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);
463 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
464 #endif /* CONFIG_SPARSEMEM */
465 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
469 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
470 * @page: The page within the block of interest
471 * @pfn: The target page frame number
472 * @end_bitidx: The last bit of interest to retrieve
473 * @mask: mask of bits that the caller is interested in
475 * Return: pageblock_bits flags
477 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
479 unsigned long end_bitidx,
482 unsigned long *bitmap;
483 unsigned long bitidx, word_bitidx;
486 bitmap = get_pageblock_bitmap(page, pfn);
487 bitidx = pfn_to_bitidx(page, pfn);
488 word_bitidx = bitidx / BITS_PER_LONG;
489 bitidx &= (BITS_PER_LONG-1);
491 word = bitmap[word_bitidx];
492 bitidx += end_bitidx;
493 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
496 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
497 unsigned long end_bitidx,
500 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
503 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
505 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
509 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
510 * @page: The page within the block of interest
511 * @flags: The flags to set
512 * @pfn: The target page frame number
513 * @end_bitidx: The last bit of interest
514 * @mask: mask of bits that the caller is interested in
516 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
518 unsigned long end_bitidx,
521 unsigned long *bitmap;
522 unsigned long bitidx, word_bitidx;
523 unsigned long old_word, word;
525 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
526 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
528 bitmap = get_pageblock_bitmap(page, pfn);
529 bitidx = pfn_to_bitidx(page, pfn);
530 word_bitidx = bitidx / BITS_PER_LONG;
531 bitidx &= (BITS_PER_LONG-1);
533 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
535 bitidx += end_bitidx;
536 mask <<= (BITS_PER_LONG - bitidx - 1);
537 flags <<= (BITS_PER_LONG - bitidx - 1);
539 word = READ_ONCE(bitmap[word_bitidx]);
541 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
542 if (word == old_word)
548 void set_pageblock_migratetype(struct page *page, int migratetype)
550 if (unlikely(page_group_by_mobility_disabled &&
551 migratetype < MIGRATE_PCPTYPES))
552 migratetype = MIGRATE_UNMOVABLE;
554 set_pageblock_flags_group(page, (unsigned long)migratetype,
555 PB_migrate, PB_migrate_end);
558 #ifdef CONFIG_DEBUG_VM
559 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
563 unsigned long pfn = page_to_pfn(page);
564 unsigned long sp, start_pfn;
567 seq = zone_span_seqbegin(zone);
568 start_pfn = zone->zone_start_pfn;
569 sp = zone->spanned_pages;
570 if (!zone_spans_pfn(zone, pfn))
572 } while (zone_span_seqretry(zone, seq));
575 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
576 pfn, zone_to_nid(zone), zone->name,
577 start_pfn, start_pfn + sp);
582 static int page_is_consistent(struct zone *zone, struct page *page)
584 if (!pfn_valid_within(page_to_pfn(page)))
586 if (zone != page_zone(page))
592 * Temporary debugging check for pages not lying within a given zone.
594 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
596 if (page_outside_zone_boundaries(zone, page))
598 if (!page_is_consistent(zone, page))
604 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
610 static void bad_page(struct page *page, const char *reason)
612 static unsigned long resume;
613 static unsigned long nr_shown;
614 static unsigned long nr_unshown;
617 * Allow a burst of 60 reports, then keep quiet for that minute;
618 * or allow a steady drip of one report per second.
620 if (nr_shown == 60) {
621 if (time_before(jiffies, resume)) {
627 "BUG: Bad page state: %lu messages suppressed\n",
634 resume = jiffies + 60 * HZ;
636 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
637 current->comm, page_to_pfn(page));
638 __dump_page(page, reason);
639 dump_page_owner(page);
644 /* Leave bad fields for debug, except PageBuddy could make trouble */
645 page_mapcount_reset(page); /* remove PageBuddy */
646 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
650 * Higher-order pages are called "compound pages". They are structured thusly:
652 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
654 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
655 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
657 * The first tail page's ->compound_dtor holds the offset in array of compound
658 * page destructors. See compound_page_dtors.
660 * The first tail page's ->compound_order holds the order of allocation.
661 * This usage means that zero-order pages may not be compound.
664 void free_compound_page(struct page *page)
666 mem_cgroup_uncharge(page);
667 __free_pages_ok(page, compound_order(page));
670 void prep_compound_page(struct page *page, unsigned int order)
673 int nr_pages = 1 << order;
675 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
676 set_compound_order(page, order);
678 for (i = 1; i < nr_pages; i++) {
679 struct page *p = page + i;
680 set_page_count(p, 0);
681 p->mapping = TAIL_MAPPING;
682 set_compound_head(p, page);
684 atomic_set(compound_mapcount_ptr(page), -1);
685 if (hpage_pincount_available(page))
686 atomic_set(compound_pincount_ptr(page), 0);
689 #ifdef CONFIG_DEBUG_PAGEALLOC
690 unsigned int _debug_guardpage_minorder;
692 bool _debug_pagealloc_enabled_early __read_mostly
693 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
694 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
695 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
696 EXPORT_SYMBOL(_debug_pagealloc_enabled);
698 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
700 static int __init early_debug_pagealloc(char *buf)
702 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
704 early_param("debug_pagealloc", early_debug_pagealloc);
706 void init_debug_pagealloc(void)
708 if (!debug_pagealloc_enabled())
711 static_branch_enable(&_debug_pagealloc_enabled);
713 if (!debug_guardpage_minorder())
716 static_branch_enable(&_debug_guardpage_enabled);
719 static int __init debug_guardpage_minorder_setup(char *buf)
723 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
724 pr_err("Bad debug_guardpage_minorder value\n");
727 _debug_guardpage_minorder = res;
728 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
731 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
733 static inline bool set_page_guard(struct zone *zone, struct page *page,
734 unsigned int order, int migratetype)
736 if (!debug_guardpage_enabled())
739 if (order >= debug_guardpage_minorder())
742 __SetPageGuard(page);
743 INIT_LIST_HEAD(&page->lru);
744 set_page_private(page, order);
745 /* Guard pages are not available for any usage */
746 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
751 static inline void clear_page_guard(struct zone *zone, struct page *page,
752 unsigned int order, int migratetype)
754 if (!debug_guardpage_enabled())
757 __ClearPageGuard(page);
759 set_page_private(page, 0);
760 if (!is_migrate_isolate(migratetype))
761 __mod_zone_freepage_state(zone, (1 << order), migratetype);
764 static inline bool set_page_guard(struct zone *zone, struct page *page,
765 unsigned int order, int migratetype) { return false; }
766 static inline void clear_page_guard(struct zone *zone, struct page *page,
767 unsigned int order, int migratetype) {}
770 static inline void set_page_order(struct page *page, unsigned int order)
772 set_page_private(page, order);
773 __SetPageBuddy(page);
777 * This function checks whether a page is free && is the buddy
778 * we can coalesce a page and its buddy if
779 * (a) the buddy is not in a hole (check before calling!) &&
780 * (b) the buddy is in the buddy system &&
781 * (c) a page and its buddy have the same order &&
782 * (d) a page and its buddy are in the same zone.
784 * For recording whether a page is in the buddy system, we set PageBuddy.
785 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
787 * For recording page's order, we use page_private(page).
789 static inline bool page_is_buddy(struct page *page, struct page *buddy,
792 if (!page_is_guard(buddy) && !PageBuddy(buddy))
795 if (page_order(buddy) != order)
799 * zone check is done late to avoid uselessly calculating
800 * zone/node ids for pages that could never merge.
802 if (page_zone_id(page) != page_zone_id(buddy))
805 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
810 #ifdef CONFIG_COMPACTION
811 static inline struct capture_control *task_capc(struct zone *zone)
813 struct capture_control *capc = current->capture_control;
815 return unlikely(capc) &&
816 !(current->flags & PF_KTHREAD) &&
818 capc->cc->zone == zone ? capc : NULL;
822 compaction_capture(struct capture_control *capc, struct page *page,
823 int order, int migratetype)
825 if (!capc || order != capc->cc->order)
828 /* Do not accidentally pollute CMA or isolated regions*/
829 if (is_migrate_cma(migratetype) ||
830 is_migrate_isolate(migratetype))
834 * Do not let lower order allocations polluate a movable pageblock.
835 * This might let an unmovable request use a reclaimable pageblock
836 * and vice-versa but no more than normal fallback logic which can
837 * have trouble finding a high-order free page.
839 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
847 static inline struct capture_control *task_capc(struct zone *zone)
853 compaction_capture(struct capture_control *capc, struct page *page,
854 int order, int migratetype)
858 #endif /* CONFIG_COMPACTION */
860 /* Used for pages not on another list */
861 static inline void add_to_free_list(struct page *page, struct zone *zone,
862 unsigned int order, int migratetype)
864 struct free_area *area = &zone->free_area[order];
866 list_add(&page->lru, &area->free_list[migratetype]);
870 /* Used for pages not on another list */
871 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
872 unsigned int order, int migratetype)
874 struct free_area *area = &zone->free_area[order];
876 list_add_tail(&page->lru, &area->free_list[migratetype]);
880 /* Used for pages which are on another list */
881 static inline void move_to_free_list(struct page *page, struct zone *zone,
882 unsigned int order, int migratetype)
884 struct free_area *area = &zone->free_area[order];
886 list_move(&page->lru, &area->free_list[migratetype]);
889 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
892 /* clear reported state and update reported page count */
893 if (page_reported(page))
894 __ClearPageReported(page);
896 list_del(&page->lru);
897 __ClearPageBuddy(page);
898 set_page_private(page, 0);
899 zone->free_area[order].nr_free--;
903 * If this is not the largest possible page, check if the buddy
904 * of the next-highest order is free. If it is, it's possible
905 * that pages are being freed that will coalesce soon. In case,
906 * that is happening, add the free page to the tail of the list
907 * so it's less likely to be used soon and more likely to be merged
908 * as a higher order page
911 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
912 struct page *page, unsigned int order)
914 struct page *higher_page, *higher_buddy;
915 unsigned long combined_pfn;
917 if (order >= MAX_ORDER - 2)
920 if (!pfn_valid_within(buddy_pfn))
923 combined_pfn = buddy_pfn & pfn;
924 higher_page = page + (combined_pfn - pfn);
925 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
926 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
928 return pfn_valid_within(buddy_pfn) &&
929 page_is_buddy(higher_page, higher_buddy, order + 1);
933 * Freeing function for a buddy system allocator.
935 * The concept of a buddy system is to maintain direct-mapped table
936 * (containing bit values) for memory blocks of various "orders".
937 * The bottom level table contains the map for the smallest allocatable
938 * units of memory (here, pages), and each level above it describes
939 * pairs of units from the levels below, hence, "buddies".
940 * At a high level, all that happens here is marking the table entry
941 * at the bottom level available, and propagating the changes upward
942 * as necessary, plus some accounting needed to play nicely with other
943 * parts of the VM system.
944 * At each level, we keep a list of pages, which are heads of continuous
945 * free pages of length of (1 << order) and marked with PageBuddy.
946 * Page's order is recorded in page_private(page) field.
947 * So when we are allocating or freeing one, we can derive the state of the
948 * other. That is, if we allocate a small block, and both were
949 * free, the remainder of the region must be split into blocks.
950 * If a block is freed, and its buddy is also free, then this
951 * triggers coalescing into a block of larger size.
956 static inline void __free_one_page(struct page *page,
958 struct zone *zone, unsigned int order,
959 int migratetype, bool report)
961 struct capture_control *capc = task_capc(zone);
962 unsigned long buddy_pfn;
963 unsigned long combined_pfn;
964 unsigned int max_order;
968 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
970 VM_BUG_ON(!zone_is_initialized(zone));
971 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
973 VM_BUG_ON(migratetype == -1);
974 if (likely(!is_migrate_isolate(migratetype)))
975 __mod_zone_freepage_state(zone, 1 << order, migratetype);
977 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
978 VM_BUG_ON_PAGE(bad_range(zone, page), page);
981 while (order < max_order - 1) {
982 if (compaction_capture(capc, page, order, migratetype)) {
983 __mod_zone_freepage_state(zone, -(1 << order),
987 buddy_pfn = __find_buddy_pfn(pfn, order);
988 buddy = page + (buddy_pfn - pfn);
990 if (!pfn_valid_within(buddy_pfn))
992 if (!page_is_buddy(page, buddy, order))
995 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
996 * merge with it and move up one order.
998 if (page_is_guard(buddy))
999 clear_page_guard(zone, buddy, order, migratetype);
1001 del_page_from_free_list(buddy, zone, order);
1002 combined_pfn = buddy_pfn & pfn;
1003 page = page + (combined_pfn - pfn);
1007 if (max_order < MAX_ORDER) {
1008 /* If we are here, it means order is >= pageblock_order.
1009 * We want to prevent merge between freepages on isolate
1010 * pageblock and normal pageblock. Without this, pageblock
1011 * isolation could cause incorrect freepage or CMA accounting.
1013 * We don't want to hit this code for the more frequent
1014 * low-order merging.
1016 if (unlikely(has_isolate_pageblock(zone))) {
1019 buddy_pfn = __find_buddy_pfn(pfn, order);
1020 buddy = page + (buddy_pfn - pfn);
1021 buddy_mt = get_pageblock_migratetype(buddy);
1023 if (migratetype != buddy_mt
1024 && (is_migrate_isolate(migratetype) ||
1025 is_migrate_isolate(buddy_mt)))
1029 goto continue_merging;
1033 set_page_order(page, order);
1035 if (is_shuffle_order(order))
1036 to_tail = shuffle_pick_tail();
1038 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1041 add_to_free_list_tail(page, zone, order, migratetype);
1043 add_to_free_list(page, zone, order, migratetype);
1045 /* Notify page reporting subsystem of freed page */
1047 page_reporting_notify_free(order);
1051 * A bad page could be due to a number of fields. Instead of multiple branches,
1052 * try and check multiple fields with one check. The caller must do a detailed
1053 * check if necessary.
1055 static inline bool page_expected_state(struct page *page,
1056 unsigned long check_flags)
1058 if (unlikely(atomic_read(&page->_mapcount) != -1))
1061 if (unlikely((unsigned long)page->mapping |
1062 page_ref_count(page) |
1064 (unsigned long)page->mem_cgroup |
1066 (page->flags & check_flags)))
1072 static const char *page_bad_reason(struct page *page, unsigned long flags)
1074 const char *bad_reason = NULL;
1076 if (unlikely(atomic_read(&page->_mapcount) != -1))
1077 bad_reason = "nonzero mapcount";
1078 if (unlikely(page->mapping != NULL))
1079 bad_reason = "non-NULL mapping";
1080 if (unlikely(page_ref_count(page) != 0))
1081 bad_reason = "nonzero _refcount";
1082 if (unlikely(page->flags & flags)) {
1083 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1084 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1086 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1089 if (unlikely(page->mem_cgroup))
1090 bad_reason = "page still charged to cgroup";
1095 static void check_free_page_bad(struct page *page)
1098 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1101 static inline int check_free_page(struct page *page)
1103 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1106 /* Something has gone sideways, find it */
1107 check_free_page_bad(page);
1111 static int free_tail_pages_check(struct page *head_page, struct page *page)
1116 * We rely page->lru.next never has bit 0 set, unless the page
1117 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1119 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1121 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1125 switch (page - head_page) {
1127 /* the first tail page: ->mapping may be compound_mapcount() */
1128 if (unlikely(compound_mapcount(page))) {
1129 bad_page(page, "nonzero compound_mapcount");
1135 * the second tail page: ->mapping is
1136 * deferred_list.next -- ignore value.
1140 if (page->mapping != TAIL_MAPPING) {
1141 bad_page(page, "corrupted mapping in tail page");
1146 if (unlikely(!PageTail(page))) {
1147 bad_page(page, "PageTail not set");
1150 if (unlikely(compound_head(page) != head_page)) {
1151 bad_page(page, "compound_head not consistent");
1156 page->mapping = NULL;
1157 clear_compound_head(page);
1161 static void kernel_init_free_pages(struct page *page, int numpages)
1165 for (i = 0; i < numpages; i++)
1166 clear_highpage(page + i);
1169 static __always_inline bool free_pages_prepare(struct page *page,
1170 unsigned int order, bool check_free)
1174 VM_BUG_ON_PAGE(PageTail(page), page);
1176 trace_mm_page_free(page, order);
1179 * Check tail pages before head page information is cleared to
1180 * avoid checking PageCompound for order-0 pages.
1182 if (unlikely(order)) {
1183 bool compound = PageCompound(page);
1186 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1189 ClearPageDoubleMap(page);
1190 for (i = 1; i < (1 << order); i++) {
1192 bad += free_tail_pages_check(page, page + i);
1193 if (unlikely(check_free_page(page + i))) {
1197 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1200 if (PageMappingFlags(page))
1201 page->mapping = NULL;
1202 if (memcg_kmem_enabled() && PageKmemcg(page))
1203 __memcg_kmem_uncharge_page(page, order);
1205 bad += check_free_page(page);
1209 page_cpupid_reset_last(page);
1210 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1211 reset_page_owner(page, order);
1213 if (!PageHighMem(page)) {
1214 debug_check_no_locks_freed(page_address(page),
1215 PAGE_SIZE << order);
1216 debug_check_no_obj_freed(page_address(page),
1217 PAGE_SIZE << order);
1219 if (want_init_on_free())
1220 kernel_init_free_pages(page, 1 << order);
1222 kernel_poison_pages(page, 1 << order, 0);
1224 * arch_free_page() can make the page's contents inaccessible. s390
1225 * does this. So nothing which can access the page's contents should
1226 * happen after this.
1228 arch_free_page(page, order);
1230 if (debug_pagealloc_enabled_static())
1231 kernel_map_pages(page, 1 << order, 0);
1233 kasan_free_nondeferred_pages(page, order);
1238 #ifdef CONFIG_DEBUG_VM
1240 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1241 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1242 * moved from pcp lists to free lists.
1244 static bool free_pcp_prepare(struct page *page)
1246 return free_pages_prepare(page, 0, true);
1249 static bool bulkfree_pcp_prepare(struct page *page)
1251 if (debug_pagealloc_enabled_static())
1252 return check_free_page(page);
1258 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1259 * moving from pcp lists to free list in order to reduce overhead. With
1260 * debug_pagealloc enabled, they are checked also immediately when being freed
1263 static bool free_pcp_prepare(struct page *page)
1265 if (debug_pagealloc_enabled_static())
1266 return free_pages_prepare(page, 0, true);
1268 return free_pages_prepare(page, 0, false);
1271 static bool bulkfree_pcp_prepare(struct page *page)
1273 return check_free_page(page);
1275 #endif /* CONFIG_DEBUG_VM */
1277 static inline void prefetch_buddy(struct page *page)
1279 unsigned long pfn = page_to_pfn(page);
1280 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1281 struct page *buddy = page + (buddy_pfn - pfn);
1287 * Frees a number of pages from the PCP lists
1288 * Assumes all pages on list are in same zone, and of same order.
1289 * count is the number of pages to free.
1291 * If the zone was previously in an "all pages pinned" state then look to
1292 * see if this freeing clears that state.
1294 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1295 * pinned" detection logic.
1297 static void free_pcppages_bulk(struct zone *zone, int count,
1298 struct per_cpu_pages *pcp)
1300 int migratetype = 0;
1302 int prefetch_nr = 0;
1303 bool isolated_pageblocks;
1304 struct page *page, *tmp;
1308 struct list_head *list;
1311 * Remove pages from lists in a round-robin fashion. A
1312 * batch_free count is maintained that is incremented when an
1313 * empty list is encountered. This is so more pages are freed
1314 * off fuller lists instead of spinning excessively around empty
1319 if (++migratetype == MIGRATE_PCPTYPES)
1321 list = &pcp->lists[migratetype];
1322 } while (list_empty(list));
1324 /* This is the only non-empty list. Free them all. */
1325 if (batch_free == MIGRATE_PCPTYPES)
1329 page = list_last_entry(list, struct page, lru);
1330 /* must delete to avoid corrupting pcp list */
1331 list_del(&page->lru);
1334 if (bulkfree_pcp_prepare(page))
1337 list_add_tail(&page->lru, &head);
1340 * We are going to put the page back to the global
1341 * pool, prefetch its buddy to speed up later access
1342 * under zone->lock. It is believed the overhead of
1343 * an additional test and calculating buddy_pfn here
1344 * can be offset by reduced memory latency later. To
1345 * avoid excessive prefetching due to large count, only
1346 * prefetch buddy for the first pcp->batch nr of pages.
1348 if (prefetch_nr++ < pcp->batch)
1349 prefetch_buddy(page);
1350 } while (--count && --batch_free && !list_empty(list));
1353 spin_lock(&zone->lock);
1354 isolated_pageblocks = has_isolate_pageblock(zone);
1357 * Use safe version since after __free_one_page(),
1358 * page->lru.next will not point to original list.
1360 list_for_each_entry_safe(page, tmp, &head, lru) {
1361 int mt = get_pcppage_migratetype(page);
1362 /* MIGRATE_ISOLATE page should not go to pcplists */
1363 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1364 /* Pageblock could have been isolated meanwhile */
1365 if (unlikely(isolated_pageblocks))
1366 mt = get_pageblock_migratetype(page);
1368 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1369 trace_mm_page_pcpu_drain(page, 0, mt);
1371 spin_unlock(&zone->lock);
1374 static void free_one_page(struct zone *zone,
1375 struct page *page, unsigned long pfn,
1379 spin_lock(&zone->lock);
1380 if (unlikely(has_isolate_pageblock(zone) ||
1381 is_migrate_isolate(migratetype))) {
1382 migratetype = get_pfnblock_migratetype(page, pfn);
1384 __free_one_page(page, pfn, zone, order, migratetype, true);
1385 spin_unlock(&zone->lock);
1388 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1389 unsigned long zone, int nid)
1391 mm_zero_struct_page(page);
1392 set_page_links(page, zone, nid, pfn);
1393 init_page_count(page);
1394 page_mapcount_reset(page);
1395 page_cpupid_reset_last(page);
1396 page_kasan_tag_reset(page);
1398 INIT_LIST_HEAD(&page->lru);
1399 #ifdef WANT_PAGE_VIRTUAL
1400 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1401 if (!is_highmem_idx(zone))
1402 set_page_address(page, __va(pfn << PAGE_SHIFT));
1406 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1407 static void __meminit init_reserved_page(unsigned long pfn)
1412 if (!early_page_uninitialised(pfn))
1415 nid = early_pfn_to_nid(pfn);
1416 pgdat = NODE_DATA(nid);
1418 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1419 struct zone *zone = &pgdat->node_zones[zid];
1421 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1424 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1427 static inline void init_reserved_page(unsigned long pfn)
1430 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1433 * Initialised pages do not have PageReserved set. This function is
1434 * called for each range allocated by the bootmem allocator and
1435 * marks the pages PageReserved. The remaining valid pages are later
1436 * sent to the buddy page allocator.
1438 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1440 unsigned long start_pfn = PFN_DOWN(start);
1441 unsigned long end_pfn = PFN_UP(end);
1443 for (; start_pfn < end_pfn; start_pfn++) {
1444 if (pfn_valid(start_pfn)) {
1445 struct page *page = pfn_to_page(start_pfn);
1447 init_reserved_page(start_pfn);
1449 /* Avoid false-positive PageTail() */
1450 INIT_LIST_HEAD(&page->lru);
1453 * no need for atomic set_bit because the struct
1454 * page is not visible yet so nobody should
1457 __SetPageReserved(page);
1462 static void __free_pages_ok(struct page *page, unsigned int order)
1464 unsigned long flags;
1466 unsigned long pfn = page_to_pfn(page);
1468 if (!free_pages_prepare(page, order, true))
1471 migratetype = get_pfnblock_migratetype(page, pfn);
1472 local_irq_save(flags);
1473 __count_vm_events(PGFREE, 1 << order);
1474 free_one_page(page_zone(page), page, pfn, order, migratetype);
1475 local_irq_restore(flags);
1478 void __free_pages_core(struct page *page, unsigned int order)
1480 unsigned int nr_pages = 1 << order;
1481 struct page *p = page;
1485 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1487 __ClearPageReserved(p);
1488 set_page_count(p, 0);
1490 __ClearPageReserved(p);
1491 set_page_count(p, 0);
1493 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1494 set_page_refcounted(page);
1495 __free_pages(page, order);
1498 #ifdef CONFIG_NEED_MULTIPLE_NODES
1500 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1502 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1505 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1507 int __meminit __early_pfn_to_nid(unsigned long pfn,
1508 struct mminit_pfnnid_cache *state)
1510 unsigned long start_pfn, end_pfn;
1513 if (state->last_start <= pfn && pfn < state->last_end)
1514 return state->last_nid;
1516 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1517 if (nid != NUMA_NO_NODE) {
1518 state->last_start = start_pfn;
1519 state->last_end = end_pfn;
1520 state->last_nid = nid;
1525 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1527 int __meminit early_pfn_to_nid(unsigned long pfn)
1529 static DEFINE_SPINLOCK(early_pfn_lock);
1532 spin_lock(&early_pfn_lock);
1533 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1535 nid = first_online_node;
1536 spin_unlock(&early_pfn_lock);
1540 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1542 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1545 if (early_page_uninitialised(pfn))
1547 __free_pages_core(page, order);
1551 * Check that the whole (or subset of) a pageblock given by the interval of
1552 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1553 * with the migration of free compaction scanner. The scanners then need to
1554 * use only pfn_valid_within() check for arches that allow holes within
1557 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1559 * It's possible on some configurations to have a setup like node0 node1 node0
1560 * i.e. it's possible that all pages within a zones range of pages do not
1561 * belong to a single zone. We assume that a border between node0 and node1
1562 * can occur within a single pageblock, but not a node0 node1 node0
1563 * interleaving within a single pageblock. It is therefore sufficient to check
1564 * the first and last page of a pageblock and avoid checking each individual
1565 * page in a pageblock.
1567 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1568 unsigned long end_pfn, struct zone *zone)
1570 struct page *start_page;
1571 struct page *end_page;
1573 /* end_pfn is one past the range we are checking */
1576 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1579 start_page = pfn_to_online_page(start_pfn);
1583 if (page_zone(start_page) != zone)
1586 end_page = pfn_to_page(end_pfn);
1588 /* This gives a shorter code than deriving page_zone(end_page) */
1589 if (page_zone_id(start_page) != page_zone_id(end_page))
1595 void set_zone_contiguous(struct zone *zone)
1597 unsigned long block_start_pfn = zone->zone_start_pfn;
1598 unsigned long block_end_pfn;
1600 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1601 for (; block_start_pfn < zone_end_pfn(zone);
1602 block_start_pfn = block_end_pfn,
1603 block_end_pfn += pageblock_nr_pages) {
1605 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1607 if (!__pageblock_pfn_to_page(block_start_pfn,
1608 block_end_pfn, zone))
1613 /* We confirm that there is no hole */
1614 zone->contiguous = true;
1617 void clear_zone_contiguous(struct zone *zone)
1619 zone->contiguous = false;
1622 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1623 static void __init deferred_free_range(unsigned long pfn,
1624 unsigned long nr_pages)
1632 page = pfn_to_page(pfn);
1634 /* Free a large naturally-aligned chunk if possible */
1635 if (nr_pages == pageblock_nr_pages &&
1636 (pfn & (pageblock_nr_pages - 1)) == 0) {
1637 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1638 __free_pages_core(page, pageblock_order);
1642 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1643 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1644 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1645 __free_pages_core(page, 0);
1649 /* Completion tracking for deferred_init_memmap() threads */
1650 static atomic_t pgdat_init_n_undone __initdata;
1651 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1653 static inline void __init pgdat_init_report_one_done(void)
1655 if (atomic_dec_and_test(&pgdat_init_n_undone))
1656 complete(&pgdat_init_all_done_comp);
1660 * Returns true if page needs to be initialized or freed to buddy allocator.
1662 * First we check if pfn is valid on architectures where it is possible to have
1663 * holes within pageblock_nr_pages. On systems where it is not possible, this
1664 * function is optimized out.
1666 * Then, we check if a current large page is valid by only checking the validity
1669 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1671 if (!pfn_valid_within(pfn))
1673 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1679 * Free pages to buddy allocator. Try to free aligned pages in
1680 * pageblock_nr_pages sizes.
1682 static void __init deferred_free_pages(unsigned long pfn,
1683 unsigned long end_pfn)
1685 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1686 unsigned long nr_free = 0;
1688 for (; pfn < end_pfn; pfn++) {
1689 if (!deferred_pfn_valid(pfn)) {
1690 deferred_free_range(pfn - nr_free, nr_free);
1692 } else if (!(pfn & nr_pgmask)) {
1693 deferred_free_range(pfn - nr_free, nr_free);
1699 /* Free the last block of pages to allocator */
1700 deferred_free_range(pfn - nr_free, nr_free);
1704 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1705 * by performing it only once every pageblock_nr_pages.
1706 * Return number of pages initialized.
1708 static unsigned long __init deferred_init_pages(struct zone *zone,
1710 unsigned long end_pfn)
1712 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1713 int nid = zone_to_nid(zone);
1714 unsigned long nr_pages = 0;
1715 int zid = zone_idx(zone);
1716 struct page *page = NULL;
1718 for (; pfn < end_pfn; pfn++) {
1719 if (!deferred_pfn_valid(pfn)) {
1722 } else if (!page || !(pfn & nr_pgmask)) {
1723 page = pfn_to_page(pfn);
1727 __init_single_page(page, pfn, zid, nid);
1734 * This function is meant to pre-load the iterator for the zone init.
1735 * Specifically it walks through the ranges until we are caught up to the
1736 * first_init_pfn value and exits there. If we never encounter the value we
1737 * return false indicating there are no valid ranges left.
1740 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1741 unsigned long *spfn, unsigned long *epfn,
1742 unsigned long first_init_pfn)
1747 * Start out by walking through the ranges in this zone that have
1748 * already been initialized. We don't need to do anything with them
1749 * so we just need to flush them out of the system.
1751 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1752 if (*epfn <= first_init_pfn)
1754 if (*spfn < first_init_pfn)
1755 *spfn = first_init_pfn;
1764 * Initialize and free pages. We do it in two loops: first we initialize
1765 * struct page, then free to buddy allocator, because while we are
1766 * freeing pages we can access pages that are ahead (computing buddy
1767 * page in __free_one_page()).
1769 * In order to try and keep some memory in the cache we have the loop
1770 * broken along max page order boundaries. This way we will not cause
1771 * any issues with the buddy page computation.
1773 static unsigned long __init
1774 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1775 unsigned long *end_pfn)
1777 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1778 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1779 unsigned long nr_pages = 0;
1782 /* First we loop through and initialize the page values */
1783 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1786 if (mo_pfn <= *start_pfn)
1789 t = min(mo_pfn, *end_pfn);
1790 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1792 if (mo_pfn < *end_pfn) {
1793 *start_pfn = mo_pfn;
1798 /* Reset values and now loop through freeing pages as needed */
1801 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1807 t = min(mo_pfn, epfn);
1808 deferred_free_pages(spfn, t);
1818 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1821 unsigned long spfn, epfn;
1822 struct zone *zone = arg;
1825 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1828 * Initialize and free pages in MAX_ORDER sized increments so that we
1829 * can avoid introducing any issues with the buddy allocator.
1831 while (spfn < end_pfn) {
1832 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1837 /* An arch may override for more concurrency. */
1839 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1844 /* Initialise remaining memory on a node */
1845 static int __init deferred_init_memmap(void *data)
1847 pg_data_t *pgdat = data;
1848 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1849 unsigned long spfn = 0, epfn = 0;
1850 unsigned long first_init_pfn, flags;
1851 unsigned long start = jiffies;
1853 int zid, max_threads;
1856 /* Bind memory initialisation thread to a local node if possible */
1857 if (!cpumask_empty(cpumask))
1858 set_cpus_allowed_ptr(current, cpumask);
1860 pgdat_resize_lock(pgdat, &flags);
1861 first_init_pfn = pgdat->first_deferred_pfn;
1862 if (first_init_pfn == ULONG_MAX) {
1863 pgdat_resize_unlock(pgdat, &flags);
1864 pgdat_init_report_one_done();
1868 /* Sanity check boundaries */
1869 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1870 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1871 pgdat->first_deferred_pfn = ULONG_MAX;
1874 * Once we unlock here, the zone cannot be grown anymore, thus if an
1875 * interrupt thread must allocate this early in boot, zone must be
1876 * pre-grown prior to start of deferred page initialization.
1878 pgdat_resize_unlock(pgdat, &flags);
1880 /* Only the highest zone is deferred so find it */
1881 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1882 zone = pgdat->node_zones + zid;
1883 if (first_init_pfn < zone_end_pfn(zone))
1887 /* If the zone is empty somebody else may have cleared out the zone */
1888 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1892 max_threads = deferred_page_init_max_threads(cpumask);
1894 while (spfn < epfn) {
1895 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1896 struct padata_mt_job job = {
1897 .thread_fn = deferred_init_memmap_chunk,
1900 .size = epfn_align - spfn,
1901 .align = PAGES_PER_SECTION,
1902 .min_chunk = PAGES_PER_SECTION,
1903 .max_threads = max_threads,
1906 padata_do_multithreaded(&job);
1907 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1911 /* Sanity check that the next zone really is unpopulated */
1912 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1914 pr_info("node %d deferred pages initialised in %ums\n",
1915 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1917 pgdat_init_report_one_done();
1922 * If this zone has deferred pages, try to grow it by initializing enough
1923 * deferred pages to satisfy the allocation specified by order, rounded up to
1924 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1925 * of SECTION_SIZE bytes by initializing struct pages in increments of
1926 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1928 * Return true when zone was grown, otherwise return false. We return true even
1929 * when we grow less than requested, to let the caller decide if there are
1930 * enough pages to satisfy the allocation.
1932 * Note: We use noinline because this function is needed only during boot, and
1933 * it is called from a __ref function _deferred_grow_zone. This way we are
1934 * making sure that it is not inlined into permanent text section.
1936 static noinline bool __init
1937 deferred_grow_zone(struct zone *zone, unsigned int order)
1939 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1940 pg_data_t *pgdat = zone->zone_pgdat;
1941 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1942 unsigned long spfn, epfn, flags;
1943 unsigned long nr_pages = 0;
1946 /* Only the last zone may have deferred pages */
1947 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1950 pgdat_resize_lock(pgdat, &flags);
1953 * If someone grew this zone while we were waiting for spinlock, return
1954 * true, as there might be enough pages already.
1956 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1957 pgdat_resize_unlock(pgdat, &flags);
1961 /* If the zone is empty somebody else may have cleared out the zone */
1962 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1963 first_deferred_pfn)) {
1964 pgdat->first_deferred_pfn = ULONG_MAX;
1965 pgdat_resize_unlock(pgdat, &flags);
1966 /* Retry only once. */
1967 return first_deferred_pfn != ULONG_MAX;
1971 * Initialize and free pages in MAX_ORDER sized increments so
1972 * that we can avoid introducing any issues with the buddy
1975 while (spfn < epfn) {
1976 /* update our first deferred PFN for this section */
1977 first_deferred_pfn = spfn;
1979 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1980 touch_nmi_watchdog();
1982 /* We should only stop along section boundaries */
1983 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1986 /* If our quota has been met we can stop here */
1987 if (nr_pages >= nr_pages_needed)
1991 pgdat->first_deferred_pfn = spfn;
1992 pgdat_resize_unlock(pgdat, &flags);
1994 return nr_pages > 0;
1998 * deferred_grow_zone() is __init, but it is called from
1999 * get_page_from_freelist() during early boot until deferred_pages permanently
2000 * disables this call. This is why we have refdata wrapper to avoid warning,
2001 * and to ensure that the function body gets unloaded.
2004 _deferred_grow_zone(struct zone *zone, unsigned int order)
2006 return deferred_grow_zone(zone, order);
2009 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2011 void __init page_alloc_init_late(void)
2016 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2018 /* There will be num_node_state(N_MEMORY) threads */
2019 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2020 for_each_node_state(nid, N_MEMORY) {
2021 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2024 /* Block until all are initialised */
2025 wait_for_completion(&pgdat_init_all_done_comp);
2028 * The number of managed pages has changed due to the initialisation
2029 * so the pcpu batch and high limits needs to be updated or the limits
2030 * will be artificially small.
2032 for_each_populated_zone(zone)
2033 zone_pcp_update(zone);
2036 * We initialized the rest of the deferred pages. Permanently disable
2037 * on-demand struct page initialization.
2039 static_branch_disable(&deferred_pages);
2041 /* Reinit limits that are based on free pages after the kernel is up */
2042 files_maxfiles_init();
2045 /* Discard memblock private memory */
2048 for_each_node_state(nid, N_MEMORY)
2049 shuffle_free_memory(NODE_DATA(nid));
2051 for_each_populated_zone(zone)
2052 set_zone_contiguous(zone);
2056 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2057 void __init init_cma_reserved_pageblock(struct page *page)
2059 unsigned i = pageblock_nr_pages;
2060 struct page *p = page;
2063 __ClearPageReserved(p);
2064 set_page_count(p, 0);
2067 set_pageblock_migratetype(page, MIGRATE_CMA);
2069 if (pageblock_order >= MAX_ORDER) {
2070 i = pageblock_nr_pages;
2073 set_page_refcounted(p);
2074 __free_pages(p, MAX_ORDER - 1);
2075 p += MAX_ORDER_NR_PAGES;
2076 } while (i -= MAX_ORDER_NR_PAGES);
2078 set_page_refcounted(page);
2079 __free_pages(page, pageblock_order);
2082 adjust_managed_page_count(page, pageblock_nr_pages);
2087 * The order of subdivision here is critical for the IO subsystem.
2088 * Please do not alter this order without good reasons and regression
2089 * testing. Specifically, as large blocks of memory are subdivided,
2090 * the order in which smaller blocks are delivered depends on the order
2091 * they're subdivided in this function. This is the primary factor
2092 * influencing the order in which pages are delivered to the IO
2093 * subsystem according to empirical testing, and this is also justified
2094 * by considering the behavior of a buddy system containing a single
2095 * large block of memory acted on by a series of small allocations.
2096 * This behavior is a critical factor in sglist merging's success.
2100 static inline void expand(struct zone *zone, struct page *page,
2101 int low, int high, int migratetype)
2103 unsigned long size = 1 << high;
2105 while (high > low) {
2108 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2111 * Mark as guard pages (or page), that will allow to
2112 * merge back to allocator when buddy will be freed.
2113 * Corresponding page table entries will not be touched,
2114 * pages will stay not present in virtual address space
2116 if (set_page_guard(zone, &page[size], high, migratetype))
2119 add_to_free_list(&page[size], zone, high, migratetype);
2120 set_page_order(&page[size], high);
2124 static void check_new_page_bad(struct page *page)
2126 if (unlikely(page->flags & __PG_HWPOISON)) {
2127 /* Don't complain about hwpoisoned pages */
2128 page_mapcount_reset(page); /* remove PageBuddy */
2133 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2137 * This page is about to be returned from the page allocator
2139 static inline int check_new_page(struct page *page)
2141 if (likely(page_expected_state(page,
2142 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2145 check_new_page_bad(page);
2149 static inline bool free_pages_prezeroed(void)
2151 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2152 page_poisoning_enabled()) || want_init_on_free();
2155 #ifdef CONFIG_DEBUG_VM
2157 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2158 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2159 * also checked when pcp lists are refilled from the free lists.
2161 static inline bool check_pcp_refill(struct page *page)
2163 if (debug_pagealloc_enabled_static())
2164 return check_new_page(page);
2169 static inline bool check_new_pcp(struct page *page)
2171 return check_new_page(page);
2175 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2176 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2177 * enabled, they are also checked when being allocated from the pcp lists.
2179 static inline bool check_pcp_refill(struct page *page)
2181 return check_new_page(page);
2183 static inline bool check_new_pcp(struct page *page)
2185 if (debug_pagealloc_enabled_static())
2186 return check_new_page(page);
2190 #endif /* CONFIG_DEBUG_VM */
2192 static bool check_new_pages(struct page *page, unsigned int order)
2195 for (i = 0; i < (1 << order); i++) {
2196 struct page *p = page + i;
2198 if (unlikely(check_new_page(p)))
2205 inline void post_alloc_hook(struct page *page, unsigned int order,
2208 set_page_private(page, 0);
2209 set_page_refcounted(page);
2211 arch_alloc_page(page, order);
2212 if (debug_pagealloc_enabled_static())
2213 kernel_map_pages(page, 1 << order, 1);
2214 kasan_alloc_pages(page, order);
2215 kernel_poison_pages(page, 1 << order, 1);
2216 set_page_owner(page, order, gfp_flags);
2219 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2220 unsigned int alloc_flags)
2222 post_alloc_hook(page, order, gfp_flags);
2224 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2225 kernel_init_free_pages(page, 1 << order);
2227 if (order && (gfp_flags & __GFP_COMP))
2228 prep_compound_page(page, order);
2231 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2232 * allocate the page. The expectation is that the caller is taking
2233 * steps that will free more memory. The caller should avoid the page
2234 * being used for !PFMEMALLOC purposes.
2236 if (alloc_flags & ALLOC_NO_WATERMARKS)
2237 set_page_pfmemalloc(page);
2239 clear_page_pfmemalloc(page);
2243 * Go through the free lists for the given migratetype and remove
2244 * the smallest available page from the freelists
2246 static __always_inline
2247 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2250 unsigned int current_order;
2251 struct free_area *area;
2254 /* Find a page of the appropriate size in the preferred list */
2255 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2256 area = &(zone->free_area[current_order]);
2257 page = get_page_from_free_area(area, migratetype);
2260 del_page_from_free_list(page, zone, current_order);
2261 expand(zone, page, order, current_order, migratetype);
2262 set_pcppage_migratetype(page, migratetype);
2271 * This array describes the order lists are fallen back to when
2272 * the free lists for the desirable migrate type are depleted
2274 static int fallbacks[MIGRATE_TYPES][4] = {
2275 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2276 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2277 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2279 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2281 #ifdef CONFIG_MEMORY_ISOLATION
2282 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2287 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2290 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2293 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2294 unsigned int order) { return NULL; }
2298 * Move the free pages in a range to the free lists of the requested type.
2299 * Note that start_page and end_pages are not aligned on a pageblock
2300 * boundary. If alignment is required, use move_freepages_block()
2302 static int move_freepages(struct zone *zone,
2303 struct page *start_page, struct page *end_page,
2304 int migratetype, int *num_movable)
2308 int pages_moved = 0;
2310 for (page = start_page; page <= end_page;) {
2311 if (!pfn_valid_within(page_to_pfn(page))) {
2316 if (!PageBuddy(page)) {
2318 * We assume that pages that could be isolated for
2319 * migration are movable. But we don't actually try
2320 * isolating, as that would be expensive.
2323 (PageLRU(page) || __PageMovable(page)))
2330 /* Make sure we are not inadvertently changing nodes */
2331 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2332 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2334 order = page_order(page);
2335 move_to_free_list(page, zone, order, migratetype);
2337 pages_moved += 1 << order;
2343 int move_freepages_block(struct zone *zone, struct page *page,
2344 int migratetype, int *num_movable)
2346 unsigned long start_pfn, end_pfn;
2347 struct page *start_page, *end_page;
2352 start_pfn = page_to_pfn(page);
2353 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2354 start_page = pfn_to_page(start_pfn);
2355 end_page = start_page + pageblock_nr_pages - 1;
2356 end_pfn = start_pfn + pageblock_nr_pages - 1;
2358 /* Do not cross zone boundaries */
2359 if (!zone_spans_pfn(zone, start_pfn))
2361 if (!zone_spans_pfn(zone, end_pfn))
2364 return move_freepages(zone, start_page, end_page, migratetype,
2368 static void change_pageblock_range(struct page *pageblock_page,
2369 int start_order, int migratetype)
2371 int nr_pageblocks = 1 << (start_order - pageblock_order);
2373 while (nr_pageblocks--) {
2374 set_pageblock_migratetype(pageblock_page, migratetype);
2375 pageblock_page += pageblock_nr_pages;
2380 * When we are falling back to another migratetype during allocation, try to
2381 * steal extra free pages from the same pageblocks to satisfy further
2382 * allocations, instead of polluting multiple pageblocks.
2384 * If we are stealing a relatively large buddy page, it is likely there will
2385 * be more free pages in the pageblock, so try to steal them all. For
2386 * reclaimable and unmovable allocations, we steal regardless of page size,
2387 * as fragmentation caused by those allocations polluting movable pageblocks
2388 * is worse than movable allocations stealing from unmovable and reclaimable
2391 static bool can_steal_fallback(unsigned int order, int start_mt)
2394 * Leaving this order check is intended, although there is
2395 * relaxed order check in next check. The reason is that
2396 * we can actually steal whole pageblock if this condition met,
2397 * but, below check doesn't guarantee it and that is just heuristic
2398 * so could be changed anytime.
2400 if (order >= pageblock_order)
2403 if (order >= pageblock_order / 2 ||
2404 start_mt == MIGRATE_RECLAIMABLE ||
2405 start_mt == MIGRATE_UNMOVABLE ||
2406 page_group_by_mobility_disabled)
2412 static inline void boost_watermark(struct zone *zone)
2414 unsigned long max_boost;
2416 if (!watermark_boost_factor)
2419 * Don't bother in zones that are unlikely to produce results.
2420 * On small machines, including kdump capture kernels running
2421 * in a small area, boosting the watermark can cause an out of
2422 * memory situation immediately.
2424 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2427 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2428 watermark_boost_factor, 10000);
2431 * high watermark may be uninitialised if fragmentation occurs
2432 * very early in boot so do not boost. We do not fall
2433 * through and boost by pageblock_nr_pages as failing
2434 * allocations that early means that reclaim is not going
2435 * to help and it may even be impossible to reclaim the
2436 * boosted watermark resulting in a hang.
2441 max_boost = max(pageblock_nr_pages, max_boost);
2443 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2448 * This function implements actual steal behaviour. If order is large enough,
2449 * we can steal whole pageblock. If not, we first move freepages in this
2450 * pageblock to our migratetype and determine how many already-allocated pages
2451 * are there in the pageblock with a compatible migratetype. If at least half
2452 * of pages are free or compatible, we can change migratetype of the pageblock
2453 * itself, so pages freed in the future will be put on the correct free list.
2455 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2456 unsigned int alloc_flags, int start_type, bool whole_block)
2458 unsigned int current_order = page_order(page);
2459 int free_pages, movable_pages, alike_pages;
2462 old_block_type = get_pageblock_migratetype(page);
2465 * This can happen due to races and we want to prevent broken
2466 * highatomic accounting.
2468 if (is_migrate_highatomic(old_block_type))
2471 /* Take ownership for orders >= pageblock_order */
2472 if (current_order >= pageblock_order) {
2473 change_pageblock_range(page, current_order, start_type);
2478 * Boost watermarks to increase reclaim pressure to reduce the
2479 * likelihood of future fallbacks. Wake kswapd now as the node
2480 * may be balanced overall and kswapd will not wake naturally.
2482 boost_watermark(zone);
2483 if (alloc_flags & ALLOC_KSWAPD)
2484 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2486 /* We are not allowed to try stealing from the whole block */
2490 free_pages = move_freepages_block(zone, page, start_type,
2493 * Determine how many pages are compatible with our allocation.
2494 * For movable allocation, it's the number of movable pages which
2495 * we just obtained. For other types it's a bit more tricky.
2497 if (start_type == MIGRATE_MOVABLE) {
2498 alike_pages = movable_pages;
2501 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2502 * to MOVABLE pageblock, consider all non-movable pages as
2503 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2504 * vice versa, be conservative since we can't distinguish the
2505 * exact migratetype of non-movable pages.
2507 if (old_block_type == MIGRATE_MOVABLE)
2508 alike_pages = pageblock_nr_pages
2509 - (free_pages + movable_pages);
2514 /* moving whole block can fail due to zone boundary conditions */
2519 * If a sufficient number of pages in the block are either free or of
2520 * comparable migratability as our allocation, claim the whole block.
2522 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2523 page_group_by_mobility_disabled)
2524 set_pageblock_migratetype(page, start_type);
2529 move_to_free_list(page, zone, current_order, start_type);
2533 * Check whether there is a suitable fallback freepage with requested order.
2534 * If only_stealable is true, this function returns fallback_mt only if
2535 * we can steal other freepages all together. This would help to reduce
2536 * fragmentation due to mixed migratetype pages in one pageblock.
2538 int find_suitable_fallback(struct free_area *area, unsigned int order,
2539 int migratetype, bool only_stealable, bool *can_steal)
2544 if (area->nr_free == 0)
2549 fallback_mt = fallbacks[migratetype][i];
2550 if (fallback_mt == MIGRATE_TYPES)
2553 if (free_area_empty(area, fallback_mt))
2556 if (can_steal_fallback(order, migratetype))
2559 if (!only_stealable)
2570 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2571 * there are no empty page blocks that contain a page with a suitable order
2573 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2574 unsigned int alloc_order)
2577 unsigned long max_managed, flags;
2580 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2581 * Check is race-prone but harmless.
2583 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2584 if (zone->nr_reserved_highatomic >= max_managed)
2587 spin_lock_irqsave(&zone->lock, flags);
2589 /* Recheck the nr_reserved_highatomic limit under the lock */
2590 if (zone->nr_reserved_highatomic >= max_managed)
2594 mt = get_pageblock_migratetype(page);
2595 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2596 && !is_migrate_cma(mt)) {
2597 zone->nr_reserved_highatomic += pageblock_nr_pages;
2598 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2599 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2603 spin_unlock_irqrestore(&zone->lock, flags);
2607 * Used when an allocation is about to fail under memory pressure. This
2608 * potentially hurts the reliability of high-order allocations when under
2609 * intense memory pressure but failed atomic allocations should be easier
2610 * to recover from than an OOM.
2612 * If @force is true, try to unreserve a pageblock even though highatomic
2613 * pageblock is exhausted.
2615 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2618 struct zonelist *zonelist = ac->zonelist;
2619 unsigned long flags;
2626 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2629 * Preserve at least one pageblock unless memory pressure
2632 if (!force && zone->nr_reserved_highatomic <=
2636 spin_lock_irqsave(&zone->lock, flags);
2637 for (order = 0; order < MAX_ORDER; order++) {
2638 struct free_area *area = &(zone->free_area[order]);
2640 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2645 * In page freeing path, migratetype change is racy so
2646 * we can counter several free pages in a pageblock
2647 * in this loop althoug we changed the pageblock type
2648 * from highatomic to ac->migratetype. So we should
2649 * adjust the count once.
2651 if (is_migrate_highatomic_page(page)) {
2653 * It should never happen but changes to
2654 * locking could inadvertently allow a per-cpu
2655 * drain to add pages to MIGRATE_HIGHATOMIC
2656 * while unreserving so be safe and watch for
2659 zone->nr_reserved_highatomic -= min(
2661 zone->nr_reserved_highatomic);
2665 * Convert to ac->migratetype and avoid the normal
2666 * pageblock stealing heuristics. Minimally, the caller
2667 * is doing the work and needs the pages. More
2668 * importantly, if the block was always converted to
2669 * MIGRATE_UNMOVABLE or another type then the number
2670 * of pageblocks that cannot be completely freed
2673 set_pageblock_migratetype(page, ac->migratetype);
2674 ret = move_freepages_block(zone, page, ac->migratetype,
2677 spin_unlock_irqrestore(&zone->lock, flags);
2681 spin_unlock_irqrestore(&zone->lock, flags);
2688 * Try finding a free buddy page on the fallback list and put it on the free
2689 * list of requested migratetype, possibly along with other pages from the same
2690 * block, depending on fragmentation avoidance heuristics. Returns true if
2691 * fallback was found so that __rmqueue_smallest() can grab it.
2693 * The use of signed ints for order and current_order is a deliberate
2694 * deviation from the rest of this file, to make the for loop
2695 * condition simpler.
2697 static __always_inline bool
2698 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2699 unsigned int alloc_flags)
2701 struct free_area *area;
2703 int min_order = order;
2709 * Do not steal pages from freelists belonging to other pageblocks
2710 * i.e. orders < pageblock_order. If there are no local zones free,
2711 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2713 if (alloc_flags & ALLOC_NOFRAGMENT)
2714 min_order = pageblock_order;
2717 * Find the largest available free page in the other list. This roughly
2718 * approximates finding the pageblock with the most free pages, which
2719 * would be too costly to do exactly.
2721 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2723 area = &(zone->free_area[current_order]);
2724 fallback_mt = find_suitable_fallback(area, current_order,
2725 start_migratetype, false, &can_steal);
2726 if (fallback_mt == -1)
2730 * We cannot steal all free pages from the pageblock and the
2731 * requested migratetype is movable. In that case it's better to
2732 * steal and split the smallest available page instead of the
2733 * largest available page, because even if the next movable
2734 * allocation falls back into a different pageblock than this
2735 * one, it won't cause permanent fragmentation.
2737 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2738 && current_order > order)
2747 for (current_order = order; current_order < MAX_ORDER;
2749 area = &(zone->free_area[current_order]);
2750 fallback_mt = find_suitable_fallback(area, current_order,
2751 start_migratetype, false, &can_steal);
2752 if (fallback_mt != -1)
2757 * This should not happen - we already found a suitable fallback
2758 * when looking for the largest page.
2760 VM_BUG_ON(current_order == MAX_ORDER);
2763 page = get_page_from_free_area(area, fallback_mt);
2765 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2768 trace_mm_page_alloc_extfrag(page, order, current_order,
2769 start_migratetype, fallback_mt);
2776 * Do the hard work of removing an element from the buddy allocator.
2777 * Call me with the zone->lock already held.
2779 static __always_inline struct page *
2780 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2781 unsigned int alloc_flags)
2787 * Balance movable allocations between regular and CMA areas by
2788 * allocating from CMA when over half of the zone's free memory
2789 * is in the CMA area.
2791 if (migratetype == MIGRATE_MOVABLE &&
2792 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2793 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2794 page = __rmqueue_cma_fallback(zone, order);
2800 page = __rmqueue_smallest(zone, order, migratetype);
2801 if (unlikely(!page)) {
2802 if (migratetype == MIGRATE_MOVABLE)
2803 page = __rmqueue_cma_fallback(zone, order);
2805 if (!page && __rmqueue_fallback(zone, order, migratetype,
2810 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2815 * Obtain a specified number of elements from the buddy allocator, all under
2816 * a single hold of the lock, for efficiency. Add them to the supplied list.
2817 * Returns the number of new pages which were placed at *list.
2819 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2820 unsigned long count, struct list_head *list,
2821 int migratetype, unsigned int alloc_flags)
2825 spin_lock(&zone->lock);
2826 for (i = 0; i < count; ++i) {
2827 struct page *page = __rmqueue(zone, order, migratetype,
2829 if (unlikely(page == NULL))
2832 if (unlikely(check_pcp_refill(page)))
2836 * Split buddy pages returned by expand() are received here in
2837 * physical page order. The page is added to the tail of
2838 * caller's list. From the callers perspective, the linked list
2839 * is ordered by page number under some conditions. This is
2840 * useful for IO devices that can forward direction from the
2841 * head, thus also in the physical page order. This is useful
2842 * for IO devices that can merge IO requests if the physical
2843 * pages are ordered properly.
2845 list_add_tail(&page->lru, list);
2847 if (is_migrate_cma(get_pcppage_migratetype(page)))
2848 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2853 * i pages were removed from the buddy list even if some leak due
2854 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2855 * on i. Do not confuse with 'alloced' which is the number of
2856 * pages added to the pcp list.
2858 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2859 spin_unlock(&zone->lock);
2865 * Called from the vmstat counter updater to drain pagesets of this
2866 * currently executing processor on remote nodes after they have
2869 * Note that this function must be called with the thread pinned to
2870 * a single processor.
2872 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2874 unsigned long flags;
2875 int to_drain, batch;
2877 local_irq_save(flags);
2878 batch = READ_ONCE(pcp->batch);
2879 to_drain = min(pcp->count, batch);
2881 free_pcppages_bulk(zone, to_drain, pcp);
2882 local_irq_restore(flags);
2887 * Drain pcplists of the indicated processor and zone.
2889 * The processor must either be the current processor and the
2890 * thread pinned to the current processor or a processor that
2893 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2895 unsigned long flags;
2896 struct per_cpu_pageset *pset;
2897 struct per_cpu_pages *pcp;
2899 local_irq_save(flags);
2900 pset = per_cpu_ptr(zone->pageset, cpu);
2904 free_pcppages_bulk(zone, pcp->count, pcp);
2905 local_irq_restore(flags);
2909 * Drain pcplists of all zones on the indicated processor.
2911 * The processor must either be the current processor and the
2912 * thread pinned to the current processor or a processor that
2915 static void drain_pages(unsigned int cpu)
2919 for_each_populated_zone(zone) {
2920 drain_pages_zone(cpu, zone);
2925 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2927 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2928 * the single zone's pages.
2930 void drain_local_pages(struct zone *zone)
2932 int cpu = smp_processor_id();
2935 drain_pages_zone(cpu, zone);
2940 static void drain_local_pages_wq(struct work_struct *work)
2942 struct pcpu_drain *drain;
2944 drain = container_of(work, struct pcpu_drain, work);
2947 * drain_all_pages doesn't use proper cpu hotplug protection so
2948 * we can race with cpu offline when the WQ can move this from
2949 * a cpu pinned worker to an unbound one. We can operate on a different
2950 * cpu which is allright but we also have to make sure to not move to
2954 drain_local_pages(drain->zone);
2959 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2961 * When zone parameter is non-NULL, spill just the single zone's pages.
2963 * Note that this can be extremely slow as the draining happens in a workqueue.
2965 void drain_all_pages(struct zone *zone)
2970 * Allocate in the BSS so we wont require allocation in
2971 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2973 static cpumask_t cpus_with_pcps;
2976 * Make sure nobody triggers this path before mm_percpu_wq is fully
2979 if (WARN_ON_ONCE(!mm_percpu_wq))
2983 * Do not drain if one is already in progress unless it's specific to
2984 * a zone. Such callers are primarily CMA and memory hotplug and need
2985 * the drain to be complete when the call returns.
2987 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2990 mutex_lock(&pcpu_drain_mutex);
2994 * We don't care about racing with CPU hotplug event
2995 * as offline notification will cause the notified
2996 * cpu to drain that CPU pcps and on_each_cpu_mask
2997 * disables preemption as part of its processing
2999 for_each_online_cpu(cpu) {
3000 struct per_cpu_pageset *pcp;
3002 bool has_pcps = false;
3005 pcp = per_cpu_ptr(zone->pageset, cpu);
3009 for_each_populated_zone(z) {
3010 pcp = per_cpu_ptr(z->pageset, cpu);
3011 if (pcp->pcp.count) {
3019 cpumask_set_cpu(cpu, &cpus_with_pcps);
3021 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3024 for_each_cpu(cpu, &cpus_with_pcps) {
3025 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3028 INIT_WORK(&drain->work, drain_local_pages_wq);
3029 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3031 for_each_cpu(cpu, &cpus_with_pcps)
3032 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3034 mutex_unlock(&pcpu_drain_mutex);
3037 #ifdef CONFIG_HIBERNATION
3040 * Touch the watchdog for every WD_PAGE_COUNT pages.
3042 #define WD_PAGE_COUNT (128*1024)
3044 void mark_free_pages(struct zone *zone)
3046 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3047 unsigned long flags;
3048 unsigned int order, t;
3051 if (zone_is_empty(zone))
3054 spin_lock_irqsave(&zone->lock, flags);
3056 max_zone_pfn = zone_end_pfn(zone);
3057 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3058 if (pfn_valid(pfn)) {
3059 page = pfn_to_page(pfn);
3061 if (!--page_count) {
3062 touch_nmi_watchdog();
3063 page_count = WD_PAGE_COUNT;
3066 if (page_zone(page) != zone)
3069 if (!swsusp_page_is_forbidden(page))
3070 swsusp_unset_page_free(page);
3073 for_each_migratetype_order(order, t) {
3074 list_for_each_entry(page,
3075 &zone->free_area[order].free_list[t], lru) {
3078 pfn = page_to_pfn(page);
3079 for (i = 0; i < (1UL << order); i++) {
3080 if (!--page_count) {
3081 touch_nmi_watchdog();
3082 page_count = WD_PAGE_COUNT;
3084 swsusp_set_page_free(pfn_to_page(pfn + i));
3088 spin_unlock_irqrestore(&zone->lock, flags);
3090 #endif /* CONFIG_PM */
3092 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3096 if (!free_pcp_prepare(page))
3099 migratetype = get_pfnblock_migratetype(page, pfn);
3100 set_pcppage_migratetype(page, migratetype);
3104 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3106 struct zone *zone = page_zone(page);
3107 struct per_cpu_pages *pcp;
3110 migratetype = get_pcppage_migratetype(page);
3111 __count_vm_event(PGFREE);
3114 * We only track unmovable, reclaimable and movable on pcp lists.
3115 * Free ISOLATE pages back to the allocator because they are being
3116 * offlined but treat HIGHATOMIC as movable pages so we can get those
3117 * areas back if necessary. Otherwise, we may have to free
3118 * excessively into the page allocator
3120 if (migratetype >= MIGRATE_PCPTYPES) {
3121 if (unlikely(is_migrate_isolate(migratetype))) {
3122 free_one_page(zone, page, pfn, 0, migratetype);
3125 migratetype = MIGRATE_MOVABLE;
3128 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3129 list_add(&page->lru, &pcp->lists[migratetype]);
3131 if (pcp->count >= pcp->high) {
3132 unsigned long batch = READ_ONCE(pcp->batch);
3133 free_pcppages_bulk(zone, batch, pcp);
3138 * Free a 0-order page
3140 void free_unref_page(struct page *page)
3142 unsigned long flags;
3143 unsigned long pfn = page_to_pfn(page);
3145 if (!free_unref_page_prepare(page, pfn))
3148 local_irq_save(flags);
3149 free_unref_page_commit(page, pfn);
3150 local_irq_restore(flags);
3154 * Free a list of 0-order pages
3156 void free_unref_page_list(struct list_head *list)
3158 struct page *page, *next;
3159 unsigned long flags, pfn;
3160 int batch_count = 0;
3162 /* Prepare pages for freeing */
3163 list_for_each_entry_safe(page, next, list, lru) {
3164 pfn = page_to_pfn(page);
3165 if (!free_unref_page_prepare(page, pfn))
3166 list_del(&page->lru);
3167 set_page_private(page, pfn);
3170 local_irq_save(flags);
3171 list_for_each_entry_safe(page, next, list, lru) {
3172 unsigned long pfn = page_private(page);
3174 set_page_private(page, 0);
3175 trace_mm_page_free_batched(page);
3176 free_unref_page_commit(page, pfn);
3179 * Guard against excessive IRQ disabled times when we get
3180 * a large list of pages to free.
3182 if (++batch_count == SWAP_CLUSTER_MAX) {
3183 local_irq_restore(flags);
3185 local_irq_save(flags);
3188 local_irq_restore(flags);
3192 * split_page takes a non-compound higher-order page, and splits it into
3193 * n (1<<order) sub-pages: page[0..n]
3194 * Each sub-page must be freed individually.
3196 * Note: this is probably too low level an operation for use in drivers.
3197 * Please consult with lkml before using this in your driver.
3199 void split_page(struct page *page, unsigned int order)
3203 VM_BUG_ON_PAGE(PageCompound(page), page);
3204 VM_BUG_ON_PAGE(!page_count(page), page);
3206 for (i = 1; i < (1 << order); i++)
3207 set_page_refcounted(page + i);
3208 split_page_owner(page, order);
3210 EXPORT_SYMBOL_GPL(split_page);
3212 int __isolate_free_page(struct page *page, unsigned int order)
3214 unsigned long watermark;
3218 BUG_ON(!PageBuddy(page));
3220 zone = page_zone(page);
3221 mt = get_pageblock_migratetype(page);
3223 if (!is_migrate_isolate(mt)) {
3225 * Obey watermarks as if the page was being allocated. We can
3226 * emulate a high-order watermark check with a raised order-0
3227 * watermark, because we already know our high-order page
3230 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3231 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3234 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3237 /* Remove page from free list */
3239 del_page_from_free_list(page, zone, order);
3242 * Set the pageblock if the isolated page is at least half of a
3245 if (order >= pageblock_order - 1) {
3246 struct page *endpage = page + (1 << order) - 1;
3247 for (; page < endpage; page += pageblock_nr_pages) {
3248 int mt = get_pageblock_migratetype(page);
3249 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3250 && !is_migrate_highatomic(mt))
3251 set_pageblock_migratetype(page,
3257 return 1UL << order;
3261 * __putback_isolated_page - Return a now-isolated page back where we got it
3262 * @page: Page that was isolated
3263 * @order: Order of the isolated page
3264 * @mt: The page's pageblock's migratetype
3266 * This function is meant to return a page pulled from the free lists via
3267 * __isolate_free_page back to the free lists they were pulled from.
3269 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3271 struct zone *zone = page_zone(page);
3273 /* zone lock should be held when this function is called */
3274 lockdep_assert_held(&zone->lock);
3276 /* Return isolated page to tail of freelist. */
3277 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3281 * Update NUMA hit/miss statistics
3283 * Must be called with interrupts disabled.
3285 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3288 enum numa_stat_item local_stat = NUMA_LOCAL;
3290 /* skip numa counters update if numa stats is disabled */
3291 if (!static_branch_likely(&vm_numa_stat_key))
3294 if (zone_to_nid(z) != numa_node_id())
3295 local_stat = NUMA_OTHER;
3297 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3298 __inc_numa_state(z, NUMA_HIT);
3300 __inc_numa_state(z, NUMA_MISS);
3301 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3303 __inc_numa_state(z, local_stat);
3307 /* Remove page from the per-cpu list, caller must protect the list */
3308 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3309 unsigned int alloc_flags,
3310 struct per_cpu_pages *pcp,
3311 struct list_head *list)
3316 if (list_empty(list)) {
3317 pcp->count += rmqueue_bulk(zone, 0,
3319 migratetype, alloc_flags);
3320 if (unlikely(list_empty(list)))
3324 page = list_first_entry(list, struct page, lru);
3325 list_del(&page->lru);
3327 } while (check_new_pcp(page));
3332 /* Lock and remove page from the per-cpu list */
3333 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3334 struct zone *zone, gfp_t gfp_flags,
3335 int migratetype, unsigned int alloc_flags)
3337 struct per_cpu_pages *pcp;
3338 struct list_head *list;
3340 unsigned long flags;
3342 local_irq_save(flags);
3343 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3344 list = &pcp->lists[migratetype];
3345 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3347 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3348 zone_statistics(preferred_zone, zone);
3350 local_irq_restore(flags);
3355 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3358 struct page *rmqueue(struct zone *preferred_zone,
3359 struct zone *zone, unsigned int order,
3360 gfp_t gfp_flags, unsigned int alloc_flags,
3363 unsigned long flags;
3366 if (likely(order == 0)) {
3367 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3368 migratetype, alloc_flags);
3373 * We most definitely don't want callers attempting to
3374 * allocate greater than order-1 page units with __GFP_NOFAIL.
3376 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3377 spin_lock_irqsave(&zone->lock, flags);
3381 if (alloc_flags & ALLOC_HARDER) {
3382 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3384 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3387 page = __rmqueue(zone, order, migratetype, alloc_flags);
3388 } while (page && check_new_pages(page, order));
3389 spin_unlock(&zone->lock);
3392 __mod_zone_freepage_state(zone, -(1 << order),
3393 get_pcppage_migratetype(page));
3395 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3396 zone_statistics(preferred_zone, zone);
3397 local_irq_restore(flags);
3400 /* Separate test+clear to avoid unnecessary atomics */
3401 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3402 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3403 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3406 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3410 local_irq_restore(flags);
3414 #ifdef CONFIG_FAIL_PAGE_ALLOC
3417 struct fault_attr attr;
3419 bool ignore_gfp_highmem;
3420 bool ignore_gfp_reclaim;
3422 } fail_page_alloc = {
3423 .attr = FAULT_ATTR_INITIALIZER,
3424 .ignore_gfp_reclaim = true,
3425 .ignore_gfp_highmem = true,
3429 static int __init setup_fail_page_alloc(char *str)
3431 return setup_fault_attr(&fail_page_alloc.attr, str);
3433 __setup("fail_page_alloc=", setup_fail_page_alloc);
3435 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3437 if (order < fail_page_alloc.min_order)
3439 if (gfp_mask & __GFP_NOFAIL)
3441 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3443 if (fail_page_alloc.ignore_gfp_reclaim &&
3444 (gfp_mask & __GFP_DIRECT_RECLAIM))
3447 return should_fail(&fail_page_alloc.attr, 1 << order);
3450 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3452 static int __init fail_page_alloc_debugfs(void)
3454 umode_t mode = S_IFREG | 0600;
3457 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3458 &fail_page_alloc.attr);
3460 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3461 &fail_page_alloc.ignore_gfp_reclaim);
3462 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3463 &fail_page_alloc.ignore_gfp_highmem);
3464 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3469 late_initcall(fail_page_alloc_debugfs);
3471 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3473 #else /* CONFIG_FAIL_PAGE_ALLOC */
3475 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3480 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3482 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3484 return __should_fail_alloc_page(gfp_mask, order);
3486 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3488 static inline long __zone_watermark_unusable_free(struct zone *z,
3489 unsigned int order, unsigned int alloc_flags)
3491 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3492 long unusable_free = (1 << order) - 1;
3495 * If the caller does not have rights to ALLOC_HARDER then subtract
3496 * the high-atomic reserves. This will over-estimate the size of the
3497 * atomic reserve but it avoids a search.
3499 if (likely(!alloc_harder))
3500 unusable_free += z->nr_reserved_highatomic;
3503 /* If allocation can't use CMA areas don't use free CMA pages */
3504 if (!(alloc_flags & ALLOC_CMA))
3505 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3508 return unusable_free;
3512 * Return true if free base pages are above 'mark'. For high-order checks it
3513 * will return true of the order-0 watermark is reached and there is at least
3514 * one free page of a suitable size. Checking now avoids taking the zone lock
3515 * to check in the allocation paths if no pages are free.
3517 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3518 int highest_zoneidx, unsigned int alloc_flags,
3523 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3525 /* free_pages may go negative - that's OK */
3526 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3528 if (alloc_flags & ALLOC_HIGH)
3531 if (unlikely(alloc_harder)) {
3533 * OOM victims can try even harder than normal ALLOC_HARDER
3534 * users on the grounds that it's definitely going to be in
3535 * the exit path shortly and free memory. Any allocation it
3536 * makes during the free path will be small and short-lived.
3538 if (alloc_flags & ALLOC_OOM)
3545 * Check watermarks for an order-0 allocation request. If these
3546 * are not met, then a high-order request also cannot go ahead
3547 * even if a suitable page happened to be free.
3549 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3552 /* If this is an order-0 request then the watermark is fine */
3556 /* For a high-order request, check at least one suitable page is free */
3557 for (o = order; o < MAX_ORDER; o++) {
3558 struct free_area *area = &z->free_area[o];
3564 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3565 if (!free_area_empty(area, mt))
3570 if ((alloc_flags & ALLOC_CMA) &&
3571 !free_area_empty(area, MIGRATE_CMA)) {
3575 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3581 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3582 int highest_zoneidx, unsigned int alloc_flags)
3584 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3585 zone_page_state(z, NR_FREE_PAGES));
3588 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3589 unsigned long mark, int highest_zoneidx,
3590 unsigned int alloc_flags, gfp_t gfp_mask)
3594 free_pages = zone_page_state(z, NR_FREE_PAGES);
3597 * Fast check for order-0 only. If this fails then the reserves
3598 * need to be calculated.
3603 fast_free = free_pages;
3604 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3605 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3609 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3613 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3614 * when checking the min watermark. The min watermark is the
3615 * point where boosting is ignored so that kswapd is woken up
3616 * when below the low watermark.
3618 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3619 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3620 mark = z->_watermark[WMARK_MIN];
3621 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3622 alloc_flags, free_pages);
3628 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3629 unsigned long mark, int highest_zoneidx)
3631 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3633 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3634 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3636 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3641 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3643 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3644 node_reclaim_distance;
3646 #else /* CONFIG_NUMA */
3647 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3651 #endif /* CONFIG_NUMA */
3654 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3655 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3656 * premature use of a lower zone may cause lowmem pressure problems that
3657 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3658 * probably too small. It only makes sense to spread allocations to avoid
3659 * fragmentation between the Normal and DMA32 zones.
3661 static inline unsigned int
3662 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3664 unsigned int alloc_flags;
3667 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3670 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3672 #ifdef CONFIG_ZONE_DMA32
3676 if (zone_idx(zone) != ZONE_NORMAL)
3680 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3681 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3682 * on UMA that if Normal is populated then so is DMA32.
3684 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3685 if (nr_online_nodes > 1 && !populated_zone(--zone))
3688 alloc_flags |= ALLOC_NOFRAGMENT;
3689 #endif /* CONFIG_ZONE_DMA32 */
3694 * get_page_from_freelist goes through the zonelist trying to allocate
3697 static struct page *
3698 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3699 const struct alloc_context *ac)
3703 struct pglist_data *last_pgdat_dirty_limit = NULL;
3708 * Scan zonelist, looking for a zone with enough free.
3709 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3711 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3712 z = ac->preferred_zoneref;
3713 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist,
3714 ac->highest_zoneidx, ac->nodemask) {
3718 if (cpusets_enabled() &&
3719 (alloc_flags & ALLOC_CPUSET) &&
3720 !__cpuset_zone_allowed(zone, gfp_mask))
3723 * When allocating a page cache page for writing, we
3724 * want to get it from a node that is within its dirty
3725 * limit, such that no single node holds more than its
3726 * proportional share of globally allowed dirty pages.
3727 * The dirty limits take into account the node's
3728 * lowmem reserves and high watermark so that kswapd
3729 * should be able to balance it without having to
3730 * write pages from its LRU list.
3732 * XXX: For now, allow allocations to potentially
3733 * exceed the per-node dirty limit in the slowpath
3734 * (spread_dirty_pages unset) before going into reclaim,
3735 * which is important when on a NUMA setup the allowed
3736 * nodes are together not big enough to reach the
3737 * global limit. The proper fix for these situations
3738 * will require awareness of nodes in the
3739 * dirty-throttling and the flusher threads.
3741 if (ac->spread_dirty_pages) {
3742 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3745 if (!node_dirty_ok(zone->zone_pgdat)) {
3746 last_pgdat_dirty_limit = zone->zone_pgdat;
3751 if (no_fallback && nr_online_nodes > 1 &&
3752 zone != ac->preferred_zoneref->zone) {
3756 * If moving to a remote node, retry but allow
3757 * fragmenting fallbacks. Locality is more important
3758 * than fragmentation avoidance.
3760 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3761 if (zone_to_nid(zone) != local_nid) {
3762 alloc_flags &= ~ALLOC_NOFRAGMENT;
3767 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3768 if (!zone_watermark_fast(zone, order, mark,
3769 ac->highest_zoneidx, alloc_flags,
3773 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3775 * Watermark failed for this zone, but see if we can
3776 * grow this zone if it contains deferred pages.
3778 if (static_branch_unlikely(&deferred_pages)) {
3779 if (_deferred_grow_zone(zone, order))
3783 /* Checked here to keep the fast path fast */
3784 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3785 if (alloc_flags & ALLOC_NO_WATERMARKS)
3788 if (node_reclaim_mode == 0 ||
3789 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3792 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3794 case NODE_RECLAIM_NOSCAN:
3797 case NODE_RECLAIM_FULL:
3798 /* scanned but unreclaimable */
3801 /* did we reclaim enough */
3802 if (zone_watermark_ok(zone, order, mark,
3803 ac->highest_zoneidx, alloc_flags))
3811 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3812 gfp_mask, alloc_flags, ac->migratetype);
3814 prep_new_page(page, order, gfp_mask, alloc_flags);
3817 * If this is a high-order atomic allocation then check
3818 * if the pageblock should be reserved for the future
3820 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3821 reserve_highatomic_pageblock(page, zone, order);
3825 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3826 /* Try again if zone has deferred pages */
3827 if (static_branch_unlikely(&deferred_pages)) {
3828 if (_deferred_grow_zone(zone, order))
3836 * It's possible on a UMA machine to get through all zones that are
3837 * fragmented. If avoiding fragmentation, reset and try again.
3840 alloc_flags &= ~ALLOC_NOFRAGMENT;
3847 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3849 unsigned int filter = SHOW_MEM_FILTER_NODES;
3852 * This documents exceptions given to allocations in certain
3853 * contexts that are allowed to allocate outside current's set
3856 if (!(gfp_mask & __GFP_NOMEMALLOC))
3857 if (tsk_is_oom_victim(current) ||
3858 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3859 filter &= ~SHOW_MEM_FILTER_NODES;
3860 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3861 filter &= ~SHOW_MEM_FILTER_NODES;
3863 show_mem(filter, nodemask);
3866 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3868 struct va_format vaf;
3870 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3872 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3875 va_start(args, fmt);
3878 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3879 current->comm, &vaf, gfp_mask, &gfp_mask,
3880 nodemask_pr_args(nodemask));
3883 cpuset_print_current_mems_allowed();
3886 warn_alloc_show_mem(gfp_mask, nodemask);
3889 static inline struct page *
3890 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3891 unsigned int alloc_flags,
3892 const struct alloc_context *ac)
3896 page = get_page_from_freelist(gfp_mask, order,
3897 alloc_flags|ALLOC_CPUSET, ac);
3899 * fallback to ignore cpuset restriction if our nodes
3903 page = get_page_from_freelist(gfp_mask, order,
3909 static inline struct page *
3910 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3911 const struct alloc_context *ac, unsigned long *did_some_progress)
3913 struct oom_control oc = {
3914 .zonelist = ac->zonelist,
3915 .nodemask = ac->nodemask,
3917 .gfp_mask = gfp_mask,
3922 *did_some_progress = 0;
3925 * Acquire the oom lock. If that fails, somebody else is
3926 * making progress for us.
3928 if (!mutex_trylock(&oom_lock)) {
3929 *did_some_progress = 1;
3930 schedule_timeout_uninterruptible(1);
3935 * Go through the zonelist yet one more time, keep very high watermark
3936 * here, this is only to catch a parallel oom killing, we must fail if
3937 * we're still under heavy pressure. But make sure that this reclaim
3938 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3939 * allocation which will never fail due to oom_lock already held.
3941 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3942 ~__GFP_DIRECT_RECLAIM, order,
3943 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3947 /* Coredumps can quickly deplete all memory reserves */
3948 if (current->flags & PF_DUMPCORE)
3950 /* The OOM killer will not help higher order allocs */
3951 if (order > PAGE_ALLOC_COSTLY_ORDER)
3954 * We have already exhausted all our reclaim opportunities without any
3955 * success so it is time to admit defeat. We will skip the OOM killer
3956 * because it is very likely that the caller has a more reasonable
3957 * fallback than shooting a random task.
3959 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3961 /* The OOM killer does not needlessly kill tasks for lowmem */
3962 if (ac->highest_zoneidx < ZONE_NORMAL)
3964 if (pm_suspended_storage())
3967 * XXX: GFP_NOFS allocations should rather fail than rely on
3968 * other request to make a forward progress.
3969 * We are in an unfortunate situation where out_of_memory cannot
3970 * do much for this context but let's try it to at least get
3971 * access to memory reserved if the current task is killed (see
3972 * out_of_memory). Once filesystems are ready to handle allocation
3973 * failures more gracefully we should just bail out here.
3976 /* The OOM killer may not free memory on a specific node */
3977 if (gfp_mask & __GFP_THISNODE)
3980 /* Exhausted what can be done so it's blame time */
3981 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3982 *did_some_progress = 1;
3985 * Help non-failing allocations by giving them access to memory
3988 if (gfp_mask & __GFP_NOFAIL)
3989 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3990 ALLOC_NO_WATERMARKS, ac);
3993 mutex_unlock(&oom_lock);
3998 * Maximum number of compaction retries wit a progress before OOM
3999 * killer is consider as the only way to move forward.
4001 #define MAX_COMPACT_RETRIES 16
4003 #ifdef CONFIG_COMPACTION
4004 /* Try memory compaction for high-order allocations before reclaim */
4005 static struct page *
4006 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4007 unsigned int alloc_flags, const struct alloc_context *ac,
4008 enum compact_priority prio, enum compact_result *compact_result)
4010 struct page *page = NULL;
4011 unsigned long pflags;
4012 unsigned int noreclaim_flag;
4017 psi_memstall_enter(&pflags);
4018 noreclaim_flag = memalloc_noreclaim_save();
4020 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4023 memalloc_noreclaim_restore(noreclaim_flag);
4024 psi_memstall_leave(&pflags);
4027 * At least in one zone compaction wasn't deferred or skipped, so let's
4028 * count a compaction stall
4030 count_vm_event(COMPACTSTALL);
4032 /* Prep a captured page if available */
4034 prep_new_page(page, order, gfp_mask, alloc_flags);
4036 /* Try get a page from the freelist if available */
4038 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4041 struct zone *zone = page_zone(page);
4043 zone->compact_blockskip_flush = false;
4044 compaction_defer_reset(zone, order, true);
4045 count_vm_event(COMPACTSUCCESS);
4050 * It's bad if compaction run occurs and fails. The most likely reason
4051 * is that pages exist, but not enough to satisfy watermarks.
4053 count_vm_event(COMPACTFAIL);
4061 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4062 enum compact_result compact_result,
4063 enum compact_priority *compact_priority,
4064 int *compaction_retries)
4066 int max_retries = MAX_COMPACT_RETRIES;
4069 int retries = *compaction_retries;
4070 enum compact_priority priority = *compact_priority;
4075 if (compaction_made_progress(compact_result))
4076 (*compaction_retries)++;
4079 * compaction considers all the zone as desperately out of memory
4080 * so it doesn't really make much sense to retry except when the
4081 * failure could be caused by insufficient priority
4083 if (compaction_failed(compact_result))
4084 goto check_priority;
4087 * compaction was skipped because there are not enough order-0 pages
4088 * to work with, so we retry only if it looks like reclaim can help.
4090 if (compaction_needs_reclaim(compact_result)) {
4091 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4096 * make sure the compaction wasn't deferred or didn't bail out early
4097 * due to locks contention before we declare that we should give up.
4098 * But the next retry should use a higher priority if allowed, so
4099 * we don't just keep bailing out endlessly.
4101 if (compaction_withdrawn(compact_result)) {
4102 goto check_priority;
4106 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4107 * costly ones because they are de facto nofail and invoke OOM
4108 * killer to move on while costly can fail and users are ready
4109 * to cope with that. 1/4 retries is rather arbitrary but we
4110 * would need much more detailed feedback from compaction to
4111 * make a better decision.
4113 if (order > PAGE_ALLOC_COSTLY_ORDER)
4115 if (*compaction_retries <= max_retries) {
4121 * Make sure there are attempts at the highest priority if we exhausted
4122 * all retries or failed at the lower priorities.
4125 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4126 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4128 if (*compact_priority > min_priority) {
4129 (*compact_priority)--;
4130 *compaction_retries = 0;
4134 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4138 static inline struct page *
4139 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4140 unsigned int alloc_flags, const struct alloc_context *ac,
4141 enum compact_priority prio, enum compact_result *compact_result)
4143 *compact_result = COMPACT_SKIPPED;
4148 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4149 enum compact_result compact_result,
4150 enum compact_priority *compact_priority,
4151 int *compaction_retries)
4156 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4160 * There are setups with compaction disabled which would prefer to loop
4161 * inside the allocator rather than hit the oom killer prematurely.
4162 * Let's give them a good hope and keep retrying while the order-0
4163 * watermarks are OK.
4165 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4166 ac->highest_zoneidx, ac->nodemask) {
4167 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4168 ac->highest_zoneidx, alloc_flags))
4173 #endif /* CONFIG_COMPACTION */
4175 #ifdef CONFIG_LOCKDEP
4176 static struct lockdep_map __fs_reclaim_map =
4177 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4179 static bool __need_fs_reclaim(gfp_t gfp_mask)
4181 gfp_mask = current_gfp_context(gfp_mask);
4183 /* no reclaim without waiting on it */
4184 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4187 /* this guy won't enter reclaim */
4188 if (current->flags & PF_MEMALLOC)
4191 /* We're only interested __GFP_FS allocations for now */
4192 if (!(gfp_mask & __GFP_FS))
4195 if (gfp_mask & __GFP_NOLOCKDEP)
4201 void __fs_reclaim_acquire(void)
4203 lock_map_acquire(&__fs_reclaim_map);
4206 void __fs_reclaim_release(void)
4208 lock_map_release(&__fs_reclaim_map);
4211 void fs_reclaim_acquire(gfp_t gfp_mask)
4213 if (__need_fs_reclaim(gfp_mask))
4214 __fs_reclaim_acquire();
4216 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4218 void fs_reclaim_release(gfp_t gfp_mask)
4220 if (__need_fs_reclaim(gfp_mask))
4221 __fs_reclaim_release();
4223 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4226 /* Perform direct synchronous page reclaim */
4228 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4229 const struct alloc_context *ac)
4232 unsigned int noreclaim_flag;
4233 unsigned long pflags;
4237 /* We now go into synchronous reclaim */
4238 cpuset_memory_pressure_bump();
4239 psi_memstall_enter(&pflags);
4240 fs_reclaim_acquire(gfp_mask);
4241 noreclaim_flag = memalloc_noreclaim_save();
4243 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4246 memalloc_noreclaim_restore(noreclaim_flag);
4247 fs_reclaim_release(gfp_mask);
4248 psi_memstall_leave(&pflags);
4255 /* The really slow allocator path where we enter direct reclaim */
4256 static inline struct page *
4257 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4258 unsigned int alloc_flags, const struct alloc_context *ac,
4259 unsigned long *did_some_progress)
4261 struct page *page = NULL;
4262 bool drained = false;
4264 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4265 if (unlikely(!(*did_some_progress)))
4269 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4272 * If an allocation failed after direct reclaim, it could be because
4273 * pages are pinned on the per-cpu lists or in high alloc reserves.
4274 * Shrink them them and try again
4276 if (!page && !drained) {
4277 unreserve_highatomic_pageblock(ac, false);
4278 drain_all_pages(NULL);
4286 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4287 const struct alloc_context *ac)
4291 pg_data_t *last_pgdat = NULL;
4292 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4294 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4296 if (last_pgdat != zone->zone_pgdat)
4297 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4298 last_pgdat = zone->zone_pgdat;
4302 static inline unsigned int
4303 gfp_to_alloc_flags(gfp_t gfp_mask)
4305 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4308 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4309 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4310 * to save two branches.
4312 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4313 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4316 * The caller may dip into page reserves a bit more if the caller
4317 * cannot run direct reclaim, or if the caller has realtime scheduling
4318 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4319 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4321 alloc_flags |= (__force int)
4322 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4324 if (gfp_mask & __GFP_ATOMIC) {
4326 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4327 * if it can't schedule.
4329 if (!(gfp_mask & __GFP_NOMEMALLOC))
4330 alloc_flags |= ALLOC_HARDER;
4332 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4333 * comment for __cpuset_node_allowed().
4335 alloc_flags &= ~ALLOC_CPUSET;
4336 } else if (unlikely(rt_task(current)) && !in_interrupt())
4337 alloc_flags |= ALLOC_HARDER;
4340 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4341 alloc_flags |= ALLOC_CMA;
4346 static bool oom_reserves_allowed(struct task_struct *tsk)
4348 if (!tsk_is_oom_victim(tsk))
4352 * !MMU doesn't have oom reaper so give access to memory reserves
4353 * only to the thread with TIF_MEMDIE set
4355 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4362 * Distinguish requests which really need access to full memory
4363 * reserves from oom victims which can live with a portion of it
4365 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4367 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4369 if (gfp_mask & __GFP_MEMALLOC)
4370 return ALLOC_NO_WATERMARKS;
4371 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4372 return ALLOC_NO_WATERMARKS;
4373 if (!in_interrupt()) {
4374 if (current->flags & PF_MEMALLOC)
4375 return ALLOC_NO_WATERMARKS;
4376 else if (oom_reserves_allowed(current))
4383 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4385 return !!__gfp_pfmemalloc_flags(gfp_mask);
4389 * Checks whether it makes sense to retry the reclaim to make a forward progress
4390 * for the given allocation request.
4392 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4393 * without success, or when we couldn't even meet the watermark if we
4394 * reclaimed all remaining pages on the LRU lists.
4396 * Returns true if a retry is viable or false to enter the oom path.
4399 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4400 struct alloc_context *ac, int alloc_flags,
4401 bool did_some_progress, int *no_progress_loops)
4408 * Costly allocations might have made a progress but this doesn't mean
4409 * their order will become available due to high fragmentation so
4410 * always increment the no progress counter for them
4412 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4413 *no_progress_loops = 0;
4415 (*no_progress_loops)++;
4418 * Make sure we converge to OOM if we cannot make any progress
4419 * several times in the row.
4421 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4422 /* Before OOM, exhaust highatomic_reserve */
4423 return unreserve_highatomic_pageblock(ac, true);
4427 * Keep reclaiming pages while there is a chance this will lead
4428 * somewhere. If none of the target zones can satisfy our allocation
4429 * request even if all reclaimable pages are considered then we are
4430 * screwed and have to go OOM.
4432 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4433 ac->highest_zoneidx, ac->nodemask) {
4434 unsigned long available;
4435 unsigned long reclaimable;
4436 unsigned long min_wmark = min_wmark_pages(zone);
4439 available = reclaimable = zone_reclaimable_pages(zone);
4440 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4443 * Would the allocation succeed if we reclaimed all
4444 * reclaimable pages?
4446 wmark = __zone_watermark_ok(zone, order, min_wmark,
4447 ac->highest_zoneidx, alloc_flags, available);
4448 trace_reclaim_retry_zone(z, order, reclaimable,
4449 available, min_wmark, *no_progress_loops, wmark);
4452 * If we didn't make any progress and have a lot of
4453 * dirty + writeback pages then we should wait for
4454 * an IO to complete to slow down the reclaim and
4455 * prevent from pre mature OOM
4457 if (!did_some_progress) {
4458 unsigned long write_pending;
4460 write_pending = zone_page_state_snapshot(zone,
4461 NR_ZONE_WRITE_PENDING);
4463 if (2 * write_pending > reclaimable) {
4464 congestion_wait(BLK_RW_ASYNC, HZ/10);
4476 * Memory allocation/reclaim might be called from a WQ context and the
4477 * current implementation of the WQ concurrency control doesn't
4478 * recognize that a particular WQ is congested if the worker thread is
4479 * looping without ever sleeping. Therefore we have to do a short sleep
4480 * here rather than calling cond_resched().
4482 if (current->flags & PF_WQ_WORKER)
4483 schedule_timeout_uninterruptible(1);
4490 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4493 * It's possible that cpuset's mems_allowed and the nodemask from
4494 * mempolicy don't intersect. This should be normally dealt with by
4495 * policy_nodemask(), but it's possible to race with cpuset update in
4496 * such a way the check therein was true, and then it became false
4497 * before we got our cpuset_mems_cookie here.
4498 * This assumes that for all allocations, ac->nodemask can come only
4499 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4500 * when it does not intersect with the cpuset restrictions) or the
4501 * caller can deal with a violated nodemask.
4503 if (cpusets_enabled() && ac->nodemask &&
4504 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4505 ac->nodemask = NULL;
4510 * When updating a task's mems_allowed or mempolicy nodemask, it is
4511 * possible to race with parallel threads in such a way that our
4512 * allocation can fail while the mask is being updated. If we are about
4513 * to fail, check if the cpuset changed during allocation and if so,
4516 if (read_mems_allowed_retry(cpuset_mems_cookie))
4522 static inline struct page *
4523 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4524 struct alloc_context *ac)
4526 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4527 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4528 struct page *page = NULL;
4529 unsigned int alloc_flags;
4530 unsigned long did_some_progress;
4531 enum compact_priority compact_priority;
4532 enum compact_result compact_result;
4533 int compaction_retries;
4534 int no_progress_loops;
4535 unsigned int cpuset_mems_cookie;
4539 * We also sanity check to catch abuse of atomic reserves being used by
4540 * callers that are not in atomic context.
4542 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4543 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4544 gfp_mask &= ~__GFP_ATOMIC;
4547 compaction_retries = 0;
4548 no_progress_loops = 0;
4549 compact_priority = DEF_COMPACT_PRIORITY;
4550 cpuset_mems_cookie = read_mems_allowed_begin();
4553 * The fast path uses conservative alloc_flags to succeed only until
4554 * kswapd needs to be woken up, and to avoid the cost of setting up
4555 * alloc_flags precisely. So we do that now.
4557 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4560 * We need to recalculate the starting point for the zonelist iterator
4561 * because we might have used different nodemask in the fast path, or
4562 * there was a cpuset modification and we are retrying - otherwise we
4563 * could end up iterating over non-eligible zones endlessly.
4565 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4566 ac->highest_zoneidx, ac->nodemask);
4567 if (!ac->preferred_zoneref->zone)
4570 if (alloc_flags & ALLOC_KSWAPD)
4571 wake_all_kswapds(order, gfp_mask, ac);
4574 * The adjusted alloc_flags might result in immediate success, so try
4577 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4582 * For costly allocations, try direct compaction first, as it's likely
4583 * that we have enough base pages and don't need to reclaim. For non-
4584 * movable high-order allocations, do that as well, as compaction will
4585 * try prevent permanent fragmentation by migrating from blocks of the
4587 * Don't try this for allocations that are allowed to ignore
4588 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4590 if (can_direct_reclaim &&
4592 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4593 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4594 page = __alloc_pages_direct_compact(gfp_mask, order,
4596 INIT_COMPACT_PRIORITY,
4602 * Checks for costly allocations with __GFP_NORETRY, which
4603 * includes some THP page fault allocations
4605 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4607 * If allocating entire pageblock(s) and compaction
4608 * failed because all zones are below low watermarks
4609 * or is prohibited because it recently failed at this
4610 * order, fail immediately unless the allocator has
4611 * requested compaction and reclaim retry.
4614 * - potentially very expensive because zones are far
4615 * below their low watermarks or this is part of very
4616 * bursty high order allocations,
4617 * - not guaranteed to help because isolate_freepages()
4618 * may not iterate over freed pages as part of its
4620 * - unlikely to make entire pageblocks free on its
4623 if (compact_result == COMPACT_SKIPPED ||
4624 compact_result == COMPACT_DEFERRED)
4628 * Looks like reclaim/compaction is worth trying, but
4629 * sync compaction could be very expensive, so keep
4630 * using async compaction.
4632 compact_priority = INIT_COMPACT_PRIORITY;
4637 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4638 if (alloc_flags & ALLOC_KSWAPD)
4639 wake_all_kswapds(order, gfp_mask, ac);
4641 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4643 alloc_flags = reserve_flags;
4646 * Reset the nodemask and zonelist iterators if memory policies can be
4647 * ignored. These allocations are high priority and system rather than
4650 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4651 ac->nodemask = NULL;
4652 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4653 ac->highest_zoneidx, ac->nodemask);
4656 /* Attempt with potentially adjusted zonelist and alloc_flags */
4657 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4661 /* Caller is not willing to reclaim, we can't balance anything */
4662 if (!can_direct_reclaim)
4665 /* Avoid recursion of direct reclaim */
4666 if (current->flags & PF_MEMALLOC)
4669 /* Try direct reclaim and then allocating */
4670 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4671 &did_some_progress);
4675 /* Try direct compaction and then allocating */
4676 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4677 compact_priority, &compact_result);
4681 /* Do not loop if specifically requested */
4682 if (gfp_mask & __GFP_NORETRY)
4686 * Do not retry costly high order allocations unless they are
4687 * __GFP_RETRY_MAYFAIL
4689 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4692 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4693 did_some_progress > 0, &no_progress_loops))
4697 * It doesn't make any sense to retry for the compaction if the order-0
4698 * reclaim is not able to make any progress because the current
4699 * implementation of the compaction depends on the sufficient amount
4700 * of free memory (see __compaction_suitable)
4702 if (did_some_progress > 0 &&
4703 should_compact_retry(ac, order, alloc_flags,
4704 compact_result, &compact_priority,
4705 &compaction_retries))
4709 /* Deal with possible cpuset update races before we start OOM killing */
4710 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4713 /* Reclaim has failed us, start killing things */
4714 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4718 /* Avoid allocations with no watermarks from looping endlessly */
4719 if (tsk_is_oom_victim(current) &&
4720 (alloc_flags == ALLOC_OOM ||
4721 (gfp_mask & __GFP_NOMEMALLOC)))
4724 /* Retry as long as the OOM killer is making progress */
4725 if (did_some_progress) {
4726 no_progress_loops = 0;
4731 /* Deal with possible cpuset update races before we fail */
4732 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4736 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4739 if (gfp_mask & __GFP_NOFAIL) {
4741 * All existing users of the __GFP_NOFAIL are blockable, so warn
4742 * of any new users that actually require GFP_NOWAIT
4744 if (WARN_ON_ONCE(!can_direct_reclaim))
4748 * PF_MEMALLOC request from this context is rather bizarre
4749 * because we cannot reclaim anything and only can loop waiting
4750 * for somebody to do a work for us
4752 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4755 * non failing costly orders are a hard requirement which we
4756 * are not prepared for much so let's warn about these users
4757 * so that we can identify them and convert them to something
4760 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4763 * Help non-failing allocations by giving them access to memory
4764 * reserves but do not use ALLOC_NO_WATERMARKS because this
4765 * could deplete whole memory reserves which would just make
4766 * the situation worse
4768 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4776 warn_alloc(gfp_mask, ac->nodemask,
4777 "page allocation failure: order:%u", order);
4782 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4783 int preferred_nid, nodemask_t *nodemask,
4784 struct alloc_context *ac, gfp_t *alloc_mask,
4785 unsigned int *alloc_flags)
4787 ac->highest_zoneidx = gfp_zone(gfp_mask);
4788 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4789 ac->nodemask = nodemask;
4790 ac->migratetype = gfp_migratetype(gfp_mask);
4792 if (cpusets_enabled()) {
4793 *alloc_mask |= __GFP_HARDWALL;
4795 ac->nodemask = &cpuset_current_mems_allowed;
4797 *alloc_flags |= ALLOC_CPUSET;
4800 fs_reclaim_acquire(gfp_mask);
4801 fs_reclaim_release(gfp_mask);
4803 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4805 if (should_fail_alloc_page(gfp_mask, order))
4808 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4809 *alloc_flags |= ALLOC_CMA;
4814 /* Determine whether to spread dirty pages and what the first usable zone */
4815 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4817 /* Dirty zone balancing only done in the fast path */
4818 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4821 * The preferred zone is used for statistics but crucially it is
4822 * also used as the starting point for the zonelist iterator. It
4823 * may get reset for allocations that ignore memory policies.
4825 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4826 ac->highest_zoneidx, ac->nodemask);
4830 * This is the 'heart' of the zoned buddy allocator.
4833 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4834 nodemask_t *nodemask)
4837 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4838 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4839 struct alloc_context ac = { };
4842 * There are several places where we assume that the order value is sane
4843 * so bail out early if the request is out of bound.
4845 if (unlikely(order >= MAX_ORDER)) {
4846 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4850 gfp_mask &= gfp_allowed_mask;
4851 alloc_mask = gfp_mask;
4852 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4855 finalise_ac(gfp_mask, &ac);
4858 * Forbid the first pass from falling back to types that fragment
4859 * memory until all local zones are considered.
4861 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4863 /* First allocation attempt */
4864 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4869 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4870 * resp. GFP_NOIO which has to be inherited for all allocation requests
4871 * from a particular context which has been marked by
4872 * memalloc_no{fs,io}_{save,restore}.
4874 alloc_mask = current_gfp_context(gfp_mask);
4875 ac.spread_dirty_pages = false;
4878 * Restore the original nodemask if it was potentially replaced with
4879 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4881 ac.nodemask = nodemask;
4883 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4886 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4887 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4888 __free_pages(page, order);
4892 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4896 EXPORT_SYMBOL(__alloc_pages_nodemask);
4899 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4900 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4901 * you need to access high mem.
4903 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4907 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4910 return (unsigned long) page_address(page);
4912 EXPORT_SYMBOL(__get_free_pages);
4914 unsigned long get_zeroed_page(gfp_t gfp_mask)
4916 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4918 EXPORT_SYMBOL(get_zeroed_page);
4920 static inline void free_the_page(struct page *page, unsigned int order)
4922 if (order == 0) /* Via pcp? */
4923 free_unref_page(page);
4925 __free_pages_ok(page, order);
4928 void __free_pages(struct page *page, unsigned int order)
4930 if (put_page_testzero(page))
4931 free_the_page(page, order);
4933 EXPORT_SYMBOL(__free_pages);
4935 void free_pages(unsigned long addr, unsigned int order)
4938 VM_BUG_ON(!virt_addr_valid((void *)addr));
4939 __free_pages(virt_to_page((void *)addr), order);
4943 EXPORT_SYMBOL(free_pages);
4947 * An arbitrary-length arbitrary-offset area of memory which resides
4948 * within a 0 or higher order page. Multiple fragments within that page
4949 * are individually refcounted, in the page's reference counter.
4951 * The page_frag functions below provide a simple allocation framework for
4952 * page fragments. This is used by the network stack and network device
4953 * drivers to provide a backing region of memory for use as either an
4954 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4956 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4959 struct page *page = NULL;
4960 gfp_t gfp = gfp_mask;
4962 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4963 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4965 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4966 PAGE_FRAG_CACHE_MAX_ORDER);
4967 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4969 if (unlikely(!page))
4970 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4972 nc->va = page ? page_address(page) : NULL;
4977 void __page_frag_cache_drain(struct page *page, unsigned int count)
4979 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4981 if (page_ref_sub_and_test(page, count))
4982 free_the_page(page, compound_order(page));
4984 EXPORT_SYMBOL(__page_frag_cache_drain);
4986 void *page_frag_alloc(struct page_frag_cache *nc,
4987 unsigned int fragsz, gfp_t gfp_mask)
4989 unsigned int size = PAGE_SIZE;
4993 if (unlikely(!nc->va)) {
4995 page = __page_frag_cache_refill(nc, gfp_mask);
4999 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5000 /* if size can vary use size else just use PAGE_SIZE */
5003 /* Even if we own the page, we do not use atomic_set().
5004 * This would break get_page_unless_zero() users.
5006 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5008 /* reset page count bias and offset to start of new frag */
5009 nc->pfmemalloc = page_is_pfmemalloc(page);
5010 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5014 offset = nc->offset - fragsz;
5015 if (unlikely(offset < 0)) {
5016 page = virt_to_page(nc->va);
5018 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5021 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5022 /* if size can vary use size else just use PAGE_SIZE */
5025 /* OK, page count is 0, we can safely set it */
5026 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5028 /* reset page count bias and offset to start of new frag */
5029 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5030 offset = size - fragsz;
5034 nc->offset = offset;
5036 return nc->va + offset;
5038 EXPORT_SYMBOL(page_frag_alloc);
5041 * Frees a page fragment allocated out of either a compound or order 0 page.
5043 void page_frag_free(void *addr)
5045 struct page *page = virt_to_head_page(addr);
5047 if (unlikely(put_page_testzero(page)))
5048 free_the_page(page, compound_order(page));
5050 EXPORT_SYMBOL(page_frag_free);
5052 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5056 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5057 unsigned long used = addr + PAGE_ALIGN(size);
5059 split_page(virt_to_page((void *)addr), order);
5060 while (used < alloc_end) {
5065 return (void *)addr;
5069 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5070 * @size: the number of bytes to allocate
5071 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5073 * This function is similar to alloc_pages(), except that it allocates the
5074 * minimum number of pages to satisfy the request. alloc_pages() can only
5075 * allocate memory in power-of-two pages.
5077 * This function is also limited by MAX_ORDER.
5079 * Memory allocated by this function must be released by free_pages_exact().
5081 * Return: pointer to the allocated area or %NULL in case of error.
5083 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5085 unsigned int order = get_order(size);
5088 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5089 gfp_mask &= ~__GFP_COMP;
5091 addr = __get_free_pages(gfp_mask, order);
5092 return make_alloc_exact(addr, order, size);
5094 EXPORT_SYMBOL(alloc_pages_exact);
5097 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5099 * @nid: the preferred node ID where memory should be allocated
5100 * @size: the number of bytes to allocate
5101 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5103 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5106 * Return: pointer to the allocated area or %NULL in case of error.
5108 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5110 unsigned int order = get_order(size);
5113 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5114 gfp_mask &= ~__GFP_COMP;
5116 p = alloc_pages_node(nid, gfp_mask, order);
5119 return make_alloc_exact((unsigned long)page_address(p), order, size);
5123 * free_pages_exact - release memory allocated via alloc_pages_exact()
5124 * @virt: the value returned by alloc_pages_exact.
5125 * @size: size of allocation, same value as passed to alloc_pages_exact().
5127 * Release the memory allocated by a previous call to alloc_pages_exact.
5129 void free_pages_exact(void *virt, size_t size)
5131 unsigned long addr = (unsigned long)virt;
5132 unsigned long end = addr + PAGE_ALIGN(size);
5134 while (addr < end) {
5139 EXPORT_SYMBOL(free_pages_exact);
5142 * nr_free_zone_pages - count number of pages beyond high watermark
5143 * @offset: The zone index of the highest zone
5145 * nr_free_zone_pages() counts the number of pages which are beyond the
5146 * high watermark within all zones at or below a given zone index. For each
5147 * zone, the number of pages is calculated as:
5149 * nr_free_zone_pages = managed_pages - high_pages
5151 * Return: number of pages beyond high watermark.
5153 static unsigned long nr_free_zone_pages(int offset)
5158 /* Just pick one node, since fallback list is circular */
5159 unsigned long sum = 0;
5161 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5163 for_each_zone_zonelist(zone, z, zonelist, offset) {
5164 unsigned long size = zone_managed_pages(zone);
5165 unsigned long high = high_wmark_pages(zone);
5174 * nr_free_buffer_pages - count number of pages beyond high watermark
5176 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5177 * watermark within ZONE_DMA and ZONE_NORMAL.
5179 * Return: number of pages beyond high watermark within ZONE_DMA and
5182 unsigned long nr_free_buffer_pages(void)
5184 return nr_free_zone_pages(gfp_zone(GFP_USER));
5186 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5188 static inline void show_node(struct zone *zone)
5190 if (IS_ENABLED(CONFIG_NUMA))
5191 printk("Node %d ", zone_to_nid(zone));
5194 long si_mem_available(void)
5197 unsigned long pagecache;
5198 unsigned long wmark_low = 0;
5199 unsigned long pages[NR_LRU_LISTS];
5200 unsigned long reclaimable;
5204 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5205 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5208 wmark_low += low_wmark_pages(zone);
5211 * Estimate the amount of memory available for userspace allocations,
5212 * without causing swapping.
5214 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5217 * Not all the page cache can be freed, otherwise the system will
5218 * start swapping. Assume at least half of the page cache, or the
5219 * low watermark worth of cache, needs to stay.
5221 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5222 pagecache -= min(pagecache / 2, wmark_low);
5223 available += pagecache;
5226 * Part of the reclaimable slab and other kernel memory consists of
5227 * items that are in use, and cannot be freed. Cap this estimate at the
5230 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5231 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5232 available += reclaimable - min(reclaimable / 2, wmark_low);
5238 EXPORT_SYMBOL_GPL(si_mem_available);
5240 void si_meminfo(struct sysinfo *val)
5242 val->totalram = totalram_pages();
5243 val->sharedram = global_node_page_state(NR_SHMEM);
5244 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5245 val->bufferram = nr_blockdev_pages();
5246 val->totalhigh = totalhigh_pages();
5247 val->freehigh = nr_free_highpages();
5248 val->mem_unit = PAGE_SIZE;
5251 EXPORT_SYMBOL(si_meminfo);
5254 void si_meminfo_node(struct sysinfo *val, int nid)
5256 int zone_type; /* needs to be signed */
5257 unsigned long managed_pages = 0;
5258 unsigned long managed_highpages = 0;
5259 unsigned long free_highpages = 0;
5260 pg_data_t *pgdat = NODE_DATA(nid);
5262 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5263 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5264 val->totalram = managed_pages;
5265 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5266 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5267 #ifdef CONFIG_HIGHMEM
5268 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5269 struct zone *zone = &pgdat->node_zones[zone_type];
5271 if (is_highmem(zone)) {
5272 managed_highpages += zone_managed_pages(zone);
5273 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5276 val->totalhigh = managed_highpages;
5277 val->freehigh = free_highpages;
5279 val->totalhigh = managed_highpages;
5280 val->freehigh = free_highpages;
5282 val->mem_unit = PAGE_SIZE;
5287 * Determine whether the node should be displayed or not, depending on whether
5288 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5290 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5292 if (!(flags & SHOW_MEM_FILTER_NODES))
5296 * no node mask - aka implicit memory numa policy. Do not bother with
5297 * the synchronization - read_mems_allowed_begin - because we do not
5298 * have to be precise here.
5301 nodemask = &cpuset_current_mems_allowed;
5303 return !node_isset(nid, *nodemask);
5306 #define K(x) ((x) << (PAGE_SHIFT-10))
5308 static void show_migration_types(unsigned char type)
5310 static const char types[MIGRATE_TYPES] = {
5311 [MIGRATE_UNMOVABLE] = 'U',
5312 [MIGRATE_MOVABLE] = 'M',
5313 [MIGRATE_RECLAIMABLE] = 'E',
5314 [MIGRATE_HIGHATOMIC] = 'H',
5316 [MIGRATE_CMA] = 'C',
5318 #ifdef CONFIG_MEMORY_ISOLATION
5319 [MIGRATE_ISOLATE] = 'I',
5322 char tmp[MIGRATE_TYPES + 1];
5326 for (i = 0; i < MIGRATE_TYPES; i++) {
5327 if (type & (1 << i))
5332 printk(KERN_CONT "(%s) ", tmp);
5336 * Show free area list (used inside shift_scroll-lock stuff)
5337 * We also calculate the percentage fragmentation. We do this by counting the
5338 * memory on each free list with the exception of the first item on the list.
5341 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5344 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5346 unsigned long free_pcp = 0;
5351 for_each_populated_zone(zone) {
5352 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5355 for_each_online_cpu(cpu)
5356 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5359 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5360 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5361 " unevictable:%lu dirty:%lu writeback:%lu\n"
5362 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5363 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5364 " free:%lu free_pcp:%lu free_cma:%lu\n",
5365 global_node_page_state(NR_ACTIVE_ANON),
5366 global_node_page_state(NR_INACTIVE_ANON),
5367 global_node_page_state(NR_ISOLATED_ANON),
5368 global_node_page_state(NR_ACTIVE_FILE),
5369 global_node_page_state(NR_INACTIVE_FILE),
5370 global_node_page_state(NR_ISOLATED_FILE),
5371 global_node_page_state(NR_UNEVICTABLE),
5372 global_node_page_state(NR_FILE_DIRTY),
5373 global_node_page_state(NR_WRITEBACK),
5374 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5375 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5376 global_node_page_state(NR_FILE_MAPPED),
5377 global_node_page_state(NR_SHMEM),
5378 global_zone_page_state(NR_PAGETABLE),
5379 global_zone_page_state(NR_BOUNCE),
5380 global_zone_page_state(NR_FREE_PAGES),
5382 global_zone_page_state(NR_FREE_CMA_PAGES));
5384 for_each_online_pgdat(pgdat) {
5385 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5389 " active_anon:%lukB"
5390 " inactive_anon:%lukB"
5391 " active_file:%lukB"
5392 " inactive_file:%lukB"
5393 " unevictable:%lukB"
5394 " isolated(anon):%lukB"
5395 " isolated(file):%lukB"
5400 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5402 " shmem_pmdmapped: %lukB"
5405 " writeback_tmp:%lukB"
5406 " kernel_stack:%lukB"
5407 #ifdef CONFIG_SHADOW_CALL_STACK
5408 " shadow_call_stack:%lukB"
5410 " all_unreclaimable? %s"
5413 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5414 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5415 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5416 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5417 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5418 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5419 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5420 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5421 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5422 K(node_page_state(pgdat, NR_WRITEBACK)),
5423 K(node_page_state(pgdat, NR_SHMEM)),
5424 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5425 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5426 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5428 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5430 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5431 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5432 #ifdef CONFIG_SHADOW_CALL_STACK
5433 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5435 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5439 for_each_populated_zone(zone) {
5442 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5446 for_each_online_cpu(cpu)
5447 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5456 " reserved_highatomic:%luKB"
5457 " active_anon:%lukB"
5458 " inactive_anon:%lukB"
5459 " active_file:%lukB"
5460 " inactive_file:%lukB"
5461 " unevictable:%lukB"
5462 " writepending:%lukB"
5473 K(zone_page_state(zone, NR_FREE_PAGES)),
5474 K(min_wmark_pages(zone)),
5475 K(low_wmark_pages(zone)),
5476 K(high_wmark_pages(zone)),
5477 K(zone->nr_reserved_highatomic),
5478 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5479 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5480 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5481 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5482 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5483 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5484 K(zone->present_pages),
5485 K(zone_managed_pages(zone)),
5486 K(zone_page_state(zone, NR_MLOCK)),
5487 K(zone_page_state(zone, NR_PAGETABLE)),
5488 K(zone_page_state(zone, NR_BOUNCE)),
5490 K(this_cpu_read(zone->pageset->pcp.count)),
5491 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5492 printk("lowmem_reserve[]:");
5493 for (i = 0; i < MAX_NR_ZONES; i++)
5494 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5495 printk(KERN_CONT "\n");
5498 for_each_populated_zone(zone) {
5500 unsigned long nr[MAX_ORDER], flags, total = 0;
5501 unsigned char types[MAX_ORDER];
5503 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5506 printk(KERN_CONT "%s: ", zone->name);
5508 spin_lock_irqsave(&zone->lock, flags);
5509 for (order = 0; order < MAX_ORDER; order++) {
5510 struct free_area *area = &zone->free_area[order];
5513 nr[order] = area->nr_free;
5514 total += nr[order] << order;
5517 for (type = 0; type < MIGRATE_TYPES; type++) {
5518 if (!free_area_empty(area, type))
5519 types[order] |= 1 << type;
5522 spin_unlock_irqrestore(&zone->lock, flags);
5523 for (order = 0; order < MAX_ORDER; order++) {
5524 printk(KERN_CONT "%lu*%lukB ",
5525 nr[order], K(1UL) << order);
5527 show_migration_types(types[order]);
5529 printk(KERN_CONT "= %lukB\n", K(total));
5532 hugetlb_show_meminfo();
5534 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5536 show_swap_cache_info();
5539 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5541 zoneref->zone = zone;
5542 zoneref->zone_idx = zone_idx(zone);
5546 * Builds allocation fallback zone lists.
5548 * Add all populated zones of a node to the zonelist.
5550 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5553 enum zone_type zone_type = MAX_NR_ZONES;
5558 zone = pgdat->node_zones + zone_type;
5559 if (managed_zone(zone)) {
5560 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5561 check_highest_zone(zone_type);
5563 } while (zone_type);
5570 static int __parse_numa_zonelist_order(char *s)
5573 * We used to support different zonlists modes but they turned
5574 * out to be just not useful. Let's keep the warning in place
5575 * if somebody still use the cmd line parameter so that we do
5576 * not fail it silently
5578 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5579 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5585 char numa_zonelist_order[] = "Node";
5588 * sysctl handler for numa_zonelist_order
5590 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5591 void *buffer, size_t *length, loff_t *ppos)
5594 return __parse_numa_zonelist_order(buffer);
5595 return proc_dostring(table, write, buffer, length, ppos);
5599 #define MAX_NODE_LOAD (nr_online_nodes)
5600 static int node_load[MAX_NUMNODES];
5603 * find_next_best_node - find the next node that should appear in a given node's fallback list
5604 * @node: node whose fallback list we're appending
5605 * @used_node_mask: nodemask_t of already used nodes
5607 * We use a number of factors to determine which is the next node that should
5608 * appear on a given node's fallback list. The node should not have appeared
5609 * already in @node's fallback list, and it should be the next closest node
5610 * according to the distance array (which contains arbitrary distance values
5611 * from each node to each node in the system), and should also prefer nodes
5612 * with no CPUs, since presumably they'll have very little allocation pressure
5613 * on them otherwise.
5615 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5617 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5620 int min_val = INT_MAX;
5621 int best_node = NUMA_NO_NODE;
5622 const struct cpumask *tmp = cpumask_of_node(0);
5624 /* Use the local node if we haven't already */
5625 if (!node_isset(node, *used_node_mask)) {
5626 node_set(node, *used_node_mask);
5630 for_each_node_state(n, N_MEMORY) {
5632 /* Don't want a node to appear more than once */
5633 if (node_isset(n, *used_node_mask))
5636 /* Use the distance array to find the distance */
5637 val = node_distance(node, n);
5639 /* Penalize nodes under us ("prefer the next node") */
5642 /* Give preference to headless and unused nodes */
5643 tmp = cpumask_of_node(n);
5644 if (!cpumask_empty(tmp))
5645 val += PENALTY_FOR_NODE_WITH_CPUS;
5647 /* Slight preference for less loaded node */
5648 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5649 val += node_load[n];
5651 if (val < min_val) {
5658 node_set(best_node, *used_node_mask);
5665 * Build zonelists ordered by node and zones within node.
5666 * This results in maximum locality--normal zone overflows into local
5667 * DMA zone, if any--but risks exhausting DMA zone.
5669 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5672 struct zoneref *zonerefs;
5675 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5677 for (i = 0; i < nr_nodes; i++) {
5680 pg_data_t *node = NODE_DATA(node_order[i]);
5682 nr_zones = build_zonerefs_node(node, zonerefs);
5683 zonerefs += nr_zones;
5685 zonerefs->zone = NULL;
5686 zonerefs->zone_idx = 0;
5690 * Build gfp_thisnode zonelists
5692 static void build_thisnode_zonelists(pg_data_t *pgdat)
5694 struct zoneref *zonerefs;
5697 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5698 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5699 zonerefs += nr_zones;
5700 zonerefs->zone = NULL;
5701 zonerefs->zone_idx = 0;
5705 * Build zonelists ordered by zone and nodes within zones.
5706 * This results in conserving DMA zone[s] until all Normal memory is
5707 * exhausted, but results in overflowing to remote node while memory
5708 * may still exist in local DMA zone.
5711 static void build_zonelists(pg_data_t *pgdat)
5713 static int node_order[MAX_NUMNODES];
5714 int node, load, nr_nodes = 0;
5715 nodemask_t used_mask = NODE_MASK_NONE;
5716 int local_node, prev_node;
5718 /* NUMA-aware ordering of nodes */
5719 local_node = pgdat->node_id;
5720 load = nr_online_nodes;
5721 prev_node = local_node;
5723 memset(node_order, 0, sizeof(node_order));
5724 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5726 * We don't want to pressure a particular node.
5727 * So adding penalty to the first node in same
5728 * distance group to make it round-robin.
5730 if (node_distance(local_node, node) !=
5731 node_distance(local_node, prev_node))
5732 node_load[node] = load;
5734 node_order[nr_nodes++] = node;
5739 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5740 build_thisnode_zonelists(pgdat);
5743 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5745 * Return node id of node used for "local" allocations.
5746 * I.e., first node id of first zone in arg node's generic zonelist.
5747 * Used for initializing percpu 'numa_mem', which is used primarily
5748 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5750 int local_memory_node(int node)
5754 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5755 gfp_zone(GFP_KERNEL),
5757 return zone_to_nid(z->zone);
5761 static void setup_min_unmapped_ratio(void);
5762 static void setup_min_slab_ratio(void);
5763 #else /* CONFIG_NUMA */
5765 static void build_zonelists(pg_data_t *pgdat)
5767 int node, local_node;
5768 struct zoneref *zonerefs;
5771 local_node = pgdat->node_id;
5773 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5774 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5775 zonerefs += nr_zones;
5778 * Now we build the zonelist so that it contains the zones
5779 * of all the other nodes.
5780 * We don't want to pressure a particular node, so when
5781 * building the zones for node N, we make sure that the
5782 * zones coming right after the local ones are those from
5783 * node N+1 (modulo N)
5785 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5786 if (!node_online(node))
5788 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5789 zonerefs += nr_zones;
5791 for (node = 0; node < local_node; node++) {
5792 if (!node_online(node))
5794 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5795 zonerefs += nr_zones;
5798 zonerefs->zone = NULL;
5799 zonerefs->zone_idx = 0;
5802 #endif /* CONFIG_NUMA */
5805 * Boot pageset table. One per cpu which is going to be used for all
5806 * zones and all nodes. The parameters will be set in such a way
5807 * that an item put on a list will immediately be handed over to
5808 * the buddy list. This is safe since pageset manipulation is done
5809 * with interrupts disabled.
5811 * The boot_pagesets must be kept even after bootup is complete for
5812 * unused processors and/or zones. They do play a role for bootstrapping
5813 * hotplugged processors.
5815 * zoneinfo_show() and maybe other functions do
5816 * not check if the processor is online before following the pageset pointer.
5817 * Other parts of the kernel may not check if the zone is available.
5819 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5820 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5821 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5823 static void __build_all_zonelists(void *data)
5826 int __maybe_unused cpu;
5827 pg_data_t *self = data;
5828 static DEFINE_SPINLOCK(lock);
5833 memset(node_load, 0, sizeof(node_load));
5837 * This node is hotadded and no memory is yet present. So just
5838 * building zonelists is fine - no need to touch other nodes.
5840 if (self && !node_online(self->node_id)) {
5841 build_zonelists(self);
5843 for_each_online_node(nid) {
5844 pg_data_t *pgdat = NODE_DATA(nid);
5846 build_zonelists(pgdat);
5849 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5851 * We now know the "local memory node" for each node--
5852 * i.e., the node of the first zone in the generic zonelist.
5853 * Set up numa_mem percpu variable for on-line cpus. During
5854 * boot, only the boot cpu should be on-line; we'll init the
5855 * secondary cpus' numa_mem as they come on-line. During
5856 * node/memory hotplug, we'll fixup all on-line cpus.
5858 for_each_online_cpu(cpu)
5859 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5866 static noinline void __init
5867 build_all_zonelists_init(void)
5871 __build_all_zonelists(NULL);
5874 * Initialize the boot_pagesets that are going to be used
5875 * for bootstrapping processors. The real pagesets for
5876 * each zone will be allocated later when the per cpu
5877 * allocator is available.
5879 * boot_pagesets are used also for bootstrapping offline
5880 * cpus if the system is already booted because the pagesets
5881 * are needed to initialize allocators on a specific cpu too.
5882 * F.e. the percpu allocator needs the page allocator which
5883 * needs the percpu allocator in order to allocate its pagesets
5884 * (a chicken-egg dilemma).
5886 for_each_possible_cpu(cpu)
5887 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5889 mminit_verify_zonelist();
5890 cpuset_init_current_mems_allowed();
5894 * unless system_state == SYSTEM_BOOTING.
5896 * __ref due to call of __init annotated helper build_all_zonelists_init
5897 * [protected by SYSTEM_BOOTING].
5899 void __ref build_all_zonelists(pg_data_t *pgdat)
5901 unsigned long vm_total_pages;
5903 if (system_state == SYSTEM_BOOTING) {
5904 build_all_zonelists_init();
5906 __build_all_zonelists(pgdat);
5907 /* cpuset refresh routine should be here */
5909 /* Get the number of free pages beyond high watermark in all zones. */
5910 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5912 * Disable grouping by mobility if the number of pages in the
5913 * system is too low to allow the mechanism to work. It would be
5914 * more accurate, but expensive to check per-zone. This check is
5915 * made on memory-hotadd so a system can start with mobility
5916 * disabled and enable it later
5918 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5919 page_group_by_mobility_disabled = 1;
5921 page_group_by_mobility_disabled = 0;
5923 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5925 page_group_by_mobility_disabled ? "off" : "on",
5928 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5932 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5933 static bool __meminit
5934 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5936 static struct memblock_region *r;
5938 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5939 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5940 for_each_memblock(memory, r) {
5941 if (*pfn < memblock_region_memory_end_pfn(r))
5945 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5946 memblock_is_mirror(r)) {
5947 *pfn = memblock_region_memory_end_pfn(r);
5955 * Initially all pages are reserved - free ones are freed
5956 * up by memblock_free_all() once the early boot process is
5957 * done. Non-atomic initialization, single-pass.
5959 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5960 unsigned long start_pfn, enum memmap_context context,
5961 struct vmem_altmap *altmap)
5963 unsigned long pfn, end_pfn = start_pfn + size;
5966 if (highest_memmap_pfn < end_pfn - 1)
5967 highest_memmap_pfn = end_pfn - 1;
5969 #ifdef CONFIG_ZONE_DEVICE
5971 * Honor reservation requested by the driver for this ZONE_DEVICE
5972 * memory. We limit the total number of pages to initialize to just
5973 * those that might contain the memory mapping. We will defer the
5974 * ZONE_DEVICE page initialization until after we have released
5977 if (zone == ZONE_DEVICE) {
5981 if (start_pfn == altmap->base_pfn)
5982 start_pfn += altmap->reserve;
5983 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5987 for (pfn = start_pfn; pfn < end_pfn; ) {
5989 * There can be holes in boot-time mem_map[]s handed to this
5990 * function. They do not exist on hotplugged memory.
5992 if (context == MEMMAP_EARLY) {
5993 if (overlap_memmap_init(zone, &pfn))
5995 if (defer_init(nid, pfn, end_pfn))
5999 page = pfn_to_page(pfn);
6000 __init_single_page(page, pfn, zone, nid);
6001 if (context == MEMMAP_HOTPLUG)
6002 __SetPageReserved(page);
6005 * Mark the block movable so that blocks are reserved for
6006 * movable at startup. This will force kernel allocations
6007 * to reserve their blocks rather than leaking throughout
6008 * the address space during boot when many long-lived
6009 * kernel allocations are made.
6011 * bitmap is created for zone's valid pfn range. but memmap
6012 * can be created for invalid pages (for alignment)
6013 * check here not to call set_pageblock_migratetype() against
6016 if (!(pfn & (pageblock_nr_pages - 1))) {
6017 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6024 #ifdef CONFIG_ZONE_DEVICE
6025 void __ref memmap_init_zone_device(struct zone *zone,
6026 unsigned long start_pfn,
6027 unsigned long nr_pages,
6028 struct dev_pagemap *pgmap)
6030 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6031 struct pglist_data *pgdat = zone->zone_pgdat;
6032 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6033 unsigned long zone_idx = zone_idx(zone);
6034 unsigned long start = jiffies;
6035 int nid = pgdat->node_id;
6037 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6041 * The call to memmap_init_zone should have already taken care
6042 * of the pages reserved for the memmap, so we can just jump to
6043 * the end of that region and start processing the device pages.
6046 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6047 nr_pages = end_pfn - start_pfn;
6050 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6051 struct page *page = pfn_to_page(pfn);
6053 __init_single_page(page, pfn, zone_idx, nid);
6056 * Mark page reserved as it will need to wait for onlining
6057 * phase for it to be fully associated with a zone.
6059 * We can use the non-atomic __set_bit operation for setting
6060 * the flag as we are still initializing the pages.
6062 __SetPageReserved(page);
6065 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6066 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6067 * ever freed or placed on a driver-private list.
6069 page->pgmap = pgmap;
6070 page->zone_device_data = NULL;
6073 * Mark the block movable so that blocks are reserved for
6074 * movable at startup. This will force kernel allocations
6075 * to reserve their blocks rather than leaking throughout
6076 * the address space during boot when many long-lived
6077 * kernel allocations are made.
6079 * bitmap is created for zone's valid pfn range. but memmap
6080 * can be created for invalid pages (for alignment)
6081 * check here not to call set_pageblock_migratetype() against
6084 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6085 * because this is done early in section_activate()
6087 if (!(pfn & (pageblock_nr_pages - 1))) {
6088 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6093 pr_info("%s initialised %lu pages in %ums\n", __func__,
6094 nr_pages, jiffies_to_msecs(jiffies - start));
6098 static void __meminit zone_init_free_lists(struct zone *zone)
6100 unsigned int order, t;
6101 for_each_migratetype_order(order, t) {
6102 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6103 zone->free_area[order].nr_free = 0;
6107 void __meminit __weak memmap_init(unsigned long size, int nid,
6109 unsigned long range_start_pfn)
6111 unsigned long start_pfn, end_pfn;
6112 unsigned long range_end_pfn = range_start_pfn + size;
6115 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6116 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6117 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6119 if (end_pfn > start_pfn) {
6120 size = end_pfn - start_pfn;
6121 memmap_init_zone(size, nid, zone, start_pfn,
6122 MEMMAP_EARLY, NULL);
6127 static int zone_batchsize(struct zone *zone)
6133 * The per-cpu-pages pools are set to around 1000th of the
6136 batch = zone_managed_pages(zone) / 1024;
6137 /* But no more than a meg. */
6138 if (batch * PAGE_SIZE > 1024 * 1024)
6139 batch = (1024 * 1024) / PAGE_SIZE;
6140 batch /= 4; /* We effectively *= 4 below */
6145 * Clamp the batch to a 2^n - 1 value. Having a power
6146 * of 2 value was found to be more likely to have
6147 * suboptimal cache aliasing properties in some cases.
6149 * For example if 2 tasks are alternately allocating
6150 * batches of pages, one task can end up with a lot
6151 * of pages of one half of the possible page colors
6152 * and the other with pages of the other colors.
6154 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6159 /* The deferral and batching of frees should be suppressed under NOMMU
6162 * The problem is that NOMMU needs to be able to allocate large chunks
6163 * of contiguous memory as there's no hardware page translation to
6164 * assemble apparent contiguous memory from discontiguous pages.
6166 * Queueing large contiguous runs of pages for batching, however,
6167 * causes the pages to actually be freed in smaller chunks. As there
6168 * can be a significant delay between the individual batches being
6169 * recycled, this leads to the once large chunks of space being
6170 * fragmented and becoming unavailable for high-order allocations.
6177 * pcp->high and pcp->batch values are related and dependent on one another:
6178 * ->batch must never be higher then ->high.
6179 * The following function updates them in a safe manner without read side
6182 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6183 * those fields changing asynchronously (acording the the above rule).
6185 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6186 * outside of boot time (or some other assurance that no concurrent updaters
6189 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6190 unsigned long batch)
6192 /* start with a fail safe value for batch */
6196 /* Update high, then batch, in order */
6203 /* a companion to pageset_set_high() */
6204 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6206 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6209 static void pageset_init(struct per_cpu_pageset *p)
6211 struct per_cpu_pages *pcp;
6214 memset(p, 0, sizeof(*p));
6217 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6218 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6221 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6224 pageset_set_batch(p, batch);
6228 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6229 * to the value high for the pageset p.
6231 static void pageset_set_high(struct per_cpu_pageset *p,
6234 unsigned long batch = max(1UL, high / 4);
6235 if ((high / 4) > (PAGE_SHIFT * 8))
6236 batch = PAGE_SHIFT * 8;
6238 pageset_update(&p->pcp, high, batch);
6241 static void pageset_set_high_and_batch(struct zone *zone,
6242 struct per_cpu_pageset *pcp)
6244 if (percpu_pagelist_fraction)
6245 pageset_set_high(pcp,
6246 (zone_managed_pages(zone) /
6247 percpu_pagelist_fraction));
6249 pageset_set_batch(pcp, zone_batchsize(zone));
6252 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6254 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6257 pageset_set_high_and_batch(zone, pcp);
6260 void __meminit setup_zone_pageset(struct zone *zone)
6263 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6264 for_each_possible_cpu(cpu)
6265 zone_pageset_init(zone, cpu);
6269 * Allocate per cpu pagesets and initialize them.
6270 * Before this call only boot pagesets were available.
6272 void __init setup_per_cpu_pageset(void)
6274 struct pglist_data *pgdat;
6276 int __maybe_unused cpu;
6278 for_each_populated_zone(zone)
6279 setup_zone_pageset(zone);
6283 * Unpopulated zones continue using the boot pagesets.
6284 * The numa stats for these pagesets need to be reset.
6285 * Otherwise, they will end up skewing the stats of
6286 * the nodes these zones are associated with.
6288 for_each_possible_cpu(cpu) {
6289 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6290 memset(pcp->vm_numa_stat_diff, 0,
6291 sizeof(pcp->vm_numa_stat_diff));
6295 for_each_online_pgdat(pgdat)
6296 pgdat->per_cpu_nodestats =
6297 alloc_percpu(struct per_cpu_nodestat);
6300 static __meminit void zone_pcp_init(struct zone *zone)
6303 * per cpu subsystem is not up at this point. The following code
6304 * relies on the ability of the linker to provide the
6305 * offset of a (static) per cpu variable into the per cpu area.
6307 zone->pageset = &boot_pageset;
6309 if (populated_zone(zone))
6310 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6311 zone->name, zone->present_pages,
6312 zone_batchsize(zone));
6315 void __meminit init_currently_empty_zone(struct zone *zone,
6316 unsigned long zone_start_pfn,
6319 struct pglist_data *pgdat = zone->zone_pgdat;
6320 int zone_idx = zone_idx(zone) + 1;
6322 if (zone_idx > pgdat->nr_zones)
6323 pgdat->nr_zones = zone_idx;
6325 zone->zone_start_pfn = zone_start_pfn;
6327 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6328 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6330 (unsigned long)zone_idx(zone),
6331 zone_start_pfn, (zone_start_pfn + size));
6333 zone_init_free_lists(zone);
6334 zone->initialized = 1;
6338 * get_pfn_range_for_nid - Return the start and end page frames for a node
6339 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6340 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6341 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6343 * It returns the start and end page frame of a node based on information
6344 * provided by memblock_set_node(). If called for a node
6345 * with no available memory, a warning is printed and the start and end
6348 void __init get_pfn_range_for_nid(unsigned int nid,
6349 unsigned long *start_pfn, unsigned long *end_pfn)
6351 unsigned long this_start_pfn, this_end_pfn;
6357 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6358 *start_pfn = min(*start_pfn, this_start_pfn);
6359 *end_pfn = max(*end_pfn, this_end_pfn);
6362 if (*start_pfn == -1UL)
6367 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6368 * assumption is made that zones within a node are ordered in monotonic
6369 * increasing memory addresses so that the "highest" populated zone is used
6371 static void __init find_usable_zone_for_movable(void)
6374 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6375 if (zone_index == ZONE_MOVABLE)
6378 if (arch_zone_highest_possible_pfn[zone_index] >
6379 arch_zone_lowest_possible_pfn[zone_index])
6383 VM_BUG_ON(zone_index == -1);
6384 movable_zone = zone_index;
6388 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6389 * because it is sized independent of architecture. Unlike the other zones,
6390 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6391 * in each node depending on the size of each node and how evenly kernelcore
6392 * is distributed. This helper function adjusts the zone ranges
6393 * provided by the architecture for a given node by using the end of the
6394 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6395 * zones within a node are in order of monotonic increases memory addresses
6397 static void __init adjust_zone_range_for_zone_movable(int nid,
6398 unsigned long zone_type,
6399 unsigned long node_start_pfn,
6400 unsigned long node_end_pfn,
6401 unsigned long *zone_start_pfn,
6402 unsigned long *zone_end_pfn)
6404 /* Only adjust if ZONE_MOVABLE is on this node */
6405 if (zone_movable_pfn[nid]) {
6406 /* Size ZONE_MOVABLE */
6407 if (zone_type == ZONE_MOVABLE) {
6408 *zone_start_pfn = zone_movable_pfn[nid];
6409 *zone_end_pfn = min(node_end_pfn,
6410 arch_zone_highest_possible_pfn[movable_zone]);
6412 /* Adjust for ZONE_MOVABLE starting within this range */
6413 } else if (!mirrored_kernelcore &&
6414 *zone_start_pfn < zone_movable_pfn[nid] &&
6415 *zone_end_pfn > zone_movable_pfn[nid]) {
6416 *zone_end_pfn = zone_movable_pfn[nid];
6418 /* Check if this whole range is within ZONE_MOVABLE */
6419 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6420 *zone_start_pfn = *zone_end_pfn;
6425 * Return the number of pages a zone spans in a node, including holes
6426 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6428 static unsigned long __init zone_spanned_pages_in_node(int nid,
6429 unsigned long zone_type,
6430 unsigned long node_start_pfn,
6431 unsigned long node_end_pfn,
6432 unsigned long *zone_start_pfn,
6433 unsigned long *zone_end_pfn)
6435 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6436 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6437 /* When hotadd a new node from cpu_up(), the node should be empty */
6438 if (!node_start_pfn && !node_end_pfn)
6441 /* Get the start and end of the zone */
6442 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6443 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6444 adjust_zone_range_for_zone_movable(nid, zone_type,
6445 node_start_pfn, node_end_pfn,
6446 zone_start_pfn, zone_end_pfn);
6448 /* Check that this node has pages within the zone's required range */
6449 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6452 /* Move the zone boundaries inside the node if necessary */
6453 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6454 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6456 /* Return the spanned pages */
6457 return *zone_end_pfn - *zone_start_pfn;
6461 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6462 * then all holes in the requested range will be accounted for.
6464 unsigned long __init __absent_pages_in_range(int nid,
6465 unsigned long range_start_pfn,
6466 unsigned long range_end_pfn)
6468 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6469 unsigned long start_pfn, end_pfn;
6472 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6473 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6474 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6475 nr_absent -= end_pfn - start_pfn;
6481 * absent_pages_in_range - Return number of page frames in holes within a range
6482 * @start_pfn: The start PFN to start searching for holes
6483 * @end_pfn: The end PFN to stop searching for holes
6485 * Return: the number of pages frames in memory holes within a range.
6487 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6488 unsigned long end_pfn)
6490 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6493 /* Return the number of page frames in holes in a zone on a node */
6494 static unsigned long __init zone_absent_pages_in_node(int nid,
6495 unsigned long zone_type,
6496 unsigned long node_start_pfn,
6497 unsigned long node_end_pfn)
6499 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6500 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6501 unsigned long zone_start_pfn, zone_end_pfn;
6502 unsigned long nr_absent;
6504 /* When hotadd a new node from cpu_up(), the node should be empty */
6505 if (!node_start_pfn && !node_end_pfn)
6508 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6509 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6511 adjust_zone_range_for_zone_movable(nid, zone_type,
6512 node_start_pfn, node_end_pfn,
6513 &zone_start_pfn, &zone_end_pfn);
6514 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6517 * ZONE_MOVABLE handling.
6518 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6521 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6522 unsigned long start_pfn, end_pfn;
6523 struct memblock_region *r;
6525 for_each_memblock(memory, r) {
6526 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6527 zone_start_pfn, zone_end_pfn);
6528 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6529 zone_start_pfn, zone_end_pfn);
6531 if (zone_type == ZONE_MOVABLE &&
6532 memblock_is_mirror(r))
6533 nr_absent += end_pfn - start_pfn;
6535 if (zone_type == ZONE_NORMAL &&
6536 !memblock_is_mirror(r))
6537 nr_absent += end_pfn - start_pfn;
6544 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6545 unsigned long node_start_pfn,
6546 unsigned long node_end_pfn)
6548 unsigned long realtotalpages = 0, totalpages = 0;
6551 for (i = 0; i < MAX_NR_ZONES; i++) {
6552 struct zone *zone = pgdat->node_zones + i;
6553 unsigned long zone_start_pfn, zone_end_pfn;
6554 unsigned long spanned, absent;
6555 unsigned long size, real_size;
6557 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6562 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6567 real_size = size - absent;
6570 zone->zone_start_pfn = zone_start_pfn;
6572 zone->zone_start_pfn = 0;
6573 zone->spanned_pages = size;
6574 zone->present_pages = real_size;
6577 realtotalpages += real_size;
6580 pgdat->node_spanned_pages = totalpages;
6581 pgdat->node_present_pages = realtotalpages;
6582 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6586 #ifndef CONFIG_SPARSEMEM
6588 * Calculate the size of the zone->blockflags rounded to an unsigned long
6589 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6590 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6591 * round what is now in bits to nearest long in bits, then return it in
6594 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6596 unsigned long usemapsize;
6598 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6599 usemapsize = roundup(zonesize, pageblock_nr_pages);
6600 usemapsize = usemapsize >> pageblock_order;
6601 usemapsize *= NR_PAGEBLOCK_BITS;
6602 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6604 return usemapsize / 8;
6607 static void __ref setup_usemap(struct pglist_data *pgdat,
6609 unsigned long zone_start_pfn,
6610 unsigned long zonesize)
6612 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6613 zone->pageblock_flags = NULL;
6615 zone->pageblock_flags =
6616 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6618 if (!zone->pageblock_flags)
6619 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6620 usemapsize, zone->name, pgdat->node_id);
6624 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6625 unsigned long zone_start_pfn, unsigned long zonesize) {}
6626 #endif /* CONFIG_SPARSEMEM */
6628 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6630 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6631 void __init set_pageblock_order(void)
6635 /* Check that pageblock_nr_pages has not already been setup */
6636 if (pageblock_order)
6639 if (HPAGE_SHIFT > PAGE_SHIFT)
6640 order = HUGETLB_PAGE_ORDER;
6642 order = MAX_ORDER - 1;
6645 * Assume the largest contiguous order of interest is a huge page.
6646 * This value may be variable depending on boot parameters on IA64 and
6649 pageblock_order = order;
6651 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6654 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6655 * is unused as pageblock_order is set at compile-time. See
6656 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6659 void __init set_pageblock_order(void)
6663 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6665 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6666 unsigned long present_pages)
6668 unsigned long pages = spanned_pages;
6671 * Provide a more accurate estimation if there are holes within
6672 * the zone and SPARSEMEM is in use. If there are holes within the
6673 * zone, each populated memory region may cost us one or two extra
6674 * memmap pages due to alignment because memmap pages for each
6675 * populated regions may not be naturally aligned on page boundary.
6676 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6678 if (spanned_pages > present_pages + (present_pages >> 4) &&
6679 IS_ENABLED(CONFIG_SPARSEMEM))
6680 pages = present_pages;
6682 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6685 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6686 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6688 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6690 spin_lock_init(&ds_queue->split_queue_lock);
6691 INIT_LIST_HEAD(&ds_queue->split_queue);
6692 ds_queue->split_queue_len = 0;
6695 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6698 #ifdef CONFIG_COMPACTION
6699 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6701 init_waitqueue_head(&pgdat->kcompactd_wait);
6704 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6707 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6709 pgdat_resize_init(pgdat);
6711 pgdat_init_split_queue(pgdat);
6712 pgdat_init_kcompactd(pgdat);
6714 init_waitqueue_head(&pgdat->kswapd_wait);
6715 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6717 pgdat_page_ext_init(pgdat);
6718 spin_lock_init(&pgdat->lru_lock);
6719 lruvec_init(&pgdat->__lruvec);
6722 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6723 unsigned long remaining_pages)
6725 atomic_long_set(&zone->managed_pages, remaining_pages);
6726 zone_set_nid(zone, nid);
6727 zone->name = zone_names[idx];
6728 zone->zone_pgdat = NODE_DATA(nid);
6729 spin_lock_init(&zone->lock);
6730 zone_seqlock_init(zone);
6731 zone_pcp_init(zone);
6735 * Set up the zone data structures
6736 * - init pgdat internals
6737 * - init all zones belonging to this node
6739 * NOTE: this function is only called during memory hotplug
6741 #ifdef CONFIG_MEMORY_HOTPLUG
6742 void __ref free_area_init_core_hotplug(int nid)
6745 pg_data_t *pgdat = NODE_DATA(nid);
6747 pgdat_init_internals(pgdat);
6748 for (z = 0; z < MAX_NR_ZONES; z++)
6749 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6754 * Set up the zone data structures:
6755 * - mark all pages reserved
6756 * - mark all memory queues empty
6757 * - clear the memory bitmaps
6759 * NOTE: pgdat should get zeroed by caller.
6760 * NOTE: this function is only called during early init.
6762 static void __init free_area_init_core(struct pglist_data *pgdat)
6765 int nid = pgdat->node_id;
6767 pgdat_init_internals(pgdat);
6768 pgdat->per_cpu_nodestats = &boot_nodestats;
6770 for (j = 0; j < MAX_NR_ZONES; j++) {
6771 struct zone *zone = pgdat->node_zones + j;
6772 unsigned long size, freesize, memmap_pages;
6773 unsigned long zone_start_pfn = zone->zone_start_pfn;
6775 size = zone->spanned_pages;
6776 freesize = zone->present_pages;
6779 * Adjust freesize so that it accounts for how much memory
6780 * is used by this zone for memmap. This affects the watermark
6781 * and per-cpu initialisations
6783 memmap_pages = calc_memmap_size(size, freesize);
6784 if (!is_highmem_idx(j)) {
6785 if (freesize >= memmap_pages) {
6786 freesize -= memmap_pages;
6789 " %s zone: %lu pages used for memmap\n",
6790 zone_names[j], memmap_pages);
6792 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6793 zone_names[j], memmap_pages, freesize);
6796 /* Account for reserved pages */
6797 if (j == 0 && freesize > dma_reserve) {
6798 freesize -= dma_reserve;
6799 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6800 zone_names[0], dma_reserve);
6803 if (!is_highmem_idx(j))
6804 nr_kernel_pages += freesize;
6805 /* Charge for highmem memmap if there are enough kernel pages */
6806 else if (nr_kernel_pages > memmap_pages * 2)
6807 nr_kernel_pages -= memmap_pages;
6808 nr_all_pages += freesize;
6811 * Set an approximate value for lowmem here, it will be adjusted
6812 * when the bootmem allocator frees pages into the buddy system.
6813 * And all highmem pages will be managed by the buddy system.
6815 zone_init_internals(zone, j, nid, freesize);
6820 set_pageblock_order();
6821 setup_usemap(pgdat, zone, zone_start_pfn, size);
6822 init_currently_empty_zone(zone, zone_start_pfn, size);
6823 memmap_init(size, nid, j, zone_start_pfn);
6827 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6828 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6830 unsigned long __maybe_unused start = 0;
6831 unsigned long __maybe_unused offset = 0;
6833 /* Skip empty nodes */
6834 if (!pgdat->node_spanned_pages)
6837 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6838 offset = pgdat->node_start_pfn - start;
6839 /* ia64 gets its own node_mem_map, before this, without bootmem */
6840 if (!pgdat->node_mem_map) {
6841 unsigned long size, end;
6845 * The zone's endpoints aren't required to be MAX_ORDER
6846 * aligned but the node_mem_map endpoints must be in order
6847 * for the buddy allocator to function correctly.
6849 end = pgdat_end_pfn(pgdat);
6850 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6851 size = (end - start) * sizeof(struct page);
6852 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6855 panic("Failed to allocate %ld bytes for node %d memory map\n",
6856 size, pgdat->node_id);
6857 pgdat->node_mem_map = map + offset;
6859 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6860 __func__, pgdat->node_id, (unsigned long)pgdat,
6861 (unsigned long)pgdat->node_mem_map);
6862 #ifndef CONFIG_NEED_MULTIPLE_NODES
6864 * With no DISCONTIG, the global mem_map is just set as node 0's
6866 if (pgdat == NODE_DATA(0)) {
6867 mem_map = NODE_DATA(0)->node_mem_map;
6868 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6874 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6875 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6877 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6878 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6880 pgdat->first_deferred_pfn = ULONG_MAX;
6883 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6886 static void __init free_area_init_node(int nid)
6888 pg_data_t *pgdat = NODE_DATA(nid);
6889 unsigned long start_pfn = 0;
6890 unsigned long end_pfn = 0;
6892 /* pg_data_t should be reset to zero when it's allocated */
6893 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6895 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6897 pgdat->node_id = nid;
6898 pgdat->node_start_pfn = start_pfn;
6899 pgdat->per_cpu_nodestats = NULL;
6901 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6902 (u64)start_pfn << PAGE_SHIFT,
6903 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6904 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6906 alloc_node_mem_map(pgdat);
6907 pgdat_set_deferred_range(pgdat);
6909 free_area_init_core(pgdat);
6912 void __init free_area_init_memoryless_node(int nid)
6914 free_area_init_node(nid);
6917 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6919 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6920 * PageReserved(). Return the number of struct pages that were initialized.
6922 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6927 for (pfn = spfn; pfn < epfn; pfn++) {
6928 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6929 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6930 + pageblock_nr_pages - 1;
6934 * Use a fake node/zone (0) for now. Some of these pages
6935 * (in memblock.reserved but not in memblock.memory) will
6936 * get re-initialized via reserve_bootmem_region() later.
6938 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6939 __SetPageReserved(pfn_to_page(pfn));
6947 * Only struct pages that are backed by physical memory are zeroed and
6948 * initialized by going through __init_single_page(). But, there are some
6949 * struct pages which are reserved in memblock allocator and their fields
6950 * may be accessed (for example page_to_pfn() on some configuration accesses
6951 * flags). We must explicitly initialize those struct pages.
6953 * This function also addresses a similar issue where struct pages are left
6954 * uninitialized because the physical address range is not covered by
6955 * memblock.memory or memblock.reserved. That could happen when memblock
6956 * layout is manually configured via memmap=, or when the highest physical
6957 * address (max_pfn) does not end on a section boundary.
6959 static void __init init_unavailable_mem(void)
6961 phys_addr_t start, end;
6963 phys_addr_t next = 0;
6966 * Loop through unavailable ranges not covered by memblock.memory.
6969 for_each_mem_range(i, &memblock.memory, NULL,
6970 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6972 pgcnt += init_unavailable_range(PFN_DOWN(next),
6978 * Early sections always have a fully populated memmap for the whole
6979 * section - see pfn_valid(). If the last section has holes at the
6980 * end and that section is marked "online", the memmap will be
6981 * considered initialized. Make sure that memmap has a well defined
6984 pgcnt += init_unavailable_range(PFN_DOWN(next),
6985 round_up(max_pfn, PAGES_PER_SECTION));
6988 * Struct pages that do not have backing memory. This could be because
6989 * firmware is using some of this memory, or for some other reasons.
6992 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6995 static inline void __init init_unavailable_mem(void)
6998 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7000 #if MAX_NUMNODES > 1
7002 * Figure out the number of possible node ids.
7004 void __init setup_nr_node_ids(void)
7006 unsigned int highest;
7008 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7009 nr_node_ids = highest + 1;
7014 * node_map_pfn_alignment - determine the maximum internode alignment
7016 * This function should be called after node map is populated and sorted.
7017 * It calculates the maximum power of two alignment which can distinguish
7020 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7021 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7022 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7023 * shifted, 1GiB is enough and this function will indicate so.
7025 * This is used to test whether pfn -> nid mapping of the chosen memory
7026 * model has fine enough granularity to avoid incorrect mapping for the
7027 * populated node map.
7029 * Return: the determined alignment in pfn's. 0 if there is no alignment
7030 * requirement (single node).
7032 unsigned long __init node_map_pfn_alignment(void)
7034 unsigned long accl_mask = 0, last_end = 0;
7035 unsigned long start, end, mask;
7036 int last_nid = NUMA_NO_NODE;
7039 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7040 if (!start || last_nid < 0 || last_nid == nid) {
7047 * Start with a mask granular enough to pin-point to the
7048 * start pfn and tick off bits one-by-one until it becomes
7049 * too coarse to separate the current node from the last.
7051 mask = ~((1 << __ffs(start)) - 1);
7052 while (mask && last_end <= (start & (mask << 1)))
7055 /* accumulate all internode masks */
7059 /* convert mask to number of pages */
7060 return ~accl_mask + 1;
7064 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7066 * Return: the minimum PFN based on information provided via
7067 * memblock_set_node().
7069 unsigned long __init find_min_pfn_with_active_regions(void)
7071 return PHYS_PFN(memblock_start_of_DRAM());
7075 * early_calculate_totalpages()
7076 * Sum pages in active regions for movable zone.
7077 * Populate N_MEMORY for calculating usable_nodes.
7079 static unsigned long __init early_calculate_totalpages(void)
7081 unsigned long totalpages = 0;
7082 unsigned long start_pfn, end_pfn;
7085 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7086 unsigned long pages = end_pfn - start_pfn;
7088 totalpages += pages;
7090 node_set_state(nid, N_MEMORY);
7096 * Find the PFN the Movable zone begins in each node. Kernel memory
7097 * is spread evenly between nodes as long as the nodes have enough
7098 * memory. When they don't, some nodes will have more kernelcore than
7101 static void __init find_zone_movable_pfns_for_nodes(void)
7104 unsigned long usable_startpfn;
7105 unsigned long kernelcore_node, kernelcore_remaining;
7106 /* save the state before borrow the nodemask */
7107 nodemask_t saved_node_state = node_states[N_MEMORY];
7108 unsigned long totalpages = early_calculate_totalpages();
7109 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7110 struct memblock_region *r;
7112 /* Need to find movable_zone earlier when movable_node is specified. */
7113 find_usable_zone_for_movable();
7116 * If movable_node is specified, ignore kernelcore and movablecore
7119 if (movable_node_is_enabled()) {
7120 for_each_memblock(memory, r) {
7121 if (!memblock_is_hotpluggable(r))
7124 nid = memblock_get_region_node(r);
7126 usable_startpfn = PFN_DOWN(r->base);
7127 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7128 min(usable_startpfn, zone_movable_pfn[nid]) :
7136 * If kernelcore=mirror is specified, ignore movablecore option
7138 if (mirrored_kernelcore) {
7139 bool mem_below_4gb_not_mirrored = false;
7141 for_each_memblock(memory, r) {
7142 if (memblock_is_mirror(r))
7145 nid = memblock_get_region_node(r);
7147 usable_startpfn = memblock_region_memory_base_pfn(r);
7149 if (usable_startpfn < 0x100000) {
7150 mem_below_4gb_not_mirrored = true;
7154 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7155 min(usable_startpfn, zone_movable_pfn[nid]) :
7159 if (mem_below_4gb_not_mirrored)
7160 pr_warn("This configuration results in unmirrored kernel memory.\n");
7166 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7167 * amount of necessary memory.
7169 if (required_kernelcore_percent)
7170 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7172 if (required_movablecore_percent)
7173 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7177 * If movablecore= was specified, calculate what size of
7178 * kernelcore that corresponds so that memory usable for
7179 * any allocation type is evenly spread. If both kernelcore
7180 * and movablecore are specified, then the value of kernelcore
7181 * will be used for required_kernelcore if it's greater than
7182 * what movablecore would have allowed.
7184 if (required_movablecore) {
7185 unsigned long corepages;
7188 * Round-up so that ZONE_MOVABLE is at least as large as what
7189 * was requested by the user
7191 required_movablecore =
7192 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7193 required_movablecore = min(totalpages, required_movablecore);
7194 corepages = totalpages - required_movablecore;
7196 required_kernelcore = max(required_kernelcore, corepages);
7200 * If kernelcore was not specified or kernelcore size is larger
7201 * than totalpages, there is no ZONE_MOVABLE.
7203 if (!required_kernelcore || required_kernelcore >= totalpages)
7206 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7207 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7210 /* Spread kernelcore memory as evenly as possible throughout nodes */
7211 kernelcore_node = required_kernelcore / usable_nodes;
7212 for_each_node_state(nid, N_MEMORY) {
7213 unsigned long start_pfn, end_pfn;
7216 * Recalculate kernelcore_node if the division per node
7217 * now exceeds what is necessary to satisfy the requested
7218 * amount of memory for the kernel
7220 if (required_kernelcore < kernelcore_node)
7221 kernelcore_node = required_kernelcore / usable_nodes;
7224 * As the map is walked, we track how much memory is usable
7225 * by the kernel using kernelcore_remaining. When it is
7226 * 0, the rest of the node is usable by ZONE_MOVABLE
7228 kernelcore_remaining = kernelcore_node;
7230 /* Go through each range of PFNs within this node */
7231 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7232 unsigned long size_pages;
7234 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7235 if (start_pfn >= end_pfn)
7238 /* Account for what is only usable for kernelcore */
7239 if (start_pfn < usable_startpfn) {
7240 unsigned long kernel_pages;
7241 kernel_pages = min(end_pfn, usable_startpfn)
7244 kernelcore_remaining -= min(kernel_pages,
7245 kernelcore_remaining);
7246 required_kernelcore -= min(kernel_pages,
7247 required_kernelcore);
7249 /* Continue if range is now fully accounted */
7250 if (end_pfn <= usable_startpfn) {
7253 * Push zone_movable_pfn to the end so
7254 * that if we have to rebalance
7255 * kernelcore across nodes, we will
7256 * not double account here
7258 zone_movable_pfn[nid] = end_pfn;
7261 start_pfn = usable_startpfn;
7265 * The usable PFN range for ZONE_MOVABLE is from
7266 * start_pfn->end_pfn. Calculate size_pages as the
7267 * number of pages used as kernelcore
7269 size_pages = end_pfn - start_pfn;
7270 if (size_pages > kernelcore_remaining)
7271 size_pages = kernelcore_remaining;
7272 zone_movable_pfn[nid] = start_pfn + size_pages;
7275 * Some kernelcore has been met, update counts and
7276 * break if the kernelcore for this node has been
7279 required_kernelcore -= min(required_kernelcore,
7281 kernelcore_remaining -= size_pages;
7282 if (!kernelcore_remaining)
7288 * If there is still required_kernelcore, we do another pass with one
7289 * less node in the count. This will push zone_movable_pfn[nid] further
7290 * along on the nodes that still have memory until kernelcore is
7294 if (usable_nodes && required_kernelcore > usable_nodes)
7298 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7299 for (nid = 0; nid < MAX_NUMNODES; nid++)
7300 zone_movable_pfn[nid] =
7301 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7304 /* restore the node_state */
7305 node_states[N_MEMORY] = saved_node_state;
7308 /* Any regular or high memory on that node ? */
7309 static void check_for_memory(pg_data_t *pgdat, int nid)
7311 enum zone_type zone_type;
7313 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7314 struct zone *zone = &pgdat->node_zones[zone_type];
7315 if (populated_zone(zone)) {
7316 if (IS_ENABLED(CONFIG_HIGHMEM))
7317 node_set_state(nid, N_HIGH_MEMORY);
7318 if (zone_type <= ZONE_NORMAL)
7319 node_set_state(nid, N_NORMAL_MEMORY);
7326 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7327 * such cases we allow max_zone_pfn sorted in the descending order
7329 bool __weak arch_has_descending_max_zone_pfns(void)
7335 * free_area_init - Initialise all pg_data_t and zone data
7336 * @max_zone_pfn: an array of max PFNs for each zone
7338 * This will call free_area_init_node() for each active node in the system.
7339 * Using the page ranges provided by memblock_set_node(), the size of each
7340 * zone in each node and their holes is calculated. If the maximum PFN
7341 * between two adjacent zones match, it is assumed that the zone is empty.
7342 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7343 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7344 * starts where the previous one ended. For example, ZONE_DMA32 starts
7345 * at arch_max_dma_pfn.
7347 void __init free_area_init(unsigned long *max_zone_pfn)
7349 unsigned long start_pfn, end_pfn;
7353 /* Record where the zone boundaries are */
7354 memset(arch_zone_lowest_possible_pfn, 0,
7355 sizeof(arch_zone_lowest_possible_pfn));
7356 memset(arch_zone_highest_possible_pfn, 0,
7357 sizeof(arch_zone_highest_possible_pfn));
7359 start_pfn = find_min_pfn_with_active_regions();
7360 descending = arch_has_descending_max_zone_pfns();
7362 for (i = 0; i < MAX_NR_ZONES; i++) {
7364 zone = MAX_NR_ZONES - i - 1;
7368 if (zone == ZONE_MOVABLE)
7371 end_pfn = max(max_zone_pfn[zone], start_pfn);
7372 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7373 arch_zone_highest_possible_pfn[zone] = end_pfn;
7375 start_pfn = end_pfn;
7378 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7379 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7380 find_zone_movable_pfns_for_nodes();
7382 /* Print out the zone ranges */
7383 pr_info("Zone ranges:\n");
7384 for (i = 0; i < MAX_NR_ZONES; i++) {
7385 if (i == ZONE_MOVABLE)
7387 pr_info(" %-8s ", zone_names[i]);
7388 if (arch_zone_lowest_possible_pfn[i] ==
7389 arch_zone_highest_possible_pfn[i])
7392 pr_cont("[mem %#018Lx-%#018Lx]\n",
7393 (u64)arch_zone_lowest_possible_pfn[i]
7395 ((u64)arch_zone_highest_possible_pfn[i]
7396 << PAGE_SHIFT) - 1);
7399 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7400 pr_info("Movable zone start for each node\n");
7401 for (i = 0; i < MAX_NUMNODES; i++) {
7402 if (zone_movable_pfn[i])
7403 pr_info(" Node %d: %#018Lx\n", i,
7404 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7408 * Print out the early node map, and initialize the
7409 * subsection-map relative to active online memory ranges to
7410 * enable future "sub-section" extensions of the memory map.
7412 pr_info("Early memory node ranges\n");
7413 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7414 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7415 (u64)start_pfn << PAGE_SHIFT,
7416 ((u64)end_pfn << PAGE_SHIFT) - 1);
7417 subsection_map_init(start_pfn, end_pfn - start_pfn);
7420 /* Initialise every node */
7421 mminit_verify_pageflags_layout();
7422 setup_nr_node_ids();
7423 init_unavailable_mem();
7424 for_each_online_node(nid) {
7425 pg_data_t *pgdat = NODE_DATA(nid);
7426 free_area_init_node(nid);
7428 /* Any memory on that node */
7429 if (pgdat->node_present_pages)
7430 node_set_state(nid, N_MEMORY);
7431 check_for_memory(pgdat, nid);
7435 static int __init cmdline_parse_core(char *p, unsigned long *core,
7436 unsigned long *percent)
7438 unsigned long long coremem;
7444 /* Value may be a percentage of total memory, otherwise bytes */
7445 coremem = simple_strtoull(p, &endptr, 0);
7446 if (*endptr == '%') {
7447 /* Paranoid check for percent values greater than 100 */
7448 WARN_ON(coremem > 100);
7452 coremem = memparse(p, &p);
7453 /* Paranoid check that UL is enough for the coremem value */
7454 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7456 *core = coremem >> PAGE_SHIFT;
7463 * kernelcore=size sets the amount of memory for use for allocations that
7464 * cannot be reclaimed or migrated.
7466 static int __init cmdline_parse_kernelcore(char *p)
7468 /* parse kernelcore=mirror */
7469 if (parse_option_str(p, "mirror")) {
7470 mirrored_kernelcore = true;
7474 return cmdline_parse_core(p, &required_kernelcore,
7475 &required_kernelcore_percent);
7479 * movablecore=size sets the amount of memory for use for allocations that
7480 * can be reclaimed or migrated.
7482 static int __init cmdline_parse_movablecore(char *p)
7484 return cmdline_parse_core(p, &required_movablecore,
7485 &required_movablecore_percent);
7488 early_param("kernelcore", cmdline_parse_kernelcore);
7489 early_param("movablecore", cmdline_parse_movablecore);
7491 void adjust_managed_page_count(struct page *page, long count)
7493 atomic_long_add(count, &page_zone(page)->managed_pages);
7494 totalram_pages_add(count);
7495 #ifdef CONFIG_HIGHMEM
7496 if (PageHighMem(page))
7497 totalhigh_pages_add(count);
7500 EXPORT_SYMBOL(adjust_managed_page_count);
7502 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7505 unsigned long pages = 0;
7507 start = (void *)PAGE_ALIGN((unsigned long)start);
7508 end = (void *)((unsigned long)end & PAGE_MASK);
7509 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7510 struct page *page = virt_to_page(pos);
7511 void *direct_map_addr;
7514 * 'direct_map_addr' might be different from 'pos'
7515 * because some architectures' virt_to_page()
7516 * work with aliases. Getting the direct map
7517 * address ensures that we get a _writeable_
7518 * alias for the memset().
7520 direct_map_addr = page_address(page);
7521 if ((unsigned int)poison <= 0xFF)
7522 memset(direct_map_addr, poison, PAGE_SIZE);
7524 free_reserved_page(page);
7528 pr_info("Freeing %s memory: %ldK\n",
7529 s, pages << (PAGE_SHIFT - 10));
7534 #ifdef CONFIG_HIGHMEM
7535 void free_highmem_page(struct page *page)
7537 __free_reserved_page(page);
7538 totalram_pages_inc();
7539 atomic_long_inc(&page_zone(page)->managed_pages);
7540 totalhigh_pages_inc();
7545 void __init mem_init_print_info(const char *str)
7547 unsigned long physpages, codesize, datasize, rosize, bss_size;
7548 unsigned long init_code_size, init_data_size;
7550 physpages = get_num_physpages();
7551 codesize = _etext - _stext;
7552 datasize = _edata - _sdata;
7553 rosize = __end_rodata - __start_rodata;
7554 bss_size = __bss_stop - __bss_start;
7555 init_data_size = __init_end - __init_begin;
7556 init_code_size = _einittext - _sinittext;
7559 * Detect special cases and adjust section sizes accordingly:
7560 * 1) .init.* may be embedded into .data sections
7561 * 2) .init.text.* may be out of [__init_begin, __init_end],
7562 * please refer to arch/tile/kernel/vmlinux.lds.S.
7563 * 3) .rodata.* may be embedded into .text or .data sections.
7565 #define adj_init_size(start, end, size, pos, adj) \
7567 if (start <= pos && pos < end && size > adj) \
7571 adj_init_size(__init_begin, __init_end, init_data_size,
7572 _sinittext, init_code_size);
7573 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7574 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7575 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7576 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7578 #undef adj_init_size
7580 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7581 #ifdef CONFIG_HIGHMEM
7585 nr_free_pages() << (PAGE_SHIFT - 10),
7586 physpages << (PAGE_SHIFT - 10),
7587 codesize >> 10, datasize >> 10, rosize >> 10,
7588 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7589 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7590 totalcma_pages << (PAGE_SHIFT - 10),
7591 #ifdef CONFIG_HIGHMEM
7592 totalhigh_pages() << (PAGE_SHIFT - 10),
7594 str ? ", " : "", str ? str : "");
7598 * set_dma_reserve - set the specified number of pages reserved in the first zone
7599 * @new_dma_reserve: The number of pages to mark reserved
7601 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7602 * In the DMA zone, a significant percentage may be consumed by kernel image
7603 * and other unfreeable allocations which can skew the watermarks badly. This
7604 * function may optionally be used to account for unfreeable pages in the
7605 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7606 * smaller per-cpu batchsize.
7608 void __init set_dma_reserve(unsigned long new_dma_reserve)
7610 dma_reserve = new_dma_reserve;
7613 static int page_alloc_cpu_dead(unsigned int cpu)
7616 lru_add_drain_cpu(cpu);
7620 * Spill the event counters of the dead processor
7621 * into the current processors event counters.
7622 * This artificially elevates the count of the current
7625 vm_events_fold_cpu(cpu);
7628 * Zero the differential counters of the dead processor
7629 * so that the vm statistics are consistent.
7631 * This is only okay since the processor is dead and cannot
7632 * race with what we are doing.
7634 cpu_vm_stats_fold(cpu);
7639 int hashdist = HASHDIST_DEFAULT;
7641 static int __init set_hashdist(char *str)
7645 hashdist = simple_strtoul(str, &str, 0);
7648 __setup("hashdist=", set_hashdist);
7651 void __init page_alloc_init(void)
7656 if (num_node_state(N_MEMORY) == 1)
7660 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7661 "mm/page_alloc:dead", NULL,
7662 page_alloc_cpu_dead);
7667 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7668 * or min_free_kbytes changes.
7670 static void calculate_totalreserve_pages(void)
7672 struct pglist_data *pgdat;
7673 unsigned long reserve_pages = 0;
7674 enum zone_type i, j;
7676 for_each_online_pgdat(pgdat) {
7678 pgdat->totalreserve_pages = 0;
7680 for (i = 0; i < MAX_NR_ZONES; i++) {
7681 struct zone *zone = pgdat->node_zones + i;
7683 unsigned long managed_pages = zone_managed_pages(zone);
7685 /* Find valid and maximum lowmem_reserve in the zone */
7686 for (j = i; j < MAX_NR_ZONES; j++) {
7687 if (zone->lowmem_reserve[j] > max)
7688 max = zone->lowmem_reserve[j];
7691 /* we treat the high watermark as reserved pages. */
7692 max += high_wmark_pages(zone);
7694 if (max > managed_pages)
7695 max = managed_pages;
7697 pgdat->totalreserve_pages += max;
7699 reserve_pages += max;
7702 totalreserve_pages = reserve_pages;
7706 * setup_per_zone_lowmem_reserve - called whenever
7707 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7708 * has a correct pages reserved value, so an adequate number of
7709 * pages are left in the zone after a successful __alloc_pages().
7711 static void setup_per_zone_lowmem_reserve(void)
7713 struct pglist_data *pgdat;
7714 enum zone_type j, idx;
7716 for_each_online_pgdat(pgdat) {
7717 for (j = 0; j < MAX_NR_ZONES; j++) {
7718 struct zone *zone = pgdat->node_zones + j;
7719 unsigned long managed_pages = zone_managed_pages(zone);
7721 zone->lowmem_reserve[j] = 0;
7725 struct zone *lower_zone;
7728 lower_zone = pgdat->node_zones + idx;
7730 if (!sysctl_lowmem_reserve_ratio[idx] ||
7731 !zone_managed_pages(lower_zone)) {
7732 lower_zone->lowmem_reserve[j] = 0;
7735 lower_zone->lowmem_reserve[j] =
7736 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7738 managed_pages += zone_managed_pages(lower_zone);
7743 /* update totalreserve_pages */
7744 calculate_totalreserve_pages();
7747 static void __setup_per_zone_wmarks(void)
7749 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7750 unsigned long lowmem_pages = 0;
7752 unsigned long flags;
7754 /* Calculate total number of !ZONE_HIGHMEM pages */
7755 for_each_zone(zone) {
7756 if (!is_highmem(zone))
7757 lowmem_pages += zone_managed_pages(zone);
7760 for_each_zone(zone) {
7763 spin_lock_irqsave(&zone->lock, flags);
7764 tmp = (u64)pages_min * zone_managed_pages(zone);
7765 do_div(tmp, lowmem_pages);
7766 if (is_highmem(zone)) {
7768 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7769 * need highmem pages, so cap pages_min to a small
7772 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7773 * deltas control async page reclaim, and so should
7774 * not be capped for highmem.
7776 unsigned long min_pages;
7778 min_pages = zone_managed_pages(zone) / 1024;
7779 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7780 zone->_watermark[WMARK_MIN] = min_pages;
7783 * If it's a lowmem zone, reserve a number of pages
7784 * proportionate to the zone's size.
7786 zone->_watermark[WMARK_MIN] = tmp;
7790 * Set the kswapd watermarks distance according to the
7791 * scale factor in proportion to available memory, but
7792 * ensure a minimum size on small systems.
7794 tmp = max_t(u64, tmp >> 2,
7795 mult_frac(zone_managed_pages(zone),
7796 watermark_scale_factor, 10000));
7798 zone->watermark_boost = 0;
7799 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7800 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7802 spin_unlock_irqrestore(&zone->lock, flags);
7805 /* update totalreserve_pages */
7806 calculate_totalreserve_pages();
7810 * setup_per_zone_wmarks - called when min_free_kbytes changes
7811 * or when memory is hot-{added|removed}
7813 * Ensures that the watermark[min,low,high] values for each zone are set
7814 * correctly with respect to min_free_kbytes.
7816 void setup_per_zone_wmarks(void)
7818 static DEFINE_SPINLOCK(lock);
7821 __setup_per_zone_wmarks();
7826 * Initialise min_free_kbytes.
7828 * For small machines we want it small (128k min). For large machines
7829 * we want it large (256MB max). But it is not linear, because network
7830 * bandwidth does not increase linearly with machine size. We use
7832 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7833 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7849 int __meminit init_per_zone_wmark_min(void)
7851 unsigned long lowmem_kbytes;
7852 int new_min_free_kbytes;
7854 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7855 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7857 if (new_min_free_kbytes > user_min_free_kbytes) {
7858 min_free_kbytes = new_min_free_kbytes;
7859 if (min_free_kbytes < 128)
7860 min_free_kbytes = 128;
7861 if (min_free_kbytes > 262144)
7862 min_free_kbytes = 262144;
7864 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7865 new_min_free_kbytes, user_min_free_kbytes);
7867 setup_per_zone_wmarks();
7868 refresh_zone_stat_thresholds();
7869 setup_per_zone_lowmem_reserve();
7872 setup_min_unmapped_ratio();
7873 setup_min_slab_ratio();
7878 core_initcall(init_per_zone_wmark_min)
7881 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7882 * that we can call two helper functions whenever min_free_kbytes
7885 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7886 void *buffer, size_t *length, loff_t *ppos)
7890 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7895 user_min_free_kbytes = min_free_kbytes;
7896 setup_per_zone_wmarks();
7901 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7902 void *buffer, size_t *length, loff_t *ppos)
7906 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7911 setup_per_zone_wmarks();
7917 static void setup_min_unmapped_ratio(void)
7922 for_each_online_pgdat(pgdat)
7923 pgdat->min_unmapped_pages = 0;
7926 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7927 sysctl_min_unmapped_ratio) / 100;
7931 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7932 void *buffer, size_t *length, loff_t *ppos)
7936 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7940 setup_min_unmapped_ratio();
7945 static void setup_min_slab_ratio(void)
7950 for_each_online_pgdat(pgdat)
7951 pgdat->min_slab_pages = 0;
7954 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7955 sysctl_min_slab_ratio) / 100;
7958 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7959 void *buffer, size_t *length, loff_t *ppos)
7963 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7967 setup_min_slab_ratio();
7974 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7975 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7976 * whenever sysctl_lowmem_reserve_ratio changes.
7978 * The reserve ratio obviously has absolutely no relation with the
7979 * minimum watermarks. The lowmem reserve ratio can only make sense
7980 * if in function of the boot time zone sizes.
7982 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7983 void *buffer, size_t *length, loff_t *ppos)
7987 proc_dointvec_minmax(table, write, buffer, length, ppos);
7989 for (i = 0; i < MAX_NR_ZONES; i++) {
7990 if (sysctl_lowmem_reserve_ratio[i] < 1)
7991 sysctl_lowmem_reserve_ratio[i] = 0;
7994 setup_per_zone_lowmem_reserve();
7998 static void __zone_pcp_update(struct zone *zone)
8002 for_each_possible_cpu(cpu)
8003 pageset_set_high_and_batch(zone,
8004 per_cpu_ptr(zone->pageset, cpu));
8008 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8009 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8010 * pagelist can have before it gets flushed back to buddy allocator.
8012 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8013 void *buffer, size_t *length, loff_t *ppos)
8016 int old_percpu_pagelist_fraction;
8019 mutex_lock(&pcp_batch_high_lock);
8020 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8022 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8023 if (!write || ret < 0)
8026 /* Sanity checking to avoid pcp imbalance */
8027 if (percpu_pagelist_fraction &&
8028 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8029 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8035 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8038 for_each_populated_zone(zone)
8039 __zone_pcp_update(zone);
8041 mutex_unlock(&pcp_batch_high_lock);
8045 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8047 * Returns the number of pages that arch has reserved but
8048 * is not known to alloc_large_system_hash().
8050 static unsigned long __init arch_reserved_kernel_pages(void)
8057 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8058 * machines. As memory size is increased the scale is also increased but at
8059 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8060 * quadruples the scale is increased by one, which means the size of hash table
8061 * only doubles, instead of quadrupling as well.
8062 * Because 32-bit systems cannot have large physical memory, where this scaling
8063 * makes sense, it is disabled on such platforms.
8065 #if __BITS_PER_LONG > 32
8066 #define ADAPT_SCALE_BASE (64ul << 30)
8067 #define ADAPT_SCALE_SHIFT 2
8068 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8072 * allocate a large system hash table from bootmem
8073 * - it is assumed that the hash table must contain an exact power-of-2
8074 * quantity of entries
8075 * - limit is the number of hash buckets, not the total allocation size
8077 void *__init alloc_large_system_hash(const char *tablename,
8078 unsigned long bucketsize,
8079 unsigned long numentries,
8082 unsigned int *_hash_shift,
8083 unsigned int *_hash_mask,
8084 unsigned long low_limit,
8085 unsigned long high_limit)
8087 unsigned long long max = high_limit;
8088 unsigned long log2qty, size;
8093 /* allow the kernel cmdline to have a say */
8095 /* round applicable memory size up to nearest megabyte */
8096 numentries = nr_kernel_pages;
8097 numentries -= arch_reserved_kernel_pages();
8099 /* It isn't necessary when PAGE_SIZE >= 1MB */
8100 if (PAGE_SHIFT < 20)
8101 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8103 #if __BITS_PER_LONG > 32
8105 unsigned long adapt;
8107 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8108 adapt <<= ADAPT_SCALE_SHIFT)
8113 /* limit to 1 bucket per 2^scale bytes of low memory */
8114 if (scale > PAGE_SHIFT)
8115 numentries >>= (scale - PAGE_SHIFT);
8117 numentries <<= (PAGE_SHIFT - scale);
8119 /* Make sure we've got at least a 0-order allocation.. */
8120 if (unlikely(flags & HASH_SMALL)) {
8121 /* Makes no sense without HASH_EARLY */
8122 WARN_ON(!(flags & HASH_EARLY));
8123 if (!(numentries >> *_hash_shift)) {
8124 numentries = 1UL << *_hash_shift;
8125 BUG_ON(!numentries);
8127 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8128 numentries = PAGE_SIZE / bucketsize;
8130 numentries = roundup_pow_of_two(numentries);
8132 /* limit allocation size to 1/16 total memory by default */
8134 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8135 do_div(max, bucketsize);
8137 max = min(max, 0x80000000ULL);
8139 if (numentries < low_limit)
8140 numentries = low_limit;
8141 if (numentries > max)
8144 log2qty = ilog2(numentries);
8146 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8149 size = bucketsize << log2qty;
8150 if (flags & HASH_EARLY) {
8151 if (flags & HASH_ZERO)
8152 table = memblock_alloc(size, SMP_CACHE_BYTES);
8154 table = memblock_alloc_raw(size,
8156 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8157 table = __vmalloc(size, gfp_flags);
8161 * If bucketsize is not a power-of-two, we may free
8162 * some pages at the end of hash table which
8163 * alloc_pages_exact() automatically does
8165 table = alloc_pages_exact(size, gfp_flags);
8166 kmemleak_alloc(table, size, 1, gfp_flags);
8168 } while (!table && size > PAGE_SIZE && --log2qty);
8171 panic("Failed to allocate %s hash table\n", tablename);
8173 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8174 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8175 virt ? "vmalloc" : "linear");
8178 *_hash_shift = log2qty;
8180 *_hash_mask = (1 << log2qty) - 1;
8186 * This function checks whether pageblock includes unmovable pages or not.
8188 * PageLRU check without isolation or lru_lock could race so that
8189 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8190 * check without lock_page also may miss some movable non-lru pages at
8191 * race condition. So you can't expect this function should be exact.
8193 * Returns a page without holding a reference. If the caller wants to
8194 * dereference that page (e.g., dumping), it has to make sure that that it
8195 * cannot get removed (e.g., via memory unplug) concurrently.
8198 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8199 int migratetype, int flags)
8201 unsigned long iter = 0;
8202 unsigned long pfn = page_to_pfn(page);
8205 * TODO we could make this much more efficient by not checking every
8206 * page in the range if we know all of them are in MOVABLE_ZONE and
8207 * that the movable zone guarantees that pages are migratable but
8208 * the later is not the case right now unfortunatelly. E.g. movablecore
8209 * can still lead to having bootmem allocations in zone_movable.
8212 if (is_migrate_cma_page(page)) {
8214 * CMA allocations (alloc_contig_range) really need to mark
8215 * isolate CMA pageblocks even when they are not movable in fact
8216 * so consider them movable here.
8218 if (is_migrate_cma(migratetype))
8224 for (; iter < pageblock_nr_pages; iter++) {
8225 if (!pfn_valid_within(pfn + iter))
8228 page = pfn_to_page(pfn + iter);
8230 if (PageReserved(page))
8234 * If the zone is movable and we have ruled out all reserved
8235 * pages then it should be reasonably safe to assume the rest
8238 if (zone_idx(zone) == ZONE_MOVABLE)
8242 * Hugepages are not in LRU lists, but they're movable.
8243 * THPs are on the LRU, but need to be counted as #small pages.
8244 * We need not scan over tail pages because we don't
8245 * handle each tail page individually in migration.
8247 if (PageHuge(page) || PageTransCompound(page)) {
8248 struct page *head = compound_head(page);
8249 unsigned int skip_pages;
8251 if (PageHuge(page)) {
8252 if (!hugepage_migration_supported(page_hstate(head)))
8254 } else if (!PageLRU(head) && !__PageMovable(head)) {
8258 skip_pages = compound_nr(head) - (page - head);
8259 iter += skip_pages - 1;
8264 * We can't use page_count without pin a page
8265 * because another CPU can free compound page.
8266 * This check already skips compound tails of THP
8267 * because their page->_refcount is zero at all time.
8269 if (!page_ref_count(page)) {
8270 if (PageBuddy(page))
8271 iter += (1 << page_order(page)) - 1;
8276 * The HWPoisoned page may be not in buddy system, and
8277 * page_count() is not 0.
8279 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8283 * We treat all PageOffline() pages as movable when offlining
8284 * to give drivers a chance to decrement their reference count
8285 * in MEM_GOING_OFFLINE in order to indicate that these pages
8286 * can be offlined as there are no direct references anymore.
8287 * For actually unmovable PageOffline() where the driver does
8288 * not support this, we will fail later when trying to actually
8289 * move these pages that still have a reference count > 0.
8290 * (false negatives in this function only)
8292 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8295 if (__PageMovable(page) || PageLRU(page))
8299 * If there are RECLAIMABLE pages, we need to check
8300 * it. But now, memory offline itself doesn't call
8301 * shrink_node_slabs() and it still to be fixed.
8304 * If the page is not RAM, page_count()should be 0.
8305 * we don't need more check. This is an _used_ not-movable page.
8307 * The problematic thing here is PG_reserved pages. PG_reserved
8308 * is set to both of a memory hole page and a _used_ kernel
8316 #ifdef CONFIG_CONTIG_ALLOC
8317 static unsigned long pfn_max_align_down(unsigned long pfn)
8319 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8320 pageblock_nr_pages) - 1);
8323 static unsigned long pfn_max_align_up(unsigned long pfn)
8325 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8326 pageblock_nr_pages));
8329 /* [start, end) must belong to a single zone. */
8330 static int __alloc_contig_migrate_range(struct compact_control *cc,
8331 unsigned long start, unsigned long end)
8333 /* This function is based on compact_zone() from compaction.c. */
8334 unsigned int nr_reclaimed;
8335 unsigned long pfn = start;
8336 unsigned int tries = 0;
8341 while (pfn < end || !list_empty(&cc->migratepages)) {
8342 if (fatal_signal_pending(current)) {
8347 if (list_empty(&cc->migratepages)) {
8348 cc->nr_migratepages = 0;
8349 pfn = isolate_migratepages_range(cc, pfn, end);
8355 } else if (++tries == 5) {
8356 ret = ret < 0 ? ret : -EBUSY;
8360 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8362 cc->nr_migratepages -= nr_reclaimed;
8364 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8365 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8368 putback_movable_pages(&cc->migratepages);
8375 * alloc_contig_range() -- tries to allocate given range of pages
8376 * @start: start PFN to allocate
8377 * @end: one-past-the-last PFN to allocate
8378 * @migratetype: migratetype of the underlaying pageblocks (either
8379 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8380 * in range must have the same migratetype and it must
8381 * be either of the two.
8382 * @gfp_mask: GFP mask to use during compaction
8384 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8385 * aligned. The PFN range must belong to a single zone.
8387 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8388 * pageblocks in the range. Once isolated, the pageblocks should not
8389 * be modified by others.
8391 * Return: zero on success or negative error code. On success all
8392 * pages which PFN is in [start, end) are allocated for the caller and
8393 * need to be freed with free_contig_range().
8395 int alloc_contig_range(unsigned long start, unsigned long end,
8396 unsigned migratetype, gfp_t gfp_mask)
8398 unsigned long outer_start, outer_end;
8402 struct compact_control cc = {
8403 .nr_migratepages = 0,
8405 .zone = page_zone(pfn_to_page(start)),
8406 .mode = MIGRATE_SYNC,
8407 .ignore_skip_hint = true,
8408 .no_set_skip_hint = true,
8409 .gfp_mask = current_gfp_context(gfp_mask),
8410 .alloc_contig = true,
8412 INIT_LIST_HEAD(&cc.migratepages);
8415 * What we do here is we mark all pageblocks in range as
8416 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8417 * have different sizes, and due to the way page allocator
8418 * work, we align the range to biggest of the two pages so
8419 * that page allocator won't try to merge buddies from
8420 * different pageblocks and change MIGRATE_ISOLATE to some
8421 * other migration type.
8423 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8424 * migrate the pages from an unaligned range (ie. pages that
8425 * we are interested in). This will put all the pages in
8426 * range back to page allocator as MIGRATE_ISOLATE.
8428 * When this is done, we take the pages in range from page
8429 * allocator removing them from the buddy system. This way
8430 * page allocator will never consider using them.
8432 * This lets us mark the pageblocks back as
8433 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8434 * aligned range but not in the unaligned, original range are
8435 * put back to page allocator so that buddy can use them.
8438 ret = start_isolate_page_range(pfn_max_align_down(start),
8439 pfn_max_align_up(end), migratetype, 0);
8444 * In case of -EBUSY, we'd like to know which page causes problem.
8445 * So, just fall through. test_pages_isolated() has a tracepoint
8446 * which will report the busy page.
8448 * It is possible that busy pages could become available before
8449 * the call to test_pages_isolated, and the range will actually be
8450 * allocated. So, if we fall through be sure to clear ret so that
8451 * -EBUSY is not accidentally used or returned to caller.
8453 ret = __alloc_contig_migrate_range(&cc, start, end);
8454 if (ret && ret != -EBUSY)
8459 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8460 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8461 * more, all pages in [start, end) are free in page allocator.
8462 * What we are going to do is to allocate all pages from
8463 * [start, end) (that is remove them from page allocator).
8465 * The only problem is that pages at the beginning and at the
8466 * end of interesting range may be not aligned with pages that
8467 * page allocator holds, ie. they can be part of higher order
8468 * pages. Because of this, we reserve the bigger range and
8469 * once this is done free the pages we are not interested in.
8471 * We don't have to hold zone->lock here because the pages are
8472 * isolated thus they won't get removed from buddy.
8475 lru_add_drain_all();
8478 outer_start = start;
8479 while (!PageBuddy(pfn_to_page(outer_start))) {
8480 if (++order >= MAX_ORDER) {
8481 outer_start = start;
8484 outer_start &= ~0UL << order;
8487 if (outer_start != start) {
8488 order = page_order(pfn_to_page(outer_start));
8491 * outer_start page could be small order buddy page and
8492 * it doesn't include start page. Adjust outer_start
8493 * in this case to report failed page properly
8494 * on tracepoint in test_pages_isolated()
8496 if (outer_start + (1UL << order) <= start)
8497 outer_start = start;
8500 /* Make sure the range is really isolated. */
8501 if (test_pages_isolated(outer_start, end, 0)) {
8502 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8503 __func__, outer_start, end);
8508 /* Grab isolated pages from freelists. */
8509 outer_end = isolate_freepages_range(&cc, outer_start, end);
8515 /* Free head and tail (if any) */
8516 if (start != outer_start)
8517 free_contig_range(outer_start, start - outer_start);
8518 if (end != outer_end)
8519 free_contig_range(end, outer_end - end);
8522 undo_isolate_page_range(pfn_max_align_down(start),
8523 pfn_max_align_up(end), migratetype);
8526 EXPORT_SYMBOL(alloc_contig_range);
8528 static int __alloc_contig_pages(unsigned long start_pfn,
8529 unsigned long nr_pages, gfp_t gfp_mask)
8531 unsigned long end_pfn = start_pfn + nr_pages;
8533 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8537 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8538 unsigned long nr_pages)
8540 unsigned long i, end_pfn = start_pfn + nr_pages;
8543 for (i = start_pfn; i < end_pfn; i++) {
8544 page = pfn_to_online_page(i);
8548 if (page_zone(page) != z)
8551 if (PageReserved(page))
8554 if (page_count(page) > 0)
8563 static bool zone_spans_last_pfn(const struct zone *zone,
8564 unsigned long start_pfn, unsigned long nr_pages)
8566 unsigned long last_pfn = start_pfn + nr_pages - 1;
8568 return zone_spans_pfn(zone, last_pfn);
8572 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8573 * @nr_pages: Number of contiguous pages to allocate
8574 * @gfp_mask: GFP mask to limit search and used during compaction
8576 * @nodemask: Mask for other possible nodes
8578 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8579 * on an applicable zonelist to find a contiguous pfn range which can then be
8580 * tried for allocation with alloc_contig_range(). This routine is intended
8581 * for allocation requests which can not be fulfilled with the buddy allocator.
8583 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8584 * power of two then the alignment is guaranteed to be to the given nr_pages
8585 * (e.g. 1GB request would be aligned to 1GB).
8587 * Allocated pages can be freed with free_contig_range() or by manually calling
8588 * __free_page() on each allocated page.
8590 * Return: pointer to contiguous pages on success, or NULL if not successful.
8592 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8593 int nid, nodemask_t *nodemask)
8595 unsigned long ret, pfn, flags;
8596 struct zonelist *zonelist;
8600 zonelist = node_zonelist(nid, gfp_mask);
8601 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8602 gfp_zone(gfp_mask), nodemask) {
8603 spin_lock_irqsave(&zone->lock, flags);
8605 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8606 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8607 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8609 * We release the zone lock here because
8610 * alloc_contig_range() will also lock the zone
8611 * at some point. If there's an allocation
8612 * spinning on this lock, it may win the race
8613 * and cause alloc_contig_range() to fail...
8615 spin_unlock_irqrestore(&zone->lock, flags);
8616 ret = __alloc_contig_pages(pfn, nr_pages,
8619 return pfn_to_page(pfn);
8620 spin_lock_irqsave(&zone->lock, flags);
8624 spin_unlock_irqrestore(&zone->lock, flags);
8628 #endif /* CONFIG_CONTIG_ALLOC */
8630 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8632 unsigned int count = 0;
8634 for (; nr_pages--; pfn++) {
8635 struct page *page = pfn_to_page(pfn);
8637 count += page_count(page) != 1;
8640 WARN(count != 0, "%d pages are still in use!\n", count);
8642 EXPORT_SYMBOL(free_contig_range);
8645 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8646 * page high values need to be recalulated.
8648 void __meminit zone_pcp_update(struct zone *zone)
8650 mutex_lock(&pcp_batch_high_lock);
8651 __zone_pcp_update(zone);
8652 mutex_unlock(&pcp_batch_high_lock);
8655 void zone_pcp_reset(struct zone *zone)
8657 unsigned long flags;
8659 struct per_cpu_pageset *pset;
8661 /* avoid races with drain_pages() */
8662 local_irq_save(flags);
8663 if (zone->pageset != &boot_pageset) {
8664 for_each_online_cpu(cpu) {
8665 pset = per_cpu_ptr(zone->pageset, cpu);
8666 drain_zonestat(zone, pset);
8668 free_percpu(zone->pageset);
8669 zone->pageset = &boot_pageset;
8671 local_irq_restore(flags);
8674 #ifdef CONFIG_MEMORY_HOTREMOVE
8676 * All pages in the range must be in a single zone and isolated
8677 * before calling this.
8680 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8686 unsigned long flags;
8687 unsigned long offlined_pages = 0;
8689 /* find the first valid pfn */
8690 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8694 return offlined_pages;
8696 offline_mem_sections(pfn, end_pfn);
8697 zone = page_zone(pfn_to_page(pfn));
8698 spin_lock_irqsave(&zone->lock, flags);
8700 while (pfn < end_pfn) {
8701 if (!pfn_valid(pfn)) {
8705 page = pfn_to_page(pfn);
8707 * The HWPoisoned page may be not in buddy system, and
8708 * page_count() is not 0.
8710 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8716 * At this point all remaining PageOffline() pages have a
8717 * reference count of 0 and can simply be skipped.
8719 if (PageOffline(page)) {
8720 BUG_ON(page_count(page));
8721 BUG_ON(PageBuddy(page));
8727 BUG_ON(page_count(page));
8728 BUG_ON(!PageBuddy(page));
8729 order = page_order(page);
8730 offlined_pages += 1 << order;
8731 del_page_from_free_list(page, zone, order);
8732 pfn += (1 << order);
8734 spin_unlock_irqrestore(&zone->lock, flags);
8736 return offlined_pages;
8740 bool is_free_buddy_page(struct page *page)
8742 struct zone *zone = page_zone(page);
8743 unsigned long pfn = page_to_pfn(page);
8744 unsigned long flags;
8747 spin_lock_irqsave(&zone->lock, flags);
8748 for (order = 0; order < MAX_ORDER; order++) {
8749 struct page *page_head = page - (pfn & ((1 << order) - 1));
8751 if (PageBuddy(page_head) && page_order(page_head) >= order)
8754 spin_unlock_irqrestore(&zone->lock, flags);
8756 return order < MAX_ORDER;
8759 #ifdef CONFIG_MEMORY_FAILURE
8761 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8762 * test is performed under the zone lock to prevent a race against page
8765 bool set_hwpoison_free_buddy_page(struct page *page)
8767 struct zone *zone = page_zone(page);
8768 unsigned long pfn = page_to_pfn(page);
8769 unsigned long flags;
8771 bool hwpoisoned = false;
8773 spin_lock_irqsave(&zone->lock, flags);
8774 for (order = 0; order < MAX_ORDER; order++) {
8775 struct page *page_head = page - (pfn & ((1 << order) - 1));
8777 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8778 if (!TestSetPageHWPoison(page))
8783 spin_unlock_irqrestore(&zone->lock, flags);