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);
691 if (hpage_pincount_available(page))
692 atomic_set(compound_pincount_ptr(page), 0);
695 #ifdef CONFIG_DEBUG_PAGEALLOC
696 unsigned int _debug_guardpage_minorder;
698 bool _debug_pagealloc_enabled_early __read_mostly
699 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
700 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
701 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
702 EXPORT_SYMBOL(_debug_pagealloc_enabled);
704 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
706 static int __init early_debug_pagealloc(char *buf)
708 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
710 early_param("debug_pagealloc", early_debug_pagealloc);
712 void init_debug_pagealloc(void)
714 if (!debug_pagealloc_enabled())
717 static_branch_enable(&_debug_pagealloc_enabled);
719 if (!debug_guardpage_minorder())
722 static_branch_enable(&_debug_guardpage_enabled);
725 static int __init debug_guardpage_minorder_setup(char *buf)
729 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
730 pr_err("Bad debug_guardpage_minorder value\n");
733 _debug_guardpage_minorder = res;
734 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
737 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
739 static inline bool set_page_guard(struct zone *zone, struct page *page,
740 unsigned int order, int migratetype)
742 if (!debug_guardpage_enabled())
745 if (order >= debug_guardpage_minorder())
748 __SetPageGuard(page);
749 INIT_LIST_HEAD(&page->lru);
750 set_page_private(page, order);
751 /* Guard pages are not available for any usage */
752 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
757 static inline void clear_page_guard(struct zone *zone, struct page *page,
758 unsigned int order, int migratetype)
760 if (!debug_guardpage_enabled())
763 __ClearPageGuard(page);
765 set_page_private(page, 0);
766 if (!is_migrate_isolate(migratetype))
767 __mod_zone_freepage_state(zone, (1 << order), migratetype);
770 static inline bool set_page_guard(struct zone *zone, struct page *page,
771 unsigned int order, int migratetype) { return false; }
772 static inline void clear_page_guard(struct zone *zone, struct page *page,
773 unsigned int order, int migratetype) {}
776 static inline void set_page_order(struct page *page, unsigned int order)
778 set_page_private(page, order);
779 __SetPageBuddy(page);
783 * This function checks whether a page is free && is the buddy
784 * we can coalesce a page and its buddy if
785 * (a) the buddy is not in a hole (check before calling!) &&
786 * (b) the buddy is in the buddy system &&
787 * (c) a page and its buddy have the same order &&
788 * (d) a page and its buddy are in the same zone.
790 * For recording whether a page is in the buddy system, we set PageBuddy.
791 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
793 * For recording page's order, we use page_private(page).
795 static inline int page_is_buddy(struct page *page, struct page *buddy,
798 if (page_is_guard(buddy) && page_order(buddy) == order) {
799 if (page_zone_id(page) != page_zone_id(buddy))
802 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
807 if (PageBuddy(buddy) && page_order(buddy) == order) {
809 * zone check is done late to avoid uselessly
810 * calculating zone/node ids for pages that could
813 if (page_zone_id(page) != page_zone_id(buddy))
816 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
823 #ifdef CONFIG_COMPACTION
824 static inline struct capture_control *task_capc(struct zone *zone)
826 struct capture_control *capc = current->capture_control;
829 !(current->flags & PF_KTHREAD) &&
831 capc->cc->zone == zone &&
832 capc->cc->direct_compaction ? capc : NULL;
836 compaction_capture(struct capture_control *capc, struct page *page,
837 int order, int migratetype)
839 if (!capc || order != capc->cc->order)
842 /* Do not accidentally pollute CMA or isolated regions*/
843 if (is_migrate_cma(migratetype) ||
844 is_migrate_isolate(migratetype))
848 * Do not let lower order allocations polluate a movable pageblock.
849 * This might let an unmovable request use a reclaimable pageblock
850 * and vice-versa but no more than normal fallback logic which can
851 * have trouble finding a high-order free page.
853 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
861 static inline struct capture_control *task_capc(struct zone *zone)
867 compaction_capture(struct capture_control *capc, struct page *page,
868 int order, int migratetype)
872 #endif /* CONFIG_COMPACTION */
875 * Freeing function for a buddy system allocator.
877 * The concept of a buddy system is to maintain direct-mapped table
878 * (containing bit values) for memory blocks of various "orders".
879 * The bottom level table contains the map for the smallest allocatable
880 * units of memory (here, pages), and each level above it describes
881 * pairs of units from the levels below, hence, "buddies".
882 * At a high level, all that happens here is marking the table entry
883 * at the bottom level available, and propagating the changes upward
884 * as necessary, plus some accounting needed to play nicely with other
885 * parts of the VM system.
886 * At each level, we keep a list of pages, which are heads of continuous
887 * free pages of length of (1 << order) and marked with PageBuddy.
888 * Page's order is recorded in page_private(page) field.
889 * So when we are allocating or freeing one, we can derive the state of the
890 * other. That is, if we allocate a small block, and both were
891 * free, the remainder of the region must be split into blocks.
892 * If a block is freed, and its buddy is also free, then this
893 * triggers coalescing into a block of larger size.
898 static inline void __free_one_page(struct page *page,
900 struct zone *zone, unsigned int order,
903 unsigned long combined_pfn;
904 unsigned long uninitialized_var(buddy_pfn);
906 unsigned int max_order;
907 struct capture_control *capc = task_capc(zone);
909 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
911 VM_BUG_ON(!zone_is_initialized(zone));
912 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
914 VM_BUG_ON(migratetype == -1);
915 if (likely(!is_migrate_isolate(migratetype)))
916 __mod_zone_freepage_state(zone, 1 << order, migratetype);
918 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
919 VM_BUG_ON_PAGE(bad_range(zone, page), page);
922 while (order < max_order - 1) {
923 if (compaction_capture(capc, page, order, migratetype)) {
924 __mod_zone_freepage_state(zone, -(1 << order),
928 buddy_pfn = __find_buddy_pfn(pfn, order);
929 buddy = page + (buddy_pfn - pfn);
931 if (!pfn_valid_within(buddy_pfn))
933 if (!page_is_buddy(page, buddy, order))
936 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
937 * merge with it and move up one order.
939 if (page_is_guard(buddy))
940 clear_page_guard(zone, buddy, order, migratetype);
942 del_page_from_free_area(buddy, &zone->free_area[order]);
943 combined_pfn = buddy_pfn & pfn;
944 page = page + (combined_pfn - pfn);
948 if (max_order < MAX_ORDER) {
949 /* If we are here, it means order is >= pageblock_order.
950 * We want to prevent merge between freepages on isolate
951 * pageblock and normal pageblock. Without this, pageblock
952 * isolation could cause incorrect freepage or CMA accounting.
954 * We don't want to hit this code for the more frequent
957 if (unlikely(has_isolate_pageblock(zone))) {
960 buddy_pfn = __find_buddy_pfn(pfn, order);
961 buddy = page + (buddy_pfn - pfn);
962 buddy_mt = get_pageblock_migratetype(buddy);
964 if (migratetype != buddy_mt
965 && (is_migrate_isolate(migratetype) ||
966 is_migrate_isolate(buddy_mt)))
970 goto continue_merging;
974 set_page_order(page, order);
977 * If this is not the largest possible page, check if the buddy
978 * of the next-highest order is free. If it is, it's possible
979 * that pages are being freed that will coalesce soon. In case,
980 * that is happening, add the free page to the tail of the list
981 * so it's less likely to be used soon and more likely to be merged
982 * as a higher order page
984 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)
985 && !is_shuffle_order(order)) {
986 struct page *higher_page, *higher_buddy;
987 combined_pfn = buddy_pfn & pfn;
988 higher_page = page + (combined_pfn - pfn);
989 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
990 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
991 if (pfn_valid_within(buddy_pfn) &&
992 page_is_buddy(higher_page, higher_buddy, order + 1)) {
993 add_to_free_area_tail(page, &zone->free_area[order],
999 if (is_shuffle_order(order))
1000 add_to_free_area_random(page, &zone->free_area[order],
1003 add_to_free_area(page, &zone->free_area[order], migratetype);
1008 * A bad page could be due to a number of fields. Instead of multiple branches,
1009 * try and check multiple fields with one check. The caller must do a detailed
1010 * check if necessary.
1012 static inline bool page_expected_state(struct page *page,
1013 unsigned long check_flags)
1015 if (unlikely(atomic_read(&page->_mapcount) != -1))
1018 if (unlikely((unsigned long)page->mapping |
1019 page_ref_count(page) |
1021 (unsigned long)page->mem_cgroup |
1023 (page->flags & check_flags)))
1029 static void free_pages_check_bad(struct page *page)
1031 const char *bad_reason;
1032 unsigned long bad_flags;
1037 if (unlikely(atomic_read(&page->_mapcount) != -1))
1038 bad_reason = "nonzero mapcount";
1039 if (unlikely(page->mapping != NULL))
1040 bad_reason = "non-NULL mapping";
1041 if (unlikely(page_ref_count(page) != 0))
1042 bad_reason = "nonzero _refcount";
1043 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1044 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1045 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1048 if (unlikely(page->mem_cgroup))
1049 bad_reason = "page still charged to cgroup";
1051 bad_page(page, bad_reason, bad_flags);
1054 static inline int free_pages_check(struct page *page)
1056 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1059 /* Something has gone sideways, find it */
1060 free_pages_check_bad(page);
1064 static int free_tail_pages_check(struct page *head_page, struct page *page)
1069 * We rely page->lru.next never has bit 0 set, unless the page
1070 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1072 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1074 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1078 switch (page - head_page) {
1080 /* the first tail page: ->mapping may be compound_mapcount() */
1081 if (unlikely(compound_mapcount(page))) {
1082 bad_page(page, "nonzero compound_mapcount", 0);
1088 * the second tail page: ->mapping is
1089 * deferred_list.next -- ignore value.
1093 if (page->mapping != TAIL_MAPPING) {
1094 bad_page(page, "corrupted mapping in tail page", 0);
1099 if (unlikely(!PageTail(page))) {
1100 bad_page(page, "PageTail not set", 0);
1103 if (unlikely(compound_head(page) != head_page)) {
1104 bad_page(page, "compound_head not consistent", 0);
1109 page->mapping = NULL;
1110 clear_compound_head(page);
1114 static void kernel_init_free_pages(struct page *page, int numpages)
1118 for (i = 0; i < numpages; i++)
1119 clear_highpage(page + i);
1122 static __always_inline bool free_pages_prepare(struct page *page,
1123 unsigned int order, bool check_free)
1127 VM_BUG_ON_PAGE(PageTail(page), page);
1129 trace_mm_page_free(page, order);
1132 * Check tail pages before head page information is cleared to
1133 * avoid checking PageCompound for order-0 pages.
1135 if (unlikely(order)) {
1136 bool compound = PageCompound(page);
1139 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1142 ClearPageDoubleMap(page);
1143 for (i = 1; i < (1 << order); i++) {
1145 bad += free_tail_pages_check(page, page + i);
1146 if (unlikely(free_pages_check(page + i))) {
1150 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1153 if (PageMappingFlags(page))
1154 page->mapping = NULL;
1155 if (memcg_kmem_enabled() && PageKmemcg(page))
1156 __memcg_kmem_uncharge(page, order);
1158 bad += free_pages_check(page);
1162 page_cpupid_reset_last(page);
1163 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1164 reset_page_owner(page, order);
1166 if (!PageHighMem(page)) {
1167 debug_check_no_locks_freed(page_address(page),
1168 PAGE_SIZE << order);
1169 debug_check_no_obj_freed(page_address(page),
1170 PAGE_SIZE << order);
1172 if (want_init_on_free())
1173 kernel_init_free_pages(page, 1 << order);
1175 kernel_poison_pages(page, 1 << order, 0);
1177 * arch_free_page() can make the page's contents inaccessible. s390
1178 * does this. So nothing which can access the page's contents should
1179 * happen after this.
1181 arch_free_page(page, order);
1183 if (debug_pagealloc_enabled_static())
1184 kernel_map_pages(page, 1 << order, 0);
1186 kasan_free_nondeferred_pages(page, order);
1191 #ifdef CONFIG_DEBUG_VM
1193 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1194 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1195 * moved from pcp lists to free lists.
1197 static bool free_pcp_prepare(struct page *page)
1199 return free_pages_prepare(page, 0, true);
1202 static bool bulkfree_pcp_prepare(struct page *page)
1204 if (debug_pagealloc_enabled_static())
1205 return free_pages_check(page);
1211 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1212 * moving from pcp lists to free list in order to reduce overhead. With
1213 * debug_pagealloc enabled, they are checked also immediately when being freed
1216 static bool free_pcp_prepare(struct page *page)
1218 if (debug_pagealloc_enabled_static())
1219 return free_pages_prepare(page, 0, true);
1221 return free_pages_prepare(page, 0, false);
1224 static bool bulkfree_pcp_prepare(struct page *page)
1226 return free_pages_check(page);
1228 #endif /* CONFIG_DEBUG_VM */
1230 static inline void prefetch_buddy(struct page *page)
1232 unsigned long pfn = page_to_pfn(page);
1233 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1234 struct page *buddy = page + (buddy_pfn - pfn);
1240 * Frees a number of pages from the PCP lists
1241 * Assumes all pages on list are in same zone, and of same order.
1242 * count is the number of pages to free.
1244 * If the zone was previously in an "all pages pinned" state then look to
1245 * see if this freeing clears that state.
1247 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1248 * pinned" detection logic.
1250 static void free_pcppages_bulk(struct zone *zone, int count,
1251 struct per_cpu_pages *pcp)
1253 int migratetype = 0;
1255 int prefetch_nr = 0;
1256 bool isolated_pageblocks;
1257 struct page *page, *tmp;
1261 struct list_head *list;
1264 * Remove pages from lists in a round-robin fashion. A
1265 * batch_free count is maintained that is incremented when an
1266 * empty list is encountered. This is so more pages are freed
1267 * off fuller lists instead of spinning excessively around empty
1272 if (++migratetype == MIGRATE_PCPTYPES)
1274 list = &pcp->lists[migratetype];
1275 } while (list_empty(list));
1277 /* This is the only non-empty list. Free them all. */
1278 if (batch_free == MIGRATE_PCPTYPES)
1282 page = list_last_entry(list, struct page, lru);
1283 /* must delete to avoid corrupting pcp list */
1284 list_del(&page->lru);
1287 if (bulkfree_pcp_prepare(page))
1290 list_add_tail(&page->lru, &head);
1293 * We are going to put the page back to the global
1294 * pool, prefetch its buddy to speed up later access
1295 * under zone->lock. It is believed the overhead of
1296 * an additional test and calculating buddy_pfn here
1297 * can be offset by reduced memory latency later. To
1298 * avoid excessive prefetching due to large count, only
1299 * prefetch buddy for the first pcp->batch nr of pages.
1301 if (prefetch_nr++ < pcp->batch)
1302 prefetch_buddy(page);
1303 } while (--count && --batch_free && !list_empty(list));
1306 spin_lock(&zone->lock);
1307 isolated_pageblocks = has_isolate_pageblock(zone);
1310 * Use safe version since after __free_one_page(),
1311 * page->lru.next will not point to original list.
1313 list_for_each_entry_safe(page, tmp, &head, lru) {
1314 int mt = get_pcppage_migratetype(page);
1315 /* MIGRATE_ISOLATE page should not go to pcplists */
1316 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1317 /* Pageblock could have been isolated meanwhile */
1318 if (unlikely(isolated_pageblocks))
1319 mt = get_pageblock_migratetype(page);
1321 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1322 trace_mm_page_pcpu_drain(page, 0, mt);
1324 spin_unlock(&zone->lock);
1327 static void free_one_page(struct zone *zone,
1328 struct page *page, unsigned long pfn,
1332 spin_lock(&zone->lock);
1333 if (unlikely(has_isolate_pageblock(zone) ||
1334 is_migrate_isolate(migratetype))) {
1335 migratetype = get_pfnblock_migratetype(page, pfn);
1337 __free_one_page(page, pfn, zone, order, migratetype);
1338 spin_unlock(&zone->lock);
1341 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1342 unsigned long zone, int nid)
1344 mm_zero_struct_page(page);
1345 set_page_links(page, zone, nid, pfn);
1346 init_page_count(page);
1347 page_mapcount_reset(page);
1348 page_cpupid_reset_last(page);
1349 page_kasan_tag_reset(page);
1351 INIT_LIST_HEAD(&page->lru);
1352 #ifdef WANT_PAGE_VIRTUAL
1353 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1354 if (!is_highmem_idx(zone))
1355 set_page_address(page, __va(pfn << PAGE_SHIFT));
1359 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1360 static void __meminit init_reserved_page(unsigned long pfn)
1365 if (!early_page_uninitialised(pfn))
1368 nid = early_pfn_to_nid(pfn);
1369 pgdat = NODE_DATA(nid);
1371 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1372 struct zone *zone = &pgdat->node_zones[zid];
1374 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1377 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1380 static inline void init_reserved_page(unsigned long pfn)
1383 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1386 * Initialised pages do not have PageReserved set. This function is
1387 * called for each range allocated by the bootmem allocator and
1388 * marks the pages PageReserved. The remaining valid pages are later
1389 * sent to the buddy page allocator.
1391 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1393 unsigned long start_pfn = PFN_DOWN(start);
1394 unsigned long end_pfn = PFN_UP(end);
1396 for (; start_pfn < end_pfn; start_pfn++) {
1397 if (pfn_valid(start_pfn)) {
1398 struct page *page = pfn_to_page(start_pfn);
1400 init_reserved_page(start_pfn);
1402 /* Avoid false-positive PageTail() */
1403 INIT_LIST_HEAD(&page->lru);
1406 * no need for atomic set_bit because the struct
1407 * page is not visible yet so nobody should
1410 __SetPageReserved(page);
1415 static void __free_pages_ok(struct page *page, unsigned int order)
1417 unsigned long flags;
1419 unsigned long pfn = page_to_pfn(page);
1421 if (!free_pages_prepare(page, order, true))
1424 migratetype = get_pfnblock_migratetype(page, pfn);
1425 local_irq_save(flags);
1426 __count_vm_events(PGFREE, 1 << order);
1427 free_one_page(page_zone(page), page, pfn, order, migratetype);
1428 local_irq_restore(flags);
1431 void __free_pages_core(struct page *page, unsigned int order)
1433 unsigned int nr_pages = 1 << order;
1434 struct page *p = page;
1438 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1440 __ClearPageReserved(p);
1441 set_page_count(p, 0);
1443 __ClearPageReserved(p);
1444 set_page_count(p, 0);
1446 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1447 set_page_refcounted(page);
1448 __free_pages(page, order);
1451 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1452 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1454 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1456 int __meminit early_pfn_to_nid(unsigned long pfn)
1458 static DEFINE_SPINLOCK(early_pfn_lock);
1461 spin_lock(&early_pfn_lock);
1462 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1464 nid = first_online_node;
1465 spin_unlock(&early_pfn_lock);
1471 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1472 /* Only safe to use early in boot when initialisation is single-threaded */
1473 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1477 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1478 if (nid >= 0 && nid != node)
1484 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1491 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1494 if (early_page_uninitialised(pfn))
1496 __free_pages_core(page, order);
1500 * Check that the whole (or subset of) a pageblock given by the interval of
1501 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1502 * with the migration of free compaction scanner. The scanners then need to
1503 * use only pfn_valid_within() check for arches that allow holes within
1506 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1508 * It's possible on some configurations to have a setup like node0 node1 node0
1509 * i.e. it's possible that all pages within a zones range of pages do not
1510 * belong to a single zone. We assume that a border between node0 and node1
1511 * can occur within a single pageblock, but not a node0 node1 node0
1512 * interleaving within a single pageblock. It is therefore sufficient to check
1513 * the first and last page of a pageblock and avoid checking each individual
1514 * page in a pageblock.
1516 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1517 unsigned long end_pfn, struct zone *zone)
1519 struct page *start_page;
1520 struct page *end_page;
1522 /* end_pfn is one past the range we are checking */
1525 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1528 start_page = pfn_to_online_page(start_pfn);
1532 if (page_zone(start_page) != zone)
1535 end_page = pfn_to_page(end_pfn);
1537 /* This gives a shorter code than deriving page_zone(end_page) */
1538 if (page_zone_id(start_page) != page_zone_id(end_page))
1544 void set_zone_contiguous(struct zone *zone)
1546 unsigned long block_start_pfn = zone->zone_start_pfn;
1547 unsigned long block_end_pfn;
1549 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1550 for (; block_start_pfn < zone_end_pfn(zone);
1551 block_start_pfn = block_end_pfn,
1552 block_end_pfn += pageblock_nr_pages) {
1554 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1556 if (!__pageblock_pfn_to_page(block_start_pfn,
1557 block_end_pfn, zone))
1561 /* We confirm that there is no hole */
1562 zone->contiguous = true;
1565 void clear_zone_contiguous(struct zone *zone)
1567 zone->contiguous = false;
1570 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1571 static void __init deferred_free_range(unsigned long pfn,
1572 unsigned long nr_pages)
1580 page = pfn_to_page(pfn);
1582 /* Free a large naturally-aligned chunk if possible */
1583 if (nr_pages == pageblock_nr_pages &&
1584 (pfn & (pageblock_nr_pages - 1)) == 0) {
1585 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1586 __free_pages_core(page, pageblock_order);
1590 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1591 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1592 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1593 __free_pages_core(page, 0);
1597 /* Completion tracking for deferred_init_memmap() threads */
1598 static atomic_t pgdat_init_n_undone __initdata;
1599 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1601 static inline void __init pgdat_init_report_one_done(void)
1603 if (atomic_dec_and_test(&pgdat_init_n_undone))
1604 complete(&pgdat_init_all_done_comp);
1608 * Returns true if page needs to be initialized or freed to buddy allocator.
1610 * First we check if pfn is valid on architectures where it is possible to have
1611 * holes within pageblock_nr_pages. On systems where it is not possible, this
1612 * function is optimized out.
1614 * Then, we check if a current large page is valid by only checking the validity
1617 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1619 if (!pfn_valid_within(pfn))
1621 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1627 * Free pages to buddy allocator. Try to free aligned pages in
1628 * pageblock_nr_pages sizes.
1630 static void __init deferred_free_pages(unsigned long pfn,
1631 unsigned long end_pfn)
1633 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1634 unsigned long nr_free = 0;
1636 for (; pfn < end_pfn; pfn++) {
1637 if (!deferred_pfn_valid(pfn)) {
1638 deferred_free_range(pfn - nr_free, nr_free);
1640 } else if (!(pfn & nr_pgmask)) {
1641 deferred_free_range(pfn - nr_free, nr_free);
1643 touch_nmi_watchdog();
1648 /* Free the last block of pages to allocator */
1649 deferred_free_range(pfn - nr_free, nr_free);
1653 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1654 * by performing it only once every pageblock_nr_pages.
1655 * Return number of pages initialized.
1657 static unsigned long __init deferred_init_pages(struct zone *zone,
1659 unsigned long end_pfn)
1661 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1662 int nid = zone_to_nid(zone);
1663 unsigned long nr_pages = 0;
1664 int zid = zone_idx(zone);
1665 struct page *page = NULL;
1667 for (; pfn < end_pfn; pfn++) {
1668 if (!deferred_pfn_valid(pfn)) {
1671 } else if (!page || !(pfn & nr_pgmask)) {
1672 page = pfn_to_page(pfn);
1673 touch_nmi_watchdog();
1677 __init_single_page(page, pfn, zid, nid);
1684 * This function is meant to pre-load the iterator for the zone init.
1685 * Specifically it walks through the ranges until we are caught up to the
1686 * first_init_pfn value and exits there. If we never encounter the value we
1687 * return false indicating there are no valid ranges left.
1690 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1691 unsigned long *spfn, unsigned long *epfn,
1692 unsigned long first_init_pfn)
1697 * Start out by walking through the ranges in this zone that have
1698 * already been initialized. We don't need to do anything with them
1699 * so we just need to flush them out of the system.
1701 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1702 if (*epfn <= first_init_pfn)
1704 if (*spfn < first_init_pfn)
1705 *spfn = first_init_pfn;
1714 * Initialize and free pages. We do it in two loops: first we initialize
1715 * struct page, then free to buddy allocator, because while we are
1716 * freeing pages we can access pages that are ahead (computing buddy
1717 * page in __free_one_page()).
1719 * In order to try and keep some memory in the cache we have the loop
1720 * broken along max page order boundaries. This way we will not cause
1721 * any issues with the buddy page computation.
1723 static unsigned long __init
1724 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1725 unsigned long *end_pfn)
1727 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1728 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1729 unsigned long nr_pages = 0;
1732 /* First we loop through and initialize the page values */
1733 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1736 if (mo_pfn <= *start_pfn)
1739 t = min(mo_pfn, *end_pfn);
1740 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1742 if (mo_pfn < *end_pfn) {
1743 *start_pfn = mo_pfn;
1748 /* Reset values and now loop through freeing pages as needed */
1751 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1757 t = min(mo_pfn, epfn);
1758 deferred_free_pages(spfn, t);
1767 /* Initialise remaining memory on a node */
1768 static int __init deferred_init_memmap(void *data)
1770 pg_data_t *pgdat = data;
1771 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1772 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1773 unsigned long first_init_pfn, flags;
1774 unsigned long start = jiffies;
1779 /* Bind memory initialisation thread to a local node if possible */
1780 if (!cpumask_empty(cpumask))
1781 set_cpus_allowed_ptr(current, cpumask);
1783 pgdat_resize_lock(pgdat, &flags);
1784 first_init_pfn = pgdat->first_deferred_pfn;
1785 if (first_init_pfn == ULONG_MAX) {
1786 pgdat_resize_unlock(pgdat, &flags);
1787 pgdat_init_report_one_done();
1791 /* Sanity check boundaries */
1792 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1793 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1794 pgdat->first_deferred_pfn = ULONG_MAX;
1796 /* Only the highest zone is deferred so find it */
1797 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1798 zone = pgdat->node_zones + zid;
1799 if (first_init_pfn < zone_end_pfn(zone))
1803 /* If the zone is empty somebody else may have cleared out the zone */
1804 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1809 * Initialize and free pages in MAX_ORDER sized increments so
1810 * that we can avoid introducing any issues with the buddy
1814 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1816 pgdat_resize_unlock(pgdat, &flags);
1818 /* Sanity check that the next zone really is unpopulated */
1819 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1821 pr_info("node %d initialised, %lu pages in %ums\n",
1822 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1824 pgdat_init_report_one_done();
1829 * If this zone has deferred pages, try to grow it by initializing enough
1830 * deferred pages to satisfy the allocation specified by order, rounded up to
1831 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1832 * of SECTION_SIZE bytes by initializing struct pages in increments of
1833 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1835 * Return true when zone was grown, otherwise return false. We return true even
1836 * when we grow less than requested, to let the caller decide if there are
1837 * enough pages to satisfy the allocation.
1839 * Note: We use noinline because this function is needed only during boot, and
1840 * it is called from a __ref function _deferred_grow_zone. This way we are
1841 * making sure that it is not inlined into permanent text section.
1843 static noinline bool __init
1844 deferred_grow_zone(struct zone *zone, unsigned int order)
1846 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1847 pg_data_t *pgdat = zone->zone_pgdat;
1848 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1849 unsigned long spfn, epfn, flags;
1850 unsigned long nr_pages = 0;
1853 /* Only the last zone may have deferred pages */
1854 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1857 pgdat_resize_lock(pgdat, &flags);
1860 * If deferred pages have been initialized while we were waiting for
1861 * the lock, return true, as the zone was grown. The caller will retry
1862 * this zone. We won't return to this function since the caller also
1863 * has this static branch.
1865 if (!static_branch_unlikely(&deferred_pages)) {
1866 pgdat_resize_unlock(pgdat, &flags);
1871 * If someone grew this zone while we were waiting for spinlock, return
1872 * true, as there might be enough pages already.
1874 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1875 pgdat_resize_unlock(pgdat, &flags);
1879 /* If the zone is empty somebody else may have cleared out the zone */
1880 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1881 first_deferred_pfn)) {
1882 pgdat->first_deferred_pfn = ULONG_MAX;
1883 pgdat_resize_unlock(pgdat, &flags);
1884 /* Retry only once. */
1885 return first_deferred_pfn != ULONG_MAX;
1889 * Initialize and free pages in MAX_ORDER sized increments so
1890 * that we can avoid introducing any issues with the buddy
1893 while (spfn < epfn) {
1894 /* update our first deferred PFN for this section */
1895 first_deferred_pfn = spfn;
1897 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1899 /* We should only stop along section boundaries */
1900 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1903 /* If our quota has been met we can stop here */
1904 if (nr_pages >= nr_pages_needed)
1908 pgdat->first_deferred_pfn = spfn;
1909 pgdat_resize_unlock(pgdat, &flags);
1911 return nr_pages > 0;
1915 * deferred_grow_zone() is __init, but it is called from
1916 * get_page_from_freelist() during early boot until deferred_pages permanently
1917 * disables this call. This is why we have refdata wrapper to avoid warning,
1918 * and to ensure that the function body gets unloaded.
1921 _deferred_grow_zone(struct zone *zone, unsigned int order)
1923 return deferred_grow_zone(zone, order);
1926 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1928 void __init page_alloc_init_late(void)
1933 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1935 /* There will be num_node_state(N_MEMORY) threads */
1936 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1937 for_each_node_state(nid, N_MEMORY) {
1938 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1941 /* Block until all are initialised */
1942 wait_for_completion(&pgdat_init_all_done_comp);
1945 * The number of managed pages has changed due to the initialisation
1946 * so the pcpu batch and high limits needs to be updated or the limits
1947 * will be artificially small.
1949 for_each_populated_zone(zone)
1950 zone_pcp_update(zone);
1953 * We initialized the rest of the deferred pages. Permanently disable
1954 * on-demand struct page initialization.
1956 static_branch_disable(&deferred_pages);
1958 /* Reinit limits that are based on free pages after the kernel is up */
1959 files_maxfiles_init();
1962 /* Discard memblock private memory */
1965 for_each_node_state(nid, N_MEMORY)
1966 shuffle_free_memory(NODE_DATA(nid));
1968 for_each_populated_zone(zone)
1969 set_zone_contiguous(zone);
1973 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1974 void __init init_cma_reserved_pageblock(struct page *page)
1976 unsigned i = pageblock_nr_pages;
1977 struct page *p = page;
1980 __ClearPageReserved(p);
1981 set_page_count(p, 0);
1984 set_pageblock_migratetype(page, MIGRATE_CMA);
1986 if (pageblock_order >= MAX_ORDER) {
1987 i = pageblock_nr_pages;
1990 set_page_refcounted(p);
1991 __free_pages(p, MAX_ORDER - 1);
1992 p += MAX_ORDER_NR_PAGES;
1993 } while (i -= MAX_ORDER_NR_PAGES);
1995 set_page_refcounted(page);
1996 __free_pages(page, pageblock_order);
1999 adjust_managed_page_count(page, pageblock_nr_pages);
2004 * The order of subdivision here is critical for the IO subsystem.
2005 * Please do not alter this order without good reasons and regression
2006 * testing. Specifically, as large blocks of memory are subdivided,
2007 * the order in which smaller blocks are delivered depends on the order
2008 * they're subdivided in this function. This is the primary factor
2009 * influencing the order in which pages are delivered to the IO
2010 * subsystem according to empirical testing, and this is also justified
2011 * by considering the behavior of a buddy system containing a single
2012 * large block of memory acted on by a series of small allocations.
2013 * This behavior is a critical factor in sglist merging's success.
2017 static inline void expand(struct zone *zone, struct page *page,
2018 int low, int high, struct free_area *area,
2021 unsigned long size = 1 << high;
2023 while (high > low) {
2027 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2030 * Mark as guard pages (or page), that will allow to
2031 * merge back to allocator when buddy will be freed.
2032 * Corresponding page table entries will not be touched,
2033 * pages will stay not present in virtual address space
2035 if (set_page_guard(zone, &page[size], high, migratetype))
2038 add_to_free_area(&page[size], area, migratetype);
2039 set_page_order(&page[size], high);
2043 static void check_new_page_bad(struct page *page)
2045 const char *bad_reason = NULL;
2046 unsigned long bad_flags = 0;
2048 if (unlikely(atomic_read(&page->_mapcount) != -1))
2049 bad_reason = "nonzero mapcount";
2050 if (unlikely(page->mapping != NULL))
2051 bad_reason = "non-NULL mapping";
2052 if (unlikely(page_ref_count(page) != 0))
2053 bad_reason = "nonzero _refcount";
2054 if (unlikely(page->flags & __PG_HWPOISON)) {
2055 bad_reason = "HWPoisoned (hardware-corrupted)";
2056 bad_flags = __PG_HWPOISON;
2057 /* Don't complain about hwpoisoned pages */
2058 page_mapcount_reset(page); /* remove PageBuddy */
2061 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2062 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2063 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2066 if (unlikely(page->mem_cgroup))
2067 bad_reason = "page still charged to cgroup";
2069 bad_page(page, bad_reason, bad_flags);
2073 * This page is about to be returned from the page allocator
2075 static inline int check_new_page(struct page *page)
2077 if (likely(page_expected_state(page,
2078 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2081 check_new_page_bad(page);
2085 static inline bool free_pages_prezeroed(void)
2087 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2088 page_poisoning_enabled()) || want_init_on_free();
2091 #ifdef CONFIG_DEBUG_VM
2093 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2094 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2095 * also checked when pcp lists are refilled from the free lists.
2097 static inline bool check_pcp_refill(struct page *page)
2099 if (debug_pagealloc_enabled_static())
2100 return check_new_page(page);
2105 static inline bool check_new_pcp(struct page *page)
2107 return check_new_page(page);
2111 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2112 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2113 * enabled, they are also checked when being allocated from the pcp lists.
2115 static inline bool check_pcp_refill(struct page *page)
2117 return check_new_page(page);
2119 static inline bool check_new_pcp(struct page *page)
2121 if (debug_pagealloc_enabled_static())
2122 return check_new_page(page);
2126 #endif /* CONFIG_DEBUG_VM */
2128 static bool check_new_pages(struct page *page, unsigned int order)
2131 for (i = 0; i < (1 << order); i++) {
2132 struct page *p = page + i;
2134 if (unlikely(check_new_page(p)))
2141 inline void post_alloc_hook(struct page *page, unsigned int order,
2144 set_page_private(page, 0);
2145 set_page_refcounted(page);
2147 arch_alloc_page(page, order);
2148 if (debug_pagealloc_enabled_static())
2149 kernel_map_pages(page, 1 << order, 1);
2150 kasan_alloc_pages(page, order);
2151 kernel_poison_pages(page, 1 << order, 1);
2152 set_page_owner(page, order, gfp_flags);
2155 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2156 unsigned int alloc_flags)
2158 post_alloc_hook(page, order, gfp_flags);
2160 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2161 kernel_init_free_pages(page, 1 << order);
2163 if (order && (gfp_flags & __GFP_COMP))
2164 prep_compound_page(page, order);
2167 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2168 * allocate the page. The expectation is that the caller is taking
2169 * steps that will free more memory. The caller should avoid the page
2170 * being used for !PFMEMALLOC purposes.
2172 if (alloc_flags & ALLOC_NO_WATERMARKS)
2173 set_page_pfmemalloc(page);
2175 clear_page_pfmemalloc(page);
2179 * Go through the free lists for the given migratetype and remove
2180 * the smallest available page from the freelists
2182 static __always_inline
2183 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2186 unsigned int current_order;
2187 struct free_area *area;
2190 /* Find a page of the appropriate size in the preferred list */
2191 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2192 area = &(zone->free_area[current_order]);
2193 page = get_page_from_free_area(area, migratetype);
2196 del_page_from_free_area(page, area);
2197 expand(zone, page, order, current_order, area, migratetype);
2198 set_pcppage_migratetype(page, migratetype);
2207 * This array describes the order lists are fallen back to when
2208 * the free lists for the desirable migrate type are depleted
2210 static int fallbacks[MIGRATE_TYPES][4] = {
2211 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2212 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2213 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2215 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2217 #ifdef CONFIG_MEMORY_ISOLATION
2218 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2223 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2226 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2229 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2230 unsigned int order) { return NULL; }
2234 * Move the free pages in a range to the free lists of the requested type.
2235 * Note that start_page and end_pages are not aligned on a pageblock
2236 * boundary. If alignment is required, use move_freepages_block()
2238 static int move_freepages(struct zone *zone,
2239 struct page *start_page, struct page *end_page,
2240 int migratetype, int *num_movable)
2244 int pages_moved = 0;
2246 for (page = start_page; page <= end_page;) {
2247 if (!pfn_valid_within(page_to_pfn(page))) {
2252 if (!PageBuddy(page)) {
2254 * We assume that pages that could be isolated for
2255 * migration are movable. But we don't actually try
2256 * isolating, as that would be expensive.
2259 (PageLRU(page) || __PageMovable(page)))
2266 /* Make sure we are not inadvertently changing nodes */
2267 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2268 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2270 order = page_order(page);
2271 move_to_free_area(page, &zone->free_area[order], migratetype);
2273 pages_moved += 1 << order;
2279 int move_freepages_block(struct zone *zone, struct page *page,
2280 int migratetype, int *num_movable)
2282 unsigned long start_pfn, end_pfn;
2283 struct page *start_page, *end_page;
2288 start_pfn = page_to_pfn(page);
2289 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2290 start_page = pfn_to_page(start_pfn);
2291 end_page = start_page + pageblock_nr_pages - 1;
2292 end_pfn = start_pfn + pageblock_nr_pages - 1;
2294 /* Do not cross zone boundaries */
2295 if (!zone_spans_pfn(zone, start_pfn))
2297 if (!zone_spans_pfn(zone, end_pfn))
2300 return move_freepages(zone, start_page, end_page, migratetype,
2304 static void change_pageblock_range(struct page *pageblock_page,
2305 int start_order, int migratetype)
2307 int nr_pageblocks = 1 << (start_order - pageblock_order);
2309 while (nr_pageblocks--) {
2310 set_pageblock_migratetype(pageblock_page, migratetype);
2311 pageblock_page += pageblock_nr_pages;
2316 * When we are falling back to another migratetype during allocation, try to
2317 * steal extra free pages from the same pageblocks to satisfy further
2318 * allocations, instead of polluting multiple pageblocks.
2320 * If we are stealing a relatively large buddy page, it is likely there will
2321 * be more free pages in the pageblock, so try to steal them all. For
2322 * reclaimable and unmovable allocations, we steal regardless of page size,
2323 * as fragmentation caused by those allocations polluting movable pageblocks
2324 * is worse than movable allocations stealing from unmovable and reclaimable
2327 static bool can_steal_fallback(unsigned int order, int start_mt)
2330 * Leaving this order check is intended, although there is
2331 * relaxed order check in next check. The reason is that
2332 * we can actually steal whole pageblock if this condition met,
2333 * but, below check doesn't guarantee it and that is just heuristic
2334 * so could be changed anytime.
2336 if (order >= pageblock_order)
2339 if (order >= pageblock_order / 2 ||
2340 start_mt == MIGRATE_RECLAIMABLE ||
2341 start_mt == MIGRATE_UNMOVABLE ||
2342 page_group_by_mobility_disabled)
2348 static inline void boost_watermark(struct zone *zone)
2350 unsigned long max_boost;
2352 if (!watermark_boost_factor)
2355 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2356 watermark_boost_factor, 10000);
2359 * high watermark may be uninitialised if fragmentation occurs
2360 * very early in boot so do not boost. We do not fall
2361 * through and boost by pageblock_nr_pages as failing
2362 * allocations that early means that reclaim is not going
2363 * to help and it may even be impossible to reclaim the
2364 * boosted watermark resulting in a hang.
2369 max_boost = max(pageblock_nr_pages, max_boost);
2371 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2376 * This function implements actual steal behaviour. If order is large enough,
2377 * we can steal whole pageblock. If not, we first move freepages in this
2378 * pageblock to our migratetype and determine how many already-allocated pages
2379 * are there in the pageblock with a compatible migratetype. If at least half
2380 * of pages are free or compatible, we can change migratetype of the pageblock
2381 * itself, so pages freed in the future will be put on the correct free list.
2383 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2384 unsigned int alloc_flags, int start_type, bool whole_block)
2386 unsigned int current_order = page_order(page);
2387 struct free_area *area;
2388 int free_pages, movable_pages, alike_pages;
2391 old_block_type = get_pageblock_migratetype(page);
2394 * This can happen due to races and we want to prevent broken
2395 * highatomic accounting.
2397 if (is_migrate_highatomic(old_block_type))
2400 /* Take ownership for orders >= pageblock_order */
2401 if (current_order >= pageblock_order) {
2402 change_pageblock_range(page, current_order, start_type);
2407 * Boost watermarks to increase reclaim pressure to reduce the
2408 * likelihood of future fallbacks. Wake kswapd now as the node
2409 * may be balanced overall and kswapd will not wake naturally.
2411 boost_watermark(zone);
2412 if (alloc_flags & ALLOC_KSWAPD)
2413 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2415 /* We are not allowed to try stealing from the whole block */
2419 free_pages = move_freepages_block(zone, page, start_type,
2422 * Determine how many pages are compatible with our allocation.
2423 * For movable allocation, it's the number of movable pages which
2424 * we just obtained. For other types it's a bit more tricky.
2426 if (start_type == MIGRATE_MOVABLE) {
2427 alike_pages = movable_pages;
2430 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2431 * to MOVABLE pageblock, consider all non-movable pages as
2432 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2433 * vice versa, be conservative since we can't distinguish the
2434 * exact migratetype of non-movable pages.
2436 if (old_block_type == MIGRATE_MOVABLE)
2437 alike_pages = pageblock_nr_pages
2438 - (free_pages + movable_pages);
2443 /* moving whole block can fail due to zone boundary conditions */
2448 * If a sufficient number of pages in the block are either free or of
2449 * comparable migratability as our allocation, claim the whole block.
2451 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2452 page_group_by_mobility_disabled)
2453 set_pageblock_migratetype(page, start_type);
2458 area = &zone->free_area[current_order];
2459 move_to_free_area(page, area, start_type);
2463 * Check whether there is a suitable fallback freepage with requested order.
2464 * If only_stealable is true, this function returns fallback_mt only if
2465 * we can steal other freepages all together. This would help to reduce
2466 * fragmentation due to mixed migratetype pages in one pageblock.
2468 int find_suitable_fallback(struct free_area *area, unsigned int order,
2469 int migratetype, bool only_stealable, bool *can_steal)
2474 if (area->nr_free == 0)
2479 fallback_mt = fallbacks[migratetype][i];
2480 if (fallback_mt == MIGRATE_TYPES)
2483 if (free_area_empty(area, fallback_mt))
2486 if (can_steal_fallback(order, migratetype))
2489 if (!only_stealable)
2500 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2501 * there are no empty page blocks that contain a page with a suitable order
2503 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2504 unsigned int alloc_order)
2507 unsigned long max_managed, flags;
2510 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2511 * Check is race-prone but harmless.
2513 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2514 if (zone->nr_reserved_highatomic >= max_managed)
2517 spin_lock_irqsave(&zone->lock, flags);
2519 /* Recheck the nr_reserved_highatomic limit under the lock */
2520 if (zone->nr_reserved_highatomic >= max_managed)
2524 mt = get_pageblock_migratetype(page);
2525 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2526 && !is_migrate_cma(mt)) {
2527 zone->nr_reserved_highatomic += pageblock_nr_pages;
2528 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2529 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2533 spin_unlock_irqrestore(&zone->lock, flags);
2537 * Used when an allocation is about to fail under memory pressure. This
2538 * potentially hurts the reliability of high-order allocations when under
2539 * intense memory pressure but failed atomic allocations should be easier
2540 * to recover from than an OOM.
2542 * If @force is true, try to unreserve a pageblock even though highatomic
2543 * pageblock is exhausted.
2545 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2548 struct zonelist *zonelist = ac->zonelist;
2549 unsigned long flags;
2556 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2559 * Preserve at least one pageblock unless memory pressure
2562 if (!force && zone->nr_reserved_highatomic <=
2566 spin_lock_irqsave(&zone->lock, flags);
2567 for (order = 0; order < MAX_ORDER; order++) {
2568 struct free_area *area = &(zone->free_area[order]);
2570 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2575 * In page freeing path, migratetype change is racy so
2576 * we can counter several free pages in a pageblock
2577 * in this loop althoug we changed the pageblock type
2578 * from highatomic to ac->migratetype. So we should
2579 * adjust the count once.
2581 if (is_migrate_highatomic_page(page)) {
2583 * It should never happen but changes to
2584 * locking could inadvertently allow a per-cpu
2585 * drain to add pages to MIGRATE_HIGHATOMIC
2586 * while unreserving so be safe and watch for
2589 zone->nr_reserved_highatomic -= min(
2591 zone->nr_reserved_highatomic);
2595 * Convert to ac->migratetype and avoid the normal
2596 * pageblock stealing heuristics. Minimally, the caller
2597 * is doing the work and needs the pages. More
2598 * importantly, if the block was always converted to
2599 * MIGRATE_UNMOVABLE or another type then the number
2600 * of pageblocks that cannot be completely freed
2603 set_pageblock_migratetype(page, ac->migratetype);
2604 ret = move_freepages_block(zone, page, ac->migratetype,
2607 spin_unlock_irqrestore(&zone->lock, flags);
2611 spin_unlock_irqrestore(&zone->lock, flags);
2618 * Try finding a free buddy page on the fallback list and put it on the free
2619 * list of requested migratetype, possibly along with other pages from the same
2620 * block, depending on fragmentation avoidance heuristics. Returns true if
2621 * fallback was found so that __rmqueue_smallest() can grab it.
2623 * The use of signed ints for order and current_order is a deliberate
2624 * deviation from the rest of this file, to make the for loop
2625 * condition simpler.
2627 static __always_inline bool
2628 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2629 unsigned int alloc_flags)
2631 struct free_area *area;
2633 int min_order = order;
2639 * Do not steal pages from freelists belonging to other pageblocks
2640 * i.e. orders < pageblock_order. If there are no local zones free,
2641 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2643 if (alloc_flags & ALLOC_NOFRAGMENT)
2644 min_order = pageblock_order;
2647 * Find the largest available free page in the other list. This roughly
2648 * approximates finding the pageblock with the most free pages, which
2649 * would be too costly to do exactly.
2651 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2653 area = &(zone->free_area[current_order]);
2654 fallback_mt = find_suitable_fallback(area, current_order,
2655 start_migratetype, false, &can_steal);
2656 if (fallback_mt == -1)
2660 * We cannot steal all free pages from the pageblock and the
2661 * requested migratetype is movable. In that case it's better to
2662 * steal and split the smallest available page instead of the
2663 * largest available page, because even if the next movable
2664 * allocation falls back into a different pageblock than this
2665 * one, it won't cause permanent fragmentation.
2667 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2668 && current_order > order)
2677 for (current_order = order; current_order < MAX_ORDER;
2679 area = &(zone->free_area[current_order]);
2680 fallback_mt = find_suitable_fallback(area, current_order,
2681 start_migratetype, false, &can_steal);
2682 if (fallback_mt != -1)
2687 * This should not happen - we already found a suitable fallback
2688 * when looking for the largest page.
2690 VM_BUG_ON(current_order == MAX_ORDER);
2693 page = get_page_from_free_area(area, fallback_mt);
2695 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2698 trace_mm_page_alloc_extfrag(page, order, current_order,
2699 start_migratetype, fallback_mt);
2706 * Do the hard work of removing an element from the buddy allocator.
2707 * Call me with the zone->lock already held.
2709 static __always_inline struct page *
2710 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2711 unsigned int alloc_flags)
2716 page = __rmqueue_smallest(zone, order, migratetype);
2717 if (unlikely(!page)) {
2718 if (migratetype == MIGRATE_MOVABLE)
2719 page = __rmqueue_cma_fallback(zone, order);
2721 if (!page && __rmqueue_fallback(zone, order, migratetype,
2726 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2731 * Obtain a specified number of elements from the buddy allocator, all under
2732 * a single hold of the lock, for efficiency. Add them to the supplied list.
2733 * Returns the number of new pages which were placed at *list.
2735 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2736 unsigned long count, struct list_head *list,
2737 int migratetype, unsigned int alloc_flags)
2741 spin_lock(&zone->lock);
2742 for (i = 0; i < count; ++i) {
2743 struct page *page = __rmqueue(zone, order, migratetype,
2745 if (unlikely(page == NULL))
2748 if (unlikely(check_pcp_refill(page)))
2752 * Split buddy pages returned by expand() are received here in
2753 * physical page order. The page is added to the tail of
2754 * caller's list. From the callers perspective, the linked list
2755 * is ordered by page number under some conditions. This is
2756 * useful for IO devices that can forward direction from the
2757 * head, thus also in the physical page order. This is useful
2758 * for IO devices that can merge IO requests if the physical
2759 * pages are ordered properly.
2761 list_add_tail(&page->lru, list);
2763 if (is_migrate_cma(get_pcppage_migratetype(page)))
2764 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2769 * i pages were removed from the buddy list even if some leak due
2770 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2771 * on i. Do not confuse with 'alloced' which is the number of
2772 * pages added to the pcp list.
2774 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2775 spin_unlock(&zone->lock);
2781 * Called from the vmstat counter updater to drain pagesets of this
2782 * currently executing processor on remote nodes after they have
2785 * Note that this function must be called with the thread pinned to
2786 * a single processor.
2788 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2790 unsigned long flags;
2791 int to_drain, batch;
2793 local_irq_save(flags);
2794 batch = READ_ONCE(pcp->batch);
2795 to_drain = min(pcp->count, batch);
2797 free_pcppages_bulk(zone, to_drain, pcp);
2798 local_irq_restore(flags);
2803 * Drain pcplists of the indicated processor and zone.
2805 * The processor must either be the current processor and the
2806 * thread pinned to the current processor or a processor that
2809 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2811 unsigned long flags;
2812 struct per_cpu_pageset *pset;
2813 struct per_cpu_pages *pcp;
2815 local_irq_save(flags);
2816 pset = per_cpu_ptr(zone->pageset, cpu);
2820 free_pcppages_bulk(zone, pcp->count, pcp);
2821 local_irq_restore(flags);
2825 * Drain pcplists of all zones on the indicated processor.
2827 * The processor must either be the current processor and the
2828 * thread pinned to the current processor or a processor that
2831 static void drain_pages(unsigned int cpu)
2835 for_each_populated_zone(zone) {
2836 drain_pages_zone(cpu, zone);
2841 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2843 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2844 * the single zone's pages.
2846 void drain_local_pages(struct zone *zone)
2848 int cpu = smp_processor_id();
2851 drain_pages_zone(cpu, zone);
2856 static void drain_local_pages_wq(struct work_struct *work)
2858 struct pcpu_drain *drain;
2860 drain = container_of(work, struct pcpu_drain, work);
2863 * drain_all_pages doesn't use proper cpu hotplug protection so
2864 * we can race with cpu offline when the WQ can move this from
2865 * a cpu pinned worker to an unbound one. We can operate on a different
2866 * cpu which is allright but we also have to make sure to not move to
2870 drain_local_pages(drain->zone);
2875 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2877 * When zone parameter is non-NULL, spill just the single zone's pages.
2879 * Note that this can be extremely slow as the draining happens in a workqueue.
2881 void drain_all_pages(struct zone *zone)
2886 * Allocate in the BSS so we wont require allocation in
2887 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2889 static cpumask_t cpus_with_pcps;
2892 * Make sure nobody triggers this path before mm_percpu_wq is fully
2895 if (WARN_ON_ONCE(!mm_percpu_wq))
2899 * Do not drain if one is already in progress unless it's specific to
2900 * a zone. Such callers are primarily CMA and memory hotplug and need
2901 * the drain to be complete when the call returns.
2903 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2906 mutex_lock(&pcpu_drain_mutex);
2910 * We don't care about racing with CPU hotplug event
2911 * as offline notification will cause the notified
2912 * cpu to drain that CPU pcps and on_each_cpu_mask
2913 * disables preemption as part of its processing
2915 for_each_online_cpu(cpu) {
2916 struct per_cpu_pageset *pcp;
2918 bool has_pcps = false;
2921 pcp = per_cpu_ptr(zone->pageset, cpu);
2925 for_each_populated_zone(z) {
2926 pcp = per_cpu_ptr(z->pageset, cpu);
2927 if (pcp->pcp.count) {
2935 cpumask_set_cpu(cpu, &cpus_with_pcps);
2937 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2940 for_each_cpu(cpu, &cpus_with_pcps) {
2941 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2944 INIT_WORK(&drain->work, drain_local_pages_wq);
2945 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2947 for_each_cpu(cpu, &cpus_with_pcps)
2948 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2950 mutex_unlock(&pcpu_drain_mutex);
2953 #ifdef CONFIG_HIBERNATION
2956 * Touch the watchdog for every WD_PAGE_COUNT pages.
2958 #define WD_PAGE_COUNT (128*1024)
2960 void mark_free_pages(struct zone *zone)
2962 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2963 unsigned long flags;
2964 unsigned int order, t;
2967 if (zone_is_empty(zone))
2970 spin_lock_irqsave(&zone->lock, flags);
2972 max_zone_pfn = zone_end_pfn(zone);
2973 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2974 if (pfn_valid(pfn)) {
2975 page = pfn_to_page(pfn);
2977 if (!--page_count) {
2978 touch_nmi_watchdog();
2979 page_count = WD_PAGE_COUNT;
2982 if (page_zone(page) != zone)
2985 if (!swsusp_page_is_forbidden(page))
2986 swsusp_unset_page_free(page);
2989 for_each_migratetype_order(order, t) {
2990 list_for_each_entry(page,
2991 &zone->free_area[order].free_list[t], lru) {
2994 pfn = page_to_pfn(page);
2995 for (i = 0; i < (1UL << order); i++) {
2996 if (!--page_count) {
2997 touch_nmi_watchdog();
2998 page_count = WD_PAGE_COUNT;
3000 swsusp_set_page_free(pfn_to_page(pfn + i));
3004 spin_unlock_irqrestore(&zone->lock, flags);
3006 #endif /* CONFIG_PM */
3008 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3012 if (!free_pcp_prepare(page))
3015 migratetype = get_pfnblock_migratetype(page, pfn);
3016 set_pcppage_migratetype(page, migratetype);
3020 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3022 struct zone *zone = page_zone(page);
3023 struct per_cpu_pages *pcp;
3026 migratetype = get_pcppage_migratetype(page);
3027 __count_vm_event(PGFREE);
3030 * We only track unmovable, reclaimable and movable on pcp lists.
3031 * Free ISOLATE pages back to the allocator because they are being
3032 * offlined but treat HIGHATOMIC as movable pages so we can get those
3033 * areas back if necessary. Otherwise, we may have to free
3034 * excessively into the page allocator
3036 if (migratetype >= MIGRATE_PCPTYPES) {
3037 if (unlikely(is_migrate_isolate(migratetype))) {
3038 free_one_page(zone, page, pfn, 0, migratetype);
3041 migratetype = MIGRATE_MOVABLE;
3044 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3045 list_add(&page->lru, &pcp->lists[migratetype]);
3047 if (pcp->count >= pcp->high) {
3048 unsigned long batch = READ_ONCE(pcp->batch);
3049 free_pcppages_bulk(zone, batch, pcp);
3054 * Free a 0-order page
3056 void free_unref_page(struct page *page)
3058 unsigned long flags;
3059 unsigned long pfn = page_to_pfn(page);
3061 if (!free_unref_page_prepare(page, pfn))
3064 local_irq_save(flags);
3065 free_unref_page_commit(page, pfn);
3066 local_irq_restore(flags);
3070 * Free a list of 0-order pages
3072 void free_unref_page_list(struct list_head *list)
3074 struct page *page, *next;
3075 unsigned long flags, pfn;
3076 int batch_count = 0;
3078 /* Prepare pages for freeing */
3079 list_for_each_entry_safe(page, next, list, lru) {
3080 pfn = page_to_pfn(page);
3081 if (!free_unref_page_prepare(page, pfn))
3082 list_del(&page->lru);
3083 set_page_private(page, pfn);
3086 local_irq_save(flags);
3087 list_for_each_entry_safe(page, next, list, lru) {
3088 unsigned long pfn = page_private(page);
3090 set_page_private(page, 0);
3091 trace_mm_page_free_batched(page);
3092 free_unref_page_commit(page, pfn);
3095 * Guard against excessive IRQ disabled times when we get
3096 * a large list of pages to free.
3098 if (++batch_count == SWAP_CLUSTER_MAX) {
3099 local_irq_restore(flags);
3101 local_irq_save(flags);
3104 local_irq_restore(flags);
3108 * split_page takes a non-compound higher-order page, and splits it into
3109 * n (1<<order) sub-pages: page[0..n]
3110 * Each sub-page must be freed individually.
3112 * Note: this is probably too low level an operation for use in drivers.
3113 * Please consult with lkml before using this in your driver.
3115 void split_page(struct page *page, unsigned int order)
3119 VM_BUG_ON_PAGE(PageCompound(page), page);
3120 VM_BUG_ON_PAGE(!page_count(page), page);
3122 for (i = 1; i < (1 << order); i++)
3123 set_page_refcounted(page + i);
3124 split_page_owner(page, order);
3126 EXPORT_SYMBOL_GPL(split_page);
3128 int __isolate_free_page(struct page *page, unsigned int order)
3130 struct free_area *area = &page_zone(page)->free_area[order];
3131 unsigned long watermark;
3135 BUG_ON(!PageBuddy(page));
3137 zone = page_zone(page);
3138 mt = get_pageblock_migratetype(page);
3140 if (!is_migrate_isolate(mt)) {
3142 * Obey watermarks as if the page was being allocated. We can
3143 * emulate a high-order watermark check with a raised order-0
3144 * watermark, because we already know our high-order page
3147 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3148 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3151 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3154 /* Remove page from free list */
3156 del_page_from_free_area(page, area);
3159 * Set the pageblock if the isolated page is at least half of a
3162 if (order >= pageblock_order - 1) {
3163 struct page *endpage = page + (1 << order) - 1;
3164 for (; page < endpage; page += pageblock_nr_pages) {
3165 int mt = get_pageblock_migratetype(page);
3166 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3167 && !is_migrate_highatomic(mt))
3168 set_pageblock_migratetype(page,
3174 return 1UL << order;
3178 * Update NUMA hit/miss statistics
3180 * Must be called with interrupts disabled.
3182 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3185 enum numa_stat_item local_stat = NUMA_LOCAL;
3187 /* skip numa counters update if numa stats is disabled */
3188 if (!static_branch_likely(&vm_numa_stat_key))
3191 if (zone_to_nid(z) != numa_node_id())
3192 local_stat = NUMA_OTHER;
3194 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3195 __inc_numa_state(z, NUMA_HIT);
3197 __inc_numa_state(z, NUMA_MISS);
3198 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3200 __inc_numa_state(z, local_stat);
3204 /* Remove page from the per-cpu list, caller must protect the list */
3205 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3206 unsigned int alloc_flags,
3207 struct per_cpu_pages *pcp,
3208 struct list_head *list)
3213 if (list_empty(list)) {
3214 pcp->count += rmqueue_bulk(zone, 0,
3216 migratetype, alloc_flags);
3217 if (unlikely(list_empty(list)))
3221 page = list_first_entry(list, struct page, lru);
3222 list_del(&page->lru);
3224 } while (check_new_pcp(page));
3229 /* Lock and remove page from the per-cpu list */
3230 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3231 struct zone *zone, gfp_t gfp_flags,
3232 int migratetype, unsigned int alloc_flags)
3234 struct per_cpu_pages *pcp;
3235 struct list_head *list;
3237 unsigned long flags;
3239 local_irq_save(flags);
3240 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3241 list = &pcp->lists[migratetype];
3242 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3244 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3245 zone_statistics(preferred_zone, zone);
3247 local_irq_restore(flags);
3252 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3255 struct page *rmqueue(struct zone *preferred_zone,
3256 struct zone *zone, unsigned int order,
3257 gfp_t gfp_flags, unsigned int alloc_flags,
3260 unsigned long flags;
3263 if (likely(order == 0)) {
3264 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3265 migratetype, alloc_flags);
3270 * We most definitely don't want callers attempting to
3271 * allocate greater than order-1 page units with __GFP_NOFAIL.
3273 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3274 spin_lock_irqsave(&zone->lock, flags);
3278 if (alloc_flags & ALLOC_HARDER) {
3279 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3281 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3284 page = __rmqueue(zone, order, migratetype, alloc_flags);
3285 } while (page && check_new_pages(page, order));
3286 spin_unlock(&zone->lock);
3289 __mod_zone_freepage_state(zone, -(1 << order),
3290 get_pcppage_migratetype(page));
3292 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3293 zone_statistics(preferred_zone, zone);
3294 local_irq_restore(flags);
3297 /* Separate test+clear to avoid unnecessary atomics */
3298 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3299 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3300 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3303 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3307 local_irq_restore(flags);
3311 #ifdef CONFIG_FAIL_PAGE_ALLOC
3314 struct fault_attr attr;
3316 bool ignore_gfp_highmem;
3317 bool ignore_gfp_reclaim;
3319 } fail_page_alloc = {
3320 .attr = FAULT_ATTR_INITIALIZER,
3321 .ignore_gfp_reclaim = true,
3322 .ignore_gfp_highmem = true,
3326 static int __init setup_fail_page_alloc(char *str)
3328 return setup_fault_attr(&fail_page_alloc.attr, str);
3330 __setup("fail_page_alloc=", setup_fail_page_alloc);
3332 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3334 if (order < fail_page_alloc.min_order)
3336 if (gfp_mask & __GFP_NOFAIL)
3338 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3340 if (fail_page_alloc.ignore_gfp_reclaim &&
3341 (gfp_mask & __GFP_DIRECT_RECLAIM))
3344 return should_fail(&fail_page_alloc.attr, 1 << order);
3347 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3349 static int __init fail_page_alloc_debugfs(void)
3351 umode_t mode = S_IFREG | 0600;
3354 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3355 &fail_page_alloc.attr);
3357 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3358 &fail_page_alloc.ignore_gfp_reclaim);
3359 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3360 &fail_page_alloc.ignore_gfp_highmem);
3361 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3366 late_initcall(fail_page_alloc_debugfs);
3368 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3370 #else /* CONFIG_FAIL_PAGE_ALLOC */
3372 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3377 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3379 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3381 return __should_fail_alloc_page(gfp_mask, order);
3383 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3386 * Return true if free base pages are above 'mark'. For high-order checks it
3387 * will return true of the order-0 watermark is reached and there is at least
3388 * one free page of a suitable size. Checking now avoids taking the zone lock
3389 * to check in the allocation paths if no pages are free.
3391 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3392 int classzone_idx, unsigned int alloc_flags,
3397 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3399 /* free_pages may go negative - that's OK */
3400 free_pages -= (1 << order) - 1;
3402 if (alloc_flags & ALLOC_HIGH)
3406 * If the caller does not have rights to ALLOC_HARDER then subtract
3407 * the high-atomic reserves. This will over-estimate the size of the
3408 * atomic reserve but it avoids a search.
3410 if (likely(!alloc_harder)) {
3411 free_pages -= z->nr_reserved_highatomic;
3414 * OOM victims can try even harder than normal ALLOC_HARDER
3415 * users on the grounds that it's definitely going to be in
3416 * the exit path shortly and free memory. Any allocation it
3417 * makes during the free path will be small and short-lived.
3419 if (alloc_flags & ALLOC_OOM)
3427 /* If allocation can't use CMA areas don't use free CMA pages */
3428 if (!(alloc_flags & ALLOC_CMA))
3429 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3433 * Check watermarks for an order-0 allocation request. If these
3434 * are not met, then a high-order request also cannot go ahead
3435 * even if a suitable page happened to be free.
3437 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3440 /* If this is an order-0 request then the watermark is fine */
3444 /* For a high-order request, check at least one suitable page is free */
3445 for (o = order; o < MAX_ORDER; o++) {
3446 struct free_area *area = &z->free_area[o];
3452 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3453 if (!free_area_empty(area, mt))
3458 if ((alloc_flags & ALLOC_CMA) &&
3459 !free_area_empty(area, MIGRATE_CMA)) {
3464 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3470 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3471 int classzone_idx, unsigned int alloc_flags)
3473 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3474 zone_page_state(z, NR_FREE_PAGES));
3477 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3478 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3480 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3484 /* If allocation can't use CMA areas don't use free CMA pages */
3485 if (!(alloc_flags & ALLOC_CMA))
3486 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3490 * Fast check for order-0 only. If this fails then the reserves
3491 * need to be calculated. There is a corner case where the check
3492 * passes but only the high-order atomic reserve are free. If
3493 * the caller is !atomic then it'll uselessly search the free
3494 * list. That corner case is then slower but it is harmless.
3496 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3499 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3503 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3504 unsigned long mark, int classzone_idx)
3506 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3508 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3509 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3511 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3516 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3518 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3519 node_reclaim_distance;
3521 #else /* CONFIG_NUMA */
3522 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3526 #endif /* CONFIG_NUMA */
3529 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3530 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3531 * premature use of a lower zone may cause lowmem pressure problems that
3532 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3533 * probably too small. It only makes sense to spread allocations to avoid
3534 * fragmentation between the Normal and DMA32 zones.
3536 static inline unsigned int
3537 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3539 unsigned int alloc_flags = 0;
3541 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3542 alloc_flags |= ALLOC_KSWAPD;
3544 #ifdef CONFIG_ZONE_DMA32
3548 if (zone_idx(zone) != ZONE_NORMAL)
3552 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3553 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3554 * on UMA that if Normal is populated then so is DMA32.
3556 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3557 if (nr_online_nodes > 1 && !populated_zone(--zone))
3560 alloc_flags |= ALLOC_NOFRAGMENT;
3561 #endif /* CONFIG_ZONE_DMA32 */
3566 * get_page_from_freelist goes through the zonelist trying to allocate
3569 static struct page *
3570 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3571 const struct alloc_context *ac)
3575 struct pglist_data *last_pgdat_dirty_limit = NULL;
3580 * Scan zonelist, looking for a zone with enough free.
3581 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3583 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3584 z = ac->preferred_zoneref;
3585 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3590 if (cpusets_enabled() &&
3591 (alloc_flags & ALLOC_CPUSET) &&
3592 !__cpuset_zone_allowed(zone, gfp_mask))
3595 * When allocating a page cache page for writing, we
3596 * want to get it from a node that is within its dirty
3597 * limit, such that no single node holds more than its
3598 * proportional share of globally allowed dirty pages.
3599 * The dirty limits take into account the node's
3600 * lowmem reserves and high watermark so that kswapd
3601 * should be able to balance it without having to
3602 * write pages from its LRU list.
3604 * XXX: For now, allow allocations to potentially
3605 * exceed the per-node dirty limit in the slowpath
3606 * (spread_dirty_pages unset) before going into reclaim,
3607 * which is important when on a NUMA setup the allowed
3608 * nodes are together not big enough to reach the
3609 * global limit. The proper fix for these situations
3610 * will require awareness of nodes in the
3611 * dirty-throttling and the flusher threads.
3613 if (ac->spread_dirty_pages) {
3614 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3617 if (!node_dirty_ok(zone->zone_pgdat)) {
3618 last_pgdat_dirty_limit = zone->zone_pgdat;
3623 if (no_fallback && nr_online_nodes > 1 &&
3624 zone != ac->preferred_zoneref->zone) {
3628 * If moving to a remote node, retry but allow
3629 * fragmenting fallbacks. Locality is more important
3630 * than fragmentation avoidance.
3632 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3633 if (zone_to_nid(zone) != local_nid) {
3634 alloc_flags &= ~ALLOC_NOFRAGMENT;
3639 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3640 if (!zone_watermark_fast(zone, order, mark,
3641 ac_classzone_idx(ac), alloc_flags)) {
3644 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3646 * Watermark failed for this zone, but see if we can
3647 * grow this zone if it contains deferred pages.
3649 if (static_branch_unlikely(&deferred_pages)) {
3650 if (_deferred_grow_zone(zone, order))
3654 /* Checked here to keep the fast path fast */
3655 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3656 if (alloc_flags & ALLOC_NO_WATERMARKS)
3659 if (node_reclaim_mode == 0 ||
3660 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3663 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3665 case NODE_RECLAIM_NOSCAN:
3668 case NODE_RECLAIM_FULL:
3669 /* scanned but unreclaimable */
3672 /* did we reclaim enough */
3673 if (zone_watermark_ok(zone, order, mark,
3674 ac_classzone_idx(ac), alloc_flags))
3682 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3683 gfp_mask, alloc_flags, ac->migratetype);
3685 prep_new_page(page, order, gfp_mask, alloc_flags);
3688 * If this is a high-order atomic allocation then check
3689 * if the pageblock should be reserved for the future
3691 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3692 reserve_highatomic_pageblock(page, zone, order);
3696 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3697 /* Try again if zone has deferred pages */
3698 if (static_branch_unlikely(&deferred_pages)) {
3699 if (_deferred_grow_zone(zone, order))
3707 * It's possible on a UMA machine to get through all zones that are
3708 * fragmented. If avoiding fragmentation, reset and try again.
3711 alloc_flags &= ~ALLOC_NOFRAGMENT;
3718 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3720 unsigned int filter = SHOW_MEM_FILTER_NODES;
3723 * This documents exceptions given to allocations in certain
3724 * contexts that are allowed to allocate outside current's set
3727 if (!(gfp_mask & __GFP_NOMEMALLOC))
3728 if (tsk_is_oom_victim(current) ||
3729 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3730 filter &= ~SHOW_MEM_FILTER_NODES;
3731 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3732 filter &= ~SHOW_MEM_FILTER_NODES;
3734 show_mem(filter, nodemask);
3737 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3739 struct va_format vaf;
3741 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3743 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3746 va_start(args, fmt);
3749 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3750 current->comm, &vaf, gfp_mask, &gfp_mask,
3751 nodemask_pr_args(nodemask));
3754 cpuset_print_current_mems_allowed();
3757 warn_alloc_show_mem(gfp_mask, nodemask);
3760 static inline struct page *
3761 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3762 unsigned int alloc_flags,
3763 const struct alloc_context *ac)
3767 page = get_page_from_freelist(gfp_mask, order,
3768 alloc_flags|ALLOC_CPUSET, ac);
3770 * fallback to ignore cpuset restriction if our nodes
3774 page = get_page_from_freelist(gfp_mask, order,
3780 static inline struct page *
3781 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3782 const struct alloc_context *ac, unsigned long *did_some_progress)
3784 struct oom_control oc = {
3785 .zonelist = ac->zonelist,
3786 .nodemask = ac->nodemask,
3788 .gfp_mask = gfp_mask,
3793 *did_some_progress = 0;
3796 * Acquire the oom lock. If that fails, somebody else is
3797 * making progress for us.
3799 if (!mutex_trylock(&oom_lock)) {
3800 *did_some_progress = 1;
3801 schedule_timeout_uninterruptible(1);
3806 * Go through the zonelist yet one more time, keep very high watermark
3807 * here, this is only to catch a parallel oom killing, we must fail if
3808 * we're still under heavy pressure. But make sure that this reclaim
3809 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3810 * allocation which will never fail due to oom_lock already held.
3812 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3813 ~__GFP_DIRECT_RECLAIM, order,
3814 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3818 /* Coredumps can quickly deplete all memory reserves */
3819 if (current->flags & PF_DUMPCORE)
3821 /* The OOM killer will not help higher order allocs */
3822 if (order > PAGE_ALLOC_COSTLY_ORDER)
3825 * We have already exhausted all our reclaim opportunities without any
3826 * success so it is time to admit defeat. We will skip the OOM killer
3827 * because it is very likely that the caller has a more reasonable
3828 * fallback than shooting a random task.
3830 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3832 /* The OOM killer does not needlessly kill tasks for lowmem */
3833 if (ac->high_zoneidx < ZONE_NORMAL)
3835 if (pm_suspended_storage())
3838 * XXX: GFP_NOFS allocations should rather fail than rely on
3839 * other request to make a forward progress.
3840 * We are in an unfortunate situation where out_of_memory cannot
3841 * do much for this context but let's try it to at least get
3842 * access to memory reserved if the current task is killed (see
3843 * out_of_memory). Once filesystems are ready to handle allocation
3844 * failures more gracefully we should just bail out here.
3847 /* The OOM killer may not free memory on a specific node */
3848 if (gfp_mask & __GFP_THISNODE)
3851 /* Exhausted what can be done so it's blame time */
3852 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3853 *did_some_progress = 1;
3856 * Help non-failing allocations by giving them access to memory
3859 if (gfp_mask & __GFP_NOFAIL)
3860 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3861 ALLOC_NO_WATERMARKS, ac);
3864 mutex_unlock(&oom_lock);
3869 * Maximum number of compaction retries wit a progress before OOM
3870 * killer is consider as the only way to move forward.
3872 #define MAX_COMPACT_RETRIES 16
3874 #ifdef CONFIG_COMPACTION
3875 /* Try memory compaction for high-order allocations before reclaim */
3876 static struct page *
3877 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3878 unsigned int alloc_flags, const struct alloc_context *ac,
3879 enum compact_priority prio, enum compact_result *compact_result)
3881 struct page *page = NULL;
3882 unsigned long pflags;
3883 unsigned int noreclaim_flag;
3888 psi_memstall_enter(&pflags);
3889 noreclaim_flag = memalloc_noreclaim_save();
3891 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3894 memalloc_noreclaim_restore(noreclaim_flag);
3895 psi_memstall_leave(&pflags);
3898 * At least in one zone compaction wasn't deferred or skipped, so let's
3899 * count a compaction stall
3901 count_vm_event(COMPACTSTALL);
3903 /* Prep a captured page if available */
3905 prep_new_page(page, order, gfp_mask, alloc_flags);
3907 /* Try get a page from the freelist if available */
3909 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3912 struct zone *zone = page_zone(page);
3914 zone->compact_blockskip_flush = false;
3915 compaction_defer_reset(zone, order, true);
3916 count_vm_event(COMPACTSUCCESS);
3921 * It's bad if compaction run occurs and fails. The most likely reason
3922 * is that pages exist, but not enough to satisfy watermarks.
3924 count_vm_event(COMPACTFAIL);
3932 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3933 enum compact_result compact_result,
3934 enum compact_priority *compact_priority,
3935 int *compaction_retries)
3937 int max_retries = MAX_COMPACT_RETRIES;
3940 int retries = *compaction_retries;
3941 enum compact_priority priority = *compact_priority;
3946 if (compaction_made_progress(compact_result))
3947 (*compaction_retries)++;
3950 * compaction considers all the zone as desperately out of memory
3951 * so it doesn't really make much sense to retry except when the
3952 * failure could be caused by insufficient priority
3954 if (compaction_failed(compact_result))
3955 goto check_priority;
3958 * compaction was skipped because there are not enough order-0 pages
3959 * to work with, so we retry only if it looks like reclaim can help.
3961 if (compaction_needs_reclaim(compact_result)) {
3962 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3967 * make sure the compaction wasn't deferred or didn't bail out early
3968 * due to locks contention before we declare that we should give up.
3969 * But the next retry should use a higher priority if allowed, so
3970 * we don't just keep bailing out endlessly.
3972 if (compaction_withdrawn(compact_result)) {
3973 goto check_priority;
3977 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3978 * costly ones because they are de facto nofail and invoke OOM
3979 * killer to move on while costly can fail and users are ready
3980 * to cope with that. 1/4 retries is rather arbitrary but we
3981 * would need much more detailed feedback from compaction to
3982 * make a better decision.
3984 if (order > PAGE_ALLOC_COSTLY_ORDER)
3986 if (*compaction_retries <= max_retries) {
3992 * Make sure there are attempts at the highest priority if we exhausted
3993 * all retries or failed at the lower priorities.
3996 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3997 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3999 if (*compact_priority > min_priority) {
4000 (*compact_priority)--;
4001 *compaction_retries = 0;
4005 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4009 static inline struct page *
4010 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4011 unsigned int alloc_flags, const struct alloc_context *ac,
4012 enum compact_priority prio, enum compact_result *compact_result)
4014 *compact_result = COMPACT_SKIPPED;
4019 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4020 enum compact_result compact_result,
4021 enum compact_priority *compact_priority,
4022 int *compaction_retries)
4027 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4031 * There are setups with compaction disabled which would prefer to loop
4032 * inside the allocator rather than hit the oom killer prematurely.
4033 * Let's give them a good hope and keep retrying while the order-0
4034 * watermarks are OK.
4036 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4038 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4039 ac_classzone_idx(ac), alloc_flags))
4044 #endif /* CONFIG_COMPACTION */
4046 #ifdef CONFIG_LOCKDEP
4047 static struct lockdep_map __fs_reclaim_map =
4048 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4050 static bool __need_fs_reclaim(gfp_t gfp_mask)
4052 gfp_mask = current_gfp_context(gfp_mask);
4054 /* no reclaim without waiting on it */
4055 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4058 /* this guy won't enter reclaim */
4059 if (current->flags & PF_MEMALLOC)
4062 /* We're only interested __GFP_FS allocations for now */
4063 if (!(gfp_mask & __GFP_FS))
4066 if (gfp_mask & __GFP_NOLOCKDEP)
4072 void __fs_reclaim_acquire(void)
4074 lock_map_acquire(&__fs_reclaim_map);
4077 void __fs_reclaim_release(void)
4079 lock_map_release(&__fs_reclaim_map);
4082 void fs_reclaim_acquire(gfp_t gfp_mask)
4084 if (__need_fs_reclaim(gfp_mask))
4085 __fs_reclaim_acquire();
4087 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4089 void fs_reclaim_release(gfp_t gfp_mask)
4091 if (__need_fs_reclaim(gfp_mask))
4092 __fs_reclaim_release();
4094 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4097 /* Perform direct synchronous page reclaim */
4099 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4100 const struct alloc_context *ac)
4103 unsigned int noreclaim_flag;
4104 unsigned long pflags;
4108 /* We now go into synchronous reclaim */
4109 cpuset_memory_pressure_bump();
4110 psi_memstall_enter(&pflags);
4111 fs_reclaim_acquire(gfp_mask);
4112 noreclaim_flag = memalloc_noreclaim_save();
4114 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4117 memalloc_noreclaim_restore(noreclaim_flag);
4118 fs_reclaim_release(gfp_mask);
4119 psi_memstall_leave(&pflags);
4126 /* The really slow allocator path where we enter direct reclaim */
4127 static inline struct page *
4128 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4129 unsigned int alloc_flags, const struct alloc_context *ac,
4130 unsigned long *did_some_progress)
4132 struct page *page = NULL;
4133 bool drained = false;
4135 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4136 if (unlikely(!(*did_some_progress)))
4140 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4143 * If an allocation failed after direct reclaim, it could be because
4144 * pages are pinned on the per-cpu lists or in high alloc reserves.
4145 * Shrink them them and try again
4147 if (!page && !drained) {
4148 unreserve_highatomic_pageblock(ac, false);
4149 drain_all_pages(NULL);
4157 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4158 const struct alloc_context *ac)
4162 pg_data_t *last_pgdat = NULL;
4163 enum zone_type high_zoneidx = ac->high_zoneidx;
4165 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4167 if (last_pgdat != zone->zone_pgdat)
4168 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4169 last_pgdat = zone->zone_pgdat;
4173 static inline unsigned int
4174 gfp_to_alloc_flags(gfp_t gfp_mask)
4176 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4178 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4179 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4182 * The caller may dip into page reserves a bit more if the caller
4183 * cannot run direct reclaim, or if the caller has realtime scheduling
4184 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4185 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4187 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4189 if (gfp_mask & __GFP_ATOMIC) {
4191 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4192 * if it can't schedule.
4194 if (!(gfp_mask & __GFP_NOMEMALLOC))
4195 alloc_flags |= ALLOC_HARDER;
4197 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4198 * comment for __cpuset_node_allowed().
4200 alloc_flags &= ~ALLOC_CPUSET;
4201 } else if (unlikely(rt_task(current)) && !in_interrupt())
4202 alloc_flags |= ALLOC_HARDER;
4204 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4205 alloc_flags |= ALLOC_KSWAPD;
4208 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4209 alloc_flags |= ALLOC_CMA;
4214 static bool oom_reserves_allowed(struct task_struct *tsk)
4216 if (!tsk_is_oom_victim(tsk))
4220 * !MMU doesn't have oom reaper so give access to memory reserves
4221 * only to the thread with TIF_MEMDIE set
4223 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4230 * Distinguish requests which really need access to full memory
4231 * reserves from oom victims which can live with a portion of it
4233 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4235 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4237 if (gfp_mask & __GFP_MEMALLOC)
4238 return ALLOC_NO_WATERMARKS;
4239 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4240 return ALLOC_NO_WATERMARKS;
4241 if (!in_interrupt()) {
4242 if (current->flags & PF_MEMALLOC)
4243 return ALLOC_NO_WATERMARKS;
4244 else if (oom_reserves_allowed(current))
4251 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4253 return !!__gfp_pfmemalloc_flags(gfp_mask);
4257 * Checks whether it makes sense to retry the reclaim to make a forward progress
4258 * for the given allocation request.
4260 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4261 * without success, or when we couldn't even meet the watermark if we
4262 * reclaimed all remaining pages on the LRU lists.
4264 * Returns true if a retry is viable or false to enter the oom path.
4267 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4268 struct alloc_context *ac, int alloc_flags,
4269 bool did_some_progress, int *no_progress_loops)
4276 * Costly allocations might have made a progress but this doesn't mean
4277 * their order will become available due to high fragmentation so
4278 * always increment the no progress counter for them
4280 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4281 *no_progress_loops = 0;
4283 (*no_progress_loops)++;
4286 * Make sure we converge to OOM if we cannot make any progress
4287 * several times in the row.
4289 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4290 /* Before OOM, exhaust highatomic_reserve */
4291 return unreserve_highatomic_pageblock(ac, true);
4295 * Keep reclaiming pages while there is a chance this will lead
4296 * somewhere. If none of the target zones can satisfy our allocation
4297 * request even if all reclaimable pages are considered then we are
4298 * screwed and have to go OOM.
4300 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4302 unsigned long available;
4303 unsigned long reclaimable;
4304 unsigned long min_wmark = min_wmark_pages(zone);
4307 available = reclaimable = zone_reclaimable_pages(zone);
4308 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4311 * Would the allocation succeed if we reclaimed all
4312 * reclaimable pages?
4314 wmark = __zone_watermark_ok(zone, order, min_wmark,
4315 ac_classzone_idx(ac), alloc_flags, available);
4316 trace_reclaim_retry_zone(z, order, reclaimable,
4317 available, min_wmark, *no_progress_loops, wmark);
4320 * If we didn't make any progress and have a lot of
4321 * dirty + writeback pages then we should wait for
4322 * an IO to complete to slow down the reclaim and
4323 * prevent from pre mature OOM
4325 if (!did_some_progress) {
4326 unsigned long write_pending;
4328 write_pending = zone_page_state_snapshot(zone,
4329 NR_ZONE_WRITE_PENDING);
4331 if (2 * write_pending > reclaimable) {
4332 congestion_wait(BLK_RW_ASYNC, HZ/10);
4344 * Memory allocation/reclaim might be called from a WQ context and the
4345 * current implementation of the WQ concurrency control doesn't
4346 * recognize that a particular WQ is congested if the worker thread is
4347 * looping without ever sleeping. Therefore we have to do a short sleep
4348 * here rather than calling cond_resched().
4350 if (current->flags & PF_WQ_WORKER)
4351 schedule_timeout_uninterruptible(1);
4358 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4361 * It's possible that cpuset's mems_allowed and the nodemask from
4362 * mempolicy don't intersect. This should be normally dealt with by
4363 * policy_nodemask(), but it's possible to race with cpuset update in
4364 * such a way the check therein was true, and then it became false
4365 * before we got our cpuset_mems_cookie here.
4366 * This assumes that for all allocations, ac->nodemask can come only
4367 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4368 * when it does not intersect with the cpuset restrictions) or the
4369 * caller can deal with a violated nodemask.
4371 if (cpusets_enabled() && ac->nodemask &&
4372 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4373 ac->nodemask = NULL;
4378 * When updating a task's mems_allowed or mempolicy nodemask, it is
4379 * possible to race with parallel threads in such a way that our
4380 * allocation can fail while the mask is being updated. If we are about
4381 * to fail, check if the cpuset changed during allocation and if so,
4384 if (read_mems_allowed_retry(cpuset_mems_cookie))
4390 static inline struct page *
4391 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4392 struct alloc_context *ac)
4394 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4395 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4396 struct page *page = NULL;
4397 unsigned int alloc_flags;
4398 unsigned long did_some_progress;
4399 enum compact_priority compact_priority;
4400 enum compact_result compact_result;
4401 int compaction_retries;
4402 int no_progress_loops;
4403 unsigned int cpuset_mems_cookie;
4407 * We also sanity check to catch abuse of atomic reserves being used by
4408 * callers that are not in atomic context.
4410 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4411 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4412 gfp_mask &= ~__GFP_ATOMIC;
4415 compaction_retries = 0;
4416 no_progress_loops = 0;
4417 compact_priority = DEF_COMPACT_PRIORITY;
4418 cpuset_mems_cookie = read_mems_allowed_begin();
4421 * The fast path uses conservative alloc_flags to succeed only until
4422 * kswapd needs to be woken up, and to avoid the cost of setting up
4423 * alloc_flags precisely. So we do that now.
4425 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4428 * We need to recalculate the starting point for the zonelist iterator
4429 * because we might have used different nodemask in the fast path, or
4430 * there was a cpuset modification and we are retrying - otherwise we
4431 * could end up iterating over non-eligible zones endlessly.
4433 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4434 ac->high_zoneidx, ac->nodemask);
4435 if (!ac->preferred_zoneref->zone)
4438 if (alloc_flags & ALLOC_KSWAPD)
4439 wake_all_kswapds(order, gfp_mask, ac);
4442 * The adjusted alloc_flags might result in immediate success, so try
4445 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4450 * For costly allocations, try direct compaction first, as it's likely
4451 * that we have enough base pages and don't need to reclaim. For non-
4452 * movable high-order allocations, do that as well, as compaction will
4453 * try prevent permanent fragmentation by migrating from blocks of the
4455 * Don't try this for allocations that are allowed to ignore
4456 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4458 if (can_direct_reclaim &&
4460 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4461 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4462 page = __alloc_pages_direct_compact(gfp_mask, order,
4464 INIT_COMPACT_PRIORITY,
4470 * Checks for costly allocations with __GFP_NORETRY, which
4471 * includes some THP page fault allocations
4473 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4475 * If allocating entire pageblock(s) and compaction
4476 * failed because all zones are below low watermarks
4477 * or is prohibited because it recently failed at this
4478 * order, fail immediately unless the allocator has
4479 * requested compaction and reclaim retry.
4482 * - potentially very expensive because zones are far
4483 * below their low watermarks or this is part of very
4484 * bursty high order allocations,
4485 * - not guaranteed to help because isolate_freepages()
4486 * may not iterate over freed pages as part of its
4488 * - unlikely to make entire pageblocks free on its
4491 if (compact_result == COMPACT_SKIPPED ||
4492 compact_result == COMPACT_DEFERRED)
4496 * Looks like reclaim/compaction is worth trying, but
4497 * sync compaction could be very expensive, so keep
4498 * using async compaction.
4500 compact_priority = INIT_COMPACT_PRIORITY;
4505 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4506 if (alloc_flags & ALLOC_KSWAPD)
4507 wake_all_kswapds(order, gfp_mask, ac);
4509 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4511 alloc_flags = reserve_flags;
4514 * Reset the nodemask and zonelist iterators if memory policies can be
4515 * ignored. These allocations are high priority and system rather than
4518 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4519 ac->nodemask = NULL;
4520 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4521 ac->high_zoneidx, ac->nodemask);
4524 /* Attempt with potentially adjusted zonelist and alloc_flags */
4525 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4529 /* Caller is not willing to reclaim, we can't balance anything */
4530 if (!can_direct_reclaim)
4533 /* Avoid recursion of direct reclaim */
4534 if (current->flags & PF_MEMALLOC)
4537 /* Try direct reclaim and then allocating */
4538 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4539 &did_some_progress);
4543 /* Try direct compaction and then allocating */
4544 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4545 compact_priority, &compact_result);
4549 /* Do not loop if specifically requested */
4550 if (gfp_mask & __GFP_NORETRY)
4554 * Do not retry costly high order allocations unless they are
4555 * __GFP_RETRY_MAYFAIL
4557 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4560 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4561 did_some_progress > 0, &no_progress_loops))
4565 * It doesn't make any sense to retry for the compaction if the order-0
4566 * reclaim is not able to make any progress because the current
4567 * implementation of the compaction depends on the sufficient amount
4568 * of free memory (see __compaction_suitable)
4570 if (did_some_progress > 0 &&
4571 should_compact_retry(ac, order, alloc_flags,
4572 compact_result, &compact_priority,
4573 &compaction_retries))
4577 /* Deal with possible cpuset update races before we start OOM killing */
4578 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4581 /* Reclaim has failed us, start killing things */
4582 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4586 /* Avoid allocations with no watermarks from looping endlessly */
4587 if (tsk_is_oom_victim(current) &&
4588 (alloc_flags == ALLOC_OOM ||
4589 (gfp_mask & __GFP_NOMEMALLOC)))
4592 /* Retry as long as the OOM killer is making progress */
4593 if (did_some_progress) {
4594 no_progress_loops = 0;
4599 /* Deal with possible cpuset update races before we fail */
4600 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4604 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4607 if (gfp_mask & __GFP_NOFAIL) {
4609 * All existing users of the __GFP_NOFAIL are blockable, so warn
4610 * of any new users that actually require GFP_NOWAIT
4612 if (WARN_ON_ONCE(!can_direct_reclaim))
4616 * PF_MEMALLOC request from this context is rather bizarre
4617 * because we cannot reclaim anything and only can loop waiting
4618 * for somebody to do a work for us
4620 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4623 * non failing costly orders are a hard requirement which we
4624 * are not prepared for much so let's warn about these users
4625 * so that we can identify them and convert them to something
4628 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4631 * Help non-failing allocations by giving them access to memory
4632 * reserves but do not use ALLOC_NO_WATERMARKS because this
4633 * could deplete whole memory reserves which would just make
4634 * the situation worse
4636 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4644 warn_alloc(gfp_mask, ac->nodemask,
4645 "page allocation failure: order:%u", order);
4650 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4651 int preferred_nid, nodemask_t *nodemask,
4652 struct alloc_context *ac, gfp_t *alloc_mask,
4653 unsigned int *alloc_flags)
4655 ac->high_zoneidx = gfp_zone(gfp_mask);
4656 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4657 ac->nodemask = nodemask;
4658 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4660 if (cpusets_enabled()) {
4661 *alloc_mask |= __GFP_HARDWALL;
4663 ac->nodemask = &cpuset_current_mems_allowed;
4665 *alloc_flags |= ALLOC_CPUSET;
4668 fs_reclaim_acquire(gfp_mask);
4669 fs_reclaim_release(gfp_mask);
4671 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4673 if (should_fail_alloc_page(gfp_mask, order))
4676 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4677 *alloc_flags |= ALLOC_CMA;
4682 /* Determine whether to spread dirty pages and what the first usable zone */
4683 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4685 /* Dirty zone balancing only done in the fast path */
4686 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4689 * The preferred zone is used for statistics but crucially it is
4690 * also used as the starting point for the zonelist iterator. It
4691 * may get reset for allocations that ignore memory policies.
4693 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4694 ac->high_zoneidx, ac->nodemask);
4698 * This is the 'heart' of the zoned buddy allocator.
4701 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4702 nodemask_t *nodemask)
4705 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4706 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4707 struct alloc_context ac = { };
4710 * There are several places where we assume that the order value is sane
4711 * so bail out early if the request is out of bound.
4713 if (unlikely(order >= MAX_ORDER)) {
4714 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4718 gfp_mask &= gfp_allowed_mask;
4719 alloc_mask = gfp_mask;
4720 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4723 finalise_ac(gfp_mask, &ac);
4726 * Forbid the first pass from falling back to types that fragment
4727 * memory until all local zones are considered.
4729 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4731 /* First allocation attempt */
4732 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4737 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4738 * resp. GFP_NOIO which has to be inherited for all allocation requests
4739 * from a particular context which has been marked by
4740 * memalloc_no{fs,io}_{save,restore}.
4742 alloc_mask = current_gfp_context(gfp_mask);
4743 ac.spread_dirty_pages = false;
4746 * Restore the original nodemask if it was potentially replaced with
4747 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4749 if (unlikely(ac.nodemask != nodemask))
4750 ac.nodemask = nodemask;
4752 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4755 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4756 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4757 __free_pages(page, order);
4761 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4765 EXPORT_SYMBOL(__alloc_pages_nodemask);
4768 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4769 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4770 * you need to access high mem.
4772 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4776 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4779 return (unsigned long) page_address(page);
4781 EXPORT_SYMBOL(__get_free_pages);
4783 unsigned long get_zeroed_page(gfp_t gfp_mask)
4785 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4787 EXPORT_SYMBOL(get_zeroed_page);
4789 static inline void free_the_page(struct page *page, unsigned int order)
4791 if (order == 0) /* Via pcp? */
4792 free_unref_page(page);
4794 __free_pages_ok(page, order);
4797 void __free_pages(struct page *page, unsigned int order)
4799 if (put_page_testzero(page))
4800 free_the_page(page, order);
4802 EXPORT_SYMBOL(__free_pages);
4804 void free_pages(unsigned long addr, unsigned int order)
4807 VM_BUG_ON(!virt_addr_valid((void *)addr));
4808 __free_pages(virt_to_page((void *)addr), order);
4812 EXPORT_SYMBOL(free_pages);
4816 * An arbitrary-length arbitrary-offset area of memory which resides
4817 * within a 0 or higher order page. Multiple fragments within that page
4818 * are individually refcounted, in the page's reference counter.
4820 * The page_frag functions below provide a simple allocation framework for
4821 * page fragments. This is used by the network stack and network device
4822 * drivers to provide a backing region of memory for use as either an
4823 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4825 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4828 struct page *page = NULL;
4829 gfp_t gfp = gfp_mask;
4831 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4832 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4834 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4835 PAGE_FRAG_CACHE_MAX_ORDER);
4836 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4838 if (unlikely(!page))
4839 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4841 nc->va = page ? page_address(page) : NULL;
4846 void __page_frag_cache_drain(struct page *page, unsigned int count)
4848 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4850 if (page_ref_sub_and_test(page, count))
4851 free_the_page(page, compound_order(page));
4853 EXPORT_SYMBOL(__page_frag_cache_drain);
4855 void *page_frag_alloc(struct page_frag_cache *nc,
4856 unsigned int fragsz, gfp_t gfp_mask)
4858 unsigned int size = PAGE_SIZE;
4862 if (unlikely(!nc->va)) {
4864 page = __page_frag_cache_refill(nc, gfp_mask);
4868 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4869 /* if size can vary use size else just use PAGE_SIZE */
4872 /* Even if we own the page, we do not use atomic_set().
4873 * This would break get_page_unless_zero() users.
4875 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4877 /* reset page count bias and offset to start of new frag */
4878 nc->pfmemalloc = page_is_pfmemalloc(page);
4879 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4883 offset = nc->offset - fragsz;
4884 if (unlikely(offset < 0)) {
4885 page = virt_to_page(nc->va);
4887 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4890 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4891 /* if size can vary use size else just use PAGE_SIZE */
4894 /* OK, page count is 0, we can safely set it */
4895 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4897 /* reset page count bias and offset to start of new frag */
4898 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4899 offset = size - fragsz;
4903 nc->offset = offset;
4905 return nc->va + offset;
4907 EXPORT_SYMBOL(page_frag_alloc);
4910 * Frees a page fragment allocated out of either a compound or order 0 page.
4912 void page_frag_free(void *addr)
4914 struct page *page = virt_to_head_page(addr);
4916 if (unlikely(put_page_testzero(page)))
4917 free_the_page(page, compound_order(page));
4919 EXPORT_SYMBOL(page_frag_free);
4921 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4925 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4926 unsigned long used = addr + PAGE_ALIGN(size);
4928 split_page(virt_to_page((void *)addr), order);
4929 while (used < alloc_end) {
4934 return (void *)addr;
4938 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4939 * @size: the number of bytes to allocate
4940 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4942 * This function is similar to alloc_pages(), except that it allocates the
4943 * minimum number of pages to satisfy the request. alloc_pages() can only
4944 * allocate memory in power-of-two pages.
4946 * This function is also limited by MAX_ORDER.
4948 * Memory allocated by this function must be released by free_pages_exact().
4950 * Return: pointer to the allocated area or %NULL in case of error.
4952 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4954 unsigned int order = get_order(size);
4957 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4958 gfp_mask &= ~__GFP_COMP;
4960 addr = __get_free_pages(gfp_mask, order);
4961 return make_alloc_exact(addr, order, size);
4963 EXPORT_SYMBOL(alloc_pages_exact);
4966 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4968 * @nid: the preferred node ID where memory should be allocated
4969 * @size: the number of bytes to allocate
4970 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4972 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4975 * Return: pointer to the allocated area or %NULL in case of error.
4977 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4979 unsigned int order = get_order(size);
4982 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4983 gfp_mask &= ~__GFP_COMP;
4985 p = alloc_pages_node(nid, gfp_mask, order);
4988 return make_alloc_exact((unsigned long)page_address(p), order, size);
4992 * free_pages_exact - release memory allocated via alloc_pages_exact()
4993 * @virt: the value returned by alloc_pages_exact.
4994 * @size: size of allocation, same value as passed to alloc_pages_exact().
4996 * Release the memory allocated by a previous call to alloc_pages_exact.
4998 void free_pages_exact(void *virt, size_t size)
5000 unsigned long addr = (unsigned long)virt;
5001 unsigned long end = addr + PAGE_ALIGN(size);
5003 while (addr < end) {
5008 EXPORT_SYMBOL(free_pages_exact);
5011 * nr_free_zone_pages - count number of pages beyond high watermark
5012 * @offset: The zone index of the highest zone
5014 * nr_free_zone_pages() counts the number of pages which are beyond the
5015 * high watermark within all zones at or below a given zone index. For each
5016 * zone, the number of pages is calculated as:
5018 * nr_free_zone_pages = managed_pages - high_pages
5020 * Return: number of pages beyond high watermark.
5022 static unsigned long nr_free_zone_pages(int offset)
5027 /* Just pick one node, since fallback list is circular */
5028 unsigned long sum = 0;
5030 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5032 for_each_zone_zonelist(zone, z, zonelist, offset) {
5033 unsigned long size = zone_managed_pages(zone);
5034 unsigned long high = high_wmark_pages(zone);
5043 * nr_free_buffer_pages - count number of pages beyond high watermark
5045 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5046 * watermark within ZONE_DMA and ZONE_NORMAL.
5048 * Return: number of pages beyond high watermark within ZONE_DMA and
5051 unsigned long nr_free_buffer_pages(void)
5053 return nr_free_zone_pages(gfp_zone(GFP_USER));
5055 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5058 * nr_free_pagecache_pages - count number of pages beyond high watermark
5060 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5061 * high watermark within all zones.
5063 * Return: number of pages beyond high watermark within all zones.
5065 unsigned long nr_free_pagecache_pages(void)
5067 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5070 static inline void show_node(struct zone *zone)
5072 if (IS_ENABLED(CONFIG_NUMA))
5073 printk("Node %d ", zone_to_nid(zone));
5076 long si_mem_available(void)
5079 unsigned long pagecache;
5080 unsigned long wmark_low = 0;
5081 unsigned long pages[NR_LRU_LISTS];
5082 unsigned long reclaimable;
5086 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5087 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5090 wmark_low += low_wmark_pages(zone);
5093 * Estimate the amount of memory available for userspace allocations,
5094 * without causing swapping.
5096 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5099 * Not all the page cache can be freed, otherwise the system will
5100 * start swapping. Assume at least half of the page cache, or the
5101 * low watermark worth of cache, needs to stay.
5103 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5104 pagecache -= min(pagecache / 2, wmark_low);
5105 available += pagecache;
5108 * Part of the reclaimable slab and other kernel memory consists of
5109 * items that are in use, and cannot be freed. Cap this estimate at the
5112 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5113 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5114 available += reclaimable - min(reclaimable / 2, wmark_low);
5120 EXPORT_SYMBOL_GPL(si_mem_available);
5122 void si_meminfo(struct sysinfo *val)
5124 val->totalram = totalram_pages();
5125 val->sharedram = global_node_page_state(NR_SHMEM);
5126 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5127 val->bufferram = nr_blockdev_pages();
5128 val->totalhigh = totalhigh_pages();
5129 val->freehigh = nr_free_highpages();
5130 val->mem_unit = PAGE_SIZE;
5133 EXPORT_SYMBOL(si_meminfo);
5136 void si_meminfo_node(struct sysinfo *val, int nid)
5138 int zone_type; /* needs to be signed */
5139 unsigned long managed_pages = 0;
5140 unsigned long managed_highpages = 0;
5141 unsigned long free_highpages = 0;
5142 pg_data_t *pgdat = NODE_DATA(nid);
5144 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5145 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5146 val->totalram = managed_pages;
5147 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5148 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5149 #ifdef CONFIG_HIGHMEM
5150 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5151 struct zone *zone = &pgdat->node_zones[zone_type];
5153 if (is_highmem(zone)) {
5154 managed_highpages += zone_managed_pages(zone);
5155 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5158 val->totalhigh = managed_highpages;
5159 val->freehigh = free_highpages;
5161 val->totalhigh = managed_highpages;
5162 val->freehigh = free_highpages;
5164 val->mem_unit = PAGE_SIZE;
5169 * Determine whether the node should be displayed or not, depending on whether
5170 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5172 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5174 if (!(flags & SHOW_MEM_FILTER_NODES))
5178 * no node mask - aka implicit memory numa policy. Do not bother with
5179 * the synchronization - read_mems_allowed_begin - because we do not
5180 * have to be precise here.
5183 nodemask = &cpuset_current_mems_allowed;
5185 return !node_isset(nid, *nodemask);
5188 #define K(x) ((x) << (PAGE_SHIFT-10))
5190 static void show_migration_types(unsigned char type)
5192 static const char types[MIGRATE_TYPES] = {
5193 [MIGRATE_UNMOVABLE] = 'U',
5194 [MIGRATE_MOVABLE] = 'M',
5195 [MIGRATE_RECLAIMABLE] = 'E',
5196 [MIGRATE_HIGHATOMIC] = 'H',
5198 [MIGRATE_CMA] = 'C',
5200 #ifdef CONFIG_MEMORY_ISOLATION
5201 [MIGRATE_ISOLATE] = 'I',
5204 char tmp[MIGRATE_TYPES + 1];
5208 for (i = 0; i < MIGRATE_TYPES; i++) {
5209 if (type & (1 << i))
5214 printk(KERN_CONT "(%s) ", tmp);
5218 * Show free area list (used inside shift_scroll-lock stuff)
5219 * We also calculate the percentage fragmentation. We do this by counting the
5220 * memory on each free list with the exception of the first item on the list.
5223 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5226 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5228 unsigned long free_pcp = 0;
5233 for_each_populated_zone(zone) {
5234 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5237 for_each_online_cpu(cpu)
5238 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5241 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5242 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5243 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5244 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5245 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5246 " free:%lu free_pcp:%lu free_cma:%lu\n",
5247 global_node_page_state(NR_ACTIVE_ANON),
5248 global_node_page_state(NR_INACTIVE_ANON),
5249 global_node_page_state(NR_ISOLATED_ANON),
5250 global_node_page_state(NR_ACTIVE_FILE),
5251 global_node_page_state(NR_INACTIVE_FILE),
5252 global_node_page_state(NR_ISOLATED_FILE),
5253 global_node_page_state(NR_UNEVICTABLE),
5254 global_node_page_state(NR_FILE_DIRTY),
5255 global_node_page_state(NR_WRITEBACK),
5256 global_node_page_state(NR_UNSTABLE_NFS),
5257 global_node_page_state(NR_SLAB_RECLAIMABLE),
5258 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5259 global_node_page_state(NR_FILE_MAPPED),
5260 global_node_page_state(NR_SHMEM),
5261 global_zone_page_state(NR_PAGETABLE),
5262 global_zone_page_state(NR_BOUNCE),
5263 global_zone_page_state(NR_FREE_PAGES),
5265 global_zone_page_state(NR_FREE_CMA_PAGES));
5267 for_each_online_pgdat(pgdat) {
5268 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5272 " active_anon:%lukB"
5273 " inactive_anon:%lukB"
5274 " active_file:%lukB"
5275 " inactive_file:%lukB"
5276 " unevictable:%lukB"
5277 " isolated(anon):%lukB"
5278 " isolated(file):%lukB"
5283 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5285 " shmem_pmdmapped: %lukB"
5288 " writeback_tmp:%lukB"
5290 " all_unreclaimable? %s"
5293 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5294 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5295 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5296 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5297 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5298 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5299 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5300 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5301 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5302 K(node_page_state(pgdat, NR_WRITEBACK)),
5303 K(node_page_state(pgdat, NR_SHMEM)),
5304 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5305 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5306 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5308 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5310 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5311 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5312 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5316 for_each_populated_zone(zone) {
5319 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5323 for_each_online_cpu(cpu)
5324 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5333 " reserved_highatomic:%luKB"
5334 " active_anon:%lukB"
5335 " inactive_anon:%lukB"
5336 " active_file:%lukB"
5337 " inactive_file:%lukB"
5338 " unevictable:%lukB"
5339 " writepending:%lukB"
5343 " kernel_stack:%lukB"
5351 K(zone_page_state(zone, NR_FREE_PAGES)),
5352 K(min_wmark_pages(zone)),
5353 K(low_wmark_pages(zone)),
5354 K(high_wmark_pages(zone)),
5355 K(zone->nr_reserved_highatomic),
5356 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5357 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5358 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5359 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5360 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5361 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5362 K(zone->present_pages),
5363 K(zone_managed_pages(zone)),
5364 K(zone_page_state(zone, NR_MLOCK)),
5365 zone_page_state(zone, NR_KERNEL_STACK_KB),
5366 K(zone_page_state(zone, NR_PAGETABLE)),
5367 K(zone_page_state(zone, NR_BOUNCE)),
5369 K(this_cpu_read(zone->pageset->pcp.count)),
5370 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5371 printk("lowmem_reserve[]:");
5372 for (i = 0; i < MAX_NR_ZONES; i++)
5373 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5374 printk(KERN_CONT "\n");
5377 for_each_populated_zone(zone) {
5379 unsigned long nr[MAX_ORDER], flags, total = 0;
5380 unsigned char types[MAX_ORDER];
5382 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5385 printk(KERN_CONT "%s: ", zone->name);
5387 spin_lock_irqsave(&zone->lock, flags);
5388 for (order = 0; order < MAX_ORDER; order++) {
5389 struct free_area *area = &zone->free_area[order];
5392 nr[order] = area->nr_free;
5393 total += nr[order] << order;
5396 for (type = 0; type < MIGRATE_TYPES; type++) {
5397 if (!free_area_empty(area, type))
5398 types[order] |= 1 << type;
5401 spin_unlock_irqrestore(&zone->lock, flags);
5402 for (order = 0; order < MAX_ORDER; order++) {
5403 printk(KERN_CONT "%lu*%lukB ",
5404 nr[order], K(1UL) << order);
5406 show_migration_types(types[order]);
5408 printk(KERN_CONT "= %lukB\n", K(total));
5411 hugetlb_show_meminfo();
5413 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5415 show_swap_cache_info();
5418 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5420 zoneref->zone = zone;
5421 zoneref->zone_idx = zone_idx(zone);
5425 * Builds allocation fallback zone lists.
5427 * Add all populated zones of a node to the zonelist.
5429 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5432 enum zone_type zone_type = MAX_NR_ZONES;
5437 zone = pgdat->node_zones + zone_type;
5438 if (managed_zone(zone)) {
5439 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5440 check_highest_zone(zone_type);
5442 } while (zone_type);
5449 static int __parse_numa_zonelist_order(char *s)
5452 * We used to support different zonlists modes but they turned
5453 * out to be just not useful. Let's keep the warning in place
5454 * if somebody still use the cmd line parameter so that we do
5455 * not fail it silently
5457 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5458 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5464 static __init int setup_numa_zonelist_order(char *s)
5469 return __parse_numa_zonelist_order(s);
5471 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5473 char numa_zonelist_order[] = "Node";
5476 * sysctl handler for numa_zonelist_order
5478 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5479 void __user *buffer, size_t *length,
5486 return proc_dostring(table, write, buffer, length, ppos);
5487 str = memdup_user_nul(buffer, 16);
5489 return PTR_ERR(str);
5491 ret = __parse_numa_zonelist_order(str);
5497 #define MAX_NODE_LOAD (nr_online_nodes)
5498 static int node_load[MAX_NUMNODES];
5501 * find_next_best_node - find the next node that should appear in a given node's fallback list
5502 * @node: node whose fallback list we're appending
5503 * @used_node_mask: nodemask_t of already used nodes
5505 * We use a number of factors to determine which is the next node that should
5506 * appear on a given node's fallback list. The node should not have appeared
5507 * already in @node's fallback list, and it should be the next closest node
5508 * according to the distance array (which contains arbitrary distance values
5509 * from each node to each node in the system), and should also prefer nodes
5510 * with no CPUs, since presumably they'll have very little allocation pressure
5511 * on them otherwise.
5513 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5515 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5518 int min_val = INT_MAX;
5519 int best_node = NUMA_NO_NODE;
5520 const struct cpumask *tmp = cpumask_of_node(0);
5522 /* Use the local node if we haven't already */
5523 if (!node_isset(node, *used_node_mask)) {
5524 node_set(node, *used_node_mask);
5528 for_each_node_state(n, N_MEMORY) {
5530 /* Don't want a node to appear more than once */
5531 if (node_isset(n, *used_node_mask))
5534 /* Use the distance array to find the distance */
5535 val = node_distance(node, n);
5537 /* Penalize nodes under us ("prefer the next node") */
5540 /* Give preference to headless and unused nodes */
5541 tmp = cpumask_of_node(n);
5542 if (!cpumask_empty(tmp))
5543 val += PENALTY_FOR_NODE_WITH_CPUS;
5545 /* Slight preference for less loaded node */
5546 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5547 val += node_load[n];
5549 if (val < min_val) {
5556 node_set(best_node, *used_node_mask);
5563 * Build zonelists ordered by node and zones within node.
5564 * This results in maximum locality--normal zone overflows into local
5565 * DMA zone, if any--but risks exhausting DMA zone.
5567 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5570 struct zoneref *zonerefs;
5573 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5575 for (i = 0; i < nr_nodes; i++) {
5578 pg_data_t *node = NODE_DATA(node_order[i]);
5580 nr_zones = build_zonerefs_node(node, zonerefs);
5581 zonerefs += nr_zones;
5583 zonerefs->zone = NULL;
5584 zonerefs->zone_idx = 0;
5588 * Build gfp_thisnode zonelists
5590 static void build_thisnode_zonelists(pg_data_t *pgdat)
5592 struct zoneref *zonerefs;
5595 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5596 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5597 zonerefs += nr_zones;
5598 zonerefs->zone = NULL;
5599 zonerefs->zone_idx = 0;
5603 * Build zonelists ordered by zone and nodes within zones.
5604 * This results in conserving DMA zone[s] until all Normal memory is
5605 * exhausted, but results in overflowing to remote node while memory
5606 * may still exist in local DMA zone.
5609 static void build_zonelists(pg_data_t *pgdat)
5611 static int node_order[MAX_NUMNODES];
5612 int node, load, nr_nodes = 0;
5613 nodemask_t used_mask;
5614 int local_node, prev_node;
5616 /* NUMA-aware ordering of nodes */
5617 local_node = pgdat->node_id;
5618 load = nr_online_nodes;
5619 prev_node = local_node;
5620 nodes_clear(used_mask);
5622 memset(node_order, 0, sizeof(node_order));
5623 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5625 * We don't want to pressure a particular node.
5626 * So adding penalty to the first node in same
5627 * distance group to make it round-robin.
5629 if (node_distance(local_node, node) !=
5630 node_distance(local_node, prev_node))
5631 node_load[node] = load;
5633 node_order[nr_nodes++] = node;
5638 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5639 build_thisnode_zonelists(pgdat);
5642 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5644 * Return node id of node used for "local" allocations.
5645 * I.e., first node id of first zone in arg node's generic zonelist.
5646 * Used for initializing percpu 'numa_mem', which is used primarily
5647 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5649 int local_memory_node(int node)
5653 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5654 gfp_zone(GFP_KERNEL),
5656 return zone_to_nid(z->zone);
5660 static void setup_min_unmapped_ratio(void);
5661 static void setup_min_slab_ratio(void);
5662 #else /* CONFIG_NUMA */
5664 static void build_zonelists(pg_data_t *pgdat)
5666 int node, local_node;
5667 struct zoneref *zonerefs;
5670 local_node = pgdat->node_id;
5672 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5673 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5674 zonerefs += nr_zones;
5677 * Now we build the zonelist so that it contains the zones
5678 * of all the other nodes.
5679 * We don't want to pressure a particular node, so when
5680 * building the zones for node N, we make sure that the
5681 * zones coming right after the local ones are those from
5682 * node N+1 (modulo N)
5684 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5685 if (!node_online(node))
5687 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5688 zonerefs += nr_zones;
5690 for (node = 0; node < local_node; node++) {
5691 if (!node_online(node))
5693 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5694 zonerefs += nr_zones;
5697 zonerefs->zone = NULL;
5698 zonerefs->zone_idx = 0;
5701 #endif /* CONFIG_NUMA */
5704 * Boot pageset table. One per cpu which is going to be used for all
5705 * zones and all nodes. The parameters will be set in such a way
5706 * that an item put on a list will immediately be handed over to
5707 * the buddy list. This is safe since pageset manipulation is done
5708 * with interrupts disabled.
5710 * The boot_pagesets must be kept even after bootup is complete for
5711 * unused processors and/or zones. They do play a role for bootstrapping
5712 * hotplugged processors.
5714 * zoneinfo_show() and maybe other functions do
5715 * not check if the processor is online before following the pageset pointer.
5716 * Other parts of the kernel may not check if the zone is available.
5718 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5719 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5720 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5722 static void __build_all_zonelists(void *data)
5725 int __maybe_unused cpu;
5726 pg_data_t *self = data;
5727 static DEFINE_SPINLOCK(lock);
5732 memset(node_load, 0, sizeof(node_load));
5736 * This node is hotadded and no memory is yet present. So just
5737 * building zonelists is fine - no need to touch other nodes.
5739 if (self && !node_online(self->node_id)) {
5740 build_zonelists(self);
5742 for_each_online_node(nid) {
5743 pg_data_t *pgdat = NODE_DATA(nid);
5745 build_zonelists(pgdat);
5748 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5750 * We now know the "local memory node" for each node--
5751 * i.e., the node of the first zone in the generic zonelist.
5752 * Set up numa_mem percpu variable for on-line cpus. During
5753 * boot, only the boot cpu should be on-line; we'll init the
5754 * secondary cpus' numa_mem as they come on-line. During
5755 * node/memory hotplug, we'll fixup all on-line cpus.
5757 for_each_online_cpu(cpu)
5758 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5765 static noinline void __init
5766 build_all_zonelists_init(void)
5770 __build_all_zonelists(NULL);
5773 * Initialize the boot_pagesets that are going to be used
5774 * for bootstrapping processors. The real pagesets for
5775 * each zone will be allocated later when the per cpu
5776 * allocator is available.
5778 * boot_pagesets are used also for bootstrapping offline
5779 * cpus if the system is already booted because the pagesets
5780 * are needed to initialize allocators on a specific cpu too.
5781 * F.e. the percpu allocator needs the page allocator which
5782 * needs the percpu allocator in order to allocate its pagesets
5783 * (a chicken-egg dilemma).
5785 for_each_possible_cpu(cpu)
5786 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5788 mminit_verify_zonelist();
5789 cpuset_init_current_mems_allowed();
5793 * unless system_state == SYSTEM_BOOTING.
5795 * __ref due to call of __init annotated helper build_all_zonelists_init
5796 * [protected by SYSTEM_BOOTING].
5798 void __ref build_all_zonelists(pg_data_t *pgdat)
5800 if (system_state == SYSTEM_BOOTING) {
5801 build_all_zonelists_init();
5803 __build_all_zonelists(pgdat);
5804 /* cpuset refresh routine should be here */
5806 vm_total_pages = nr_free_pagecache_pages();
5808 * Disable grouping by mobility if the number of pages in the
5809 * system is too low to allow the mechanism to work. It would be
5810 * more accurate, but expensive to check per-zone. This check is
5811 * made on memory-hotadd so a system can start with mobility
5812 * disabled and enable it later
5814 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5815 page_group_by_mobility_disabled = 1;
5817 page_group_by_mobility_disabled = 0;
5819 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5821 page_group_by_mobility_disabled ? "off" : "on",
5824 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5828 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5829 static bool __meminit
5830 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5832 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5833 static struct memblock_region *r;
5835 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5836 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5837 for_each_memblock(memory, r) {
5838 if (*pfn < memblock_region_memory_end_pfn(r))
5842 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5843 memblock_is_mirror(r)) {
5844 *pfn = memblock_region_memory_end_pfn(r);
5852 #ifdef CONFIG_SPARSEMEM
5853 /* Skip PFNs that belong to non-present sections */
5854 static inline __meminit unsigned long next_pfn(unsigned long pfn)
5856 const unsigned long section_nr = pfn_to_section_nr(++pfn);
5858 if (present_section_nr(section_nr))
5860 return section_nr_to_pfn(next_present_section_nr(section_nr));
5863 static inline __meminit unsigned long next_pfn(unsigned long pfn)
5870 * Initially all pages are reserved - free ones are freed
5871 * up by memblock_free_all() once the early boot process is
5872 * done. Non-atomic initialization, single-pass.
5874 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5875 unsigned long start_pfn, enum memmap_context context,
5876 struct vmem_altmap *altmap)
5878 unsigned long pfn, end_pfn = start_pfn + size;
5881 if (highest_memmap_pfn < end_pfn - 1)
5882 highest_memmap_pfn = end_pfn - 1;
5884 #ifdef CONFIG_ZONE_DEVICE
5886 * Honor reservation requested by the driver for this ZONE_DEVICE
5887 * memory. We limit the total number of pages to initialize to just
5888 * those that might contain the memory mapping. We will defer the
5889 * ZONE_DEVICE page initialization until after we have released
5892 if (zone == ZONE_DEVICE) {
5896 if (start_pfn == altmap->base_pfn)
5897 start_pfn += altmap->reserve;
5898 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5902 for (pfn = start_pfn; pfn < end_pfn; ) {
5904 * There can be holes in boot-time mem_map[]s handed to this
5905 * function. They do not exist on hotplugged memory.
5907 if (context == MEMMAP_EARLY) {
5908 if (!early_pfn_valid(pfn)) {
5909 pfn = next_pfn(pfn);
5912 if (!early_pfn_in_nid(pfn, nid)) {
5916 if (overlap_memmap_init(zone, &pfn))
5918 if (defer_init(nid, pfn, end_pfn))
5922 page = pfn_to_page(pfn);
5923 __init_single_page(page, pfn, zone, nid);
5924 if (context == MEMMAP_HOTPLUG)
5925 __SetPageReserved(page);
5928 * Mark the block movable so that blocks are reserved for
5929 * movable at startup. This will force kernel allocations
5930 * to reserve their blocks rather than leaking throughout
5931 * the address space during boot when many long-lived
5932 * kernel allocations are made.
5934 * bitmap is created for zone's valid pfn range. but memmap
5935 * can be created for invalid pages (for alignment)
5936 * check here not to call set_pageblock_migratetype() against
5939 if (!(pfn & (pageblock_nr_pages - 1))) {
5940 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5947 #ifdef CONFIG_ZONE_DEVICE
5948 void __ref memmap_init_zone_device(struct zone *zone,
5949 unsigned long start_pfn,
5950 unsigned long nr_pages,
5951 struct dev_pagemap *pgmap)
5953 unsigned long pfn, end_pfn = start_pfn + nr_pages;
5954 struct pglist_data *pgdat = zone->zone_pgdat;
5955 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
5956 unsigned long zone_idx = zone_idx(zone);
5957 unsigned long start = jiffies;
5958 int nid = pgdat->node_id;
5960 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
5964 * The call to memmap_init_zone should have already taken care
5965 * of the pages reserved for the memmap, so we can just jump to
5966 * the end of that region and start processing the device pages.
5969 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5970 nr_pages = end_pfn - start_pfn;
5973 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5974 struct page *page = pfn_to_page(pfn);
5976 __init_single_page(page, pfn, zone_idx, nid);
5979 * Mark page reserved as it will need to wait for onlining
5980 * phase for it to be fully associated with a zone.
5982 * We can use the non-atomic __set_bit operation for setting
5983 * the flag as we are still initializing the pages.
5985 __SetPageReserved(page);
5988 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5989 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5990 * ever freed or placed on a driver-private list.
5992 page->pgmap = pgmap;
5993 page->zone_device_data = NULL;
5996 * Mark the block movable so that blocks are reserved for
5997 * movable at startup. This will force kernel allocations
5998 * to reserve their blocks rather than leaking throughout
5999 * the address space during boot when many long-lived
6000 * kernel allocations are made.
6002 * bitmap is created for zone's valid pfn range. but memmap
6003 * can be created for invalid pages (for alignment)
6004 * check here not to call set_pageblock_migratetype() against
6007 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6008 * because this is done early in section_activate()
6010 if (!(pfn & (pageblock_nr_pages - 1))) {
6011 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6016 pr_info("%s initialised %lu pages in %ums\n", __func__,
6017 nr_pages, jiffies_to_msecs(jiffies - start));
6021 static void __meminit zone_init_free_lists(struct zone *zone)
6023 unsigned int order, t;
6024 for_each_migratetype_order(order, t) {
6025 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6026 zone->free_area[order].nr_free = 0;
6030 void __meminit __weak memmap_init(unsigned long size, int nid,
6031 unsigned long zone, unsigned long start_pfn)
6033 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6036 static int zone_batchsize(struct zone *zone)
6042 * The per-cpu-pages pools are set to around 1000th of the
6045 batch = zone_managed_pages(zone) / 1024;
6046 /* But no more than a meg. */
6047 if (batch * PAGE_SIZE > 1024 * 1024)
6048 batch = (1024 * 1024) / PAGE_SIZE;
6049 batch /= 4; /* We effectively *= 4 below */
6054 * Clamp the batch to a 2^n - 1 value. Having a power
6055 * of 2 value was found to be more likely to have
6056 * suboptimal cache aliasing properties in some cases.
6058 * For example if 2 tasks are alternately allocating
6059 * batches of pages, one task can end up with a lot
6060 * of pages of one half of the possible page colors
6061 * and the other with pages of the other colors.
6063 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6068 /* The deferral and batching of frees should be suppressed under NOMMU
6071 * The problem is that NOMMU needs to be able to allocate large chunks
6072 * of contiguous memory as there's no hardware page translation to
6073 * assemble apparent contiguous memory from discontiguous pages.
6075 * Queueing large contiguous runs of pages for batching, however,
6076 * causes the pages to actually be freed in smaller chunks. As there
6077 * can be a significant delay between the individual batches being
6078 * recycled, this leads to the once large chunks of space being
6079 * fragmented and becoming unavailable for high-order allocations.
6086 * pcp->high and pcp->batch values are related and dependent on one another:
6087 * ->batch must never be higher then ->high.
6088 * The following function updates them in a safe manner without read side
6091 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6092 * those fields changing asynchronously (acording the the above rule).
6094 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6095 * outside of boot time (or some other assurance that no concurrent updaters
6098 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6099 unsigned long batch)
6101 /* start with a fail safe value for batch */
6105 /* Update high, then batch, in order */
6112 /* a companion to pageset_set_high() */
6113 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6115 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6118 static void pageset_init(struct per_cpu_pageset *p)
6120 struct per_cpu_pages *pcp;
6123 memset(p, 0, sizeof(*p));
6126 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6127 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6130 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6133 pageset_set_batch(p, batch);
6137 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6138 * to the value high for the pageset p.
6140 static void pageset_set_high(struct per_cpu_pageset *p,
6143 unsigned long batch = max(1UL, high / 4);
6144 if ((high / 4) > (PAGE_SHIFT * 8))
6145 batch = PAGE_SHIFT * 8;
6147 pageset_update(&p->pcp, high, batch);
6150 static void pageset_set_high_and_batch(struct zone *zone,
6151 struct per_cpu_pageset *pcp)
6153 if (percpu_pagelist_fraction)
6154 pageset_set_high(pcp,
6155 (zone_managed_pages(zone) /
6156 percpu_pagelist_fraction));
6158 pageset_set_batch(pcp, zone_batchsize(zone));
6161 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6163 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6166 pageset_set_high_and_batch(zone, pcp);
6169 void __meminit setup_zone_pageset(struct zone *zone)
6172 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6173 for_each_possible_cpu(cpu)
6174 zone_pageset_init(zone, cpu);
6178 * Allocate per cpu pagesets and initialize them.
6179 * Before this call only boot pagesets were available.
6181 void __init setup_per_cpu_pageset(void)
6183 struct pglist_data *pgdat;
6186 for_each_populated_zone(zone)
6187 setup_zone_pageset(zone);
6189 for_each_online_pgdat(pgdat)
6190 pgdat->per_cpu_nodestats =
6191 alloc_percpu(struct per_cpu_nodestat);
6194 static __meminit void zone_pcp_init(struct zone *zone)
6197 * per cpu subsystem is not up at this point. The following code
6198 * relies on the ability of the linker to provide the
6199 * offset of a (static) per cpu variable into the per cpu area.
6201 zone->pageset = &boot_pageset;
6203 if (populated_zone(zone))
6204 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6205 zone->name, zone->present_pages,
6206 zone_batchsize(zone));
6209 void __meminit init_currently_empty_zone(struct zone *zone,
6210 unsigned long zone_start_pfn,
6213 struct pglist_data *pgdat = zone->zone_pgdat;
6214 int zone_idx = zone_idx(zone) + 1;
6216 if (zone_idx > pgdat->nr_zones)
6217 pgdat->nr_zones = zone_idx;
6219 zone->zone_start_pfn = zone_start_pfn;
6221 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6222 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6224 (unsigned long)zone_idx(zone),
6225 zone_start_pfn, (zone_start_pfn + size));
6227 zone_init_free_lists(zone);
6228 zone->initialized = 1;
6231 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6232 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6235 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6237 int __meminit __early_pfn_to_nid(unsigned long pfn,
6238 struct mminit_pfnnid_cache *state)
6240 unsigned long start_pfn, end_pfn;
6243 if (state->last_start <= pfn && pfn < state->last_end)
6244 return state->last_nid;
6246 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6247 if (nid != NUMA_NO_NODE) {
6248 state->last_start = start_pfn;
6249 state->last_end = end_pfn;
6250 state->last_nid = nid;
6255 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6258 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6259 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6260 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6262 * If an architecture guarantees that all ranges registered contain no holes
6263 * and may be freed, this this function may be used instead of calling
6264 * memblock_free_early_nid() manually.
6266 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6268 unsigned long start_pfn, end_pfn;
6271 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6272 start_pfn = min(start_pfn, max_low_pfn);
6273 end_pfn = min(end_pfn, max_low_pfn);
6275 if (start_pfn < end_pfn)
6276 memblock_free_early_nid(PFN_PHYS(start_pfn),
6277 (end_pfn - start_pfn) << PAGE_SHIFT,
6283 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6284 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6286 * If an architecture guarantees that all ranges registered contain no holes and may
6287 * be freed, this function may be used instead of calling memory_present() manually.
6289 void __init sparse_memory_present_with_active_regions(int nid)
6291 unsigned long start_pfn, end_pfn;
6294 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6295 memory_present(this_nid, start_pfn, end_pfn);
6299 * get_pfn_range_for_nid - Return the start and end page frames for a node
6300 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6301 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6302 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6304 * It returns the start and end page frame of a node based on information
6305 * provided by memblock_set_node(). If called for a node
6306 * with no available memory, a warning is printed and the start and end
6309 void __init get_pfn_range_for_nid(unsigned int nid,
6310 unsigned long *start_pfn, unsigned long *end_pfn)
6312 unsigned long this_start_pfn, this_end_pfn;
6318 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6319 *start_pfn = min(*start_pfn, this_start_pfn);
6320 *end_pfn = max(*end_pfn, this_end_pfn);
6323 if (*start_pfn == -1UL)
6328 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6329 * assumption is made that zones within a node are ordered in monotonic
6330 * increasing memory addresses so that the "highest" populated zone is used
6332 static void __init find_usable_zone_for_movable(void)
6335 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6336 if (zone_index == ZONE_MOVABLE)
6339 if (arch_zone_highest_possible_pfn[zone_index] >
6340 arch_zone_lowest_possible_pfn[zone_index])
6344 VM_BUG_ON(zone_index == -1);
6345 movable_zone = zone_index;
6349 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6350 * because it is sized independent of architecture. Unlike the other zones,
6351 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6352 * in each node depending on the size of each node and how evenly kernelcore
6353 * is distributed. This helper function adjusts the zone ranges
6354 * provided by the architecture for a given node by using the end of the
6355 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6356 * zones within a node are in order of monotonic increases memory addresses
6358 static void __init adjust_zone_range_for_zone_movable(int nid,
6359 unsigned long zone_type,
6360 unsigned long node_start_pfn,
6361 unsigned long node_end_pfn,
6362 unsigned long *zone_start_pfn,
6363 unsigned long *zone_end_pfn)
6365 /* Only adjust if ZONE_MOVABLE is on this node */
6366 if (zone_movable_pfn[nid]) {
6367 /* Size ZONE_MOVABLE */
6368 if (zone_type == ZONE_MOVABLE) {
6369 *zone_start_pfn = zone_movable_pfn[nid];
6370 *zone_end_pfn = min(node_end_pfn,
6371 arch_zone_highest_possible_pfn[movable_zone]);
6373 /* Adjust for ZONE_MOVABLE starting within this range */
6374 } else if (!mirrored_kernelcore &&
6375 *zone_start_pfn < zone_movable_pfn[nid] &&
6376 *zone_end_pfn > zone_movable_pfn[nid]) {
6377 *zone_end_pfn = zone_movable_pfn[nid];
6379 /* Check if this whole range is within ZONE_MOVABLE */
6380 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6381 *zone_start_pfn = *zone_end_pfn;
6386 * Return the number of pages a zone spans in a node, including holes
6387 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6389 static unsigned long __init zone_spanned_pages_in_node(int nid,
6390 unsigned long zone_type,
6391 unsigned long node_start_pfn,
6392 unsigned long node_end_pfn,
6393 unsigned long *zone_start_pfn,
6394 unsigned long *zone_end_pfn,
6395 unsigned long *ignored)
6397 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6398 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6399 /* When hotadd a new node from cpu_up(), the node should be empty */
6400 if (!node_start_pfn && !node_end_pfn)
6403 /* Get the start and end of the zone */
6404 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6405 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6406 adjust_zone_range_for_zone_movable(nid, zone_type,
6407 node_start_pfn, node_end_pfn,
6408 zone_start_pfn, zone_end_pfn);
6410 /* Check that this node has pages within the zone's required range */
6411 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6414 /* Move the zone boundaries inside the node if necessary */
6415 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6416 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6418 /* Return the spanned pages */
6419 return *zone_end_pfn - *zone_start_pfn;
6423 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6424 * then all holes in the requested range will be accounted for.
6426 unsigned long __init __absent_pages_in_range(int nid,
6427 unsigned long range_start_pfn,
6428 unsigned long range_end_pfn)
6430 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6431 unsigned long start_pfn, end_pfn;
6434 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6435 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6436 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6437 nr_absent -= end_pfn - start_pfn;
6443 * absent_pages_in_range - Return number of page frames in holes within a range
6444 * @start_pfn: The start PFN to start searching for holes
6445 * @end_pfn: The end PFN to stop searching for holes
6447 * Return: the number of pages frames in memory holes within a range.
6449 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6450 unsigned long end_pfn)
6452 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6455 /* Return the number of page frames in holes in a zone on a node */
6456 static unsigned long __init zone_absent_pages_in_node(int nid,
6457 unsigned long zone_type,
6458 unsigned long node_start_pfn,
6459 unsigned long node_end_pfn,
6460 unsigned long *ignored)
6462 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6463 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6464 unsigned long zone_start_pfn, zone_end_pfn;
6465 unsigned long nr_absent;
6467 /* When hotadd a new node from cpu_up(), the node should be empty */
6468 if (!node_start_pfn && !node_end_pfn)
6471 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6472 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6474 adjust_zone_range_for_zone_movable(nid, zone_type,
6475 node_start_pfn, node_end_pfn,
6476 &zone_start_pfn, &zone_end_pfn);
6477 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6480 * ZONE_MOVABLE handling.
6481 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6484 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6485 unsigned long start_pfn, end_pfn;
6486 struct memblock_region *r;
6488 for_each_memblock(memory, r) {
6489 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6490 zone_start_pfn, zone_end_pfn);
6491 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6492 zone_start_pfn, zone_end_pfn);
6494 if (zone_type == ZONE_MOVABLE &&
6495 memblock_is_mirror(r))
6496 nr_absent += end_pfn - start_pfn;
6498 if (zone_type == ZONE_NORMAL &&
6499 !memblock_is_mirror(r))
6500 nr_absent += end_pfn - start_pfn;
6507 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6508 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6509 unsigned long zone_type,
6510 unsigned long node_start_pfn,
6511 unsigned long node_end_pfn,
6512 unsigned long *zone_start_pfn,
6513 unsigned long *zone_end_pfn,
6514 unsigned long *zones_size)
6518 *zone_start_pfn = node_start_pfn;
6519 for (zone = 0; zone < zone_type; zone++)
6520 *zone_start_pfn += zones_size[zone];
6522 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6524 return zones_size[zone_type];
6527 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6528 unsigned long zone_type,
6529 unsigned long node_start_pfn,
6530 unsigned long node_end_pfn,
6531 unsigned long *zholes_size)
6536 return zholes_size[zone_type];
6539 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6541 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6542 unsigned long node_start_pfn,
6543 unsigned long node_end_pfn,
6544 unsigned long *zones_size,
6545 unsigned long *zholes_size)
6547 unsigned long realtotalpages = 0, totalpages = 0;
6550 for (i = 0; i < MAX_NR_ZONES; i++) {
6551 struct zone *zone = pgdat->node_zones + i;
6552 unsigned long zone_start_pfn, zone_end_pfn;
6553 unsigned long size, real_size;
6555 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6561 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6562 node_start_pfn, node_end_pfn,
6565 zone->zone_start_pfn = zone_start_pfn;
6567 zone->zone_start_pfn = 0;
6568 zone->spanned_pages = size;
6569 zone->present_pages = real_size;
6572 realtotalpages += real_size;
6575 pgdat->node_spanned_pages = totalpages;
6576 pgdat->node_present_pages = realtotalpages;
6577 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6581 #ifndef CONFIG_SPARSEMEM
6583 * Calculate the size of the zone->blockflags rounded to an unsigned long
6584 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6585 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6586 * round what is now in bits to nearest long in bits, then return it in
6589 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6591 unsigned long usemapsize;
6593 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6594 usemapsize = roundup(zonesize, pageblock_nr_pages);
6595 usemapsize = usemapsize >> pageblock_order;
6596 usemapsize *= NR_PAGEBLOCK_BITS;
6597 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6599 return usemapsize / 8;
6602 static void __ref setup_usemap(struct pglist_data *pgdat,
6604 unsigned long zone_start_pfn,
6605 unsigned long zonesize)
6607 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6608 zone->pageblock_flags = NULL;
6610 zone->pageblock_flags =
6611 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6613 if (!zone->pageblock_flags)
6614 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6615 usemapsize, zone->name, pgdat->node_id);
6619 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6620 unsigned long zone_start_pfn, unsigned long zonesize) {}
6621 #endif /* CONFIG_SPARSEMEM */
6623 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6625 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6626 void __init set_pageblock_order(void)
6630 /* Check that pageblock_nr_pages has not already been setup */
6631 if (pageblock_order)
6634 if (HPAGE_SHIFT > PAGE_SHIFT)
6635 order = HUGETLB_PAGE_ORDER;
6637 order = MAX_ORDER - 1;
6640 * Assume the largest contiguous order of interest is a huge page.
6641 * This value may be variable depending on boot parameters on IA64 and
6644 pageblock_order = order;
6646 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6649 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6650 * is unused as pageblock_order is set at compile-time. See
6651 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6654 void __init set_pageblock_order(void)
6658 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6660 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6661 unsigned long present_pages)
6663 unsigned long pages = spanned_pages;
6666 * Provide a more accurate estimation if there are holes within
6667 * the zone and SPARSEMEM is in use. If there are holes within the
6668 * zone, each populated memory region may cost us one or two extra
6669 * memmap pages due to alignment because memmap pages for each
6670 * populated regions may not be naturally aligned on page boundary.
6671 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6673 if (spanned_pages > present_pages + (present_pages >> 4) &&
6674 IS_ENABLED(CONFIG_SPARSEMEM))
6675 pages = present_pages;
6677 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6680 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6681 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6683 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6685 spin_lock_init(&ds_queue->split_queue_lock);
6686 INIT_LIST_HEAD(&ds_queue->split_queue);
6687 ds_queue->split_queue_len = 0;
6690 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6693 #ifdef CONFIG_COMPACTION
6694 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6696 init_waitqueue_head(&pgdat->kcompactd_wait);
6699 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6702 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6704 pgdat_resize_init(pgdat);
6706 pgdat_init_split_queue(pgdat);
6707 pgdat_init_kcompactd(pgdat);
6709 init_waitqueue_head(&pgdat->kswapd_wait);
6710 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6712 pgdat_page_ext_init(pgdat);
6713 spin_lock_init(&pgdat->lru_lock);
6714 lruvec_init(&pgdat->__lruvec);
6717 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6718 unsigned long remaining_pages)
6720 atomic_long_set(&zone->managed_pages, remaining_pages);
6721 zone_set_nid(zone, nid);
6722 zone->name = zone_names[idx];
6723 zone->zone_pgdat = NODE_DATA(nid);
6724 spin_lock_init(&zone->lock);
6725 zone_seqlock_init(zone);
6726 zone_pcp_init(zone);
6730 * Set up the zone data structures
6731 * - init pgdat internals
6732 * - init all zones belonging to this node
6734 * NOTE: this function is only called during memory hotplug
6736 #ifdef CONFIG_MEMORY_HOTPLUG
6737 void __ref free_area_init_core_hotplug(int nid)
6740 pg_data_t *pgdat = NODE_DATA(nid);
6742 pgdat_init_internals(pgdat);
6743 for (z = 0; z < MAX_NR_ZONES; z++)
6744 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6749 * Set up the zone data structures:
6750 * - mark all pages reserved
6751 * - mark all memory queues empty
6752 * - clear the memory bitmaps
6754 * NOTE: pgdat should get zeroed by caller.
6755 * NOTE: this function is only called during early init.
6757 static void __init free_area_init_core(struct pglist_data *pgdat)
6760 int nid = pgdat->node_id;
6762 pgdat_init_internals(pgdat);
6763 pgdat->per_cpu_nodestats = &boot_nodestats;
6765 for (j = 0; j < MAX_NR_ZONES; j++) {
6766 struct zone *zone = pgdat->node_zones + j;
6767 unsigned long size, freesize, memmap_pages;
6768 unsigned long zone_start_pfn = zone->zone_start_pfn;
6770 size = zone->spanned_pages;
6771 freesize = zone->present_pages;
6774 * Adjust freesize so that it accounts for how much memory
6775 * is used by this zone for memmap. This affects the watermark
6776 * and per-cpu initialisations
6778 memmap_pages = calc_memmap_size(size, freesize);
6779 if (!is_highmem_idx(j)) {
6780 if (freesize >= memmap_pages) {
6781 freesize -= memmap_pages;
6784 " %s zone: %lu pages used for memmap\n",
6785 zone_names[j], memmap_pages);
6787 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6788 zone_names[j], memmap_pages, freesize);
6791 /* Account for reserved pages */
6792 if (j == 0 && freesize > dma_reserve) {
6793 freesize -= dma_reserve;
6794 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6795 zone_names[0], dma_reserve);
6798 if (!is_highmem_idx(j))
6799 nr_kernel_pages += freesize;
6800 /* Charge for highmem memmap if there are enough kernel pages */
6801 else if (nr_kernel_pages > memmap_pages * 2)
6802 nr_kernel_pages -= memmap_pages;
6803 nr_all_pages += freesize;
6806 * Set an approximate value for lowmem here, it will be adjusted
6807 * when the bootmem allocator frees pages into the buddy system.
6808 * And all highmem pages will be managed by the buddy system.
6810 zone_init_internals(zone, j, nid, freesize);
6815 set_pageblock_order();
6816 setup_usemap(pgdat, zone, zone_start_pfn, size);
6817 init_currently_empty_zone(zone, zone_start_pfn, size);
6818 memmap_init(size, nid, j, zone_start_pfn);
6822 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6823 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6825 unsigned long __maybe_unused start = 0;
6826 unsigned long __maybe_unused offset = 0;
6828 /* Skip empty nodes */
6829 if (!pgdat->node_spanned_pages)
6832 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6833 offset = pgdat->node_start_pfn - start;
6834 /* ia64 gets its own node_mem_map, before this, without bootmem */
6835 if (!pgdat->node_mem_map) {
6836 unsigned long size, end;
6840 * The zone's endpoints aren't required to be MAX_ORDER
6841 * aligned but the node_mem_map endpoints must be in order
6842 * for the buddy allocator to function correctly.
6844 end = pgdat_end_pfn(pgdat);
6845 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6846 size = (end - start) * sizeof(struct page);
6847 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6850 panic("Failed to allocate %ld bytes for node %d memory map\n",
6851 size, pgdat->node_id);
6852 pgdat->node_mem_map = map + offset;
6854 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6855 __func__, pgdat->node_id, (unsigned long)pgdat,
6856 (unsigned long)pgdat->node_mem_map);
6857 #ifndef CONFIG_NEED_MULTIPLE_NODES
6859 * With no DISCONTIG, the global mem_map is just set as node 0's
6861 if (pgdat == NODE_DATA(0)) {
6862 mem_map = NODE_DATA(0)->node_mem_map;
6863 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6864 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6866 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6871 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6872 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6874 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6875 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6877 pgdat->first_deferred_pfn = ULONG_MAX;
6880 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6883 void __init free_area_init_node(int nid, unsigned long *zones_size,
6884 unsigned long node_start_pfn,
6885 unsigned long *zholes_size)
6887 pg_data_t *pgdat = NODE_DATA(nid);
6888 unsigned long start_pfn = 0;
6889 unsigned long end_pfn = 0;
6891 /* pg_data_t should be reset to zero when it's allocated */
6892 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6894 pgdat->node_id = nid;
6895 pgdat->node_start_pfn = node_start_pfn;
6896 pgdat->per_cpu_nodestats = NULL;
6897 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6898 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6899 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6900 (u64)start_pfn << PAGE_SHIFT,
6901 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6903 start_pfn = node_start_pfn;
6905 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6906 zones_size, zholes_size);
6908 alloc_node_mem_map(pgdat);
6909 pgdat_set_deferred_range(pgdat);
6911 free_area_init_core(pgdat);
6914 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6916 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6917 * PageReserved(). Return the number of struct pages that were initialized.
6919 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6924 for (pfn = spfn; pfn < epfn; pfn++) {
6925 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6926 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6927 + pageblock_nr_pages - 1;
6931 * Use a fake node/zone (0) for now. Some of these pages
6932 * (in memblock.reserved but not in memblock.memory) will
6933 * get re-initialized via reserve_bootmem_region() later.
6935 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6936 __SetPageReserved(pfn_to_page(pfn));
6944 * Only struct pages that are backed by physical memory are zeroed and
6945 * initialized by going through __init_single_page(). But, there are some
6946 * struct pages which are reserved in memblock allocator and their fields
6947 * may be accessed (for example page_to_pfn() on some configuration accesses
6948 * flags). We must explicitly initialize those struct pages.
6950 * This function also addresses a similar issue where struct pages are left
6951 * uninitialized because the physical address range is not covered by
6952 * memblock.memory or memblock.reserved. That could happen when memblock
6953 * layout is manually configured via memmap=, or when the highest physical
6954 * address (max_pfn) does not end on a section boundary.
6956 static void __init init_unavailable_mem(void)
6958 phys_addr_t start, end;
6960 phys_addr_t next = 0;
6963 * Loop through unavailable ranges not covered by memblock.memory.
6966 for_each_mem_range(i, &memblock.memory, NULL,
6967 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6969 pgcnt += init_unavailable_range(PFN_DOWN(next),
6975 * Early sections always have a fully populated memmap for the whole
6976 * section - see pfn_valid(). If the last section has holes at the
6977 * end and that section is marked "online", the memmap will be
6978 * considered initialized. Make sure that memmap has a well defined
6981 pgcnt += init_unavailable_range(PFN_DOWN(next),
6982 round_up(max_pfn, PAGES_PER_SECTION));
6985 * Struct pages that do not have backing memory. This could be because
6986 * firmware is using some of this memory, or for some other reasons.
6989 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6992 static inline void __init init_unavailable_mem(void)
6995 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6997 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6999 #if MAX_NUMNODES > 1
7001 * Figure out the number of possible node ids.
7003 void __init setup_nr_node_ids(void)
7005 unsigned int highest;
7007 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7008 nr_node_ids = highest + 1;
7013 * node_map_pfn_alignment - determine the maximum internode alignment
7015 * This function should be called after node map is populated and sorted.
7016 * It calculates the maximum power of two alignment which can distinguish
7019 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7020 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7021 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7022 * shifted, 1GiB is enough and this function will indicate so.
7024 * This is used to test whether pfn -> nid mapping of the chosen memory
7025 * model has fine enough granularity to avoid incorrect mapping for the
7026 * populated node map.
7028 * Return: the determined alignment in pfn's. 0 if there is no alignment
7029 * requirement (single node).
7031 unsigned long __init node_map_pfn_alignment(void)
7033 unsigned long accl_mask = 0, last_end = 0;
7034 unsigned long start, end, mask;
7035 int last_nid = NUMA_NO_NODE;
7038 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7039 if (!start || last_nid < 0 || last_nid == nid) {
7046 * Start with a mask granular enough to pin-point to the
7047 * start pfn and tick off bits one-by-one until it becomes
7048 * too coarse to separate the current node from the last.
7050 mask = ~((1 << __ffs(start)) - 1);
7051 while (mask && last_end <= (start & (mask << 1)))
7054 /* accumulate all internode masks */
7058 /* convert mask to number of pages */
7059 return ~accl_mask + 1;
7062 /* Find the lowest pfn for a node */
7063 static unsigned long __init find_min_pfn_for_node(int nid)
7065 unsigned long min_pfn = ULONG_MAX;
7066 unsigned long start_pfn;
7069 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7070 min_pfn = min(min_pfn, start_pfn);
7072 if (min_pfn == ULONG_MAX) {
7073 pr_warn("Could not find start_pfn for node %d\n", nid);
7081 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7083 * Return: the minimum PFN based on information provided via
7084 * memblock_set_node().
7086 unsigned long __init find_min_pfn_with_active_regions(void)
7088 return find_min_pfn_for_node(MAX_NUMNODES);
7092 * early_calculate_totalpages()
7093 * Sum pages in active regions for movable zone.
7094 * Populate N_MEMORY for calculating usable_nodes.
7096 static unsigned long __init early_calculate_totalpages(void)
7098 unsigned long totalpages = 0;
7099 unsigned long start_pfn, end_pfn;
7102 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7103 unsigned long pages = end_pfn - start_pfn;
7105 totalpages += pages;
7107 node_set_state(nid, N_MEMORY);
7113 * Find the PFN the Movable zone begins in each node. Kernel memory
7114 * is spread evenly between nodes as long as the nodes have enough
7115 * memory. When they don't, some nodes will have more kernelcore than
7118 static void __init find_zone_movable_pfns_for_nodes(void)
7121 unsigned long usable_startpfn;
7122 unsigned long kernelcore_node, kernelcore_remaining;
7123 /* save the state before borrow the nodemask */
7124 nodemask_t saved_node_state = node_states[N_MEMORY];
7125 unsigned long totalpages = early_calculate_totalpages();
7126 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7127 struct memblock_region *r;
7129 /* Need to find movable_zone earlier when movable_node is specified. */
7130 find_usable_zone_for_movable();
7133 * If movable_node is specified, ignore kernelcore and movablecore
7136 if (movable_node_is_enabled()) {
7137 for_each_memblock(memory, r) {
7138 if (!memblock_is_hotpluggable(r))
7143 usable_startpfn = PFN_DOWN(r->base);
7144 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7145 min(usable_startpfn, zone_movable_pfn[nid]) :
7153 * If kernelcore=mirror is specified, ignore movablecore option
7155 if (mirrored_kernelcore) {
7156 bool mem_below_4gb_not_mirrored = false;
7158 for_each_memblock(memory, r) {
7159 if (memblock_is_mirror(r))
7164 usable_startpfn = memblock_region_memory_base_pfn(r);
7166 if (usable_startpfn < 0x100000) {
7167 mem_below_4gb_not_mirrored = true;
7171 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7172 min(usable_startpfn, zone_movable_pfn[nid]) :
7176 if (mem_below_4gb_not_mirrored)
7177 pr_warn("This configuration results in unmirrored kernel memory.");
7183 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7184 * amount of necessary memory.
7186 if (required_kernelcore_percent)
7187 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7189 if (required_movablecore_percent)
7190 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7194 * If movablecore= was specified, calculate what size of
7195 * kernelcore that corresponds so that memory usable for
7196 * any allocation type is evenly spread. If both kernelcore
7197 * and movablecore are specified, then the value of kernelcore
7198 * will be used for required_kernelcore if it's greater than
7199 * what movablecore would have allowed.
7201 if (required_movablecore) {
7202 unsigned long corepages;
7205 * Round-up so that ZONE_MOVABLE is at least as large as what
7206 * was requested by the user
7208 required_movablecore =
7209 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7210 required_movablecore = min(totalpages, required_movablecore);
7211 corepages = totalpages - required_movablecore;
7213 required_kernelcore = max(required_kernelcore, corepages);
7217 * If kernelcore was not specified or kernelcore size is larger
7218 * than totalpages, there is no ZONE_MOVABLE.
7220 if (!required_kernelcore || required_kernelcore >= totalpages)
7223 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7224 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7227 /* Spread kernelcore memory as evenly as possible throughout nodes */
7228 kernelcore_node = required_kernelcore / usable_nodes;
7229 for_each_node_state(nid, N_MEMORY) {
7230 unsigned long start_pfn, end_pfn;
7233 * Recalculate kernelcore_node if the division per node
7234 * now exceeds what is necessary to satisfy the requested
7235 * amount of memory for the kernel
7237 if (required_kernelcore < kernelcore_node)
7238 kernelcore_node = required_kernelcore / usable_nodes;
7241 * As the map is walked, we track how much memory is usable
7242 * by the kernel using kernelcore_remaining. When it is
7243 * 0, the rest of the node is usable by ZONE_MOVABLE
7245 kernelcore_remaining = kernelcore_node;
7247 /* Go through each range of PFNs within this node */
7248 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7249 unsigned long size_pages;
7251 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7252 if (start_pfn >= end_pfn)
7255 /* Account for what is only usable for kernelcore */
7256 if (start_pfn < usable_startpfn) {
7257 unsigned long kernel_pages;
7258 kernel_pages = min(end_pfn, usable_startpfn)
7261 kernelcore_remaining -= min(kernel_pages,
7262 kernelcore_remaining);
7263 required_kernelcore -= min(kernel_pages,
7264 required_kernelcore);
7266 /* Continue if range is now fully accounted */
7267 if (end_pfn <= usable_startpfn) {
7270 * Push zone_movable_pfn to the end so
7271 * that if we have to rebalance
7272 * kernelcore across nodes, we will
7273 * not double account here
7275 zone_movable_pfn[nid] = end_pfn;
7278 start_pfn = usable_startpfn;
7282 * The usable PFN range for ZONE_MOVABLE is from
7283 * start_pfn->end_pfn. Calculate size_pages as the
7284 * number of pages used as kernelcore
7286 size_pages = end_pfn - start_pfn;
7287 if (size_pages > kernelcore_remaining)
7288 size_pages = kernelcore_remaining;
7289 zone_movable_pfn[nid] = start_pfn + size_pages;
7292 * Some kernelcore has been met, update counts and
7293 * break if the kernelcore for this node has been
7296 required_kernelcore -= min(required_kernelcore,
7298 kernelcore_remaining -= size_pages;
7299 if (!kernelcore_remaining)
7305 * If there is still required_kernelcore, we do another pass with one
7306 * less node in the count. This will push zone_movable_pfn[nid] further
7307 * along on the nodes that still have memory until kernelcore is
7311 if (usable_nodes && required_kernelcore > usable_nodes)
7315 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7316 for (nid = 0; nid < MAX_NUMNODES; nid++)
7317 zone_movable_pfn[nid] =
7318 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7321 /* restore the node_state */
7322 node_states[N_MEMORY] = saved_node_state;
7325 /* Any regular or high memory on that node ? */
7326 static void check_for_memory(pg_data_t *pgdat, int nid)
7328 enum zone_type zone_type;
7330 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7331 struct zone *zone = &pgdat->node_zones[zone_type];
7332 if (populated_zone(zone)) {
7333 if (IS_ENABLED(CONFIG_HIGHMEM))
7334 node_set_state(nid, N_HIGH_MEMORY);
7335 if (zone_type <= ZONE_NORMAL)
7336 node_set_state(nid, N_NORMAL_MEMORY);
7343 * free_area_init_nodes - Initialise all pg_data_t and zone data
7344 * @max_zone_pfn: an array of max PFNs for each zone
7346 * This will call free_area_init_node() for each active node in the system.
7347 * Using the page ranges provided by memblock_set_node(), the size of each
7348 * zone in each node and their holes is calculated. If the maximum PFN
7349 * between two adjacent zones match, it is assumed that the zone is empty.
7350 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7351 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7352 * starts where the previous one ended. For example, ZONE_DMA32 starts
7353 * at arch_max_dma_pfn.
7355 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7357 unsigned long start_pfn, end_pfn;
7360 /* Record where the zone boundaries are */
7361 memset(arch_zone_lowest_possible_pfn, 0,
7362 sizeof(arch_zone_lowest_possible_pfn));
7363 memset(arch_zone_highest_possible_pfn, 0,
7364 sizeof(arch_zone_highest_possible_pfn));
7366 start_pfn = find_min_pfn_with_active_regions();
7368 for (i = 0; i < MAX_NR_ZONES; i++) {
7369 if (i == ZONE_MOVABLE)
7372 end_pfn = max(max_zone_pfn[i], start_pfn);
7373 arch_zone_lowest_possible_pfn[i] = start_pfn;
7374 arch_zone_highest_possible_pfn[i] = end_pfn;
7376 start_pfn = end_pfn;
7379 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7380 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7381 find_zone_movable_pfns_for_nodes();
7383 /* Print out the zone ranges */
7384 pr_info("Zone ranges:\n");
7385 for (i = 0; i < MAX_NR_ZONES; i++) {
7386 if (i == ZONE_MOVABLE)
7388 pr_info(" %-8s ", zone_names[i]);
7389 if (arch_zone_lowest_possible_pfn[i] ==
7390 arch_zone_highest_possible_pfn[i])
7393 pr_cont("[mem %#018Lx-%#018Lx]\n",
7394 (u64)arch_zone_lowest_possible_pfn[i]
7396 ((u64)arch_zone_highest_possible_pfn[i]
7397 << PAGE_SHIFT) - 1);
7400 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7401 pr_info("Movable zone start for each node\n");
7402 for (i = 0; i < MAX_NUMNODES; i++) {
7403 if (zone_movable_pfn[i])
7404 pr_info(" Node %d: %#018Lx\n", i,
7405 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7409 * Print out the early node map, and initialize the
7410 * subsection-map relative to active online memory ranges to
7411 * enable future "sub-section" extensions of the memory map.
7413 pr_info("Early memory node ranges\n");
7414 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7415 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7416 (u64)start_pfn << PAGE_SHIFT,
7417 ((u64)end_pfn << PAGE_SHIFT) - 1);
7418 subsection_map_init(start_pfn, end_pfn - start_pfn);
7421 /* Initialise every node */
7422 mminit_verify_pageflags_layout();
7423 setup_nr_node_ids();
7424 init_unavailable_mem();
7425 for_each_online_node(nid) {
7426 pg_data_t *pgdat = NODE_DATA(nid);
7427 free_area_init_node(nid, NULL,
7428 find_min_pfn_for_node(nid), NULL);
7430 /* Any memory on that node */
7431 if (pgdat->node_present_pages)
7432 node_set_state(nid, N_MEMORY);
7433 check_for_memory(pgdat, nid);
7437 static int __init cmdline_parse_core(char *p, unsigned long *core,
7438 unsigned long *percent)
7440 unsigned long long coremem;
7446 /* Value may be a percentage of total memory, otherwise bytes */
7447 coremem = simple_strtoull(p, &endptr, 0);
7448 if (*endptr == '%') {
7449 /* Paranoid check for percent values greater than 100 */
7450 WARN_ON(coremem > 100);
7454 coremem = memparse(p, &p);
7455 /* Paranoid check that UL is enough for the coremem value */
7456 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7458 *core = coremem >> PAGE_SHIFT;
7465 * kernelcore=size sets the amount of memory for use for allocations that
7466 * cannot be reclaimed or migrated.
7468 static int __init cmdline_parse_kernelcore(char *p)
7470 /* parse kernelcore=mirror */
7471 if (parse_option_str(p, "mirror")) {
7472 mirrored_kernelcore = true;
7476 return cmdline_parse_core(p, &required_kernelcore,
7477 &required_kernelcore_percent);
7481 * movablecore=size sets the amount of memory for use for allocations that
7482 * can be reclaimed or migrated.
7484 static int __init cmdline_parse_movablecore(char *p)
7486 return cmdline_parse_core(p, &required_movablecore,
7487 &required_movablecore_percent);
7490 early_param("kernelcore", cmdline_parse_kernelcore);
7491 early_param("movablecore", cmdline_parse_movablecore);
7493 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7495 void adjust_managed_page_count(struct page *page, long count)
7497 atomic_long_add(count, &page_zone(page)->managed_pages);
7498 totalram_pages_add(count);
7499 #ifdef CONFIG_HIGHMEM
7500 if (PageHighMem(page))
7501 totalhigh_pages_add(count);
7504 EXPORT_SYMBOL(adjust_managed_page_count);
7506 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7509 unsigned long pages = 0;
7511 start = (void *)PAGE_ALIGN((unsigned long)start);
7512 end = (void *)((unsigned long)end & PAGE_MASK);
7513 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7514 struct page *page = virt_to_page(pos);
7515 void *direct_map_addr;
7518 * 'direct_map_addr' might be different from 'pos'
7519 * because some architectures' virt_to_page()
7520 * work with aliases. Getting the direct map
7521 * address ensures that we get a _writeable_
7522 * alias for the memset().
7524 direct_map_addr = page_address(page);
7525 if ((unsigned int)poison <= 0xFF)
7526 memset(direct_map_addr, poison, PAGE_SIZE);
7528 free_reserved_page(page);
7532 pr_info("Freeing %s memory: %ldK\n",
7533 s, pages << (PAGE_SHIFT - 10));
7538 #ifdef CONFIG_HIGHMEM
7539 void free_highmem_page(struct page *page)
7541 __free_reserved_page(page);
7542 totalram_pages_inc();
7543 atomic_long_inc(&page_zone(page)->managed_pages);
7544 totalhigh_pages_inc();
7549 void __init mem_init_print_info(const char *str)
7551 unsigned long physpages, codesize, datasize, rosize, bss_size;
7552 unsigned long init_code_size, init_data_size;
7554 physpages = get_num_physpages();
7555 codesize = _etext - _stext;
7556 datasize = _edata - _sdata;
7557 rosize = __end_rodata - __start_rodata;
7558 bss_size = __bss_stop - __bss_start;
7559 init_data_size = __init_end - __init_begin;
7560 init_code_size = _einittext - _sinittext;
7563 * Detect special cases and adjust section sizes accordingly:
7564 * 1) .init.* may be embedded into .data sections
7565 * 2) .init.text.* may be out of [__init_begin, __init_end],
7566 * please refer to arch/tile/kernel/vmlinux.lds.S.
7567 * 3) .rodata.* may be embedded into .text or .data sections.
7569 #define adj_init_size(start, end, size, pos, adj) \
7571 if (start <= pos && pos < end && size > adj) \
7575 adj_init_size(__init_begin, __init_end, init_data_size,
7576 _sinittext, init_code_size);
7577 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7578 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7579 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7580 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7582 #undef adj_init_size
7584 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7585 #ifdef CONFIG_HIGHMEM
7589 nr_free_pages() << (PAGE_SHIFT - 10),
7590 physpages << (PAGE_SHIFT - 10),
7591 codesize >> 10, datasize >> 10, rosize >> 10,
7592 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7593 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7594 totalcma_pages << (PAGE_SHIFT - 10),
7595 #ifdef CONFIG_HIGHMEM
7596 totalhigh_pages() << (PAGE_SHIFT - 10),
7598 str ? ", " : "", str ? str : "");
7602 * set_dma_reserve - set the specified number of pages reserved in the first zone
7603 * @new_dma_reserve: The number of pages to mark reserved
7605 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7606 * In the DMA zone, a significant percentage may be consumed by kernel image
7607 * and other unfreeable allocations which can skew the watermarks badly. This
7608 * function may optionally be used to account for unfreeable pages in the
7609 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7610 * smaller per-cpu batchsize.
7612 void __init set_dma_reserve(unsigned long new_dma_reserve)
7614 dma_reserve = new_dma_reserve;
7617 void __init free_area_init(unsigned long *zones_size)
7619 init_unavailable_mem();
7620 free_area_init_node(0, zones_size,
7621 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7624 static int page_alloc_cpu_dead(unsigned int cpu)
7627 lru_add_drain_cpu(cpu);
7631 * Spill the event counters of the dead processor
7632 * into the current processors event counters.
7633 * This artificially elevates the count of the current
7636 vm_events_fold_cpu(cpu);
7639 * Zero the differential counters of the dead processor
7640 * so that the vm statistics are consistent.
7642 * This is only okay since the processor is dead and cannot
7643 * race with what we are doing.
7645 cpu_vm_stats_fold(cpu);
7650 int hashdist = HASHDIST_DEFAULT;
7652 static int __init set_hashdist(char *str)
7656 hashdist = simple_strtoul(str, &str, 0);
7659 __setup("hashdist=", set_hashdist);
7662 void __init page_alloc_init(void)
7667 if (num_node_state(N_MEMORY) == 1)
7671 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7672 "mm/page_alloc:dead", NULL,
7673 page_alloc_cpu_dead);
7678 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7679 * or min_free_kbytes changes.
7681 static void calculate_totalreserve_pages(void)
7683 struct pglist_data *pgdat;
7684 unsigned long reserve_pages = 0;
7685 enum zone_type i, j;
7687 for_each_online_pgdat(pgdat) {
7689 pgdat->totalreserve_pages = 0;
7691 for (i = 0; i < MAX_NR_ZONES; i++) {
7692 struct zone *zone = pgdat->node_zones + i;
7694 unsigned long managed_pages = zone_managed_pages(zone);
7696 /* Find valid and maximum lowmem_reserve in the zone */
7697 for (j = i; j < MAX_NR_ZONES; j++) {
7698 if (zone->lowmem_reserve[j] > max)
7699 max = zone->lowmem_reserve[j];
7702 /* we treat the high watermark as reserved pages. */
7703 max += high_wmark_pages(zone);
7705 if (max > managed_pages)
7706 max = managed_pages;
7708 pgdat->totalreserve_pages += max;
7710 reserve_pages += max;
7713 totalreserve_pages = reserve_pages;
7717 * setup_per_zone_lowmem_reserve - called whenever
7718 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7719 * has a correct pages reserved value, so an adequate number of
7720 * pages are left in the zone after a successful __alloc_pages().
7722 static void setup_per_zone_lowmem_reserve(void)
7724 struct pglist_data *pgdat;
7725 enum zone_type j, idx;
7727 for_each_online_pgdat(pgdat) {
7728 for (j = 0; j < MAX_NR_ZONES; j++) {
7729 struct zone *zone = pgdat->node_zones + j;
7730 unsigned long managed_pages = zone_managed_pages(zone);
7732 zone->lowmem_reserve[j] = 0;
7736 struct zone *lower_zone;
7739 lower_zone = pgdat->node_zones + idx;
7741 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7742 sysctl_lowmem_reserve_ratio[idx] = 0;
7743 lower_zone->lowmem_reserve[j] = 0;
7745 lower_zone->lowmem_reserve[j] =
7746 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7748 managed_pages += zone_managed_pages(lower_zone);
7753 /* update totalreserve_pages */
7754 calculate_totalreserve_pages();
7757 static void __setup_per_zone_wmarks(void)
7759 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7760 unsigned long lowmem_pages = 0;
7762 unsigned long flags;
7764 /* Calculate total number of !ZONE_HIGHMEM pages */
7765 for_each_zone(zone) {
7766 if (!is_highmem(zone))
7767 lowmem_pages += zone_managed_pages(zone);
7770 for_each_zone(zone) {
7773 spin_lock_irqsave(&zone->lock, flags);
7774 tmp = (u64)pages_min * zone_managed_pages(zone);
7775 do_div(tmp, lowmem_pages);
7776 if (is_highmem(zone)) {
7778 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7779 * need highmem pages, so cap pages_min to a small
7782 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7783 * deltas control async page reclaim, and so should
7784 * not be capped for highmem.
7786 unsigned long min_pages;
7788 min_pages = zone_managed_pages(zone) / 1024;
7789 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7790 zone->_watermark[WMARK_MIN] = min_pages;
7793 * If it's a lowmem zone, reserve a number of pages
7794 * proportionate to the zone's size.
7796 zone->_watermark[WMARK_MIN] = tmp;
7800 * Set the kswapd watermarks distance according to the
7801 * scale factor in proportion to available memory, but
7802 * ensure a minimum size on small systems.
7804 tmp = max_t(u64, tmp >> 2,
7805 mult_frac(zone_managed_pages(zone),
7806 watermark_scale_factor, 10000));
7808 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7809 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7810 zone->watermark_boost = 0;
7812 spin_unlock_irqrestore(&zone->lock, flags);
7815 /* update totalreserve_pages */
7816 calculate_totalreserve_pages();
7820 * setup_per_zone_wmarks - called when min_free_kbytes changes
7821 * or when memory is hot-{added|removed}
7823 * Ensures that the watermark[min,low,high] values for each zone are set
7824 * correctly with respect to min_free_kbytes.
7826 void setup_per_zone_wmarks(void)
7828 static DEFINE_SPINLOCK(lock);
7831 __setup_per_zone_wmarks();
7836 * Initialise min_free_kbytes.
7838 * For small machines we want it small (128k min). For large machines
7839 * we want it large (64MB max). But it is not linear, because network
7840 * bandwidth does not increase linearly with machine size. We use
7842 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7843 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7859 int __meminit init_per_zone_wmark_min(void)
7861 unsigned long lowmem_kbytes;
7862 int new_min_free_kbytes;
7864 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7865 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7867 if (new_min_free_kbytes > user_min_free_kbytes) {
7868 min_free_kbytes = new_min_free_kbytes;
7869 if (min_free_kbytes < 128)
7870 min_free_kbytes = 128;
7871 if (min_free_kbytes > 65536)
7872 min_free_kbytes = 65536;
7874 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7875 new_min_free_kbytes, user_min_free_kbytes);
7877 setup_per_zone_wmarks();
7878 refresh_zone_stat_thresholds();
7879 setup_per_zone_lowmem_reserve();
7882 setup_min_unmapped_ratio();
7883 setup_min_slab_ratio();
7888 core_initcall(init_per_zone_wmark_min)
7891 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7892 * that we can call two helper functions whenever min_free_kbytes
7895 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7896 void __user *buffer, size_t *length, loff_t *ppos)
7900 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7905 user_min_free_kbytes = min_free_kbytes;
7906 setup_per_zone_wmarks();
7911 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7912 void __user *buffer, size_t *length, loff_t *ppos)
7916 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7923 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7924 void __user *buffer, size_t *length, loff_t *ppos)
7928 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7933 setup_per_zone_wmarks();
7939 static void setup_min_unmapped_ratio(void)
7944 for_each_online_pgdat(pgdat)
7945 pgdat->min_unmapped_pages = 0;
7948 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7949 sysctl_min_unmapped_ratio) / 100;
7953 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7954 void __user *buffer, size_t *length, loff_t *ppos)
7958 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7962 setup_min_unmapped_ratio();
7967 static void setup_min_slab_ratio(void)
7972 for_each_online_pgdat(pgdat)
7973 pgdat->min_slab_pages = 0;
7976 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7977 sysctl_min_slab_ratio) / 100;
7980 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7981 void __user *buffer, size_t *length, loff_t *ppos)
7985 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7989 setup_min_slab_ratio();
7996 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7997 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7998 * whenever sysctl_lowmem_reserve_ratio changes.
8000 * The reserve ratio obviously has absolutely no relation with the
8001 * minimum watermarks. The lowmem reserve ratio can only make sense
8002 * if in function of the boot time zone sizes.
8004 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8005 void __user *buffer, size_t *length, loff_t *ppos)
8007 proc_dointvec_minmax(table, write, buffer, length, ppos);
8008 setup_per_zone_lowmem_reserve();
8012 static void __zone_pcp_update(struct zone *zone)
8016 for_each_possible_cpu(cpu)
8017 pageset_set_high_and_batch(zone,
8018 per_cpu_ptr(zone->pageset, cpu));
8022 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8023 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8024 * pagelist can have before it gets flushed back to buddy allocator.
8026 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8027 void __user *buffer, size_t *length, loff_t *ppos)
8030 int old_percpu_pagelist_fraction;
8033 mutex_lock(&pcp_batch_high_lock);
8034 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8036 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8037 if (!write || ret < 0)
8040 /* Sanity checking to avoid pcp imbalance */
8041 if (percpu_pagelist_fraction &&
8042 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8043 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8049 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8052 for_each_populated_zone(zone)
8053 __zone_pcp_update(zone);
8055 mutex_unlock(&pcp_batch_high_lock);
8059 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8061 * Returns the number of pages that arch has reserved but
8062 * is not known to alloc_large_system_hash().
8064 static unsigned long __init arch_reserved_kernel_pages(void)
8071 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8072 * machines. As memory size is increased the scale is also increased but at
8073 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8074 * quadruples the scale is increased by one, which means the size of hash table
8075 * only doubles, instead of quadrupling as well.
8076 * Because 32-bit systems cannot have large physical memory, where this scaling
8077 * makes sense, it is disabled on such platforms.
8079 #if __BITS_PER_LONG > 32
8080 #define ADAPT_SCALE_BASE (64ul << 30)
8081 #define ADAPT_SCALE_SHIFT 2
8082 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8086 * allocate a large system hash table from bootmem
8087 * - it is assumed that the hash table must contain an exact power-of-2
8088 * quantity of entries
8089 * - limit is the number of hash buckets, not the total allocation size
8091 void *__init alloc_large_system_hash(const char *tablename,
8092 unsigned long bucketsize,
8093 unsigned long numentries,
8096 unsigned int *_hash_shift,
8097 unsigned int *_hash_mask,
8098 unsigned long low_limit,
8099 unsigned long high_limit)
8101 unsigned long long max = high_limit;
8102 unsigned long log2qty, size;
8107 /* allow the kernel cmdline to have a say */
8109 /* round applicable memory size up to nearest megabyte */
8110 numentries = nr_kernel_pages;
8111 numentries -= arch_reserved_kernel_pages();
8113 /* It isn't necessary when PAGE_SIZE >= 1MB */
8114 if (PAGE_SHIFT < 20)
8115 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8117 #if __BITS_PER_LONG > 32
8119 unsigned long adapt;
8121 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8122 adapt <<= ADAPT_SCALE_SHIFT)
8127 /* limit to 1 bucket per 2^scale bytes of low memory */
8128 if (scale > PAGE_SHIFT)
8129 numentries >>= (scale - PAGE_SHIFT);
8131 numentries <<= (PAGE_SHIFT - scale);
8133 /* Make sure we've got at least a 0-order allocation.. */
8134 if (unlikely(flags & HASH_SMALL)) {
8135 /* Makes no sense without HASH_EARLY */
8136 WARN_ON(!(flags & HASH_EARLY));
8137 if (!(numentries >> *_hash_shift)) {
8138 numentries = 1UL << *_hash_shift;
8139 BUG_ON(!numentries);
8141 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8142 numentries = PAGE_SIZE / bucketsize;
8144 numentries = roundup_pow_of_two(numentries);
8146 /* limit allocation size to 1/16 total memory by default */
8148 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8149 do_div(max, bucketsize);
8151 max = min(max, 0x80000000ULL);
8153 if (numentries < low_limit)
8154 numentries = low_limit;
8155 if (numentries > max)
8158 log2qty = ilog2(numentries);
8160 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8163 size = bucketsize << log2qty;
8164 if (flags & HASH_EARLY) {
8165 if (flags & HASH_ZERO)
8166 table = memblock_alloc(size, SMP_CACHE_BYTES);
8168 table = memblock_alloc_raw(size,
8170 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8171 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8175 * If bucketsize is not a power-of-two, we may free
8176 * some pages at the end of hash table which
8177 * alloc_pages_exact() automatically does
8179 table = alloc_pages_exact(size, gfp_flags);
8180 kmemleak_alloc(table, size, 1, gfp_flags);
8182 } while (!table && size > PAGE_SIZE && --log2qty);
8185 panic("Failed to allocate %s hash table\n", tablename);
8187 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8188 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8189 virt ? "vmalloc" : "linear");
8192 *_hash_shift = log2qty;
8194 *_hash_mask = (1 << log2qty) - 1;
8200 * This function checks whether pageblock includes unmovable pages or not.
8202 * PageLRU check without isolation or lru_lock could race so that
8203 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8204 * check without lock_page also may miss some movable non-lru pages at
8205 * race condition. So you can't expect this function should be exact.
8207 * Returns a page without holding a reference. If the caller wants to
8208 * dereference that page (e.g., dumping), it has to make sure that that it
8209 * cannot get removed (e.g., via memory unplug) concurrently.
8212 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8213 int migratetype, int flags)
8215 unsigned long iter = 0;
8216 unsigned long pfn = page_to_pfn(page);
8219 * TODO we could make this much more efficient by not checking every
8220 * page in the range if we know all of them are in MOVABLE_ZONE and
8221 * that the movable zone guarantees that pages are migratable but
8222 * the later is not the case right now unfortunatelly. E.g. movablecore
8223 * can still lead to having bootmem allocations in zone_movable.
8226 if (is_migrate_cma_page(page)) {
8228 * CMA allocations (alloc_contig_range) really need to mark
8229 * isolate CMA pageblocks even when they are not movable in fact
8230 * so consider them movable here.
8232 if (is_migrate_cma(migratetype))
8238 for (; iter < pageblock_nr_pages; iter++) {
8239 if (!pfn_valid_within(pfn + iter))
8242 page = pfn_to_page(pfn + iter);
8244 if (PageReserved(page))
8248 * If the zone is movable and we have ruled out all reserved
8249 * pages then it should be reasonably safe to assume the rest
8252 if (zone_idx(zone) == ZONE_MOVABLE)
8256 * Hugepages are not in LRU lists, but they're movable.
8257 * We need not scan over tail pages because we don't
8258 * handle each tail page individually in migration.
8260 if (PageHuge(page)) {
8261 struct page *head = compound_head(page);
8262 unsigned int skip_pages;
8264 if (!hugepage_migration_supported(page_hstate(head)))
8267 skip_pages = compound_nr(head) - (page - head);
8268 iter += skip_pages - 1;
8273 * We can't use page_count without pin a page
8274 * because another CPU can free compound page.
8275 * This check already skips compound tails of THP
8276 * because their page->_refcount is zero at all time.
8278 if (!page_ref_count(page)) {
8279 if (PageBuddy(page))
8280 iter += (1 << page_order(page)) - 1;
8285 * The HWPoisoned page may be not in buddy system, and
8286 * page_count() is not 0.
8288 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8291 if (__PageMovable(page) || PageLRU(page))
8295 * If there are RECLAIMABLE pages, we need to check
8296 * it. But now, memory offline itself doesn't call
8297 * shrink_node_slabs() and it still to be fixed.
8300 * If the page is not RAM, page_count()should be 0.
8301 * we don't need more check. This is an _used_ not-movable page.
8303 * The problematic thing here is PG_reserved pages. PG_reserved
8304 * is set to both of a memory hole page and a _used_ kernel
8312 #ifdef CONFIG_CONTIG_ALLOC
8313 static unsigned long pfn_max_align_down(unsigned long pfn)
8315 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8316 pageblock_nr_pages) - 1);
8319 static unsigned long pfn_max_align_up(unsigned long pfn)
8321 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8322 pageblock_nr_pages));
8325 /* [start, end) must belong to a single zone. */
8326 static int __alloc_contig_migrate_range(struct compact_control *cc,
8327 unsigned long start, unsigned long end)
8329 /* This function is based on compact_zone() from compaction.c. */
8330 unsigned long nr_reclaimed;
8331 unsigned long pfn = start;
8332 unsigned int tries = 0;
8337 while (pfn < end || !list_empty(&cc->migratepages)) {
8338 if (fatal_signal_pending(current)) {
8343 if (list_empty(&cc->migratepages)) {
8344 cc->nr_migratepages = 0;
8345 pfn = isolate_migratepages_range(cc, pfn, end);
8351 } else if (++tries == 5) {
8352 ret = ret < 0 ? ret : -EBUSY;
8356 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8358 cc->nr_migratepages -= nr_reclaimed;
8360 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8361 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8364 putback_movable_pages(&cc->migratepages);
8371 * alloc_contig_range() -- tries to allocate given range of pages
8372 * @start: start PFN to allocate
8373 * @end: one-past-the-last PFN to allocate
8374 * @migratetype: migratetype of the underlaying pageblocks (either
8375 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8376 * in range must have the same migratetype and it must
8377 * be either of the two.
8378 * @gfp_mask: GFP mask to use during compaction
8380 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8381 * aligned. The PFN range must belong to a single zone.
8383 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8384 * pageblocks in the range. Once isolated, the pageblocks should not
8385 * be modified by others.
8387 * Return: zero on success or negative error code. On success all
8388 * pages which PFN is in [start, end) are allocated for the caller and
8389 * need to be freed with free_contig_range().
8391 int alloc_contig_range(unsigned long start, unsigned long end,
8392 unsigned migratetype, gfp_t gfp_mask)
8394 unsigned long outer_start, outer_end;
8398 struct compact_control cc = {
8399 .nr_migratepages = 0,
8401 .zone = page_zone(pfn_to_page(start)),
8402 .mode = MIGRATE_SYNC,
8403 .ignore_skip_hint = true,
8404 .no_set_skip_hint = true,
8405 .gfp_mask = current_gfp_context(gfp_mask),
8407 INIT_LIST_HEAD(&cc.migratepages);
8410 * What we do here is we mark all pageblocks in range as
8411 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8412 * have different sizes, and due to the way page allocator
8413 * work, we align the range to biggest of the two pages so
8414 * that page allocator won't try to merge buddies from
8415 * different pageblocks and change MIGRATE_ISOLATE to some
8416 * other migration type.
8418 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8419 * migrate the pages from an unaligned range (ie. pages that
8420 * we are interested in). This will put all the pages in
8421 * range back to page allocator as MIGRATE_ISOLATE.
8423 * When this is done, we take the pages in range from page
8424 * allocator removing them from the buddy system. This way
8425 * page allocator will never consider using them.
8427 * This lets us mark the pageblocks back as
8428 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8429 * aligned range but not in the unaligned, original range are
8430 * put back to page allocator so that buddy can use them.
8433 ret = start_isolate_page_range(pfn_max_align_down(start),
8434 pfn_max_align_up(end), migratetype, 0);
8439 * In case of -EBUSY, we'd like to know which page causes problem.
8440 * So, just fall through. test_pages_isolated() has a tracepoint
8441 * which will report the busy page.
8443 * It is possible that busy pages could become available before
8444 * the call to test_pages_isolated, and the range will actually be
8445 * allocated. So, if we fall through be sure to clear ret so that
8446 * -EBUSY is not accidentally used or returned to caller.
8448 ret = __alloc_contig_migrate_range(&cc, start, end);
8449 if (ret && ret != -EBUSY)
8454 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8455 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8456 * more, all pages in [start, end) are free in page allocator.
8457 * What we are going to do is to allocate all pages from
8458 * [start, end) (that is remove them from page allocator).
8460 * The only problem is that pages at the beginning and at the
8461 * end of interesting range may be not aligned with pages that
8462 * page allocator holds, ie. they can be part of higher order
8463 * pages. Because of this, we reserve the bigger range and
8464 * once this is done free the pages we are not interested in.
8466 * We don't have to hold zone->lock here because the pages are
8467 * isolated thus they won't get removed from buddy.
8470 lru_add_drain_all();
8473 outer_start = start;
8474 while (!PageBuddy(pfn_to_page(outer_start))) {
8475 if (++order >= MAX_ORDER) {
8476 outer_start = start;
8479 outer_start &= ~0UL << order;
8482 if (outer_start != start) {
8483 order = page_order(pfn_to_page(outer_start));
8486 * outer_start page could be small order buddy page and
8487 * it doesn't include start page. Adjust outer_start
8488 * in this case to report failed page properly
8489 * on tracepoint in test_pages_isolated()
8491 if (outer_start + (1UL << order) <= start)
8492 outer_start = start;
8495 /* Make sure the range is really isolated. */
8496 if (test_pages_isolated(outer_start, end, 0)) {
8497 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8498 __func__, outer_start, end);
8503 /* Grab isolated pages from freelists. */
8504 outer_end = isolate_freepages_range(&cc, outer_start, end);
8510 /* Free head and tail (if any) */
8511 if (start != outer_start)
8512 free_contig_range(outer_start, start - outer_start);
8513 if (end != outer_end)
8514 free_contig_range(end, outer_end - end);
8517 undo_isolate_page_range(pfn_max_align_down(start),
8518 pfn_max_align_up(end), migratetype);
8522 static int __alloc_contig_pages(unsigned long start_pfn,
8523 unsigned long nr_pages, gfp_t gfp_mask)
8525 unsigned long end_pfn = start_pfn + nr_pages;
8527 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8531 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8532 unsigned long nr_pages)
8534 unsigned long i, end_pfn = start_pfn + nr_pages;
8537 for (i = start_pfn; i < end_pfn; i++) {
8538 page = pfn_to_online_page(i);
8542 if (page_zone(page) != z)
8545 if (PageReserved(page))
8548 if (page_count(page) > 0)
8557 static bool zone_spans_last_pfn(const struct zone *zone,
8558 unsigned long start_pfn, unsigned long nr_pages)
8560 unsigned long last_pfn = start_pfn + nr_pages - 1;
8562 return zone_spans_pfn(zone, last_pfn);
8566 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8567 * @nr_pages: Number of contiguous pages to allocate
8568 * @gfp_mask: GFP mask to limit search and used during compaction
8570 * @nodemask: Mask for other possible nodes
8572 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8573 * on an applicable zonelist to find a contiguous pfn range which can then be
8574 * tried for allocation with alloc_contig_range(). This routine is intended
8575 * for allocation requests which can not be fulfilled with the buddy allocator.
8577 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8578 * power of two then the alignment is guaranteed to be to the given nr_pages
8579 * (e.g. 1GB request would be aligned to 1GB).
8581 * Allocated pages can be freed with free_contig_range() or by manually calling
8582 * __free_page() on each allocated page.
8584 * Return: pointer to contiguous pages on success, or NULL if not successful.
8586 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8587 int nid, nodemask_t *nodemask)
8589 unsigned long ret, pfn, flags;
8590 struct zonelist *zonelist;
8594 zonelist = node_zonelist(nid, gfp_mask);
8595 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8596 gfp_zone(gfp_mask), nodemask) {
8597 spin_lock_irqsave(&zone->lock, flags);
8599 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8600 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8601 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8603 * We release the zone lock here because
8604 * alloc_contig_range() will also lock the zone
8605 * at some point. If there's an allocation
8606 * spinning on this lock, it may win the race
8607 * and cause alloc_contig_range() to fail...
8609 spin_unlock_irqrestore(&zone->lock, flags);
8610 ret = __alloc_contig_pages(pfn, nr_pages,
8613 return pfn_to_page(pfn);
8614 spin_lock_irqsave(&zone->lock, flags);
8618 spin_unlock_irqrestore(&zone->lock, flags);
8622 #endif /* CONFIG_CONTIG_ALLOC */
8624 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8626 unsigned int count = 0;
8628 for (; nr_pages--; pfn++) {
8629 struct page *page = pfn_to_page(pfn);
8631 count += page_count(page) != 1;
8634 WARN(count != 0, "%d pages are still in use!\n", count);
8638 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8639 * page high values need to be recalulated.
8641 void __meminit zone_pcp_update(struct zone *zone)
8643 mutex_lock(&pcp_batch_high_lock);
8644 __zone_pcp_update(zone);
8645 mutex_unlock(&pcp_batch_high_lock);
8648 void zone_pcp_reset(struct zone *zone)
8650 unsigned long flags;
8652 struct per_cpu_pageset *pset;
8654 /* avoid races with drain_pages() */
8655 local_irq_save(flags);
8656 if (zone->pageset != &boot_pageset) {
8657 for_each_online_cpu(cpu) {
8658 pset = per_cpu_ptr(zone->pageset, cpu);
8659 drain_zonestat(zone, pset);
8661 free_percpu(zone->pageset);
8662 zone->pageset = &boot_pageset;
8664 local_irq_restore(flags);
8667 #ifdef CONFIG_MEMORY_HOTREMOVE
8669 * All pages in the range must be in a single zone and isolated
8670 * before calling this.
8673 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8679 unsigned long flags;
8680 unsigned long offlined_pages = 0;
8682 /* find the first valid pfn */
8683 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8687 return offlined_pages;
8689 offline_mem_sections(pfn, end_pfn);
8690 zone = page_zone(pfn_to_page(pfn));
8691 spin_lock_irqsave(&zone->lock, flags);
8693 while (pfn < end_pfn) {
8694 if (!pfn_valid(pfn)) {
8698 page = pfn_to_page(pfn);
8700 * The HWPoisoned page may be not in buddy system, and
8701 * page_count() is not 0.
8703 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8709 BUG_ON(page_count(page));
8710 BUG_ON(!PageBuddy(page));
8711 order = page_order(page);
8712 offlined_pages += 1 << order;
8713 del_page_from_free_area(page, &zone->free_area[order]);
8714 pfn += (1 << order);
8716 spin_unlock_irqrestore(&zone->lock, flags);
8718 return offlined_pages;
8722 bool is_free_buddy_page(struct page *page)
8724 struct zone *zone = page_zone(page);
8725 unsigned long pfn = page_to_pfn(page);
8726 unsigned long flags;
8729 spin_lock_irqsave(&zone->lock, flags);
8730 for (order = 0; order < MAX_ORDER; order++) {
8731 struct page *page_head = page - (pfn & ((1 << order) - 1));
8733 if (PageBuddy(page_head) && page_order(page_head) >= order)
8736 spin_unlock_irqrestore(&zone->lock, flags);
8738 return order < MAX_ORDER;
8741 #ifdef CONFIG_MEMORY_FAILURE
8743 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8744 * test is performed under the zone lock to prevent a race against page
8747 bool set_hwpoison_free_buddy_page(struct page *page)
8749 struct zone *zone = page_zone(page);
8750 unsigned long pfn = page_to_pfn(page);
8751 unsigned long flags;
8753 bool hwpoisoned = false;
8755 spin_lock_irqsave(&zone->lock, flags);
8756 for (order = 0; order < MAX_ORDER; order++) {
8757 struct page *page_head = page - (pfn & ((1 << order) - 1));
8759 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8760 if (!TestSetPageHWPoison(page))
8765 spin_unlock_irqrestore(&zone->lock, flags);