free_unref_page_prepare(page, pfn))
3062 local_irq_save(flags);
3063 free_unref_page_commit(page, pfn);
3064 local_irq_restore(flags);
3068 * Free a list of 0-order pages
3070 void free_unref_page_list(struct list_head *list)
3072 struct page *page, *next;
3073 unsigned long flags, pfn;
3074 int batch_count = 0;
3076 /* Prepare pages for freeing */
3077 list_for_each_entry_safe(page, next, list, lru) {
3078 pfn = page_to_pfn(page);
3079 if (!free_unref_page_prepare(page, pfn))
3080 list_del(&page->lru);
3081 set_page_private(page, pfn);
3084 local_irq_save(flags);
3085 list_for_each_entry_safe(page, next, list, lru) {
3086 unsigned long pfn = page_private(page);
3088 set_page_private(page, 0);
3089 trace_mm_page_free_batched(page);
3090 free_unref_page_commit(page, pfn);
3093 * Guard against excessive IRQ disabled times when we get
3094 * a large list of pages to free.
3096 if (++batch_count == SWAP_CLUSTER_MAX) {
3097 local_irq_restore(flags);
3099 local_irq_save(flags);
3102 local_irq_restore(flags);
3106 * split_page takes a non-compound higher-order page, and splits it into
3107 * n (1<<order) sub-pages: page[0..n]
3108 * Each sub-page must be freed individually.
3110 * Note: this is probably too low level an operation for use in drivers.
3111 * Please consult with lkml before using this in your driver.
3113 void split_page(struct page *page, unsigned int order)
3117 VM_BUG_ON_PAGE(PageCompound(page), page);
3118 VM_BUG_ON_PAGE(!page_count(page), page);
3120 for (i = 1; i < (1 << order); i++)
3121 set_page_refcounted(page + i);
3122 split_page_owner(page, order);
3124 EXPORT_SYMBOL_GPL(split_page);
3126 int __isolate_free_page(struct page *page, unsigned int order)
3128 struct free_area *area = &page_zone(page)->free_area[order];
3129 unsigned long watermark;
3133 BUG_ON(!PageBuddy(page));
3135 zone = page_zone(page);
3136 mt = get_pageblock_migratetype(page);
3138 if (!is_migrate_isolate(mt)) {
3140 * Obey watermarks as if the page was being allocated. We can
3141 * emulate a high-order watermark check with a raised order-0
3142 * watermark, because we already know our high-order page
3145 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3146 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3149 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3152 /* Remove page from free list */
3154 del_page_from_free_area(page, area);
3157 * Set the pageblock if the isolated page is at least half of a
3160 if (order >= pageblock_order - 1) {
3161 struct page *endpage = page + (1 << order) - 1;
3162 for (; page < endpage; page += pageblock_nr_pages) {
3163 int mt = get_pageblock_migratetype(page);
3164 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3165 && !is_migrate_highatomic(mt))
3166 set_pageblock_migratetype(page,
3172 return 1UL << order;
3176 * Update NUMA hit/miss statistics
3178 * Must be called with interrupts disabled.
3180 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3183 enum numa_stat_item local_stat = NUMA_LOCAL;
3185 /* skip numa counters update if numa stats is disabled */
3186 if (!static_branch_likely(&vm_numa_stat_key))
3189 if (zone_to_nid(z) != numa_node_id())
3190 local_stat = NUMA_OTHER;
3192 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3193 __inc_numa_state(z, NUMA_HIT);
3195 __inc_numa_state(z, NUMA_MISS);
3196 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3198 __inc_numa_state(z, local_stat);
3202 /* Remove page from the per-cpu list, caller must protect the list */
3203 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3204 unsigned int alloc_flags,
3205 struct per_cpu_pages *pcp,
3206 struct list_head *list)
3211 if (list_empty(list)) {
3212 pcp->count += rmqueue_bulk(zone, 0,
3214 migratetype, alloc_flags);
3215 if (unlikely(list_empty(list)))
3219 page = list_first_entry(list, struct page, lru);
3220 list_del(&page->lru);
3222 } while (check_new_pcp(page));
3227 /* Lock and remove page from the per-cpu list */
3228 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3229 struct zone *zone, gfp_t gfp_flags,
3230 int migratetype, unsigned int alloc_flags)
3232 struct per_cpu_pages *pcp;
3233 struct list_head *list;
3235 unsigned long flags;
3237 local_irq_save(flags);
3238 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3239 list = &pcp->lists[migratetype];
3240 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3242 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3243 zone_statistics(preferred_zone, zone);
3245 local_irq_restore(flags);
3250 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3253 struct page *rmqueue(struct zone *preferred_zone,
3254 struct zone *zone, unsigned int order,
3255 gfp_t gfp_flags, unsigned int alloc_flags,
3258 unsigned long flags;
3261 if (likely(order == 0)) {
3262 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3263 migratetype, alloc_flags);
3268 * We most definitely don't want callers attempting to
3269 * allocate greater than order-1 page units with __GFP_NOFAIL.
3271 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3272 spin_lock_irqsave(&zone->lock, flags);
3276 if (alloc_flags & ALLOC_HARDER) {
3277 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3279 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3282 page = __rmqueue(zone, order, migratetype, alloc_flags);
3283 } while (page && check_new_pages(page, order));
3284 spin_unlock(&zone->lock);
3287 __mod_zone_freepage_state(zone, -(1 << order),
3288 get_pcppage_migratetype(page));
3290 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3291 zone_statistics(preferred_zone, zone);
3292 local_irq_restore(flags);
3295 /* Separate test+clear to avoid unnecessary atomics */
3296 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3297 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3298 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3301 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3305 local_irq_restore(flags);
3309 #ifdef CONFIG_FAIL_PAGE_ALLOC
3312 struct fault_attr attr;
3314 bool ignore_gfp_highmem;
3315 bool ignore_gfp_reclaim;
3317 } fail_page_alloc = {
3318 .attr = FAULT_ATTR_INITIALIZER,
3319 .ignore_gfp_reclaim = true,
3320 .ignore_gfp_highmem = true,
3324 static int __init setup_fail_page_alloc(char *str)
3326 return setup_fault_attr(&fail_page_alloc.attr, str);
3328 __setup("fail_page_alloc=", setup_fail_page_alloc);
3330 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3332 if (order < fail_page_alloc.min_order)
3334 if (gfp_mask & __GFP_NOFAIL)
3336 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3338 if (fail_page_alloc.ignore_gfp_reclaim &&
3339 (gfp_mask & __GFP_DIRECT_RECLAIM))
3342 return should_fail(&fail_page_alloc.attr, 1 << order);
3345 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3347 static int __init fail_page_alloc_debugfs(void)
3349 umode_t mode = S_IFREG | 0600;
3352 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3353 &fail_page_alloc.attr);
3355 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3356 &fail_page_alloc.ignore_gfp_reclaim);
3357 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3358 &fail_page_alloc.ignore_gfp_highmem);
3359 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3364 late_initcall(fail_page_alloc_debugfs);
3366 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3368 #else /* CONFIG_FAIL_PAGE_ALLOC */
3370 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3375 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3377 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3379 return __should_fail_alloc_page(gfp_mask, order);
3381 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3384 * Return true if free base pages are above 'mark'. For high-order checks it
3385 * will return true of the order-0 watermark is reached and there is at least
3386 * one free page of a suitable size. Checking now avoids taking the zone lock
3387 * to check in the allocation paths if no pages are free.
3389 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3390 int classzone_idx, unsigned int alloc_flags,
3395 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3397 /* free_pages may go negative - that's OK */
3398 free_pages -= (1 << order) - 1;
3400 if (alloc_flags & ALLOC_HIGH)
3404 * If the caller does not have rights to ALLOC_HARDER then subtract
3405 * the high-atomic reserves. This will over-estimate the size of the
3406 * atomic reserve but it avoids a search.
3408 if (likely(!alloc_harder)) {
3409 free_pages -= z->nr_reserved_highatomic;
3412 * OOM victims can try even harder than normal ALLOC_HARDER
3413 * users on the grounds that it's definitely going to be in
3414 * the exit path shortly and free memory. Any allocation it
3415 * makes during the free path will be small and short-lived.
3417 if (alloc_flags & ALLOC_OOM)
3425 /* If allocation can't use CMA areas don't use free CMA pages */
3426 if (!(alloc_flags & ALLOC_CMA))
3427 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3431 * Check watermarks for an order-0 allocation request. If these
3432 * are not met, then a high-order request also cannot go ahead
3433 * even if a suitable page happened to be free.
3435 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3438 /* If this is an order-0 request then the watermark is fine */
3442 /* For a high-order request, check at least one suitable page is free */
3443 for (o = order; o < MAX_ORDER; o++) {
3444 struct free_area *area = &z->free_area[o];
3450 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3451 if (!free_area_empty(area, mt))
3456 if ((alloc_flags & ALLOC_CMA) &&
3457 !free_area_empty(area, MIGRATE_CMA)) {
3462 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3468 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3469 int classzone_idx, unsigned int alloc_flags)
3471 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3472 zone_page_state(z, NR_FREE_PAGES));
3475 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3476 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3478 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3482 /* If allocation can't use CMA areas don't use free CMA pages */
3483 if (!(alloc_flags & ALLOC_CMA))
3484 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3488 * Fast check for order-0 only. If this fails then the reserves
3489 * need to be calculated. There is a corner case where the check
3490 * passes but only the high-order atomic reserve are free. If
3491 * the caller is !atomic then it'll uselessly search the free
3492 * list. That corner case is then slower but it is harmless.
3494 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3497 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3501 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3502 unsigned long mark, int classzone_idx)
3504 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3506 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3507 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3509 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3514 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3516 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3517 node_reclaim_distance;
3519 #else /* CONFIG_NUMA */
3520 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3524 #endif /* CONFIG_NUMA */
3527 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3528 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3529 * premature use of a lower zone may cause lowmem pressure problems that
3530 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3531 * probably too small. It only makes sense to spread allocations to avoid
3532 * fragmentation between the Normal and DMA32 zones.
3534 static inline unsigned int
3535 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3537 unsigned int alloc_flags = 0;
3539 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3540 alloc_flags |= ALLOC_KSWAPD;
3542 #ifdef CONFIG_ZONE_DMA32
3546 if (zone_idx(zone) != ZONE_NORMAL)
3550 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3551 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3552 * on UMA that if Normal is populated then so is DMA32.
3554 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3555 if (nr_online_nodes > 1 && !populated_zone(--zone))
3558 alloc_flags |= ALLOC_NOFRAGMENT;
3559 #endif /* CONFIG_ZONE_DMA32 */
3564 * get_page_from_freelist goes through the zonelist trying to allocate
3567 static struct page *
3568 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3569 const struct alloc_context *ac)
3573 struct pglist_data *last_pgdat_dirty_limit = NULL;
3578 * Scan zonelist, looking for a zone with enough free.
3579 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3581 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3582 z = ac->preferred_zoneref;
3583 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3588 if (cpusets_enabled() &&
3589 (alloc_flags & ALLOC_CPUSET) &&
3590 !__cpuset_zone_allowed(zone, gfp_mask))
3593 * When allocating a page cache page for writing, we
3594 * want to get it from a node that is within its dirty
3595 * limit, such that no single node holds more than its
3596 * proportional share of globally allowed dirty pages.
3597 * The dirty limits take into account the node's
3598 * lowmem reserves and high watermark so that kswapd
3599 * should be able to balance it without having to
3600 * write pages from its LRU list.
3602 * XXX: For now, allow allocations to potentially
3603 * exceed the per-node dirty limit in the slowpath
3604 * (spread_dirty_pages unset) before going into reclaim,
3605 * which is important when on a NUMA setup the allowed
3606 * nodes are together not big enough to reach the
3607 * global limit. The proper fix for these situations
3608 * will require awareness of nodes in the
3609 * dirty-throttling and the flusher threads.
3611 if (ac->spread_dirty_pages) {
3612 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3615 if (!node_dirty_ok(zone->zone_pgdat)) {
3616 last_pgdat_dirty_limit = zone->zone_pgdat;
3621 if (no_fallback && nr_online_nodes > 1 &&
3622 zone != ac->preferred_zoneref->zone) {
3626 * If moving to a remote node, retry but allow
3627 * fragmenting fallbacks. Locality is more important
3628 * than fragmentation avoidance.
3630 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3631 if (zone_to_nid(zone) != local_nid) {
3632 alloc_flags &= ~ALLOC_NOFRAGMENT;
3637 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3638 if (!zone_watermark_fast(zone, order, mark,
3639 ac_classzone_idx(ac), alloc_flags)) {
3642 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3644 * Watermark failed for this zone, but see if we can
3645 * grow this zone if it contains deferred pages.
3647 if (static_branch_unlikely(&deferred_pages)) {
3648 if (_deferred_grow_zone(zone, order))
3652 /* Checked here to keep the fast path fast */
3653 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3654 if (alloc_flags & ALLOC_NO_WATERMARKS)
3657 if (node_reclaim_mode == 0 ||
3658 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3661 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3663 case NODE_RECLAIM_NOSCAN:
3666 case NODE_RECLAIM_FULL:
3667 /* scanned but unreclaimable */
3670 /* did we reclaim enough */
3671 if (zone_watermark_ok(zone, order, mark,
3672 ac_classzone_idx(ac), alloc_flags))
3680 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3681 gfp_mask, alloc_flags, ac->migratetype);
3683 prep_new_page(page, order, gfp_mask, alloc_flags);
3686 * If this is a high-order atomic allocation then check
3687 * if the pageblock should be reserved for the future
3689 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3690 reserve_highatomic_pageblock(page, zone, order);
3694 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3695 /* Try again if zone has deferred pages */
3696 if (static_branch_unlikely(&deferred_pages)) {
3697 if (_deferred_grow_zone(zone, order))
3705 * It's possible on a UMA machine to get through all zones that are
3706 * fragmented. If avoiding fragmentation, reset and try again.
3709 alloc_flags &= ~ALLOC_NOFRAGMENT;
3716 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3718 unsigned int filter = SHOW_MEM_FILTER_NODES;
3721 * This documents exceptions given to allocations in certain
3722 * contexts that are allowed to allocate outside current's set
3725 if (!(gfp_mask & __GFP_NOMEMALLOC))
3726 if (tsk_is_oom_victim(current) ||
3727 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3728 filter &= ~SHOW_MEM_FILTER_NODES;
3729 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3730 filter &= ~SHOW_MEM_FILTER_NODES;
3732 show_mem(filter, nodemask);
3735 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3737 struct va_format vaf;
3739 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3741 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3744 va_start(args, fmt);
3747 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3748 current->comm, &vaf, gfp_mask, &gfp_mask,
3749 nodemask_pr_args(nodemask));
3752 cpuset_print_current_mems_allowed();
3755 warn_alloc_show_mem(gfp_mask, nodemask);
3758 static inline struct page *
3759 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3760 unsigned int alloc_flags,
3761 const struct alloc_context *ac)
3765 page = get_page_from_freelist(gfp_mask, order,
3766 alloc_flags|ALLOC_CPUSET, ac);
3768 * fallback to ignore cpuset restriction if our nodes
3772 page = get_page_from_freelist(gfp_mask, order,
3778 static inline struct page *
3779 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3780 const struct alloc_context *ac, unsigned long *did_some_progress)
3782 struct oom_control oc = {
3783 .zonelist = ac->zonelist,
3784 .nodemask = ac->nodemask,
3786 .gfp_mask = gfp_mask,
3791 *did_some_progress = 0;
3794 * Acquire the oom lock. If that fails, somebody else is
3795 * making progress for us.
3797 if (!mutex_trylock(&oom_lock)) {
3798 *did_some_progress = 1;
3799 schedule_timeout_uninterruptible(1);
3804 * Go through the zonelist yet one more time, keep very high watermark
3805 * here, this is only to catch a parallel oom killing, we must fail if
3806 * we're still under heavy pressure. But make sure that this reclaim
3807 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3808 * allocation which will never fail due to oom_lock already held.
3810 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3811 ~__GFP_DIRECT_RECLAIM, order,
3812 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3816 /* Coredumps can quickly deplete all memory reserves */
3817 if (current->flags & PF_DUMPCORE)
3819 /* The OOM killer will not help higher order allocs */
3820 if (order > PAGE_ALLOC_COSTLY_ORDER)
3823 * We have already exhausted all our reclaim opportunities without any
3824 * success so it is time to admit defeat. We will skip the OOM killer
3825 * because it is very likely that the caller has a more reasonable
3826 * fallback than shooting a random task.
3828 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3830 /* The OOM killer does not needlessly kill tasks for lowmem */
3831 if (ac->high_zoneidx < ZONE_NORMAL)
3833 if (pm_suspended_storage())
3836 * XXX: GFP_NOFS allocations should rather fail than rely on
3837 * other request to make a forward progress.
3838 * We are in an unfortunate situation where out_of_memory cannot
3839 * do much for this context but let's try it to at least get
3840 * access to memory reserved if the current task is killed (see
3841 * out_of_memory). Once filesystems are ready to handle allocation
3842 * failures more gracefully we should just bail out here.
3845 /* The OOM killer may not free memory on a specific node */
3846 if (gfp_mask & __GFP_THISNODE)
3849 /* Exhausted what can be done so it's blame time */
3850 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3851 *did_some_progress = 1;
3854 * Help non-failing allocations by giving them access to memory
3857 if (gfp_mask & __GFP_NOFAIL)
3858 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3859 ALLOC_NO_WATERMARKS, ac);
3862 mutex_unlock(&oom_lock);
3867 * Maximum number of compaction retries wit a progress before OOM
3868 * killer is consider as the only way to move forward.
3870 #define MAX_COMPACT_RETRIES 16
3872 #ifdef CONFIG_COMPACTION
3873 /* Try memory compaction for high-order allocations before reclaim */
3874 static struct page *
3875 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3876 unsigned int alloc_flags, const struct alloc_context *ac,
3877 enum compact_priority prio, enum compact_result *compact_result)
3879 struct page *page = NULL;
3880 unsigned long pflags;
3881 unsigned int noreclaim_flag;
3886 psi_memstall_enter(&pflags);
3887 noreclaim_flag = memalloc_noreclaim_save();
3889 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3892 memalloc_noreclaim_restore(noreclaim_flag);
3893 psi_memstall_leave(&pflags);
3896 * At least in one zone compaction wasn't deferred or skipped, so let's
3897 * count a compaction stall
3899 count_vm_event(COMPACTSTALL);
3901 /* Prep a captured page if available */
3903 prep_new_page(page, order, gfp_mask, alloc_flags);
3905 /* Try get a page from the freelist if available */
3907 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3910 struct zone *zone = page_zone(page);
3912 zone->compact_blockskip_flush = false;
3913 compaction_defer_reset(zone, order, true);
3914 count_vm_event(COMPACTSUCCESS);
3919 * It's bad if compaction run occurs and fails. The most likely reason
3920 * is that pages exist, but not enough to satisfy watermarks.
3922 count_vm_event(COMPACTFAIL);
3930 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3931 enum compact_result compact_result,
3932 enum compact_priority *compact_priority,
3933 int *compaction_retries)
3935 int max_retries = MAX_COMPACT_RETRIES;
3938 int retries = *compaction_retries;
3939 enum compact_priority priority = *compact_priority;
3944 if (compaction_made_progress(compact_result))
3945 (*compaction_retries)++;
3948 * compaction considers all the zone as desperately out of memory
3949 * so it doesn't really make much sense to retry except when the
3950 * failure could be caused by insufficient priority
3952 if (compaction_failed(compact_result))
3953 goto check_priority;
3956 * compaction was skipped because there are not enough order-0 pages
3957 * to work with, so we retry only if it looks like reclaim can help.
3959 if (compaction_needs_reclaim(compact_result)) {
3960 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3965 * make sure the compaction wasn't deferred or didn't bail out early
3966 * due to locks contention before we declare that we should give up.
3967 * But the next retry should use a higher priority if allowed, so
3968 * we don't just keep bailing out endlessly.
3970 if (compaction_withdrawn(compact_result)) {
3971 goto check_priority;
3975 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3976 * costly ones because they are de facto nofail and invoke OOM
3977 * killer to move on while costly can fail and users are ready
3978 * to cope with that. 1/4 retries is rather arbitrary but we
3979 * would need much more detailed feedback from compaction to
3980 * make a better decision.
3982 if (order > PAGE_ALLOC_COSTLY_ORDER)
3984 if (*compaction_retries <= max_retries) {
3990 * Make sure there are attempts at the highest priority if we exhausted
3991 * all retries or failed at the lower priorities.
3994 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3995 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3997 if (*compact_priority > min_priority) {
3998 (*compact_priority)--;
3999 *compaction_retries = 0;
4003 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4007 static inline struct page *
4008 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4009 unsigned int alloc_flags, const struct alloc_context *ac,
4010 enum compact_priority prio, enum compact_result *compact_result)
4012 *compact_result = COMPACT_SKIPPED;
4017 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4018 enum compact_result compact_result,
4019 enum compact_priority *compact_priority,
4020 int *compaction_retries)
4025 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4029 * There are setups with compaction disabled which would prefer to loop
4030 * inside the allocator rather than hit the oom killer prematurely.
4031 * Let's give them a good hope and keep retrying while the order-0
4032 * watermarks are OK.
4034 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4036 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4037 ac_classzone_idx(ac), alloc_flags))
4042 #endif /* CONFIG_COMPACTION */
4044 #ifdef CONFIG_LOCKDEP
4045 static struct lockdep_map __fs_reclaim_map =
4046 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4048 static bool __need_fs_reclaim(gfp_t gfp_mask)
4050 gfp_mask = current_gfp_context(gfp_mask);
4052 /* no reclaim without waiting on it */
4053 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4056 /* this guy won't enter reclaim */
4057 if (current->flags & PF_MEMALLOC)
4060 /* We're only interested __GFP_FS allocations for now */
4061 if (!(gfp_mask & __GFP_FS))
4064 if (gfp_mask & __GFP_NOLOCKDEP)
4070 void __fs_reclaim_acquire(void)
4072 lock_map_acquire(&__fs_reclaim_map);
4075 void __fs_reclaim_release(void)
4077 lock_map_release(&__fs_reclaim_map);
4080 void fs_reclaim_acquire(gfp_t gfp_mask)
4082 if (__need_fs_reclaim(gfp_mask))
4083 __fs_reclaim_acquire();
4085 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4087 void fs_reclaim_release(gfp_t gfp_mask)
4089 if (__need_fs_reclaim(gfp_mask))
4090 __fs_reclaim_release();
4092 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4095 /* Perform direct synchronous page reclaim */
4097 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4098 const struct alloc_context *ac)
4101 unsigned int noreclaim_flag;
4102 unsigned long pflags;
4106 /* We now go into synchronous reclaim */
4107 cpuset_memory_pressure_bump();
4108 psi_memstall_enter(&pflags);
4109 fs_reclaim_acquire(gfp_mask);
4110 noreclaim_flag = memalloc_noreclaim_save();
4112 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4115 memalloc_noreclaim_restore(noreclaim_flag);
4116 fs_reclaim_release(gfp_mask);
4117 psi_memstall_leave(&pflags);
4124 /* The really slow allocator path where we enter direct reclaim */
4125 static inline struct page *
4126 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4127 unsigned int alloc_flags, const struct alloc_context *ac,
4128 unsigned long *did_some_progress)
4130 struct page *page = NULL;
4131 bool drained = false;
4133 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4134 if (unlikely(!(*did_some_progress)))
4138 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4141 * If an allocation failed after direct reclaim, it could be because
4142 * pages are pinned on the per-cpu lists or in high alloc reserves.
4143 * Shrink them them and try again
4145 if (!page && !drained) {
4146 unreserve_highatomic_pageblock(ac, false);
4147 drain_all_pages(NULL);
4155 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4156 const struct alloc_context *ac)
4160 pg_data_t *last_pgdat = NULL;
4161 enum zone_type high_zoneidx = ac->high_zoneidx;
4163 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4165 if (last_pgdat != zone->zone_pgdat)
4166 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4167 last_pgdat = zone->zone_pgdat;
4171 static inline unsigned int
4172 gfp_to_alloc_flags(gfp_t gfp_mask)
4174 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4176 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4177 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4180 * The caller may dip into page reserves a bit more if the caller
4181 * cannot run direct reclaim, or if the caller has realtime scheduling
4182 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4183 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4185 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4187 if (gfp_mask & __GFP_ATOMIC) {
4189 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4190 * if it can't schedule.
4192 if (!(gfp_mask & __GFP_NOMEMALLOC))
4193 alloc_flags |= ALLOC_HARDER;
4195 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4196 * comment for __cpuset_node_allowed().
4198 alloc_flags &= ~ALLOC_CPUSET;
4199 } else if (unlikely(rt_task(current)) && !in_interrupt())
4200 alloc_flags |= ALLOC_HARDER;
4202 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4203 alloc_flags |= ALLOC_KSWAPD;
4206 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4207 alloc_flags |= ALLOC_CMA;
4212 static bool oom_reserves_allowed(struct task_struct *tsk)
4214 if (!tsk_is_oom_victim(tsk))
4218 * !MMU doesn't have oom reaper so give access to memory reserves
4219 * only to the thread with TIF_MEMDIE set
4221 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4228 * Distinguish requests which really need access to full memory
4229 * reserves from oom victims which can live with a portion of it
4231 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4233 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4235 if (gfp_mask & __GFP_MEMALLOC)
4236 return ALLOC_NO_WATERMARKS;
4237 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4238 return ALLOC_NO_WATERMARKS;
4239 if (!in_interrupt()) {
4240 if (current->flags & PF_MEMALLOC)
4241 return ALLOC_NO_WATERMARKS;
4242 else if (oom_reserves_allowed(current))
4249 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4251 return !!__gfp_pfmemalloc_flags(gfp_mask);
4255 * Checks whether it makes sense to retry the reclaim to make a forward progress
4256 * for the given allocation request.
4258 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4259 * without success, or when we couldn't even meet the watermark if we
4260 * reclaimed all remaining pages on the LRU lists.
4262 * Returns true if a retry is viable or false to enter the oom path.
4265 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4266 struct alloc_context *ac, int alloc_flags,
4267 bool did_some_progress, int *no_progress_loops)
4274 * Costly allocations might have made a progress but this doesn't mean
4275 * their order will become available due to high fragmentation so
4276 * always increment the no progress counter for them
4278 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4279 *no_progress_loops = 0;
4281 (*no_progress_loops)++;
4284 * Make sure we converge to OOM if we cannot make any progress
4285 * several times in the row.
4287 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4288 /* Before OOM, exhaust highatomic_reserve */
4289 return unreserve_highatomic_pageblock(ac, true);
4293 * Keep reclaiming pages while there is a chance this will lead
4294 * somewhere. If none of the target zones can satisfy our allocation
4295 * request even if all reclaimable pages are considered then we are
4296 * screwed and have to go OOM.
4298 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4300 unsigned long available;
4301 unsigned long reclaimable;
4302 unsigned long min_wmark = min_wmark_pages(zone);
4305 available = reclaimable = zone_reclaimable_pages(zone);
4306 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4309 * Would the allocation succeed if we reclaimed all
4310 * reclaimable pages?
4312 wmark = __zone_watermark_ok(zone, order, min_wmark,
4313 ac_classzone_idx(ac), alloc_flags, available);
4314 trace_reclaim_retry_zone(z, order, reclaimable,
4315 available, min_wmark, *no_progress_loops, wmark);
4318 * If we didn't make any progress and have a lot of
4319 * dirty + writeback pages then we should wait for
4320 * an IO to complete to slow down the reclaim and
4321 * prevent from pre mature OOM
4323 if (!did_some_progress) {
4324 unsigned long write_pending;
4326 write_pending = zone_page_state_snapshot(zone,
4327 NR_ZONE_WRITE_PENDING);
4329 if (2 * write_pending > reclaimable) {
4330 congestion_wait(BLK_RW_ASYNC, HZ/10);
4342 * Memory allocation/reclaim might be called from a WQ context and the
4343 * current implementation of the WQ concurrency control doesn't
4344 * recognize that a particular WQ is congested if the worker thread is
4345 * looping without ever sleeping. Therefore we have to do a short sleep
4346 * here rather than calling cond_resched().
4348 if (current->flags & PF_WQ_WORKER)
4349 schedule_timeout_uninterruptible(1);
4356 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4359 * It's possible that cpuset's mems_allowed and the nodemask from
4360 * mempolicy don't intersect. This should be normally dealt with by
4361 * policy_nodemask(), but it's possible to race with cpuset update in
4362 * such a way the check therein was true, and then it became false
4363 * before we got our cpuset_mems_cookie here.
4364 * This assumes that for all allocations, ac->nodemask can come only
4365 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4366 * when it does not intersect with the cpuset restrictions) or the
4367 * caller can deal with a violated nodemask.
4369 if (cpusets_enabled() && ac->nodemask &&
4370 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4371 ac->nodemask = NULL;
4376 * When updating a task's mems_allowed or mempolicy nodemask, it is
4377 * possible to race with parallel threads in such a way that our
4378 * allocation can fail while the mask is being updated. If we are about
4379 * to fail, check if the cpuset changed during allocation and if so,
4382 if (read_mems_allowed_retry(cpuset_mems_cookie))
4388 static inline struct page *
4389 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4390 struct alloc_context *ac)
4392 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4393 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4394 struct page *page = NULL;
4395 unsigned int alloc_flags;
4396 unsigned long did_some_progress;
4397 enum compact_priority compact_priority;
4398 enum compact_result compact_result;
4399 int compaction_retries;
4400 int no_progress_loops;
4401 unsigned int cpuset_mems_cookie;
4405 * We also sanity check to catch abuse of atomic reserves being used by
4406 * callers that are not in atomic context.
4408 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4409 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4410 gfp_mask &= ~__GFP_ATOMIC;
4413 compaction_retries = 0;
4414 no_progress_loops = 0;
4415 compact_priority = DEF_COMPACT_PRIORITY;
4416 cpuset_mems_cookie = read_mems_allowed_begin();
4419 * The fast path uses conservative alloc_flags to succeed only until
4420 * kswapd needs to be woken up, and to avoid the cost of setting up
4421 * alloc_flags precisely. So we do that now.
4423 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4426 * We need to recalculate the starting point for the zonelist iterator
4427 * because we might have used different nodemask in the fast path, or
4428 * there was a cpuset modification and we are retrying - otherwise we
4429 * could end up iterating over non-eligible zones endlessly.
4431 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4432 ac->high_zoneidx, ac->nodemask);
4433 if (!ac->preferred_zoneref->zone)
4436 if (alloc_flags & ALLOC_KSWAPD)
4437 wake_all_kswapds(order, gfp_mask, ac);
4440 * The adjusted alloc_flags might result in immediate success, so try
4443 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4448 * For costly allocations, try direct compaction first, as it's likely
4449 * that we have enough base pages and don't need to reclaim. For non-
4450 * movable high-order allocations, do that as well, as compaction will
4451 * try prevent permanent fragmentation by migrating from blocks of the
4453 * Don't try this for allocations that are allowed to ignore
4454 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4456 if (can_direct_reclaim &&
4458 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4459 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4460 page = __alloc_pages_direct_compact(gfp_mask, order,
4462 INIT_COMPACT_PRIORITY,
4468 * Checks for costly allocations with __GFP_NORETRY, which
4469 * includes some THP page fault allocations
4471 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4473 * If allocating entire pageblock(s) and compaction
4474 * failed because all zones are below low watermarks
4475 * or is prohibited because it recently failed at this
4476 * order, fail immediately unless the allocator has
4477 * requested compaction and reclaim retry.
4480 * - potentially very expensive because zones are far
4481 * below their low watermarks or this is part of very
4482 * bursty high order allocations,
4483 * - not guaranteed to help because isolate_freepages()
4484 * may not iterate over freed pages as part of its
4486 * - unlikely to make entire pageblocks free on its
4489 if (compact_result == COMPACT_SKIPPED ||
4490 compact_result == COMPACT_DEFERRED)
4494 * Looks like reclaim/compaction is worth trying, but
4495 * sync compaction could be very expensive, so keep
4496 * using async compaction.
4498 compact_priority = INIT_COMPACT_PRIORITY;
4503 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4504 if (alloc_flags & ALLOC_KSWAPD)
4505 wake_all_kswapds(order, gfp_mask, ac);
4507 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4509 alloc_flags = reserve_flags;
4512 * Reset the nodemask and zonelist iterators if memory policies can be
4513 * ignored. These allocations are high priority and system rather than
4516 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4517 ac->nodemask = NULL;
4518 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4519 ac->high_zoneidx, ac->nodemask);
4522 /* Attempt with potentially adjusted zonelist and alloc_flags */
4523 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4527 /* Caller is not willing to reclaim, we can't balance anything */
4528 if (!can_direct_reclaim)
4531 /* Avoid recursion of direct reclaim */
4532 if (current->flags & PF_MEMALLOC)
4535 /* Try direct reclaim and then allocating */
4536 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4537 &did_some_progress);
4541 /* Try direct compaction and then allocating */
4542 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4543 compact_priority, &compact_result);
4547 /* Do not loop if specifically requested */
4548 if (gfp_mask & __GFP_NORETRY)
4552 * Do not retry costly high order allocations unless they are
4553 * __GFP_RETRY_MAYFAIL
4555 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4558 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4559 did_some_progress > 0, &no_progress_loops))
4563 * It doesn't make any sense to retry for the compaction if the order-0
4564 * reclaim is not able to make any progress because the current
4565 * implementation of the compaction depends on the sufficient amount
4566 * of free memory (see __compaction_suitable)
4568 if (did_some_progress > 0 &&
4569 should_compact_retry(ac, order, alloc_flags,
4570 compact_result, &compact_priority,
4571 &compaction_retries))
4575 /* Deal with possible cpuset update races before we start OOM killing */
4576 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4579 /* Reclaim has failed us, start killing things */
4580 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4584 /* Avoid allocations with no watermarks from looping endlessly */
4585 if (tsk_is_oom_victim(current) &&
4586 (alloc_flags == ALLOC_OOM ||
4587 (gfp_mask & __GFP_NOMEMALLOC)))
4590 /* Retry as long as the OOM killer is making progress */
4591 if (did_some_progress) {
4592 no_progress_loops = 0;
4597 /* Deal with possible cpuset update races before we fail */
4598 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4602 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4605 if (gfp_mask & __GFP_NOFAIL) {
4607 * All existing users of the __GFP_NOFAIL are blockable, so warn
4608 * of any new users that actually require GFP_NOWAIT
4610 if (WARN_ON_ONCE(!can_direct_reclaim))
4614 * PF_MEMALLOC request from this context is rather bizarre
4615 * because we cannot reclaim anything and only can loop waiting
4616 * for somebody to do a work for us
4618 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4621 * non failing costly orders are a hard requirement which we
4622 * are not prepared for much so let's warn about these users
4623 * so that we can identify them and convert them to something
4626 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4629 * Help non-failing allocations by giving them access to memory
4630 * reserves but do not use ALLOC_NO_WATERMARKS because this
4631 * could deplete whole memory reserves which would just make
4632 * the situation worse
4634 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4642 warn_alloc(gfp_mask, ac->nodemask,
4643 "page allocation failure: order:%u", order);
4648 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4649 int preferred_nid, nodemask_t *nodemask,
4650 struct alloc_context *ac, gfp_t *alloc_mask,
4651 unsigned int *alloc_flags)
4653 ac->high_zoneidx = gfp_zone(gfp_mask);
4654 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4655 ac->nodemask = nodemask;
4656 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4658 if (cpusets_enabled()) {
4659 *alloc_mask |= __GFP_HARDWALL;
4661 ac->nodemask = &cpuset_current_mems_allowed;
4663 *alloc_flags |= ALLOC_CPUSET;
4666 fs_reclaim_acquire(gfp_mask);
4667 fs_reclaim_release(gfp_mask);
4669 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4671 if (should_fail_alloc_page(gfp_mask, order))
4674 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4675 *alloc_flags |= ALLOC_CMA;
4680 /* Determine whether to spread dirty pages and what the first usable zone */
4681 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4683 /* Dirty zone balancing only done in the fast path */
4684 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4687 * The preferred zone is used for statistics but crucially it is
4688 * also used as the starting point for the zonelist iterator. It
4689 * may get reset for allocations that ignore memory policies.
4691 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4692 ac->high_zoneidx, ac->nodemask);
4696 * This is the 'heart' of the zoned buddy allocator.
4699 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4700 nodemask_t *nodemask)
4703 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4704 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4705 struct alloc_context ac = { };
4708 * There are several places where we assume that the order value is sane
4709 * so bail out early if the request is out of bound.
4711 if (unlikely(order >= MAX_ORDER)) {
4712 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4716 gfp_mask &= gfp_allowed_mask;
4717 alloc_mask = gfp_mask;
4718 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4721 finalise_ac(gfp_mask, &ac);
4724 * Forbid the first pass from falling back to types that fragment
4725 * memory until all local zones are considered.
4727 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4729 /* First allocation attempt */
4730 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4735 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4736 * resp. GFP_NOIO which has to be inherited for all allocation requests
4737 * from a particular context which has been marked by
4738 * memalloc_no{fs,io}_{save,restore}.
4740 alloc_mask = current_gfp_context(gfp_mask);
4741 ac.spread_dirty_pages = false;
4744 * Restore the original nodemask if it was potentially replaced with
4745 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4747 if (unlikely(ac.nodemask != nodemask))
4748 ac.nodemask = nodemask;
4750 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4753 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4754 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4755 __free_pages(page, order);
4759 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4763 EXPORT_SYMBOL(__alloc_pages_nodemask);
4766 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4767 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4768 * you need to access high mem.
4770 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4774 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4777 return (unsigned long) page_address(page);
4779 EXPORT_SYMBOL(__get_free_pages);
4781 unsigned long get_zeroed_page(gfp_t gfp_mask)
4783 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4785 EXPORT_SYMBOL(get_zeroed_page);
4787 static inline void free_the_page(struct page *page, unsigned int order)
4789 if (order == 0) /* Via pcp? */
4790 free_unref_page(page);
4792 __free_pages_ok(page, order);
4795 void __free_pages(struct page *page, unsigned int order)
4797 if (put_page_testzero(page))
4798 free_the_page(page, order);
4800 EXPORT_SYMBOL(__free_pages);
4802 void free_pages(unsigned long addr, unsigned int order)
4805 VM_BUG_ON(!virt_addr_valid((void *)addr));
4806 __free_pages(virt_to_page((void *)addr), order);
4810 EXPORT_SYMBOL(free_pages);
4814 * An arbitrary-length arbitrary-offset area of memory which resides
4815 * within a 0 or higher order page. Multiple fragments within that page
4816 * are individually refcounted, in the page's reference counter.
4818 * The page_frag functions below provide a simple allocation framework for
4819 * page fragments. This is used by the network stack and network device
4820 * drivers to provide a backing region of memory for use as either an
4821 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4823 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4826 struct page *page = NULL;
4827 gfp_t gfp = gfp_mask;
4829 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4830 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4832 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4833 PAGE_FRAG_CACHE_MAX_ORDER);
4834 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4836 if (unlikely(!page))
4837 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4839 nc->va = page ? page_address(page) : NULL;
4844 void __page_frag_cache_drain(struct page *page, unsigned int count)
4846 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4848 if (page_ref_sub_and_test(page, count))
4849 free_the_page(page, compound_order(page));
4851 EXPORT_SYMBOL(__page_frag_cache_drain);
4853 void *page_frag_alloc(struct page_frag_cache *nc,
4854 unsigned int fragsz, gfp_t gfp_mask)
4856 unsigned int size = PAGE_SIZE;
4860 if (unlikely(!nc->va)) {
4862 page = __page_frag_cache_refill(nc, gfp_mask);
4866 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4867 /* if size can vary use size else just use PAGE_SIZE */
4870 /* Even if we own the page, we do not use atomic_set().
4871 * This would break get_page_unless_zero() users.
4873 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4875 /* reset page count bias and offset to start of new frag */
4876 nc->pfmemalloc = page_is_pfmemalloc(page);
4877 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4881 offset = nc->offset - fragsz;
4882 if (unlikely(offset < 0)) {
4883 page = virt_to_page(nc->va);
4885 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4888 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4889 /* if size can vary use size else just use PAGE_SIZE */
4892 /* OK, page count is 0, we can safely set it */
4893 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4895 /* reset page count bias and offset to start of new frag */
4896 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4897 offset = size - fragsz;
4901 nc->offset = offset;
4903 return nc->va + offset;
4905 EXPORT_SYMBOL(page_frag_alloc);
4908 * Frees a page fragment allocated out of either a compound or order 0 page.
4910 void page_frag_free(void *addr)
4912 struct page *page = virt_to_head_page(addr);
4914 if (unlikely(put_page_testzero(page)))
4915 free_the_page(page, compound_order(page));
4917 EXPORT_SYMBOL(page_frag_free);
4919 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4923 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4924 unsigned long used = addr + PAGE_ALIGN(size);
4926 split_page(virt_to_page((void *)addr), order);
4927 while (used < alloc_end) {
4932 return (void *)addr;
4936 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4937 * @size: the number of bytes to allocate
4938 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4940 * This function is similar to alloc_pages(), except that it allocates the
4941 * minimum number of pages to satisfy the request. alloc_pages() can only
4942 * allocate memory in power-of-two pages.
4944 * This function is also limited by MAX_ORDER.
4946 * Memory allocated by this function must be released by free_pages_exact().
4948 * Return: pointer to the allocated area or %NULL in case of error.
4950 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4952 unsigned int order = get_order(size);
4955 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4956 gfp_mask &= ~__GFP_COMP;
4958 addr = __get_free_pages(gfp_mask, order);
4959 return make_alloc_exact(addr, order, size);
4961 EXPORT_SYMBOL(alloc_pages_exact);
4964 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4966 * @nid: the preferred node ID where memory should be allocated
4967 * @size: the number of bytes to allocate
4968 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4970 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4973 * Return: pointer to the allocated area or %NULL in case of error.
4975 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4977 unsigned int order = get_order(size);
4980 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4981 gfp_mask &= ~__GFP_COMP;
4983 p = alloc_pages_node(nid, gfp_mask, order);
4986 return make_alloc_exact((unsigned long)page_address(p), order, size);
4990 * free_pages_exact - release memory allocated via alloc_pages_exact()
4991 * @virt: the value returned by alloc_pages_exact.
4992 * @size: size of allocation, same value as passed to alloc_pages_exact().
4994 * Release the memory allocated by a previous call to alloc_pages_exact.
4996 void free_pages_exact(void *virt, size_t size)
4998 unsigned long addr = (unsigned long)virt;
4999 unsigned long end = addr + PAGE_ALIGN(size);
5001 while (addr < end) {
5006 EXPORT_SYMBOL(free_pages_exact);
5009 * nr_free_zone_pages - count number of pages beyond high watermark
5010 * @offset: The zone index of the highest zone
5012 * nr_free_zone_pages() counts the number of pages which are beyond the
5013 * high watermark within all zones at or below a given zone index. For each
5014 * zone, the number of pages is calculated as:
5016 * nr_free_zone_pages = managed_pages - high_pages
5018 * Return: number of pages beyond high watermark.
5020 static unsigned long nr_free_zone_pages(int offset)
5025 /* Just pick one node, since fallback list is circular */
5026 unsigned long sum = 0;
5028 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5030 for_each_zone_zonelist(zone, z, zonelist, offset) {
5031 unsigned long size = zone_managed_pages(zone);
5032 unsigned long high = high_wmark_pages(zone);
5041 * nr_free_buffer_pages - count number of pages beyond high watermark
5043 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5044 * watermark within ZONE_DMA and ZONE_NORMAL.
5046 * Return: number of pages beyond high watermark within ZONE_DMA and
5049 unsigned long nr_free_buffer_pages(void)
5051 return nr_free_zone_pages(gfp_zone(GFP_USER));
5053 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5056 * nr_free_pagecache_pages - count number of pages beyond high watermark
5058 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5059 * high watermark within all zones.
5061 * Return: number of pages beyond high watermark within all zones.
5063 unsigned long nr_free_pagecache_pages(void)
5065 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5068 static inline void show_node(struct zone *zone)
5070 if (IS_ENABLED(CONFIG_NUMA))
5071 printk("Node %d ", zone_to_nid(zone));
5074 long si_mem_available(void)
5077 unsigned long pagecache;
5078 unsigned long wmark_low = 0;
5079 unsigned long pages[NR_LRU_LISTS];
5080 unsigned long reclaimable;
5084 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5085 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5088 wmark_low += low_wmark_pages(zone);
5091 * Estimate the amount of memory available for userspace allocations,
5092 * without causing swapping.
5094 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5097 * Not all the page cache can be freed, otherwise the system will
5098 * start swapping. Assume at least half of the page cache, or the
5099 * low watermark worth of cache, needs to stay.
5101 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5102 pagecache -= min(pagecache / 2, wmark_low);
5103 available += pagecache;
5106 * Part of the reclaimable slab and other kernel memory consists of
5107 * items that are in use, and cannot be freed. Cap this estimate at the
5110 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5111 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5112 available += reclaimable - min(reclaimable / 2, wmark_low);
5118 EXPORT_SYMBOL_GPL(si_mem_available);
5120 void si_meminfo(struct sysinfo *val)
5122 val->totalram = totalram_pages();
5123 val->sharedram = global_node_page_state(NR_SHMEM);
5124 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5125 val->bufferram = nr_blockdev_pages();
5126 val->totalhigh = totalhigh_pages();
5127 val->freehigh = nr_free_highpages();
5128 val->mem_unit = PAGE_SIZE;
5131 EXPORT_SYMBOL(si_meminfo);
5134 void si_meminfo_node(struct sysinfo *val, int nid)
5136 int zone_type; /* needs to be signed */
5137 unsigned long managed_pages = 0;
5138 unsigned long managed_highpages = 0;
5139 unsigned long free_highpages = 0;
5140 pg_data_t *pgdat = NODE_DATA(nid);
5142 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5143 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5144 val->totalram = managed_pages;
5145 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5146 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5147 #ifdef CONFIG_HIGHMEM
5148 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5149 struct zone *zone = &pgdat->node_zones[zone_type];
5151 if (is_highmem(zone)) {
5152 managed_highpages += zone_managed_pages(zone);
5153 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5156 val->totalhigh = managed_highpages;
5157 val->freehigh = free_highpages;
5159 val->totalhigh = managed_highpages;
5160 val->freehigh = free_highpages;
5162 val->mem_unit = PAGE_SIZE;
5167 * Determine whether the node should be displayed or not, depending on whether
5168 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5170 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5172 if (!(flags & SHOW_MEM_FILTER_NODES))
5176 * no node mask - aka implicit memory numa policy. Do not bother with
5177 * the synchronization - read_mems_allowed_begin - because we do not
5178 * have to be precise here.
5181 nodemask = &cpuset_current_mems_allowed;
5183 return !node_isset(nid, *nodemask);
5186 #define K(x) ((x) << (PAGE_SHIFT-10))
5188 static void show_migration_types(unsigned char type)
5190 static const char types[MIGRATE_TYPES] = {
5191 [MIGRATE_UNMOVABLE] = 'U',
5192 [MIGRATE_MOVABLE] = 'M',
5193 [MIGRATE_RECLAIMABLE] = 'E',
5194 [MIGRATE_HIGHATOMIC] = 'H',
5196 [MIGRATE_CMA] = 'C',
5198 #ifdef CONFIG_MEMORY_ISOLATION
5199 [MIGRATE_ISOLATE] = 'I',
5202 char tmp[MIGRATE_TYPES + 1];
5206 for (i = 0; i < MIGRATE_TYPES; i++) {
5207 if (type & (1 << i))
5212 printk(KERN_CONT "(%s) ", tmp);
5216 * Show free area list (used inside shift_scroll-lock stuff)
5217 * We also calculate the percentage fragmentation. We do this by counting the
5218 * memory on each free list with the exception of the first item on the list.
5221 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5224 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5226 unsigned long free_pcp = 0;
5231 for_each_populated_zone(zone) {
5232 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5235 for_each_online_cpu(cpu)
5236 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5239 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5240 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5241 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5242 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5243 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5244 " free:%lu free_pcp:%lu free_cma:%lu\n",
5245 global_node_page_state(NR_ACTIVE_ANON),
5246 global_node_page_state(NR_INACTIVE_ANON),
5247 global_node_page_state(NR_ISOLATED_ANON),
5248 global_node_page_state(NR_ACTIVE_FILE),
5249 global_node_page_state(NR_INACTIVE_FILE),
5250 global_node_page_state(NR_ISOLATED_FILE),
5251 global_node_page_state(NR_UNEVICTABLE),
5252 global_node_page_state(NR_FILE_DIRTY),
5253 global_node_page_state(NR_WRITEBACK),
5254 global_node_page_state(NR_UNSTABLE_NFS),
5255 global_node_page_state(NR_SLAB_RECLAIMABLE),
5256 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5257 global_node_page_state(NR_FILE_MAPPED),
5258 global_node_page_state(NR_SHMEM),
5259 global_zone_page_state(NR_PAGETABLE),
5260 global_zone_page_state(NR_BOUNCE),
5261 global_zone_page_state(NR_FREE_PAGES),
5263 global_zone_page_state(NR_FREE_CMA_PAGES));
5265 for_each_online_pgdat(pgdat) {
5266 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5270 " active_anon:%lukB"
5271 " inactive_anon:%lukB"
5272 " active_file:%lukB"
5273 " inactive_file:%lukB"
5274 " unevictable:%lukB"
5275 " isolated(anon):%lukB"
5276 " isolated(file):%lukB"
5281 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5283 " shmem_pmdmapped: %lukB"
5286 " writeback_tmp:%lukB"
5288 " all_unreclaimable? %s"
5291 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5292 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5293 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5294 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5295 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5296 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5297 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5298 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5299 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5300 K(node_page_state(pgdat, NR_WRITEBACK)),
5301 K(node_page_state(pgdat, NR_SHMEM)),
5302 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5303 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5304 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5306 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5308 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5309 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5310 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5314 for_each_populated_zone(zone) {
5317 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5321 for_each_online_cpu(cpu)
5322 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5331 " reserved_highatomic:%luKB"
5332 " active_anon:%lukB"
5333 " inactive_anon:%lukB"
5334 " active_file:%lukB"
5335 " inactive_file:%lukB"
5336 " unevictable:%lukB"
5337 " writepending:%lukB"
5341 " kernel_stack:%lukB"
5349 K(zone_page_state(zone, NR_FREE_PAGES)),
5350 K(min_wmark_pages(zone)),
5351 K(low_wmark_pages(zone)),
5352 K(high_wmark_pages(zone)),
5353 K(zone->nr_reserved_highatomic),
5354 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5355 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5356 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5357 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5358 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5359 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5360 K(zone->present_pages),
5361 K(zone_managed_pages(zone)),
5362 K(zone_page_state(zone, NR_MLOCK)),
5363 zone_page_state(zone, NR_KERNEL_STACK_KB),
5364 K(zone_page_state(zone, NR_PAGETABLE)),
5365 K(zone_page_state(zone, NR_BOUNCE)),
5367 K(this_cpu_read(zone->pageset->pcp.count)),
5368 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5369 printk("lowmem_reserve[]:");
5370 for (i = 0; i < MAX_NR_ZONES; i++)
5371 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5372 printk(KERN_CONT "\n");
5375 for_each_populated_zone(zone) {
5377 unsigned long nr[MAX_ORDER], flags, total = 0;
5378 unsigned char types[MAX_ORDER];
5380 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5383 printk(KERN_CONT "%s: ", zone->name);
5385 spin_lock_irqsave(&zone->lock, flags);
5386 for (order = 0; order < MAX_ORDER; order++) {
5387 struct free_area *area = &zone->free_area[order];
5390 nr[order] = area->nr_free;
5391 total += nr[order] << order;
5394 for (type = 0; type < MIGRATE_TYPES; type++) {
5395 if (!free_area_empty(area, type))
5396 types[order] |= 1 << type;
5399 spin_unlock_irqrestore(&zone->lock, flags);
5400 for (order = 0; order < MAX_ORDER; order++) {
5401 printk(KERN_CONT "%lu*%lukB ",
5402 nr[order], K(1UL) << order);
5404 show_migration_types(types[order]);
5406 printk(KERN_CONT "= %lukB\n", K(total));
5409 hugetlb_show_meminfo();
5411 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5413 show_swap_cache_info();
5416 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5418 zoneref->zone = zone;
5419 zoneref->zone_idx = zone_idx(zone);
5423 * Builds allocation fallback zone lists.
5425 * Add all populated zones of a node to the zonelist.
5427 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5430 enum zone_type zone_type = MAX_NR_ZONES;
5435 zone = pgdat->node_zones + zone_type;
5436 if (managed_zone(zone)) {
5437 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5438 check_highest_zone(zone_type);
5440 } while (zone_type);
5447 static int __parse_numa_zonelist_order(char *s)
5450 * We used to support different zonlists modes but they turned
5451 * out to be just not useful. Let's keep the warning in place
5452 * if somebody still use the cmd line parameter so that we do
5453 * not fail it silently
5455 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5456 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5462 static __init int setup_numa_zonelist_order(char *s)
5467 return __parse_numa_zonelist_order(s);
5469 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5471 char numa_zonelist_order[] = "Node";
5474 * sysctl handler for numa_zonelist_order
5476 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5477 void __user *buffer, size_t *length,
5484 return proc_dostring(table, write, buffer, length, ppos);
5485 str = memdup_user_nul(buffer, 16);
5487 return PTR_ERR(str);
5489 ret = __parse_numa_zonelist_order(str);
5495 #define MAX_NODE_LOAD (nr_online_nodes)
5496 static int node_load[MAX_NUMNODES];
5499 * find_next_best_node - find the next node that should appear in a given node's fallback list
5500 * @node: node whose fallback list we're appending
5501 * @used_node_mask: nodemask_t of already used nodes
5503 * We use a number of factors to determine which is the next node that should
5504 * appear on a given node's fallback list. The node should not have appeared
5505 * already in @node's fallback list, and it should be the next closest node
5506 * according to the distance array (which contains arbitrary distance values
5507 * from each node to each node in the system), and should also prefer nodes
5508 * with no CPUs, since presumably they'll have very little allocation pressure
5509 * on them otherwise.
5511 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5513 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5516 int min_val = INT_MAX;
5517 int best_node = NUMA_NO_NODE;
5518 const struct cpumask *tmp = cpumask_of_node(0);
5520 /* Use the local node if we haven't already */
5521 if (!node_isset(node, *used_node_mask)) {
5522 node_set(node, *used_node_mask);
5526 for_each_node_state(n, N_MEMORY) {
5528 /* Don't want a node to appear more than once */
5529 if (node_isset(n, *used_node_mask))
5532 /* Use the distance array to find the distance */
5533 val = node_distance(node, n);
5535 /* Penalize nodes under us ("prefer the next node") */
5538 /* Give preference to headless and unused nodes */
5539 tmp = cpumask_of_node(n);
5540 if (!cpumask_empty(tmp))
5541 val += PENALTY_FOR_NODE_WITH_CPUS;
5543 /* Slight preference for less loaded node */
5544 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5545 val += node_load[n];
5547 if (val < min_val) {
5554 node_set(best_node, *used_node_mask);
5561 * Build zonelists ordered by node and zones within node.
5562 * This results in maximum locality--normal zone overflows into local
5563 * DMA zone, if any--but risks exhausting DMA zone.
5565 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5568 struct zoneref *zonerefs;
5571 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5573 for (i = 0; i < nr_nodes; i++) {
5576 pg_data_t *node = NODE_DATA(node_order[i]);
5578 nr_zones = build_zonerefs_node(node, zonerefs);
5579 zonerefs += nr_zones;
5581 zonerefs->zone = NULL;
5582 zonerefs->zone_idx = 0;
5586 * Build gfp_thisnode zonelists
5588 static void build_thisnode_zonelists(pg_data_t *pgdat)
5590 struct zoneref *zonerefs;
5593 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5594 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5595 zonerefs += nr_zones;
5596 zonerefs->zone = NULL;
5597 zonerefs->zone_idx = 0;
5601 * Build zonelists ordered by zone and nodes within zones.
5602 * This results in conserving DMA zone[s] until all Normal memory is
5603 * exhausted, but results in overflowing to remote node while memory
5604 * may still exist in local DMA zone.
5607 static void build_zonelists(pg_data_t *pgdat)
5609 static int node_order[MAX_NUMNODES];
5610 int node, load, nr_nodes = 0;
5611 nodemask_t used_mask;
5612 int local_node, prev_node;
5614 /* NUMA-aware ordering of nodes */
5615 local_node = pgdat->node_id;
5616 load = nr_online_nodes;
5617 prev_node = local_node;
5618 nodes_clear(used_mask);
5620 memset(node_order, 0, sizeof(node_order));
5621 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5623 * We don't want to pressure a particular node.
5624 * So adding penalty to the first node in same
5625 * distance group to make it round-robin.
5627 if (node_distance(local_node, node) !=
5628 node_distance(local_node, prev_node))
5629 node_load[node] = load;
5631 node_order[nr_nodes++] = node;
5636 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5637 build_thisnode_zonelists(pgdat);
5640 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5642 * Return node id of node used for "local" allocations.
5643 * I.e., first node id of first zone in arg node's generic zonelist.
5644 * Used for initializing percpu 'numa_mem', which is used primarily
5645 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5647 int local_memory_node(int node)
5651 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5652 gfp_zone(GFP_KERNEL),
5654 return zone_to_nid(z->zone);
5658 static void setup_min_unmapped_ratio(void);
5659 static void setup_min_slab_ratio(void);
5660 #else /* CONFIG_NUMA */
5662 static void build_zonelists(pg_data_t *pgdat)
5664 int node, local_node;
5665 struct zoneref *zonerefs;
5668 local_node = pgdat->node_id;
5670 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5671 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5672 zonerefs += nr_zones;
5675 * Now we build the zonelist so that it contains the zones
5676 * of all the other nodes.
5677 * We don't want to pressure a particular node, so when
5678 * building the zones for node N, we make sure that the
5679 * zones coming right after the local ones are those from
5680 * node N+1 (modulo N)
5682 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5683 if (!node_online(node))
5685 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5686 zonerefs += nr_zones;
5688 for (node = 0; node < local_node; node++) {
5689 if (!node_online(node))
5691 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5692 zonerefs += nr_zones;
5695 zonerefs->zone = NULL;
5696 zonerefs->zone_idx = 0;
5699 #endif /* CONFIG_NUMA */
5702 * Boot pageset table. One per cpu which is going to be used for all
5703 * zones and all nodes. The parameters will be set in such a way
5704 * that an item put on a list will immediately be handed over to
5705 * the buddy list. This is safe since pageset manipulation is done
5706 * with interrupts disabled.
5708 * The boot_pagesets must be kept even after bootup is complete for
5709 * unused processors and/or zones. They do play a role for bootstrapping
5710 * hotplugged processors.
5712 * zoneinfo_show() and maybe other functions do
5713 * not check if the processor is online before following the pageset pointer.
5714 * Other parts of the kernel may not check if the zone is available.
5716 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5717 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5718 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5720 static void __build_all_zonelists(void *data)
5723 int __maybe_unused cpu;
5724 pg_data_t *self = data;
5725 static DEFINE_SPINLOCK(lock);
5730 memset(node_load, 0, sizeof(node_load));
5734 * This node is hotadded and no memory is yet present. So just
5735 * building zonelists is fine - no need to touch other nodes.
5737 if (self && !node_online(self->node_id)) {
5738 build_zonelists(self);
5740 for_each_online_node(nid) {
5741 pg_data_t *pgdat = NODE_DATA(nid);
5743 build_zonelists(pgdat);
5746 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5748 * We now know the "local memory node" for each node--
5749 * i.e., the node of the first zone in the generic zonelist.
5750 * Set up numa_mem percpu variable for on-line cpus. During
5751 * boot, only the boot cpu should be on-line; we'll init the
5752 * secondary cpus' numa_mem as they come on-line. During
5753 * node/memory hotplug, we'll fixup all on-line cpus.
5755 for_each_online_cpu(cpu)
5756 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5763 static noinline void __init
5764 build_all_zonelists_init(void)
5768 __build_all_zonelists(NULL);
5771 * Initialize the boot_pagesets that are going to be used
5772 * for bootstrapping processors. The real pagesets for
5773 * each zone will be allocated later when the per cpu
5774 * allocator is available.
5776 * boot_pagesets are used also for bootstrapping offline
5777 * cpus if the system is already booted because the pagesets
5778 * are needed to initialize allocators on a specific cpu too.
5779 * F.e. the percpu allocator needs the page allocator which
5780 * needs the percpu allocator in order to allocate its pagesets
5781 * (a chicken-egg dilemma).
5783 for_each_possible_cpu(cpu)
5784 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5786 mminit_verify_zonelist();
5787 cpuset_init_current_mems_allowed();
5791 * unless system_state == SYSTEM_BOOTING.
5793 * __ref due to call of __init annotated helper build_all_zonelists_init
5794 * [protected by SYSTEM_BOOTING].
5796 void __ref build_all_zonelists(pg_data_t *pgdat)
5798 if (system_state == SYSTEM_BOOTING) {
5799 build_all_zonelists_init();
5801 __build_all_zonelists(pgdat);
5802 /* cpuset refresh routine should be here */
5804 vm_total_pages = nr_free_pagecache_pages();
5806 * Disable grouping by mobility if the number of pages in the
5807 * system is too low to allow the mechanism to work. It would be
5808 * more accurate, but expensive to check per-zone. This check is
5809 * made on memory-hotadd so a system can start with mobility
5810 * disabled and enable it later
5812 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5813 page_group_by_mobility_disabled = 1;
5815 page_group_by_mobility_disabled = 0;
5817 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5819 page_group_by_mobility_disabled ? "off" : "on",
5822 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5826 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5827 static bool __meminit
5828 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5830 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5831 static struct memblock_region *r;
5833 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5834 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5835 for_each_memblock(memory, r) {
5836 if (*pfn < memblock_region_memory_end_pfn(r))
5840 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5841 memblock_is_mirror(r)) {
5842 *pfn = memblock_region_memory_end_pfn(r);
5850 #ifdef CONFIG_SPARSEMEM
5851 /* Skip PFNs that belong to non-present sections */
5852 static inline __meminit unsigned long next_pfn(unsigned long pfn)
5854 const unsigned long section_nr = pfn_to_section_nr(++pfn);
5856 if (present_section_nr(section_nr))
5858 return section_nr_to_pfn(next_present_section_nr(section_nr));
5861 static inline __meminit unsigned long next_pfn(unsigned long pfn)
5868 * Initially all pages are reserved - free ones are freed
5869 * up by memblock_free_all() once the early boot process is
5870 * done. Non-atomic initialization, single-pass.
5872 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5873 unsigned long start_pfn, enum memmap_context context,
5874 struct vmem_altmap *altmap)
5876 unsigned long pfn, end_pfn = start_pfn + size;
5879 if (highest_memmap_pfn < end_pfn - 1)
5880 highest_memmap_pfn = end_pfn - 1;
5882 #ifdef CONFIG_ZONE_DEVICE
5884 * Honor reservation requested by the driver for this ZONE_DEVICE
5885 * memory. We limit the total number of pages to initialize to just
5886 * those that might contain the memory mapping. We will defer the
5887 * ZONE_DEVICE page initialization until after we have released
5890 if (zone == ZONE_DEVICE) {
5894 if (start_pfn == altmap->base_pfn)
5895 start_pfn += altmap->reserve;
5896 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5900 for (pfn = start_pfn; pfn < end_pfn; ) {
5902 * There can be holes in boot-time mem_map[]s handed to this
5903 * function. They do not exist on hotplugged memory.
5905 if (context == MEMMAP_EARLY) {
5906 if (!early_pfn_valid(pfn)) {
5907 pfn = next_pfn(pfn);
5910 if (!early_pfn_in_nid(pfn, nid)) {
5914 if (overlap_memmap_init(zone, &pfn))
5916 if (defer_init(nid, pfn, end_pfn))
5920 page = pfn_to_page(pfn);
5921 __init_single_page(page, pfn, zone, nid);
5922 if (context == MEMMAP_HOTPLUG)
5923 __SetPageReserved(page);
5926 * Mark the block movable so that blocks are reserved for
5927 * movable at startup. This will force kernel allocations
5928 * to reserve their blocks rather than leaking throughout
5929 * the address space during boot when many long-lived
5930 * kernel allocations are made.
5932 * bitmap is created for zone's valid pfn range. but memmap
5933 * can be created for invalid pages (for alignment)
5934 * check here not to call set_pageblock_migratetype() against
5937 if (!(pfn & (pageblock_nr_pages - 1))) {
5938 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5945 #ifdef CONFIG_ZONE_DEVICE
5946 void __ref memmap_init_zone_device(struct zone *zone,
5947 unsigned long start_pfn,
5948 unsigned long nr_pages,
5949 struct dev_pagemap *pgmap)
5951 unsigned long pfn, end_pfn = start_pfn + nr_pages;
5952 struct pglist_data *pgdat = zone->zone_pgdat;
5953 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
5954 unsigned long zone_idx = zone_idx(zone);
5955 unsigned long start = jiffies;
5956 int nid = pgdat->node_id;
5958 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
5962 * The call to memmap_init_zone should have already taken care
5963 * of the pages reserved for the memmap, so we can just jump to
5964 * the end of that region and start processing the device pages.
5967 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5968 nr_pages = end_pfn - start_pfn;
5971 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5972 struct page *page = pfn_to_page(pfn);
5974 __init_single_page(page, pfn, zone_idx, nid);
5977 * Mark page reserved as it will need to wait for onlining
5978 * phase for it to be fully associated with a zone.
5980 * We can use the non-atomic __set_bit operation for setting
5981 * the flag as we are still initializing the pages.
5983 __SetPageReserved(page);
5986 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5987 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5988 * ever freed or placed on a driver-private list.
5990 page->pgmap = pgmap;
5991 page->zone_device_data = NULL;
5994 * Mark the block movable so that blocks are reserved for
5995 * movable at startup. This will force kernel allocations
5996 * to reserve their blocks rather than leaking throughout
5997 * the address space during boot when many long-lived
5998 * kernel allocations are made.
6000 * bitmap is created for zone's valid pfn range. but memmap
6001 * can be created for invalid pages (for alignment)
6002 * check here not to call set_pageblock_migratetype() against
6005 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6006 * because this is done early in section_activate()
6008 if (!(pfn & (pageblock_nr_pages - 1))) {
6009 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6014 pr_info("%s initialised %lu pages in %ums\n", __func__,
6015 nr_pages, jiffies_to_msecs(jiffies - start));
6019 static void __meminit zone_init_free_lists(struct zone *zone)
6021 unsigned int order, t;
6022 for_each_migratetype_order(order, t) {
6023 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6024 zone->free_area[order].nr_free = 0;
6028 void __meminit __weak memmap_init(unsigned long size, int nid,
6029 unsigned long zone, unsigned long start_pfn)
6031 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6034 static int zone_batchsize(struct zone *zone)
6040 * The per-cpu-pages pools are set to around 1000th of the
6043 batch = zone_managed_pages(zone) / 1024;
6044 /* But no more than a meg. */
6045 if (batch * PAGE_SIZE > 1024 * 1024)
6046 batch = (1024 * 1024) / PAGE_SIZE;
6047 batch /= 4; /* We effectively *= 4 below */
6052 * Clamp the batch to a 2^n - 1 value. Having a power
6053 * of 2 value was found to be more likely to have
6054 * suboptimal cache aliasing properties in some cases.
6056 * For example if 2 tasks are alternately allocating
6057 * batches of pages, one task can end up with a lot
6058 * of pages of one half of the possible page colors
6059 * and the other with pages of the other colors.
6061 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6066 /* The deferral and batching of frees should be suppressed under NOMMU
6069 * The problem is that NOMMU needs to be able to allocate large chunks
6070 * of contiguous memory as there's no hardware page translation to
6071 * assemble apparent contiguous memory from discontiguous pages.
6073 * Queueing large contiguous runs of pages for batching, however,
6074 * causes the pages to actually be freed in smaller chunks. As there
6075 * can be a significant delay between the individual batches being
6076 * recycled, this leads to the once large chunks of space being
6077 * fragmented and becoming unavailable for high-order allocations.
6084 * pcp->high and pcp->batch values are related and dependent on one another:
6085 * ->batch must never be higher then ->high.
6086 * The following function updates them in a safe manner without read side
6089 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6090 * those fields changing asynchronously (acording the the above rule).
6092 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6093 * outside of boot time (or some other assurance that no concurrent updaters
6096 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6097 unsigned long batch)
6099 /* start with a fail safe value for batch */
6103 /* Update high, then batch, in order */
6110 /* a companion to pageset_set_high() */
6111 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6113 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6116 static void pageset_init(struct per_cpu_pageset *p)
6118 struct per_cpu_pages *pcp;
6121 memset(p, 0, sizeof(*p));
6124 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6125 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6128 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6131 pageset_set_batch(p, batch);
6135 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6136 * to the value high for the pageset p.
6138 static void pageset_set_high(struct per_cpu_pageset *p,
6141 unsigned long batch = max(1UL, high / 4);
6142 if ((high / 4) > (PAGE_SHIFT * 8))
6143 batch = PAGE_SHIFT * 8;
6145 pageset_update(&p->pcp, high, batch);
6148 static void pageset_set_high_and_batch(struct zone *zone,
6149 struct per_cpu_pageset *pcp)
6151 if (percpu_pagelist_fraction)
6152 pageset_set_high(pcp,
6153 (zone_managed_pages(zone) /
6154 percpu_pagelist_fraction));
6156 pageset_set_batch(pcp, zone_batchsize(zone));
6159 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6161 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6164 pageset_set_high_and_batch(zone, pcp);
6167 void __meminit setup_zone_pageset(struct zone *zone)
6170 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6171 for_each_possible_cpu(cpu)
6172 zone_pageset_init(zone, cpu);
6176 * Allocate per cpu pagesets and initialize them.
6177 * Before this call only boot pagesets were available.
6179 void __init setup_per_cpu_pageset(void)
6181 struct pglist_data *pgdat;
6184 for_each_populated_zone(zone)
6185 setup_zone_pageset(zone);
6187 for_each_online_pgdat(pgdat)
6188 pgdat->per_cpu_nodestats =
6189 alloc_percpu(struct per_cpu_nodestat);
6192 static __meminit void zone_pcp_init(struct zone *zone)
6195 * per cpu subsystem is not up at this point. The following code
6196 * relies on the ability of the linker to provide the
6197 * offset of a (static) per cpu variable into the per cpu area.
6199 zone->pageset = &boot_pageset;
6201 if (populated_zone(zone))
6202 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6203 zone->name, zone->present_pages,
6204 zone_batchsize(zone));
6207 void __meminit init_currently_empty_zone(struct zone *zone,
6208 unsigned long zone_start_pfn,
6211 struct pglist_data *pgdat = zone->zone_pgdat;
6212 int zone_idx = zone_idx(zone) + 1;
6214 if (zone_idx > pgdat->nr_zones)
6215 pgdat->nr_zones = zone_idx;
6217 zone->zone_start_pfn = zone_start_pfn;
6219 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6220 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6222 (unsigned long)zone_idx(zone),
6223 zone_start_pfn, (zone_start_pfn + size));
6225 zone_init_free_lists(zone);
6226 zone->initialized = 1;
6229 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6230 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6233 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6235 int __meminit __early_pfn_to_nid(unsigned long pfn,
6236 struct mminit_pfnnid_cache *state)
6238 unsigned long start_pfn, end_pfn;
6241 if (state->last_start <= pfn && pfn < state->last_end)
6242 return state->last_nid;
6244 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6245 if (nid != NUMA_NO_NODE) {
6246 state->last_start = start_pfn;
6247 state->last_end = end_pfn;
6248 state->last_nid = nid;
6253 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6256 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6257 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6258 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6260 * If an architecture guarantees that all ranges registered contain no holes
6261 * and may be freed, this this function may be used instead of calling
6262 * memblock_free_early_nid() manually.
6264 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6266 unsigned long start_pfn, end_pfn;
6269 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6270 start_pfn = min(start_pfn, max_low_pfn);
6271 end_pfn = min(end_pfn, max_low_pfn);
6273 if (start_pfn < end_pfn)
6274 memblock_free_early_nid(PFN_PHYS(start_pfn),
6275 (end_pfn - start_pfn) << PAGE_SHIFT,
6281 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6282 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6284 * If an architecture guarantees that all ranges registered contain no holes and may
6285 * be freed, this function may be used instead of calling memory_present() manually.
6287 void __init sparse_memory_present_with_active_regions(int nid)
6289 unsigned long start_pfn, end_pfn;
6292 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6293 memory_present(this_nid, start_pfn, end_pfn);
6297 * get_pfn_range_for_nid - Return the start and end page frames for a node
6298 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6299 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6300 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6302 * It returns the start and end page frame of a node based on information
6303 * provided by memblock_set_node(). If called for a node
6304 * with no available memory, a warning is printed and the start and end
6307 void __init get_pfn_range_for_nid(unsigned int nid,
6308 unsigned long *start_pfn, unsigned long *end_pfn)
6310 unsigned long this_start_pfn, this_end_pfn;
6316 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6317 *start_pfn = min(*start_pfn, this_start_pfn);
6318 *end_pfn = max(*end_pfn, this_end_pfn);
6321 if (*start_pfn == -1UL)
6326 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6327 * assumption is made that zones within a node are ordered in monotonic
6328 * increasing memory addresses so that the "highest" populated zone is used
6330 static void __init find_usable_zone_for_movable(void)
6333 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6334 if (zone_index == ZONE_MOVABLE)
6337 if (arch_zone_highest_possible_pfn[zone_index] >
6338 arch_zone_lowest_possible_pfn[zone_index])
6342 VM_BUG_ON(zone_index == -1);
6343 movable_zone = zone_index;
6347 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6348 * because it is sized independent of architecture. Unlike the other zones,
6349 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6350 * in each node depending on the size of each node and how evenly kernelcore
6351 * is distributed. This helper function adjusts the zone ranges
6352 * provided by the architecture for a given node by using the end of the
6353 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6354 * zones within a node are in order of monotonic increases memory addresses
6356 static void __init adjust_zone_range_for_zone_movable(int nid,
6357 unsigned long zone_type,
6358 unsigned long node_start_pfn,
6359 unsigned long node_end_pfn,
6360 unsigned long *zone_start_pfn,
6361 unsigned long *zone_end_pfn)
6363 /* Only adjust if ZONE_MOVABLE is on this node */
6364 if (zone_movable_pfn[nid]) {
6365 /* Size ZONE_MOVABLE */
6366 if (zone_type == ZONE_MOVABLE) {
6367 *zone_start_pfn = zone_movable_pfn[nid];
6368 *zone_end_pfn = min(node_end_pfn,
6369 arch_zone_highest_possible_pfn[movable_zone]);
6371 /* Adjust for ZONE_MOVABLE starting within this range */
6372 } else if (!mirrored_kernelcore &&
6373 *zone_start_pfn < zone_movable_pfn[nid] &&
6374 *zone_end_pfn > zone_movable_pfn[nid]) {
6375 *zone_end_pfn = zone_movable_pfn[nid];
6377 /* Check if this whole range is within ZONE_MOVABLE */
6378 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6379 *zone_start_pfn = *zone_end_pfn;
6384 * Return the number of pages a zone spans in a node, including holes
6385 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6387 static unsigned long __init zone_spanned_pages_in_node(int nid,
6388 unsigned long zone_type,
6389 unsigned long node_start_pfn,
6390 unsigned long node_end_pfn,
6391 unsigned long *zone_start_pfn,
6392 unsigned long *zone_end_pfn,
6393 unsigned long *ignored)
6395 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6396 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6397 /* When hotadd a new node from cpu_up(), the node should be empty */
6398 if (!node_start_pfn && !node_end_pfn)
6401 /* Get the start and end of the zone */
6402 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6403 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6404 adjust_zone_range_for_zone_movable(nid, zone_type,
6405 node_start_pfn, node_end_pfn,
6406 zone_start_pfn, zone_end_pfn);
6408 /* Check that this node has pages within the zone's required range */
6409 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6412 /* Move the zone boundaries inside the node if necessary */
6413 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6414 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6416 /* Return the spanned pages */
6417 return *zone_end_pfn - *zone_start_pfn;
6421 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6422 * then all holes in the requested range will be accounted for.
6424 unsigned long __init __absent_pages_in_range(int nid,
6425 unsigned long range_start_pfn,
6426 unsigned long range_end_pfn)
6428 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6429 unsigned long start_pfn, end_pfn;
6432 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6433 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6434 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6435 nr_absent -= end_pfn - start_pfn;
6441 * absent_pages_in_range - Return number of page frames in holes within a range
6442 * @start_pfn: The start PFN to start searching for holes
6443 * @end_pfn: The end PFN to stop searching for holes
6445 * Return: the number of pages frames in memory holes within a range.
6447 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6448 unsigned long end_pfn)
6450 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6453 /* Return the number of page frames in holes in a zone on a node */
6454 static unsigned long __init zone_absent_pages_in_node(int nid,
6455 unsigned long zone_type,
6456 unsigned long node_start_pfn,
6457 unsigned long node_end_pfn,
6458 unsigned long *ignored)
6460 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6461 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6462 unsigned long zone_start_pfn, zone_end_pfn;
6463 unsigned long nr_absent;
6465 /* When hotadd a new node from cpu_up(), the node should be empty */
6466 if (!node_start_pfn && !node_end_pfn)
6469 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6470 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6472 adjust_zone_range_for_zone_movable(nid, zone_type,
6473 node_start_pfn, node_end_pfn,
6474 &zone_start_pfn, &zone_end_pfn);
6475 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6478 * ZONE_MOVABLE handling.
6479 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6482 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6483 unsigned long start_pfn, end_pfn;
6484 struct memblock_region *r;
6486 for_each_memblock(memory, r) {
6487 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6488 zone_start_pfn, zone_end_pfn);
6489 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6490 zone_start_pfn, zone_end_pfn);
6492 if (zone_type == ZONE_MOVABLE &&
6493 memblock_is_mirror(r))
6494 nr_absent += end_pfn - start_pfn;
6496 if (zone_type == ZONE_NORMAL &&
6497 !memblock_is_mirror(r))
6498 nr_absent += end_pfn - start_pfn;
6505 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6506 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6507 unsigned long zone_type,
6508 unsigned long node_start_pfn,
6509 unsigned long node_end_pfn,
6510 unsigned long *zone_start_pfn,
6511 unsigned long *zone_end_pfn,
6512 unsigned long *zones_size)
6516 *zone_start_pfn = node_start_pfn;
6517 for (zone = 0; zone < zone_type; zone++)
6518 *zone_start_pfn += zones_size[zone];
6520 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6522 return zones_size[zone_type];
6525 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6526 unsigned long zone_type,
6527 unsigned long node_start_pfn,
6528 unsigned long node_end_pfn,
6529 unsigned long *zholes_size)
6534 return zholes_size[zone_type];
6537 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6539 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6540 unsigned long node_start_pfn,
6541 unsigned long node_end_pfn,
6542 unsigned long *zones_size,
6543 unsigned long *zholes_size)
6545 unsigned long realtotalpages = 0, totalpages = 0;
6548 for (i = 0; i < MAX_NR_ZONES; i++) {
6549 struct zone *zone = pgdat->node_zones + i;
6550 unsigned long zone_start_pfn, zone_end_pfn;
6551 unsigned long size, real_size;
6553 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6559 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6560 node_start_pfn, node_end_pfn,
6563 zone->zone_start_pfn = zone_start_pfn;
6565 zone->zone_start_pfn = 0;
6566 zone->spanned_pages = size;
6567 zone->present_pages = real_size;
6570 realtotalpages += real_size;
6573 pgdat->node_spanned_pages = totalpages;
6574 pgdat->node_present_pages = realtotalpages;
6575 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6579 #ifndef CONFIG_SPARSEMEM
6581 * Calculate the size of the zone->blockflags rounded to an unsigned long
6582 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6583 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6584 * round what is now in bits to nearest long in bits, then return it in
6587 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6589 unsigned long usemapsize;
6591 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6592 usemapsize = roundup(zonesize, pageblock_nr_pages);
6593 usemapsize = usemapsize >> pageblock_order;
6594 usemapsize *= NR_PAGEBLOCK_BITS;
6595 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6597 return usemapsize / 8;
6600 static void __ref setup_usemap(struct pglist_data *pgdat,
6602 unsigned long zone_start_pfn,
6603 unsigned long zonesize)
6605 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6606 zone->pageblock_flags = NULL;
6608 zone->pageblock_flags =
6609 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6611 if (!zone->pageblock_flags)
6612 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6613 usemapsize, zone->name, pgdat->node_id);
6617 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6618 unsigned long zone_start_pfn, unsigned long zonesize) {}
6619 #endif /* CONFIG_SPARSEMEM */
6621 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6623 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6624 void __init set_pageblock_order(void)
6628 /* Check that pageblock_nr_pages has not already been setup */
6629 if (pageblock_order)
6632 if (HPAGE_SHIFT > PAGE_SHIFT)
6633 order = HUGETLB_PAGE_ORDER;
6635 order = MAX_ORDER - 1;
6638 * Assume the largest contiguous order of interest is a huge page.
6639 * This value may be variable depending on boot parameters on IA64 and
6642 pageblock_order = order;
6644 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6647 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6648 * is unused as pageblock_order is set at compile-time. See
6649 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6652 void __init set_pageblock_order(void)
6656 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6658 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6659 unsigned long present_pages)
6661 unsigned long pages = spanned_pages;
6664 * Provide a more accurate estimation if there are holes within
6665 * the zone and SPARSEMEM is in use. If there are holes within the
6666 * zone, each populated memory region may cost us one or two extra
6667 * memmap pages due to alignment because memmap pages for each
6668 * populated regions may not be naturally aligned on page boundary.
6669 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6671 if (spanned_pages > present_pages + (present_pages >> 4) &&
6672 IS_ENABLED(CONFIG_SPARSEMEM))
6673 pages = present_pages;
6675 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6678 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6679 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6681 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6683 spin_lock_init(&ds_queue->split_queue_lock);
6684 INIT_LIST_HEAD(&ds_queue->split_queue);
6685 ds_queue->split_queue_len = 0;
6688 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6691 #ifdef CONFIG_COMPACTION
6692 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6694 init_waitqueue_head(&pgdat->kcompactd_wait);
6697 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6700 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6702 pgdat_resize_init(pgdat);
6704 pgdat_init_split_queue(pgdat);
6705 pgdat_init_kcompactd(pgdat);
6707 init_waitqueue_head(&pgdat->kswapd_wait);
6708 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6710 pgdat_page_ext_init(pgdat);
6711 spin_lock_init(&pgdat->lru_lock);
6712 lruvec_init(&pgdat->__lruvec);
6715 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6716 unsigned long remaining_pages)
6718 atomic_long_set(&zone->managed_pages, remaining_pages);
6719 zone_set_nid(zone, nid);
6720 zone->name = zone_names[idx];
6721 zone->zone_pgdat = NODE_DATA(nid);
6722 spin_lock_init(&zone->lock);
6723 zone_seqlock_init(zone);
6724 zone_pcp_init(zone);
6728 * Set up the zone data structures
6729 * - init pgdat internals
6730 * - init all zones belonging to this node
6732 * NOTE: this function is only called during memory hotplug
6734 #ifdef CONFIG_MEMORY_HOTPLUG
6735 void __ref free_area_init_core_hotplug(int nid)
6738 pg_data_t *pgdat = NODE_DATA(nid);
6740 pgdat_init_internals(pgdat);
6741 for (z = 0; z < MAX_NR_ZONES; z++)
6742 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6747 * Set up the zone data structures:
6748 * - mark all pages reserved
6749 * - mark all memory queues empty
6750 * - clear the memory bitmaps
6752 * NOTE: pgdat should get zeroed by caller.
6753 * NOTE: this function is only called during early init.
6755 static void __init free_area_init_core(struct pglist_data *pgdat)
6758 int nid = pgdat->node_id;
6760 pgdat_init_internals(pgdat);
6761 pgdat->per_cpu_nodestats = &boot_nodestats;
6763 for (j = 0; j < MAX_NR_ZONES; j++) {
6764 struct zone *zone = pgdat->node_zones + j;
6765 unsigned long size, freesize, memmap_pages;
6766 unsigned long zone_start_pfn = zone->zone_start_pfn;
6768 size = zone->spanned_pages;
6769 freesize = zone->present_pages;
6772 * Adjust freesize so that it accounts for how much memory
6773 * is used by this zone for memmap. This affects the watermark
6774 * and per-cpu initialisations
6776 memmap_pages = calc_memmap_size(size, freesize);
6777 if (!is_highmem_idx(j)) {
6778 if (freesize >= memmap_pages) {
6779 freesize -= memmap_pages;
6782 " %s zone: %lu pages used for memmap\n",
6783 zone_names[j], memmap_pages);
6785 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6786 zone_names[j], memmap_pages, freesize);
6789 /* Account for reserved pages */
6790 if (j == 0 && freesize > dma_reserve) {
6791 freesize -= dma_reserve;
6792 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6793 zone_names[0], dma_reserve);
6796 if (!is_highmem_idx(j))
6797 nr_kernel_pages += freesize;
6798 /* Charge for highmem memmap if there are enough kernel pages */
6799 else if (nr_kernel_pages > memmap_pages * 2)
6800 nr_kernel_pages -= memmap_pages;
6801 nr_all_pages += freesize;
6804 * Set an approximate value for lowmem here, it will be adjusted
6805 * when the bootmem allocator frees pages into the buddy system.
6806 * And all highmem pages will be managed by the buddy system.
6808 zone_init_internals(zone, j, nid, freesize);
6813 set_pageblock_order();
6814 setup_usemap(pgdat, zone, zone_start_pfn, size);
6815 init_currently_empty_zone(zone, zone_start_pfn, size);
6816 memmap_init(size, nid, j, zone_start_pfn);
6820 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6821 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6823 unsigned long __maybe_unused start = 0;
6824 unsigned long __maybe_unused offset = 0;
6826 /* Skip empty nodes */
6827 if (!pgdat->node_spanned_pages)
6830 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6831 offset = pgdat->node_start_pfn - start;
6832 /* ia64 gets its own node_mem_map, before this, without bootmem */
6833 if (!pgdat->node_mem_map) {
6834 unsigned long size, end;
6838 * The zone's endpoints aren't required to be MAX_ORDER
6839 * aligned but the node_mem_map endpoints must be in order
6840 * for the buddy allocator to function correctly.
6842 end = pgdat_end_pfn(pgdat);
6843 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6844 size = (end - start) * sizeof(struct page);
6845 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6848 panic("Failed to allocate %ld bytes for node %d memory map\n",
6849 size, pgdat->node_id);
6850 pgdat->node_mem_map = map + offset;
6852 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6853 __func__, pgdat->node_id, (unsigned long)pgdat,
6854 (unsigned long)pgdat->node_mem_map);
6855 #ifndef CONFIG_NEED_MULTIPLE_NODES
6857 * With no DISCONTIG, the global mem_map is just set as node 0's
6859 if (pgdat == NODE_DATA(0)) {
6860 mem_map = NODE_DATA(0)->node_mem_map;
6861 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6862 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6864 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6869 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6870 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6872 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6873 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6875 pgdat->first_deferred_pfn = ULONG_MAX;
6878 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6881 void __init free_area_init_node(int nid, unsigned long *zones_size,
6882 unsigned long node_start_pfn,
6883 unsigned long *zholes_size)
6885 pg_data_t *pgdat = NODE_DATA(nid);
6886 unsigned long start_pfn = 0;
6887 unsigned long end_pfn = 0;
6889 /* pg_data_t should be reset to zero when it's allocated */
6890 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6892 pgdat->node_id = nid;
6893 pgdat->node_start_pfn = node_start_pfn;
6894 pgdat->per_cpu_nodestats = NULL;
6895 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6896 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6897 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6898 (u64)start_pfn << PAGE_SHIFT,
6899 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6901 start_pfn = node_start_pfn;
6903 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6904 zones_size, zholes_size);
6906 alloc_node_mem_map(pgdat);
6907 pgdat_set_deferred_range(pgdat);
6909 free_area_init_core(pgdat);
6912 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6914 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6915 * PageReserved(). Return the number of struct pages that were initialized.
6917 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6922 for (pfn = spfn; pfn < epfn; pfn++) {
6923 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6924 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6925 + pageblock_nr_pages - 1;
6929 * Use a fake node/zone (0) for now. Some of these pages
6930 * (in memblock.reserved but not in memblock.memory) will
6931 * get re-initialized via reserve_bootmem_region() later.
6933 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6934 __SetPageReserved(pfn_to_page(pfn));
6942 * Only struct pages that are backed by physical memory are zeroed and
6943 * initialized by going through __init_single_page(). But, there are some
6944 * struct pages which are reserved in memblock allocator and their fields
6945 * may be accessed (for example page_to_pfn() on some configuration accesses
6946 * flags). We must explicitly initialize those struct pages.
6948 * This function also addresses a similar issue where struct pages are left
6949 * uninitialized because the physical address range is not covered by
6950 * memblock.memory or memblock.reserved. That could happen when memblock
6951 * layout is manually configured via memmap=, or when the highest physical
6952 * address (max_pfn) does not end on a section boundary.
6954 static void __init init_unavailable_mem(void)
6956 phys_addr_t start, end;
6958 phys_addr_t next = 0;
6961 * Loop through unavailable ranges not covered by memblock.memory.
6964 for_each_mem_range(i, &memblock.memory, NULL,
6965 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6967 pgcnt += init_unavailable_range(PFN_DOWN(next),
6973 * Early sections always have a fully populated memmap for the whole
6974 * section - see pfn_valid(). If the last section has holes at the
6975 * end and that section is marked "online", the memmap will be
6976 * considered initialized. Make sure that memmap has a well defined
6979 pgcnt += init_unavailable_range(PFN_DOWN(next),
6980 round_up(max_pfn, PAGES_PER_SECTION));
6983 * Struct pages that do not have backing memory. This could be because
6984 * firmware is using some of this memory, or for some other reasons.
6987 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6990 static inline void __init init_unavailable_mem(void)
6993 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6995 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6997 #if MAX_NUMNODES > 1
6999 * Figure out the number of possible node ids.
7001 void __init setup_nr_node_ids(void)
7003 unsigned int highest;
7005 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7006 nr_node_ids = highest + 1;
7011 * node_map_pfn_alignment - determine the maximum internode alignment
7013 * This function should be called after node map is populated and sorted.
7014 * It calculates the maximum power of two alignment which can distinguish
7017 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7018 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7019 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7020 * shifted, 1GiB is enough and this function will indicate so.
7022 * This is used to test whether pfn -> nid mapping of the chosen memory
7023 * model has fine enough granularity to avoid incorrect mapping for the
7024 * populated node map.
7026 * Return: the determined alignment in pfn's. 0 if there is no alignment
7027 * requirement (single node).
7029 unsigned long __init node_map_pfn_alignment(void)
7031 unsigned long accl_mask = 0, last_end = 0;
7032 unsigned long start, end, mask;
7033 int last_nid = NUMA_NO_NODE;
7036 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7037 if (!start || last_nid < 0 || last_nid == nid) {
7044 * Start with a mask granular enough to pin-point to the
7045 * start pfn and tick off bits one-by-one until it becomes
7046 * too coarse to separate the current node from the last.
7048 mask = ~((1 << __ffs(start)) - 1);
7049 while (mask && last_end <= (start & (mask << 1)))
7052 /* accumulate all internode masks */
7056 /* convert mask to number of pages */
7057 return ~accl_mask + 1;
7060 /* Find the lowest pfn for a node */
7061 static unsigned long __init find_min_pfn_for_node(int nid)
7063 unsigned long min_pfn = ULONG_MAX;
7064 unsigned long start_pfn;
7067 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7068 min_pfn = min(min_pfn, start_pfn);
7070 if (min_pfn == ULONG_MAX) {
7071 pr_warn("Could not find start_pfn for node %d\n", nid);
7079 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7081 * Return: the minimum PFN based on information provided via
7082 * memblock_set_node().
7084 unsigned long __init find_min_pfn_with_active_regions(void)
7086 return find_min_pfn_for_node(MAX_NUMNODES);
7090 * early_calculate_totalpages()
7091 * Sum pages in active regions for movable zone.
7092 * Populate N_MEMORY for calculating usable_nodes.
7094 static unsigned long __init early_calculate_totalpages(void)
7096 unsigned long totalpages = 0;
7097 unsigned long start_pfn, end_pfn;
7100 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7101 unsigned long pages = end_pfn - start_pfn;
7103 totalpages += pages;
7105 node_set_state(nid, N_MEMORY);
7111 * Find the PFN the Movable zone begins in each node. Kernel memory
7112 * is spread evenly between nodes as long as the nodes have enough
7113 * memory. When they don't, some nodes will have more kernelcore than
7116 static void __init find_zone_movable_pfns_for_nodes(void)
7119 unsigned long usable_startpfn;
7120 unsigned long kernelcore_node, kernelcore_remaining;
7121 /* save the state before borrow the nodemask */
7122 nodemask_t saved_node_state = node_states[N_MEMORY];
7123 unsigned long totalpages = early_calculate_totalpages();
7124 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7125 struct memblock_region *r;
7127 /* Need to find movable_zone earlier when movable_node is specified. */
7128 find_usable_zone_for_movable();
7131 * If movable_node is specified, ignore kernelcore and movablecore
7134 if (movable_node_is_enabled()) {
7135 for_each_memblock(memory, r) {
7136 if (!memblock_is_hotpluggable(r))
7141 usable_startpfn = PFN_DOWN(r->base);
7142 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7143 min(usable_startpfn, zone_movable_pfn[nid]) :
7151 * If kernelcore=mirror is specified, ignore movablecore option
7153 if (mirrored_kernelcore) {
7154 bool mem_below_4gb_not_mirrored = false;
7156 for_each_memblock(memory, r) {
7157 if (memblock_is_mirror(r))
7162 usable_startpfn = memblock_region_memory_base_pfn(r);
7164 if (usable_startpfn < 0x100000) {
7165 mem_below_4gb_not_mirrored = true;
7169 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7170 min(usable_startpfn, zone_movable_pfn[nid]) :
7174 if (mem_below_4gb_not_mirrored)
7175 pr_warn("This configuration results in unmirrored kernel memory.");
7181 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7182 * amount of necessary memory.
7184 if (required_kernelcore_percent)
7185 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7187 if (required_movablecore_percent)
7188 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7192 * If movablecore= was specified, calculate what size of
7193 * kernelcore that corresponds so that memory usable for
7194 * any allocation type is evenly spread. If both kernelcore
7195 * and movablecore are specified, then the value of kernelcore
7196 * will be used for required_kernelcore if it's greater than
7197 * what movablecore would have allowed.
7199 if (required_movablecore) {
7200 unsigned long corepages;
7203 * Round-up so that ZONE_MOVABLE is at least as large as what
7204 * was requested by the user
7206 required_movablecore =
7207 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7208 required_movablecore = min(totalpages, required_movablecore);
7209 corepages = totalpages - required_movablecore;
7211 required_kernelcore = max(required_kernelcore, corepages);
7215 * If kernelcore was not specified or kernelcore size is larger
7216 * than totalpages, there is no ZONE_MOVABLE.
7218 if (!required_kernelcore || required_kernelcore >= totalpages)
7221 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7222 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7225 /* Spread kernelcore memory as evenly as possible throughout nodes */
7226 kernelcore_node = required_kernelcore / usable_nodes;
7227 for_each_node_state(nid, N_MEMORY) {
7228 unsigned long start_pfn, end_pfn;
7231 * Recalculate kernelcore_node if the division per node
7232 * now exceeds what is necessary to satisfy the requested
7233 * amount of memory for the kernel
7235 if (required_kernelcore < kernelcore_node)
7236 kernelcore_node = required_kernelcore / usable_nodes;
7239 * As the map is walked, we track how much memory is usable
7240 * by the kernel using kernelcore_remaining. When it is
7241 * 0, the rest of the node is usable by ZONE_MOVABLE
7243 kernelcore_remaining = kernelcore_node;
7245 /* Go through each range of PFNs within this node */
7246 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7247 unsigned long size_pages;
7249 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7250 if (start_pfn >= end_pfn)
7253 /* Account for what is only usable for kernelcore */
7254 if (start_pfn < usable_startpfn) {
7255 unsigned long kernel_pages;
7256 kernel_pages = min(end_pfn, usable_startpfn)
7259 kernelcore_remaining -= min(kernel_pages,
7260 kernelcore_remaining);
7261 required_kernelcore -= min(kernel_pages,
7262 required_kernelcore);
7264 /* Continue if range is now fully accounted */
7265 if (end_pfn <= usable_startpfn) {
7268 * Push zone_movable_pfn to the end so
7269 * that if we have to rebalance
7270 * kernelcore across nodes, we will
7271 * not double account here
7273 zone_movable_pfn[nid] = end_pfn;
7276 start_pfn = usable_startpfn;
7280 * The usable PFN range for ZONE_MOVABLE is from
7281 * start_pfn->end_pfn. Calculate size_pages as the
7282 * number of pages used as kernelcore
7284 size_pages = end_pfn - start_pfn;
7285 if (size_pages > kernelcore_remaining)
7286 size_pages = kernelcore_remaining;
7287 zone_movable_pfn[nid] = start_pfn + size_pages;
7290 * Some kernelcore has been met, update counts and
7291 * break if the kernelcore for this node has been
7294 required_kernelcore -= min(required_kernelcore,
7296 kernelcore_remaining -= size_pages;
7297 if (!kernelcore_remaining)
7303 * If there is still required_kernelcore, we do another pass with one
7304 * less node in the count. This will push zone_movable_pfn[nid] further
7305 * along on the nodes that still have memory until kernelcore is
7309 if (usable_nodes && required_kernelcore > usable_nodes)
7313 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7314 for (nid = 0; nid < MAX_NUMNODES; nid++)
7315 zone_movable_pfn[nid] =
7316 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7319 /* restore the node_state */
7320 node_states[N_MEMORY] = saved_node_state;
7323 /* Any regular or high memory on that node ? */
7324 static void check_for_memory(pg_data_t *pgdat, int nid)
7326 enum zone_type zone_type;
7328 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7329 struct zone *zone = &pgdat->node_zones[zone_type];
7330 if (populated_zone(zone)) {
7331 if (IS_ENABLED(CONFIG_HIGHMEM))
7332 node_set_state(nid, N_HIGH_MEMORY);
7333 if (zone_type <= ZONE_NORMAL)
7334 node_set_state(nid, N_NORMAL_MEMORY);
7341 * free_area_init_nodes - Initialise all pg_data_t and zone data
7342 * @max_zone_pfn: an array of max PFNs for each zone
7344 * This will call free_area_init_node() for each active node in the system.
7345 * Using the page ranges provided by memblock_set_node(), the size of each
7346 * zone in each node and their holes is calculated. If the maximum PFN
7347 * between two adjacent zones match, it is assumed that the zone is empty.
7348 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7349 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7350 * starts where the previous one ended. For example, ZONE_DMA32 starts
7351 * at arch_max_dma_pfn.
7353 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7355 unsigned long start_pfn, end_pfn;
7358 /* Record where the zone boundaries are */
7359 memset(arch_zone_lowest_possible_pfn, 0,
7360 sizeof(arch_zone_lowest_possible_pfn));
7361 memset(arch_zone_highest_possible_pfn, 0,
7362 sizeof(arch_zone_highest_possible_pfn));
7364 start_pfn = find_min_pfn_with_active_regions();
7366 for (i = 0; i < MAX_NR_ZONES; i++) {
7367 if (i == ZONE_MOVABLE)
7370 end_pfn = max(max_zone_pfn[i], start_pfn);
7371 arch_zone_lowest_possible_pfn[i] = start_pfn;
7372 arch_zone_highest_possible_pfn[i] = end_pfn;
7374 start_pfn = end_pfn;
7377 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7378 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7379 find_zone_movable_pfns_for_nodes();
7381 /* Print out the zone ranges */
7382 pr_info("Zone ranges:\n");
7383 for (i = 0; i < MAX_NR_ZONES; i++) {
7384 if (i == ZONE_MOVABLE)
7386 pr_info(" %-8s ", zone_names[i]);
7387 if (arch_zone_lowest_possible_pfn[i] ==
7388 arch_zone_highest_possible_pfn[i])
7391 pr_cont("[mem %#018Lx-%#018Lx]\n",
7392 (u64)arch_zone_lowest_possible_pfn[i]
7394 ((u64)arch_zone_highest_possible_pfn[i]
7395 << PAGE_SHIFT) - 1);
7398 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7399 pr_info("Movable zone start for each node\n");
7400 for (i = 0; i < MAX_NUMNODES; i++) {
7401 if (zone_movable_pfn[i])
7402 pr_info(" Node %d: %#018Lx\n", i,
7403 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7407 * Print out the early node map, and initialize the
7408 * subsection-map relative to active online memory ranges to
7409 * enable future "sub-section" extensions of the memory map.
7411 pr_info("Early memory node ranges\n");
7412 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7413 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7414 (u64)start_pfn << PAGE_SHIFT,
7415 ((u64)end_pfn << PAGE_SHIFT) - 1);
7416 subsection_map_init(start_pfn, end_pfn - start_pfn);
7419 /* Initialise every node */
7420 mminit_verify_pageflags_layout();
7421 setup_nr_node_ids();
7422 init_unavailable_mem();
7423 for_each_online_node(nid) {
7424 pg_data_t *pgdat = NODE_DATA(nid);
7425 free_area_init_node(nid, NULL,
7426 find_min_pfn_for_node(nid), NULL);
7428 /* Any memory on that node */
7429 if (pgdat->node_present_pages)
7430 node_set_state(nid, N_MEMORY);
7431 check_for_memory(pgdat, nid);
7435 static int __init cmdline_parse_core(char *p, unsigned long *core,
7436 unsigned long *percent)
7438 unsigned long long coremem;
7444 /* Value may be a percentage of total memory, otherwise bytes */
7445 coremem = simple_strtoull(p, &endptr, 0);
7446 if (*endptr == '%') {
7447 /* Paranoid check for percent values greater than 100 */
7448 WARN_ON(coremem > 100);
7452 coremem = memparse(p, &p);
7453 /* Paranoid check that UL is enough for the coremem value */
7454 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7456 *core = coremem >> PAGE_SHIFT;
7463 * kernelcore=size sets the amount of memory for use for allocations that
7464 * cannot be reclaimed or migrated.
7466 static int __init cmdline_parse_kernelcore(char *p)
7468 /* parse kernelcore=mirror */
7469 if (parse_option_str(p, "mirror")) {
7470 mirrored_kernelcore = true;
7474 return cmdline_parse_core(p, &required_kernelcore,
7475 &required_kernelcore_percent);
7479 * movablecore=size sets the amount of memory for use for allocations that
7480 * can be reclaimed or migrated.
7482 static int __init cmdline_parse_movablecore(char *p)
7484 return cmdline_parse_core(p, &required_movablecore,
7485 &required_movablecore_percent);
7488 early_param("kernelcore", cmdline_parse_kernelcore);
7489 early_param("movablecore", cmdline_parse_movablecore);
7491 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7493 void adjust_managed_page_count(struct page *page, long count)
7495 atomic_long_add(count, &page_zone(page)->managed_pages);
7496 totalram_pages_add(count);
7497 #ifdef CONFIG_HIGHMEM
7498 if (PageHighMem(page))
7499 totalhigh_pages_add(count);
7502 EXPORT_SYMBOL(adjust_managed_page_count);
7504 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7507 unsigned long pages = 0;
7509 start = (void *)PAGE_ALIGN((unsigned long)start);
7510 end = (void *)((unsigned long)end & PAGE_MASK);
7511 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7512 struct page *page = virt_to_page(pos);
7513 void *direct_map_addr;
7516 * 'direct_map_addr' might be different from 'pos'
7517 * because some architectures' virt_to_page()
7518 * work with aliases. Getting the direct map
7519 * address ensures that we get a _writeable_
7520 * alias for the memset().
7522 direct_map_addr = page_address(page);
7523 if ((unsigned int)poison <= 0xFF)
7524 memset(direct_map_addr, poison, PAGE_SIZE);
7526 free_reserved_page(page);
7530 pr_info("Freeing %s memory: %ldK\n",
7531 s, pages << (PAGE_SHIFT - 10));
7536 #ifdef CONFIG_HIGHMEM
7537 void free_highmem_page(struct page *page)
7539 __free_reserved_page(page);
7540 totalram_pages_inc();
7541 atomic_long_inc(&page_zone(page)->managed_pages);
7542 totalhigh_pages_inc();
7547 void __init mem_init_print_info(const char *str)
7549 unsigned long physpages, codesize, datasize, rosize, bss_size;
7550 unsigned long init_code_size, init_data_size;
7552 physpages = get_num_physpages();
7553 codesize = _etext - _stext;
7554 datasize = _edata - _sdata;
7555 rosize = __end_rodata - __start_rodata;
7556 bss_size = __bss_stop - __bss_start;
7557 init_data_size = __init_end - __init_begin;
7558 init_code_size = _einittext - _sinittext;
7561 * Detect special cases and adjust section sizes accordingly:
7562 * 1) .init.* may be embedded into .data sections
7563 * 2) .init.text.* may be out of [__init_begin, __init_end],
7564 * please refer to arch/tile/kernel/vmlinux.lds.S.
7565 * 3) .rodata.* may be embedded into .text or .data sections.
7567 #define adj_init_size(start, end, size, pos, adj) \
7569 if (start <= pos && pos < end && size > adj) \
7573 adj_init_size(__init_begin, __init_end, init_data_size,
7574 _sinittext, init_code_size);
7575 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7576 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7577 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7578 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7580 #undef adj_init_size
7582 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7583 #ifdef CONFIG_HIGHMEM
7587 nr_free_pages() << (PAGE_SHIFT - 10),
7588 physpages << (PAGE_SHIFT - 10),
7589 codesize >> 10, datasize >> 10, rosize >> 10,
7590 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7591 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7592 totalcma_pages << (PAGE_SHIFT - 10),
7593 #ifdef CONFIG_HIGHMEM
7594 totalhigh_pages() << (PAGE_SHIFT - 10),
7596 str ? ", " : "", str ? str : "");
7600 * set_dma_reserve - set the specified number of pages reserved in the first zone
7601 * @new_dma_reserve: The number of pages to mark reserved
7603 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7604 * In the DMA zone, a significant percentage may be consumed by kernel image
7605 * and other unfreeable allocations which can skew the watermarks badly. This
7606 * function may optionally be used to account for unfreeable pages in the
7607 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7608 * smaller per-cpu batchsize.
7610 void __init set_dma_reserve(unsigned long new_dma_reserve)
7612 dma_reserve = new_dma_reserve;
7615 void __init free_area_init(unsigned long *zones_size)
7617 init_unavailable_mem();
7618 free_area_init_node(0, zones_size,
7619 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7622 static int page_alloc_cpu_dead(unsigned int cpu)
7625 lru_add_drain_cpu(cpu);
7629 * Spill the event counters of the dead processor
7630 * into the current processors event counters.
7631 * This artificially elevates the count of the current
7634 vm_events_fold_cpu(cpu);
7637 * Zero the differential counters of the dead processor
7638 * so that the vm statistics are consistent.
7640 * This is only okay since the processor is dead and cannot
7641 * race with what we are doing.
7643 cpu_vm_stats_fold(cpu);
7648 int hashdist = HASHDIST_DEFAULT;
7650 static int __init set_hashdist(char *str)
7654 hashdist = simple_strtoul(str, &str, 0);
7657 __setup("hashdist=", set_hashdist);
7660 void __init page_alloc_init(void)
7665 if (num_node_state(N_MEMORY) == 1)
7669 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7670 "mm/page_alloc:dead", NULL,
7671 page_alloc_cpu_dead);
7676 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7677 * or min_free_kbytes changes.
7679 static void calculate_totalreserve_pages(void)
7681 struct pglist_data *pgdat;
7682 unsigned long reserve_pages = 0;
7683 enum zone_type i, j;
7685 for_each_online_pgdat(pgdat) {
7687 pgdat->totalreserve_pages = 0;
7689 for (i = 0; i < MAX_NR_ZONES; i++) {
7690 struct zone *zone = pgdat->node_zones + i;
7692 unsigned long managed_pages = zone_managed_pages(zone);
7694 /* Find valid and maximum lowmem_reserve in the zone */
7695 for (j = i; j < MAX_NR_ZONES; j++) {
7696 if (zone->lowmem_reserve[j] > max)
7697 max = zone->lowmem_reserve[j];
7700 /* we treat the high watermark as reserved pages. */
7701 max += high_wmark_pages(zone);
7703 if (max > managed_pages)
7704 max = managed_pages;
7706 pgdat->totalreserve_pages += max;
7708 reserve_pages += max;
7711 totalreserve_pages = reserve_pages;
7715 * setup_per_zone_lowmem_reserve - called whenever
7716 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7717 * has a correct pages reserved value, so an adequate number of
7718 * pages are left in the zone after a successful __alloc_pages().
7720 static void setup_per_zone_lowmem_reserve(void)
7722 struct pglist_data *pgdat;
7723 enum zone_type j, idx;
7725 for_each_online_pgdat(pgdat) {
7726 for (j = 0; j < MAX_NR_ZONES; j++) {
7727 struct zone *zone = pgdat->node_zones + j;
7728 unsigned long managed_pages = zone_managed_pages(zone);
7730 zone->lowmem_reserve[j] = 0;
7734 struct zone *lower_zone;
7737 lower_zone = pgdat->node_zones + idx;
7739 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7740 sysctl_lowmem_reserve_ratio[idx] = 0;
7741 lower_zone->lowmem_reserve[j] = 0;
7743 lower_zone->lowmem_reserve[j] =
7744 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7746 managed_pages += zone_managed_pages(lower_zone);
7751 /* update totalreserve_pages */
7752 calculate_totalreserve_pages();
7755 static void __setup_per_zone_wmarks(void)
7757 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7758 unsigned long lowmem_pages = 0;
7760 unsigned long flags;
7762 /* Calculate total number of !ZONE_HIGHMEM pages */
7763 for_each_zone(zone) {
7764 if (!is_highmem(zone))
7765 lowmem_pages += zone_managed_pages(zone);
7768 for_each_zone(zone) {
7771 spin_lock_irqsave(&zone->lock, flags);
7772 tmp = (u64)pages_min * zone_managed_pages(zone);
7773 do_div(tmp, lowmem_pages);
7774 if (is_highmem(zone)) {
7776 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7777 * need highmem pages, so cap pages_min to a small
7780 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7781 * deltas control async page reclaim, and so should
7782 * not be capped for highmem.
7784 unsigned long min_pages;
7786 min_pages = zone_managed_pages(zone) / 1024;
7787 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7788 zone->_watermark[WMARK_MIN] = min_pages;
7791 * If it's a lowmem zone, reserve a number of pages
7792 * proportionate to the zone's size.
7794 zone->_watermark[WMARK_MIN] = tmp;
7798 * Set the kswapd watermarks distance according to the
7799 * scale factor in proportion to available memory, but
7800 * ensure a minimum size on small systems.
7802 tmp = max_t(u64, tmp >> 2,
7803 mult_frac(zone_managed_pages(zone),
7804 watermark_scale_factor, 10000));
7806 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7807 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7808 zone->watermark_boost = 0;
7810 spin_unlock_irqrestore(&zone->lock, flags);
7813 /* update totalreserve_pages */
7814 calculate_totalreserve_pages();
7818 * setup_per_zone_wmarks - called when min_free_kbytes changes
7819 * or when memory is hot-{added|removed}
7821 * Ensures that the watermark[min,low,high] values for each zone are set
7822 * correctly with respect to min_free_kbytes.
7824 void setup_per_zone_wmarks(void)
7826 static DEFINE_SPINLOCK(lock);
7829 __setup_per_zone_wmarks();
7834 * Initialise min_free_kbytes.
7836 * For small machines we want it small (128k min). For large machines
7837 * we want it large (64MB max). But it is not linear, because network
7838 * bandwidth does not increase linearly with machine size. We use
7840 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7841 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7857 int __meminit init_per_zone_wmark_min(void)
7859 unsigned long lowmem_kbytes;
7860 int new_min_free_kbytes;
7862 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7863 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7865 if (new_min_free_kbytes > user_min_free_kbytes) {
7866 min_free_kbytes = new_min_free_kbytes;
7867 if (min_free_kbytes < 128)
7868 min_free_kbytes = 128;
7869 if (min_free_kbytes > 65536)
7870 min_free_kbytes = 65536;
7872 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7873 new_min_free_kbytes, user_min_free_kbytes);
7875 setup_per_zone_wmarks();
7876 refresh_zone_stat_thresholds();
7877 setup_per_zone_lowmem_reserve();
7880 setup_min_unmapped_ratio();
7881 setup_min_slab_ratio();
7886 core_initcall(init_per_zone_wmark_min)
7889 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7890 * that we can call two helper functions whenever min_free_kbytes
7893 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7894 void __user *buffer, size_t *length, loff_t *ppos)
7898 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7903 user_min_free_kbytes = min_free_kbytes;
7904 setup_per_zone_wmarks();
7909 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7910 void __user *buffer, size_t *length, loff_t *ppos)
7914 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7921 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7922 void __user *buffer, size_t *length, loff_t *ppos)
7926 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7931 setup_per_zone_wmarks();
7937 static void setup_min_unmapped_ratio(void)
7942 for_each_online_pgdat(pgdat)
7943 pgdat->min_unmapped_pages = 0;
7946 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7947 sysctl_min_unmapped_ratio) / 100;
7951 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7952 void __user *buffer, size_t *length, loff_t *ppos)
7956 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7960 setup_min_unmapped_ratio();
7965 static void setup_min_slab_ratio(void)
7970 for_each_online_pgdat(pgdat)
7971 pgdat->min_slab_pages = 0;
7974 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7975 sysctl_min_slab_ratio) / 100;
7978 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7979 void __user *buffer, size_t *length, loff_t *ppos)
7983 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7987 setup_min_slab_ratio();
7994 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7995 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7996 * whenever sysctl_lowmem_reserve_ratio changes.
7998 * The reserve ratio obviously has absolutely no relation with the
7999 * minimum watermarks. The lowmem reserve ratio can only make sense
8000 * if in function of the boot time zone sizes.
8002 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8003 void __user *buffer, size_t *length, loff_t *ppos)
8005 proc_dointvec_minmax(table, write, buffer, length, ppos);
8006 setup_per_zone_lowmem_reserve();
8010 static void __zone_pcp_update(struct zone *zone)
8014 for_each_possible_cpu(cpu)
8015 pageset_set_high_and_batch(zone,
8016 per_cpu_ptr(zone->pageset, cpu));
8020 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8021 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8022 * pagelist can have before it gets flushed back to buddy allocator.
8024 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8025 void __user *buffer, size_t *length, loff_t *ppos)
8028 int old_percpu_pagelist_fraction;
8031 mutex_lock(&pcp_batch_high_lock);
8032 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8034 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8035 if (!write || ret < 0)
8038 /* Sanity checking to avoid pcp imbalance */
8039 if (percpu_pagelist_fraction &&
8040 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8041 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8047 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8050 for_each_populated_zone(zone)
8051 __zone_pcp_update(zone);
8053 mutex_unlock(&pcp_batch_high_lock);
8057 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8059 * Returns the number of pages that arch has reserved but
8060 * is not known to alloc_large_system_hash().
8062 static unsigned long __init arch_reserved_kernel_pages(void)
8069 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8070 * machines. As memory size is increased the scale is also increased but at
8071 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8072 * quadruples the scale is increased by one, which means the size of hash table
8073 * only doubles, instead of quadrupling as well.
8074 * Because 32-bit systems cannot have large physical memory, where this scaling
8075 * makes sense, it is disabled on such platforms.
8077 #if __BITS_PER_LONG > 32
8078 #define ADAPT_SCALE_BASE (64ul << 30)
8079 #define ADAPT_SCALE_SHIFT 2
8080 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8084 * allocate a large system hash table from bootmem
8085 * - it is assumed that the hash table must contain an exact power-of-2
8086 * quantity of entries
8087 * - limit is the number of hash buckets, not the total allocation size
8089 void *__init alloc_large_system_hash(const char *tablename,
8090 unsigned long bucketsize,
8091 unsigned long numentries,
8094 unsigned int *_hash_shift,
8095 unsigned int *_hash_mask,
8096 unsigned long low_limit,
8097 unsigned long high_limit)
8099 unsigned long long max = high_limit;
8100 unsigned long log2qty, size;
8105 /* allow the kernel cmdline to have a say */
8107 /* round applicable memory size up to nearest megabyte */
8108 numentries = nr_kernel_pages;
8109 numentries -= arch_reserved_kernel_pages();
8111 /* It isn't necessary when PAGE_SIZE >= 1MB */
8112 if (PAGE_SHIFT < 20)
8113 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8115 #if __BITS_PER_LONG > 32
8117 unsigned long adapt;
8119 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8120 adapt <<= ADAPT_SCALE_SHIFT)
8125 /* limit to 1 bucket per 2^scale bytes of low memory */
8126 if (scale > PAGE_SHIFT)
8127 numentries >>= (scale - PAGE_SHIFT);
8129 numentries <<= (PAGE_SHIFT - scale);
8131 /* Make sure we've got at least a 0-order allocation.. */
8132 if (unlikely(flags & HASH_SMALL)) {
8133 /* Makes no sense without HASH_EARLY */
8134 WARN_ON(!(flags & HASH_EARLY));
8135 if (!(numentries >> *_hash_shift)) {
8136 numentries = 1UL << *_hash_shift;
8137 BUG_ON(!numentries);
8139 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8140 numentries = PAGE_SIZE / bucketsize;
8142 numentries = roundup_pow_of_two(numentries);
8144 /* limit allocation size to 1/16 total memory by default */
8146 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8147 do_div(max, bucketsize);
8149 max = min(max, 0x80000000ULL);
8151 if (numentries < low_limit)
8152 numentries = low_limit;
8153 if (numentries > max)
8156 log2qty = ilog2(numentries);
8158 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8161 size = bucketsize << log2qty;
8162 if (flags & HASH_EARLY) {
8163 if (flags & HASH_ZERO)
8164 table = memblock_alloc(size, SMP_CACHE_BYTES);
8166 table = memblock_alloc_raw(size,
8168 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8169 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8173 * If bucketsize is not a power-of-two, we may free
8174 * some pages at the end of hash table which
8175 * alloc_pages_exact() automatically does
8177 table = alloc_pages_exact(size, gfp_flags);
8178 kmemleak_alloc(table, size, 1, gfp_flags);
8180 } while (!table && size > PAGE_SIZE && --log2qty);
8183 panic("Failed to allocate %s hash table\n", tablename);
8185 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8186 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8187 virt ? "vmalloc" : "linear");
8190 *_hash_shift = log2qty;
8192 *_hash_mask = (1 << log2qty) - 1;
8198 * This function checks whether pageblock includes unmovable pages or not.
8200 * PageLRU check without isolation or lru_lock could race so that
8201 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8202 * check without lock_page also may miss some movable non-lru pages at
8203 * race condition. So you can't expect this function should be exact.
8205 * Returns a page without holding a reference. If the caller wants to
8206 * dereference that page (e.g., dumping), it has to make sure that that it
8207 * cannot get removed (e.g., via memory unplug) concurrently.
8210 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8211 int migratetype, int flags)
8213 unsigned long iter = 0;
8214 unsigned long pfn = page_to_pfn(page);
8217 * TODO we could make this much more efficient by not checking every
8218 * page in the range if we know all of them are in MOVABLE_ZONE and
8219 * that the movable zone guarantees that pages are migratable but
8220 * the later is not the case right now unfortunatelly. E.g. movablecore
8221 * can still lead to having bootmem allocations in zone_movable.
8224 if (is_migrate_cma_page(page)) {
8226 * CMA allocations (alloc_contig_range) really need to mark
8227 * isolate CMA pageblocks even when they are not movable in fact
8228 * so consider them movable here.
8230 if (is_migrate_cma(migratetype))
8236 for (; iter < pageblock_nr_pages; iter++) {
8237 if (!pfn_valid_within(pfn + iter))
8240 page = pfn_to_page(pfn + iter);
8242 if (PageReserved(page))
8246 * If the zone is movable and we have ruled out all reserved
8247 * pages then it should be reasonably safe to assume the rest
8250 if (zone_idx(zone) == ZONE_MOVABLE)
8254 * Hugepages are not in LRU lists, but they're movable.
8255 * We need not scan over tail pages because we don't
8256 * handle each tail page individually in migration.
8258 if (PageHuge(page)) {
8259 struct page *head = compound_head(page);
8260 unsigned int skip_pages;
8262 if (!hugepage_migration_supported(page_hstate(head)))
8265 skip_pages = compound_nr(head) - (page - head);
8266 iter += skip_pages - 1;
8271 * We can't use page_count without pin a page
8272 * because another CPU can free compound page.
8273 * This check already skips compound tails of THP
8274 * because their page->_refcount is zero at all time.
8276 if (!page_ref_count(page)) {
8277 if (PageBuddy(page))
8278 iter += (1 << page_order(page)) - 1;
8283 * The HWPoisoned page may be not in buddy system, and
8284 * page_count() is not 0.
8286 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8289 if (__PageMovable(page) || PageLRU(page))
8293 * If there are RECLAIMABLE pages, we need to check
8294 * it. But now, memory offline itself doesn't call
8295 * shrink_node_slabs() and it still to be fixed.
8298 * If the page is not RAM, page_count()should be 0.
8299 * we don't need more check. This is an _used_ not-movable page.
8301 * The problematic thing here is PG_reserved pages. PG_reserved
8302 * is set to both of a memory hole page and a _used_ kernel
8310 #ifdef CONFIG_CONTIG_ALLOC
8311 static unsigned long pfn_max_align_down(unsigned long pfn)
8313 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8314 pageblock_nr_pages) - 1);
8317 static unsigned long pfn_max_align_up(unsigned long pfn)
8319 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8320 pageblock_nr_pages));
8323 /* [start, end) must belong to a single zone. */
8324 static int __alloc_contig_migrate_range(struct compact_control *cc,
8325 unsigned long start, unsigned long end)
8327 /* This function is based on compact_zone() from compaction.c. */
8328 unsigned long nr_reclaimed;
8329 unsigned long pfn = start;
8330 unsigned int tries = 0;
8335 while (pfn < end || !list_empty(&cc->migratepages)) {
8336 if (fatal_signal_pending(current)) {
8341 if (list_empty(&cc->migratepages)) {
8342 cc->nr_migratepages = 0;
8343 pfn = isolate_migratepages_range(cc, pfn, end);
8349 } else if (++tries == 5) {
8350 ret = ret < 0 ? ret : -EBUSY;
8354 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8356 cc->nr_migratepages -= nr_reclaimed;
8358 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8359 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8362 putback_movable_pages(&cc->migratepages);
8369 * alloc_contig_range() -- tries to allocate given range of pages
8370 * @start: start PFN to allocate
8371 * @end: one-past-the-last PFN to allocate
8372 * @migratetype: migratetype of the underlaying pageblocks (either
8373 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8374 * in range must have the same migratetype and it must
8375 * be either of the two.
8376 * @gfp_mask: GFP mask to use during compaction
8378 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8379 * aligned. The PFN range must belong to a single zone.
8381 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8382 * pageblocks in the range. Once isolated, the pageblocks should not
8383 * be modified by others.
8385 * Return: zero on success or negative error code. On success all
8386 * pages which PFN is in [start, end) are allocated for the caller and
8387 * need to be freed with free_contig_range().
8389 int alloc_contig_range(unsigned long start, unsigned long end,
8390 unsigned migratetype, gfp_t gfp_mask)
8392 unsigned long outer_start, outer_end;
8396 struct compact_control cc = {
8397 .nr_migratepages = 0,
8399 .zone = page_zone(pfn_to_page(start)),
8400 .mode = MIGRATE_SYNC,
8401 .ignore_skip_hint = true,
8402 .no_set_skip_hint = true,
8403 .gfp_mask = current_gfp_context(gfp_mask),
8405 INIT_LIST_HEAD(&cc.migratepages);
8408 * What we do here is we mark all pageblocks in range as
8409 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8410 * have different sizes, and due to the way page allocator
8411 * work, we align the range to biggest of the two pages so
8412 * that page allocator won't try to merge buddies from
8413 * different pageblocks and change MIGRATE_ISOLATE to some
8414 * other migration type.
8416 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8417 * migrate the pages from an unaligned range (ie. pages that
8418 * we are interested in). This will put all the pages in
8419 * range back to page allocator as MIGRATE_ISOLATE.
8421 * When this is done, we take the pages in range from page
8422 * allocator removing them from the buddy system. This way
8423 * page allocator will never consider using them.
8425 * This lets us mark the pageblocks back as
8426 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8427 * aligned range but not in the unaligned, original range are
8428 * put back to page allocator so that buddy can use them.
8431 ret = start_isolate_page_range(pfn_max_align_down(start),
8432 pfn_max_align_up(end), migratetype, 0);
8437 * In case of -EBUSY, we'd like to know which page causes problem.
8438 * So, just fall through. test_pages_isolated() has a tracepoint
8439 * which will report the busy page.
8441 * It is possible that busy pages could become available before
8442 * the call to test_pages_isolated, and the range will actually be
8443 * allocated. So, if we fall through be sure to clear ret so that
8444 * -EBUSY is not accidentally used or returned to caller.
8446 ret = __alloc_contig_migrate_range(&cc, start, end);
8447 if (ret && ret != -EBUSY)
8452 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8453 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8454 * more, all pages in [start, end) are free in page allocator.
8455 * What we are going to do is to allocate all pages from
8456 * [start, end) (that is remove them from page allocator).
8458 * The only problem is that pages at the beginning and at the
8459 * end of interesting range may be not aligned with pages that
8460 * page allocator holds, ie. they can be part of higher order
8461 * pages. Because of this, we reserve the bigger range and
8462 * once this is done free the pages we are not interested in.
8464 * We don't have to hold zone->lock here because the pages are
8465 * isolated thus they won't get removed from buddy.
8468 lru_add_drain_all();
8471 outer_start = start;
8472 while (!PageBuddy(pfn_to_page(outer_start))) {
8473 if (++order >= MAX_ORDER) {
8474 outer_start = start;
8477 outer_start &= ~0UL << order;
8480 if (outer_start != start) {
8481 order = page_order(pfn_to_page(outer_start));
8484 * outer_start page could be small order buddy page and
8485 * it doesn't include start page. Adjust outer_start
8486 * in this case to report failed page properly
8487 * on tracepoint in test_pages_isolated()
8489 if (outer_start + (1UL << order) <= start)
8490 outer_start = start;
8493 /* Make sure the range is really isolated. */
8494 if (test_pages_isolated(outer_start, end, 0)) {
8495 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8496 __func__, outer_start, end);
8501 /* Grab isolated pages from freelists. */
8502 outer_end = isolate_freepages_range(&cc, outer_start, end);
8508 /* Free head and tail (if any) */
8509 if (start != outer_start)
8510 free_contig_range(outer_start, start - outer_start);
8511 if (end != outer_end)
8512 free_contig_range(end, outer_end - end);
8515 undo_isolate_page_range(pfn_max_align_down(start),
8516 pfn_max_align_up(end), migratetype);
8520 static int __alloc_contig_pages(unsigned long start_pfn,
8521 unsigned long nr_pages, gfp_t gfp_mask)
8523 unsigned long end_pfn = start_pfn + nr_pages;
8525 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8529 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8530 unsigned long nr_pages)
8532 unsigned long i, end_pfn = start_pfn + nr_pages;
8535 for (i = start_pfn; i < end_pfn; i++) {
8536 page = pfn_to_online_page(i);
8540 if (page_zone(page) != z)
8543 if (PageReserved(page))
8546 if (page_count(page) > 0)
8555 static bool zone_spans_last_pfn(const struct zone *zone,
8556 unsigned long start_pfn, unsigned long nr_pages)
8558 unsigned long last_pfn = start_pfn + nr_pages - 1;
8560 return zone_spans_pfn(zone, last_pfn);
8564 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8565 * @nr_pages: Number of contiguous pages to allocate
8566 * @gfp_mask: GFP mask to limit search and used during compaction
8568 * @nodemask: Mask for other possible nodes
8570 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8571 * on an applicable zonelist to find a contiguous pfn range which can then be
8572 * tried for allocation with alloc_contig_range(). This routine is intended
8573 * for allocation requests which can not be fulfilled with the buddy allocator.
8575 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8576 * power of two then the alignment is guaranteed to be to the given nr_pages
8577 * (e.g. 1GB request would be aligned to 1GB).
8579 * Allocated pages can be freed with free_contig_range() or by manually calling
8580 * __free_page() on each allocated page.
8582 * Return: pointer to contiguous pages on success, or NULL if not successful.
8584 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8585 int nid, nodemask_t *nodemask)
8587 unsigned long ret, pfn, flags;
8588 struct zonelist *zonelist;
8592 zonelist = node_zonelist(nid, gfp_mask);
8593 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8594 gfp_zone(gfp_mask), nodemask) {
8595 spin_lock_irqsave(&zone->lock, flags);
8597 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8598 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8599 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8601 * We release the zone lock here because
8602 * alloc_contig_range() will also lock the zone
8603 * at some point. If there's an allocation
8604 * spinning on this lock, it may win the race
8605 * and cause alloc_contig_range() to fail...
8607 spin_unlock_irqrestore(&zone->lock, flags);
8608 ret = __alloc_contig_pages(pfn, nr_pages,
8611 return pfn_to_page(pfn);
8612 spin_lock_irqsave(&zone->lock, flags);
8616 spin_unlock_irqrestore(&zone->lock, flags);
8620 #endif /* CONFIG_CONTIG_ALLOC */
8622 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8624 unsigned int count = 0;
8626 for (; nr_pages--; pfn++) {
8627 struct page *page = pfn_to_page(pfn);
8629 count += page_count(page) != 1;
8632 WARN(count != 0, "%d pages are still in use!\n", count);
8636 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8637 * page high values need to be recalulated.
8639 void __meminit zone_pcp_update(struct zone *zone)
8641 mutex_lock(&pcp_batch_high_lock);
8642 __zone_pcp_update(zone);
8643 mutex_unlock(&pcp_batch_high_lock);
8646 void zone_pcp_reset(struct zone *zone)
8648 unsigned long flags;
8650 struct per_cpu_pageset *pset;
8652 /* avoid races with drain_pages() */
8653 local_irq_save(flags);
8654 if (zone->pageset != &boot_pageset) {
8655 for_each_online_cpu(cpu) {
8656 pset = per_cpu_ptr(zone->pageset, cpu);
8657 drain_zonestat(zone, pset);
8659 free_percpu(zone->pageset);
8660 zone->pageset = &boot_pageset;
8662 local_irq_restore(flags);
8665 #ifdef CONFIG_MEMORY_HOTREMOVE
8667 * All pages in the range must be in a single zone and isolated
8668 * before calling this.
8671 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8677 unsigned long flags;
8678 unsigned long offlined_pages = 0;
8680 /* find the first valid pfn */
8681 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8685 return offlined_pages;
8687 offline_mem_sections(pfn, end_pfn);
8688 zone = page_zone(pfn_to_page(pfn));
8689 spin_lock_irqsave(&zone->lock, flags);
8691 while (pfn < end_pfn) {
8692 if (!pfn_valid(pfn)) {
8696 page = pfn_to_page(pfn);
8698 * The HWPoisoned page may be not in buddy system, and
8699 * page_count() is not 0.
8701 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8707 BUG_ON(page_count(page));
8708 BUG_ON(!PageBuddy(page));
8709 order = page_order(page);
8710 offlined_pages += 1 << order;
8711 del_page_from_free_area(page, &zone->free_area[order]);
8712 pfn += (1 << order);
8714 spin_unlock_irqrestore(&zone->lock, flags);
8716 return offlined_pages;
8720 bool is_free_buddy_page(struct page *page)
8722 struct zone *zone = page_zone(page);
8723 unsigned long pfn = page_to_pfn(page);
8724 unsigned long flags;
8727 spin_lock_irqsave(&zone->lock, flags);
8728 for (order = 0; order < MAX_ORDER; order++) {
8729 struct page *page_head = page - (pfn & ((1 << order) - 1));
8731 if (PageBuddy(page_head) && page_order(page_head) >= order)
8734 spin_unlock_irqrestore(&zone->lock, flags);
8736 return order < MAX_ORDER;
8739 #ifdef CONFIG_MEMORY_FAILURE
8741 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8742 * test is performed under the zone lock to prevent a race against page
8745 bool set_hwpoison_free_buddy_page(struct page *page)
8747 struct zone *zone = page_zone(page);
8748 unsigned long pfn = page_to_pfn(page);
8749 unsigned long flags;
8751 bool hwpoisoned = false;
8753 spin_lock_irqsave(&zone->lock, flags);
8754 for (order = 0; order < MAX_ORDER; order++) {
8755 struct page *page_head = page - (pfn & ((1 << order) - 1));
8757 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8758 if (!TestSetPageHWPoison(page))
8763 spin_unlock_irqrestore(&zone->lock, flags);