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>
71 #include <linux/padata.h>
73 #include <asm/sections.h>
74 #include <asm/tlbflush.h>
75 #include <asm/div64.h>
78 #include "page_reporting.h"
80 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
81 static DEFINE_MUTEX(pcp_batch_high_lock);
82 #define MIN_PERCPU_PAGELIST_FRACTION (8)
84 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
85 DEFINE_PER_CPU(int, numa_node);
86 EXPORT_PER_CPU_SYMBOL(numa_node);
89 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
91 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
93 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
94 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
95 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
96 * defined in <linux/topology.h>.
98 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
99 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
102 /* work_structs for global per-cpu drains */
105 struct work_struct work;
107 static DEFINE_MUTEX(pcpu_drain_mutex);
108 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
110 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
111 volatile unsigned long latent_entropy __latent_entropy;
112 EXPORT_SYMBOL(latent_entropy);
116 * Array of node states.
118 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
119 [N_POSSIBLE] = NODE_MASK_ALL,
120 [N_ONLINE] = { { [0] = 1UL } },
122 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
123 #ifdef CONFIG_HIGHMEM
124 [N_HIGH_MEMORY] = { { [0] = 1UL } },
126 [N_MEMORY] = { { [0] = 1UL } },
127 [N_CPU] = { { [0] = 1UL } },
130 EXPORT_SYMBOL(node_states);
132 atomic_long_t _totalram_pages __read_mostly;
133 EXPORT_SYMBOL(_totalram_pages);
134 unsigned long totalreserve_pages __read_mostly;
135 unsigned long totalcma_pages __read_mostly;
137 int percpu_pagelist_fraction;
138 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
139 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
140 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
142 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
144 EXPORT_SYMBOL(init_on_alloc);
146 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
147 DEFINE_STATIC_KEY_TRUE(init_on_free);
149 DEFINE_STATIC_KEY_FALSE(init_on_free);
151 EXPORT_SYMBOL(init_on_free);
153 static int __init early_init_on_alloc(char *buf)
160 ret = kstrtobool(buf, &bool_result);
161 if (bool_result && page_poisoning_enabled())
162 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
164 static_branch_enable(&init_on_alloc);
166 static_branch_disable(&init_on_alloc);
169 early_param("init_on_alloc", early_init_on_alloc);
171 static int __init early_init_on_free(char *buf)
178 ret = kstrtobool(buf, &bool_result);
179 if (bool_result && page_poisoning_enabled())
180 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
182 static_branch_enable(&init_on_free);
184 static_branch_disable(&init_on_free);
187 early_param("init_on_free", early_init_on_free);
190 * A cached value of the page's pageblock's migratetype, used when the page is
191 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
192 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
193 * Also the migratetype set in the page does not necessarily match the pcplist
194 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
195 * other index - this ensures that it will be put on the correct CMA freelist.
197 static inline int get_pcppage_migratetype(struct page *page)
202 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
204 page->index = migratetype;
207 #ifdef CONFIG_PM_SLEEP
209 * The following functions are used by the suspend/hibernate code to temporarily
210 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
211 * while devices are suspended. To avoid races with the suspend/hibernate code,
212 * they should always be called with system_transition_mutex held
213 * (gfp_allowed_mask also should only be modified with system_transition_mutex
214 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
215 * with that modification).
218 static gfp_t saved_gfp_mask;
220 void pm_restore_gfp_mask(void)
222 WARN_ON(!mutex_is_locked(&system_transition_mutex));
223 if (saved_gfp_mask) {
224 gfp_allowed_mask = saved_gfp_mask;
229 void pm_restrict_gfp_mask(void)
231 WARN_ON(!mutex_is_locked(&system_transition_mutex));
232 WARN_ON(saved_gfp_mask);
233 saved_gfp_mask = gfp_allowed_mask;
234 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
237 bool pm_suspended_storage(void)
239 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
243 #endif /* CONFIG_PM_SLEEP */
245 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
246 unsigned int pageblock_order __read_mostly;
249 static void __free_pages_ok(struct page *page, unsigned int order);
252 * results with 256, 32 in the lowmem_reserve sysctl:
253 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
254 * 1G machine -> (16M dma, 784M normal, 224M high)
255 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
256 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
257 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
259 * TBD: should special case ZONE_DMA32 machines here - in those we normally
260 * don't need any ZONE_NORMAL reservation
262 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
263 #ifdef CONFIG_ZONE_DMA
266 #ifdef CONFIG_ZONE_DMA32
270 #ifdef CONFIG_HIGHMEM
276 static char * const zone_names[MAX_NR_ZONES] = {
277 #ifdef CONFIG_ZONE_DMA
280 #ifdef CONFIG_ZONE_DMA32
284 #ifdef CONFIG_HIGHMEM
288 #ifdef CONFIG_ZONE_DEVICE
293 const char * const migratetype_names[MIGRATE_TYPES] = {
301 #ifdef CONFIG_MEMORY_ISOLATION
306 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
307 [NULL_COMPOUND_DTOR] = NULL,
308 [COMPOUND_PAGE_DTOR] = free_compound_page,
309 #ifdef CONFIG_HUGETLB_PAGE
310 [HUGETLB_PAGE_DTOR] = free_huge_page,
312 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
313 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
317 int min_free_kbytes = 1024;
318 int user_min_free_kbytes = -1;
319 #ifdef CONFIG_DISCONTIGMEM
321 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
322 * are not on separate NUMA nodes. Functionally this works but with
323 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
324 * quite small. By default, do not boost watermarks on discontigmem as in
325 * many cases very high-order allocations like THP are likely to be
326 * unsupported and the premature reclaim offsets the advantage of long-term
327 * fragmentation avoidance.
329 int watermark_boost_factor __read_mostly;
331 int watermark_boost_factor __read_mostly = 15000;
333 int watermark_scale_factor = 10;
335 static unsigned long nr_kernel_pages __initdata;
336 static unsigned long nr_all_pages __initdata;
337 static unsigned long dma_reserve __initdata;
339 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
340 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
341 static unsigned long required_kernelcore __initdata;
342 static unsigned long required_kernelcore_percent __initdata;
343 static unsigned long required_movablecore __initdata;
344 static unsigned long required_movablecore_percent __initdata;
345 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
346 static bool mirrored_kernelcore __meminitdata;
348 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
350 EXPORT_SYMBOL(movable_zone);
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)
613 static unsigned long resume;
614 static unsigned long nr_shown;
615 static unsigned long nr_unshown;
618 * Allow a burst of 60 reports, then keep quiet for that minute;
619 * or allow a steady drip of one report per second.
621 if (nr_shown == 60) {
622 if (time_before(jiffies, resume)) {
628 "BUG: Bad page state: %lu messages suppressed\n",
635 resume = jiffies + 60 * HZ;
637 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
638 current->comm, page_to_pfn(page));
639 __dump_page(page, reason);
640 dump_page_owner(page);
645 /* Leave bad fields for debug, except PageBuddy could make trouble */
646 page_mapcount_reset(page); /* remove PageBuddy */
647 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
651 * Higher-order pages are called "compound pages". They are structured thusly:
653 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
655 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
656 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
658 * The first tail page's ->compound_dtor holds the offset in array of compound
659 * page destructors. See compound_page_dtors.
661 * The first tail page's ->compound_order holds the order of allocation.
662 * This usage means that zero-order pages may not be compound.
665 void free_compound_page(struct page *page)
667 mem_cgroup_uncharge(page);
668 __free_pages_ok(page, compound_order(page));
671 void prep_compound_page(struct page *page, unsigned int order)
674 int nr_pages = 1 << order;
676 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
677 set_compound_order(page, order);
679 for (i = 1; i < nr_pages; i++) {
680 struct page *p = page + i;
681 set_page_count(p, 0);
682 p->mapping = TAIL_MAPPING;
683 set_compound_head(p, page);
685 atomic_set(compound_mapcount_ptr(page), -1);
686 if (hpage_pincount_available(page))
687 atomic_set(compound_pincount_ptr(page), 0);
690 #ifdef CONFIG_DEBUG_PAGEALLOC
691 unsigned int _debug_guardpage_minorder;
693 bool _debug_pagealloc_enabled_early __read_mostly
694 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
695 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
696 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
697 EXPORT_SYMBOL(_debug_pagealloc_enabled);
699 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
701 static int __init early_debug_pagealloc(char *buf)
703 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
705 early_param("debug_pagealloc", early_debug_pagealloc);
707 void init_debug_pagealloc(void)
709 if (!debug_pagealloc_enabled())
712 static_branch_enable(&_debug_pagealloc_enabled);
714 if (!debug_guardpage_minorder())
717 static_branch_enable(&_debug_guardpage_enabled);
720 static int __init debug_guardpage_minorder_setup(char *buf)
724 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
725 pr_err("Bad debug_guardpage_minorder value\n");
728 _debug_guardpage_minorder = res;
729 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
732 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
734 static inline bool set_page_guard(struct zone *zone, struct page *page,
735 unsigned int order, int migratetype)
737 if (!debug_guardpage_enabled())
740 if (order >= debug_guardpage_minorder())
743 __SetPageGuard(page);
744 INIT_LIST_HEAD(&page->lru);
745 set_page_private(page, order);
746 /* Guard pages are not available for any usage */
747 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
752 static inline void clear_page_guard(struct zone *zone, struct page *page,
753 unsigned int order, int migratetype)
755 if (!debug_guardpage_enabled())
758 __ClearPageGuard(page);
760 set_page_private(page, 0);
761 if (!is_migrate_isolate(migratetype))
762 __mod_zone_freepage_state(zone, (1 << order), migratetype);
765 static inline bool set_page_guard(struct zone *zone, struct page *page,
766 unsigned int order, int migratetype) { return false; }
767 static inline void clear_page_guard(struct zone *zone, struct page *page,
768 unsigned int order, int migratetype) {}
771 static inline void set_page_order(struct page *page, unsigned int order)
773 set_page_private(page, order);
774 __SetPageBuddy(page);
778 * This function checks whether a page is free && is the buddy
779 * we can coalesce a page and its buddy if
780 * (a) the buddy is not in a hole (check before calling!) &&
781 * (b) the buddy is in the buddy system &&
782 * (c) a page and its buddy have the same order &&
783 * (d) a page and its buddy are in the same zone.
785 * For recording whether a page is in the buddy system, we set PageBuddy.
786 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
788 * For recording page's order, we use page_private(page).
790 static inline bool page_is_buddy(struct page *page, struct page *buddy,
793 if (!page_is_guard(buddy) && !PageBuddy(buddy))
796 if (page_order(buddy) != order)
800 * zone check is done late to avoid uselessly calculating
801 * zone/node ids for pages that could never merge.
803 if (page_zone_id(page) != page_zone_id(buddy))
806 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
811 #ifdef CONFIG_COMPACTION
812 static inline struct capture_control *task_capc(struct zone *zone)
814 struct capture_control *capc = current->capture_control;
817 !(current->flags & PF_KTHREAD) &&
819 capc->cc->zone == zone &&
820 capc->cc->direct_compaction ? capc : NULL;
824 compaction_capture(struct capture_control *capc, struct page *page,
825 int order, int migratetype)
827 if (!capc || order != capc->cc->order)
830 /* Do not accidentally pollute CMA or isolated regions*/
831 if (is_migrate_cma(migratetype) ||
832 is_migrate_isolate(migratetype))
836 * Do not let lower order allocations polluate a movable pageblock.
837 * This might let an unmovable request use a reclaimable pageblock
838 * and vice-versa but no more than normal fallback logic which can
839 * have trouble finding a high-order free page.
841 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
849 static inline struct capture_control *task_capc(struct zone *zone)
855 compaction_capture(struct capture_control *capc, struct page *page,
856 int order, int migratetype)
860 #endif /* CONFIG_COMPACTION */
862 /* Used for pages not on another list */
863 static inline void add_to_free_list(struct page *page, struct zone *zone,
864 unsigned int order, int migratetype)
866 struct free_area *area = &zone->free_area[order];
868 list_add(&page->lru, &area->free_list[migratetype]);
872 /* Used for pages not on another list */
873 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
874 unsigned int order, int migratetype)
876 struct free_area *area = &zone->free_area[order];
878 list_add_tail(&page->lru, &area->free_list[migratetype]);
882 /* Used for pages which are on another list */
883 static inline void move_to_free_list(struct page *page, struct zone *zone,
884 unsigned int order, int migratetype)
886 struct free_area *area = &zone->free_area[order];
888 list_move(&page->lru, &area->free_list[migratetype]);
891 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
894 /* clear reported state and update reported page count */
895 if (page_reported(page))
896 __ClearPageReported(page);
898 list_del(&page->lru);
899 __ClearPageBuddy(page);
900 set_page_private(page, 0);
901 zone->free_area[order].nr_free--;
905 * If this is not the largest possible page, check if the buddy
906 * of the next-highest order is free. If it is, it's possible
907 * that pages are being freed that will coalesce soon. In case,
908 * that is happening, add the free page to the tail of the list
909 * so it's less likely to be used soon and more likely to be merged
910 * as a higher order page
913 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
914 struct page *page, unsigned int order)
916 struct page *higher_page, *higher_buddy;
917 unsigned long combined_pfn;
919 if (order >= MAX_ORDER - 2)
922 if (!pfn_valid_within(buddy_pfn))
925 combined_pfn = buddy_pfn & pfn;
926 higher_page = page + (combined_pfn - pfn);
927 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
928 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
930 return pfn_valid_within(buddy_pfn) &&
931 page_is_buddy(higher_page, higher_buddy, order + 1);
935 * Freeing function for a buddy system allocator.
937 * The concept of a buddy system is to maintain direct-mapped table
938 * (containing bit values) for memory blocks of various "orders".
939 * The bottom level table contains the map for the smallest allocatable
940 * units of memory (here, pages), and each level above it describes
941 * pairs of units from the levels below, hence, "buddies".
942 * At a high level, all that happens here is marking the table entry
943 * at the bottom level available, and propagating the changes upward
944 * as necessary, plus some accounting needed to play nicely with other
945 * parts of the VM system.
946 * At each level, we keep a list of pages, which are heads of continuous
947 * free pages of length of (1 << order) and marked with PageBuddy.
948 * Page's order is recorded in page_private(page) field.
949 * So when we are allocating or freeing one, we can derive the state of the
950 * other. That is, if we allocate a small block, and both were
951 * free, the remainder of the region must be split into blocks.
952 * If a block is freed, and its buddy is also free, then this
953 * triggers coalescing into a block of larger size.
958 static inline void __free_one_page(struct page *page,
960 struct zone *zone, unsigned int order,
961 int migratetype, bool report)
963 struct capture_control *capc = task_capc(zone);
964 unsigned long uninitialized_var(buddy_pfn);
965 unsigned long combined_pfn;
966 unsigned int max_order;
970 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
972 VM_BUG_ON(!zone_is_initialized(zone));
973 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
975 VM_BUG_ON(migratetype == -1);
976 if (likely(!is_migrate_isolate(migratetype)))
977 __mod_zone_freepage_state(zone, 1 << order, migratetype);
979 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
980 VM_BUG_ON_PAGE(bad_range(zone, page), page);
983 while (order < max_order - 1) {
984 if (compaction_capture(capc, page, order, migratetype)) {
985 __mod_zone_freepage_state(zone, -(1 << order),
989 buddy_pfn = __find_buddy_pfn(pfn, order);
990 buddy = page + (buddy_pfn - pfn);
992 if (!pfn_valid_within(buddy_pfn))
994 if (!page_is_buddy(page, buddy, order))
997 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
998 * merge with it and move up one order.
1000 if (page_is_guard(buddy))
1001 clear_page_guard(zone, buddy, order, migratetype);
1003 del_page_from_free_list(buddy, zone, order);
1004 combined_pfn = buddy_pfn & pfn;
1005 page = page + (combined_pfn - pfn);
1009 if (max_order < MAX_ORDER) {
1010 /* If we are here, it means order is >= pageblock_order.
1011 * We want to prevent merge between freepages on isolate
1012 * pageblock and normal pageblock. Without this, pageblock
1013 * isolation could cause incorrect freepage or CMA accounting.
1015 * We don't want to hit this code for the more frequent
1016 * low-order merging.
1018 if (unlikely(has_isolate_pageblock(zone))) {
1021 buddy_pfn = __find_buddy_pfn(pfn, order);
1022 buddy = page + (buddy_pfn - pfn);
1023 buddy_mt = get_pageblock_migratetype(buddy);
1025 if (migratetype != buddy_mt
1026 && (is_migrate_isolate(migratetype) ||
1027 is_migrate_isolate(buddy_mt)))
1031 goto continue_merging;
1035 set_page_order(page, order);
1037 if (is_shuffle_order(order))
1038 to_tail = shuffle_pick_tail();
1040 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1043 add_to_free_list_tail(page, zone, order, migratetype);
1045 add_to_free_list(page, zone, order, migratetype);
1047 /* Notify page reporting subsystem of freed page */
1049 page_reporting_notify_free(order);
1053 * A bad page could be due to a number of fields. Instead of multiple branches,
1054 * try and check multiple fields with one check. The caller must do a detailed
1055 * check if necessary.
1057 static inline bool page_expected_state(struct page *page,
1058 unsigned long check_flags)
1060 if (unlikely(atomic_read(&page->_mapcount) != -1))
1063 if (unlikely((unsigned long)page->mapping |
1064 page_ref_count(page) |
1066 (unsigned long)page->mem_cgroup |
1068 (page->flags & check_flags)))
1074 static const char *page_bad_reason(struct page *page, unsigned long flags)
1076 const char *bad_reason = NULL;
1078 if (unlikely(atomic_read(&page->_mapcount) != -1))
1079 bad_reason = "nonzero mapcount";
1080 if (unlikely(page->mapping != NULL))
1081 bad_reason = "non-NULL mapping";
1082 if (unlikely(page_ref_count(page) != 0))
1083 bad_reason = "nonzero _refcount";
1084 if (unlikely(page->flags & flags)) {
1085 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1086 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1088 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1091 if (unlikely(page->mem_cgroup))
1092 bad_reason = "page still charged to cgroup";
1097 static void check_free_page_bad(struct page *page)
1100 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1103 static inline int check_free_page(struct page *page)
1105 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1108 /* Something has gone sideways, find it */
1109 check_free_page_bad(page);
1113 static int free_tail_pages_check(struct page *head_page, struct page *page)
1118 * We rely page->lru.next never has bit 0 set, unless the page
1119 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1121 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1123 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1127 switch (page - head_page) {
1129 /* the first tail page: ->mapping may be compound_mapcount() */
1130 if (unlikely(compound_mapcount(page))) {
1131 bad_page(page, "nonzero compound_mapcount");
1137 * the second tail page: ->mapping is
1138 * deferred_list.next -- ignore value.
1142 if (page->mapping != TAIL_MAPPING) {
1143 bad_page(page, "corrupted mapping in tail page");
1148 if (unlikely(!PageTail(page))) {
1149 bad_page(page, "PageTail not set");
1152 if (unlikely(compound_head(page) != head_page)) {
1153 bad_page(page, "compound_head not consistent");
1158 page->mapping = NULL;
1159 clear_compound_head(page);
1163 static void kernel_init_free_pages(struct page *page, int numpages)
1167 for (i = 0; i < numpages; i++)
1168 clear_highpage(page + i);
1171 static __always_inline bool free_pages_prepare(struct page *page,
1172 unsigned int order, bool check_free)
1176 VM_BUG_ON_PAGE(PageTail(page), page);
1178 trace_mm_page_free(page, order);
1181 * Check tail pages before head page information is cleared to
1182 * avoid checking PageCompound for order-0 pages.
1184 if (unlikely(order)) {
1185 bool compound = PageCompound(page);
1188 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1191 ClearPageDoubleMap(page);
1192 for (i = 1; i < (1 << order); i++) {
1194 bad += free_tail_pages_check(page, page + i);
1195 if (unlikely(check_free_page(page + i))) {
1199 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1202 if (PageMappingFlags(page))
1203 page->mapping = NULL;
1204 if (memcg_kmem_enabled() && PageKmemcg(page))
1205 __memcg_kmem_uncharge_page(page, order);
1207 bad += check_free_page(page);
1211 page_cpupid_reset_last(page);
1212 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1213 reset_page_owner(page, order);
1215 if (!PageHighMem(page)) {
1216 debug_check_no_locks_freed(page_address(page),
1217 PAGE_SIZE << order);
1218 debug_check_no_obj_freed(page_address(page),
1219 PAGE_SIZE << order);
1221 if (want_init_on_free())
1222 kernel_init_free_pages(page, 1 << order);
1224 kernel_poison_pages(page, 1 << order, 0);
1226 * arch_free_page() can make the page's contents inaccessible. s390
1227 * does this. So nothing which can access the page's contents should
1228 * happen after this.
1230 arch_free_page(page, order);
1232 if (debug_pagealloc_enabled_static())
1233 kernel_map_pages(page, 1 << order, 0);
1235 kasan_free_nondeferred_pages(page, order);
1240 #ifdef CONFIG_DEBUG_VM
1242 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1243 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1244 * moved from pcp lists to free lists.
1246 static bool free_pcp_prepare(struct page *page)
1248 return free_pages_prepare(page, 0, true);
1251 static bool bulkfree_pcp_prepare(struct page *page)
1253 if (debug_pagealloc_enabled_static())
1254 return check_free_page(page);
1260 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1261 * moving from pcp lists to free list in order to reduce overhead. With
1262 * debug_pagealloc enabled, they are checked also immediately when being freed
1265 static bool free_pcp_prepare(struct page *page)
1267 if (debug_pagealloc_enabled_static())
1268 return free_pages_prepare(page, 0, true);
1270 return free_pages_prepare(page, 0, false);
1273 static bool bulkfree_pcp_prepare(struct page *page)
1275 return check_free_page(page);
1277 #endif /* CONFIG_DEBUG_VM */
1279 static inline void prefetch_buddy(struct page *page)
1281 unsigned long pfn = page_to_pfn(page);
1282 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1283 struct page *buddy = page + (buddy_pfn - pfn);
1289 * Frees a number of pages from the PCP lists
1290 * Assumes all pages on list are in same zone, and of same order.
1291 * count is the number of pages to free.
1293 * If the zone was previously in an "all pages pinned" state then look to
1294 * see if this freeing clears that state.
1296 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1297 * pinned" detection logic.
1299 static void free_pcppages_bulk(struct zone *zone, int count,
1300 struct per_cpu_pages *pcp)
1302 int migratetype = 0;
1304 int prefetch_nr = 0;
1305 bool isolated_pageblocks;
1306 struct page *page, *tmp;
1310 struct list_head *list;
1313 * Remove pages from lists in a round-robin fashion. A
1314 * batch_free count is maintained that is incremented when an
1315 * empty list is encountered. This is so more pages are freed
1316 * off fuller lists instead of spinning excessively around empty
1321 if (++migratetype == MIGRATE_PCPTYPES)
1323 list = &pcp->lists[migratetype];
1324 } while (list_empty(list));
1326 /* This is the only non-empty list. Free them all. */
1327 if (batch_free == MIGRATE_PCPTYPES)
1331 page = list_last_entry(list, struct page, lru);
1332 /* must delete to avoid corrupting pcp list */
1333 list_del(&page->lru);
1336 if (bulkfree_pcp_prepare(page))
1339 list_add_tail(&page->lru, &head);
1342 * We are going to put the page back to the global
1343 * pool, prefetch its buddy to speed up later access
1344 * under zone->lock. It is believed the overhead of
1345 * an additional test and calculating buddy_pfn here
1346 * can be offset by reduced memory latency later. To
1347 * avoid excessive prefetching due to large count, only
1348 * prefetch buddy for the first pcp->batch nr of pages.
1350 if (prefetch_nr++ < pcp->batch)
1351 prefetch_buddy(page);
1352 } while (--count && --batch_free && !list_empty(list));
1355 spin_lock(&zone->lock);
1356 isolated_pageblocks = has_isolate_pageblock(zone);
1359 * Use safe version since after __free_one_page(),
1360 * page->lru.next will not point to original list.
1362 list_for_each_entry_safe(page, tmp, &head, lru) {
1363 int mt = get_pcppage_migratetype(page);
1364 /* MIGRATE_ISOLATE page should not go to pcplists */
1365 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1366 /* Pageblock could have been isolated meanwhile */
1367 if (unlikely(isolated_pageblocks))
1368 mt = get_pageblock_migratetype(page);
1370 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1371 trace_mm_page_pcpu_drain(page, 0, mt);
1373 spin_unlock(&zone->lock);
1376 static void free_one_page(struct zone *zone,
1377 struct page *page, unsigned long pfn,
1381 spin_lock(&zone->lock);
1382 if (unlikely(has_isolate_pageblock(zone) ||
1383 is_migrate_isolate(migratetype))) {
1384 migratetype = get_pfnblock_migratetype(page, pfn);
1386 __free_one_page(page, pfn, zone, order, migratetype, true);
1387 spin_unlock(&zone->lock);
1390 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1391 unsigned long zone, int nid)
1393 mm_zero_struct_page(page);
1394 set_page_links(page, zone, nid, pfn);
1395 init_page_count(page);
1396 page_mapcount_reset(page);
1397 page_cpupid_reset_last(page);
1398 page_kasan_tag_reset(page);
1400 INIT_LIST_HEAD(&page->lru);
1401 #ifdef WANT_PAGE_VIRTUAL
1402 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1403 if (!is_highmem_idx(zone))
1404 set_page_address(page, __va(pfn << PAGE_SHIFT));
1408 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1409 static void __meminit init_reserved_page(unsigned long pfn)
1414 if (!early_page_uninitialised(pfn))
1417 nid = early_pfn_to_nid(pfn);
1418 pgdat = NODE_DATA(nid);
1420 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1421 struct zone *zone = &pgdat->node_zones[zid];
1423 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1426 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1429 static inline void init_reserved_page(unsigned long pfn)
1432 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1435 * Initialised pages do not have PageReserved set. This function is
1436 * called for each range allocated by the bootmem allocator and
1437 * marks the pages PageReserved. The remaining valid pages are later
1438 * sent to the buddy page allocator.
1440 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1442 unsigned long start_pfn = PFN_DOWN(start);
1443 unsigned long end_pfn = PFN_UP(end);
1445 for (; start_pfn < end_pfn; start_pfn++) {
1446 if (pfn_valid(start_pfn)) {
1447 struct page *page = pfn_to_page(start_pfn);
1449 init_reserved_page(start_pfn);
1451 /* Avoid false-positive PageTail() */
1452 INIT_LIST_HEAD(&page->lru);
1455 * no need for atomic set_bit because the struct
1456 * page is not visible yet so nobody should
1459 __SetPageReserved(page);
1464 static void __free_pages_ok(struct page *page, unsigned int order)
1466 unsigned long flags;
1468 unsigned long pfn = page_to_pfn(page);
1470 if (!free_pages_prepare(page, order, true))
1473 migratetype = get_pfnblock_migratetype(page, pfn);
1474 local_irq_save(flags);
1475 __count_vm_events(PGFREE, 1 << order);
1476 free_one_page(page_zone(page), page, pfn, order, migratetype);
1477 local_irq_restore(flags);
1480 void __free_pages_core(struct page *page, unsigned int order)
1482 unsigned int nr_pages = 1 << order;
1483 struct page *p = page;
1487 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1489 __ClearPageReserved(p);
1490 set_page_count(p, 0);
1492 __ClearPageReserved(p);
1493 set_page_count(p, 0);
1495 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1496 set_page_refcounted(page);
1497 __free_pages(page, order);
1500 #ifdef CONFIG_NEED_MULTIPLE_NODES
1502 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1504 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1507 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1509 int __meminit __early_pfn_to_nid(unsigned long pfn,
1510 struct mminit_pfnnid_cache *state)
1512 unsigned long start_pfn, end_pfn;
1515 if (state->last_start <= pfn && pfn < state->last_end)
1516 return state->last_nid;
1518 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1519 if (nid != NUMA_NO_NODE) {
1520 state->last_start = start_pfn;
1521 state->last_end = end_pfn;
1522 state->last_nid = nid;
1527 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1529 int __meminit early_pfn_to_nid(unsigned long pfn)
1531 static DEFINE_SPINLOCK(early_pfn_lock);
1534 spin_lock(&early_pfn_lock);
1535 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1537 nid = first_online_node;
1538 spin_unlock(&early_pfn_lock);
1542 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1544 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1547 if (early_page_uninitialised(pfn))
1549 __free_pages_core(page, order);
1553 * Check that the whole (or subset of) a pageblock given by the interval of
1554 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1555 * with the migration of free compaction scanner. The scanners then need to
1556 * use only pfn_valid_within() check for arches that allow holes within
1559 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1561 * It's possible on some configurations to have a setup like node0 node1 node0
1562 * i.e. it's possible that all pages within a zones range of pages do not
1563 * belong to a single zone. We assume that a border between node0 and node1
1564 * can occur within a single pageblock, but not a node0 node1 node0
1565 * interleaving within a single pageblock. It is therefore sufficient to check
1566 * the first and last page of a pageblock and avoid checking each individual
1567 * page in a pageblock.
1569 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1570 unsigned long end_pfn, struct zone *zone)
1572 struct page *start_page;
1573 struct page *end_page;
1575 /* end_pfn is one past the range we are checking */
1578 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1581 start_page = pfn_to_online_page(start_pfn);
1585 if (page_zone(start_page) != zone)
1588 end_page = pfn_to_page(end_pfn);
1590 /* This gives a shorter code than deriving page_zone(end_page) */
1591 if (page_zone_id(start_page) != page_zone_id(end_page))
1597 void set_zone_contiguous(struct zone *zone)
1599 unsigned long block_start_pfn = zone->zone_start_pfn;
1600 unsigned long block_end_pfn;
1602 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1603 for (; block_start_pfn < zone_end_pfn(zone);
1604 block_start_pfn = block_end_pfn,
1605 block_end_pfn += pageblock_nr_pages) {
1607 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1609 if (!__pageblock_pfn_to_page(block_start_pfn,
1610 block_end_pfn, zone))
1615 /* We confirm that there is no hole */
1616 zone->contiguous = true;
1619 void clear_zone_contiguous(struct zone *zone)
1621 zone->contiguous = false;
1624 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1625 static void __init deferred_free_range(unsigned long pfn,
1626 unsigned long nr_pages)
1634 page = pfn_to_page(pfn);
1636 /* Free a large naturally-aligned chunk if possible */
1637 if (nr_pages == pageblock_nr_pages &&
1638 (pfn & (pageblock_nr_pages - 1)) == 0) {
1639 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1640 __free_pages_core(page, pageblock_order);
1644 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1645 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1646 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1647 __free_pages_core(page, 0);
1651 /* Completion tracking for deferred_init_memmap() threads */
1652 static atomic_t pgdat_init_n_undone __initdata;
1653 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1655 static inline void __init pgdat_init_report_one_done(void)
1657 if (atomic_dec_and_test(&pgdat_init_n_undone))
1658 complete(&pgdat_init_all_done_comp);
1662 * Returns true if page needs to be initialized or freed to buddy allocator.
1664 * First we check if pfn is valid on architectures where it is possible to have
1665 * holes within pageblock_nr_pages. On systems where it is not possible, this
1666 * function is optimized out.
1668 * Then, we check if a current large page is valid by only checking the validity
1671 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1673 if (!pfn_valid_within(pfn))
1675 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1681 * Free pages to buddy allocator. Try to free aligned pages in
1682 * pageblock_nr_pages sizes.
1684 static void __init deferred_free_pages(unsigned long pfn,
1685 unsigned long end_pfn)
1687 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1688 unsigned long nr_free = 0;
1690 for (; pfn < end_pfn; pfn++) {
1691 if (!deferred_pfn_valid(pfn)) {
1692 deferred_free_range(pfn - nr_free, nr_free);
1694 } else if (!(pfn & nr_pgmask)) {
1695 deferred_free_range(pfn - nr_free, nr_free);
1701 /* Free the last block of pages to allocator */
1702 deferred_free_range(pfn - nr_free, nr_free);
1706 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1707 * by performing it only once every pageblock_nr_pages.
1708 * Return number of pages initialized.
1710 static unsigned long __init deferred_init_pages(struct zone *zone,
1712 unsigned long end_pfn)
1714 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1715 int nid = zone_to_nid(zone);
1716 unsigned long nr_pages = 0;
1717 int zid = zone_idx(zone);
1718 struct page *page = NULL;
1720 for (; pfn < end_pfn; pfn++) {
1721 if (!deferred_pfn_valid(pfn)) {
1724 } else if (!page || !(pfn & nr_pgmask)) {
1725 page = pfn_to_page(pfn);
1729 __init_single_page(page, pfn, zid, nid);
1736 * This function is meant to pre-load the iterator for the zone init.
1737 * Specifically it walks through the ranges until we are caught up to the
1738 * first_init_pfn value and exits there. If we never encounter the value we
1739 * return false indicating there are no valid ranges left.
1742 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1743 unsigned long *spfn, unsigned long *epfn,
1744 unsigned long first_init_pfn)
1749 * Start out by walking through the ranges in this zone that have
1750 * already been initialized. We don't need to do anything with them
1751 * so we just need to flush them out of the system.
1753 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1754 if (*epfn <= first_init_pfn)
1756 if (*spfn < first_init_pfn)
1757 *spfn = first_init_pfn;
1766 * Initialize and free pages. We do it in two loops: first we initialize
1767 * struct page, then free to buddy allocator, because while we are
1768 * freeing pages we can access pages that are ahead (computing buddy
1769 * page in __free_one_page()).
1771 * In order to try and keep some memory in the cache we have the loop
1772 * broken along max page order boundaries. This way we will not cause
1773 * any issues with the buddy page computation.
1775 static unsigned long __init
1776 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1777 unsigned long *end_pfn)
1779 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1780 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1781 unsigned long nr_pages = 0;
1784 /* First we loop through and initialize the page values */
1785 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1788 if (mo_pfn <= *start_pfn)
1791 t = min(mo_pfn, *end_pfn);
1792 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1794 if (mo_pfn < *end_pfn) {
1795 *start_pfn = mo_pfn;
1800 /* Reset values and now loop through freeing pages as needed */
1803 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1809 t = min(mo_pfn, epfn);
1810 deferred_free_pages(spfn, t);
1820 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1823 unsigned long spfn, epfn;
1824 struct zone *zone = arg;
1827 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1830 * Initialize and free pages in MAX_ORDER sized increments so that we
1831 * can avoid introducing any issues with the buddy allocator.
1833 while (spfn < end_pfn) {
1834 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1839 /* Initialise remaining memory on a node */
1840 static int __init deferred_init_memmap(void *data)
1842 pg_data_t *pgdat = data;
1843 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1844 unsigned long spfn = 0, epfn = 0;
1845 unsigned long first_init_pfn, flags;
1846 unsigned long start = jiffies;
1848 int zid, max_threads;
1851 /* Bind memory initialisation thread to a local node if possible */
1852 if (!cpumask_empty(cpumask))
1853 set_cpus_allowed_ptr(current, cpumask);
1855 pgdat_resize_lock(pgdat, &flags);
1856 first_init_pfn = pgdat->first_deferred_pfn;
1857 if (first_init_pfn == ULONG_MAX) {
1858 pgdat_resize_unlock(pgdat, &flags);
1859 pgdat_init_report_one_done();
1863 /* Sanity check boundaries */
1864 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1865 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1866 pgdat->first_deferred_pfn = ULONG_MAX;
1869 * Once we unlock here, the zone cannot be grown anymore, thus if an
1870 * interrupt thread must allocate this early in boot, zone must be
1871 * pre-grown prior to start of deferred page initialization.
1873 pgdat_resize_unlock(pgdat, &flags);
1875 /* Only the highest zone is deferred so find it */
1876 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1877 zone = pgdat->node_zones + zid;
1878 if (first_init_pfn < zone_end_pfn(zone))
1882 /* If the zone is empty somebody else may have cleared out the zone */
1883 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1888 * More CPUs always led to greater speedups on tested systems, up to
1889 * all the nodes' CPUs. Use all since the system is otherwise idle now.
1891 max_threads = max(cpumask_weight(cpumask), 1u);
1893 while (spfn < epfn) {
1894 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1895 struct padata_mt_job job = {
1896 .thread_fn = deferred_init_memmap_chunk,
1899 .size = epfn_align - spfn,
1900 .align = PAGES_PER_SECTION,
1901 .min_chunk = PAGES_PER_SECTION,
1902 .max_threads = max_threads,
1905 padata_do_multithreaded(&job);
1906 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1910 /* Sanity check that the next zone really is unpopulated */
1911 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1913 pr_info("node %d deferred pages initialised in %ums\n",
1914 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1916 pgdat_init_report_one_done();
1921 * If this zone has deferred pages, try to grow it by initializing enough
1922 * deferred pages to satisfy the allocation specified by order, rounded up to
1923 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1924 * of SECTION_SIZE bytes by initializing struct pages in increments of
1925 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1927 * Return true when zone was grown, otherwise return false. We return true even
1928 * when we grow less than requested, to let the caller decide if there are
1929 * enough pages to satisfy the allocation.
1931 * Note: We use noinline because this function is needed only during boot, and
1932 * it is called from a __ref function _deferred_grow_zone. This way we are
1933 * making sure that it is not inlined into permanent text section.
1935 static noinline bool __init
1936 deferred_grow_zone(struct zone *zone, unsigned int order)
1938 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1939 pg_data_t *pgdat = zone->zone_pgdat;
1940 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1941 unsigned long spfn, epfn, flags;
1942 unsigned long nr_pages = 0;
1945 /* Only the last zone may have deferred pages */
1946 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1949 pgdat_resize_lock(pgdat, &flags);
1952 * If someone grew this zone while we were waiting for spinlock, return
1953 * true, as there might be enough pages already.
1955 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1956 pgdat_resize_unlock(pgdat, &flags);
1960 /* If the zone is empty somebody else may have cleared out the zone */
1961 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1962 first_deferred_pfn)) {
1963 pgdat->first_deferred_pfn = ULONG_MAX;
1964 pgdat_resize_unlock(pgdat, &flags);
1965 /* Retry only once. */
1966 return first_deferred_pfn != ULONG_MAX;
1970 * Initialize and free pages in MAX_ORDER sized increments so
1971 * that we can avoid introducing any issues with the buddy
1974 while (spfn < epfn) {
1975 /* update our first deferred PFN for this section */
1976 first_deferred_pfn = spfn;
1978 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1979 touch_nmi_watchdog();
1981 /* We should only stop along section boundaries */
1982 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1985 /* If our quota has been met we can stop here */
1986 if (nr_pages >= nr_pages_needed)
1990 pgdat->first_deferred_pfn = spfn;
1991 pgdat_resize_unlock(pgdat, &flags);
1993 return nr_pages > 0;
1997 * deferred_grow_zone() is __init, but it is called from
1998 * get_page_from_freelist() during early boot until deferred_pages permanently
1999 * disables this call. This is why we have refdata wrapper to avoid warning,
2000 * and to ensure that the function body gets unloaded.
2003 _deferred_grow_zone(struct zone *zone, unsigned int order)
2005 return deferred_grow_zone(zone, order);
2008 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2010 void __init page_alloc_init_late(void)
2015 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2017 /* There will be num_node_state(N_MEMORY) threads */
2018 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2019 for_each_node_state(nid, N_MEMORY) {
2020 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2023 /* Block until all are initialised */
2024 wait_for_completion(&pgdat_init_all_done_comp);
2027 * The number of managed pages has changed due to the initialisation
2028 * so the pcpu batch and high limits needs to be updated or the limits
2029 * will be artificially small.
2031 for_each_populated_zone(zone)
2032 zone_pcp_update(zone);
2035 * We initialized the rest of the deferred pages. Permanently disable
2036 * on-demand struct page initialization.
2038 static_branch_disable(&deferred_pages);
2040 /* Reinit limits that are based on free pages after the kernel is up */
2041 files_maxfiles_init();
2044 /* Discard memblock private memory */
2047 for_each_node_state(nid, N_MEMORY)
2048 shuffle_free_memory(NODE_DATA(nid));
2050 for_each_populated_zone(zone)
2051 set_zone_contiguous(zone);
2055 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2056 void __init init_cma_reserved_pageblock(struct page *page)
2058 unsigned i = pageblock_nr_pages;
2059 struct page *p = page;
2062 __ClearPageReserved(p);
2063 set_page_count(p, 0);
2066 set_pageblock_migratetype(page, MIGRATE_CMA);
2068 if (pageblock_order >= MAX_ORDER) {
2069 i = pageblock_nr_pages;
2072 set_page_refcounted(p);
2073 __free_pages(p, MAX_ORDER - 1);
2074 p += MAX_ORDER_NR_PAGES;
2075 } while (i -= MAX_ORDER_NR_PAGES);
2077 set_page_refcounted(page);
2078 __free_pages(page, pageblock_order);
2081 adjust_managed_page_count(page, pageblock_nr_pages);
2086 * The order of subdivision here is critical for the IO subsystem.
2087 * Please do not alter this order without good reasons and regression
2088 * testing. Specifically, as large blocks of memory are subdivided,
2089 * the order in which smaller blocks are delivered depends on the order
2090 * they're subdivided in this function. This is the primary factor
2091 * influencing the order in which pages are delivered to the IO
2092 * subsystem according to empirical testing, and this is also justified
2093 * by considering the behavior of a buddy system containing a single
2094 * large block of memory acted on by a series of small allocations.
2095 * This behavior is a critical factor in sglist merging's success.
2099 static inline void expand(struct zone *zone, struct page *page,
2100 int low, int high, int migratetype)
2102 unsigned long size = 1 << high;
2104 while (high > low) {
2107 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2110 * Mark as guard pages (or page), that will allow to
2111 * merge back to allocator when buddy will be freed.
2112 * Corresponding page table entries will not be touched,
2113 * pages will stay not present in virtual address space
2115 if (set_page_guard(zone, &page[size], high, migratetype))
2118 add_to_free_list(&page[size], zone, high, migratetype);
2119 set_page_order(&page[size], high);
2123 static void check_new_page_bad(struct page *page)
2125 if (unlikely(page->flags & __PG_HWPOISON)) {
2126 /* Don't complain about hwpoisoned pages */
2127 page_mapcount_reset(page); /* remove PageBuddy */
2132 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2136 * This page is about to be returned from the page allocator
2138 static inline int check_new_page(struct page *page)
2140 if (likely(page_expected_state(page,
2141 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2144 check_new_page_bad(page);
2148 static inline bool free_pages_prezeroed(void)
2150 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2151 page_poisoning_enabled()) || want_init_on_free();
2154 #ifdef CONFIG_DEBUG_VM
2156 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2157 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2158 * also checked when pcp lists are refilled from the free lists.
2160 static inline bool check_pcp_refill(struct page *page)
2162 if (debug_pagealloc_enabled_static())
2163 return check_new_page(page);
2168 static inline bool check_new_pcp(struct page *page)
2170 return check_new_page(page);
2174 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2175 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2176 * enabled, they are also checked when being allocated from the pcp lists.
2178 static inline bool check_pcp_refill(struct page *page)
2180 return check_new_page(page);
2182 static inline bool check_new_pcp(struct page *page)
2184 if (debug_pagealloc_enabled_static())
2185 return check_new_page(page);
2189 #endif /* CONFIG_DEBUG_VM */
2191 static bool check_new_pages(struct page *page, unsigned int order)
2194 for (i = 0; i < (1 << order); i++) {
2195 struct page *p = page + i;
2197 if (unlikely(check_new_page(p)))
2204 inline void post_alloc_hook(struct page *page, unsigned int order,
2207 set_page_private(page, 0);
2208 set_page_refcounted(page);
2210 arch_alloc_page(page, order);
2211 if (debug_pagealloc_enabled_static())
2212 kernel_map_pages(page, 1 << order, 1);
2213 kasan_alloc_pages(page, order);
2214 kernel_poison_pages(page, 1 << order, 1);
2215 set_page_owner(page, order, gfp_flags);
2218 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2219 unsigned int alloc_flags)
2221 post_alloc_hook(page, order, gfp_flags);
2223 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2224 kernel_init_free_pages(page, 1 << order);
2226 if (order && (gfp_flags & __GFP_COMP))
2227 prep_compound_page(page, order);
2230 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2231 * allocate the page. The expectation is that the caller is taking
2232 * steps that will free more memory. The caller should avoid the page
2233 * being used for !PFMEMALLOC purposes.
2235 if (alloc_flags & ALLOC_NO_WATERMARKS)
2236 set_page_pfmemalloc(page);
2238 clear_page_pfmemalloc(page);
2242 * Go through the free lists for the given migratetype and remove
2243 * the smallest available page from the freelists
2245 static __always_inline
2246 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2249 unsigned int current_order;
2250 struct free_area *area;
2253 /* Find a page of the appropriate size in the preferred list */
2254 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2255 area = &(zone->free_area[current_order]);
2256 page = get_page_from_free_area(area, migratetype);
2259 del_page_from_free_list(page, zone, current_order);
2260 expand(zone, page, order, current_order, migratetype);
2261 set_pcppage_migratetype(page, migratetype);
2270 * This array describes the order lists are fallen back to when
2271 * the free lists for the desirable migrate type are depleted
2273 static int fallbacks[MIGRATE_TYPES][4] = {
2274 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2275 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2276 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2278 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2280 #ifdef CONFIG_MEMORY_ISOLATION
2281 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2286 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2289 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2292 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2293 unsigned int order) { return NULL; }
2297 * Move the free pages in a range to the free lists of the requested type.
2298 * Note that start_page and end_pages are not aligned on a pageblock
2299 * boundary. If alignment is required, use move_freepages_block()
2301 static int move_freepages(struct zone *zone,
2302 struct page *start_page, struct page *end_page,
2303 int migratetype, int *num_movable)
2307 int pages_moved = 0;
2309 for (page = start_page; page <= end_page;) {
2310 if (!pfn_valid_within(page_to_pfn(page))) {
2315 if (!PageBuddy(page)) {
2317 * We assume that pages that could be isolated for
2318 * migration are movable. But we don't actually try
2319 * isolating, as that would be expensive.
2322 (PageLRU(page) || __PageMovable(page)))
2329 /* Make sure we are not inadvertently changing nodes */
2330 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2331 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2333 order = page_order(page);
2334 move_to_free_list(page, zone, order, migratetype);
2336 pages_moved += 1 << order;
2342 int move_freepages_block(struct zone *zone, struct page *page,
2343 int migratetype, int *num_movable)
2345 unsigned long start_pfn, end_pfn;
2346 struct page *start_page, *end_page;
2351 start_pfn = page_to_pfn(page);
2352 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2353 start_page = pfn_to_page(start_pfn);
2354 end_page = start_page + pageblock_nr_pages - 1;
2355 end_pfn = start_pfn + pageblock_nr_pages - 1;
2357 /* Do not cross zone boundaries */
2358 if (!zone_spans_pfn(zone, start_pfn))
2360 if (!zone_spans_pfn(zone, end_pfn))
2363 return move_freepages(zone, start_page, end_page, migratetype,
2367 static void change_pageblock_range(struct page *pageblock_page,
2368 int start_order, int migratetype)
2370 int nr_pageblocks = 1 << (start_order - pageblock_order);
2372 while (nr_pageblocks--) {
2373 set_pageblock_migratetype(pageblock_page, migratetype);
2374 pageblock_page += pageblock_nr_pages;
2379 * When we are falling back to another migratetype during allocation, try to
2380 * steal extra free pages from the same pageblocks to satisfy further
2381 * allocations, instead of polluting multiple pageblocks.
2383 * If we are stealing a relatively large buddy page, it is likely there will
2384 * be more free pages in the pageblock, so try to steal them all. For
2385 * reclaimable and unmovable allocations, we steal regardless of page size,
2386 * as fragmentation caused by those allocations polluting movable pageblocks
2387 * is worse than movable allocations stealing from unmovable and reclaimable
2390 static bool can_steal_fallback(unsigned int order, int start_mt)
2393 * Leaving this order check is intended, although there is
2394 * relaxed order check in next check. The reason is that
2395 * we can actually steal whole pageblock if this condition met,
2396 * but, below check doesn't guarantee it and that is just heuristic
2397 * so could be changed anytime.
2399 if (order >= pageblock_order)
2402 if (order >= pageblock_order / 2 ||
2403 start_mt == MIGRATE_RECLAIMABLE ||
2404 start_mt == MIGRATE_UNMOVABLE ||
2405 page_group_by_mobility_disabled)
2411 static inline void boost_watermark(struct zone *zone)
2413 unsigned long max_boost;
2415 if (!watermark_boost_factor)
2418 * Don't bother in zones that are unlikely to produce results.
2419 * On small machines, including kdump capture kernels running
2420 * in a small area, boosting the watermark can cause an out of
2421 * memory situation immediately.
2423 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2426 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2427 watermark_boost_factor, 10000);
2430 * high watermark may be uninitialised if fragmentation occurs
2431 * very early in boot so do not boost. We do not fall
2432 * through and boost by pageblock_nr_pages as failing
2433 * allocations that early means that reclaim is not going
2434 * to help and it may even be impossible to reclaim the
2435 * boosted watermark resulting in a hang.
2440 max_boost = max(pageblock_nr_pages, max_boost);
2442 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2447 * This function implements actual steal behaviour. If order is large enough,
2448 * we can steal whole pageblock. If not, we first move freepages in this
2449 * pageblock to our migratetype and determine how many already-allocated pages
2450 * are there in the pageblock with a compatible migratetype. If at least half
2451 * of pages are free or compatible, we can change migratetype of the pageblock
2452 * itself, so pages freed in the future will be put on the correct free list.
2454 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2455 unsigned int alloc_flags, int start_type, bool whole_block)
2457 unsigned int current_order = page_order(page);
2458 int free_pages, movable_pages, alike_pages;
2461 old_block_type = get_pageblock_migratetype(page);
2464 * This can happen due to races and we want to prevent broken
2465 * highatomic accounting.
2467 if (is_migrate_highatomic(old_block_type))
2470 /* Take ownership for orders >= pageblock_order */
2471 if (current_order >= pageblock_order) {
2472 change_pageblock_range(page, current_order, start_type);
2477 * Boost watermarks to increase reclaim pressure to reduce the
2478 * likelihood of future fallbacks. Wake kswapd now as the node
2479 * may be balanced overall and kswapd will not wake naturally.
2481 boost_watermark(zone);
2482 if (alloc_flags & ALLOC_KSWAPD)
2483 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2485 /* We are not allowed to try stealing from the whole block */
2489 free_pages = move_freepages_block(zone, page, start_type,
2492 * Determine how many pages are compatible with our allocation.
2493 * For movable allocation, it's the number of movable pages which
2494 * we just obtained. For other types it's a bit more tricky.
2496 if (start_type == MIGRATE_MOVABLE) {
2497 alike_pages = movable_pages;
2500 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2501 * to MOVABLE pageblock, consider all non-movable pages as
2502 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2503 * vice versa, be conservative since we can't distinguish the
2504 * exact migratetype of non-movable pages.
2506 if (old_block_type == MIGRATE_MOVABLE)
2507 alike_pages = pageblock_nr_pages
2508 - (free_pages + movable_pages);
2513 /* moving whole block can fail due to zone boundary conditions */
2518 * If a sufficient number of pages in the block are either free or of
2519 * comparable migratability as our allocation, claim the whole block.
2521 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2522 page_group_by_mobility_disabled)
2523 set_pageblock_migratetype(page, start_type);
2528 move_to_free_list(page, zone, current_order, start_type);
2532 * Check whether there is a suitable fallback freepage with requested order.
2533 * If only_stealable is true, this function returns fallback_mt only if
2534 * we can steal other freepages all together. This would help to reduce
2535 * fragmentation due to mixed migratetype pages in one pageblock.
2537 int find_suitable_fallback(struct free_area *area, unsigned int order,
2538 int migratetype, bool only_stealable, bool *can_steal)
2543 if (area->nr_free == 0)
2548 fallback_mt = fallbacks[migratetype][i];
2549 if (fallback_mt == MIGRATE_TYPES)
2552 if (free_area_empty(area, fallback_mt))
2555 if (can_steal_fallback(order, migratetype))
2558 if (!only_stealable)
2569 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2570 * there are no empty page blocks that contain a page with a suitable order
2572 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2573 unsigned int alloc_order)
2576 unsigned long max_managed, flags;
2579 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2580 * Check is race-prone but harmless.
2582 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2583 if (zone->nr_reserved_highatomic >= max_managed)
2586 spin_lock_irqsave(&zone->lock, flags);
2588 /* Recheck the nr_reserved_highatomic limit under the lock */
2589 if (zone->nr_reserved_highatomic >= max_managed)
2593 mt = get_pageblock_migratetype(page);
2594 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2595 && !is_migrate_cma(mt)) {
2596 zone->nr_reserved_highatomic += pageblock_nr_pages;
2597 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2598 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2602 spin_unlock_irqrestore(&zone->lock, flags);
2606 * Used when an allocation is about to fail under memory pressure. This
2607 * potentially hurts the reliability of high-order allocations when under
2608 * intense memory pressure but failed atomic allocations should be easier
2609 * to recover from than an OOM.
2611 * If @force is true, try to unreserve a pageblock even though highatomic
2612 * pageblock is exhausted.
2614 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2617 struct zonelist *zonelist = ac->zonelist;
2618 unsigned long flags;
2625 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2628 * Preserve at least one pageblock unless memory pressure
2631 if (!force && zone->nr_reserved_highatomic <=
2635 spin_lock_irqsave(&zone->lock, flags);
2636 for (order = 0; order < MAX_ORDER; order++) {
2637 struct free_area *area = &(zone->free_area[order]);
2639 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2644 * In page freeing path, migratetype change is racy so
2645 * we can counter several free pages in a pageblock
2646 * in this loop althoug we changed the pageblock type
2647 * from highatomic to ac->migratetype. So we should
2648 * adjust the count once.
2650 if (is_migrate_highatomic_page(page)) {
2652 * It should never happen but changes to
2653 * locking could inadvertently allow a per-cpu
2654 * drain to add pages to MIGRATE_HIGHATOMIC
2655 * while unreserving so be safe and watch for
2658 zone->nr_reserved_highatomic -= min(
2660 zone->nr_reserved_highatomic);
2664 * Convert to ac->migratetype and avoid the normal
2665 * pageblock stealing heuristics. Minimally, the caller
2666 * is doing the work and needs the pages. More
2667 * importantly, if the block was always converted to
2668 * MIGRATE_UNMOVABLE or another type then the number
2669 * of pageblocks that cannot be completely freed
2672 set_pageblock_migratetype(page, ac->migratetype);
2673 ret = move_freepages_block(zone, page, ac->migratetype,
2676 spin_unlock_irqrestore(&zone->lock, flags);
2680 spin_unlock_irqrestore(&zone->lock, flags);
2687 * Try finding a free buddy page on the fallback list and put it on the free
2688 * list of requested migratetype, possibly along with other pages from the same
2689 * block, depending on fragmentation avoidance heuristics. Returns true if
2690 * fallback was found so that __rmqueue_smallest() can grab it.
2692 * The use of signed ints for order and current_order is a deliberate
2693 * deviation from the rest of this file, to make the for loop
2694 * condition simpler.
2696 static __always_inline bool
2697 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2698 unsigned int alloc_flags)
2700 struct free_area *area;
2702 int min_order = order;
2708 * Do not steal pages from freelists belonging to other pageblocks
2709 * i.e. orders < pageblock_order. If there are no local zones free,
2710 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2712 if (alloc_flags & ALLOC_NOFRAGMENT)
2713 min_order = pageblock_order;
2716 * Find the largest available free page in the other list. This roughly
2717 * approximates finding the pageblock with the most free pages, which
2718 * would be too costly to do exactly.
2720 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2722 area = &(zone->free_area[current_order]);
2723 fallback_mt = find_suitable_fallback(area, current_order,
2724 start_migratetype, false, &can_steal);
2725 if (fallback_mt == -1)
2729 * We cannot steal all free pages from the pageblock and the
2730 * requested migratetype is movable. In that case it's better to
2731 * steal and split the smallest available page instead of the
2732 * largest available page, because even if the next movable
2733 * allocation falls back into a different pageblock than this
2734 * one, it won't cause permanent fragmentation.
2736 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2737 && current_order > order)
2746 for (current_order = order; current_order < MAX_ORDER;
2748 area = &(zone->free_area[current_order]);
2749 fallback_mt = find_suitable_fallback(area, current_order,
2750 start_migratetype, false, &can_steal);
2751 if (fallback_mt != -1)
2756 * This should not happen - we already found a suitable fallback
2757 * when looking for the largest page.
2759 VM_BUG_ON(current_order == MAX_ORDER);
2762 page = get_page_from_free_area(area, fallback_mt);
2764 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2767 trace_mm_page_alloc_extfrag(page, order, current_order,
2768 start_migratetype, fallback_mt);
2775 * Do the hard work of removing an element from the buddy allocator.
2776 * Call me with the zone->lock already held.
2778 static __always_inline struct page *
2779 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2780 unsigned int alloc_flags)
2786 * Balance movable allocations between regular and CMA areas by
2787 * allocating from CMA when over half of the zone's free memory
2788 * is in the CMA area.
2790 if (migratetype == MIGRATE_MOVABLE &&
2791 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2792 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2793 page = __rmqueue_cma_fallback(zone, order);
2799 page = __rmqueue_smallest(zone, order, migratetype);
2800 if (unlikely(!page)) {
2801 if (migratetype == MIGRATE_MOVABLE)
2802 page = __rmqueue_cma_fallback(zone, order);
2804 if (!page && __rmqueue_fallback(zone, order, migratetype,
2809 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2814 * Obtain a specified number of elements from the buddy allocator, all under
2815 * a single hold of the lock, for efficiency. Add them to the supplied list.
2816 * Returns the number of new pages which were placed at *list.
2818 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2819 unsigned long count, struct list_head *list,
2820 int migratetype, unsigned int alloc_flags)
2824 spin_lock(&zone->lock);
2825 for (i = 0; i < count; ++i) {
2826 struct page *page = __rmqueue(zone, order, migratetype,
2828 if (unlikely(page == NULL))
2831 if (unlikely(check_pcp_refill(page)))
2835 * Split buddy pages returned by expand() are received here in
2836 * physical page order. The page is added to the tail of
2837 * caller's list. From the callers perspective, the linked list
2838 * is ordered by page number under some conditions. This is
2839 * useful for IO devices that can forward direction from the
2840 * head, thus also in the physical page order. This is useful
2841 * for IO devices that can merge IO requests if the physical
2842 * pages are ordered properly.
2844 list_add_tail(&page->lru, list);
2846 if (is_migrate_cma(get_pcppage_migratetype(page)))
2847 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2852 * i pages were removed from the buddy list even if some leak due
2853 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2854 * on i. Do not confuse with 'alloced' which is the number of
2855 * pages added to the pcp list.
2857 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2858 spin_unlock(&zone->lock);
2864 * Called from the vmstat counter updater to drain pagesets of this
2865 * currently executing processor on remote nodes after they have
2868 * Note that this function must be called with the thread pinned to
2869 * a single processor.
2871 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2873 unsigned long flags;
2874 int to_drain, batch;
2876 local_irq_save(flags);
2877 batch = READ_ONCE(pcp->batch);
2878 to_drain = min(pcp->count, batch);
2880 free_pcppages_bulk(zone, to_drain, pcp);
2881 local_irq_restore(flags);
2886 * Drain pcplists of the indicated processor and zone.
2888 * The processor must either be the current processor and the
2889 * thread pinned to the current processor or a processor that
2892 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2894 unsigned long flags;
2895 struct per_cpu_pageset *pset;
2896 struct per_cpu_pages *pcp;
2898 local_irq_save(flags);
2899 pset = per_cpu_ptr(zone->pageset, cpu);
2903 free_pcppages_bulk(zone, pcp->count, pcp);
2904 local_irq_restore(flags);
2908 * Drain pcplists of all zones on the indicated processor.
2910 * The processor must either be the current processor and the
2911 * thread pinned to the current processor or a processor that
2914 static void drain_pages(unsigned int cpu)
2918 for_each_populated_zone(zone) {
2919 drain_pages_zone(cpu, zone);
2924 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2926 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2927 * the single zone's pages.
2929 void drain_local_pages(struct zone *zone)
2931 int cpu = smp_processor_id();
2934 drain_pages_zone(cpu, zone);
2939 static void drain_local_pages_wq(struct work_struct *work)
2941 struct pcpu_drain *drain;
2943 drain = container_of(work, struct pcpu_drain, work);
2946 * drain_all_pages doesn't use proper cpu hotplug protection so
2947 * we can race with cpu offline when the WQ can move this from
2948 * a cpu pinned worker to an unbound one. We can operate on a different
2949 * cpu which is allright but we also have to make sure to not move to
2953 drain_local_pages(drain->zone);
2958 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2960 * When zone parameter is non-NULL, spill just the single zone's pages.
2962 * Note that this can be extremely slow as the draining happens in a workqueue.
2964 void drain_all_pages(struct zone *zone)
2969 * Allocate in the BSS so we wont require allocation in
2970 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2972 static cpumask_t cpus_with_pcps;
2975 * Make sure nobody triggers this path before mm_percpu_wq is fully
2978 if (WARN_ON_ONCE(!mm_percpu_wq))
2982 * Do not drain if one is already in progress unless it's specific to
2983 * a zone. Such callers are primarily CMA and memory hotplug and need
2984 * the drain to be complete when the call returns.
2986 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2989 mutex_lock(&pcpu_drain_mutex);
2993 * We don't care about racing with CPU hotplug event
2994 * as offline notification will cause the notified
2995 * cpu to drain that CPU pcps and on_each_cpu_mask
2996 * disables preemption as part of its processing
2998 for_each_online_cpu(cpu) {
2999 struct per_cpu_pageset *pcp;
3001 bool has_pcps = false;
3004 pcp = per_cpu_ptr(zone->pageset, cpu);
3008 for_each_populated_zone(z) {
3009 pcp = per_cpu_ptr(z->pageset, cpu);
3010 if (pcp->pcp.count) {
3018 cpumask_set_cpu(cpu, &cpus_with_pcps);
3020 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3023 for_each_cpu(cpu, &cpus_with_pcps) {
3024 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3027 INIT_WORK(&drain->work, drain_local_pages_wq);
3028 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3030 for_each_cpu(cpu, &cpus_with_pcps)
3031 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3033 mutex_unlock(&pcpu_drain_mutex);
3036 #ifdef CONFIG_HIBERNATION
3039 * Touch the watchdog for every WD_PAGE_COUNT pages.
3041 #define WD_PAGE_COUNT (128*1024)
3043 void mark_free_pages(struct zone *zone)
3045 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3046 unsigned long flags;
3047 unsigned int order, t;
3050 if (zone_is_empty(zone))
3053 spin_lock_irqsave(&zone->lock, flags);
3055 max_zone_pfn = zone_end_pfn(zone);
3056 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3057 if (pfn_valid(pfn)) {
3058 page = pfn_to_page(pfn);
3060 if (!--page_count) {
3061 touch_nmi_watchdog();
3062 page_count = WD_PAGE_COUNT;
3065 if (page_zone(page) != zone)
3068 if (!swsusp_page_is_forbidden(page))
3069 swsusp_unset_page_free(page);
3072 for_each_migratetype_order(order, t) {
3073 list_for_each_entry(page,
3074 &zone->free_area[order].free_list[t], lru) {
3077 pfn = page_to_pfn(page);
3078 for (i = 0; i < (1UL << order); i++) {
3079 if (!--page_count) {
3080 touch_nmi_watchdog();
3081 page_count = WD_PAGE_COUNT;
3083 swsusp_set_page_free(pfn_to_page(pfn + i));
3087 spin_unlock_irqrestore(&zone->lock, flags);
3089 #endif /* CONFIG_PM */
3091 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3095 if (!free_pcp_prepare(page))
3098 migratetype = get_pfnblock_migratetype(page, pfn);
3099 set_pcppage_migratetype(page, migratetype);
3103 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3105 struct zone *zone = page_zone(page);
3106 struct per_cpu_pages *pcp;
3109 migratetype = get_pcppage_migratetype(page);
3110 __count_vm_event(PGFREE);
3113 * We only track unmovable, reclaimable and movable on pcp lists.
3114 * Free ISOLATE pages back to the allocator because they are being
3115 * offlined but treat HIGHATOMIC as movable pages so we can get those
3116 * areas back if necessary. Otherwise, we may have to free
3117 * excessively into the page allocator
3119 if (migratetype >= MIGRATE_PCPTYPES) {
3120 if (unlikely(is_migrate_isolate(migratetype))) {
3121 free_one_page(zone, page, pfn, 0, migratetype);
3124 migratetype = MIGRATE_MOVABLE;
3127 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3128 list_add(&page->lru, &pcp->lists[migratetype]);
3130 if (pcp->count >= pcp->high) {
3131 unsigned long batch = READ_ONCE(pcp->batch);
3132 free_pcppages_bulk(zone, batch, pcp);
3137 * Free a 0-order page
3139 void free_unref_page(struct page *page)
3141 unsigned long flags;
3142 unsigned long pfn = page_to_pfn(page);
3144 if (!free_unref_page_prepare(page, pfn))
3147 local_irq_save(flags);
3148 free_unref_page_commit(page, pfn);
3149 local_irq_restore(flags);
3153 * Free a list of 0-order pages
3155 void free_unref_page_list(struct list_head *list)
3157 struct page *page, *next;
3158 unsigned long flags, pfn;
3159 int batch_count = 0;
3161 /* Prepare pages for freeing */
3162 list_for_each_entry_safe(page, next, list, lru) {
3163 pfn = page_to_pfn(page);
3164 if (!free_unref_page_prepare(page, pfn))
3165 list_del(&page->lru);
3166 set_page_private(page, pfn);
3169 local_irq_save(flags);
3170 list_for_each_entry_safe(page, next, list, lru) {
3171 unsigned long pfn = page_private(page);
3173 set_page_private(page, 0);
3174 trace_mm_page_free_batched(page);
3175 free_unref_page_commit(page, pfn);
3178 * Guard against excessive IRQ disabled times when we get
3179 * a large list of pages to free.
3181 if (++batch_count == SWAP_CLUSTER_MAX) {
3182 local_irq_restore(flags);
3184 local_irq_save(flags);
3187 local_irq_restore(flags);
3191 * split_page takes a non-compound higher-order page, and splits it into
3192 * n (1<<order) sub-pages: page[0..n]
3193 * Each sub-page must be freed individually.
3195 * Note: this is probably too low level an operation for use in drivers.
3196 * Please consult with lkml before using this in your driver.
3198 void split_page(struct page *page, unsigned int order)
3202 VM_BUG_ON_PAGE(PageCompound(page), page);
3203 VM_BUG_ON_PAGE(!page_count(page), page);
3205 for (i = 1; i < (1 << order); i++)
3206 set_page_refcounted(page + i);
3207 split_page_owner(page, order);
3209 EXPORT_SYMBOL_GPL(split_page);
3211 int __isolate_free_page(struct page *page, unsigned int order)
3213 unsigned long watermark;
3217 BUG_ON(!PageBuddy(page));
3219 zone = page_zone(page);
3220 mt = get_pageblock_migratetype(page);
3222 if (!is_migrate_isolate(mt)) {
3224 * Obey watermarks as if the page was being allocated. We can
3225 * emulate a high-order watermark check with a raised order-0
3226 * watermark, because we already know our high-order page
3229 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3230 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3233 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3236 /* Remove page from free list */
3238 del_page_from_free_list(page, zone, order);
3241 * Set the pageblock if the isolated page is at least half of a
3244 if (order >= pageblock_order - 1) {
3245 struct page *endpage = page + (1 << order) - 1;
3246 for (; page < endpage; page += pageblock_nr_pages) {
3247 int mt = get_pageblock_migratetype(page);
3248 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3249 && !is_migrate_highatomic(mt))
3250 set_pageblock_migratetype(page,
3256 return 1UL << order;
3260 * __putback_isolated_page - Return a now-isolated page back where we got it
3261 * @page: Page that was isolated
3262 * @order: Order of the isolated page
3263 * @mt: The page's pageblock's migratetype
3265 * This function is meant to return a page pulled from the free lists via
3266 * __isolate_free_page back to the free lists they were pulled from.
3268 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3270 struct zone *zone = page_zone(page);
3272 /* zone lock should be held when this function is called */
3273 lockdep_assert_held(&zone->lock);
3275 /* Return isolated page to tail of freelist. */
3276 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3280 * Update NUMA hit/miss statistics
3282 * Must be called with interrupts disabled.
3284 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3287 enum numa_stat_item local_stat = NUMA_LOCAL;
3289 /* skip numa counters update if numa stats is disabled */
3290 if (!static_branch_likely(&vm_numa_stat_key))
3293 if (zone_to_nid(z) != numa_node_id())
3294 local_stat = NUMA_OTHER;
3296 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3297 __inc_numa_state(z, NUMA_HIT);
3299 __inc_numa_state(z, NUMA_MISS);
3300 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3302 __inc_numa_state(z, local_stat);
3306 /* Remove page from the per-cpu list, caller must protect the list */
3307 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3308 unsigned int alloc_flags,
3309 struct per_cpu_pages *pcp,
3310 struct list_head *list)
3315 if (list_empty(list)) {
3316 pcp->count += rmqueue_bulk(zone, 0,
3318 migratetype, alloc_flags);
3319 if (unlikely(list_empty(list)))
3323 page = list_first_entry(list, struct page, lru);
3324 list_del(&page->lru);
3326 } while (check_new_pcp(page));
3331 /* Lock and remove page from the per-cpu list */
3332 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3333 struct zone *zone, gfp_t gfp_flags,
3334 int migratetype, unsigned int alloc_flags)
3336 struct per_cpu_pages *pcp;
3337 struct list_head *list;
3339 unsigned long flags;
3341 local_irq_save(flags);
3342 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3343 list = &pcp->lists[migratetype];
3344 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3346 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3347 zone_statistics(preferred_zone, zone);
3349 local_irq_restore(flags);
3354 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3357 struct page *rmqueue(struct zone *preferred_zone,
3358 struct zone *zone, unsigned int order,
3359 gfp_t gfp_flags, unsigned int alloc_flags,
3362 unsigned long flags;
3365 if (likely(order == 0)) {
3366 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3367 migratetype, alloc_flags);
3372 * We most definitely don't want callers attempting to
3373 * allocate greater than order-1 page units with __GFP_NOFAIL.
3375 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3376 spin_lock_irqsave(&zone->lock, flags);
3380 if (alloc_flags & ALLOC_HARDER) {
3381 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3383 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3386 page = __rmqueue(zone, order, migratetype, alloc_flags);
3387 } while (page && check_new_pages(page, order));
3388 spin_unlock(&zone->lock);
3391 __mod_zone_freepage_state(zone, -(1 << order),
3392 get_pcppage_migratetype(page));
3394 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3395 zone_statistics(preferred_zone, zone);
3396 local_irq_restore(flags);
3399 /* Separate test+clear to avoid unnecessary atomics */
3400 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3401 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3402 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3405 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3409 local_irq_restore(flags);
3413 #ifdef CONFIG_FAIL_PAGE_ALLOC
3416 struct fault_attr attr;
3418 bool ignore_gfp_highmem;
3419 bool ignore_gfp_reclaim;
3421 } fail_page_alloc = {
3422 .attr = FAULT_ATTR_INITIALIZER,
3423 .ignore_gfp_reclaim = true,
3424 .ignore_gfp_highmem = true,
3428 static int __init setup_fail_page_alloc(char *str)
3430 return setup_fault_attr(&fail_page_alloc.attr, str);
3432 __setup("fail_page_alloc=", setup_fail_page_alloc);
3434 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3436 if (order < fail_page_alloc.min_order)
3438 if (gfp_mask & __GFP_NOFAIL)
3440 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3442 if (fail_page_alloc.ignore_gfp_reclaim &&
3443 (gfp_mask & __GFP_DIRECT_RECLAIM))
3446 return should_fail(&fail_page_alloc.attr, 1 << order);
3449 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3451 static int __init fail_page_alloc_debugfs(void)
3453 umode_t mode = S_IFREG | 0600;
3456 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3457 &fail_page_alloc.attr);
3459 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3460 &fail_page_alloc.ignore_gfp_reclaim);
3461 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3462 &fail_page_alloc.ignore_gfp_highmem);
3463 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3468 late_initcall(fail_page_alloc_debugfs);
3470 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3472 #else /* CONFIG_FAIL_PAGE_ALLOC */
3474 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3479 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3481 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3483 return __should_fail_alloc_page(gfp_mask, order);
3485 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3488 * Return true if free base pages are above 'mark'. For high-order checks it
3489 * will return true of the order-0 watermark is reached and there is at least
3490 * one free page of a suitable size. Checking now avoids taking the zone lock
3491 * to check in the allocation paths if no pages are free.
3493 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3494 int highest_zoneidx, unsigned int alloc_flags,
3499 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3501 /* free_pages may go negative - that's OK */
3502 free_pages -= (1 << order) - 1;
3504 if (alloc_flags & ALLOC_HIGH)
3508 * If the caller does not have rights to ALLOC_HARDER then subtract
3509 * the high-atomic reserves. This will over-estimate the size of the
3510 * atomic reserve but it avoids a search.
3512 if (likely(!alloc_harder)) {
3513 free_pages -= z->nr_reserved_highatomic;
3516 * OOM victims can try even harder than normal ALLOC_HARDER
3517 * users on the grounds that it's definitely going to be in
3518 * the exit path shortly and free memory. Any allocation it
3519 * makes during the free path will be small and short-lived.
3521 if (alloc_flags & ALLOC_OOM)
3529 /* If allocation can't use CMA areas don't use free CMA pages */
3530 if (!(alloc_flags & ALLOC_CMA))
3531 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3535 * Check watermarks for an order-0 allocation request. If these
3536 * are not met, then a high-order request also cannot go ahead
3537 * even if a suitable page happened to be free.
3539 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3542 /* If this is an order-0 request then the watermark is fine */
3546 /* For a high-order request, check at least one suitable page is free */
3547 for (o = order; o < MAX_ORDER; o++) {
3548 struct free_area *area = &z->free_area[o];
3554 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3555 if (!free_area_empty(area, mt))
3560 if ((alloc_flags & ALLOC_CMA) &&
3561 !free_area_empty(area, MIGRATE_CMA)) {
3565 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3571 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3572 int highest_zoneidx, unsigned int alloc_flags)
3574 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3575 zone_page_state(z, NR_FREE_PAGES));
3578 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3579 unsigned long mark, int highest_zoneidx,
3580 unsigned int alloc_flags)
3582 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3586 /* If allocation can't use CMA areas don't use free CMA pages */
3587 if (!(alloc_flags & ALLOC_CMA))
3588 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3592 * Fast check for order-0 only. If this fails then the reserves
3593 * need to be calculated. There is a corner case where the check
3594 * passes but only the high-order atomic reserve are free. If
3595 * the caller is !atomic then it'll uselessly search the free
3596 * list. That corner case is then slower but it is harmless.
3598 if (!order && (free_pages - cma_pages) >
3599 mark + z->lowmem_reserve[highest_zoneidx])
3602 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3606 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3607 unsigned long mark, int highest_zoneidx)
3609 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3611 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3612 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3614 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3619 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3621 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3622 node_reclaim_distance;
3624 #else /* CONFIG_NUMA */
3625 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3629 #endif /* CONFIG_NUMA */
3632 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3633 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3634 * premature use of a lower zone may cause lowmem pressure problems that
3635 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3636 * probably too small. It only makes sense to spread allocations to avoid
3637 * fragmentation between the Normal and DMA32 zones.
3639 static inline unsigned int
3640 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3642 unsigned int alloc_flags;
3645 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3648 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3650 #ifdef CONFIG_ZONE_DMA32
3654 if (zone_idx(zone) != ZONE_NORMAL)
3658 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3659 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3660 * on UMA that if Normal is populated then so is DMA32.
3662 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3663 if (nr_online_nodes > 1 && !populated_zone(--zone))
3666 alloc_flags |= ALLOC_NOFRAGMENT;
3667 #endif /* CONFIG_ZONE_DMA32 */
3672 * get_page_from_freelist goes through the zonelist trying to allocate
3675 static struct page *
3676 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3677 const struct alloc_context *ac)
3681 struct pglist_data *last_pgdat_dirty_limit = NULL;
3686 * Scan zonelist, looking for a zone with enough free.
3687 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3689 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3690 z = ac->preferred_zoneref;
3691 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist,
3692 ac->highest_zoneidx, ac->nodemask) {
3696 if (cpusets_enabled() &&
3697 (alloc_flags & ALLOC_CPUSET) &&
3698 !__cpuset_zone_allowed(zone, gfp_mask))
3701 * When allocating a page cache page for writing, we
3702 * want to get it from a node that is within its dirty
3703 * limit, such that no single node holds more than its
3704 * proportional share of globally allowed dirty pages.
3705 * The dirty limits take into account the node's
3706 * lowmem reserves and high watermark so that kswapd
3707 * should be able to balance it without having to
3708 * write pages from its LRU list.
3710 * XXX: For now, allow allocations to potentially
3711 * exceed the per-node dirty limit in the slowpath
3712 * (spread_dirty_pages unset) before going into reclaim,
3713 * which is important when on a NUMA setup the allowed
3714 * nodes are together not big enough to reach the
3715 * global limit. The proper fix for these situations
3716 * will require awareness of nodes in the
3717 * dirty-throttling and the flusher threads.
3719 if (ac->spread_dirty_pages) {
3720 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3723 if (!node_dirty_ok(zone->zone_pgdat)) {
3724 last_pgdat_dirty_limit = zone->zone_pgdat;
3729 if (no_fallback && nr_online_nodes > 1 &&
3730 zone != ac->preferred_zoneref->zone) {
3734 * If moving to a remote node, retry but allow
3735 * fragmenting fallbacks. Locality is more important
3736 * than fragmentation avoidance.
3738 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3739 if (zone_to_nid(zone) != local_nid) {
3740 alloc_flags &= ~ALLOC_NOFRAGMENT;
3745 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3746 if (!zone_watermark_fast(zone, order, mark,
3747 ac->highest_zoneidx, alloc_flags)) {
3750 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3752 * Watermark failed for this zone, but see if we can
3753 * grow this zone if it contains deferred pages.
3755 if (static_branch_unlikely(&deferred_pages)) {
3756 if (_deferred_grow_zone(zone, order))
3760 /* Checked here to keep the fast path fast */
3761 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3762 if (alloc_flags & ALLOC_NO_WATERMARKS)
3765 if (node_reclaim_mode == 0 ||
3766 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3769 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3771 case NODE_RECLAIM_NOSCAN:
3774 case NODE_RECLAIM_FULL:
3775 /* scanned but unreclaimable */
3778 /* did we reclaim enough */
3779 if (zone_watermark_ok(zone, order, mark,
3780 ac->highest_zoneidx, alloc_flags))
3788 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3789 gfp_mask, alloc_flags, ac->migratetype);
3791 prep_new_page(page, order, gfp_mask, alloc_flags);
3794 * If this is a high-order atomic allocation then check
3795 * if the pageblock should be reserved for the future
3797 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3798 reserve_highatomic_pageblock(page, zone, order);
3802 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3803 /* Try again if zone has deferred pages */
3804 if (static_branch_unlikely(&deferred_pages)) {
3805 if (_deferred_grow_zone(zone, order))
3813 * It's possible on a UMA machine to get through all zones that are
3814 * fragmented. If avoiding fragmentation, reset and try again.
3817 alloc_flags &= ~ALLOC_NOFRAGMENT;
3824 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3826 unsigned int filter = SHOW_MEM_FILTER_NODES;
3829 * This documents exceptions given to allocations in certain
3830 * contexts that are allowed to allocate outside current's set
3833 if (!(gfp_mask & __GFP_NOMEMALLOC))
3834 if (tsk_is_oom_victim(current) ||
3835 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3836 filter &= ~SHOW_MEM_FILTER_NODES;
3837 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3838 filter &= ~SHOW_MEM_FILTER_NODES;
3840 show_mem(filter, nodemask);
3843 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3845 struct va_format vaf;
3847 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3849 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3852 va_start(args, fmt);
3855 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3856 current->comm, &vaf, gfp_mask, &gfp_mask,
3857 nodemask_pr_args(nodemask));
3860 cpuset_print_current_mems_allowed();
3863 warn_alloc_show_mem(gfp_mask, nodemask);
3866 static inline struct page *
3867 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3868 unsigned int alloc_flags,
3869 const struct alloc_context *ac)
3873 page = get_page_from_freelist(gfp_mask, order,
3874 alloc_flags|ALLOC_CPUSET, ac);
3876 * fallback to ignore cpuset restriction if our nodes
3880 page = get_page_from_freelist(gfp_mask, order,
3886 static inline struct page *
3887 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3888 const struct alloc_context *ac, unsigned long *did_some_progress)
3890 struct oom_control oc = {
3891 .zonelist = ac->zonelist,
3892 .nodemask = ac->nodemask,
3894 .gfp_mask = gfp_mask,
3899 *did_some_progress = 0;
3902 * Acquire the oom lock. If that fails, somebody else is
3903 * making progress for us.
3905 if (!mutex_trylock(&oom_lock)) {
3906 *did_some_progress = 1;
3907 schedule_timeout_uninterruptible(1);
3912 * Go through the zonelist yet one more time, keep very high watermark
3913 * here, this is only to catch a parallel oom killing, we must fail if
3914 * we're still under heavy pressure. But make sure that this reclaim
3915 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3916 * allocation which will never fail due to oom_lock already held.
3918 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3919 ~__GFP_DIRECT_RECLAIM, order,
3920 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3924 /* Coredumps can quickly deplete all memory reserves */
3925 if (current->flags & PF_DUMPCORE)
3927 /* The OOM killer will not help higher order allocs */
3928 if (order > PAGE_ALLOC_COSTLY_ORDER)
3931 * We have already exhausted all our reclaim opportunities without any
3932 * success so it is time to admit defeat. We will skip the OOM killer
3933 * because it is very likely that the caller has a more reasonable
3934 * fallback than shooting a random task.
3936 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3938 /* The OOM killer does not needlessly kill tasks for lowmem */
3939 if (ac->highest_zoneidx < ZONE_NORMAL)
3941 if (pm_suspended_storage())
3944 * XXX: GFP_NOFS allocations should rather fail than rely on
3945 * other request to make a forward progress.
3946 * We are in an unfortunate situation where out_of_memory cannot
3947 * do much for this context but let's try it to at least get
3948 * access to memory reserved if the current task is killed (see
3949 * out_of_memory). Once filesystems are ready to handle allocation
3950 * failures more gracefully we should just bail out here.
3953 /* The OOM killer may not free memory on a specific node */
3954 if (gfp_mask & __GFP_THISNODE)
3957 /* Exhausted what can be done so it's blame time */
3958 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3959 *did_some_progress = 1;
3962 * Help non-failing allocations by giving them access to memory
3965 if (gfp_mask & __GFP_NOFAIL)
3966 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3967 ALLOC_NO_WATERMARKS, ac);
3970 mutex_unlock(&oom_lock);
3975 * Maximum number of compaction retries wit a progress before OOM
3976 * killer is consider as the only way to move forward.
3978 #define MAX_COMPACT_RETRIES 16
3980 #ifdef CONFIG_COMPACTION
3981 /* Try memory compaction for high-order allocations before reclaim */
3982 static struct page *
3983 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3984 unsigned int alloc_flags, const struct alloc_context *ac,
3985 enum compact_priority prio, enum compact_result *compact_result)
3987 struct page *page = NULL;
3988 unsigned long pflags;
3989 unsigned int noreclaim_flag;
3994 psi_memstall_enter(&pflags);
3995 noreclaim_flag = memalloc_noreclaim_save();
3997 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4000 memalloc_noreclaim_restore(noreclaim_flag);
4001 psi_memstall_leave(&pflags);
4004 * At least in one zone compaction wasn't deferred or skipped, so let's
4005 * count a compaction stall
4007 count_vm_event(COMPACTSTALL);
4009 /* Prep a captured page if available */
4011 prep_new_page(page, order, gfp_mask, alloc_flags);
4013 /* Try get a page from the freelist if available */
4015 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4018 struct zone *zone = page_zone(page);
4020 zone->compact_blockskip_flush = false;
4021 compaction_defer_reset(zone, order, true);
4022 count_vm_event(COMPACTSUCCESS);
4027 * It's bad if compaction run occurs and fails. The most likely reason
4028 * is that pages exist, but not enough to satisfy watermarks.
4030 count_vm_event(COMPACTFAIL);
4038 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4039 enum compact_result compact_result,
4040 enum compact_priority *compact_priority,
4041 int *compaction_retries)
4043 int max_retries = MAX_COMPACT_RETRIES;
4046 int retries = *compaction_retries;
4047 enum compact_priority priority = *compact_priority;
4052 if (compaction_made_progress(compact_result))
4053 (*compaction_retries)++;
4056 * compaction considers all the zone as desperately out of memory
4057 * so it doesn't really make much sense to retry except when the
4058 * failure could be caused by insufficient priority
4060 if (compaction_failed(compact_result))
4061 goto check_priority;
4064 * compaction was skipped because there are not enough order-0 pages
4065 * to work with, so we retry only if it looks like reclaim can help.
4067 if (compaction_needs_reclaim(compact_result)) {
4068 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4073 * make sure the compaction wasn't deferred or didn't bail out early
4074 * due to locks contention before we declare that we should give up.
4075 * But the next retry should use a higher priority if allowed, so
4076 * we don't just keep bailing out endlessly.
4078 if (compaction_withdrawn(compact_result)) {
4079 goto check_priority;
4083 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4084 * costly ones because they are de facto nofail and invoke OOM
4085 * killer to move on while costly can fail and users are ready
4086 * to cope with that. 1/4 retries is rather arbitrary but we
4087 * would need much more detailed feedback from compaction to
4088 * make a better decision.
4090 if (order > PAGE_ALLOC_COSTLY_ORDER)
4092 if (*compaction_retries <= max_retries) {
4098 * Make sure there are attempts at the highest priority if we exhausted
4099 * all retries or failed at the lower priorities.
4102 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4103 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4105 if (*compact_priority > min_priority) {
4106 (*compact_priority)--;
4107 *compaction_retries = 0;
4111 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4115 static inline struct page *
4116 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4117 unsigned int alloc_flags, const struct alloc_context *ac,
4118 enum compact_priority prio, enum compact_result *compact_result)
4120 *compact_result = COMPACT_SKIPPED;
4125 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4126 enum compact_result compact_result,
4127 enum compact_priority *compact_priority,
4128 int *compaction_retries)
4133 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4137 * There are setups with compaction disabled which would prefer to loop
4138 * inside the allocator rather than hit the oom killer prematurely.
4139 * Let's give them a good hope and keep retrying while the order-0
4140 * watermarks are OK.
4142 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4143 ac->highest_zoneidx, ac->nodemask) {
4144 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4145 ac->highest_zoneidx, alloc_flags))
4150 #endif /* CONFIG_COMPACTION */
4152 #ifdef CONFIG_LOCKDEP
4153 static struct lockdep_map __fs_reclaim_map =
4154 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4156 static bool __need_fs_reclaim(gfp_t gfp_mask)
4158 gfp_mask = current_gfp_context(gfp_mask);
4160 /* no reclaim without waiting on it */
4161 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4164 /* this guy won't enter reclaim */
4165 if (current->flags & PF_MEMALLOC)
4168 /* We're only interested __GFP_FS allocations for now */
4169 if (!(gfp_mask & __GFP_FS))
4172 if (gfp_mask & __GFP_NOLOCKDEP)
4178 void __fs_reclaim_acquire(void)
4180 lock_map_acquire(&__fs_reclaim_map);
4183 void __fs_reclaim_release(void)
4185 lock_map_release(&__fs_reclaim_map);
4188 void fs_reclaim_acquire(gfp_t gfp_mask)
4190 if (__need_fs_reclaim(gfp_mask))
4191 __fs_reclaim_acquire();
4193 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4195 void fs_reclaim_release(gfp_t gfp_mask)
4197 if (__need_fs_reclaim(gfp_mask))
4198 __fs_reclaim_release();
4200 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4203 /* Perform direct synchronous page reclaim */
4205 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4206 const struct alloc_context *ac)
4209 unsigned int noreclaim_flag;
4210 unsigned long pflags;
4214 /* We now go into synchronous reclaim */
4215 cpuset_memory_pressure_bump();
4216 psi_memstall_enter(&pflags);
4217 fs_reclaim_acquire(gfp_mask);
4218 noreclaim_flag = memalloc_noreclaim_save();
4220 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4223 memalloc_noreclaim_restore(noreclaim_flag);
4224 fs_reclaim_release(gfp_mask);
4225 psi_memstall_leave(&pflags);
4232 /* The really slow allocator path where we enter direct reclaim */
4233 static inline struct page *
4234 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4235 unsigned int alloc_flags, const struct alloc_context *ac,
4236 unsigned long *did_some_progress)
4238 struct page *page = NULL;
4239 bool drained = false;
4241 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4242 if (unlikely(!(*did_some_progress)))
4246 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4249 * If an allocation failed after direct reclaim, it could be because
4250 * pages are pinned on the per-cpu lists or in high alloc reserves.
4251 * Shrink them them and try again
4253 if (!page && !drained) {
4254 unreserve_highatomic_pageblock(ac, false);
4255 drain_all_pages(NULL);
4263 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4264 const struct alloc_context *ac)
4268 pg_data_t *last_pgdat = NULL;
4269 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4271 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4273 if (last_pgdat != zone->zone_pgdat)
4274 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4275 last_pgdat = zone->zone_pgdat;
4279 static inline unsigned int
4280 gfp_to_alloc_flags(gfp_t gfp_mask)
4282 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4285 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4286 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4287 * to save two branches.
4289 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4290 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4293 * The caller may dip into page reserves a bit more if the caller
4294 * cannot run direct reclaim, or if the caller has realtime scheduling
4295 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4296 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4298 alloc_flags |= (__force int)
4299 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4301 if (gfp_mask & __GFP_ATOMIC) {
4303 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4304 * if it can't schedule.
4306 if (!(gfp_mask & __GFP_NOMEMALLOC))
4307 alloc_flags |= ALLOC_HARDER;
4309 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4310 * comment for __cpuset_node_allowed().
4312 alloc_flags &= ~ALLOC_CPUSET;
4313 } else if (unlikely(rt_task(current)) && !in_interrupt())
4314 alloc_flags |= ALLOC_HARDER;
4317 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4318 alloc_flags |= ALLOC_CMA;
4323 static bool oom_reserves_allowed(struct task_struct *tsk)
4325 if (!tsk_is_oom_victim(tsk))
4329 * !MMU doesn't have oom reaper so give access to memory reserves
4330 * only to the thread with TIF_MEMDIE set
4332 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4339 * Distinguish requests which really need access to full memory
4340 * reserves from oom victims which can live with a portion of it
4342 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4344 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4346 if (gfp_mask & __GFP_MEMALLOC)
4347 return ALLOC_NO_WATERMARKS;
4348 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4349 return ALLOC_NO_WATERMARKS;
4350 if (!in_interrupt()) {
4351 if (current->flags & PF_MEMALLOC)
4352 return ALLOC_NO_WATERMARKS;
4353 else if (oom_reserves_allowed(current))
4360 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4362 return !!__gfp_pfmemalloc_flags(gfp_mask);
4366 * Checks whether it makes sense to retry the reclaim to make a forward progress
4367 * for the given allocation request.
4369 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4370 * without success, or when we couldn't even meet the watermark if we
4371 * reclaimed all remaining pages on the LRU lists.
4373 * Returns true if a retry is viable or false to enter the oom path.
4376 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4377 struct alloc_context *ac, int alloc_flags,
4378 bool did_some_progress, int *no_progress_loops)
4385 * Costly allocations might have made a progress but this doesn't mean
4386 * their order will become available due to high fragmentation so
4387 * always increment the no progress counter for them
4389 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4390 *no_progress_loops = 0;
4392 (*no_progress_loops)++;
4395 * Make sure we converge to OOM if we cannot make any progress
4396 * several times in the row.
4398 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4399 /* Before OOM, exhaust highatomic_reserve */
4400 return unreserve_highatomic_pageblock(ac, true);
4404 * Keep reclaiming pages while there is a chance this will lead
4405 * somewhere. If none of the target zones can satisfy our allocation
4406 * request even if all reclaimable pages are considered then we are
4407 * screwed and have to go OOM.
4409 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4410 ac->highest_zoneidx, ac->nodemask) {
4411 unsigned long available;
4412 unsigned long reclaimable;
4413 unsigned long min_wmark = min_wmark_pages(zone);
4416 available = reclaimable = zone_reclaimable_pages(zone);
4417 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4420 * Would the allocation succeed if we reclaimed all
4421 * reclaimable pages?
4423 wmark = __zone_watermark_ok(zone, order, min_wmark,
4424 ac->highest_zoneidx, alloc_flags, available);
4425 trace_reclaim_retry_zone(z, order, reclaimable,
4426 available, min_wmark, *no_progress_loops, wmark);
4429 * If we didn't make any progress and have a lot of
4430 * dirty + writeback pages then we should wait for
4431 * an IO to complete to slow down the reclaim and
4432 * prevent from pre mature OOM
4434 if (!did_some_progress) {
4435 unsigned long write_pending;
4437 write_pending = zone_page_state_snapshot(zone,
4438 NR_ZONE_WRITE_PENDING);
4440 if (2 * write_pending > reclaimable) {
4441 congestion_wait(BLK_RW_ASYNC, HZ/10);
4453 * Memory allocation/reclaim might be called from a WQ context and the
4454 * current implementation of the WQ concurrency control doesn't
4455 * recognize that a particular WQ is congested if the worker thread is
4456 * looping without ever sleeping. Therefore we have to do a short sleep
4457 * here rather than calling cond_resched().
4459 if (current->flags & PF_WQ_WORKER)
4460 schedule_timeout_uninterruptible(1);
4467 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4470 * It's possible that cpuset's mems_allowed and the nodemask from
4471 * mempolicy don't intersect. This should be normally dealt with by
4472 * policy_nodemask(), but it's possible to race with cpuset update in
4473 * such a way the check therein was true, and then it became false
4474 * before we got our cpuset_mems_cookie here.
4475 * This assumes that for all allocations, ac->nodemask can come only
4476 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4477 * when it does not intersect with the cpuset restrictions) or the
4478 * caller can deal with a violated nodemask.
4480 if (cpusets_enabled() && ac->nodemask &&
4481 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4482 ac->nodemask = NULL;
4487 * When updating a task's mems_allowed or mempolicy nodemask, it is
4488 * possible to race with parallel threads in such a way that our
4489 * allocation can fail while the mask is being updated. If we are about
4490 * to fail, check if the cpuset changed during allocation and if so,
4493 if (read_mems_allowed_retry(cpuset_mems_cookie))
4499 static inline struct page *
4500 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4501 struct alloc_context *ac)
4503 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4504 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4505 struct page *page = NULL;
4506 unsigned int alloc_flags;
4507 unsigned long did_some_progress;
4508 enum compact_priority compact_priority;
4509 enum compact_result compact_result;
4510 int compaction_retries;
4511 int no_progress_loops;
4512 unsigned int cpuset_mems_cookie;
4516 * We also sanity check to catch abuse of atomic reserves being used by
4517 * callers that are not in atomic context.
4519 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4520 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4521 gfp_mask &= ~__GFP_ATOMIC;
4524 compaction_retries = 0;
4525 no_progress_loops = 0;
4526 compact_priority = DEF_COMPACT_PRIORITY;
4527 cpuset_mems_cookie = read_mems_allowed_begin();
4530 * The fast path uses conservative alloc_flags to succeed only until
4531 * kswapd needs to be woken up, and to avoid the cost of setting up
4532 * alloc_flags precisely. So we do that now.
4534 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4537 * We need to recalculate the starting point for the zonelist iterator
4538 * because we might have used different nodemask in the fast path, or
4539 * there was a cpuset modification and we are retrying - otherwise we
4540 * could end up iterating over non-eligible zones endlessly.
4542 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4543 ac->highest_zoneidx, ac->nodemask);
4544 if (!ac->preferred_zoneref->zone)
4547 if (alloc_flags & ALLOC_KSWAPD)
4548 wake_all_kswapds(order, gfp_mask, ac);
4551 * The adjusted alloc_flags might result in immediate success, so try
4554 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4559 * For costly allocations, try direct compaction first, as it's likely
4560 * that we have enough base pages and don't need to reclaim. For non-
4561 * movable high-order allocations, do that as well, as compaction will
4562 * try prevent permanent fragmentation by migrating from blocks of the
4564 * Don't try this for allocations that are allowed to ignore
4565 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4567 if (can_direct_reclaim &&
4569 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4570 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4571 page = __alloc_pages_direct_compact(gfp_mask, order,
4573 INIT_COMPACT_PRIORITY,
4579 * Checks for costly allocations with __GFP_NORETRY, which
4580 * includes some THP page fault allocations
4582 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4584 * If allocating entire pageblock(s) and compaction
4585 * failed because all zones are below low watermarks
4586 * or is prohibited because it recently failed at this
4587 * order, fail immediately unless the allocator has
4588 * requested compaction and reclaim retry.
4591 * - potentially very expensive because zones are far
4592 * below their low watermarks or this is part of very
4593 * bursty high order allocations,
4594 * - not guaranteed to help because isolate_freepages()
4595 * may not iterate over freed pages as part of its
4597 * - unlikely to make entire pageblocks free on its
4600 if (compact_result == COMPACT_SKIPPED ||
4601 compact_result == COMPACT_DEFERRED)
4605 * Looks like reclaim/compaction is worth trying, but
4606 * sync compaction could be very expensive, so keep
4607 * using async compaction.
4609 compact_priority = INIT_COMPACT_PRIORITY;
4614 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4615 if (alloc_flags & ALLOC_KSWAPD)
4616 wake_all_kswapds(order, gfp_mask, ac);
4618 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4620 alloc_flags = reserve_flags;
4623 * Reset the nodemask and zonelist iterators if memory policies can be
4624 * ignored. These allocations are high priority and system rather than
4627 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4628 ac->nodemask = NULL;
4629 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4630 ac->highest_zoneidx, ac->nodemask);
4633 /* Attempt with potentially adjusted zonelist and alloc_flags */
4634 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4638 /* Caller is not willing to reclaim, we can't balance anything */
4639 if (!can_direct_reclaim)
4642 /* Avoid recursion of direct reclaim */
4643 if (current->flags & PF_MEMALLOC)
4646 /* Try direct reclaim and then allocating */
4647 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4648 &did_some_progress);
4652 /* Try direct compaction and then allocating */
4653 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4654 compact_priority, &compact_result);
4658 /* Do not loop if specifically requested */
4659 if (gfp_mask & __GFP_NORETRY)
4663 * Do not retry costly high order allocations unless they are
4664 * __GFP_RETRY_MAYFAIL
4666 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4669 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4670 did_some_progress > 0, &no_progress_loops))
4674 * It doesn't make any sense to retry for the compaction if the order-0
4675 * reclaim is not able to make any progress because the current
4676 * implementation of the compaction depends on the sufficient amount
4677 * of free memory (see __compaction_suitable)
4679 if (did_some_progress > 0 &&
4680 should_compact_retry(ac, order, alloc_flags,
4681 compact_result, &compact_priority,
4682 &compaction_retries))
4686 /* Deal with possible cpuset update races before we start OOM killing */
4687 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4690 /* Reclaim has failed us, start killing things */
4691 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4695 /* Avoid allocations with no watermarks from looping endlessly */
4696 if (tsk_is_oom_victim(current) &&
4697 (alloc_flags == ALLOC_OOM ||
4698 (gfp_mask & __GFP_NOMEMALLOC)))
4701 /* Retry as long as the OOM killer is making progress */
4702 if (did_some_progress) {
4703 no_progress_loops = 0;
4708 /* Deal with possible cpuset update races before we fail */
4709 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4713 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4716 if (gfp_mask & __GFP_NOFAIL) {
4718 * All existing users of the __GFP_NOFAIL are blockable, so warn
4719 * of any new users that actually require GFP_NOWAIT
4721 if (WARN_ON_ONCE(!can_direct_reclaim))
4725 * PF_MEMALLOC request from this context is rather bizarre
4726 * because we cannot reclaim anything and only can loop waiting
4727 * for somebody to do a work for us
4729 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4732 * non failing costly orders are a hard requirement which we
4733 * are not prepared for much so let's warn about these users
4734 * so that we can identify them and convert them to something
4737 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4740 * Help non-failing allocations by giving them access to memory
4741 * reserves but do not use ALLOC_NO_WATERMARKS because this
4742 * could deplete whole memory reserves which would just make
4743 * the situation worse
4745 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4753 warn_alloc(gfp_mask, ac->nodemask,
4754 "page allocation failure: order:%u", order);
4759 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4760 int preferred_nid, nodemask_t *nodemask,
4761 struct alloc_context *ac, gfp_t *alloc_mask,
4762 unsigned int *alloc_flags)
4764 ac->highest_zoneidx = gfp_zone(gfp_mask);
4765 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4766 ac->nodemask = nodemask;
4767 ac->migratetype = gfp_migratetype(gfp_mask);
4769 if (cpusets_enabled()) {
4770 *alloc_mask |= __GFP_HARDWALL;
4772 ac->nodemask = &cpuset_current_mems_allowed;
4774 *alloc_flags |= ALLOC_CPUSET;
4777 fs_reclaim_acquire(gfp_mask);
4778 fs_reclaim_release(gfp_mask);
4780 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4782 if (should_fail_alloc_page(gfp_mask, order))
4785 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4786 *alloc_flags |= ALLOC_CMA;
4791 /* Determine whether to spread dirty pages and what the first usable zone */
4792 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4794 /* Dirty zone balancing only done in the fast path */
4795 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4798 * The preferred zone is used for statistics but crucially it is
4799 * also used as the starting point for the zonelist iterator. It
4800 * may get reset for allocations that ignore memory policies.
4802 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4803 ac->highest_zoneidx, ac->nodemask);
4807 * This is the 'heart' of the zoned buddy allocator.
4810 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4811 nodemask_t *nodemask)
4814 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4815 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4816 struct alloc_context ac = { };
4819 * There are several places where we assume that the order value is sane
4820 * so bail out early if the request is out of bound.
4822 if (unlikely(order >= MAX_ORDER)) {
4823 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4827 gfp_mask &= gfp_allowed_mask;
4828 alloc_mask = gfp_mask;
4829 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4832 finalise_ac(gfp_mask, &ac);
4835 * Forbid the first pass from falling back to types that fragment
4836 * memory until all local zones are considered.
4838 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4840 /* First allocation attempt */
4841 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4846 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4847 * resp. GFP_NOIO which has to be inherited for all allocation requests
4848 * from a particular context which has been marked by
4849 * memalloc_no{fs,io}_{save,restore}.
4851 alloc_mask = current_gfp_context(gfp_mask);
4852 ac.spread_dirty_pages = false;
4855 * Restore the original nodemask if it was potentially replaced with
4856 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4858 ac.nodemask = nodemask;
4860 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4863 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4864 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4865 __free_pages(page, order);
4869 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4873 EXPORT_SYMBOL(__alloc_pages_nodemask);
4876 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4877 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4878 * you need to access high mem.
4880 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4884 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4887 return (unsigned long) page_address(page);
4889 EXPORT_SYMBOL(__get_free_pages);
4891 unsigned long get_zeroed_page(gfp_t gfp_mask)
4893 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4895 EXPORT_SYMBOL(get_zeroed_page);
4897 static inline void free_the_page(struct page *page, unsigned int order)
4899 if (order == 0) /* Via pcp? */
4900 free_unref_page(page);
4902 __free_pages_ok(page, order);
4905 void __free_pages(struct page *page, unsigned int order)
4907 if (put_page_testzero(page))
4908 free_the_page(page, order);
4910 EXPORT_SYMBOL(__free_pages);
4912 void free_pages(unsigned long addr, unsigned int order)
4915 VM_BUG_ON(!virt_addr_valid((void *)addr));
4916 __free_pages(virt_to_page((void *)addr), order);
4920 EXPORT_SYMBOL(free_pages);
4924 * An arbitrary-length arbitrary-offset area of memory which resides
4925 * within a 0 or higher order page. Multiple fragments within that page
4926 * are individually refcounted, in the page's reference counter.
4928 * The page_frag functions below provide a simple allocation framework for
4929 * page fragments. This is used by the network stack and network device
4930 * drivers to provide a backing region of memory for use as either an
4931 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4933 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4936 struct page *page = NULL;
4937 gfp_t gfp = gfp_mask;
4939 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4940 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4942 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4943 PAGE_FRAG_CACHE_MAX_ORDER);
4944 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4946 if (unlikely(!page))
4947 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4949 nc->va = page ? page_address(page) : NULL;
4954 void __page_frag_cache_drain(struct page *page, unsigned int count)
4956 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4958 if (page_ref_sub_and_test(page, count))
4959 free_the_page(page, compound_order(page));
4961 EXPORT_SYMBOL(__page_frag_cache_drain);
4963 void *page_frag_alloc(struct page_frag_cache *nc,
4964 unsigned int fragsz, gfp_t gfp_mask)
4966 unsigned int size = PAGE_SIZE;
4970 if (unlikely(!nc->va)) {
4972 page = __page_frag_cache_refill(nc, gfp_mask);
4976 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4977 /* if size can vary use size else just use PAGE_SIZE */
4980 /* Even if we own the page, we do not use atomic_set().
4981 * This would break get_page_unless_zero() users.
4983 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4985 /* reset page count bias and offset to start of new frag */
4986 nc->pfmemalloc = page_is_pfmemalloc(page);
4987 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4991 offset = nc->offset - fragsz;
4992 if (unlikely(offset < 0)) {
4993 page = virt_to_page(nc->va);
4995 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4998 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4999 /* if size can vary use size else just use PAGE_SIZE */
5002 /* OK, page count is 0, we can safely set it */
5003 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5005 /* reset page count bias and offset to start of new frag */
5006 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5007 offset = size - fragsz;
5011 nc->offset = offset;
5013 return nc->va + offset;
5015 EXPORT_SYMBOL(page_frag_alloc);
5018 * Frees a page fragment allocated out of either a compound or order 0 page.
5020 void page_frag_free(void *addr)
5022 struct page *page = virt_to_head_page(addr);
5024 if (unlikely(put_page_testzero(page)))
5025 free_the_page(page, compound_order(page));
5027 EXPORT_SYMBOL(page_frag_free);
5029 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5033 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5034 unsigned long used = addr + PAGE_ALIGN(size);
5036 split_page(virt_to_page((void *)addr), order);
5037 while (used < alloc_end) {
5042 return (void *)addr;
5046 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5047 * @size: the number of bytes to allocate
5048 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5050 * This function is similar to alloc_pages(), except that it allocates the
5051 * minimum number of pages to satisfy the request. alloc_pages() can only
5052 * allocate memory in power-of-two pages.
5054 * This function is also limited by MAX_ORDER.
5056 * Memory allocated by this function must be released by free_pages_exact().
5058 * Return: pointer to the allocated area or %NULL in case of error.
5060 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5062 unsigned int order = get_order(size);
5065 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5066 gfp_mask &= ~__GFP_COMP;
5068 addr = __get_free_pages(gfp_mask, order);
5069 return make_alloc_exact(addr, order, size);
5071 EXPORT_SYMBOL(alloc_pages_exact);
5074 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5076 * @nid: the preferred node ID where memory should be allocated
5077 * @size: the number of bytes to allocate
5078 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5080 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5083 * Return: pointer to the allocated area or %NULL in case of error.
5085 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5087 unsigned int order = get_order(size);
5090 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5091 gfp_mask &= ~__GFP_COMP;
5093 p = alloc_pages_node(nid, gfp_mask, order);
5096 return make_alloc_exact((unsigned long)page_address(p), order, size);
5100 * free_pages_exact - release memory allocated via alloc_pages_exact()
5101 * @virt: the value returned by alloc_pages_exact.
5102 * @size: size of allocation, same value as passed to alloc_pages_exact().
5104 * Release the memory allocated by a previous call to alloc_pages_exact.
5106 void free_pages_exact(void *virt, size_t size)
5108 unsigned long addr = (unsigned long)virt;
5109 unsigned long end = addr + PAGE_ALIGN(size);
5111 while (addr < end) {
5116 EXPORT_SYMBOL(free_pages_exact);
5119 * nr_free_zone_pages - count number of pages beyond high watermark
5120 * @offset: The zone index of the highest zone
5122 * nr_free_zone_pages() counts the number of pages which are beyond the
5123 * high watermark within all zones at or below a given zone index. For each
5124 * zone, the number of pages is calculated as:
5126 * nr_free_zone_pages = managed_pages - high_pages
5128 * Return: number of pages beyond high watermark.
5130 static unsigned long nr_free_zone_pages(int offset)
5135 /* Just pick one node, since fallback list is circular */
5136 unsigned long sum = 0;
5138 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5140 for_each_zone_zonelist(zone, z, zonelist, offset) {
5141 unsigned long size = zone_managed_pages(zone);
5142 unsigned long high = high_wmark_pages(zone);
5151 * nr_free_buffer_pages - count number of pages beyond high watermark
5153 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5154 * watermark within ZONE_DMA and ZONE_NORMAL.
5156 * Return: number of pages beyond high watermark within ZONE_DMA and
5159 unsigned long nr_free_buffer_pages(void)
5161 return nr_free_zone_pages(gfp_zone(GFP_USER));
5163 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5166 * nr_free_pagecache_pages - count number of pages beyond high watermark
5168 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5169 * high watermark within all zones.
5171 * Return: number of pages beyond high watermark within all zones.
5173 unsigned long nr_free_pagecache_pages(void)
5175 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5178 static inline void show_node(struct zone *zone)
5180 if (IS_ENABLED(CONFIG_NUMA))
5181 printk("Node %d ", zone_to_nid(zone));
5184 long si_mem_available(void)
5187 unsigned long pagecache;
5188 unsigned long wmark_low = 0;
5189 unsigned long pages[NR_LRU_LISTS];
5190 unsigned long reclaimable;
5194 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5195 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5198 wmark_low += low_wmark_pages(zone);
5201 * Estimate the amount of memory available for userspace allocations,
5202 * without causing swapping.
5204 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5207 * Not all the page cache can be freed, otherwise the system will
5208 * start swapping. Assume at least half of the page cache, or the
5209 * low watermark worth of cache, needs to stay.
5211 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5212 pagecache -= min(pagecache / 2, wmark_low);
5213 available += pagecache;
5216 * Part of the reclaimable slab and other kernel memory consists of
5217 * items that are in use, and cannot be freed. Cap this estimate at the
5220 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5221 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5222 available += reclaimable - min(reclaimable / 2, wmark_low);
5228 EXPORT_SYMBOL_GPL(si_mem_available);
5230 void si_meminfo(struct sysinfo *val)
5232 val->totalram = totalram_pages();
5233 val->sharedram = global_node_page_state(NR_SHMEM);
5234 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5235 val->bufferram = nr_blockdev_pages();
5236 val->totalhigh = totalhigh_pages();
5237 val->freehigh = nr_free_highpages();
5238 val->mem_unit = PAGE_SIZE;
5241 EXPORT_SYMBOL(si_meminfo);
5244 void si_meminfo_node(struct sysinfo *val, int nid)
5246 int zone_type; /* needs to be signed */
5247 unsigned long managed_pages = 0;
5248 unsigned long managed_highpages = 0;
5249 unsigned long free_highpages = 0;
5250 pg_data_t *pgdat = NODE_DATA(nid);
5252 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5253 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5254 val->totalram = managed_pages;
5255 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5256 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5257 #ifdef CONFIG_HIGHMEM
5258 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5259 struct zone *zone = &pgdat->node_zones[zone_type];
5261 if (is_highmem(zone)) {
5262 managed_highpages += zone_managed_pages(zone);
5263 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5266 val->totalhigh = managed_highpages;
5267 val->freehigh = free_highpages;
5269 val->totalhigh = managed_highpages;
5270 val->freehigh = free_highpages;
5272 val->mem_unit = PAGE_SIZE;
5277 * Determine whether the node should be displayed or not, depending on whether
5278 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5280 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5282 if (!(flags & SHOW_MEM_FILTER_NODES))
5286 * no node mask - aka implicit memory numa policy. Do not bother with
5287 * the synchronization - read_mems_allowed_begin - because we do not
5288 * have to be precise here.
5291 nodemask = &cpuset_current_mems_allowed;
5293 return !node_isset(nid, *nodemask);
5296 #define K(x) ((x) << (PAGE_SHIFT-10))
5298 static void show_migration_types(unsigned char type)
5300 static const char types[MIGRATE_TYPES] = {
5301 [MIGRATE_UNMOVABLE] = 'U',
5302 [MIGRATE_MOVABLE] = 'M',
5303 [MIGRATE_RECLAIMABLE] = 'E',
5304 [MIGRATE_HIGHATOMIC] = 'H',
5306 [MIGRATE_CMA] = 'C',
5308 #ifdef CONFIG_MEMORY_ISOLATION
5309 [MIGRATE_ISOLATE] = 'I',
5312 char tmp[MIGRATE_TYPES + 1];
5316 for (i = 0; i < MIGRATE_TYPES; i++) {
5317 if (type & (1 << i))
5322 printk(KERN_CONT "(%s) ", tmp);
5326 * Show free area list (used inside shift_scroll-lock stuff)
5327 * We also calculate the percentage fragmentation. We do this by counting the
5328 * memory on each free list with the exception of the first item on the list.
5331 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5334 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5336 unsigned long free_pcp = 0;
5341 for_each_populated_zone(zone) {
5342 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5345 for_each_online_cpu(cpu)
5346 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5349 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5350 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5351 " unevictable:%lu dirty:%lu writeback:%lu\n"
5352 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5353 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5354 " free:%lu free_pcp:%lu free_cma:%lu\n",
5355 global_node_page_state(NR_ACTIVE_ANON),
5356 global_node_page_state(NR_INACTIVE_ANON),
5357 global_node_page_state(NR_ISOLATED_ANON),
5358 global_node_page_state(NR_ACTIVE_FILE),
5359 global_node_page_state(NR_INACTIVE_FILE),
5360 global_node_page_state(NR_ISOLATED_FILE),
5361 global_node_page_state(NR_UNEVICTABLE),
5362 global_node_page_state(NR_FILE_DIRTY),
5363 global_node_page_state(NR_WRITEBACK),
5364 global_node_page_state(NR_SLAB_RECLAIMABLE),
5365 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5366 global_node_page_state(NR_FILE_MAPPED),
5367 global_node_page_state(NR_SHMEM),
5368 global_zone_page_state(NR_PAGETABLE),
5369 global_zone_page_state(NR_BOUNCE),
5370 global_zone_page_state(NR_FREE_PAGES),
5372 global_zone_page_state(NR_FREE_CMA_PAGES));
5374 for_each_online_pgdat(pgdat) {
5375 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5379 " active_anon:%lukB"
5380 " inactive_anon:%lukB"
5381 " active_file:%lukB"
5382 " inactive_file:%lukB"
5383 " unevictable:%lukB"
5384 " isolated(anon):%lukB"
5385 " isolated(file):%lukB"
5390 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5392 " shmem_pmdmapped: %lukB"
5395 " writeback_tmp:%lukB"
5396 " all_unreclaimable? %s"
5399 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5400 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5401 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5402 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5403 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5404 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5405 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5406 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5407 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5408 K(node_page_state(pgdat, NR_WRITEBACK)),
5409 K(node_page_state(pgdat, NR_SHMEM)),
5410 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5411 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5412 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5414 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5416 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5417 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5421 for_each_populated_zone(zone) {
5424 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5428 for_each_online_cpu(cpu)
5429 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5438 " reserved_highatomic:%luKB"
5439 " active_anon:%lukB"
5440 " inactive_anon:%lukB"
5441 " active_file:%lukB"
5442 " inactive_file:%lukB"
5443 " unevictable:%lukB"
5444 " writepending:%lukB"
5448 " kernel_stack:%lukB"
5449 #ifdef CONFIG_SHADOW_CALL_STACK
5450 " shadow_call_stack:%lukB"
5459 K(zone_page_state(zone, NR_FREE_PAGES)),
5460 K(min_wmark_pages(zone)),
5461 K(low_wmark_pages(zone)),
5462 K(high_wmark_pages(zone)),
5463 K(zone->nr_reserved_highatomic),
5464 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5465 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5466 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5467 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5468 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5469 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5470 K(zone->present_pages),
5471 K(zone_managed_pages(zone)),
5472 K(zone_page_state(zone, NR_MLOCK)),
5473 zone_page_state(zone, NR_KERNEL_STACK_KB),
5474 #ifdef CONFIG_SHADOW_CALL_STACK
5475 zone_page_state(zone, NR_KERNEL_SCS_KB),
5477 K(zone_page_state(zone, NR_PAGETABLE)),
5478 K(zone_page_state(zone, NR_BOUNCE)),
5480 K(this_cpu_read(zone->pageset->pcp.count)),
5481 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5482 printk("lowmem_reserve[]:");
5483 for (i = 0; i < MAX_NR_ZONES; i++)
5484 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5485 printk(KERN_CONT "\n");
5488 for_each_populated_zone(zone) {
5490 unsigned long nr[MAX_ORDER], flags, total = 0;
5491 unsigned char types[MAX_ORDER];
5493 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5496 printk(KERN_CONT "%s: ", zone->name);
5498 spin_lock_irqsave(&zone->lock, flags);
5499 for (order = 0; order < MAX_ORDER; order++) {
5500 struct free_area *area = &zone->free_area[order];
5503 nr[order] = area->nr_free;
5504 total += nr[order] << order;
5507 for (type = 0; type < MIGRATE_TYPES; type++) {
5508 if (!free_area_empty(area, type))
5509 types[order] |= 1 << type;
5512 spin_unlock_irqrestore(&zone->lock, flags);
5513 for (order = 0; order < MAX_ORDER; order++) {
5514 printk(KERN_CONT "%lu*%lukB ",
5515 nr[order], K(1UL) << order);
5517 show_migration_types(types[order]);
5519 printk(KERN_CONT "= %lukB\n", K(total));
5522 hugetlb_show_meminfo();
5524 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5526 show_swap_cache_info();
5529 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5531 zoneref->zone = zone;
5532 zoneref->zone_idx = zone_idx(zone);
5536 * Builds allocation fallback zone lists.
5538 * Add all populated zones of a node to the zonelist.
5540 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5543 enum zone_type zone_type = MAX_NR_ZONES;
5548 zone = pgdat->node_zones + zone_type;
5549 if (managed_zone(zone)) {
5550 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5551 check_highest_zone(zone_type);
5553 } while (zone_type);
5560 static int __parse_numa_zonelist_order(char *s)
5563 * We used to support different zonlists modes but they turned
5564 * out to be just not useful. Let's keep the warning in place
5565 * if somebody still use the cmd line parameter so that we do
5566 * not fail it silently
5568 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5569 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5575 static __init int setup_numa_zonelist_order(char *s)
5580 return __parse_numa_zonelist_order(s);
5582 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5584 char numa_zonelist_order[] = "Node";
5587 * sysctl handler for numa_zonelist_order
5589 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5590 void __user *buffer, size_t *length,
5597 return proc_dostring(table, write, buffer, length, ppos);
5598 str = memdup_user_nul(buffer, 16);
5600 return PTR_ERR(str);
5602 ret = __parse_numa_zonelist_order(str);
5608 #define MAX_NODE_LOAD (nr_online_nodes)
5609 static int node_load[MAX_NUMNODES];
5612 * find_next_best_node - find the next node that should appear in a given node's fallback list
5613 * @node: node whose fallback list we're appending
5614 * @used_node_mask: nodemask_t of already used nodes
5616 * We use a number of factors to determine which is the next node that should
5617 * appear on a given node's fallback list. The node should not have appeared
5618 * already in @node's fallback list, and it should be the next closest node
5619 * according to the distance array (which contains arbitrary distance values
5620 * from each node to each node in the system), and should also prefer nodes
5621 * with no CPUs, since presumably they'll have very little allocation pressure
5622 * on them otherwise.
5624 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5626 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5629 int min_val = INT_MAX;
5630 int best_node = NUMA_NO_NODE;
5631 const struct cpumask *tmp = cpumask_of_node(0);
5633 /* Use the local node if we haven't already */
5634 if (!node_isset(node, *used_node_mask)) {
5635 node_set(node, *used_node_mask);
5639 for_each_node_state(n, N_MEMORY) {
5641 /* Don't want a node to appear more than once */
5642 if (node_isset(n, *used_node_mask))
5645 /* Use the distance array to find the distance */
5646 val = node_distance(node, n);
5648 /* Penalize nodes under us ("prefer the next node") */
5651 /* Give preference to headless and unused nodes */
5652 tmp = cpumask_of_node(n);
5653 if (!cpumask_empty(tmp))
5654 val += PENALTY_FOR_NODE_WITH_CPUS;
5656 /* Slight preference for less loaded node */
5657 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5658 val += node_load[n];
5660 if (val < min_val) {
5667 node_set(best_node, *used_node_mask);
5674 * Build zonelists ordered by node and zones within node.
5675 * This results in maximum locality--normal zone overflows into local
5676 * DMA zone, if any--but risks exhausting DMA zone.
5678 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5681 struct zoneref *zonerefs;
5684 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5686 for (i = 0; i < nr_nodes; i++) {
5689 pg_data_t *node = NODE_DATA(node_order[i]);
5691 nr_zones = build_zonerefs_node(node, zonerefs);
5692 zonerefs += nr_zones;
5694 zonerefs->zone = NULL;
5695 zonerefs->zone_idx = 0;
5699 * Build gfp_thisnode zonelists
5701 static void build_thisnode_zonelists(pg_data_t *pgdat)
5703 struct zoneref *zonerefs;
5706 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5707 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5708 zonerefs += nr_zones;
5709 zonerefs->zone = NULL;
5710 zonerefs->zone_idx = 0;
5714 * Build zonelists ordered by zone and nodes within zones.
5715 * This results in conserving DMA zone[s] until all Normal memory is
5716 * exhausted, but results in overflowing to remote node while memory
5717 * may still exist in local DMA zone.
5720 static void build_zonelists(pg_data_t *pgdat)
5722 static int node_order[MAX_NUMNODES];
5723 int node, load, nr_nodes = 0;
5724 nodemask_t used_mask = NODE_MASK_NONE;
5725 int local_node, prev_node;
5727 /* NUMA-aware ordering of nodes */
5728 local_node = pgdat->node_id;
5729 load = nr_online_nodes;
5730 prev_node = local_node;
5732 memset(node_order, 0, sizeof(node_order));
5733 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5735 * We don't want to pressure a particular node.
5736 * So adding penalty to the first node in same
5737 * distance group to make it round-robin.
5739 if (node_distance(local_node, node) !=
5740 node_distance(local_node, prev_node))
5741 node_load[node] = load;
5743 node_order[nr_nodes++] = node;
5748 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5749 build_thisnode_zonelists(pgdat);
5752 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5754 * Return node id of node used for "local" allocations.
5755 * I.e., first node id of first zone in arg node's generic zonelist.
5756 * Used for initializing percpu 'numa_mem', which is used primarily
5757 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5759 int local_memory_node(int node)
5763 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5764 gfp_zone(GFP_KERNEL),
5766 return zone_to_nid(z->zone);
5770 static void setup_min_unmapped_ratio(void);
5771 static void setup_min_slab_ratio(void);
5772 #else /* CONFIG_NUMA */
5774 static void build_zonelists(pg_data_t *pgdat)
5776 int node, local_node;
5777 struct zoneref *zonerefs;
5780 local_node = pgdat->node_id;
5782 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5783 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5784 zonerefs += nr_zones;
5787 * Now we build the zonelist so that it contains the zones
5788 * of all the other nodes.
5789 * We don't want to pressure a particular node, so when
5790 * building the zones for node N, we make sure that the
5791 * zones coming right after the local ones are those from
5792 * node N+1 (modulo N)
5794 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5795 if (!node_online(node))
5797 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5798 zonerefs += nr_zones;
5800 for (node = 0; node < local_node; node++) {
5801 if (!node_online(node))
5803 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5804 zonerefs += nr_zones;
5807 zonerefs->zone = NULL;
5808 zonerefs->zone_idx = 0;
5811 #endif /* CONFIG_NUMA */
5814 * Boot pageset table. One per cpu which is going to be used for all
5815 * zones and all nodes. The parameters will be set in such a way
5816 * that an item put on a list will immediately be handed over to
5817 * the buddy list. This is safe since pageset manipulation is done
5818 * with interrupts disabled.
5820 * The boot_pagesets must be kept even after bootup is complete for
5821 * unused processors and/or zones. They do play a role for bootstrapping
5822 * hotplugged processors.
5824 * zoneinfo_show() and maybe other functions do
5825 * not check if the processor is online before following the pageset pointer.
5826 * Other parts of the kernel may not check if the zone is available.
5828 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5829 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5830 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5832 static void __build_all_zonelists(void *data)
5835 int __maybe_unused cpu;
5836 pg_data_t *self = data;
5837 static DEFINE_SPINLOCK(lock);
5842 memset(node_load, 0, sizeof(node_load));
5846 * This node is hotadded and no memory is yet present. So just
5847 * building zonelists is fine - no need to touch other nodes.
5849 if (self && !node_online(self->node_id)) {
5850 build_zonelists(self);
5852 for_each_online_node(nid) {
5853 pg_data_t *pgdat = NODE_DATA(nid);
5855 build_zonelists(pgdat);
5858 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5860 * We now know the "local memory node" for each node--
5861 * i.e., the node of the first zone in the generic zonelist.
5862 * Set up numa_mem percpu variable for on-line cpus. During
5863 * boot, only the boot cpu should be on-line; we'll init the
5864 * secondary cpus' numa_mem as they come on-line. During
5865 * node/memory hotplug, we'll fixup all on-line cpus.
5867 for_each_online_cpu(cpu)
5868 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5875 static noinline void __init
5876 build_all_zonelists_init(void)
5880 __build_all_zonelists(NULL);
5883 * Initialize the boot_pagesets that are going to be used
5884 * for bootstrapping processors. The real pagesets for
5885 * each zone will be allocated later when the per cpu
5886 * allocator is available.
5888 * boot_pagesets are used also for bootstrapping offline
5889 * cpus if the system is already booted because the pagesets
5890 * are needed to initialize allocators on a specific cpu too.
5891 * F.e. the percpu allocator needs the page allocator which
5892 * needs the percpu allocator in order to allocate its pagesets
5893 * (a chicken-egg dilemma).
5895 for_each_possible_cpu(cpu)
5896 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5898 mminit_verify_zonelist();
5899 cpuset_init_current_mems_allowed();
5903 * unless system_state == SYSTEM_BOOTING.
5905 * __ref due to call of __init annotated helper build_all_zonelists_init
5906 * [protected by SYSTEM_BOOTING].
5908 void __ref build_all_zonelists(pg_data_t *pgdat)
5910 if (system_state == SYSTEM_BOOTING) {
5911 build_all_zonelists_init();
5913 __build_all_zonelists(pgdat);
5914 /* cpuset refresh routine should be here */
5916 vm_total_pages = nr_free_pagecache_pages();
5918 * Disable grouping by mobility if the number of pages in the
5919 * system is too low to allow the mechanism to work. It would be
5920 * more accurate, but expensive to check per-zone. This check is
5921 * made on memory-hotadd so a system can start with mobility
5922 * disabled and enable it later
5924 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5925 page_group_by_mobility_disabled = 1;
5927 page_group_by_mobility_disabled = 0;
5929 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5931 page_group_by_mobility_disabled ? "off" : "on",
5934 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5938 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5939 static bool __meminit
5940 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5942 static struct memblock_region *r;
5944 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5945 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5946 for_each_memblock(memory, r) {
5947 if (*pfn < memblock_region_memory_end_pfn(r))
5951 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5952 memblock_is_mirror(r)) {
5953 *pfn = memblock_region_memory_end_pfn(r);
5961 * Initially all pages are reserved - free ones are freed
5962 * up by memblock_free_all() once the early boot process is
5963 * done. Non-atomic initialization, single-pass.
5965 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5966 unsigned long start_pfn, enum memmap_context context,
5967 struct vmem_altmap *altmap)
5969 unsigned long pfn, end_pfn = start_pfn + size;
5972 if (highest_memmap_pfn < end_pfn - 1)
5973 highest_memmap_pfn = end_pfn - 1;
5975 #ifdef CONFIG_ZONE_DEVICE
5977 * Honor reservation requested by the driver for this ZONE_DEVICE
5978 * memory. We limit the total number of pages to initialize to just
5979 * those that might contain the memory mapping. We will defer the
5980 * ZONE_DEVICE page initialization until after we have released
5983 if (zone == ZONE_DEVICE) {
5987 if (start_pfn == altmap->base_pfn)
5988 start_pfn += altmap->reserve;
5989 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5993 for (pfn = start_pfn; pfn < end_pfn; ) {
5995 * There can be holes in boot-time mem_map[]s handed to this
5996 * function. They do not exist on hotplugged memory.
5998 if (context == MEMMAP_EARLY) {
5999 if (overlap_memmap_init(zone, &pfn))
6001 if (defer_init(nid, pfn, end_pfn))
6005 page = pfn_to_page(pfn);
6006 __init_single_page(page, pfn, zone, nid);
6007 if (context == MEMMAP_HOTPLUG)
6008 __SetPageReserved(page);
6011 * Mark the block movable so that blocks are reserved for
6012 * movable at startup. This will force kernel allocations
6013 * to reserve their blocks rather than leaking throughout
6014 * the address space during boot when many long-lived
6015 * kernel allocations are made.
6017 * bitmap is created for zone's valid pfn range. but memmap
6018 * can be created for invalid pages (for alignment)
6019 * check here not to call set_pageblock_migratetype() against
6022 if (!(pfn & (pageblock_nr_pages - 1))) {
6023 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6030 #ifdef CONFIG_ZONE_DEVICE
6031 void __ref memmap_init_zone_device(struct zone *zone,
6032 unsigned long start_pfn,
6033 unsigned long nr_pages,
6034 struct dev_pagemap *pgmap)
6036 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6037 struct pglist_data *pgdat = zone->zone_pgdat;
6038 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6039 unsigned long zone_idx = zone_idx(zone);
6040 unsigned long start = jiffies;
6041 int nid = pgdat->node_id;
6043 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6047 * The call to memmap_init_zone should have already taken care
6048 * of the pages reserved for the memmap, so we can just jump to
6049 * the end of that region and start processing the device pages.
6052 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6053 nr_pages = end_pfn - start_pfn;
6056 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6057 struct page *page = pfn_to_page(pfn);
6059 __init_single_page(page, pfn, zone_idx, nid);
6062 * Mark page reserved as it will need to wait for onlining
6063 * phase for it to be fully associated with a zone.
6065 * We can use the non-atomic __set_bit operation for setting
6066 * the flag as we are still initializing the pages.
6068 __SetPageReserved(page);
6071 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6072 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6073 * ever freed or placed on a driver-private list.
6075 page->pgmap = pgmap;
6076 page->zone_device_data = NULL;
6079 * Mark the block movable so that blocks are reserved for
6080 * movable at startup. This will force kernel allocations
6081 * to reserve their blocks rather than leaking throughout
6082 * the address space during boot when many long-lived
6083 * kernel allocations are made.
6085 * bitmap is created for zone's valid pfn range. but memmap
6086 * can be created for invalid pages (for alignment)
6087 * check here not to call set_pageblock_migratetype() against
6090 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6091 * because this is done early in section_activate()
6093 if (!(pfn & (pageblock_nr_pages - 1))) {
6094 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6099 pr_info("%s initialised %lu pages in %ums\n", __func__,
6100 nr_pages, jiffies_to_msecs(jiffies - start));
6104 static void __meminit zone_init_free_lists(struct zone *zone)
6106 unsigned int order, t;
6107 for_each_migratetype_order(order, t) {
6108 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6109 zone->free_area[order].nr_free = 0;
6113 void __meminit __weak memmap_init(unsigned long size, int nid,
6115 unsigned long range_start_pfn)
6117 unsigned long start_pfn, end_pfn;
6118 unsigned long range_end_pfn = range_start_pfn + size;
6121 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6122 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6123 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6125 if (end_pfn > start_pfn) {
6126 size = end_pfn - start_pfn;
6127 memmap_init_zone(size, nid, zone, start_pfn,
6128 MEMMAP_EARLY, NULL);
6133 static int zone_batchsize(struct zone *zone)
6139 * The per-cpu-pages pools are set to around 1000th of the
6142 batch = zone_managed_pages(zone) / 1024;
6143 /* But no more than a meg. */
6144 if (batch * PAGE_SIZE > 1024 * 1024)
6145 batch = (1024 * 1024) / PAGE_SIZE;
6146 batch /= 4; /* We effectively *= 4 below */
6151 * Clamp the batch to a 2^n - 1 value. Having a power
6152 * of 2 value was found to be more likely to have
6153 * suboptimal cache aliasing properties in some cases.
6155 * For example if 2 tasks are alternately allocating
6156 * batches of pages, one task can end up with a lot
6157 * of pages of one half of the possible page colors
6158 * and the other with pages of the other colors.
6160 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6165 /* The deferral and batching of frees should be suppressed under NOMMU
6168 * The problem is that NOMMU needs to be able to allocate large chunks
6169 * of contiguous memory as there's no hardware page translation to
6170 * assemble apparent contiguous memory from discontiguous pages.
6172 * Queueing large contiguous runs of pages for batching, however,
6173 * causes the pages to actually be freed in smaller chunks. As there
6174 * can be a significant delay between the individual batches being
6175 * recycled, this leads to the once large chunks of space being
6176 * fragmented and becoming unavailable for high-order allocations.
6183 * pcp->high and pcp->batch values are related and dependent on one another:
6184 * ->batch must never be higher then ->high.
6185 * The following function updates them in a safe manner without read side
6188 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6189 * those fields changing asynchronously (acording the the above rule).
6191 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6192 * outside of boot time (or some other assurance that no concurrent updaters
6195 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6196 unsigned long batch)
6198 /* start with a fail safe value for batch */
6202 /* Update high, then batch, in order */
6209 /* a companion to pageset_set_high() */
6210 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6212 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6215 static void pageset_init(struct per_cpu_pageset *p)
6217 struct per_cpu_pages *pcp;
6220 memset(p, 0, sizeof(*p));
6223 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6224 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6227 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6230 pageset_set_batch(p, batch);
6234 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6235 * to the value high for the pageset p.
6237 static void pageset_set_high(struct per_cpu_pageset *p,
6240 unsigned long batch = max(1UL, high / 4);
6241 if ((high / 4) > (PAGE_SHIFT * 8))
6242 batch = PAGE_SHIFT * 8;
6244 pageset_update(&p->pcp, high, batch);
6247 static void pageset_set_high_and_batch(struct zone *zone,
6248 struct per_cpu_pageset *pcp)
6250 if (percpu_pagelist_fraction)
6251 pageset_set_high(pcp,
6252 (zone_managed_pages(zone) /
6253 percpu_pagelist_fraction));
6255 pageset_set_batch(pcp, zone_batchsize(zone));
6258 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6260 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6263 pageset_set_high_and_batch(zone, pcp);
6266 void __meminit setup_zone_pageset(struct zone *zone)
6269 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6270 for_each_possible_cpu(cpu)
6271 zone_pageset_init(zone, cpu);
6275 * Allocate per cpu pagesets and initialize them.
6276 * Before this call only boot pagesets were available.
6278 void __init setup_per_cpu_pageset(void)
6280 struct pglist_data *pgdat;
6282 int __maybe_unused cpu;
6284 for_each_populated_zone(zone)
6285 setup_zone_pageset(zone);
6289 * Unpopulated zones continue using the boot pagesets.
6290 * The numa stats for these pagesets need to be reset.
6291 * Otherwise, they will end up skewing the stats of
6292 * the nodes these zones are associated with.
6294 for_each_possible_cpu(cpu) {
6295 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6296 memset(pcp->vm_numa_stat_diff, 0,
6297 sizeof(pcp->vm_numa_stat_diff));
6301 for_each_online_pgdat(pgdat)
6302 pgdat->per_cpu_nodestats =
6303 alloc_percpu(struct per_cpu_nodestat);
6306 static __meminit void zone_pcp_init(struct zone *zone)
6309 * per cpu subsystem is not up at this point. The following code
6310 * relies on the ability of the linker to provide the
6311 * offset of a (static) per cpu variable into the per cpu area.
6313 zone->pageset = &boot_pageset;
6315 if (populated_zone(zone))
6316 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6317 zone->name, zone->present_pages,
6318 zone_batchsize(zone));
6321 void __meminit init_currently_empty_zone(struct zone *zone,
6322 unsigned long zone_start_pfn,
6325 struct pglist_data *pgdat = zone->zone_pgdat;
6326 int zone_idx = zone_idx(zone) + 1;
6328 if (zone_idx > pgdat->nr_zones)
6329 pgdat->nr_zones = zone_idx;
6331 zone->zone_start_pfn = zone_start_pfn;
6333 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6334 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6336 (unsigned long)zone_idx(zone),
6337 zone_start_pfn, (zone_start_pfn + size));
6339 zone_init_free_lists(zone);
6340 zone->initialized = 1;
6344 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6345 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6347 * If an architecture guarantees that all ranges registered contain no holes and may
6348 * be freed, this function may be used instead of calling memory_present() manually.
6350 void __init sparse_memory_present_with_active_regions(int nid)
6352 unsigned long start_pfn, end_pfn;
6355 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6356 memory_present(this_nid, start_pfn, end_pfn);
6360 * get_pfn_range_for_nid - Return the start and end page frames for a node
6361 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6362 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6363 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6365 * It returns the start and end page frame of a node based on information
6366 * provided by memblock_set_node(). If called for a node
6367 * with no available memory, a warning is printed and the start and end
6370 void __init get_pfn_range_for_nid(unsigned int nid,
6371 unsigned long *start_pfn, unsigned long *end_pfn)
6373 unsigned long this_start_pfn, this_end_pfn;
6379 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6380 *start_pfn = min(*start_pfn, this_start_pfn);
6381 *end_pfn = max(*end_pfn, this_end_pfn);
6384 if (*start_pfn == -1UL)
6389 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6390 * assumption is made that zones within a node are ordered in monotonic
6391 * increasing memory addresses so that the "highest" populated zone is used
6393 static void __init find_usable_zone_for_movable(void)
6396 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6397 if (zone_index == ZONE_MOVABLE)
6400 if (arch_zone_highest_possible_pfn[zone_index] >
6401 arch_zone_lowest_possible_pfn[zone_index])
6405 VM_BUG_ON(zone_index == -1);
6406 movable_zone = zone_index;
6410 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6411 * because it is sized independent of architecture. Unlike the other zones,
6412 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6413 * in each node depending on the size of each node and how evenly kernelcore
6414 * is distributed. This helper function adjusts the zone ranges
6415 * provided by the architecture for a given node by using the end of the
6416 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6417 * zones within a node are in order of monotonic increases memory addresses
6419 static void __init adjust_zone_range_for_zone_movable(int nid,
6420 unsigned long zone_type,
6421 unsigned long node_start_pfn,
6422 unsigned long node_end_pfn,
6423 unsigned long *zone_start_pfn,
6424 unsigned long *zone_end_pfn)
6426 /* Only adjust if ZONE_MOVABLE is on this node */
6427 if (zone_movable_pfn[nid]) {
6428 /* Size ZONE_MOVABLE */
6429 if (zone_type == ZONE_MOVABLE) {
6430 *zone_start_pfn = zone_movable_pfn[nid];
6431 *zone_end_pfn = min(node_end_pfn,
6432 arch_zone_highest_possible_pfn[movable_zone]);
6434 /* Adjust for ZONE_MOVABLE starting within this range */
6435 } else if (!mirrored_kernelcore &&
6436 *zone_start_pfn < zone_movable_pfn[nid] &&
6437 *zone_end_pfn > zone_movable_pfn[nid]) {
6438 *zone_end_pfn = zone_movable_pfn[nid];
6440 /* Check if this whole range is within ZONE_MOVABLE */
6441 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6442 *zone_start_pfn = *zone_end_pfn;
6447 * Return the number of pages a zone spans in a node, including holes
6448 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6450 static unsigned long __init zone_spanned_pages_in_node(int nid,
6451 unsigned long zone_type,
6452 unsigned long node_start_pfn,
6453 unsigned long node_end_pfn,
6454 unsigned long *zone_start_pfn,
6455 unsigned long *zone_end_pfn)
6457 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6458 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6459 /* When hotadd a new node from cpu_up(), the node should be empty */
6460 if (!node_start_pfn && !node_end_pfn)
6463 /* Get the start and end of the zone */
6464 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6465 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6466 adjust_zone_range_for_zone_movable(nid, zone_type,
6467 node_start_pfn, node_end_pfn,
6468 zone_start_pfn, zone_end_pfn);
6470 /* Check that this node has pages within the zone's required range */
6471 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6474 /* Move the zone boundaries inside the node if necessary */
6475 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6476 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6478 /* Return the spanned pages */
6479 return *zone_end_pfn - *zone_start_pfn;
6483 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6484 * then all holes in the requested range will be accounted for.
6486 unsigned long __init __absent_pages_in_range(int nid,
6487 unsigned long range_start_pfn,
6488 unsigned long range_end_pfn)
6490 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6491 unsigned long start_pfn, end_pfn;
6494 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6495 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6496 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6497 nr_absent -= end_pfn - start_pfn;
6503 * absent_pages_in_range - Return number of page frames in holes within a range
6504 * @start_pfn: The start PFN to start searching for holes
6505 * @end_pfn: The end PFN to stop searching for holes
6507 * Return: the number of pages frames in memory holes within a range.
6509 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6510 unsigned long end_pfn)
6512 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6515 /* Return the number of page frames in holes in a zone on a node */
6516 static unsigned long __init zone_absent_pages_in_node(int nid,
6517 unsigned long zone_type,
6518 unsigned long node_start_pfn,
6519 unsigned long node_end_pfn)
6521 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6522 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6523 unsigned long zone_start_pfn, zone_end_pfn;
6524 unsigned long nr_absent;
6526 /* When hotadd a new node from cpu_up(), the node should be empty */
6527 if (!node_start_pfn && !node_end_pfn)
6530 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6531 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6533 adjust_zone_range_for_zone_movable(nid, zone_type,
6534 node_start_pfn, node_end_pfn,
6535 &zone_start_pfn, &zone_end_pfn);
6536 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6539 * ZONE_MOVABLE handling.
6540 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6543 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6544 unsigned long start_pfn, end_pfn;
6545 struct memblock_region *r;
6547 for_each_memblock(memory, r) {
6548 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6549 zone_start_pfn, zone_end_pfn);
6550 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6551 zone_start_pfn, zone_end_pfn);
6553 if (zone_type == ZONE_MOVABLE &&
6554 memblock_is_mirror(r))
6555 nr_absent += end_pfn - start_pfn;
6557 if (zone_type == ZONE_NORMAL &&
6558 !memblock_is_mirror(r))
6559 nr_absent += end_pfn - start_pfn;
6566 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6567 unsigned long node_start_pfn,
6568 unsigned long node_end_pfn)
6570 unsigned long realtotalpages = 0, totalpages = 0;
6573 for (i = 0; i < MAX_NR_ZONES; i++) {
6574 struct zone *zone = pgdat->node_zones + i;
6575 unsigned long zone_start_pfn, zone_end_pfn;
6576 unsigned long spanned, absent;
6577 unsigned long size, real_size;
6579 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6584 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6589 real_size = size - absent;
6592 zone->zone_start_pfn = zone_start_pfn;
6594 zone->zone_start_pfn = 0;
6595 zone->spanned_pages = size;
6596 zone->present_pages = real_size;
6599 realtotalpages += real_size;
6602 pgdat->node_spanned_pages = totalpages;
6603 pgdat->node_present_pages = realtotalpages;
6604 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6608 #ifndef CONFIG_SPARSEMEM
6610 * Calculate the size of the zone->blockflags rounded to an unsigned long
6611 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6612 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6613 * round what is now in bits to nearest long in bits, then return it in
6616 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6618 unsigned long usemapsize;
6620 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6621 usemapsize = roundup(zonesize, pageblock_nr_pages);
6622 usemapsize = usemapsize >> pageblock_order;
6623 usemapsize *= NR_PAGEBLOCK_BITS;
6624 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6626 return usemapsize / 8;
6629 static void __ref setup_usemap(struct pglist_data *pgdat,
6631 unsigned long zone_start_pfn,
6632 unsigned long zonesize)
6634 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6635 zone->pageblock_flags = NULL;
6637 zone->pageblock_flags =
6638 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6640 if (!zone->pageblock_flags)
6641 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6642 usemapsize, zone->name, pgdat->node_id);
6646 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6647 unsigned long zone_start_pfn, unsigned long zonesize) {}
6648 #endif /* CONFIG_SPARSEMEM */
6650 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6652 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6653 void __init set_pageblock_order(void)
6657 /* Check that pageblock_nr_pages has not already been setup */
6658 if (pageblock_order)
6661 if (HPAGE_SHIFT > PAGE_SHIFT)
6662 order = HUGETLB_PAGE_ORDER;
6664 order = MAX_ORDER - 1;
6667 * Assume the largest contiguous order of interest is a huge page.
6668 * This value may be variable depending on boot parameters on IA64 and
6671 pageblock_order = order;
6673 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6676 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6677 * is unused as pageblock_order is set at compile-time. See
6678 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6681 void __init set_pageblock_order(void)
6685 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6687 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6688 unsigned long present_pages)
6690 unsigned long pages = spanned_pages;
6693 * Provide a more accurate estimation if there are holes within
6694 * the zone and SPARSEMEM is in use. If there are holes within the
6695 * zone, each populated memory region may cost us one or two extra
6696 * memmap pages due to alignment because memmap pages for each
6697 * populated regions may not be naturally aligned on page boundary.
6698 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6700 if (spanned_pages > present_pages + (present_pages >> 4) &&
6701 IS_ENABLED(CONFIG_SPARSEMEM))
6702 pages = present_pages;
6704 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6707 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6708 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6710 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6712 spin_lock_init(&ds_queue->split_queue_lock);
6713 INIT_LIST_HEAD(&ds_queue->split_queue);
6714 ds_queue->split_queue_len = 0;
6717 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6720 #ifdef CONFIG_COMPACTION
6721 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6723 init_waitqueue_head(&pgdat->kcompactd_wait);
6726 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6729 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6731 pgdat_resize_init(pgdat);
6733 pgdat_init_split_queue(pgdat);
6734 pgdat_init_kcompactd(pgdat);
6736 init_waitqueue_head(&pgdat->kswapd_wait);
6737 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6739 pgdat_page_ext_init(pgdat);
6740 spin_lock_init(&pgdat->lru_lock);
6741 lruvec_init(&pgdat->__lruvec);
6744 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6745 unsigned long remaining_pages)
6747 atomic_long_set(&zone->managed_pages, remaining_pages);
6748 zone_set_nid(zone, nid);
6749 zone->name = zone_names[idx];
6750 zone->zone_pgdat = NODE_DATA(nid);
6751 spin_lock_init(&zone->lock);
6752 zone_seqlock_init(zone);
6753 zone_pcp_init(zone);
6757 * Set up the zone data structures
6758 * - init pgdat internals
6759 * - init all zones belonging to this node
6761 * NOTE: this function is only called during memory hotplug
6763 #ifdef CONFIG_MEMORY_HOTPLUG
6764 void __ref free_area_init_core_hotplug(int nid)
6767 pg_data_t *pgdat = NODE_DATA(nid);
6769 pgdat_init_internals(pgdat);
6770 for (z = 0; z < MAX_NR_ZONES; z++)
6771 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6776 * Set up the zone data structures:
6777 * - mark all pages reserved
6778 * - mark all memory queues empty
6779 * - clear the memory bitmaps
6781 * NOTE: pgdat should get zeroed by caller.
6782 * NOTE: this function is only called during early init.
6784 static void __init free_area_init_core(struct pglist_data *pgdat)
6787 int nid = pgdat->node_id;
6789 pgdat_init_internals(pgdat);
6790 pgdat->per_cpu_nodestats = &boot_nodestats;
6792 for (j = 0; j < MAX_NR_ZONES; j++) {
6793 struct zone *zone = pgdat->node_zones + j;
6794 unsigned long size, freesize, memmap_pages;
6795 unsigned long zone_start_pfn = zone->zone_start_pfn;
6797 size = zone->spanned_pages;
6798 freesize = zone->present_pages;
6801 * Adjust freesize so that it accounts for how much memory
6802 * is used by this zone for memmap. This affects the watermark
6803 * and per-cpu initialisations
6805 memmap_pages = calc_memmap_size(size, freesize);
6806 if (!is_highmem_idx(j)) {
6807 if (freesize >= memmap_pages) {
6808 freesize -= memmap_pages;
6811 " %s zone: %lu pages used for memmap\n",
6812 zone_names[j], memmap_pages);
6814 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6815 zone_names[j], memmap_pages, freesize);
6818 /* Account for reserved pages */
6819 if (j == 0 && freesize > dma_reserve) {
6820 freesize -= dma_reserve;
6821 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6822 zone_names[0], dma_reserve);
6825 if (!is_highmem_idx(j))
6826 nr_kernel_pages += freesize;
6827 /* Charge for highmem memmap if there are enough kernel pages */
6828 else if (nr_kernel_pages > memmap_pages * 2)
6829 nr_kernel_pages -= memmap_pages;
6830 nr_all_pages += freesize;
6833 * Set an approximate value for lowmem here, it will be adjusted
6834 * when the bootmem allocator frees pages into the buddy system.
6835 * And all highmem pages will be managed by the buddy system.
6837 zone_init_internals(zone, j, nid, freesize);
6842 set_pageblock_order();
6843 setup_usemap(pgdat, zone, zone_start_pfn, size);
6844 init_currently_empty_zone(zone, zone_start_pfn, size);
6845 memmap_init(size, nid, j, zone_start_pfn);
6849 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6850 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6852 unsigned long __maybe_unused start = 0;
6853 unsigned long __maybe_unused offset = 0;
6855 /* Skip empty nodes */
6856 if (!pgdat->node_spanned_pages)
6859 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6860 offset = pgdat->node_start_pfn - start;
6861 /* ia64 gets its own node_mem_map, before this, without bootmem */
6862 if (!pgdat->node_mem_map) {
6863 unsigned long size, end;
6867 * The zone's endpoints aren't required to be MAX_ORDER
6868 * aligned but the node_mem_map endpoints must be in order
6869 * for the buddy allocator to function correctly.
6871 end = pgdat_end_pfn(pgdat);
6872 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6873 size = (end - start) * sizeof(struct page);
6874 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6877 panic("Failed to allocate %ld bytes for node %d memory map\n",
6878 size, pgdat->node_id);
6879 pgdat->node_mem_map = map + offset;
6881 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6882 __func__, pgdat->node_id, (unsigned long)pgdat,
6883 (unsigned long)pgdat->node_mem_map);
6884 #ifndef CONFIG_NEED_MULTIPLE_NODES
6886 * With no DISCONTIG, the global mem_map is just set as node 0's
6888 if (pgdat == NODE_DATA(0)) {
6889 mem_map = NODE_DATA(0)->node_mem_map;
6890 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6896 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6897 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6899 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6900 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6902 pgdat->first_deferred_pfn = ULONG_MAX;
6905 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6908 static void __init free_area_init_node(int nid)
6910 pg_data_t *pgdat = NODE_DATA(nid);
6911 unsigned long start_pfn = 0;
6912 unsigned long end_pfn = 0;
6914 /* pg_data_t should be reset to zero when it's allocated */
6915 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6917 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6919 pgdat->node_id = nid;
6920 pgdat->node_start_pfn = start_pfn;
6921 pgdat->per_cpu_nodestats = NULL;
6923 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6924 (u64)start_pfn << PAGE_SHIFT,
6925 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6926 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6928 alloc_node_mem_map(pgdat);
6929 pgdat_set_deferred_range(pgdat);
6931 free_area_init_core(pgdat);
6934 void __init free_area_init_memoryless_node(int nid)
6936 free_area_init_node(nid);
6939 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6941 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6942 * PageReserved(). Return the number of struct pages that were initialized.
6944 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6949 for (pfn = spfn; pfn < epfn; pfn++) {
6950 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6951 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6952 + pageblock_nr_pages - 1;
6956 * Use a fake node/zone (0) for now. Some of these pages
6957 * (in memblock.reserved but not in memblock.memory) will
6958 * get re-initialized via reserve_bootmem_region() later.
6960 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6961 __SetPageReserved(pfn_to_page(pfn));
6969 * Only struct pages that are backed by physical memory are zeroed and
6970 * initialized by going through __init_single_page(). But, there are some
6971 * struct pages which are reserved in memblock allocator and their fields
6972 * may be accessed (for example page_to_pfn() on some configuration accesses
6973 * flags). We must explicitly initialize those struct pages.
6975 * This function also addresses a similar issue where struct pages are left
6976 * uninitialized because the physical address range is not covered by
6977 * memblock.memory or memblock.reserved. That could happen when memblock
6978 * layout is manually configured via memmap=, or when the highest physical
6979 * address (max_pfn) does not end on a section boundary.
6981 static void __init init_unavailable_mem(void)
6983 phys_addr_t start, end;
6985 phys_addr_t next = 0;
6988 * Loop through unavailable ranges not covered by memblock.memory.
6991 for_each_mem_range(i, &memblock.memory, NULL,
6992 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6994 pgcnt += init_unavailable_range(PFN_DOWN(next),
7000 * Early sections always have a fully populated memmap for the whole
7001 * section - see pfn_valid(). If the last section has holes at the
7002 * end and that section is marked "online", the memmap will be
7003 * considered initialized. Make sure that memmap has a well defined
7006 pgcnt += init_unavailable_range(PFN_DOWN(next),
7007 round_up(max_pfn, PAGES_PER_SECTION));
7010 * Struct pages that do not have backing memory. This could be because
7011 * firmware is using some of this memory, or for some other reasons.
7014 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7017 static inline void __init init_unavailable_mem(void)
7020 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7022 #if MAX_NUMNODES > 1
7024 * Figure out the number of possible node ids.
7026 void __init setup_nr_node_ids(void)
7028 unsigned int highest;
7030 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7031 nr_node_ids = highest + 1;
7036 * node_map_pfn_alignment - determine the maximum internode alignment
7038 * This function should be called after node map is populated and sorted.
7039 * It calculates the maximum power of two alignment which can distinguish
7042 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7043 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7044 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7045 * shifted, 1GiB is enough and this function will indicate so.
7047 * This is used to test whether pfn -> nid mapping of the chosen memory
7048 * model has fine enough granularity to avoid incorrect mapping for the
7049 * populated node map.
7051 * Return: the determined alignment in pfn's. 0 if there is no alignment
7052 * requirement (single node).
7054 unsigned long __init node_map_pfn_alignment(void)
7056 unsigned long accl_mask = 0, last_end = 0;
7057 unsigned long start, end, mask;
7058 int last_nid = NUMA_NO_NODE;
7061 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7062 if (!start || last_nid < 0 || last_nid == nid) {
7069 * Start with a mask granular enough to pin-point to the
7070 * start pfn and tick off bits one-by-one until it becomes
7071 * too coarse to separate the current node from the last.
7073 mask = ~((1 << __ffs(start)) - 1);
7074 while (mask && last_end <= (start & (mask << 1)))
7077 /* accumulate all internode masks */
7081 /* convert mask to number of pages */
7082 return ~accl_mask + 1;
7086 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7088 * Return: the minimum PFN based on information provided via
7089 * memblock_set_node().
7091 unsigned long __init find_min_pfn_with_active_regions(void)
7093 return PHYS_PFN(memblock_start_of_DRAM());
7097 * early_calculate_totalpages()
7098 * Sum pages in active regions for movable zone.
7099 * Populate N_MEMORY for calculating usable_nodes.
7101 static unsigned long __init early_calculate_totalpages(void)
7103 unsigned long totalpages = 0;
7104 unsigned long start_pfn, end_pfn;
7107 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7108 unsigned long pages = end_pfn - start_pfn;
7110 totalpages += pages;
7112 node_set_state(nid, N_MEMORY);
7118 * Find the PFN the Movable zone begins in each node. Kernel memory
7119 * is spread evenly between nodes as long as the nodes have enough
7120 * memory. When they don't, some nodes will have more kernelcore than
7123 static void __init find_zone_movable_pfns_for_nodes(void)
7126 unsigned long usable_startpfn;
7127 unsigned long kernelcore_node, kernelcore_remaining;
7128 /* save the state before borrow the nodemask */
7129 nodemask_t saved_node_state = node_states[N_MEMORY];
7130 unsigned long totalpages = early_calculate_totalpages();
7131 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7132 struct memblock_region *r;
7134 /* Need to find movable_zone earlier when movable_node is specified. */
7135 find_usable_zone_for_movable();
7138 * If movable_node is specified, ignore kernelcore and movablecore
7141 if (movable_node_is_enabled()) {
7142 for_each_memblock(memory, r) {
7143 if (!memblock_is_hotpluggable(r))
7146 nid = memblock_get_region_node(r);
7148 usable_startpfn = PFN_DOWN(r->base);
7149 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7150 min(usable_startpfn, zone_movable_pfn[nid]) :
7158 * If kernelcore=mirror is specified, ignore movablecore option
7160 if (mirrored_kernelcore) {
7161 bool mem_below_4gb_not_mirrored = false;
7163 for_each_memblock(memory, r) {
7164 if (memblock_is_mirror(r))
7167 nid = memblock_get_region_node(r);
7169 usable_startpfn = memblock_region_memory_base_pfn(r);
7171 if (usable_startpfn < 0x100000) {
7172 mem_below_4gb_not_mirrored = true;
7176 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7177 min(usable_startpfn, zone_movable_pfn[nid]) :
7181 if (mem_below_4gb_not_mirrored)
7182 pr_warn("This configuration results in unmirrored kernel memory.");
7188 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7189 * amount of necessary memory.
7191 if (required_kernelcore_percent)
7192 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7194 if (required_movablecore_percent)
7195 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7199 * If movablecore= was specified, calculate what size of
7200 * kernelcore that corresponds so that memory usable for
7201 * any allocation type is evenly spread. If both kernelcore
7202 * and movablecore are specified, then the value of kernelcore
7203 * will be used for required_kernelcore if it's greater than
7204 * what movablecore would have allowed.
7206 if (required_movablecore) {
7207 unsigned long corepages;
7210 * Round-up so that ZONE_MOVABLE is at least as large as what
7211 * was requested by the user
7213 required_movablecore =
7214 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7215 required_movablecore = min(totalpages, required_movablecore);
7216 corepages = totalpages - required_movablecore;
7218 required_kernelcore = max(required_kernelcore, corepages);
7222 * If kernelcore was not specified or kernelcore size is larger
7223 * than totalpages, there is no ZONE_MOVABLE.
7225 if (!required_kernelcore || required_kernelcore >= totalpages)
7228 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7229 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7232 /* Spread kernelcore memory as evenly as possible throughout nodes */
7233 kernelcore_node = required_kernelcore / usable_nodes;
7234 for_each_node_state(nid, N_MEMORY) {
7235 unsigned long start_pfn, end_pfn;
7238 * Recalculate kernelcore_node if the division per node
7239 * now exceeds what is necessary to satisfy the requested
7240 * amount of memory for the kernel
7242 if (required_kernelcore < kernelcore_node)
7243 kernelcore_node = required_kernelcore / usable_nodes;
7246 * As the map is walked, we track how much memory is usable
7247 * by the kernel using kernelcore_remaining. When it is
7248 * 0, the rest of the node is usable by ZONE_MOVABLE
7250 kernelcore_remaining = kernelcore_node;
7252 /* Go through each range of PFNs within this node */
7253 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7254 unsigned long size_pages;
7256 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7257 if (start_pfn >= end_pfn)
7260 /* Account for what is only usable for kernelcore */
7261 if (start_pfn < usable_startpfn) {
7262 unsigned long kernel_pages;
7263 kernel_pages = min(end_pfn, usable_startpfn)
7266 kernelcore_remaining -= min(kernel_pages,
7267 kernelcore_remaining);
7268 required_kernelcore -= min(kernel_pages,
7269 required_kernelcore);
7271 /* Continue if range is now fully accounted */
7272 if (end_pfn <= usable_startpfn) {
7275 * Push zone_movable_pfn to the end so
7276 * that if we have to rebalance
7277 * kernelcore across nodes, we will
7278 * not double account here
7280 zone_movable_pfn[nid] = end_pfn;
7283 start_pfn = usable_startpfn;
7287 * The usable PFN range for ZONE_MOVABLE is from
7288 * start_pfn->end_pfn. Calculate size_pages as the
7289 * number of pages used as kernelcore
7291 size_pages = end_pfn - start_pfn;
7292 if (size_pages > kernelcore_remaining)
7293 size_pages = kernelcore_remaining;
7294 zone_movable_pfn[nid] = start_pfn + size_pages;
7297 * Some kernelcore has been met, update counts and
7298 * break if the kernelcore for this node has been
7301 required_kernelcore -= min(required_kernelcore,
7303 kernelcore_remaining -= size_pages;
7304 if (!kernelcore_remaining)
7310 * If there is still required_kernelcore, we do another pass with one
7311 * less node in the count. This will push zone_movable_pfn[nid] further
7312 * along on the nodes that still have memory until kernelcore is
7316 if (usable_nodes && required_kernelcore > usable_nodes)
7320 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7321 for (nid = 0; nid < MAX_NUMNODES; nid++)
7322 zone_movable_pfn[nid] =
7323 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7326 /* restore the node_state */
7327 node_states[N_MEMORY] = saved_node_state;
7330 /* Any regular or high memory on that node ? */
7331 static void check_for_memory(pg_data_t *pgdat, int nid)
7333 enum zone_type zone_type;
7335 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7336 struct zone *zone = &pgdat->node_zones[zone_type];
7337 if (populated_zone(zone)) {
7338 if (IS_ENABLED(CONFIG_HIGHMEM))
7339 node_set_state(nid, N_HIGH_MEMORY);
7340 if (zone_type <= ZONE_NORMAL)
7341 node_set_state(nid, N_NORMAL_MEMORY);
7348 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7349 * such cases we allow max_zone_pfn sorted in the descending order
7351 bool __weak arch_has_descending_max_zone_pfns(void)
7357 * free_area_init - Initialise all pg_data_t and zone data
7358 * @max_zone_pfn: an array of max PFNs for each zone
7360 * This will call free_area_init_node() for each active node in the system.
7361 * Using the page ranges provided by memblock_set_node(), the size of each
7362 * zone in each node and their holes is calculated. If the maximum PFN
7363 * between two adjacent zones match, it is assumed that the zone is empty.
7364 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7365 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7366 * starts where the previous one ended. For example, ZONE_DMA32 starts
7367 * at arch_max_dma_pfn.
7369 void __init free_area_init(unsigned long *max_zone_pfn)
7371 unsigned long start_pfn, end_pfn;
7375 /* Record where the zone boundaries are */
7376 memset(arch_zone_lowest_possible_pfn, 0,
7377 sizeof(arch_zone_lowest_possible_pfn));
7378 memset(arch_zone_highest_possible_pfn, 0,
7379 sizeof(arch_zone_highest_possible_pfn));
7381 start_pfn = find_min_pfn_with_active_regions();
7382 descending = arch_has_descending_max_zone_pfns();
7384 for (i = 0; i < MAX_NR_ZONES; i++) {
7386 zone = MAX_NR_ZONES - i - 1;
7390 if (zone == ZONE_MOVABLE)
7393 end_pfn = max(max_zone_pfn[zone], start_pfn);
7394 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7395 arch_zone_highest_possible_pfn[zone] = end_pfn;
7397 start_pfn = end_pfn;
7400 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7401 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7402 find_zone_movable_pfns_for_nodes();
7404 /* Print out the zone ranges */
7405 pr_info("Zone ranges:\n");
7406 for (i = 0; i < MAX_NR_ZONES; i++) {
7407 if (i == ZONE_MOVABLE)
7409 pr_info(" %-8s ", zone_names[i]);
7410 if (arch_zone_lowest_possible_pfn[i] ==
7411 arch_zone_highest_possible_pfn[i])
7414 pr_cont("[mem %#018Lx-%#018Lx]\n",
7415 (u64)arch_zone_lowest_possible_pfn[i]
7417 ((u64)arch_zone_highest_possible_pfn[i]
7418 << PAGE_SHIFT) - 1);
7421 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7422 pr_info("Movable zone start for each node\n");
7423 for (i = 0; i < MAX_NUMNODES; i++) {
7424 if (zone_movable_pfn[i])
7425 pr_info(" Node %d: %#018Lx\n", i,
7426 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7430 * Print out the early node map, and initialize the
7431 * subsection-map relative to active online memory ranges to
7432 * enable future "sub-section" extensions of the memory map.
7434 pr_info("Early memory node ranges\n");
7435 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7436 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7437 (u64)start_pfn << PAGE_SHIFT,
7438 ((u64)end_pfn << PAGE_SHIFT) - 1);
7439 subsection_map_init(start_pfn, end_pfn - start_pfn);
7442 /* Initialise every node */
7443 mminit_verify_pageflags_layout();
7444 setup_nr_node_ids();
7445 init_unavailable_mem();
7446 for_each_online_node(nid) {
7447 pg_data_t *pgdat = NODE_DATA(nid);
7448 free_area_init_node(nid);
7450 /* Any memory on that node */
7451 if (pgdat->node_present_pages)
7452 node_set_state(nid, N_MEMORY);
7453 check_for_memory(pgdat, nid);
7457 static int __init cmdline_parse_core(char *p, unsigned long *core,
7458 unsigned long *percent)
7460 unsigned long long coremem;
7466 /* Value may be a percentage of total memory, otherwise bytes */
7467 coremem = simple_strtoull(p, &endptr, 0);
7468 if (*endptr == '%') {
7469 /* Paranoid check for percent values greater than 100 */
7470 WARN_ON(coremem > 100);
7474 coremem = memparse(p, &p);
7475 /* Paranoid check that UL is enough for the coremem value */
7476 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7478 *core = coremem >> PAGE_SHIFT;
7485 * kernelcore=size sets the amount of memory for use for allocations that
7486 * cannot be reclaimed or migrated.
7488 static int __init cmdline_parse_kernelcore(char *p)
7490 /* parse kernelcore=mirror */
7491 if (parse_option_str(p, "mirror")) {
7492 mirrored_kernelcore = true;
7496 return cmdline_parse_core(p, &required_kernelcore,
7497 &required_kernelcore_percent);
7501 * movablecore=size sets the amount of memory for use for allocations that
7502 * can be reclaimed or migrated.
7504 static int __init cmdline_parse_movablecore(char *p)
7506 return cmdline_parse_core(p, &required_movablecore,
7507 &required_movablecore_percent);
7510 early_param("kernelcore", cmdline_parse_kernelcore);
7511 early_param("movablecore", cmdline_parse_movablecore);
7513 void adjust_managed_page_count(struct page *page, long count)
7515 atomic_long_add(count, &page_zone(page)->managed_pages);
7516 totalram_pages_add(count);
7517 #ifdef CONFIG_HIGHMEM
7518 if (PageHighMem(page))
7519 totalhigh_pages_add(count);
7522 EXPORT_SYMBOL(adjust_managed_page_count);
7524 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7527 unsigned long pages = 0;
7529 start = (void *)PAGE_ALIGN((unsigned long)start);
7530 end = (void *)((unsigned long)end & PAGE_MASK);
7531 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7532 struct page *page = virt_to_page(pos);
7533 void *direct_map_addr;
7536 * 'direct_map_addr' might be different from 'pos'
7537 * because some architectures' virt_to_page()
7538 * work with aliases. Getting the direct map
7539 * address ensures that we get a _writeable_
7540 * alias for the memset().
7542 direct_map_addr = page_address(page);
7543 if ((unsigned int)poison <= 0xFF)
7544 memset(direct_map_addr, poison, PAGE_SIZE);
7546 free_reserved_page(page);
7550 pr_info("Freeing %s memory: %ldK\n",
7551 s, pages << (PAGE_SHIFT - 10));
7556 #ifdef CONFIG_HIGHMEM
7557 void free_highmem_page(struct page *page)
7559 __free_reserved_page(page);
7560 totalram_pages_inc();
7561 atomic_long_inc(&page_zone(page)->managed_pages);
7562 totalhigh_pages_inc();
7567 void __init mem_init_print_info(const char *str)
7569 unsigned long physpages, codesize, datasize, rosize, bss_size;
7570 unsigned long init_code_size, init_data_size;
7572 physpages = get_num_physpages();
7573 codesize = _etext - _stext;
7574 datasize = _edata - _sdata;
7575 rosize = __end_rodata - __start_rodata;
7576 bss_size = __bss_stop - __bss_start;
7577 init_data_size = __init_end - __init_begin;
7578 init_code_size = _einittext - _sinittext;
7581 * Detect special cases and adjust section sizes accordingly:
7582 * 1) .init.* may be embedded into .data sections
7583 * 2) .init.text.* may be out of [__init_begin, __init_end],
7584 * please refer to arch/tile/kernel/vmlinux.lds.S.
7585 * 3) .rodata.* may be embedded into .text or .data sections.
7587 #define adj_init_size(start, end, size, pos, adj) \
7589 if (start <= pos && pos < end && size > adj) \
7593 adj_init_size(__init_begin, __init_end, init_data_size,
7594 _sinittext, init_code_size);
7595 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7596 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7597 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7598 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7600 #undef adj_init_size
7602 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7603 #ifdef CONFIG_HIGHMEM
7607 nr_free_pages() << (PAGE_SHIFT - 10),
7608 physpages << (PAGE_SHIFT - 10),
7609 codesize >> 10, datasize >> 10, rosize >> 10,
7610 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7611 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7612 totalcma_pages << (PAGE_SHIFT - 10),
7613 #ifdef CONFIG_HIGHMEM
7614 totalhigh_pages() << (PAGE_SHIFT - 10),
7616 str ? ", " : "", str ? str : "");
7620 * set_dma_reserve - set the specified number of pages reserved in the first zone
7621 * @new_dma_reserve: The number of pages to mark reserved
7623 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7624 * In the DMA zone, a significant percentage may be consumed by kernel image
7625 * and other unfreeable allocations which can skew the watermarks badly. This
7626 * function may optionally be used to account for unfreeable pages in the
7627 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7628 * smaller per-cpu batchsize.
7630 void __init set_dma_reserve(unsigned long new_dma_reserve)
7632 dma_reserve = new_dma_reserve;
7635 static int page_alloc_cpu_dead(unsigned int cpu)
7638 lru_add_drain_cpu(cpu);
7642 * Spill the event counters of the dead processor
7643 * into the current processors event counters.
7644 * This artificially elevates the count of the current
7647 vm_events_fold_cpu(cpu);
7650 * Zero the differential counters of the dead processor
7651 * so that the vm statistics are consistent.
7653 * This is only okay since the processor is dead and cannot
7654 * race with what we are doing.
7656 cpu_vm_stats_fold(cpu);
7661 int hashdist = HASHDIST_DEFAULT;
7663 static int __init set_hashdist(char *str)
7667 hashdist = simple_strtoul(str, &str, 0);
7670 __setup("hashdist=", set_hashdist);
7673 void __init page_alloc_init(void)
7678 if (num_node_state(N_MEMORY) == 1)
7682 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7683 "mm/page_alloc:dead", NULL,
7684 page_alloc_cpu_dead);
7689 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7690 * or min_free_kbytes changes.
7692 static void calculate_totalreserve_pages(void)
7694 struct pglist_data *pgdat;
7695 unsigned long reserve_pages = 0;
7696 enum zone_type i, j;
7698 for_each_online_pgdat(pgdat) {
7700 pgdat->totalreserve_pages = 0;
7702 for (i = 0; i < MAX_NR_ZONES; i++) {
7703 struct zone *zone = pgdat->node_zones + i;
7705 unsigned long managed_pages = zone_managed_pages(zone);
7707 /* Find valid and maximum lowmem_reserve in the zone */
7708 for (j = i; j < MAX_NR_ZONES; j++) {
7709 if (zone->lowmem_reserve[j] > max)
7710 max = zone->lowmem_reserve[j];
7713 /* we treat the high watermark as reserved pages. */
7714 max += high_wmark_pages(zone);
7716 if (max > managed_pages)
7717 max = managed_pages;
7719 pgdat->totalreserve_pages += max;
7721 reserve_pages += max;
7724 totalreserve_pages = reserve_pages;
7728 * setup_per_zone_lowmem_reserve - called whenever
7729 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7730 * has a correct pages reserved value, so an adequate number of
7731 * pages are left in the zone after a successful __alloc_pages().
7733 static void setup_per_zone_lowmem_reserve(void)
7735 struct pglist_data *pgdat;
7736 enum zone_type j, idx;
7738 for_each_online_pgdat(pgdat) {
7739 for (j = 0; j < MAX_NR_ZONES; j++) {
7740 struct zone *zone = pgdat->node_zones + j;
7741 unsigned long managed_pages = zone_managed_pages(zone);
7743 zone->lowmem_reserve[j] = 0;
7747 struct zone *lower_zone;
7750 lower_zone = pgdat->node_zones + idx;
7752 if (!sysctl_lowmem_reserve_ratio[idx] ||
7753 !zone_managed_pages(lower_zone)) {
7754 lower_zone->lowmem_reserve[j] = 0;
7757 lower_zone->lowmem_reserve[j] =
7758 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7760 managed_pages += zone_managed_pages(lower_zone);
7765 /* update totalreserve_pages */
7766 calculate_totalreserve_pages();
7769 static void __setup_per_zone_wmarks(void)
7771 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7772 unsigned long lowmem_pages = 0;
7774 unsigned long flags;
7776 /* Calculate total number of !ZONE_HIGHMEM pages */
7777 for_each_zone(zone) {
7778 if (!is_highmem(zone))
7779 lowmem_pages += zone_managed_pages(zone);
7782 for_each_zone(zone) {
7785 spin_lock_irqsave(&zone->lock, flags);
7786 tmp = (u64)pages_min * zone_managed_pages(zone);
7787 do_div(tmp, lowmem_pages);
7788 if (is_highmem(zone)) {
7790 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7791 * need highmem pages, so cap pages_min to a small
7794 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7795 * deltas control async page reclaim, and so should
7796 * not be capped for highmem.
7798 unsigned long min_pages;
7800 min_pages = zone_managed_pages(zone) / 1024;
7801 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7802 zone->_watermark[WMARK_MIN] = min_pages;
7805 * If it's a lowmem zone, reserve a number of pages
7806 * proportionate to the zone's size.
7808 zone->_watermark[WMARK_MIN] = tmp;
7812 * Set the kswapd watermarks distance according to the
7813 * scale factor in proportion to available memory, but
7814 * ensure a minimum size on small systems.
7816 tmp = max_t(u64, tmp >> 2,
7817 mult_frac(zone_managed_pages(zone),
7818 watermark_scale_factor, 10000));
7820 zone->watermark_boost = 0;
7821 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7822 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7824 spin_unlock_irqrestore(&zone->lock, flags);
7827 /* update totalreserve_pages */
7828 calculate_totalreserve_pages();
7832 * setup_per_zone_wmarks - called when min_free_kbytes changes
7833 * or when memory is hot-{added|removed}
7835 * Ensures that the watermark[min,low,high] values for each zone are set
7836 * correctly with respect to min_free_kbytes.
7838 void setup_per_zone_wmarks(void)
7840 static DEFINE_SPINLOCK(lock);
7843 __setup_per_zone_wmarks();
7848 * Initialise min_free_kbytes.
7850 * For small machines we want it small (128k min). For large machines
7851 * we want it large (64MB max). But it is not linear, because network
7852 * bandwidth does not increase linearly with machine size. We use
7854 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7855 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7871 int __meminit init_per_zone_wmark_min(void)
7873 unsigned long lowmem_kbytes;
7874 int new_min_free_kbytes;
7876 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7877 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7879 if (new_min_free_kbytes > user_min_free_kbytes) {
7880 min_free_kbytes = new_min_free_kbytes;
7881 if (min_free_kbytes < 128)
7882 min_free_kbytes = 128;
7883 if (min_free_kbytes > 262144)
7884 min_free_kbytes = 262144;
7886 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7887 new_min_free_kbytes, user_min_free_kbytes);
7889 setup_per_zone_wmarks();
7890 refresh_zone_stat_thresholds();
7891 setup_per_zone_lowmem_reserve();
7894 setup_min_unmapped_ratio();
7895 setup_min_slab_ratio();
7900 core_initcall(init_per_zone_wmark_min)
7903 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7904 * that we can call two helper functions whenever min_free_kbytes
7907 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7908 void __user *buffer, size_t *length, loff_t *ppos)
7912 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7917 user_min_free_kbytes = min_free_kbytes;
7918 setup_per_zone_wmarks();
7923 int watermark_boost_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);
7935 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7936 void __user *buffer, size_t *length, loff_t *ppos)
7940 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7945 setup_per_zone_wmarks();
7951 static void setup_min_unmapped_ratio(void)
7956 for_each_online_pgdat(pgdat)
7957 pgdat->min_unmapped_pages = 0;
7960 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7961 sysctl_min_unmapped_ratio) / 100;
7965 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7966 void __user *buffer, size_t *length, loff_t *ppos)
7970 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7974 setup_min_unmapped_ratio();
7979 static void setup_min_slab_ratio(void)
7984 for_each_online_pgdat(pgdat)
7985 pgdat->min_slab_pages = 0;
7988 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7989 sysctl_min_slab_ratio) / 100;
7992 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7993 void __user *buffer, size_t *length, loff_t *ppos)
7997 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8001 setup_min_slab_ratio();
8008 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8009 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8010 * whenever sysctl_lowmem_reserve_ratio changes.
8012 * The reserve ratio obviously has absolutely no relation with the
8013 * minimum watermarks. The lowmem reserve ratio can only make sense
8014 * if in function of the boot time zone sizes.
8016 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8017 void __user *buffer, size_t *length, loff_t *ppos)
8021 proc_dointvec_minmax(table, write, buffer, length, ppos);
8023 for (i = 0; i < MAX_NR_ZONES; i++) {
8024 if (sysctl_lowmem_reserve_ratio[i] < 1)
8025 sysctl_lowmem_reserve_ratio[i] = 0;
8028 setup_per_zone_lowmem_reserve();
8032 static void __zone_pcp_update(struct zone *zone)
8036 for_each_possible_cpu(cpu)
8037 pageset_set_high_and_batch(zone,
8038 per_cpu_ptr(zone->pageset, cpu));
8042 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8043 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8044 * pagelist can have before it gets flushed back to buddy allocator.
8046 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8047 void __user *buffer, size_t *length, loff_t *ppos)
8050 int old_percpu_pagelist_fraction;
8053 mutex_lock(&pcp_batch_high_lock);
8054 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8056 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8057 if (!write || ret < 0)
8060 /* Sanity checking to avoid pcp imbalance */
8061 if (percpu_pagelist_fraction &&
8062 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8063 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8069 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8072 for_each_populated_zone(zone)
8073 __zone_pcp_update(zone);
8075 mutex_unlock(&pcp_batch_high_lock);
8079 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8081 * Returns the number of pages that arch has reserved but
8082 * is not known to alloc_large_system_hash().
8084 static unsigned long __init arch_reserved_kernel_pages(void)
8091 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8092 * machines. As memory size is increased the scale is also increased but at
8093 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8094 * quadruples the scale is increased by one, which means the size of hash table
8095 * only doubles, instead of quadrupling as well.
8096 * Because 32-bit systems cannot have large physical memory, where this scaling
8097 * makes sense, it is disabled on such platforms.
8099 #if __BITS_PER_LONG > 32
8100 #define ADAPT_SCALE_BASE (64ul << 30)
8101 #define ADAPT_SCALE_SHIFT 2
8102 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8106 * allocate a large system hash table from bootmem
8107 * - it is assumed that the hash table must contain an exact power-of-2
8108 * quantity of entries
8109 * - limit is the number of hash buckets, not the total allocation size
8111 void *__init alloc_large_system_hash(const char *tablename,
8112 unsigned long bucketsize,
8113 unsigned long numentries,
8116 unsigned int *_hash_shift,
8117 unsigned int *_hash_mask,
8118 unsigned long low_limit,
8119 unsigned long high_limit)
8121 unsigned long long max = high_limit;
8122 unsigned long log2qty, size;
8127 /* allow the kernel cmdline to have a say */
8129 /* round applicable memory size up to nearest megabyte */
8130 numentries = nr_kernel_pages;
8131 numentries -= arch_reserved_kernel_pages();
8133 /* It isn't necessary when PAGE_SIZE >= 1MB */
8134 if (PAGE_SHIFT < 20)
8135 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8137 #if __BITS_PER_LONG > 32
8139 unsigned long adapt;
8141 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8142 adapt <<= ADAPT_SCALE_SHIFT)
8147 /* limit to 1 bucket per 2^scale bytes of low memory */
8148 if (scale > PAGE_SHIFT)
8149 numentries >>= (scale - PAGE_SHIFT);
8151 numentries <<= (PAGE_SHIFT - scale);
8153 /* Make sure we've got at least a 0-order allocation.. */
8154 if (unlikely(flags & HASH_SMALL)) {
8155 /* Makes no sense without HASH_EARLY */
8156 WARN_ON(!(flags & HASH_EARLY));
8157 if (!(numentries >> *_hash_shift)) {
8158 numentries = 1UL << *_hash_shift;
8159 BUG_ON(!numentries);
8161 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8162 numentries = PAGE_SIZE / bucketsize;
8164 numentries = roundup_pow_of_two(numentries);
8166 /* limit allocation size to 1/16 total memory by default */
8168 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8169 do_div(max, bucketsize);
8171 max = min(max, 0x80000000ULL);
8173 if (numentries < low_limit)
8174 numentries = low_limit;
8175 if (numentries > max)
8178 log2qty = ilog2(numentries);
8180 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8183 size = bucketsize << log2qty;
8184 if (flags & HASH_EARLY) {
8185 if (flags & HASH_ZERO)
8186 table = memblock_alloc(size, SMP_CACHE_BYTES);
8188 table = memblock_alloc_raw(size,
8190 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8191 table = __vmalloc(size, gfp_flags);
8195 * If bucketsize is not a power-of-two, we may free
8196 * some pages at the end of hash table which
8197 * alloc_pages_exact() automatically does
8199 table = alloc_pages_exact(size, gfp_flags);
8200 kmemleak_alloc(table, size, 1, gfp_flags);
8202 } while (!table && size > PAGE_SIZE && --log2qty);
8205 panic("Failed to allocate %s hash table\n", tablename);
8207 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8208 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8209 virt ? "vmalloc" : "linear");
8212 *_hash_shift = log2qty;
8214 *_hash_mask = (1 << log2qty) - 1;
8220 * This function checks whether pageblock includes unmovable pages or not.
8222 * PageLRU check without isolation or lru_lock could race so that
8223 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8224 * check without lock_page also may miss some movable non-lru pages at
8225 * race condition. So you can't expect this function should be exact.
8227 * Returns a page without holding a reference. If the caller wants to
8228 * dereference that page (e.g., dumping), it has to make sure that that it
8229 * cannot get removed (e.g., via memory unplug) concurrently.
8232 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8233 int migratetype, int flags)
8235 unsigned long iter = 0;
8236 unsigned long pfn = page_to_pfn(page);
8239 * TODO we could make this much more efficient by not checking every
8240 * page in the range if we know all of them are in MOVABLE_ZONE and
8241 * that the movable zone guarantees that pages are migratable but
8242 * the later is not the case right now unfortunatelly. E.g. movablecore
8243 * can still lead to having bootmem allocations in zone_movable.
8246 if (is_migrate_cma_page(page)) {
8248 * CMA allocations (alloc_contig_range) really need to mark
8249 * isolate CMA pageblocks even when they are not movable in fact
8250 * so consider them movable here.
8252 if (is_migrate_cma(migratetype))
8258 for (; iter < pageblock_nr_pages; iter++) {
8259 if (!pfn_valid_within(pfn + iter))
8262 page = pfn_to_page(pfn + iter);
8264 if (PageReserved(page))
8268 * If the zone is movable and we have ruled out all reserved
8269 * pages then it should be reasonably safe to assume the rest
8272 if (zone_idx(zone) == ZONE_MOVABLE)
8276 * Hugepages are not in LRU lists, but they're movable.
8277 * THPs are on the LRU, but need to be counted as #small pages.
8278 * We need not scan over tail pages because we don't
8279 * handle each tail page individually in migration.
8281 if (PageHuge(page) || PageTransCompound(page)) {
8282 struct page *head = compound_head(page);
8283 unsigned int skip_pages;
8285 if (PageHuge(page)) {
8286 if (!hugepage_migration_supported(page_hstate(head)))
8288 } else if (!PageLRU(head) && !__PageMovable(head)) {
8292 skip_pages = compound_nr(head) - (page - head);
8293 iter += skip_pages - 1;
8298 * We can't use page_count without pin a page
8299 * because another CPU can free compound page.
8300 * This check already skips compound tails of THP
8301 * because their page->_refcount is zero at all time.
8303 if (!page_ref_count(page)) {
8304 if (PageBuddy(page))
8305 iter += (1 << page_order(page)) - 1;
8310 * The HWPoisoned page may be not in buddy system, and
8311 * page_count() is not 0.
8313 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8316 if (__PageMovable(page) || PageLRU(page))
8320 * If there are RECLAIMABLE pages, we need to check
8321 * it. But now, memory offline itself doesn't call
8322 * shrink_node_slabs() and it still to be fixed.
8325 * If the page is not RAM, page_count()should be 0.
8326 * we don't need more check. This is an _used_ not-movable page.
8328 * The problematic thing here is PG_reserved pages. PG_reserved
8329 * is set to both of a memory hole page and a _used_ kernel
8337 #ifdef CONFIG_CONTIG_ALLOC
8338 static unsigned long pfn_max_align_down(unsigned long pfn)
8340 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8341 pageblock_nr_pages) - 1);
8344 static unsigned long pfn_max_align_up(unsigned long pfn)
8346 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8347 pageblock_nr_pages));
8350 /* [start, end) must belong to a single zone. */
8351 static int __alloc_contig_migrate_range(struct compact_control *cc,
8352 unsigned long start, unsigned long end)
8354 /* This function is based on compact_zone() from compaction.c. */
8355 unsigned long nr_reclaimed;
8356 unsigned long pfn = start;
8357 unsigned int tries = 0;
8362 while (pfn < end || !list_empty(&cc->migratepages)) {
8363 if (fatal_signal_pending(current)) {
8368 if (list_empty(&cc->migratepages)) {
8369 cc->nr_migratepages = 0;
8370 pfn = isolate_migratepages_range(cc, pfn, end);
8376 } else if (++tries == 5) {
8377 ret = ret < 0 ? ret : -EBUSY;
8381 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8383 cc->nr_migratepages -= nr_reclaimed;
8385 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8386 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8389 putback_movable_pages(&cc->migratepages);
8396 * alloc_contig_range() -- tries to allocate given range of pages
8397 * @start: start PFN to allocate
8398 * @end: one-past-the-last PFN to allocate
8399 * @migratetype: migratetype of the underlaying pageblocks (either
8400 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8401 * in range must have the same migratetype and it must
8402 * be either of the two.
8403 * @gfp_mask: GFP mask to use during compaction
8405 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8406 * aligned. The PFN range must belong to a single zone.
8408 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8409 * pageblocks in the range. Once isolated, the pageblocks should not
8410 * be modified by others.
8412 * Return: zero on success or negative error code. On success all
8413 * pages which PFN is in [start, end) are allocated for the caller and
8414 * need to be freed with free_contig_range().
8416 int alloc_contig_range(unsigned long start, unsigned long end,
8417 unsigned migratetype, gfp_t gfp_mask)
8419 unsigned long outer_start, outer_end;
8423 struct compact_control cc = {
8424 .nr_migratepages = 0,
8426 .zone = page_zone(pfn_to_page(start)),
8427 .mode = MIGRATE_SYNC,
8428 .ignore_skip_hint = true,
8429 .no_set_skip_hint = true,
8430 .gfp_mask = current_gfp_context(gfp_mask),
8431 .alloc_contig = true,
8433 INIT_LIST_HEAD(&cc.migratepages);
8436 * What we do here is we mark all pageblocks in range as
8437 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8438 * have different sizes, and due to the way page allocator
8439 * work, we align the range to biggest of the two pages so
8440 * that page allocator won't try to merge buddies from
8441 * different pageblocks and change MIGRATE_ISOLATE to some
8442 * other migration type.
8444 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8445 * migrate the pages from an unaligned range (ie. pages that
8446 * we are interested in). This will put all the pages in
8447 * range back to page allocator as MIGRATE_ISOLATE.
8449 * When this is done, we take the pages in range from page
8450 * allocator removing them from the buddy system. This way
8451 * page allocator will never consider using them.
8453 * This lets us mark the pageblocks back as
8454 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8455 * aligned range but not in the unaligned, original range are
8456 * put back to page allocator so that buddy can use them.
8459 ret = start_isolate_page_range(pfn_max_align_down(start),
8460 pfn_max_align_up(end), migratetype, 0);
8465 * In case of -EBUSY, we'd like to know which page causes problem.
8466 * So, just fall through. test_pages_isolated() has a tracepoint
8467 * which will report the busy page.
8469 * It is possible that busy pages could become available before
8470 * the call to test_pages_isolated, and the range will actually be
8471 * allocated. So, if we fall through be sure to clear ret so that
8472 * -EBUSY is not accidentally used or returned to caller.
8474 ret = __alloc_contig_migrate_range(&cc, start, end);
8475 if (ret && ret != -EBUSY)
8480 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8481 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8482 * more, all pages in [start, end) are free in page allocator.
8483 * What we are going to do is to allocate all pages from
8484 * [start, end) (that is remove them from page allocator).
8486 * The only problem is that pages at the beginning and at the
8487 * end of interesting range may be not aligned with pages that
8488 * page allocator holds, ie. they can be part of higher order
8489 * pages. Because of this, we reserve the bigger range and
8490 * once this is done free the pages we are not interested in.
8492 * We don't have to hold zone->lock here because the pages are
8493 * isolated thus they won't get removed from buddy.
8496 lru_add_drain_all();
8499 outer_start = start;
8500 while (!PageBuddy(pfn_to_page(outer_start))) {
8501 if (++order >= MAX_ORDER) {
8502 outer_start = start;
8505 outer_start &= ~0UL << order;
8508 if (outer_start != start) {
8509 order = page_order(pfn_to_page(outer_start));
8512 * outer_start page could be small order buddy page and
8513 * it doesn't include start page. Adjust outer_start
8514 * in this case to report failed page properly
8515 * on tracepoint in test_pages_isolated()
8517 if (outer_start + (1UL << order) <= start)
8518 outer_start = start;
8521 /* Make sure the range is really isolated. */
8522 if (test_pages_isolated(outer_start, end, 0)) {
8523 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8524 __func__, outer_start, end);
8529 /* Grab isolated pages from freelists. */
8530 outer_end = isolate_freepages_range(&cc, outer_start, end);
8536 /* Free head and tail (if any) */
8537 if (start != outer_start)
8538 free_contig_range(outer_start, start - outer_start);
8539 if (end != outer_end)
8540 free_contig_range(end, outer_end - end);
8543 undo_isolate_page_range(pfn_max_align_down(start),
8544 pfn_max_align_up(end), migratetype);
8548 static int __alloc_contig_pages(unsigned long start_pfn,
8549 unsigned long nr_pages, gfp_t gfp_mask)
8551 unsigned long end_pfn = start_pfn + nr_pages;
8553 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8557 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8558 unsigned long nr_pages)
8560 unsigned long i, end_pfn = start_pfn + nr_pages;
8563 for (i = start_pfn; i < end_pfn; i++) {
8564 page = pfn_to_online_page(i);
8568 if (page_zone(page) != z)
8571 if (PageReserved(page))
8574 if (page_count(page) > 0)
8583 static bool zone_spans_last_pfn(const struct zone *zone,
8584 unsigned long start_pfn, unsigned long nr_pages)
8586 unsigned long last_pfn = start_pfn + nr_pages - 1;
8588 return zone_spans_pfn(zone, last_pfn);
8592 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8593 * @nr_pages: Number of contiguous pages to allocate
8594 * @gfp_mask: GFP mask to limit search and used during compaction
8596 * @nodemask: Mask for other possible nodes
8598 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8599 * on an applicable zonelist to find a contiguous pfn range which can then be
8600 * tried for allocation with alloc_contig_range(). This routine is intended
8601 * for allocation requests which can not be fulfilled with the buddy allocator.
8603 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8604 * power of two then the alignment is guaranteed to be to the given nr_pages
8605 * (e.g. 1GB request would be aligned to 1GB).
8607 * Allocated pages can be freed with free_contig_range() or by manually calling
8608 * __free_page() on each allocated page.
8610 * Return: pointer to contiguous pages on success, or NULL if not successful.
8612 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8613 int nid, nodemask_t *nodemask)
8615 unsigned long ret, pfn, flags;
8616 struct zonelist *zonelist;
8620 zonelist = node_zonelist(nid, gfp_mask);
8621 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8622 gfp_zone(gfp_mask), nodemask) {
8623 spin_lock_irqsave(&zone->lock, flags);
8625 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8626 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8627 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8629 * We release the zone lock here because
8630 * alloc_contig_range() will also lock the zone
8631 * at some point. If there's an allocation
8632 * spinning on this lock, it may win the race
8633 * and cause alloc_contig_range() to fail...
8635 spin_unlock_irqrestore(&zone->lock, flags);
8636 ret = __alloc_contig_pages(pfn, nr_pages,
8639 return pfn_to_page(pfn);
8640 spin_lock_irqsave(&zone->lock, flags);
8644 spin_unlock_irqrestore(&zone->lock, flags);
8648 #endif /* CONFIG_CONTIG_ALLOC */
8650 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8652 unsigned int count = 0;
8654 for (; nr_pages--; pfn++) {
8655 struct page *page = pfn_to_page(pfn);
8657 count += page_count(page) != 1;
8660 WARN(count != 0, "%d pages are still in use!\n", count);
8664 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8665 * page high values need to be recalulated.
8667 void __meminit zone_pcp_update(struct zone *zone)
8669 mutex_lock(&pcp_batch_high_lock);
8670 __zone_pcp_update(zone);
8671 mutex_unlock(&pcp_batch_high_lock);
8674 void zone_pcp_reset(struct zone *zone)
8676 unsigned long flags;
8678 struct per_cpu_pageset *pset;
8680 /* avoid races with drain_pages() */
8681 local_irq_save(flags);
8682 if (zone->pageset != &boot_pageset) {
8683 for_each_online_cpu(cpu) {
8684 pset = per_cpu_ptr(zone->pageset, cpu);
8685 drain_zonestat(zone, pset);
8687 free_percpu(zone->pageset);
8688 zone->pageset = &boot_pageset;
8690 local_irq_restore(flags);
8693 #ifdef CONFIG_MEMORY_HOTREMOVE
8695 * All pages in the range must be in a single zone and isolated
8696 * before calling this.
8699 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8705 unsigned long flags;
8706 unsigned long offlined_pages = 0;
8708 /* find the first valid pfn */
8709 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8713 return offlined_pages;
8715 offline_mem_sections(pfn, end_pfn);
8716 zone = page_zone(pfn_to_page(pfn));
8717 spin_lock_irqsave(&zone->lock, flags);
8719 while (pfn < end_pfn) {
8720 if (!pfn_valid(pfn)) {
8724 page = pfn_to_page(pfn);
8726 * The HWPoisoned page may be not in buddy system, and
8727 * page_count() is not 0.
8729 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8735 BUG_ON(page_count(page));
8736 BUG_ON(!PageBuddy(page));
8737 order = page_order(page);
8738 offlined_pages += 1 << order;
8739 del_page_from_free_list(page, zone, order);
8740 pfn += (1 << order);
8742 spin_unlock_irqrestore(&zone->lock, flags);
8744 return offlined_pages;
8748 bool is_free_buddy_page(struct page *page)
8750 struct zone *zone = page_zone(page);
8751 unsigned long pfn = page_to_pfn(page);
8752 unsigned long flags;
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)
8762 spin_unlock_irqrestore(&zone->lock, flags);
8764 return order < MAX_ORDER;
8767 #ifdef CONFIG_MEMORY_FAILURE
8769 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8770 * test is performed under the zone lock to prevent a race against page
8773 bool set_hwpoison_free_buddy_page(struct page *page)
8775 struct zone *zone = page_zone(page);
8776 unsigned long pfn = page_to_pfn(page);
8777 unsigned long flags;
8779 bool hwpoisoned = false;
8781 spin_lock_irqsave(&zone->lock, flags);
8782 for (order = 0; order < MAX_ORDER; order++) {
8783 struct page *page_head = page - (pfn & ((1 << order) - 1));
8785 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8786 if (!TestSetPageHWPoison(page))
8791 spin_unlock_irqrestore(&zone->lock, flags);