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 * @mask: mask of bits that the caller is interested in
474 * Return: pageblock_bits flags
476 static __always_inline
477 unsigned long __get_pfnblock_flags_mask(struct page *page,
481 unsigned long *bitmap;
482 unsigned long bitidx, word_bitidx;
485 bitmap = get_pageblock_bitmap(page, pfn);
486 bitidx = pfn_to_bitidx(page, pfn);
487 word_bitidx = bitidx / BITS_PER_LONG;
488 bitidx &= (BITS_PER_LONG-1);
490 word = bitmap[word_bitidx];
491 return (word >> bitidx) & mask;
494 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
497 return __get_pfnblock_flags_mask(page, pfn, mask);
500 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
502 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
506 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
507 * @page: The page within the block of interest
508 * @flags: The flags to set
509 * @pfn: The target page frame number
510 * @mask: mask of bits that the caller is interested in
512 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
516 unsigned long *bitmap;
517 unsigned long bitidx, word_bitidx;
518 unsigned long old_word, word;
520 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
521 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
523 bitmap = get_pageblock_bitmap(page, pfn);
524 bitidx = pfn_to_bitidx(page, pfn);
525 word_bitidx = bitidx / BITS_PER_LONG;
526 bitidx &= (BITS_PER_LONG-1);
528 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
533 word = READ_ONCE(bitmap[word_bitidx]);
535 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
536 if (word == old_word)
542 void set_pageblock_migratetype(struct page *page, int migratetype)
544 if (unlikely(page_group_by_mobility_disabled &&
545 migratetype < MIGRATE_PCPTYPES))
546 migratetype = MIGRATE_UNMOVABLE;
548 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
549 page_to_pfn(page), MIGRATETYPE_MASK);
552 #ifdef CONFIG_DEBUG_VM
553 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
557 unsigned long pfn = page_to_pfn(page);
558 unsigned long sp, start_pfn;
561 seq = zone_span_seqbegin(zone);
562 start_pfn = zone->zone_start_pfn;
563 sp = zone->spanned_pages;
564 if (!zone_spans_pfn(zone, pfn))
566 } while (zone_span_seqretry(zone, seq));
569 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
570 pfn, zone_to_nid(zone), zone->name,
571 start_pfn, start_pfn + sp);
576 static int page_is_consistent(struct zone *zone, struct page *page)
578 if (!pfn_valid_within(page_to_pfn(page)))
580 if (zone != page_zone(page))
586 * Temporary debugging check for pages not lying within a given zone.
588 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
590 if (page_outside_zone_boundaries(zone, page))
592 if (!page_is_consistent(zone, page))
598 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
604 static void bad_page(struct page *page, const char *reason)
606 static unsigned long resume;
607 static unsigned long nr_shown;
608 static unsigned long nr_unshown;
611 * Allow a burst of 60 reports, then keep quiet for that minute;
612 * or allow a steady drip of one report per second.
614 if (nr_shown == 60) {
615 if (time_before(jiffies, resume)) {
621 "BUG: Bad page state: %lu messages suppressed\n",
628 resume = jiffies + 60 * HZ;
630 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
631 current->comm, page_to_pfn(page));
632 __dump_page(page, reason);
633 dump_page_owner(page);
638 /* Leave bad fields for debug, except PageBuddy could make trouble */
639 page_mapcount_reset(page); /* remove PageBuddy */
640 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
644 * Higher-order pages are called "compound pages". They are structured thusly:
646 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
648 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
649 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
651 * The first tail page's ->compound_dtor holds the offset in array of compound
652 * page destructors. See compound_page_dtors.
654 * The first tail page's ->compound_order holds the order of allocation.
655 * This usage means that zero-order pages may not be compound.
658 void free_compound_page(struct page *page)
660 mem_cgroup_uncharge(page);
661 __free_pages_ok(page, compound_order(page));
664 void prep_compound_page(struct page *page, unsigned int order)
667 int nr_pages = 1 << order;
670 for (i = 1; i < nr_pages; i++) {
671 struct page *p = page + i;
672 set_page_count(p, 0);
673 p->mapping = TAIL_MAPPING;
674 set_compound_head(p, page);
677 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
678 set_compound_order(page, order);
679 atomic_set(compound_mapcount_ptr(page), -1);
680 if (hpage_pincount_available(page))
681 atomic_set(compound_pincount_ptr(page), 0);
684 #ifdef CONFIG_DEBUG_PAGEALLOC
685 unsigned int _debug_guardpage_minorder;
687 bool _debug_pagealloc_enabled_early __read_mostly
688 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
689 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
690 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
691 EXPORT_SYMBOL(_debug_pagealloc_enabled);
693 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
695 static int __init early_debug_pagealloc(char *buf)
697 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
699 early_param("debug_pagealloc", early_debug_pagealloc);
701 void init_debug_pagealloc(void)
703 if (!debug_pagealloc_enabled())
706 static_branch_enable(&_debug_pagealloc_enabled);
708 if (!debug_guardpage_minorder())
711 static_branch_enable(&_debug_guardpage_enabled);
714 static int __init debug_guardpage_minorder_setup(char *buf)
718 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
719 pr_err("Bad debug_guardpage_minorder value\n");
722 _debug_guardpage_minorder = res;
723 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
726 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
728 static inline bool set_page_guard(struct zone *zone, struct page *page,
729 unsigned int order, int migratetype)
731 if (!debug_guardpage_enabled())
734 if (order >= debug_guardpage_minorder())
737 __SetPageGuard(page);
738 INIT_LIST_HEAD(&page->lru);
739 set_page_private(page, order);
740 /* Guard pages are not available for any usage */
741 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
746 static inline void clear_page_guard(struct zone *zone, struct page *page,
747 unsigned int order, int migratetype)
749 if (!debug_guardpage_enabled())
752 __ClearPageGuard(page);
754 set_page_private(page, 0);
755 if (!is_migrate_isolate(migratetype))
756 __mod_zone_freepage_state(zone, (1 << order), migratetype);
759 static inline bool set_page_guard(struct zone *zone, struct page *page,
760 unsigned int order, int migratetype) { return false; }
761 static inline void clear_page_guard(struct zone *zone, struct page *page,
762 unsigned int order, int migratetype) {}
765 static inline void set_page_order(struct page *page, unsigned int order)
767 set_page_private(page, order);
768 __SetPageBuddy(page);
772 * This function checks whether a page is free && is the buddy
773 * we can coalesce a page and its buddy if
774 * (a) the buddy is not in a hole (check before calling!) &&
775 * (b) the buddy is in the buddy system &&
776 * (c) a page and its buddy have the same order &&
777 * (d) a page and its buddy are in the same zone.
779 * For recording whether a page is in the buddy system, we set PageBuddy.
780 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
782 * For recording page's order, we use page_private(page).
784 static inline bool page_is_buddy(struct page *page, struct page *buddy,
787 if (!page_is_guard(buddy) && !PageBuddy(buddy))
790 if (page_order(buddy) != order)
794 * zone check is done late to avoid uselessly calculating
795 * zone/node ids for pages that could never merge.
797 if (page_zone_id(page) != page_zone_id(buddy))
800 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
805 #ifdef CONFIG_COMPACTION
806 static inline struct capture_control *task_capc(struct zone *zone)
808 struct capture_control *capc = current->capture_control;
810 return unlikely(capc) &&
811 !(current->flags & PF_KTHREAD) &&
813 capc->cc->zone == zone ? capc : NULL;
817 compaction_capture(struct capture_control *capc, struct page *page,
818 int order, int migratetype)
820 if (!capc || order != capc->cc->order)
823 /* Do not accidentally pollute CMA or isolated regions*/
824 if (is_migrate_cma(migratetype) ||
825 is_migrate_isolate(migratetype))
829 * Do not let lower order allocations polluate a movable pageblock.
830 * This might let an unmovable request use a reclaimable pageblock
831 * and vice-versa but no more than normal fallback logic which can
832 * have trouble finding a high-order free page.
834 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
842 static inline struct capture_control *task_capc(struct zone *zone)
848 compaction_capture(struct capture_control *capc, struct page *page,
849 int order, int migratetype)
853 #endif /* CONFIG_COMPACTION */
855 /* Used for pages not on another list */
856 static inline void add_to_free_list(struct page *page, struct zone *zone,
857 unsigned int order, int migratetype)
859 struct free_area *area = &zone->free_area[order];
861 list_add(&page->lru, &area->free_list[migratetype]);
865 /* Used for pages not on another list */
866 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
867 unsigned int order, int migratetype)
869 struct free_area *area = &zone->free_area[order];
871 list_add_tail(&page->lru, &area->free_list[migratetype]);
875 /* Used for pages which are on another list */
876 static inline void move_to_free_list(struct page *page, struct zone *zone,
877 unsigned int order, int migratetype)
879 struct free_area *area = &zone->free_area[order];
881 list_move(&page->lru, &area->free_list[migratetype]);
884 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
887 /* clear reported state and update reported page count */
888 if (page_reported(page))
889 __ClearPageReported(page);
891 list_del(&page->lru);
892 __ClearPageBuddy(page);
893 set_page_private(page, 0);
894 zone->free_area[order].nr_free--;
898 * If this is not the largest possible page, check if the buddy
899 * of the next-highest order is free. If it is, it's possible
900 * that pages are being freed that will coalesce soon. In case,
901 * that is happening, add the free page to the tail of the list
902 * so it's less likely to be used soon and more likely to be merged
903 * as a higher order page
906 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
907 struct page *page, unsigned int order)
909 struct page *higher_page, *higher_buddy;
910 unsigned long combined_pfn;
912 if (order >= MAX_ORDER - 2)
915 if (!pfn_valid_within(buddy_pfn))
918 combined_pfn = buddy_pfn & pfn;
919 higher_page = page + (combined_pfn - pfn);
920 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
921 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
923 return pfn_valid_within(buddy_pfn) &&
924 page_is_buddy(higher_page, higher_buddy, order + 1);
928 * Freeing function for a buddy system allocator.
930 * The concept of a buddy system is to maintain direct-mapped table
931 * (containing bit values) for memory blocks of various "orders".
932 * The bottom level table contains the map for the smallest allocatable
933 * units of memory (here, pages), and each level above it describes
934 * pairs of units from the levels below, hence, "buddies".
935 * At a high level, all that happens here is marking the table entry
936 * at the bottom level available, and propagating the changes upward
937 * as necessary, plus some accounting needed to play nicely with other
938 * parts of the VM system.
939 * At each level, we keep a list of pages, which are heads of continuous
940 * free pages of length of (1 << order) and marked with PageBuddy.
941 * Page's order is recorded in page_private(page) field.
942 * So when we are allocating or freeing one, we can derive the state of the
943 * other. That is, if we allocate a small block, and both were
944 * free, the remainder of the region must be split into blocks.
945 * If a block is freed, and its buddy is also free, then this
946 * triggers coalescing into a block of larger size.
951 static inline void __free_one_page(struct page *page,
953 struct zone *zone, unsigned int order,
954 int migratetype, bool report)
956 struct capture_control *capc = task_capc(zone);
957 unsigned long buddy_pfn;
958 unsigned long combined_pfn;
959 unsigned int max_order;
963 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
965 VM_BUG_ON(!zone_is_initialized(zone));
966 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
968 VM_BUG_ON(migratetype == -1);
969 if (likely(!is_migrate_isolate(migratetype)))
970 __mod_zone_freepage_state(zone, 1 << order, migratetype);
972 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
973 VM_BUG_ON_PAGE(bad_range(zone, page), page);
976 while (order < max_order - 1) {
977 if (compaction_capture(capc, page, order, migratetype)) {
978 __mod_zone_freepage_state(zone, -(1 << order),
982 buddy_pfn = __find_buddy_pfn(pfn, order);
983 buddy = page + (buddy_pfn - pfn);
985 if (!pfn_valid_within(buddy_pfn))
987 if (!page_is_buddy(page, buddy, order))
990 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
991 * merge with it and move up one order.
993 if (page_is_guard(buddy))
994 clear_page_guard(zone, buddy, order, migratetype);
996 del_page_from_free_list(buddy, zone, order);
997 combined_pfn = buddy_pfn & pfn;
998 page = page + (combined_pfn - pfn);
1002 if (max_order < MAX_ORDER) {
1003 /* If we are here, it means order is >= pageblock_order.
1004 * We want to prevent merge between freepages on isolate
1005 * pageblock and normal pageblock. Without this, pageblock
1006 * isolation could cause incorrect freepage or CMA accounting.
1008 * We don't want to hit this code for the more frequent
1009 * low-order merging.
1011 if (unlikely(has_isolate_pageblock(zone))) {
1014 buddy_pfn = __find_buddy_pfn(pfn, order);
1015 buddy = page + (buddy_pfn - pfn);
1016 buddy_mt = get_pageblock_migratetype(buddy);
1018 if (migratetype != buddy_mt
1019 && (is_migrate_isolate(migratetype) ||
1020 is_migrate_isolate(buddy_mt)))
1024 goto continue_merging;
1028 set_page_order(page, order);
1030 if (is_shuffle_order(order))
1031 to_tail = shuffle_pick_tail();
1033 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1036 add_to_free_list_tail(page, zone, order, migratetype);
1038 add_to_free_list(page, zone, order, migratetype);
1040 /* Notify page reporting subsystem of freed page */
1042 page_reporting_notify_free(order);
1046 * A bad page could be due to a number of fields. Instead of multiple branches,
1047 * try and check multiple fields with one check. The caller must do a detailed
1048 * check if necessary.
1050 static inline bool page_expected_state(struct page *page,
1051 unsigned long check_flags)
1053 if (unlikely(atomic_read(&page->_mapcount) != -1))
1056 if (unlikely((unsigned long)page->mapping |
1057 page_ref_count(page) |
1059 (unsigned long)page->mem_cgroup |
1061 (page->flags & check_flags)))
1067 static const char *page_bad_reason(struct page *page, unsigned long flags)
1069 const char *bad_reason = NULL;
1071 if (unlikely(atomic_read(&page->_mapcount) != -1))
1072 bad_reason = "nonzero mapcount";
1073 if (unlikely(page->mapping != NULL))
1074 bad_reason = "non-NULL mapping";
1075 if (unlikely(page_ref_count(page) != 0))
1076 bad_reason = "nonzero _refcount";
1077 if (unlikely(page->flags & flags)) {
1078 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1079 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1081 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1084 if (unlikely(page->mem_cgroup))
1085 bad_reason = "page still charged to cgroup";
1090 static void check_free_page_bad(struct page *page)
1093 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1096 static inline int check_free_page(struct page *page)
1098 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1101 /* Something has gone sideways, find it */
1102 check_free_page_bad(page);
1106 static int free_tail_pages_check(struct page *head_page, struct page *page)
1111 * We rely page->lru.next never has bit 0 set, unless the page
1112 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1114 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1116 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1120 switch (page - head_page) {
1122 /* the first tail page: ->mapping may be compound_mapcount() */
1123 if (unlikely(compound_mapcount(page))) {
1124 bad_page(page, "nonzero compound_mapcount");
1130 * the second tail page: ->mapping is
1131 * deferred_list.next -- ignore value.
1135 if (page->mapping != TAIL_MAPPING) {
1136 bad_page(page, "corrupted mapping in tail page");
1141 if (unlikely(!PageTail(page))) {
1142 bad_page(page, "PageTail not set");
1145 if (unlikely(compound_head(page) != head_page)) {
1146 bad_page(page, "compound_head not consistent");
1151 page->mapping = NULL;
1152 clear_compound_head(page);
1156 static void kernel_init_free_pages(struct page *page, int numpages)
1160 /* s390's use of memset() could override KASAN redzones. */
1161 kasan_disable_current();
1162 for (i = 0; i < numpages; i++)
1163 clear_highpage(page + i);
1164 kasan_enable_current();
1167 static __always_inline bool free_pages_prepare(struct page *page,
1168 unsigned int order, bool check_free)
1172 VM_BUG_ON_PAGE(PageTail(page), page);
1174 trace_mm_page_free(page, order);
1177 * Check tail pages before head page information is cleared to
1178 * avoid checking PageCompound for order-0 pages.
1180 if (unlikely(order)) {
1181 bool compound = PageCompound(page);
1184 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1187 ClearPageDoubleMap(page);
1188 for (i = 1; i < (1 << order); i++) {
1190 bad += free_tail_pages_check(page, page + i);
1191 if (unlikely(check_free_page(page + i))) {
1195 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1198 if (PageMappingFlags(page))
1199 page->mapping = NULL;
1200 if (memcg_kmem_enabled() && PageKmemcg(page))
1201 __memcg_kmem_uncharge_page(page, order);
1203 bad += check_free_page(page);
1207 page_cpupid_reset_last(page);
1208 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1209 reset_page_owner(page, order);
1211 if (!PageHighMem(page)) {
1212 debug_check_no_locks_freed(page_address(page),
1213 PAGE_SIZE << order);
1214 debug_check_no_obj_freed(page_address(page),
1215 PAGE_SIZE << order);
1217 if (want_init_on_free())
1218 kernel_init_free_pages(page, 1 << order);
1220 kernel_poison_pages(page, 1 << order, 0);
1222 * arch_free_page() can make the page's contents inaccessible. s390
1223 * does this. So nothing which can access the page's contents should
1224 * happen after this.
1226 arch_free_page(page, order);
1228 if (debug_pagealloc_enabled_static())
1229 kernel_map_pages(page, 1 << order, 0);
1231 kasan_free_nondeferred_pages(page, order);
1236 #ifdef CONFIG_DEBUG_VM
1238 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1239 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1240 * moved from pcp lists to free lists.
1242 static bool free_pcp_prepare(struct page *page)
1244 return free_pages_prepare(page, 0, true);
1247 static bool bulkfree_pcp_prepare(struct page *page)
1249 if (debug_pagealloc_enabled_static())
1250 return check_free_page(page);
1256 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1257 * moving from pcp lists to free list in order to reduce overhead. With
1258 * debug_pagealloc enabled, they are checked also immediately when being freed
1261 static bool free_pcp_prepare(struct page *page)
1263 if (debug_pagealloc_enabled_static())
1264 return free_pages_prepare(page, 0, true);
1266 return free_pages_prepare(page, 0, false);
1269 static bool bulkfree_pcp_prepare(struct page *page)
1271 return check_free_page(page);
1273 #endif /* CONFIG_DEBUG_VM */
1275 static inline void prefetch_buddy(struct page *page)
1277 unsigned long pfn = page_to_pfn(page);
1278 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1279 struct page *buddy = page + (buddy_pfn - pfn);
1285 * Frees a number of pages from the PCP lists
1286 * Assumes all pages on list are in same zone, and of same order.
1287 * count is the number of pages to free.
1289 * If the zone was previously in an "all pages pinned" state then look to
1290 * see if this freeing clears that state.
1292 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1293 * pinned" detection logic.
1295 static void free_pcppages_bulk(struct zone *zone, int count,
1296 struct per_cpu_pages *pcp)
1298 int migratetype = 0;
1300 int prefetch_nr = 0;
1301 bool isolated_pageblocks;
1302 struct page *page, *tmp;
1306 struct list_head *list;
1309 * Remove pages from lists in a round-robin fashion. A
1310 * batch_free count is maintained that is incremented when an
1311 * empty list is encountered. This is so more pages are freed
1312 * off fuller lists instead of spinning excessively around empty
1317 if (++migratetype == MIGRATE_PCPTYPES)
1319 list = &pcp->lists[migratetype];
1320 } while (list_empty(list));
1322 /* This is the only non-empty list. Free them all. */
1323 if (batch_free == MIGRATE_PCPTYPES)
1327 page = list_last_entry(list, struct page, lru);
1328 /* must delete to avoid corrupting pcp list */
1329 list_del(&page->lru);
1332 if (bulkfree_pcp_prepare(page))
1335 list_add_tail(&page->lru, &head);
1338 * We are going to put the page back to the global
1339 * pool, prefetch its buddy to speed up later access
1340 * under zone->lock. It is believed the overhead of
1341 * an additional test and calculating buddy_pfn here
1342 * can be offset by reduced memory latency later. To
1343 * avoid excessive prefetching due to large count, only
1344 * prefetch buddy for the first pcp->batch nr of pages.
1346 if (prefetch_nr++ < pcp->batch)
1347 prefetch_buddy(page);
1348 } while (--count && --batch_free && !list_empty(list));
1351 spin_lock(&zone->lock);
1352 isolated_pageblocks = has_isolate_pageblock(zone);
1355 * Use safe version since after __free_one_page(),
1356 * page->lru.next will not point to original list.
1358 list_for_each_entry_safe(page, tmp, &head, lru) {
1359 int mt = get_pcppage_migratetype(page);
1360 /* MIGRATE_ISOLATE page should not go to pcplists */
1361 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1362 /* Pageblock could have been isolated meanwhile */
1363 if (unlikely(isolated_pageblocks))
1364 mt = get_pageblock_migratetype(page);
1366 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1367 trace_mm_page_pcpu_drain(page, 0, mt);
1369 spin_unlock(&zone->lock);
1372 static void free_one_page(struct zone *zone,
1373 struct page *page, unsigned long pfn,
1377 spin_lock(&zone->lock);
1378 if (unlikely(has_isolate_pageblock(zone) ||
1379 is_migrate_isolate(migratetype))) {
1380 migratetype = get_pfnblock_migratetype(page, pfn);
1382 __free_one_page(page, pfn, zone, order, migratetype, true);
1383 spin_unlock(&zone->lock);
1386 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1387 unsigned long zone, int nid)
1389 mm_zero_struct_page(page);
1390 set_page_links(page, zone, nid, pfn);
1391 init_page_count(page);
1392 page_mapcount_reset(page);
1393 page_cpupid_reset_last(page);
1394 page_kasan_tag_reset(page);
1396 INIT_LIST_HEAD(&page->lru);
1397 #ifdef WANT_PAGE_VIRTUAL
1398 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1399 if (!is_highmem_idx(zone))
1400 set_page_address(page, __va(pfn << PAGE_SHIFT));
1404 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1405 static void __meminit init_reserved_page(unsigned long pfn)
1410 if (!early_page_uninitialised(pfn))
1413 nid = early_pfn_to_nid(pfn);
1414 pgdat = NODE_DATA(nid);
1416 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1417 struct zone *zone = &pgdat->node_zones[zid];
1419 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1422 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1425 static inline void init_reserved_page(unsigned long pfn)
1428 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1431 * Initialised pages do not have PageReserved set. This function is
1432 * called for each range allocated by the bootmem allocator and
1433 * marks the pages PageReserved. The remaining valid pages are later
1434 * sent to the buddy page allocator.
1436 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1438 unsigned long start_pfn = PFN_DOWN(start);
1439 unsigned long end_pfn = PFN_UP(end);
1441 for (; start_pfn < end_pfn; start_pfn++) {
1442 if (pfn_valid(start_pfn)) {
1443 struct page *page = pfn_to_page(start_pfn);
1445 init_reserved_page(start_pfn);
1447 /* Avoid false-positive PageTail() */
1448 INIT_LIST_HEAD(&page->lru);
1451 * no need for atomic set_bit because the struct
1452 * page is not visible yet so nobody should
1455 __SetPageReserved(page);
1460 static void __free_pages_ok(struct page *page, unsigned int order)
1462 unsigned long flags;
1464 unsigned long pfn = page_to_pfn(page);
1466 if (!free_pages_prepare(page, order, true))
1469 migratetype = get_pfnblock_migratetype(page, pfn);
1470 local_irq_save(flags);
1471 __count_vm_events(PGFREE, 1 << order);
1472 free_one_page(page_zone(page), page, pfn, order, migratetype);
1473 local_irq_restore(flags);
1476 void __free_pages_core(struct page *page, unsigned int order)
1478 unsigned int nr_pages = 1 << order;
1479 struct page *p = page;
1483 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1485 __ClearPageReserved(p);
1486 set_page_count(p, 0);
1488 __ClearPageReserved(p);
1489 set_page_count(p, 0);
1491 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1492 set_page_refcounted(page);
1493 __free_pages(page, order);
1496 #ifdef CONFIG_NEED_MULTIPLE_NODES
1498 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1500 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1503 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1505 int __meminit __early_pfn_to_nid(unsigned long pfn,
1506 struct mminit_pfnnid_cache *state)
1508 unsigned long start_pfn, end_pfn;
1511 if (state->last_start <= pfn && pfn < state->last_end)
1512 return state->last_nid;
1514 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1515 if (nid != NUMA_NO_NODE) {
1516 state->last_start = start_pfn;
1517 state->last_end = end_pfn;
1518 state->last_nid = nid;
1523 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1525 int __meminit early_pfn_to_nid(unsigned long pfn)
1527 static DEFINE_SPINLOCK(early_pfn_lock);
1530 spin_lock(&early_pfn_lock);
1531 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1533 nid = first_online_node;
1534 spin_unlock(&early_pfn_lock);
1538 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1540 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1543 if (early_page_uninitialised(pfn))
1545 __free_pages_core(page, order);
1549 * Check that the whole (or subset of) a pageblock given by the interval of
1550 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1551 * with the migration of free compaction scanner. The scanners then need to
1552 * use only pfn_valid_within() check for arches that allow holes within
1555 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1557 * It's possible on some configurations to have a setup like node0 node1 node0
1558 * i.e. it's possible that all pages within a zones range of pages do not
1559 * belong to a single zone. We assume that a border between node0 and node1
1560 * can occur within a single pageblock, but not a node0 node1 node0
1561 * interleaving within a single pageblock. It is therefore sufficient to check
1562 * the first and last page of a pageblock and avoid checking each individual
1563 * page in a pageblock.
1565 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1566 unsigned long end_pfn, struct zone *zone)
1568 struct page *start_page;
1569 struct page *end_page;
1571 /* end_pfn is one past the range we are checking */
1574 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1577 start_page = pfn_to_online_page(start_pfn);
1581 if (page_zone(start_page) != zone)
1584 end_page = pfn_to_page(end_pfn);
1586 /* This gives a shorter code than deriving page_zone(end_page) */
1587 if (page_zone_id(start_page) != page_zone_id(end_page))
1593 void set_zone_contiguous(struct zone *zone)
1595 unsigned long block_start_pfn = zone->zone_start_pfn;
1596 unsigned long block_end_pfn;
1598 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1599 for (; block_start_pfn < zone_end_pfn(zone);
1600 block_start_pfn = block_end_pfn,
1601 block_end_pfn += pageblock_nr_pages) {
1603 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1605 if (!__pageblock_pfn_to_page(block_start_pfn,
1606 block_end_pfn, zone))
1611 /* We confirm that there is no hole */
1612 zone->contiguous = true;
1615 void clear_zone_contiguous(struct zone *zone)
1617 zone->contiguous = false;
1620 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1621 static void __init deferred_free_range(unsigned long pfn,
1622 unsigned long nr_pages)
1630 page = pfn_to_page(pfn);
1632 /* Free a large naturally-aligned chunk if possible */
1633 if (nr_pages == pageblock_nr_pages &&
1634 (pfn & (pageblock_nr_pages - 1)) == 0) {
1635 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1636 __free_pages_core(page, pageblock_order);
1640 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1641 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1642 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1643 __free_pages_core(page, 0);
1647 /* Completion tracking for deferred_init_memmap() threads */
1648 static atomic_t pgdat_init_n_undone __initdata;
1649 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1651 static inline void __init pgdat_init_report_one_done(void)
1653 if (atomic_dec_and_test(&pgdat_init_n_undone))
1654 complete(&pgdat_init_all_done_comp);
1658 * Returns true if page needs to be initialized or freed to buddy allocator.
1660 * First we check if pfn is valid on architectures where it is possible to have
1661 * holes within pageblock_nr_pages. On systems where it is not possible, this
1662 * function is optimized out.
1664 * Then, we check if a current large page is valid by only checking the validity
1667 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1669 if (!pfn_valid_within(pfn))
1671 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1677 * Free pages to buddy allocator. Try to free aligned pages in
1678 * pageblock_nr_pages sizes.
1680 static void __init deferred_free_pages(unsigned long pfn,
1681 unsigned long end_pfn)
1683 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1684 unsigned long nr_free = 0;
1686 for (; pfn < end_pfn; pfn++) {
1687 if (!deferred_pfn_valid(pfn)) {
1688 deferred_free_range(pfn - nr_free, nr_free);
1690 } else if (!(pfn & nr_pgmask)) {
1691 deferred_free_range(pfn - nr_free, nr_free);
1697 /* Free the last block of pages to allocator */
1698 deferred_free_range(pfn - nr_free, nr_free);
1702 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1703 * by performing it only once every pageblock_nr_pages.
1704 * Return number of pages initialized.
1706 static unsigned long __init deferred_init_pages(struct zone *zone,
1708 unsigned long end_pfn)
1710 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1711 int nid = zone_to_nid(zone);
1712 unsigned long nr_pages = 0;
1713 int zid = zone_idx(zone);
1714 struct page *page = NULL;
1716 for (; pfn < end_pfn; pfn++) {
1717 if (!deferred_pfn_valid(pfn)) {
1720 } else if (!page || !(pfn & nr_pgmask)) {
1721 page = pfn_to_page(pfn);
1725 __init_single_page(page, pfn, zid, nid);
1732 * This function is meant to pre-load the iterator for the zone init.
1733 * Specifically it walks through the ranges until we are caught up to the
1734 * first_init_pfn value and exits there. If we never encounter the value we
1735 * return false indicating there are no valid ranges left.
1738 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1739 unsigned long *spfn, unsigned long *epfn,
1740 unsigned long first_init_pfn)
1745 * Start out by walking through the ranges in this zone that have
1746 * already been initialized. We don't need to do anything with them
1747 * so we just need to flush them out of the system.
1749 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1750 if (*epfn <= first_init_pfn)
1752 if (*spfn < first_init_pfn)
1753 *spfn = first_init_pfn;
1762 * Initialize and free pages. We do it in two loops: first we initialize
1763 * struct page, then free to buddy allocator, because while we are
1764 * freeing pages we can access pages that are ahead (computing buddy
1765 * page in __free_one_page()).
1767 * In order to try and keep some memory in the cache we have the loop
1768 * broken along max page order boundaries. This way we will not cause
1769 * any issues with the buddy page computation.
1771 static unsigned long __init
1772 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1773 unsigned long *end_pfn)
1775 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1776 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1777 unsigned long nr_pages = 0;
1780 /* First we loop through and initialize the page values */
1781 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1784 if (mo_pfn <= *start_pfn)
1787 t = min(mo_pfn, *end_pfn);
1788 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1790 if (mo_pfn < *end_pfn) {
1791 *start_pfn = mo_pfn;
1796 /* Reset values and now loop through freeing pages as needed */
1799 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1805 t = min(mo_pfn, epfn);
1806 deferred_free_pages(spfn, t);
1816 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1819 unsigned long spfn, epfn;
1820 struct zone *zone = arg;
1823 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1826 * Initialize and free pages in MAX_ORDER sized increments so that we
1827 * can avoid introducing any issues with the buddy allocator.
1829 while (spfn < end_pfn) {
1830 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1835 /* An arch may override for more concurrency. */
1837 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1842 /* Initialise remaining memory on a node */
1843 static int __init deferred_init_memmap(void *data)
1845 pg_data_t *pgdat = data;
1846 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1847 unsigned long spfn = 0, epfn = 0;
1848 unsigned long first_init_pfn, flags;
1849 unsigned long start = jiffies;
1851 int zid, max_threads;
1854 /* Bind memory initialisation thread to a local node if possible */
1855 if (!cpumask_empty(cpumask))
1856 set_cpus_allowed_ptr(current, cpumask);
1858 pgdat_resize_lock(pgdat, &flags);
1859 first_init_pfn = pgdat->first_deferred_pfn;
1860 if (first_init_pfn == ULONG_MAX) {
1861 pgdat_resize_unlock(pgdat, &flags);
1862 pgdat_init_report_one_done();
1866 /* Sanity check boundaries */
1867 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1868 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1869 pgdat->first_deferred_pfn = ULONG_MAX;
1872 * Once we unlock here, the zone cannot be grown anymore, thus if an
1873 * interrupt thread must allocate this early in boot, zone must be
1874 * pre-grown prior to start of deferred page initialization.
1876 pgdat_resize_unlock(pgdat, &flags);
1878 /* Only the highest zone is deferred so find it */
1879 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1880 zone = pgdat->node_zones + zid;
1881 if (first_init_pfn < zone_end_pfn(zone))
1885 /* If the zone is empty somebody else may have cleared out the zone */
1886 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1890 max_threads = deferred_page_init_max_threads(cpumask);
1892 while (spfn < epfn) {
1893 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1894 struct padata_mt_job job = {
1895 .thread_fn = deferred_init_memmap_chunk,
1898 .size = epfn_align - spfn,
1899 .align = PAGES_PER_SECTION,
1900 .min_chunk = PAGES_PER_SECTION,
1901 .max_threads = max_threads,
1904 padata_do_multithreaded(&job);
1905 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1909 /* Sanity check that the next zone really is unpopulated */
1910 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1912 pr_info("node %d deferred pages initialised in %ums\n",
1913 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1915 pgdat_init_report_one_done();
1920 * If this zone has deferred pages, try to grow it by initializing enough
1921 * deferred pages to satisfy the allocation specified by order, rounded up to
1922 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1923 * of SECTION_SIZE bytes by initializing struct pages in increments of
1924 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1926 * Return true when zone was grown, otherwise return false. We return true even
1927 * when we grow less than requested, to let the caller decide if there are
1928 * enough pages to satisfy the allocation.
1930 * Note: We use noinline because this function is needed only during boot, and
1931 * it is called from a __ref function _deferred_grow_zone. This way we are
1932 * making sure that it is not inlined into permanent text section.
1934 static noinline bool __init
1935 deferred_grow_zone(struct zone *zone, unsigned int order)
1937 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1938 pg_data_t *pgdat = zone->zone_pgdat;
1939 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1940 unsigned long spfn, epfn, flags;
1941 unsigned long nr_pages = 0;
1944 /* Only the last zone may have deferred pages */
1945 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1948 pgdat_resize_lock(pgdat, &flags);
1951 * If someone grew this zone while we were waiting for spinlock, return
1952 * true, as there might be enough pages already.
1954 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1955 pgdat_resize_unlock(pgdat, &flags);
1959 /* If the zone is empty somebody else may have cleared out the zone */
1960 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1961 first_deferred_pfn)) {
1962 pgdat->first_deferred_pfn = ULONG_MAX;
1963 pgdat_resize_unlock(pgdat, &flags);
1964 /* Retry only once. */
1965 return first_deferred_pfn != ULONG_MAX;
1969 * Initialize and free pages in MAX_ORDER sized increments so
1970 * that we can avoid introducing any issues with the buddy
1973 while (spfn < epfn) {
1974 /* update our first deferred PFN for this section */
1975 first_deferred_pfn = spfn;
1977 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1978 touch_nmi_watchdog();
1980 /* We should only stop along section boundaries */
1981 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1984 /* If our quota has been met we can stop here */
1985 if (nr_pages >= nr_pages_needed)
1989 pgdat->first_deferred_pfn = spfn;
1990 pgdat_resize_unlock(pgdat, &flags);
1992 return nr_pages > 0;
1996 * deferred_grow_zone() is __init, but it is called from
1997 * get_page_from_freelist() during early boot until deferred_pages permanently
1998 * disables this call. This is why we have refdata wrapper to avoid warning,
1999 * and to ensure that the function body gets unloaded.
2002 _deferred_grow_zone(struct zone *zone, unsigned int order)
2004 return deferred_grow_zone(zone, order);
2007 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2009 void __init page_alloc_init_late(void)
2014 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2016 /* There will be num_node_state(N_MEMORY) threads */
2017 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2018 for_each_node_state(nid, N_MEMORY) {
2019 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2022 /* Block until all are initialised */
2023 wait_for_completion(&pgdat_init_all_done_comp);
2026 * The number of managed pages has changed due to the initialisation
2027 * so the pcpu batch and high limits needs to be updated or the limits
2028 * will be artificially small.
2030 for_each_populated_zone(zone)
2031 zone_pcp_update(zone);
2034 * We initialized the rest of the deferred pages. Permanently disable
2035 * on-demand struct page initialization.
2037 static_branch_disable(&deferred_pages);
2039 /* Reinit limits that are based on free pages after the kernel is up */
2040 files_maxfiles_init();
2043 /* Discard memblock private memory */
2046 for_each_node_state(nid, N_MEMORY)
2047 shuffle_free_memory(NODE_DATA(nid));
2049 for_each_populated_zone(zone)
2050 set_zone_contiguous(zone);
2054 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2055 void __init init_cma_reserved_pageblock(struct page *page)
2057 unsigned i = pageblock_nr_pages;
2058 struct page *p = page;
2061 __ClearPageReserved(p);
2062 set_page_count(p, 0);
2065 set_pageblock_migratetype(page, MIGRATE_CMA);
2067 if (pageblock_order >= MAX_ORDER) {
2068 i = pageblock_nr_pages;
2071 set_page_refcounted(p);
2072 __free_pages(p, MAX_ORDER - 1);
2073 p += MAX_ORDER_NR_PAGES;
2074 } while (i -= MAX_ORDER_NR_PAGES);
2076 set_page_refcounted(page);
2077 __free_pages(page, pageblock_order);
2080 adjust_managed_page_count(page, pageblock_nr_pages);
2085 * The order of subdivision here is critical for the IO subsystem.
2086 * Please do not alter this order without good reasons and regression
2087 * testing. Specifically, as large blocks of memory are subdivided,
2088 * the order in which smaller blocks are delivered depends on the order
2089 * they're subdivided in this function. This is the primary factor
2090 * influencing the order in which pages are delivered to the IO
2091 * subsystem according to empirical testing, and this is also justified
2092 * by considering the behavior of a buddy system containing a single
2093 * large block of memory acted on by a series of small allocations.
2094 * This behavior is a critical factor in sglist merging's success.
2098 static inline void expand(struct zone *zone, struct page *page,
2099 int low, int high, int migratetype)
2101 unsigned long size = 1 << high;
2103 while (high > low) {
2106 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2109 * Mark as guard pages (or page), that will allow to
2110 * merge back to allocator when buddy will be freed.
2111 * Corresponding page table entries will not be touched,
2112 * pages will stay not present in virtual address space
2114 if (set_page_guard(zone, &page[size], high, migratetype))
2117 add_to_free_list(&page[size], zone, high, migratetype);
2118 set_page_order(&page[size], high);
2122 static void check_new_page_bad(struct page *page)
2124 if (unlikely(page->flags & __PG_HWPOISON)) {
2125 /* Don't complain about hwpoisoned pages */
2126 page_mapcount_reset(page); /* remove PageBuddy */
2131 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2135 * This page is about to be returned from the page allocator
2137 static inline int check_new_page(struct page *page)
2139 if (likely(page_expected_state(page,
2140 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2143 check_new_page_bad(page);
2147 static inline bool free_pages_prezeroed(void)
2149 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2150 page_poisoning_enabled()) || want_init_on_free();
2153 #ifdef CONFIG_DEBUG_VM
2155 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2156 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2157 * also checked when pcp lists are refilled from the free lists.
2159 static inline bool check_pcp_refill(struct page *page)
2161 if (debug_pagealloc_enabled_static())
2162 return check_new_page(page);
2167 static inline bool check_new_pcp(struct page *page)
2169 return check_new_page(page);
2173 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2174 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2175 * enabled, they are also checked when being allocated from the pcp lists.
2177 static inline bool check_pcp_refill(struct page *page)
2179 return check_new_page(page);
2181 static inline bool check_new_pcp(struct page *page)
2183 if (debug_pagealloc_enabled_static())
2184 return check_new_page(page);
2188 #endif /* CONFIG_DEBUG_VM */
2190 static bool check_new_pages(struct page *page, unsigned int order)
2193 for (i = 0; i < (1 << order); i++) {
2194 struct page *p = page + i;
2196 if (unlikely(check_new_page(p)))
2203 inline void post_alloc_hook(struct page *page, unsigned int order,
2206 set_page_private(page, 0);
2207 set_page_refcounted(page);
2209 arch_alloc_page(page, order);
2210 if (debug_pagealloc_enabled_static())
2211 kernel_map_pages(page, 1 << order, 1);
2212 kasan_alloc_pages(page, order);
2213 kernel_poison_pages(page, 1 << order, 1);
2214 set_page_owner(page, order, gfp_flags);
2217 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2218 unsigned int alloc_flags)
2220 post_alloc_hook(page, order, gfp_flags);
2222 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2223 kernel_init_free_pages(page, 1 << order);
2225 if (order && (gfp_flags & __GFP_COMP))
2226 prep_compound_page(page, order);
2229 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2230 * allocate the page. The expectation is that the caller is taking
2231 * steps that will free more memory. The caller should avoid the page
2232 * being used for !PFMEMALLOC purposes.
2234 if (alloc_flags & ALLOC_NO_WATERMARKS)
2235 set_page_pfmemalloc(page);
2237 clear_page_pfmemalloc(page);
2241 * Go through the free lists for the given migratetype and remove
2242 * the smallest available page from the freelists
2244 static __always_inline
2245 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2248 unsigned int current_order;
2249 struct free_area *area;
2252 /* Find a page of the appropriate size in the preferred list */
2253 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2254 area = &(zone->free_area[current_order]);
2255 page = get_page_from_free_area(area, migratetype);
2258 del_page_from_free_list(page, zone, current_order);
2259 expand(zone, page, order, current_order, migratetype);
2260 set_pcppage_migratetype(page, migratetype);
2269 * This array describes the order lists are fallen back to when
2270 * the free lists for the desirable migrate type are depleted
2272 static int fallbacks[MIGRATE_TYPES][3] = {
2273 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2274 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2275 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2277 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2279 #ifdef CONFIG_MEMORY_ISOLATION
2280 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2285 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2288 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2291 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2292 unsigned int order) { return NULL; }
2296 * Move the free pages in a range to the free lists of the requested type.
2297 * Note that start_page and end_pages are not aligned on a pageblock
2298 * boundary. If alignment is required, use move_freepages_block()
2300 static int move_freepages(struct zone *zone,
2301 struct page *start_page, struct page *end_page,
2302 int migratetype, int *num_movable)
2306 int pages_moved = 0;
2308 for (page = start_page; page <= end_page;) {
2309 if (!pfn_valid_within(page_to_pfn(page))) {
2314 if (!PageBuddy(page)) {
2316 * We assume that pages that could be isolated for
2317 * migration are movable. But we don't actually try
2318 * isolating, as that would be expensive.
2321 (PageLRU(page) || __PageMovable(page)))
2328 /* Make sure we are not inadvertently changing nodes */
2329 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2330 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2332 order = page_order(page);
2333 move_to_free_list(page, zone, order, migratetype);
2335 pages_moved += 1 << order;
2341 int move_freepages_block(struct zone *zone, struct page *page,
2342 int migratetype, int *num_movable)
2344 unsigned long start_pfn, end_pfn;
2345 struct page *start_page, *end_page;
2350 start_pfn = page_to_pfn(page);
2351 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2352 start_page = pfn_to_page(start_pfn);
2353 end_page = start_page + pageblock_nr_pages - 1;
2354 end_pfn = start_pfn + pageblock_nr_pages - 1;
2356 /* Do not cross zone boundaries */
2357 if (!zone_spans_pfn(zone, start_pfn))
2359 if (!zone_spans_pfn(zone, end_pfn))
2362 return move_freepages(zone, start_page, end_page, migratetype,
2366 static void change_pageblock_range(struct page *pageblock_page,
2367 int start_order, int migratetype)
2369 int nr_pageblocks = 1 << (start_order - pageblock_order);
2371 while (nr_pageblocks--) {
2372 set_pageblock_migratetype(pageblock_page, migratetype);
2373 pageblock_page += pageblock_nr_pages;
2378 * When we are falling back to another migratetype during allocation, try to
2379 * steal extra free pages from the same pageblocks to satisfy further
2380 * allocations, instead of polluting multiple pageblocks.
2382 * If we are stealing a relatively large buddy page, it is likely there will
2383 * be more free pages in the pageblock, so try to steal them all. For
2384 * reclaimable and unmovable allocations, we steal regardless of page size,
2385 * as fragmentation caused by those allocations polluting movable pageblocks
2386 * is worse than movable allocations stealing from unmovable and reclaimable
2389 static bool can_steal_fallback(unsigned int order, int start_mt)
2392 * Leaving this order check is intended, although there is
2393 * relaxed order check in next check. The reason is that
2394 * we can actually steal whole pageblock if this condition met,
2395 * but, below check doesn't guarantee it and that is just heuristic
2396 * so could be changed anytime.
2398 if (order >= pageblock_order)
2401 if (order >= pageblock_order / 2 ||
2402 start_mt == MIGRATE_RECLAIMABLE ||
2403 start_mt == MIGRATE_UNMOVABLE ||
2404 page_group_by_mobility_disabled)
2410 static inline void boost_watermark(struct zone *zone)
2412 unsigned long max_boost;
2414 if (!watermark_boost_factor)
2417 * Don't bother in zones that are unlikely to produce results.
2418 * On small machines, including kdump capture kernels running
2419 * in a small area, boosting the watermark can cause an out of
2420 * memory situation immediately.
2422 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2425 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2426 watermark_boost_factor, 10000);
2429 * high watermark may be uninitialised if fragmentation occurs
2430 * very early in boot so do not boost. We do not fall
2431 * through and boost by pageblock_nr_pages as failing
2432 * allocations that early means that reclaim is not going
2433 * to help and it may even be impossible to reclaim the
2434 * boosted watermark resulting in a hang.
2439 max_boost = max(pageblock_nr_pages, max_boost);
2441 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2446 * This function implements actual steal behaviour. If order is large enough,
2447 * we can steal whole pageblock. If not, we first move freepages in this
2448 * pageblock to our migratetype and determine how many already-allocated pages
2449 * are there in the pageblock with a compatible migratetype. If at least half
2450 * of pages are free or compatible, we can change migratetype of the pageblock
2451 * itself, so pages freed in the future will be put on the correct free list.
2453 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2454 unsigned int alloc_flags, int start_type, bool whole_block)
2456 unsigned int current_order = page_order(page);
2457 int free_pages, movable_pages, alike_pages;
2460 old_block_type = get_pageblock_migratetype(page);
2463 * This can happen due to races and we want to prevent broken
2464 * highatomic accounting.
2466 if (is_migrate_highatomic(old_block_type))
2469 /* Take ownership for orders >= pageblock_order */
2470 if (current_order >= pageblock_order) {
2471 change_pageblock_range(page, current_order, start_type);
2476 * Boost watermarks to increase reclaim pressure to reduce the
2477 * likelihood of future fallbacks. Wake kswapd now as the node
2478 * may be balanced overall and kswapd will not wake naturally.
2480 boost_watermark(zone);
2481 if (alloc_flags & ALLOC_KSWAPD)
2482 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2484 /* We are not allowed to try stealing from the whole block */
2488 free_pages = move_freepages_block(zone, page, start_type,
2491 * Determine how many pages are compatible with our allocation.
2492 * For movable allocation, it's the number of movable pages which
2493 * we just obtained. For other types it's a bit more tricky.
2495 if (start_type == MIGRATE_MOVABLE) {
2496 alike_pages = movable_pages;
2499 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2500 * to MOVABLE pageblock, consider all non-movable pages as
2501 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2502 * vice versa, be conservative since we can't distinguish the
2503 * exact migratetype of non-movable pages.
2505 if (old_block_type == MIGRATE_MOVABLE)
2506 alike_pages = pageblock_nr_pages
2507 - (free_pages + movable_pages);
2512 /* moving whole block can fail due to zone boundary conditions */
2517 * If a sufficient number of pages in the block are either free or of
2518 * comparable migratability as our allocation, claim the whole block.
2520 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2521 page_group_by_mobility_disabled)
2522 set_pageblock_migratetype(page, start_type);
2527 move_to_free_list(page, zone, current_order, start_type);
2531 * Check whether there is a suitable fallback freepage with requested order.
2532 * If only_stealable is true, this function returns fallback_mt only if
2533 * we can steal other freepages all together. This would help to reduce
2534 * fragmentation due to mixed migratetype pages in one pageblock.
2536 int find_suitable_fallback(struct free_area *area, unsigned int order,
2537 int migratetype, bool only_stealable, bool *can_steal)
2542 if (area->nr_free == 0)
2547 fallback_mt = fallbacks[migratetype][i];
2548 if (fallback_mt == MIGRATE_TYPES)
2551 if (free_area_empty(area, fallback_mt))
2554 if (can_steal_fallback(order, migratetype))
2557 if (!only_stealable)
2568 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2569 * there are no empty page blocks that contain a page with a suitable order
2571 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2572 unsigned int alloc_order)
2575 unsigned long max_managed, flags;
2578 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2579 * Check is race-prone but harmless.
2581 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2582 if (zone->nr_reserved_highatomic >= max_managed)
2585 spin_lock_irqsave(&zone->lock, flags);
2587 /* Recheck the nr_reserved_highatomic limit under the lock */
2588 if (zone->nr_reserved_highatomic >= max_managed)
2592 mt = get_pageblock_migratetype(page);
2593 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2594 && !is_migrate_cma(mt)) {
2595 zone->nr_reserved_highatomic += pageblock_nr_pages;
2596 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2597 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2601 spin_unlock_irqrestore(&zone->lock, flags);
2605 * Used when an allocation is about to fail under memory pressure. This
2606 * potentially hurts the reliability of high-order allocations when under
2607 * intense memory pressure but failed atomic allocations should be easier
2608 * to recover from than an OOM.
2610 * If @force is true, try to unreserve a pageblock even though highatomic
2611 * pageblock is exhausted.
2613 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2616 struct zonelist *zonelist = ac->zonelist;
2617 unsigned long flags;
2624 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2627 * Preserve at least one pageblock unless memory pressure
2630 if (!force && zone->nr_reserved_highatomic <=
2634 spin_lock_irqsave(&zone->lock, flags);
2635 for (order = 0; order < MAX_ORDER; order++) {
2636 struct free_area *area = &(zone->free_area[order]);
2638 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2643 * In page freeing path, migratetype change is racy so
2644 * we can counter several free pages in a pageblock
2645 * in this loop althoug we changed the pageblock type
2646 * from highatomic to ac->migratetype. So we should
2647 * adjust the count once.
2649 if (is_migrate_highatomic_page(page)) {
2651 * It should never happen but changes to
2652 * locking could inadvertently allow a per-cpu
2653 * drain to add pages to MIGRATE_HIGHATOMIC
2654 * while unreserving so be safe and watch for
2657 zone->nr_reserved_highatomic -= min(
2659 zone->nr_reserved_highatomic);
2663 * Convert to ac->migratetype and avoid the normal
2664 * pageblock stealing heuristics. Minimally, the caller
2665 * is doing the work and needs the pages. More
2666 * importantly, if the block was always converted to
2667 * MIGRATE_UNMOVABLE or another type then the number
2668 * of pageblocks that cannot be completely freed
2671 set_pageblock_migratetype(page, ac->migratetype);
2672 ret = move_freepages_block(zone, page, ac->migratetype,
2675 spin_unlock_irqrestore(&zone->lock, flags);
2679 spin_unlock_irqrestore(&zone->lock, flags);
2686 * Try finding a free buddy page on the fallback list and put it on the free
2687 * list of requested migratetype, possibly along with other pages from the same
2688 * block, depending on fragmentation avoidance heuristics. Returns true if
2689 * fallback was found so that __rmqueue_smallest() can grab it.
2691 * The use of signed ints for order and current_order is a deliberate
2692 * deviation from the rest of this file, to make the for loop
2693 * condition simpler.
2695 static __always_inline bool
2696 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2697 unsigned int alloc_flags)
2699 struct free_area *area;
2701 int min_order = order;
2707 * Do not steal pages from freelists belonging to other pageblocks
2708 * i.e. orders < pageblock_order. If there are no local zones free,
2709 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2711 if (alloc_flags & ALLOC_NOFRAGMENT)
2712 min_order = pageblock_order;
2715 * Find the largest available free page in the other list. This roughly
2716 * approximates finding the pageblock with the most free pages, which
2717 * would be too costly to do exactly.
2719 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2721 area = &(zone->free_area[current_order]);
2722 fallback_mt = find_suitable_fallback(area, current_order,
2723 start_migratetype, false, &can_steal);
2724 if (fallback_mt == -1)
2728 * We cannot steal all free pages from the pageblock and the
2729 * requested migratetype is movable. In that case it's better to
2730 * steal and split the smallest available page instead of the
2731 * largest available page, because even if the next movable
2732 * allocation falls back into a different pageblock than this
2733 * one, it won't cause permanent fragmentation.
2735 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2736 && current_order > order)
2745 for (current_order = order; current_order < MAX_ORDER;
2747 area = &(zone->free_area[current_order]);
2748 fallback_mt = find_suitable_fallback(area, current_order,
2749 start_migratetype, false, &can_steal);
2750 if (fallback_mt != -1)
2755 * This should not happen - we already found a suitable fallback
2756 * when looking for the largest page.
2758 VM_BUG_ON(current_order == MAX_ORDER);
2761 page = get_page_from_free_area(area, fallback_mt);
2763 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2766 trace_mm_page_alloc_extfrag(page, order, current_order,
2767 start_migratetype, fallback_mt);
2774 * Do the hard work of removing an element from the buddy allocator.
2775 * Call me with the zone->lock already held.
2777 static __always_inline struct page *
2778 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2779 unsigned int alloc_flags)
2785 * Balance movable allocations between regular and CMA areas by
2786 * allocating from CMA when over half of the zone's free memory
2787 * is in the CMA area.
2789 if (alloc_flags & ALLOC_CMA &&
2790 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2791 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2792 page = __rmqueue_cma_fallback(zone, order);
2798 page = __rmqueue_smallest(zone, order, migratetype);
2799 if (unlikely(!page)) {
2800 if (alloc_flags & ALLOC_CMA)
2801 page = __rmqueue_cma_fallback(zone, order);
2803 if (!page && __rmqueue_fallback(zone, order, migratetype,
2808 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2813 * Obtain a specified number of elements from the buddy allocator, all under
2814 * a single hold of the lock, for efficiency. Add them to the supplied list.
2815 * Returns the number of new pages which were placed at *list.
2817 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2818 unsigned long count, struct list_head *list,
2819 int migratetype, unsigned int alloc_flags)
2823 spin_lock(&zone->lock);
2824 for (i = 0; i < count; ++i) {
2825 struct page *page = __rmqueue(zone, order, migratetype,
2827 if (unlikely(page == NULL))
2830 if (unlikely(check_pcp_refill(page)))
2834 * Split buddy pages returned by expand() are received here in
2835 * physical page order. The page is added to the tail of
2836 * caller's list. From the callers perspective, the linked list
2837 * is ordered by page number under some conditions. This is
2838 * useful for IO devices that can forward direction from the
2839 * head, thus also in the physical page order. This is useful
2840 * for IO devices that can merge IO requests if the physical
2841 * pages are ordered properly.
2843 list_add_tail(&page->lru, list);
2845 if (is_migrate_cma(get_pcppage_migratetype(page)))
2846 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2851 * i pages were removed from the buddy list even if some leak due
2852 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2853 * on i. Do not confuse with 'alloced' which is the number of
2854 * pages added to the pcp list.
2856 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2857 spin_unlock(&zone->lock);
2863 * Called from the vmstat counter updater to drain pagesets of this
2864 * currently executing processor on remote nodes after they have
2867 * Note that this function must be called with the thread pinned to
2868 * a single processor.
2870 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2872 unsigned long flags;
2873 int to_drain, batch;
2875 local_irq_save(flags);
2876 batch = READ_ONCE(pcp->batch);
2877 to_drain = min(pcp->count, batch);
2879 free_pcppages_bulk(zone, to_drain, pcp);
2880 local_irq_restore(flags);
2885 * Drain pcplists of the indicated processor and zone.
2887 * The processor must either be the current processor and the
2888 * thread pinned to the current processor or a processor that
2891 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2893 unsigned long flags;
2894 struct per_cpu_pageset *pset;
2895 struct per_cpu_pages *pcp;
2897 local_irq_save(flags);
2898 pset = per_cpu_ptr(zone->pageset, cpu);
2902 free_pcppages_bulk(zone, pcp->count, pcp);
2903 local_irq_restore(flags);
2907 * Drain pcplists of all zones on the indicated processor.
2909 * The processor must either be the current processor and the
2910 * thread pinned to the current processor or a processor that
2913 static void drain_pages(unsigned int cpu)
2917 for_each_populated_zone(zone) {
2918 drain_pages_zone(cpu, zone);
2923 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2925 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2926 * the single zone's pages.
2928 void drain_local_pages(struct zone *zone)
2930 int cpu = smp_processor_id();
2933 drain_pages_zone(cpu, zone);
2938 static void drain_local_pages_wq(struct work_struct *work)
2940 struct pcpu_drain *drain;
2942 drain = container_of(work, struct pcpu_drain, work);
2945 * drain_all_pages doesn't use proper cpu hotplug protection so
2946 * we can race with cpu offline when the WQ can move this from
2947 * a cpu pinned worker to an unbound one. We can operate on a different
2948 * cpu which is allright but we also have to make sure to not move to
2952 drain_local_pages(drain->zone);
2957 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2959 * When zone parameter is non-NULL, spill just the single zone's pages.
2961 * Note that this can be extremely slow as the draining happens in a workqueue.
2963 void drain_all_pages(struct zone *zone)
2968 * Allocate in the BSS so we wont require allocation in
2969 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2971 static cpumask_t cpus_with_pcps;
2974 * Make sure nobody triggers this path before mm_percpu_wq is fully
2977 if (WARN_ON_ONCE(!mm_percpu_wq))
2981 * Do not drain if one is already in progress unless it's specific to
2982 * a zone. Such callers are primarily CMA and memory hotplug and need
2983 * the drain to be complete when the call returns.
2985 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2988 mutex_lock(&pcpu_drain_mutex);
2992 * We don't care about racing with CPU hotplug event
2993 * as offline notification will cause the notified
2994 * cpu to drain that CPU pcps and on_each_cpu_mask
2995 * disables preemption as part of its processing
2997 for_each_online_cpu(cpu) {
2998 struct per_cpu_pageset *pcp;
3000 bool has_pcps = false;
3003 pcp = per_cpu_ptr(zone->pageset, cpu);
3007 for_each_populated_zone(z) {
3008 pcp = per_cpu_ptr(z->pageset, cpu);
3009 if (pcp->pcp.count) {
3017 cpumask_set_cpu(cpu, &cpus_with_pcps);
3019 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3022 for_each_cpu(cpu, &cpus_with_pcps) {
3023 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3026 INIT_WORK(&drain->work, drain_local_pages_wq);
3027 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3029 for_each_cpu(cpu, &cpus_with_pcps)
3030 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3032 mutex_unlock(&pcpu_drain_mutex);
3035 #ifdef CONFIG_HIBERNATION
3038 * Touch the watchdog for every WD_PAGE_COUNT pages.
3040 #define WD_PAGE_COUNT (128*1024)
3042 void mark_free_pages(struct zone *zone)
3044 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3045 unsigned long flags;
3046 unsigned int order, t;
3049 if (zone_is_empty(zone))
3052 spin_lock_irqsave(&zone->lock, flags);
3054 max_zone_pfn = zone_end_pfn(zone);
3055 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3056 if (pfn_valid(pfn)) {
3057 page = pfn_to_page(pfn);
3059 if (!--page_count) {
3060 touch_nmi_watchdog();
3061 page_count = WD_PAGE_COUNT;
3064 if (page_zone(page) != zone)
3067 if (!swsusp_page_is_forbidden(page))
3068 swsusp_unset_page_free(page);
3071 for_each_migratetype_order(order, t) {
3072 list_for_each_entry(page,
3073 &zone->free_area[order].free_list[t], lru) {
3076 pfn = page_to_pfn(page);
3077 for (i = 0; i < (1UL << order); i++) {
3078 if (!--page_count) {
3079 touch_nmi_watchdog();
3080 page_count = WD_PAGE_COUNT;
3082 swsusp_set_page_free(pfn_to_page(pfn + i));
3086 spin_unlock_irqrestore(&zone->lock, flags);
3088 #endif /* CONFIG_PM */
3090 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3094 if (!free_pcp_prepare(page))
3097 migratetype = get_pfnblock_migratetype(page, pfn);
3098 set_pcppage_migratetype(page, migratetype);
3102 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3104 struct zone *zone = page_zone(page);
3105 struct per_cpu_pages *pcp;
3108 migratetype = get_pcppage_migratetype(page);
3109 __count_vm_event(PGFREE);
3112 * We only track unmovable, reclaimable and movable on pcp lists.
3113 * Free ISOLATE pages back to the allocator because they are being
3114 * offlined but treat HIGHATOMIC as movable pages so we can get those
3115 * areas back if necessary. Otherwise, we may have to free
3116 * excessively into the page allocator
3118 if (migratetype >= MIGRATE_PCPTYPES) {
3119 if (unlikely(is_migrate_isolate(migratetype))) {
3120 free_one_page(zone, page, pfn, 0, migratetype);
3123 migratetype = MIGRATE_MOVABLE;
3126 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3127 list_add(&page->lru, &pcp->lists[migratetype]);
3129 if (pcp->count >= pcp->high) {
3130 unsigned long batch = READ_ONCE(pcp->batch);
3131 free_pcppages_bulk(zone, batch, pcp);
3136 * Free a 0-order page
3138 void free_unref_page(struct page *page)
3140 unsigned long flags;
3141 unsigned long pfn = page_to_pfn(page);
3143 if (!free_unref_page_prepare(page, pfn))
3146 local_irq_save(flags);
3147 free_unref_page_commit(page, pfn);
3148 local_irq_restore(flags);
3152 * Free a list of 0-order pages
3154 void free_unref_page_list(struct list_head *list)
3156 struct page *page, *next;
3157 unsigned long flags, pfn;
3158 int batch_count = 0;
3160 /* Prepare pages for freeing */
3161 list_for_each_entry_safe(page, next, list, lru) {
3162 pfn = page_to_pfn(page);
3163 if (!free_unref_page_prepare(page, pfn))
3164 list_del(&page->lru);
3165 set_page_private(page, pfn);
3168 local_irq_save(flags);
3169 list_for_each_entry_safe(page, next, list, lru) {
3170 unsigned long pfn = page_private(page);
3172 set_page_private(page, 0);
3173 trace_mm_page_free_batched(page);
3174 free_unref_page_commit(page, pfn);
3177 * Guard against excessive IRQ disabled times when we get
3178 * a large list of pages to free.
3180 if (++batch_count == SWAP_CLUSTER_MAX) {
3181 local_irq_restore(flags);
3183 local_irq_save(flags);
3186 local_irq_restore(flags);
3190 * split_page takes a non-compound higher-order page, and splits it into
3191 * n (1<<order) sub-pages: page[0..n]
3192 * Each sub-page must be freed individually.
3194 * Note: this is probably too low level an operation for use in drivers.
3195 * Please consult with lkml before using this in your driver.
3197 void split_page(struct page *page, unsigned int order)
3201 VM_BUG_ON_PAGE(PageCompound(page), page);
3202 VM_BUG_ON_PAGE(!page_count(page), page);
3204 for (i = 1; i < (1 << order); i++)
3205 set_page_refcounted(page + i);
3206 split_page_owner(page, order);
3208 EXPORT_SYMBOL_GPL(split_page);
3210 int __isolate_free_page(struct page *page, unsigned int order)
3212 unsigned long watermark;
3216 BUG_ON(!PageBuddy(page));
3218 zone = page_zone(page);
3219 mt = get_pageblock_migratetype(page);
3221 if (!is_migrate_isolate(mt)) {
3223 * Obey watermarks as if the page was being allocated. We can
3224 * emulate a high-order watermark check with a raised order-0
3225 * watermark, because we already know our high-order page
3228 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3229 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3232 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3235 /* Remove page from free list */
3237 del_page_from_free_list(page, zone, order);
3240 * Set the pageblock if the isolated page is at least half of a
3243 if (order >= pageblock_order - 1) {
3244 struct page *endpage = page + (1 << order) - 1;
3245 for (; page < endpage; page += pageblock_nr_pages) {
3246 int mt = get_pageblock_migratetype(page);
3247 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3248 && !is_migrate_highatomic(mt))
3249 set_pageblock_migratetype(page,
3255 return 1UL << order;
3259 * __putback_isolated_page - Return a now-isolated page back where we got it
3260 * @page: Page that was isolated
3261 * @order: Order of the isolated page
3262 * @mt: The page's pageblock's migratetype
3264 * This function is meant to return a page pulled from the free lists via
3265 * __isolate_free_page back to the free lists they were pulled from.
3267 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3269 struct zone *zone = page_zone(page);
3271 /* zone lock should be held when this function is called */
3272 lockdep_assert_held(&zone->lock);
3274 /* Return isolated page to tail of freelist. */
3275 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3279 * Update NUMA hit/miss statistics
3281 * Must be called with interrupts disabled.
3283 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3286 enum numa_stat_item local_stat = NUMA_LOCAL;
3288 /* skip numa counters update if numa stats is disabled */
3289 if (!static_branch_likely(&vm_numa_stat_key))
3292 if (zone_to_nid(z) != numa_node_id())
3293 local_stat = NUMA_OTHER;
3295 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3296 __inc_numa_state(z, NUMA_HIT);
3298 __inc_numa_state(z, NUMA_MISS);
3299 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3301 __inc_numa_state(z, local_stat);
3305 /* Remove page from the per-cpu list, caller must protect the list */
3306 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3307 unsigned int alloc_flags,
3308 struct per_cpu_pages *pcp,
3309 struct list_head *list)
3314 if (list_empty(list)) {
3315 pcp->count += rmqueue_bulk(zone, 0,
3317 migratetype, alloc_flags);
3318 if (unlikely(list_empty(list)))
3322 page = list_first_entry(list, struct page, lru);
3323 list_del(&page->lru);
3325 } while (check_new_pcp(page));
3330 /* Lock and remove page from the per-cpu list */
3331 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3332 struct zone *zone, gfp_t gfp_flags,
3333 int migratetype, unsigned int alloc_flags)
3335 struct per_cpu_pages *pcp;
3336 struct list_head *list;
3338 unsigned long flags;
3340 local_irq_save(flags);
3341 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3342 list = &pcp->lists[migratetype];
3343 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3345 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3346 zone_statistics(preferred_zone, zone);
3348 local_irq_restore(flags);
3353 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3356 struct page *rmqueue(struct zone *preferred_zone,
3357 struct zone *zone, unsigned int order,
3358 gfp_t gfp_flags, unsigned int alloc_flags,
3361 unsigned long flags;
3364 if (likely(order == 0)) {
3365 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3366 migratetype, alloc_flags);
3371 * We most definitely don't want callers attempting to
3372 * allocate greater than order-1 page units with __GFP_NOFAIL.
3374 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3375 spin_lock_irqsave(&zone->lock, flags);
3379 if (alloc_flags & ALLOC_HARDER) {
3380 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3382 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3385 page = __rmqueue(zone, order, migratetype, alloc_flags);
3386 } while (page && check_new_pages(page, order));
3387 spin_unlock(&zone->lock);
3390 __mod_zone_freepage_state(zone, -(1 << order),
3391 get_pcppage_migratetype(page));
3393 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3394 zone_statistics(preferred_zone, zone);
3395 local_irq_restore(flags);
3398 /* Separate test+clear to avoid unnecessary atomics */
3399 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3400 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3401 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3404 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3408 local_irq_restore(flags);
3412 #ifdef CONFIG_FAIL_PAGE_ALLOC
3415 struct fault_attr attr;
3417 bool ignore_gfp_highmem;
3418 bool ignore_gfp_reclaim;
3420 } fail_page_alloc = {
3421 .attr = FAULT_ATTR_INITIALIZER,
3422 .ignore_gfp_reclaim = true,
3423 .ignore_gfp_highmem = true,
3427 static int __init setup_fail_page_alloc(char *str)
3429 return setup_fault_attr(&fail_page_alloc.attr, str);
3431 __setup("fail_page_alloc=", setup_fail_page_alloc);
3433 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3435 if (order < fail_page_alloc.min_order)
3437 if (gfp_mask & __GFP_NOFAIL)
3439 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3441 if (fail_page_alloc.ignore_gfp_reclaim &&
3442 (gfp_mask & __GFP_DIRECT_RECLAIM))
3445 return should_fail(&fail_page_alloc.attr, 1 << order);
3448 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3450 static int __init fail_page_alloc_debugfs(void)
3452 umode_t mode = S_IFREG | 0600;
3455 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3456 &fail_page_alloc.attr);
3458 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3459 &fail_page_alloc.ignore_gfp_reclaim);
3460 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3461 &fail_page_alloc.ignore_gfp_highmem);
3462 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3467 late_initcall(fail_page_alloc_debugfs);
3469 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3471 #else /* CONFIG_FAIL_PAGE_ALLOC */
3473 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3478 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3480 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3482 return __should_fail_alloc_page(gfp_mask, order);
3484 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3486 static inline long __zone_watermark_unusable_free(struct zone *z,
3487 unsigned int order, unsigned int alloc_flags)
3489 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3490 long unusable_free = (1 << order) - 1;
3493 * If the caller does not have rights to ALLOC_HARDER then subtract
3494 * the high-atomic reserves. This will over-estimate the size of the
3495 * atomic reserve but it avoids a search.
3497 if (likely(!alloc_harder))
3498 unusable_free += z->nr_reserved_highatomic;
3501 /* If allocation can't use CMA areas don't use free CMA pages */
3502 if (!(alloc_flags & ALLOC_CMA))
3503 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3506 return unusable_free;
3510 * Return true if free base pages are above 'mark'. For high-order checks it
3511 * will return true of the order-0 watermark is reached and there is at least
3512 * one free page of a suitable size. Checking now avoids taking the zone lock
3513 * to check in the allocation paths if no pages are free.
3515 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3516 int highest_zoneidx, unsigned int alloc_flags,
3521 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3523 /* free_pages may go negative - that's OK */
3524 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3526 if (alloc_flags & ALLOC_HIGH)
3529 if (unlikely(alloc_harder)) {
3531 * OOM victims can try even harder than normal ALLOC_HARDER
3532 * users on the grounds that it's definitely going to be in
3533 * the exit path shortly and free memory. Any allocation it
3534 * makes during the free path will be small and short-lived.
3536 if (alloc_flags & ALLOC_OOM)
3543 * Check watermarks for an order-0 allocation request. If these
3544 * are not met, then a high-order request also cannot go ahead
3545 * even if a suitable page happened to be free.
3547 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3550 /* If this is an order-0 request then the watermark is fine */
3554 /* For a high-order request, check at least one suitable page is free */
3555 for (o = order; o < MAX_ORDER; o++) {
3556 struct free_area *area = &z->free_area[o];
3562 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3563 if (!free_area_empty(area, mt))
3568 if ((alloc_flags & ALLOC_CMA) &&
3569 !free_area_empty(area, MIGRATE_CMA)) {
3573 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3579 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3580 int highest_zoneidx, unsigned int alloc_flags)
3582 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3583 zone_page_state(z, NR_FREE_PAGES));
3586 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3587 unsigned long mark, int highest_zoneidx,
3588 unsigned int alloc_flags, gfp_t gfp_mask)
3592 free_pages = zone_page_state(z, NR_FREE_PAGES);
3595 * Fast check for order-0 only. If this fails then the reserves
3596 * need to be calculated.
3601 fast_free = free_pages;
3602 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3603 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3607 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3611 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3612 * when checking the min watermark. The min watermark is the
3613 * point where boosting is ignored so that kswapd is woken up
3614 * when below the low watermark.
3616 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3617 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3618 mark = z->_watermark[WMARK_MIN];
3619 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3620 alloc_flags, free_pages);
3626 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3627 unsigned long mark, int highest_zoneidx)
3629 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3631 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3632 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3634 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3639 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3641 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3642 node_reclaim_distance;
3644 #else /* CONFIG_NUMA */
3645 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3649 #endif /* CONFIG_NUMA */
3652 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3653 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3654 * premature use of a lower zone may cause lowmem pressure problems that
3655 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3656 * probably too small. It only makes sense to spread allocations to avoid
3657 * fragmentation between the Normal and DMA32 zones.
3659 static inline unsigned int
3660 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3662 unsigned int alloc_flags;
3665 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3668 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3670 #ifdef CONFIG_ZONE_DMA32
3674 if (zone_idx(zone) != ZONE_NORMAL)
3678 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3679 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3680 * on UMA that if Normal is populated then so is DMA32.
3682 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3683 if (nr_online_nodes > 1 && !populated_zone(--zone))
3686 alloc_flags |= ALLOC_NOFRAGMENT;
3687 #endif /* CONFIG_ZONE_DMA32 */
3691 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3692 unsigned int alloc_flags)
3695 unsigned int pflags = current->flags;
3697 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3698 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3699 alloc_flags |= ALLOC_CMA;
3706 * get_page_from_freelist goes through the zonelist trying to allocate
3709 static struct page *
3710 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3711 const struct alloc_context *ac)
3715 struct pglist_data *last_pgdat_dirty_limit = NULL;
3720 * Scan zonelist, looking for a zone with enough free.
3721 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3723 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3724 z = ac->preferred_zoneref;
3725 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist,
3726 ac->highest_zoneidx, ac->nodemask) {
3730 if (cpusets_enabled() &&
3731 (alloc_flags & ALLOC_CPUSET) &&
3732 !__cpuset_zone_allowed(zone, gfp_mask))
3735 * When allocating a page cache page for writing, we
3736 * want to get it from a node that is within its dirty
3737 * limit, such that no single node holds more than its
3738 * proportional share of globally allowed dirty pages.
3739 * The dirty limits take into account the node's
3740 * lowmem reserves and high watermark so that kswapd
3741 * should be able to balance it without having to
3742 * write pages from its LRU list.
3744 * XXX: For now, allow allocations to potentially
3745 * exceed the per-node dirty limit in the slowpath
3746 * (spread_dirty_pages unset) before going into reclaim,
3747 * which is important when on a NUMA setup the allowed
3748 * nodes are together not big enough to reach the
3749 * global limit. The proper fix for these situations
3750 * will require awareness of nodes in the
3751 * dirty-throttling and the flusher threads.
3753 if (ac->spread_dirty_pages) {
3754 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3757 if (!node_dirty_ok(zone->zone_pgdat)) {
3758 last_pgdat_dirty_limit = zone->zone_pgdat;
3763 if (no_fallback && nr_online_nodes > 1 &&
3764 zone != ac->preferred_zoneref->zone) {
3768 * If moving to a remote node, retry but allow
3769 * fragmenting fallbacks. Locality is more important
3770 * than fragmentation avoidance.
3772 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3773 if (zone_to_nid(zone) != local_nid) {
3774 alloc_flags &= ~ALLOC_NOFRAGMENT;
3779 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3780 if (!zone_watermark_fast(zone, order, mark,
3781 ac->highest_zoneidx, alloc_flags,
3785 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3787 * Watermark failed for this zone, but see if we can
3788 * grow this zone if it contains deferred pages.
3790 if (static_branch_unlikely(&deferred_pages)) {
3791 if (_deferred_grow_zone(zone, order))
3795 /* Checked here to keep the fast path fast */
3796 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3797 if (alloc_flags & ALLOC_NO_WATERMARKS)
3800 if (node_reclaim_mode == 0 ||
3801 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3804 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3806 case NODE_RECLAIM_NOSCAN:
3809 case NODE_RECLAIM_FULL:
3810 /* scanned but unreclaimable */
3813 /* did we reclaim enough */
3814 if (zone_watermark_ok(zone, order, mark,
3815 ac->highest_zoneidx, alloc_flags))
3823 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3824 gfp_mask, alloc_flags, ac->migratetype);
3826 prep_new_page(page, order, gfp_mask, alloc_flags);
3829 * If this is a high-order atomic allocation then check
3830 * if the pageblock should be reserved for the future
3832 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3833 reserve_highatomic_pageblock(page, zone, order);
3837 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3838 /* Try again if zone has deferred pages */
3839 if (static_branch_unlikely(&deferred_pages)) {
3840 if (_deferred_grow_zone(zone, order))
3848 * It's possible on a UMA machine to get through all zones that are
3849 * fragmented. If avoiding fragmentation, reset and try again.
3852 alloc_flags &= ~ALLOC_NOFRAGMENT;
3859 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3861 unsigned int filter = SHOW_MEM_FILTER_NODES;
3864 * This documents exceptions given to allocations in certain
3865 * contexts that are allowed to allocate outside current's set
3868 if (!(gfp_mask & __GFP_NOMEMALLOC))
3869 if (tsk_is_oom_victim(current) ||
3870 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3871 filter &= ~SHOW_MEM_FILTER_NODES;
3872 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3873 filter &= ~SHOW_MEM_FILTER_NODES;
3875 show_mem(filter, nodemask);
3878 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3880 struct va_format vaf;
3882 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3884 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3887 va_start(args, fmt);
3890 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3891 current->comm, &vaf, gfp_mask, &gfp_mask,
3892 nodemask_pr_args(nodemask));
3895 cpuset_print_current_mems_allowed();
3898 warn_alloc_show_mem(gfp_mask, nodemask);
3901 static inline struct page *
3902 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3903 unsigned int alloc_flags,
3904 const struct alloc_context *ac)
3908 page = get_page_from_freelist(gfp_mask, order,
3909 alloc_flags|ALLOC_CPUSET, ac);
3911 * fallback to ignore cpuset restriction if our nodes
3915 page = get_page_from_freelist(gfp_mask, order,
3921 static inline struct page *
3922 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3923 const struct alloc_context *ac, unsigned long *did_some_progress)
3925 struct oom_control oc = {
3926 .zonelist = ac->zonelist,
3927 .nodemask = ac->nodemask,
3929 .gfp_mask = gfp_mask,
3934 *did_some_progress = 0;
3937 * Acquire the oom lock. If that fails, somebody else is
3938 * making progress for us.
3940 if (!mutex_trylock(&oom_lock)) {
3941 *did_some_progress = 1;
3942 schedule_timeout_uninterruptible(1);
3947 * Go through the zonelist yet one more time, keep very high watermark
3948 * here, this is only to catch a parallel oom killing, we must fail if
3949 * we're still under heavy pressure. But make sure that this reclaim
3950 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3951 * allocation which will never fail due to oom_lock already held.
3953 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3954 ~__GFP_DIRECT_RECLAIM, order,
3955 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3959 /* Coredumps can quickly deplete all memory reserves */
3960 if (current->flags & PF_DUMPCORE)
3962 /* The OOM killer will not help higher order allocs */
3963 if (order > PAGE_ALLOC_COSTLY_ORDER)
3966 * We have already exhausted all our reclaim opportunities without any
3967 * success so it is time to admit defeat. We will skip the OOM killer
3968 * because it is very likely that the caller has a more reasonable
3969 * fallback than shooting a random task.
3971 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3973 /* The OOM killer does not needlessly kill tasks for lowmem */
3974 if (ac->highest_zoneidx < ZONE_NORMAL)
3976 if (pm_suspended_storage())
3979 * XXX: GFP_NOFS allocations should rather fail than rely on
3980 * other request to make a forward progress.
3981 * We are in an unfortunate situation where out_of_memory cannot
3982 * do much for this context but let's try it to at least get
3983 * access to memory reserved if the current task is killed (see
3984 * out_of_memory). Once filesystems are ready to handle allocation
3985 * failures more gracefully we should just bail out here.
3988 /* The OOM killer may not free memory on a specific node */
3989 if (gfp_mask & __GFP_THISNODE)
3992 /* Exhausted what can be done so it's blame time */
3993 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3994 *did_some_progress = 1;
3997 * Help non-failing allocations by giving them access to memory
4000 if (gfp_mask & __GFP_NOFAIL)
4001 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4002 ALLOC_NO_WATERMARKS, ac);
4005 mutex_unlock(&oom_lock);
4010 * Maximum number of compaction retries wit a progress before OOM
4011 * killer is consider as the only way to move forward.
4013 #define MAX_COMPACT_RETRIES 16
4015 #ifdef CONFIG_COMPACTION
4016 /* Try memory compaction for high-order allocations before reclaim */
4017 static struct page *
4018 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4019 unsigned int alloc_flags, const struct alloc_context *ac,
4020 enum compact_priority prio, enum compact_result *compact_result)
4022 struct page *page = NULL;
4023 unsigned long pflags;
4024 unsigned int noreclaim_flag;
4029 psi_memstall_enter(&pflags);
4030 noreclaim_flag = memalloc_noreclaim_save();
4032 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4035 memalloc_noreclaim_restore(noreclaim_flag);
4036 psi_memstall_leave(&pflags);
4039 * At least in one zone compaction wasn't deferred or skipped, so let's
4040 * count a compaction stall
4042 count_vm_event(COMPACTSTALL);
4044 /* Prep a captured page if available */
4046 prep_new_page(page, order, gfp_mask, alloc_flags);
4048 /* Try get a page from the freelist if available */
4050 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4053 struct zone *zone = page_zone(page);
4055 zone->compact_blockskip_flush = false;
4056 compaction_defer_reset(zone, order, true);
4057 count_vm_event(COMPACTSUCCESS);
4062 * It's bad if compaction run occurs and fails. The most likely reason
4063 * is that pages exist, but not enough to satisfy watermarks.
4065 count_vm_event(COMPACTFAIL);
4073 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4074 enum compact_result compact_result,
4075 enum compact_priority *compact_priority,
4076 int *compaction_retries)
4078 int max_retries = MAX_COMPACT_RETRIES;
4081 int retries = *compaction_retries;
4082 enum compact_priority priority = *compact_priority;
4087 if (compaction_made_progress(compact_result))
4088 (*compaction_retries)++;
4091 * compaction considers all the zone as desperately out of memory
4092 * so it doesn't really make much sense to retry except when the
4093 * failure could be caused by insufficient priority
4095 if (compaction_failed(compact_result))
4096 goto check_priority;
4099 * compaction was skipped because there are not enough order-0 pages
4100 * to work with, so we retry only if it looks like reclaim can help.
4102 if (compaction_needs_reclaim(compact_result)) {
4103 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4108 * make sure the compaction wasn't deferred or didn't bail out early
4109 * due to locks contention before we declare that we should give up.
4110 * But the next retry should use a higher priority if allowed, so
4111 * we don't just keep bailing out endlessly.
4113 if (compaction_withdrawn(compact_result)) {
4114 goto check_priority;
4118 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4119 * costly ones because they are de facto nofail and invoke OOM
4120 * killer to move on while costly can fail and users are ready
4121 * to cope with that. 1/4 retries is rather arbitrary but we
4122 * would need much more detailed feedback from compaction to
4123 * make a better decision.
4125 if (order > PAGE_ALLOC_COSTLY_ORDER)
4127 if (*compaction_retries <= max_retries) {
4133 * Make sure there are attempts at the highest priority if we exhausted
4134 * all retries or failed at the lower priorities.
4137 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4138 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4140 if (*compact_priority > min_priority) {
4141 (*compact_priority)--;
4142 *compaction_retries = 0;
4146 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4150 static inline struct page *
4151 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4152 unsigned int alloc_flags, const struct alloc_context *ac,
4153 enum compact_priority prio, enum compact_result *compact_result)
4155 *compact_result = COMPACT_SKIPPED;
4160 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4161 enum compact_result compact_result,
4162 enum compact_priority *compact_priority,
4163 int *compaction_retries)
4168 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4172 * There are setups with compaction disabled which would prefer to loop
4173 * inside the allocator rather than hit the oom killer prematurely.
4174 * Let's give them a good hope and keep retrying while the order-0
4175 * watermarks are OK.
4177 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4178 ac->highest_zoneidx, ac->nodemask) {
4179 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4180 ac->highest_zoneidx, alloc_flags))
4185 #endif /* CONFIG_COMPACTION */
4187 #ifdef CONFIG_LOCKDEP
4188 static struct lockdep_map __fs_reclaim_map =
4189 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4191 static bool __need_fs_reclaim(gfp_t gfp_mask)
4193 gfp_mask = current_gfp_context(gfp_mask);
4195 /* no reclaim without waiting on it */
4196 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4199 /* this guy won't enter reclaim */
4200 if (current->flags & PF_MEMALLOC)
4203 /* We're only interested __GFP_FS allocations for now */
4204 if (!(gfp_mask & __GFP_FS))
4207 if (gfp_mask & __GFP_NOLOCKDEP)
4213 void __fs_reclaim_acquire(void)
4215 lock_map_acquire(&__fs_reclaim_map);
4218 void __fs_reclaim_release(void)
4220 lock_map_release(&__fs_reclaim_map);
4223 void fs_reclaim_acquire(gfp_t gfp_mask)
4225 if (__need_fs_reclaim(gfp_mask))
4226 __fs_reclaim_acquire();
4228 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4230 void fs_reclaim_release(gfp_t gfp_mask)
4232 if (__need_fs_reclaim(gfp_mask))
4233 __fs_reclaim_release();
4235 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4238 /* Perform direct synchronous page reclaim */
4240 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4241 const struct alloc_context *ac)
4244 unsigned int noreclaim_flag;
4245 unsigned long pflags;
4249 /* We now go into synchronous reclaim */
4250 cpuset_memory_pressure_bump();
4251 psi_memstall_enter(&pflags);
4252 fs_reclaim_acquire(gfp_mask);
4253 noreclaim_flag = memalloc_noreclaim_save();
4255 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4258 memalloc_noreclaim_restore(noreclaim_flag);
4259 fs_reclaim_release(gfp_mask);
4260 psi_memstall_leave(&pflags);
4267 /* The really slow allocator path where we enter direct reclaim */
4268 static inline struct page *
4269 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4270 unsigned int alloc_flags, const struct alloc_context *ac,
4271 unsigned long *did_some_progress)
4273 struct page *page = NULL;
4274 bool drained = false;
4276 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4277 if (unlikely(!(*did_some_progress)))
4281 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4284 * If an allocation failed after direct reclaim, it could be because
4285 * pages are pinned on the per-cpu lists or in high alloc reserves.
4286 * Shrink them and try again
4288 if (!page && !drained) {
4289 unreserve_highatomic_pageblock(ac, false);
4290 drain_all_pages(NULL);
4298 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4299 const struct alloc_context *ac)
4303 pg_data_t *last_pgdat = NULL;
4304 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4306 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4308 if (last_pgdat != zone->zone_pgdat)
4309 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4310 last_pgdat = zone->zone_pgdat;
4314 static inline unsigned int
4315 gfp_to_alloc_flags(gfp_t gfp_mask)
4317 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4320 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4321 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4322 * to save two branches.
4324 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4325 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4328 * The caller may dip into page reserves a bit more if the caller
4329 * cannot run direct reclaim, or if the caller has realtime scheduling
4330 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4331 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4333 alloc_flags |= (__force int)
4334 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4336 if (gfp_mask & __GFP_ATOMIC) {
4338 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4339 * if it can't schedule.
4341 if (!(gfp_mask & __GFP_NOMEMALLOC))
4342 alloc_flags |= ALLOC_HARDER;
4344 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4345 * comment for __cpuset_node_allowed().
4347 alloc_flags &= ~ALLOC_CPUSET;
4348 } else if (unlikely(rt_task(current)) && !in_interrupt())
4349 alloc_flags |= ALLOC_HARDER;
4351 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4356 static bool oom_reserves_allowed(struct task_struct *tsk)
4358 if (!tsk_is_oom_victim(tsk))
4362 * !MMU doesn't have oom reaper so give access to memory reserves
4363 * only to the thread with TIF_MEMDIE set
4365 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4372 * Distinguish requests which really need access to full memory
4373 * reserves from oom victims which can live with a portion of it
4375 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4377 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4379 if (gfp_mask & __GFP_MEMALLOC)
4380 return ALLOC_NO_WATERMARKS;
4381 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4382 return ALLOC_NO_WATERMARKS;
4383 if (!in_interrupt()) {
4384 if (current->flags & PF_MEMALLOC)
4385 return ALLOC_NO_WATERMARKS;
4386 else if (oom_reserves_allowed(current))
4393 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4395 return !!__gfp_pfmemalloc_flags(gfp_mask);
4399 * Checks whether it makes sense to retry the reclaim to make a forward progress
4400 * for the given allocation request.
4402 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4403 * without success, or when we couldn't even meet the watermark if we
4404 * reclaimed all remaining pages on the LRU lists.
4406 * Returns true if a retry is viable or false to enter the oom path.
4409 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4410 struct alloc_context *ac, int alloc_flags,
4411 bool did_some_progress, int *no_progress_loops)
4418 * Costly allocations might have made a progress but this doesn't mean
4419 * their order will become available due to high fragmentation so
4420 * always increment the no progress counter for them
4422 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4423 *no_progress_loops = 0;
4425 (*no_progress_loops)++;
4428 * Make sure we converge to OOM if we cannot make any progress
4429 * several times in the row.
4431 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4432 /* Before OOM, exhaust highatomic_reserve */
4433 return unreserve_highatomic_pageblock(ac, true);
4437 * Keep reclaiming pages while there is a chance this will lead
4438 * somewhere. If none of the target zones can satisfy our allocation
4439 * request even if all reclaimable pages are considered then we are
4440 * screwed and have to go OOM.
4442 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4443 ac->highest_zoneidx, ac->nodemask) {
4444 unsigned long available;
4445 unsigned long reclaimable;
4446 unsigned long min_wmark = min_wmark_pages(zone);
4449 available = reclaimable = zone_reclaimable_pages(zone);
4450 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4453 * Would the allocation succeed if we reclaimed all
4454 * reclaimable pages?
4456 wmark = __zone_watermark_ok(zone, order, min_wmark,
4457 ac->highest_zoneidx, alloc_flags, available);
4458 trace_reclaim_retry_zone(z, order, reclaimable,
4459 available, min_wmark, *no_progress_loops, wmark);
4462 * If we didn't make any progress and have a lot of
4463 * dirty + writeback pages then we should wait for
4464 * an IO to complete to slow down the reclaim and
4465 * prevent from pre mature OOM
4467 if (!did_some_progress) {
4468 unsigned long write_pending;
4470 write_pending = zone_page_state_snapshot(zone,
4471 NR_ZONE_WRITE_PENDING);
4473 if (2 * write_pending > reclaimable) {
4474 congestion_wait(BLK_RW_ASYNC, HZ/10);
4486 * Memory allocation/reclaim might be called from a WQ context and the
4487 * current implementation of the WQ concurrency control doesn't
4488 * recognize that a particular WQ is congested if the worker thread is
4489 * looping without ever sleeping. Therefore we have to do a short sleep
4490 * here rather than calling cond_resched().
4492 if (current->flags & PF_WQ_WORKER)
4493 schedule_timeout_uninterruptible(1);
4500 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4503 * It's possible that cpuset's mems_allowed and the nodemask from
4504 * mempolicy don't intersect. This should be normally dealt with by
4505 * policy_nodemask(), but it's possible to race with cpuset update in
4506 * such a way the check therein was true, and then it became false
4507 * before we got our cpuset_mems_cookie here.
4508 * This assumes that for all allocations, ac->nodemask can come only
4509 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4510 * when it does not intersect with the cpuset restrictions) or the
4511 * caller can deal with a violated nodemask.
4513 if (cpusets_enabled() && ac->nodemask &&
4514 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4515 ac->nodemask = NULL;
4520 * When updating a task's mems_allowed or mempolicy nodemask, it is
4521 * possible to race with parallel threads in such a way that our
4522 * allocation can fail while the mask is being updated. If we are about
4523 * to fail, check if the cpuset changed during allocation and if so,
4526 if (read_mems_allowed_retry(cpuset_mems_cookie))
4532 static inline struct page *
4533 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4534 struct alloc_context *ac)
4536 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4537 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4538 struct page *page = NULL;
4539 unsigned int alloc_flags;
4540 unsigned long did_some_progress;
4541 enum compact_priority compact_priority;
4542 enum compact_result compact_result;
4543 int compaction_retries;
4544 int no_progress_loops;
4545 unsigned int cpuset_mems_cookie;
4549 * We also sanity check to catch abuse of atomic reserves being used by
4550 * callers that are not in atomic context.
4552 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4553 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4554 gfp_mask &= ~__GFP_ATOMIC;
4557 compaction_retries = 0;
4558 no_progress_loops = 0;
4559 compact_priority = DEF_COMPACT_PRIORITY;
4560 cpuset_mems_cookie = read_mems_allowed_begin();
4563 * The fast path uses conservative alloc_flags to succeed only until
4564 * kswapd needs to be woken up, and to avoid the cost of setting up
4565 * alloc_flags precisely. So we do that now.
4567 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4570 * We need to recalculate the starting point for the zonelist iterator
4571 * because we might have used different nodemask in the fast path, or
4572 * there was a cpuset modification and we are retrying - otherwise we
4573 * could end up iterating over non-eligible zones endlessly.
4575 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4576 ac->highest_zoneidx, ac->nodemask);
4577 if (!ac->preferred_zoneref->zone)
4580 if (alloc_flags & ALLOC_KSWAPD)
4581 wake_all_kswapds(order, gfp_mask, ac);
4584 * The adjusted alloc_flags might result in immediate success, so try
4587 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4592 * For costly allocations, try direct compaction first, as it's likely
4593 * that we have enough base pages and don't need to reclaim. For non-
4594 * movable high-order allocations, do that as well, as compaction will
4595 * try prevent permanent fragmentation by migrating from blocks of the
4597 * Don't try this for allocations that are allowed to ignore
4598 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4600 if (can_direct_reclaim &&
4602 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4603 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4604 page = __alloc_pages_direct_compact(gfp_mask, order,
4606 INIT_COMPACT_PRIORITY,
4612 * Checks for costly allocations with __GFP_NORETRY, which
4613 * includes some THP page fault allocations
4615 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4617 * If allocating entire pageblock(s) and compaction
4618 * failed because all zones are below low watermarks
4619 * or is prohibited because it recently failed at this
4620 * order, fail immediately unless the allocator has
4621 * requested compaction and reclaim retry.
4624 * - potentially very expensive because zones are far
4625 * below their low watermarks or this is part of very
4626 * bursty high order allocations,
4627 * - not guaranteed to help because isolate_freepages()
4628 * may not iterate over freed pages as part of its
4630 * - unlikely to make entire pageblocks free on its
4633 if (compact_result == COMPACT_SKIPPED ||
4634 compact_result == COMPACT_DEFERRED)
4638 * Looks like reclaim/compaction is worth trying, but
4639 * sync compaction could be very expensive, so keep
4640 * using async compaction.
4642 compact_priority = INIT_COMPACT_PRIORITY;
4647 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4648 if (alloc_flags & ALLOC_KSWAPD)
4649 wake_all_kswapds(order, gfp_mask, ac);
4651 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4653 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4656 * Reset the nodemask and zonelist iterators if memory policies can be
4657 * ignored. These allocations are high priority and system rather than
4660 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4661 ac->nodemask = NULL;
4662 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4663 ac->highest_zoneidx, ac->nodemask);
4666 /* Attempt with potentially adjusted zonelist and alloc_flags */
4667 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4671 /* Caller is not willing to reclaim, we can't balance anything */
4672 if (!can_direct_reclaim)
4675 /* Avoid recursion of direct reclaim */
4676 if (current->flags & PF_MEMALLOC)
4679 /* Try direct reclaim and then allocating */
4680 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4681 &did_some_progress);
4685 /* Try direct compaction and then allocating */
4686 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4687 compact_priority, &compact_result);
4691 /* Do not loop if specifically requested */
4692 if (gfp_mask & __GFP_NORETRY)
4696 * Do not retry costly high order allocations unless they are
4697 * __GFP_RETRY_MAYFAIL
4699 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4702 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4703 did_some_progress > 0, &no_progress_loops))
4707 * It doesn't make any sense to retry for the compaction if the order-0
4708 * reclaim is not able to make any progress because the current
4709 * implementation of the compaction depends on the sufficient amount
4710 * of free memory (see __compaction_suitable)
4712 if (did_some_progress > 0 &&
4713 should_compact_retry(ac, order, alloc_flags,
4714 compact_result, &compact_priority,
4715 &compaction_retries))
4719 /* Deal with possible cpuset update races before we start OOM killing */
4720 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4723 /* Reclaim has failed us, start killing things */
4724 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4728 /* Avoid allocations with no watermarks from looping endlessly */
4729 if (tsk_is_oom_victim(current) &&
4730 (alloc_flags & ALLOC_OOM ||
4731 (gfp_mask & __GFP_NOMEMALLOC)))
4734 /* Retry as long as the OOM killer is making progress */
4735 if (did_some_progress) {
4736 no_progress_loops = 0;
4741 /* Deal with possible cpuset update races before we fail */
4742 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4746 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4749 if (gfp_mask & __GFP_NOFAIL) {
4751 * All existing users of the __GFP_NOFAIL are blockable, so warn
4752 * of any new users that actually require GFP_NOWAIT
4754 if (WARN_ON_ONCE(!can_direct_reclaim))
4758 * PF_MEMALLOC request from this context is rather bizarre
4759 * because we cannot reclaim anything and only can loop waiting
4760 * for somebody to do a work for us
4762 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4765 * non failing costly orders are a hard requirement which we
4766 * are not prepared for much so let's warn about these users
4767 * so that we can identify them and convert them to something
4770 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4773 * Help non-failing allocations by giving them access to memory
4774 * reserves but do not use ALLOC_NO_WATERMARKS because this
4775 * could deplete whole memory reserves which would just make
4776 * the situation worse
4778 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4786 warn_alloc(gfp_mask, ac->nodemask,
4787 "page allocation failure: order:%u", order);
4792 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4793 int preferred_nid, nodemask_t *nodemask,
4794 struct alloc_context *ac, gfp_t *alloc_mask,
4795 unsigned int *alloc_flags)
4797 ac->highest_zoneidx = gfp_zone(gfp_mask);
4798 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4799 ac->nodemask = nodemask;
4800 ac->migratetype = gfp_migratetype(gfp_mask);
4802 if (cpusets_enabled()) {
4803 *alloc_mask |= __GFP_HARDWALL;
4805 * When we are in the interrupt context, it is irrelevant
4806 * to the current task context. It means that any node ok.
4808 if (!in_interrupt() && !ac->nodemask)
4809 ac->nodemask = &cpuset_current_mems_allowed;
4811 *alloc_flags |= ALLOC_CPUSET;
4814 fs_reclaim_acquire(gfp_mask);
4815 fs_reclaim_release(gfp_mask);
4817 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4819 if (should_fail_alloc_page(gfp_mask, order))
4822 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4827 /* Determine whether to spread dirty pages and what the first usable zone */
4828 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4830 /* Dirty zone balancing only done in the fast path */
4831 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4834 * The preferred zone is used for statistics but crucially it is
4835 * also used as the starting point for the zonelist iterator. It
4836 * may get reset for allocations that ignore memory policies.
4838 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4839 ac->highest_zoneidx, ac->nodemask);
4843 * This is the 'heart' of the zoned buddy allocator.
4846 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4847 nodemask_t *nodemask)
4850 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4851 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4852 struct alloc_context ac = { };
4855 * There are several places where we assume that the order value is sane
4856 * so bail out early if the request is out of bound.
4858 if (unlikely(order >= MAX_ORDER)) {
4859 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4863 gfp_mask &= gfp_allowed_mask;
4864 alloc_mask = gfp_mask;
4865 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4868 finalise_ac(gfp_mask, &ac);
4871 * Forbid the first pass from falling back to types that fragment
4872 * memory until all local zones are considered.
4874 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4876 /* First allocation attempt */
4877 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4882 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4883 * resp. GFP_NOIO which has to be inherited for all allocation requests
4884 * from a particular context which has been marked by
4885 * memalloc_no{fs,io}_{save,restore}.
4887 alloc_mask = current_gfp_context(gfp_mask);
4888 ac.spread_dirty_pages = false;
4891 * Restore the original nodemask if it was potentially replaced with
4892 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4894 ac.nodemask = nodemask;
4896 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4899 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4900 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4901 __free_pages(page, order);
4905 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4909 EXPORT_SYMBOL(__alloc_pages_nodemask);
4912 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4913 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4914 * you need to access high mem.
4916 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4920 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4923 return (unsigned long) page_address(page);
4925 EXPORT_SYMBOL(__get_free_pages);
4927 unsigned long get_zeroed_page(gfp_t gfp_mask)
4929 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4931 EXPORT_SYMBOL(get_zeroed_page);
4933 static inline void free_the_page(struct page *page, unsigned int order)
4935 if (order == 0) /* Via pcp? */
4936 free_unref_page(page);
4938 __free_pages_ok(page, order);
4941 void __free_pages(struct page *page, unsigned int order)
4943 if (put_page_testzero(page))
4944 free_the_page(page, order);
4946 EXPORT_SYMBOL(__free_pages);
4948 void free_pages(unsigned long addr, unsigned int order)
4951 VM_BUG_ON(!virt_addr_valid((void *)addr));
4952 __free_pages(virt_to_page((void *)addr), order);
4956 EXPORT_SYMBOL(free_pages);
4960 * An arbitrary-length arbitrary-offset area of memory which resides
4961 * within a 0 or higher order page. Multiple fragments within that page
4962 * are individually refcounted, in the page's reference counter.
4964 * The page_frag functions below provide a simple allocation framework for
4965 * page fragments. This is used by the network stack and network device
4966 * drivers to provide a backing region of memory for use as either an
4967 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4969 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4972 struct page *page = NULL;
4973 gfp_t gfp = gfp_mask;
4975 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4976 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4978 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4979 PAGE_FRAG_CACHE_MAX_ORDER);
4980 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4982 if (unlikely(!page))
4983 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4985 nc->va = page ? page_address(page) : NULL;
4990 void __page_frag_cache_drain(struct page *page, unsigned int count)
4992 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4994 if (page_ref_sub_and_test(page, count))
4995 free_the_page(page, compound_order(page));
4997 EXPORT_SYMBOL(__page_frag_cache_drain);
4999 void *page_frag_alloc(struct page_frag_cache *nc,
5000 unsigned int fragsz, gfp_t gfp_mask)
5002 unsigned int size = PAGE_SIZE;
5006 if (unlikely(!nc->va)) {
5008 page = __page_frag_cache_refill(nc, gfp_mask);
5012 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5013 /* if size can vary use size else just use PAGE_SIZE */
5016 /* Even if we own the page, we do not use atomic_set().
5017 * This would break get_page_unless_zero() users.
5019 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5021 /* reset page count bias and offset to start of new frag */
5022 nc->pfmemalloc = page_is_pfmemalloc(page);
5023 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5027 offset = nc->offset - fragsz;
5028 if (unlikely(offset < 0)) {
5029 page = virt_to_page(nc->va);
5031 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5034 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5035 /* if size can vary use size else just use PAGE_SIZE */
5038 /* OK, page count is 0, we can safely set it */
5039 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5041 /* reset page count bias and offset to start of new frag */
5042 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5043 offset = size - fragsz;
5047 nc->offset = offset;
5049 return nc->va + offset;
5051 EXPORT_SYMBOL(page_frag_alloc);
5054 * Frees a page fragment allocated out of either a compound or order 0 page.
5056 void page_frag_free(void *addr)
5058 struct page *page = virt_to_head_page(addr);
5060 if (unlikely(put_page_testzero(page)))
5061 free_the_page(page, compound_order(page));
5063 EXPORT_SYMBOL(page_frag_free);
5065 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5069 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5070 unsigned long used = addr + PAGE_ALIGN(size);
5072 split_page(virt_to_page((void *)addr), order);
5073 while (used < alloc_end) {
5078 return (void *)addr;
5082 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5083 * @size: the number of bytes to allocate
5084 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5086 * This function is similar to alloc_pages(), except that it allocates the
5087 * minimum number of pages to satisfy the request. alloc_pages() can only
5088 * allocate memory in power-of-two pages.
5090 * This function is also limited by MAX_ORDER.
5092 * Memory allocated by this function must be released by free_pages_exact().
5094 * Return: pointer to the allocated area or %NULL in case of error.
5096 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5098 unsigned int order = get_order(size);
5101 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5102 gfp_mask &= ~__GFP_COMP;
5104 addr = __get_free_pages(gfp_mask, order);
5105 return make_alloc_exact(addr, order, size);
5107 EXPORT_SYMBOL(alloc_pages_exact);
5110 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5112 * @nid: the preferred node ID where memory should be allocated
5113 * @size: the number of bytes to allocate
5114 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5116 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5119 * Return: pointer to the allocated area or %NULL in case of error.
5121 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5123 unsigned int order = get_order(size);
5126 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5127 gfp_mask &= ~__GFP_COMP;
5129 p = alloc_pages_node(nid, gfp_mask, order);
5132 return make_alloc_exact((unsigned long)page_address(p), order, size);
5136 * free_pages_exact - release memory allocated via alloc_pages_exact()
5137 * @virt: the value returned by alloc_pages_exact.
5138 * @size: size of allocation, same value as passed to alloc_pages_exact().
5140 * Release the memory allocated by a previous call to alloc_pages_exact.
5142 void free_pages_exact(void *virt, size_t size)
5144 unsigned long addr = (unsigned long)virt;
5145 unsigned long end = addr + PAGE_ALIGN(size);
5147 while (addr < end) {
5152 EXPORT_SYMBOL(free_pages_exact);
5155 * nr_free_zone_pages - count number of pages beyond high watermark
5156 * @offset: The zone index of the highest zone
5158 * nr_free_zone_pages() counts the number of pages which are beyond the
5159 * high watermark within all zones at or below a given zone index. For each
5160 * zone, the number of pages is calculated as:
5162 * nr_free_zone_pages = managed_pages - high_pages
5164 * Return: number of pages beyond high watermark.
5166 static unsigned long nr_free_zone_pages(int offset)
5171 /* Just pick one node, since fallback list is circular */
5172 unsigned long sum = 0;
5174 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5176 for_each_zone_zonelist(zone, z, zonelist, offset) {
5177 unsigned long size = zone_managed_pages(zone);
5178 unsigned long high = high_wmark_pages(zone);
5187 * nr_free_buffer_pages - count number of pages beyond high watermark
5189 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5190 * watermark within ZONE_DMA and ZONE_NORMAL.
5192 * Return: number of pages beyond high watermark within ZONE_DMA and
5195 unsigned long nr_free_buffer_pages(void)
5197 return nr_free_zone_pages(gfp_zone(GFP_USER));
5199 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5201 static inline void show_node(struct zone *zone)
5203 if (IS_ENABLED(CONFIG_NUMA))
5204 printk("Node %d ", zone_to_nid(zone));
5207 long si_mem_available(void)
5210 unsigned long pagecache;
5211 unsigned long wmark_low = 0;
5212 unsigned long pages[NR_LRU_LISTS];
5213 unsigned long reclaimable;
5217 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5218 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5221 wmark_low += low_wmark_pages(zone);
5224 * Estimate the amount of memory available for userspace allocations,
5225 * without causing swapping.
5227 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5230 * Not all the page cache can be freed, otherwise the system will
5231 * start swapping. Assume at least half of the page cache, or the
5232 * low watermark worth of cache, needs to stay.
5234 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5235 pagecache -= min(pagecache / 2, wmark_low);
5236 available += pagecache;
5239 * Part of the reclaimable slab and other kernel memory consists of
5240 * items that are in use, and cannot be freed. Cap this estimate at the
5243 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5244 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5245 available += reclaimable - min(reclaimable / 2, wmark_low);
5251 EXPORT_SYMBOL_GPL(si_mem_available);
5253 void si_meminfo(struct sysinfo *val)
5255 val->totalram = totalram_pages();
5256 val->sharedram = global_node_page_state(NR_SHMEM);
5257 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5258 val->bufferram = nr_blockdev_pages();
5259 val->totalhigh = totalhigh_pages();
5260 val->freehigh = nr_free_highpages();
5261 val->mem_unit = PAGE_SIZE;
5264 EXPORT_SYMBOL(si_meminfo);
5267 void si_meminfo_node(struct sysinfo *val, int nid)
5269 int zone_type; /* needs to be signed */
5270 unsigned long managed_pages = 0;
5271 unsigned long managed_highpages = 0;
5272 unsigned long free_highpages = 0;
5273 pg_data_t *pgdat = NODE_DATA(nid);
5275 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5276 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5277 val->totalram = managed_pages;
5278 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5279 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5280 #ifdef CONFIG_HIGHMEM
5281 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5282 struct zone *zone = &pgdat->node_zones[zone_type];
5284 if (is_highmem(zone)) {
5285 managed_highpages += zone_managed_pages(zone);
5286 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5289 val->totalhigh = managed_highpages;
5290 val->freehigh = free_highpages;
5292 val->totalhigh = managed_highpages;
5293 val->freehigh = free_highpages;
5295 val->mem_unit = PAGE_SIZE;
5300 * Determine whether the node should be displayed or not, depending on whether
5301 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5303 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5305 if (!(flags & SHOW_MEM_FILTER_NODES))
5309 * no node mask - aka implicit memory numa policy. Do not bother with
5310 * the synchronization - read_mems_allowed_begin - because we do not
5311 * have to be precise here.
5314 nodemask = &cpuset_current_mems_allowed;
5316 return !node_isset(nid, *nodemask);
5319 #define K(x) ((x) << (PAGE_SHIFT-10))
5321 static void show_migration_types(unsigned char type)
5323 static const char types[MIGRATE_TYPES] = {
5324 [MIGRATE_UNMOVABLE] = 'U',
5325 [MIGRATE_MOVABLE] = 'M',
5326 [MIGRATE_RECLAIMABLE] = 'E',
5327 [MIGRATE_HIGHATOMIC] = 'H',
5329 [MIGRATE_CMA] = 'C',
5331 #ifdef CONFIG_MEMORY_ISOLATION
5332 [MIGRATE_ISOLATE] = 'I',
5335 char tmp[MIGRATE_TYPES + 1];
5339 for (i = 0; i < MIGRATE_TYPES; i++) {
5340 if (type & (1 << i))
5345 printk(KERN_CONT "(%s) ", tmp);
5349 * Show free area list (used inside shift_scroll-lock stuff)
5350 * We also calculate the percentage fragmentation. We do this by counting the
5351 * memory on each free list with the exception of the first item on the list.
5354 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5357 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5359 unsigned long free_pcp = 0;
5364 for_each_populated_zone(zone) {
5365 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5368 for_each_online_cpu(cpu)
5369 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5372 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5373 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5374 " unevictable:%lu dirty:%lu writeback:%lu\n"
5375 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5376 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5377 " free:%lu free_pcp:%lu free_cma:%lu\n",
5378 global_node_page_state(NR_ACTIVE_ANON),
5379 global_node_page_state(NR_INACTIVE_ANON),
5380 global_node_page_state(NR_ISOLATED_ANON),
5381 global_node_page_state(NR_ACTIVE_FILE),
5382 global_node_page_state(NR_INACTIVE_FILE),
5383 global_node_page_state(NR_ISOLATED_FILE),
5384 global_node_page_state(NR_UNEVICTABLE),
5385 global_node_page_state(NR_FILE_DIRTY),
5386 global_node_page_state(NR_WRITEBACK),
5387 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5388 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5389 global_node_page_state(NR_FILE_MAPPED),
5390 global_node_page_state(NR_SHMEM),
5391 global_zone_page_state(NR_PAGETABLE),
5392 global_zone_page_state(NR_BOUNCE),
5393 global_zone_page_state(NR_FREE_PAGES),
5395 global_zone_page_state(NR_FREE_CMA_PAGES));
5397 for_each_online_pgdat(pgdat) {
5398 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5402 " active_anon:%lukB"
5403 " inactive_anon:%lukB"
5404 " active_file:%lukB"
5405 " inactive_file:%lukB"
5406 " unevictable:%lukB"
5407 " isolated(anon):%lukB"
5408 " isolated(file):%lukB"
5413 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5415 " shmem_pmdmapped: %lukB"
5418 " writeback_tmp:%lukB"
5419 " kernel_stack:%lukB"
5420 #ifdef CONFIG_SHADOW_CALL_STACK
5421 " shadow_call_stack:%lukB"
5423 " all_unreclaimable? %s"
5426 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5427 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5428 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5429 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5430 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5431 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5432 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5433 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5434 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5435 K(node_page_state(pgdat, NR_WRITEBACK)),
5436 K(node_page_state(pgdat, NR_SHMEM)),
5437 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5438 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5439 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5441 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5443 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5444 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5445 #ifdef CONFIG_SHADOW_CALL_STACK
5446 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5448 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5452 for_each_populated_zone(zone) {
5455 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5459 for_each_online_cpu(cpu)
5460 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5469 " reserved_highatomic:%luKB"
5470 " active_anon:%lukB"
5471 " inactive_anon:%lukB"
5472 " active_file:%lukB"
5473 " inactive_file:%lukB"
5474 " unevictable:%lukB"
5475 " writepending:%lukB"
5486 K(zone_page_state(zone, NR_FREE_PAGES)),
5487 K(min_wmark_pages(zone)),
5488 K(low_wmark_pages(zone)),
5489 K(high_wmark_pages(zone)),
5490 K(zone->nr_reserved_highatomic),
5491 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5492 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5493 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5494 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5495 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5496 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5497 K(zone->present_pages),
5498 K(zone_managed_pages(zone)),
5499 K(zone_page_state(zone, NR_MLOCK)),
5500 K(zone_page_state(zone, NR_PAGETABLE)),
5501 K(zone_page_state(zone, NR_BOUNCE)),
5503 K(this_cpu_read(zone->pageset->pcp.count)),
5504 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5505 printk("lowmem_reserve[]:");
5506 for (i = 0; i < MAX_NR_ZONES; i++)
5507 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5508 printk(KERN_CONT "\n");
5511 for_each_populated_zone(zone) {
5513 unsigned long nr[MAX_ORDER], flags, total = 0;
5514 unsigned char types[MAX_ORDER];
5516 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5519 printk(KERN_CONT "%s: ", zone->name);
5521 spin_lock_irqsave(&zone->lock, flags);
5522 for (order = 0; order < MAX_ORDER; order++) {
5523 struct free_area *area = &zone->free_area[order];
5526 nr[order] = area->nr_free;
5527 total += nr[order] << order;
5530 for (type = 0; type < MIGRATE_TYPES; type++) {
5531 if (!free_area_empty(area, type))
5532 types[order] |= 1 << type;
5535 spin_unlock_irqrestore(&zone->lock, flags);
5536 for (order = 0; order < MAX_ORDER; order++) {
5537 printk(KERN_CONT "%lu*%lukB ",
5538 nr[order], K(1UL) << order);
5540 show_migration_types(types[order]);
5542 printk(KERN_CONT "= %lukB\n", K(total));
5545 hugetlb_show_meminfo();
5547 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5549 show_swap_cache_info();
5552 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5554 zoneref->zone = zone;
5555 zoneref->zone_idx = zone_idx(zone);
5559 * Builds allocation fallback zone lists.
5561 * Add all populated zones of a node to the zonelist.
5563 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5566 enum zone_type zone_type = MAX_NR_ZONES;
5571 zone = pgdat->node_zones + zone_type;
5572 if (managed_zone(zone)) {
5573 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5574 check_highest_zone(zone_type);
5576 } while (zone_type);
5583 static int __parse_numa_zonelist_order(char *s)
5586 * We used to support different zonlists modes but they turned
5587 * out to be just not useful. Let's keep the warning in place
5588 * if somebody still use the cmd line parameter so that we do
5589 * not fail it silently
5591 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5592 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5598 char numa_zonelist_order[] = "Node";
5601 * sysctl handler for numa_zonelist_order
5603 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5604 void *buffer, size_t *length, loff_t *ppos)
5607 return __parse_numa_zonelist_order(buffer);
5608 return proc_dostring(table, write, buffer, length, ppos);
5612 #define MAX_NODE_LOAD (nr_online_nodes)
5613 static int node_load[MAX_NUMNODES];
5616 * find_next_best_node - find the next node that should appear in a given node's fallback list
5617 * @node: node whose fallback list we're appending
5618 * @used_node_mask: nodemask_t of already used nodes
5620 * We use a number of factors to determine which is the next node that should
5621 * appear on a given node's fallback list. The node should not have appeared
5622 * already in @node's fallback list, and it should be the next closest node
5623 * according to the distance array (which contains arbitrary distance values
5624 * from each node to each node in the system), and should also prefer nodes
5625 * with no CPUs, since presumably they'll have very little allocation pressure
5626 * on them otherwise.
5628 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5630 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5633 int min_val = INT_MAX;
5634 int best_node = NUMA_NO_NODE;
5635 const struct cpumask *tmp = cpumask_of_node(0);
5637 /* Use the local node if we haven't already */
5638 if (!node_isset(node, *used_node_mask)) {
5639 node_set(node, *used_node_mask);
5643 for_each_node_state(n, N_MEMORY) {
5645 /* Don't want a node to appear more than once */
5646 if (node_isset(n, *used_node_mask))
5649 /* Use the distance array to find the distance */
5650 val = node_distance(node, n);
5652 /* Penalize nodes under us ("prefer the next node") */
5655 /* Give preference to headless and unused nodes */
5656 tmp = cpumask_of_node(n);
5657 if (!cpumask_empty(tmp))
5658 val += PENALTY_FOR_NODE_WITH_CPUS;
5660 /* Slight preference for less loaded node */
5661 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5662 val += node_load[n];
5664 if (val < min_val) {
5671 node_set(best_node, *used_node_mask);
5678 * Build zonelists ordered by node and zones within node.
5679 * This results in maximum locality--normal zone overflows into local
5680 * DMA zone, if any--but risks exhausting DMA zone.
5682 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5685 struct zoneref *zonerefs;
5688 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5690 for (i = 0; i < nr_nodes; i++) {
5693 pg_data_t *node = NODE_DATA(node_order[i]);
5695 nr_zones = build_zonerefs_node(node, zonerefs);
5696 zonerefs += nr_zones;
5698 zonerefs->zone = NULL;
5699 zonerefs->zone_idx = 0;
5703 * Build gfp_thisnode zonelists
5705 static void build_thisnode_zonelists(pg_data_t *pgdat)
5707 struct zoneref *zonerefs;
5710 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5711 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5712 zonerefs += nr_zones;
5713 zonerefs->zone = NULL;
5714 zonerefs->zone_idx = 0;
5718 * Build zonelists ordered by zone and nodes within zones.
5719 * This results in conserving DMA zone[s] until all Normal memory is
5720 * exhausted, but results in overflowing to remote node while memory
5721 * may still exist in local DMA zone.
5724 static void build_zonelists(pg_data_t *pgdat)
5726 static int node_order[MAX_NUMNODES];
5727 int node, load, nr_nodes = 0;
5728 nodemask_t used_mask = NODE_MASK_NONE;
5729 int local_node, prev_node;
5731 /* NUMA-aware ordering of nodes */
5732 local_node = pgdat->node_id;
5733 load = nr_online_nodes;
5734 prev_node = local_node;
5736 memset(node_order, 0, sizeof(node_order));
5737 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5739 * We don't want to pressure a particular node.
5740 * So adding penalty to the first node in same
5741 * distance group to make it round-robin.
5743 if (node_distance(local_node, node) !=
5744 node_distance(local_node, prev_node))
5745 node_load[node] = load;
5747 node_order[nr_nodes++] = node;
5752 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5753 build_thisnode_zonelists(pgdat);
5756 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5758 * Return node id of node used for "local" allocations.
5759 * I.e., first node id of first zone in arg node's generic zonelist.
5760 * Used for initializing percpu 'numa_mem', which is used primarily
5761 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5763 int local_memory_node(int node)
5767 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5768 gfp_zone(GFP_KERNEL),
5770 return zone_to_nid(z->zone);
5774 static void setup_min_unmapped_ratio(void);
5775 static void setup_min_slab_ratio(void);
5776 #else /* CONFIG_NUMA */
5778 static void build_zonelists(pg_data_t *pgdat)
5780 int node, local_node;
5781 struct zoneref *zonerefs;
5784 local_node = pgdat->node_id;
5786 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5787 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5788 zonerefs += nr_zones;
5791 * Now we build the zonelist so that it contains the zones
5792 * of all the other nodes.
5793 * We don't want to pressure a particular node, so when
5794 * building the zones for node N, we make sure that the
5795 * zones coming right after the local ones are those from
5796 * node N+1 (modulo N)
5798 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5799 if (!node_online(node))
5801 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5802 zonerefs += nr_zones;
5804 for (node = 0; node < local_node; node++) {
5805 if (!node_online(node))
5807 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5808 zonerefs += nr_zones;
5811 zonerefs->zone = NULL;
5812 zonerefs->zone_idx = 0;
5815 #endif /* CONFIG_NUMA */
5818 * Boot pageset table. One per cpu which is going to be used for all
5819 * zones and all nodes. The parameters will be set in such a way
5820 * that an item put on a list will immediately be handed over to
5821 * the buddy list. This is safe since pageset manipulation is done
5822 * with interrupts disabled.
5824 * The boot_pagesets must be kept even after bootup is complete for
5825 * unused processors and/or zones. They do play a role for bootstrapping
5826 * hotplugged processors.
5828 * zoneinfo_show() and maybe other functions do
5829 * not check if the processor is online before following the pageset pointer.
5830 * Other parts of the kernel may not check if the zone is available.
5832 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5833 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5834 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5836 static void __build_all_zonelists(void *data)
5839 int __maybe_unused cpu;
5840 pg_data_t *self = data;
5841 static DEFINE_SPINLOCK(lock);
5846 memset(node_load, 0, sizeof(node_load));
5850 * This node is hotadded and no memory is yet present. So just
5851 * building zonelists is fine - no need to touch other nodes.
5853 if (self && !node_online(self->node_id)) {
5854 build_zonelists(self);
5856 for_each_online_node(nid) {
5857 pg_data_t *pgdat = NODE_DATA(nid);
5859 build_zonelists(pgdat);
5862 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5864 * We now know the "local memory node" for each node--
5865 * i.e., the node of the first zone in the generic zonelist.
5866 * Set up numa_mem percpu variable for on-line cpus. During
5867 * boot, only the boot cpu should be on-line; we'll init the
5868 * secondary cpus' numa_mem as they come on-line. During
5869 * node/memory hotplug, we'll fixup all on-line cpus.
5871 for_each_online_cpu(cpu)
5872 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5879 static noinline void __init
5880 build_all_zonelists_init(void)
5884 __build_all_zonelists(NULL);
5887 * Initialize the boot_pagesets that are going to be used
5888 * for bootstrapping processors. The real pagesets for
5889 * each zone will be allocated later when the per cpu
5890 * allocator is available.
5892 * boot_pagesets are used also for bootstrapping offline
5893 * cpus if the system is already booted because the pagesets
5894 * are needed to initialize allocators on a specific cpu too.
5895 * F.e. the percpu allocator needs the page allocator which
5896 * needs the percpu allocator in order to allocate its pagesets
5897 * (a chicken-egg dilemma).
5899 for_each_possible_cpu(cpu)
5900 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5902 mminit_verify_zonelist();
5903 cpuset_init_current_mems_allowed();
5907 * unless system_state == SYSTEM_BOOTING.
5909 * __ref due to call of __init annotated helper build_all_zonelists_init
5910 * [protected by SYSTEM_BOOTING].
5912 void __ref build_all_zonelists(pg_data_t *pgdat)
5914 unsigned long vm_total_pages;
5916 if (system_state == SYSTEM_BOOTING) {
5917 build_all_zonelists_init();
5919 __build_all_zonelists(pgdat);
5920 /* cpuset refresh routine should be here */
5922 /* Get the number of free pages beyond high watermark in all zones. */
5923 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5925 * Disable grouping by mobility if the number of pages in the
5926 * system is too low to allow the mechanism to work. It would be
5927 * more accurate, but expensive to check per-zone. This check is
5928 * made on memory-hotadd so a system can start with mobility
5929 * disabled and enable it later
5931 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5932 page_group_by_mobility_disabled = 1;
5934 page_group_by_mobility_disabled = 0;
5936 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5938 page_group_by_mobility_disabled ? "off" : "on",
5941 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5945 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5946 static bool __meminit
5947 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5949 static struct memblock_region *r;
5951 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5952 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5953 for_each_memblock(memory, r) {
5954 if (*pfn < memblock_region_memory_end_pfn(r))
5958 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5959 memblock_is_mirror(r)) {
5960 *pfn = memblock_region_memory_end_pfn(r);
5968 * Initially all pages are reserved - free ones are freed
5969 * up by memblock_free_all() once the early boot process is
5970 * done. Non-atomic initialization, single-pass.
5972 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5973 unsigned long start_pfn, enum memmap_context context,
5974 struct vmem_altmap *altmap)
5976 unsigned long pfn, end_pfn = start_pfn + size;
5979 if (highest_memmap_pfn < end_pfn - 1)
5980 highest_memmap_pfn = end_pfn - 1;
5982 #ifdef CONFIG_ZONE_DEVICE
5984 * Honor reservation requested by the driver for this ZONE_DEVICE
5985 * memory. We limit the total number of pages to initialize to just
5986 * those that might contain the memory mapping. We will defer the
5987 * ZONE_DEVICE page initialization until after we have released
5990 if (zone == ZONE_DEVICE) {
5994 if (start_pfn == altmap->base_pfn)
5995 start_pfn += altmap->reserve;
5996 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6000 for (pfn = start_pfn; pfn < end_pfn; ) {
6002 * There can be holes in boot-time mem_map[]s handed to this
6003 * function. They do not exist on hotplugged memory.
6005 if (context == MEMMAP_EARLY) {
6006 if (overlap_memmap_init(zone, &pfn))
6008 if (defer_init(nid, pfn, end_pfn))
6012 page = pfn_to_page(pfn);
6013 __init_single_page(page, pfn, zone, nid);
6014 if (context == MEMMAP_HOTPLUG)
6015 __SetPageReserved(page);
6018 * Mark the block movable so that blocks are reserved for
6019 * movable at startup. This will force kernel allocations
6020 * to reserve their blocks rather than leaking throughout
6021 * the address space during boot when many long-lived
6022 * kernel allocations are made.
6024 * bitmap is created for zone's valid pfn range. but memmap
6025 * can be created for invalid pages (for alignment)
6026 * check here not to call set_pageblock_migratetype() against
6029 if (!(pfn & (pageblock_nr_pages - 1))) {
6030 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6037 #ifdef CONFIG_ZONE_DEVICE
6038 void __ref memmap_init_zone_device(struct zone *zone,
6039 unsigned long start_pfn,
6040 unsigned long nr_pages,
6041 struct dev_pagemap *pgmap)
6043 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6044 struct pglist_data *pgdat = zone->zone_pgdat;
6045 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6046 unsigned long zone_idx = zone_idx(zone);
6047 unsigned long start = jiffies;
6048 int nid = pgdat->node_id;
6050 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6054 * The call to memmap_init_zone should have already taken care
6055 * of the pages reserved for the memmap, so we can just jump to
6056 * the end of that region and start processing the device pages.
6059 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6060 nr_pages = end_pfn - start_pfn;
6063 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6064 struct page *page = pfn_to_page(pfn);
6066 __init_single_page(page, pfn, zone_idx, nid);
6069 * Mark page reserved as it will need to wait for onlining
6070 * phase for it to be fully associated with a zone.
6072 * We can use the non-atomic __set_bit operation for setting
6073 * the flag as we are still initializing the pages.
6075 __SetPageReserved(page);
6078 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6079 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6080 * ever freed or placed on a driver-private list.
6082 page->pgmap = pgmap;
6083 page->zone_device_data = NULL;
6086 * Mark the block movable so that blocks are reserved for
6087 * movable at startup. This will force kernel allocations
6088 * to reserve their blocks rather than leaking throughout
6089 * the address space during boot when many long-lived
6090 * kernel allocations are made.
6092 * bitmap is created for zone's valid pfn range. but memmap
6093 * can be created for invalid pages (for alignment)
6094 * check here not to call set_pageblock_migratetype() against
6097 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6098 * because this is done early in section_activate()
6100 if (!(pfn & (pageblock_nr_pages - 1))) {
6101 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6106 pr_info("%s initialised %lu pages in %ums\n", __func__,
6107 nr_pages, jiffies_to_msecs(jiffies - start));
6111 static void __meminit zone_init_free_lists(struct zone *zone)
6113 unsigned int order, t;
6114 for_each_migratetype_order(order, t) {
6115 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6116 zone->free_area[order].nr_free = 0;
6120 void __meminit __weak memmap_init(unsigned long size, int nid,
6122 unsigned long range_start_pfn)
6124 unsigned long start_pfn, end_pfn;
6125 unsigned long range_end_pfn = range_start_pfn + size;
6128 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6129 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6130 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6132 if (end_pfn > start_pfn) {
6133 size = end_pfn - start_pfn;
6134 memmap_init_zone(size, nid, zone, start_pfn,
6135 MEMMAP_EARLY, NULL);
6140 static int zone_batchsize(struct zone *zone)
6146 * The per-cpu-pages pools are set to around 1000th of the
6149 batch = zone_managed_pages(zone) / 1024;
6150 /* But no more than a meg. */
6151 if (batch * PAGE_SIZE > 1024 * 1024)
6152 batch = (1024 * 1024) / PAGE_SIZE;
6153 batch /= 4; /* We effectively *= 4 below */
6158 * Clamp the batch to a 2^n - 1 value. Having a power
6159 * of 2 value was found to be more likely to have
6160 * suboptimal cache aliasing properties in some cases.
6162 * For example if 2 tasks are alternately allocating
6163 * batches of pages, one task can end up with a lot
6164 * of pages of one half of the possible page colors
6165 * and the other with pages of the other colors.
6167 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6172 /* The deferral and batching of frees should be suppressed under NOMMU
6175 * The problem is that NOMMU needs to be able to allocate large chunks
6176 * of contiguous memory as there's no hardware page translation to
6177 * assemble apparent contiguous memory from discontiguous pages.
6179 * Queueing large contiguous runs of pages for batching, however,
6180 * causes the pages to actually be freed in smaller chunks. As there
6181 * can be a significant delay between the individual batches being
6182 * recycled, this leads to the once large chunks of space being
6183 * fragmented and becoming unavailable for high-order allocations.
6190 * pcp->high and pcp->batch values are related and dependent on one another:
6191 * ->batch must never be higher then ->high.
6192 * The following function updates them in a safe manner without read side
6195 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6196 * those fields changing asynchronously (acording to the above rule).
6198 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6199 * outside of boot time (or some other assurance that no concurrent updaters
6202 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6203 unsigned long batch)
6205 /* start with a fail safe value for batch */
6209 /* Update high, then batch, in order */
6216 /* a companion to pageset_set_high() */
6217 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6219 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6222 static void pageset_init(struct per_cpu_pageset *p)
6224 struct per_cpu_pages *pcp;
6227 memset(p, 0, sizeof(*p));
6230 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6231 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6234 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6237 pageset_set_batch(p, batch);
6241 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6242 * to the value high for the pageset p.
6244 static void pageset_set_high(struct per_cpu_pageset *p,
6247 unsigned long batch = max(1UL, high / 4);
6248 if ((high / 4) > (PAGE_SHIFT * 8))
6249 batch = PAGE_SHIFT * 8;
6251 pageset_update(&p->pcp, high, batch);
6254 static void pageset_set_high_and_batch(struct zone *zone,
6255 struct per_cpu_pageset *pcp)
6257 if (percpu_pagelist_fraction)
6258 pageset_set_high(pcp,
6259 (zone_managed_pages(zone) /
6260 percpu_pagelist_fraction));
6262 pageset_set_batch(pcp, zone_batchsize(zone));
6265 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6267 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6270 pageset_set_high_and_batch(zone, pcp);
6273 void __meminit setup_zone_pageset(struct zone *zone)
6276 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6277 for_each_possible_cpu(cpu)
6278 zone_pageset_init(zone, cpu);
6282 * Allocate per cpu pagesets and initialize them.
6283 * Before this call only boot pagesets were available.
6285 void __init setup_per_cpu_pageset(void)
6287 struct pglist_data *pgdat;
6289 int __maybe_unused cpu;
6291 for_each_populated_zone(zone)
6292 setup_zone_pageset(zone);
6296 * Unpopulated zones continue using the boot pagesets.
6297 * The numa stats for these pagesets need to be reset.
6298 * Otherwise, they will end up skewing the stats of
6299 * the nodes these zones are associated with.
6301 for_each_possible_cpu(cpu) {
6302 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6303 memset(pcp->vm_numa_stat_diff, 0,
6304 sizeof(pcp->vm_numa_stat_diff));
6308 for_each_online_pgdat(pgdat)
6309 pgdat->per_cpu_nodestats =
6310 alloc_percpu(struct per_cpu_nodestat);
6313 static __meminit void zone_pcp_init(struct zone *zone)
6316 * per cpu subsystem is not up at this point. The following code
6317 * relies on the ability of the linker to provide the
6318 * offset of a (static) per cpu variable into the per cpu area.
6320 zone->pageset = &boot_pageset;
6322 if (populated_zone(zone))
6323 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6324 zone->name, zone->present_pages,
6325 zone_batchsize(zone));
6328 void __meminit init_currently_empty_zone(struct zone *zone,
6329 unsigned long zone_start_pfn,
6332 struct pglist_data *pgdat = zone->zone_pgdat;
6333 int zone_idx = zone_idx(zone) + 1;
6335 if (zone_idx > pgdat->nr_zones)
6336 pgdat->nr_zones = zone_idx;
6338 zone->zone_start_pfn = zone_start_pfn;
6340 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6341 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6343 (unsigned long)zone_idx(zone),
6344 zone_start_pfn, (zone_start_pfn + size));
6346 zone_init_free_lists(zone);
6347 zone->initialized = 1;
6351 * get_pfn_range_for_nid - Return the start and end page frames for a node
6352 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6353 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6354 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6356 * It returns the start and end page frame of a node based on information
6357 * provided by memblock_set_node(). If called for a node
6358 * with no available memory, a warning is printed and the start and end
6361 void __init get_pfn_range_for_nid(unsigned int nid,
6362 unsigned long *start_pfn, unsigned long *end_pfn)
6364 unsigned long this_start_pfn, this_end_pfn;
6370 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6371 *start_pfn = min(*start_pfn, this_start_pfn);
6372 *end_pfn = max(*end_pfn, this_end_pfn);
6375 if (*start_pfn == -1UL)
6380 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6381 * assumption is made that zones within a node are ordered in monotonic
6382 * increasing memory addresses so that the "highest" populated zone is used
6384 static void __init find_usable_zone_for_movable(void)
6387 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6388 if (zone_index == ZONE_MOVABLE)
6391 if (arch_zone_highest_possible_pfn[zone_index] >
6392 arch_zone_lowest_possible_pfn[zone_index])
6396 VM_BUG_ON(zone_index == -1);
6397 movable_zone = zone_index;
6401 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6402 * because it is sized independent of architecture. Unlike the other zones,
6403 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6404 * in each node depending on the size of each node and how evenly kernelcore
6405 * is distributed. This helper function adjusts the zone ranges
6406 * provided by the architecture for a given node by using the end of the
6407 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6408 * zones within a node are in order of monotonic increases memory addresses
6410 static void __init adjust_zone_range_for_zone_movable(int nid,
6411 unsigned long zone_type,
6412 unsigned long node_start_pfn,
6413 unsigned long node_end_pfn,
6414 unsigned long *zone_start_pfn,
6415 unsigned long *zone_end_pfn)
6417 /* Only adjust if ZONE_MOVABLE is on this node */
6418 if (zone_movable_pfn[nid]) {
6419 /* Size ZONE_MOVABLE */
6420 if (zone_type == ZONE_MOVABLE) {
6421 *zone_start_pfn = zone_movable_pfn[nid];
6422 *zone_end_pfn = min(node_end_pfn,
6423 arch_zone_highest_possible_pfn[movable_zone]);
6425 /* Adjust for ZONE_MOVABLE starting within this range */
6426 } else if (!mirrored_kernelcore &&
6427 *zone_start_pfn < zone_movable_pfn[nid] &&
6428 *zone_end_pfn > zone_movable_pfn[nid]) {
6429 *zone_end_pfn = zone_movable_pfn[nid];
6431 /* Check if this whole range is within ZONE_MOVABLE */
6432 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6433 *zone_start_pfn = *zone_end_pfn;
6438 * Return the number of pages a zone spans in a node, including holes
6439 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6441 static unsigned long __init zone_spanned_pages_in_node(int nid,
6442 unsigned long zone_type,
6443 unsigned long node_start_pfn,
6444 unsigned long node_end_pfn,
6445 unsigned long *zone_start_pfn,
6446 unsigned long *zone_end_pfn)
6448 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6449 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6450 /* When hotadd a new node from cpu_up(), the node should be empty */
6451 if (!node_start_pfn && !node_end_pfn)
6454 /* Get the start and end of the zone */
6455 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6456 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6457 adjust_zone_range_for_zone_movable(nid, zone_type,
6458 node_start_pfn, node_end_pfn,
6459 zone_start_pfn, zone_end_pfn);
6461 /* Check that this node has pages within the zone's required range */
6462 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6465 /* Move the zone boundaries inside the node if necessary */
6466 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6467 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6469 /* Return the spanned pages */
6470 return *zone_end_pfn - *zone_start_pfn;
6474 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6475 * then all holes in the requested range will be accounted for.
6477 unsigned long __init __absent_pages_in_range(int nid,
6478 unsigned long range_start_pfn,
6479 unsigned long range_end_pfn)
6481 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6482 unsigned long start_pfn, end_pfn;
6485 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6486 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6487 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6488 nr_absent -= end_pfn - start_pfn;
6494 * absent_pages_in_range - Return number of page frames in holes within a range
6495 * @start_pfn: The start PFN to start searching for holes
6496 * @end_pfn: The end PFN to stop searching for holes
6498 * Return: the number of pages frames in memory holes within a range.
6500 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6501 unsigned long end_pfn)
6503 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6506 /* Return the number of page frames in holes in a zone on a node */
6507 static unsigned long __init zone_absent_pages_in_node(int nid,
6508 unsigned long zone_type,
6509 unsigned long node_start_pfn,
6510 unsigned long node_end_pfn)
6512 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6513 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6514 unsigned long zone_start_pfn, zone_end_pfn;
6515 unsigned long nr_absent;
6517 /* When hotadd a new node from cpu_up(), the node should be empty */
6518 if (!node_start_pfn && !node_end_pfn)
6521 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6522 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6524 adjust_zone_range_for_zone_movable(nid, zone_type,
6525 node_start_pfn, node_end_pfn,
6526 &zone_start_pfn, &zone_end_pfn);
6527 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6530 * ZONE_MOVABLE handling.
6531 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6534 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6535 unsigned long start_pfn, end_pfn;
6536 struct memblock_region *r;
6538 for_each_memblock(memory, r) {
6539 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6540 zone_start_pfn, zone_end_pfn);
6541 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6542 zone_start_pfn, zone_end_pfn);
6544 if (zone_type == ZONE_MOVABLE &&
6545 memblock_is_mirror(r))
6546 nr_absent += end_pfn - start_pfn;
6548 if (zone_type == ZONE_NORMAL &&
6549 !memblock_is_mirror(r))
6550 nr_absent += end_pfn - start_pfn;
6557 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6558 unsigned long node_start_pfn,
6559 unsigned long node_end_pfn)
6561 unsigned long realtotalpages = 0, totalpages = 0;
6564 for (i = 0; i < MAX_NR_ZONES; i++) {
6565 struct zone *zone = pgdat->node_zones + i;
6566 unsigned long zone_start_pfn, zone_end_pfn;
6567 unsigned long spanned, absent;
6568 unsigned long size, real_size;
6570 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6575 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6580 real_size = size - absent;
6583 zone->zone_start_pfn = zone_start_pfn;
6585 zone->zone_start_pfn = 0;
6586 zone->spanned_pages = size;
6587 zone->present_pages = real_size;
6590 realtotalpages += real_size;
6593 pgdat->node_spanned_pages = totalpages;
6594 pgdat->node_present_pages = realtotalpages;
6595 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6599 #ifndef CONFIG_SPARSEMEM
6601 * Calculate the size of the zone->blockflags rounded to an unsigned long
6602 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6603 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6604 * round what is now in bits to nearest long in bits, then return it in
6607 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6609 unsigned long usemapsize;
6611 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6612 usemapsize = roundup(zonesize, pageblock_nr_pages);
6613 usemapsize = usemapsize >> pageblock_order;
6614 usemapsize *= NR_PAGEBLOCK_BITS;
6615 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6617 return usemapsize / 8;
6620 static void __ref setup_usemap(struct pglist_data *pgdat,
6622 unsigned long zone_start_pfn,
6623 unsigned long zonesize)
6625 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6626 zone->pageblock_flags = NULL;
6628 zone->pageblock_flags =
6629 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6631 if (!zone->pageblock_flags)
6632 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6633 usemapsize, zone->name, pgdat->node_id);
6637 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6638 unsigned long zone_start_pfn, unsigned long zonesize) {}
6639 #endif /* CONFIG_SPARSEMEM */
6641 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6643 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6644 void __init set_pageblock_order(void)
6648 /* Check that pageblock_nr_pages has not already been setup */
6649 if (pageblock_order)
6652 if (HPAGE_SHIFT > PAGE_SHIFT)
6653 order = HUGETLB_PAGE_ORDER;
6655 order = MAX_ORDER - 1;
6658 * Assume the largest contiguous order of interest is a huge page.
6659 * This value may be variable depending on boot parameters on IA64 and
6662 pageblock_order = order;
6664 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6667 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6668 * is unused as pageblock_order is set at compile-time. See
6669 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6672 void __init set_pageblock_order(void)
6676 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6678 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6679 unsigned long present_pages)
6681 unsigned long pages = spanned_pages;
6684 * Provide a more accurate estimation if there are holes within
6685 * the zone and SPARSEMEM is in use. If there are holes within the
6686 * zone, each populated memory region may cost us one or two extra
6687 * memmap pages due to alignment because memmap pages for each
6688 * populated regions may not be naturally aligned on page boundary.
6689 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6691 if (spanned_pages > present_pages + (present_pages >> 4) &&
6692 IS_ENABLED(CONFIG_SPARSEMEM))
6693 pages = present_pages;
6695 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6698 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6699 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6701 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6703 spin_lock_init(&ds_queue->split_queue_lock);
6704 INIT_LIST_HEAD(&ds_queue->split_queue);
6705 ds_queue->split_queue_len = 0;
6708 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6711 #ifdef CONFIG_COMPACTION
6712 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6714 init_waitqueue_head(&pgdat->kcompactd_wait);
6717 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6720 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6722 pgdat_resize_init(pgdat);
6724 pgdat_init_split_queue(pgdat);
6725 pgdat_init_kcompactd(pgdat);
6727 init_waitqueue_head(&pgdat->kswapd_wait);
6728 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6730 pgdat_page_ext_init(pgdat);
6731 spin_lock_init(&pgdat->lru_lock);
6732 lruvec_init(&pgdat->__lruvec);
6735 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6736 unsigned long remaining_pages)
6738 atomic_long_set(&zone->managed_pages, remaining_pages);
6739 zone_set_nid(zone, nid);
6740 zone->name = zone_names[idx];
6741 zone->zone_pgdat = NODE_DATA(nid);
6742 spin_lock_init(&zone->lock);
6743 zone_seqlock_init(zone);
6744 zone_pcp_init(zone);
6748 * Set up the zone data structures
6749 * - init pgdat internals
6750 * - init all zones belonging to this node
6752 * NOTE: this function is only called during memory hotplug
6754 #ifdef CONFIG_MEMORY_HOTPLUG
6755 void __ref free_area_init_core_hotplug(int nid)
6758 pg_data_t *pgdat = NODE_DATA(nid);
6760 pgdat_init_internals(pgdat);
6761 for (z = 0; z < MAX_NR_ZONES; z++)
6762 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6767 * Set up the zone data structures:
6768 * - mark all pages reserved
6769 * - mark all memory queues empty
6770 * - clear the memory bitmaps
6772 * NOTE: pgdat should get zeroed by caller.
6773 * NOTE: this function is only called during early init.
6775 static void __init free_area_init_core(struct pglist_data *pgdat)
6778 int nid = pgdat->node_id;
6780 pgdat_init_internals(pgdat);
6781 pgdat->per_cpu_nodestats = &boot_nodestats;
6783 for (j = 0; j < MAX_NR_ZONES; j++) {
6784 struct zone *zone = pgdat->node_zones + j;
6785 unsigned long size, freesize, memmap_pages;
6786 unsigned long zone_start_pfn = zone->zone_start_pfn;
6788 size = zone->spanned_pages;
6789 freesize = zone->present_pages;
6792 * Adjust freesize so that it accounts for how much memory
6793 * is used by this zone for memmap. This affects the watermark
6794 * and per-cpu initialisations
6796 memmap_pages = calc_memmap_size(size, freesize);
6797 if (!is_highmem_idx(j)) {
6798 if (freesize >= memmap_pages) {
6799 freesize -= memmap_pages;
6802 " %s zone: %lu pages used for memmap\n",
6803 zone_names[j], memmap_pages);
6805 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6806 zone_names[j], memmap_pages, freesize);
6809 /* Account for reserved pages */
6810 if (j == 0 && freesize > dma_reserve) {
6811 freesize -= dma_reserve;
6812 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6813 zone_names[0], dma_reserve);
6816 if (!is_highmem_idx(j))
6817 nr_kernel_pages += freesize;
6818 /* Charge for highmem memmap if there are enough kernel pages */
6819 else if (nr_kernel_pages > memmap_pages * 2)
6820 nr_kernel_pages -= memmap_pages;
6821 nr_all_pages += freesize;
6824 * Set an approximate value for lowmem here, it will be adjusted
6825 * when the bootmem allocator frees pages into the buddy system.
6826 * And all highmem pages will be managed by the buddy system.
6828 zone_init_internals(zone, j, nid, freesize);
6833 set_pageblock_order();
6834 setup_usemap(pgdat, zone, zone_start_pfn, size);
6835 init_currently_empty_zone(zone, zone_start_pfn, size);
6836 memmap_init(size, nid, j, zone_start_pfn);
6840 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6841 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6843 unsigned long __maybe_unused start = 0;
6844 unsigned long __maybe_unused offset = 0;
6846 /* Skip empty nodes */
6847 if (!pgdat->node_spanned_pages)
6850 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6851 offset = pgdat->node_start_pfn - start;
6852 /* ia64 gets its own node_mem_map, before this, without bootmem */
6853 if (!pgdat->node_mem_map) {
6854 unsigned long size, end;
6858 * The zone's endpoints aren't required to be MAX_ORDER
6859 * aligned but the node_mem_map endpoints must be in order
6860 * for the buddy allocator to function correctly.
6862 end = pgdat_end_pfn(pgdat);
6863 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6864 size = (end - start) * sizeof(struct page);
6865 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6868 panic("Failed to allocate %ld bytes for node %d memory map\n",
6869 size, pgdat->node_id);
6870 pgdat->node_mem_map = map + offset;
6872 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6873 __func__, pgdat->node_id, (unsigned long)pgdat,
6874 (unsigned long)pgdat->node_mem_map);
6875 #ifndef CONFIG_NEED_MULTIPLE_NODES
6877 * With no DISCONTIG, the global mem_map is just set as node 0's
6879 if (pgdat == NODE_DATA(0)) {
6880 mem_map = NODE_DATA(0)->node_mem_map;
6881 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6887 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6888 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6890 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6891 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6893 pgdat->first_deferred_pfn = ULONG_MAX;
6896 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6899 static void __init free_area_init_node(int nid)
6901 pg_data_t *pgdat = NODE_DATA(nid);
6902 unsigned long start_pfn = 0;
6903 unsigned long end_pfn = 0;
6905 /* pg_data_t should be reset to zero when it's allocated */
6906 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6908 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6910 pgdat->node_id = nid;
6911 pgdat->node_start_pfn = start_pfn;
6912 pgdat->per_cpu_nodestats = NULL;
6914 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6915 (u64)start_pfn << PAGE_SHIFT,
6916 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6917 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6919 alloc_node_mem_map(pgdat);
6920 pgdat_set_deferred_range(pgdat);
6922 free_area_init_core(pgdat);
6925 void __init free_area_init_memoryless_node(int nid)
6927 free_area_init_node(nid);
6930 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6932 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6933 * PageReserved(). Return the number of struct pages that were initialized.
6935 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6940 for (pfn = spfn; pfn < epfn; pfn++) {
6941 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6942 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6943 + pageblock_nr_pages - 1;
6947 * Use a fake node/zone (0) for now. Some of these pages
6948 * (in memblock.reserved but not in memblock.memory) will
6949 * get re-initialized via reserve_bootmem_region() later.
6951 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6952 __SetPageReserved(pfn_to_page(pfn));
6960 * Only struct pages that are backed by physical memory are zeroed and
6961 * initialized by going through __init_single_page(). But, there are some
6962 * struct pages which are reserved in memblock allocator and their fields
6963 * may be accessed (for example page_to_pfn() on some configuration accesses
6964 * flags). We must explicitly initialize those struct pages.
6966 * This function also addresses a similar issue where struct pages are left
6967 * uninitialized because the physical address range is not covered by
6968 * memblock.memory or memblock.reserved. That could happen when memblock
6969 * layout is manually configured via memmap=, or when the highest physical
6970 * address (max_pfn) does not end on a section boundary.
6972 static void __init init_unavailable_mem(void)
6974 phys_addr_t start, end;
6976 phys_addr_t next = 0;
6979 * Loop through unavailable ranges not covered by memblock.memory.
6982 for_each_mem_range(i, &memblock.memory, NULL,
6983 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6985 pgcnt += init_unavailable_range(PFN_DOWN(next),
6991 * Early sections always have a fully populated memmap for the whole
6992 * section - see pfn_valid(). If the last section has holes at the
6993 * end and that section is marked "online", the memmap will be
6994 * considered initialized. Make sure that memmap has a well defined
6997 pgcnt += init_unavailable_range(PFN_DOWN(next),
6998 round_up(max_pfn, PAGES_PER_SECTION));
7001 * Struct pages that do not have backing memory. This could be because
7002 * firmware is using some of this memory, or for some other reasons.
7005 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7008 static inline void __init init_unavailable_mem(void)
7011 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7013 #if MAX_NUMNODES > 1
7015 * Figure out the number of possible node ids.
7017 void __init setup_nr_node_ids(void)
7019 unsigned int highest;
7021 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7022 nr_node_ids = highest + 1;
7027 * node_map_pfn_alignment - determine the maximum internode alignment
7029 * This function should be called after node map is populated and sorted.
7030 * It calculates the maximum power of two alignment which can distinguish
7033 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7034 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7035 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7036 * shifted, 1GiB is enough and this function will indicate so.
7038 * This is used to test whether pfn -> nid mapping of the chosen memory
7039 * model has fine enough granularity to avoid incorrect mapping for the
7040 * populated node map.
7042 * Return: the determined alignment in pfn's. 0 if there is no alignment
7043 * requirement (single node).
7045 unsigned long __init node_map_pfn_alignment(void)
7047 unsigned long accl_mask = 0, last_end = 0;
7048 unsigned long start, end, mask;
7049 int last_nid = NUMA_NO_NODE;
7052 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7053 if (!start || last_nid < 0 || last_nid == nid) {
7060 * Start with a mask granular enough to pin-point to the
7061 * start pfn and tick off bits one-by-one until it becomes
7062 * too coarse to separate the current node from the last.
7064 mask = ~((1 << __ffs(start)) - 1);
7065 while (mask && last_end <= (start & (mask << 1)))
7068 /* accumulate all internode masks */
7072 /* convert mask to number of pages */
7073 return ~accl_mask + 1;
7077 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7079 * Return: the minimum PFN based on information provided via
7080 * memblock_set_node().
7082 unsigned long __init find_min_pfn_with_active_regions(void)
7084 return PHYS_PFN(memblock_start_of_DRAM());
7088 * early_calculate_totalpages()
7089 * Sum pages in active regions for movable zone.
7090 * Populate N_MEMORY for calculating usable_nodes.
7092 static unsigned long __init early_calculate_totalpages(void)
7094 unsigned long totalpages = 0;
7095 unsigned long start_pfn, end_pfn;
7098 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7099 unsigned long pages = end_pfn - start_pfn;
7101 totalpages += pages;
7103 node_set_state(nid, N_MEMORY);
7109 * Find the PFN the Movable zone begins in each node. Kernel memory
7110 * is spread evenly between nodes as long as the nodes have enough
7111 * memory. When they don't, some nodes will have more kernelcore than
7114 static void __init find_zone_movable_pfns_for_nodes(void)
7117 unsigned long usable_startpfn;
7118 unsigned long kernelcore_node, kernelcore_remaining;
7119 /* save the state before borrow the nodemask */
7120 nodemask_t saved_node_state = node_states[N_MEMORY];
7121 unsigned long totalpages = early_calculate_totalpages();
7122 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7123 struct memblock_region *r;
7125 /* Need to find movable_zone earlier when movable_node is specified. */
7126 find_usable_zone_for_movable();
7129 * If movable_node is specified, ignore kernelcore and movablecore
7132 if (movable_node_is_enabled()) {
7133 for_each_memblock(memory, r) {
7134 if (!memblock_is_hotpluggable(r))
7137 nid = memblock_get_region_node(r);
7139 usable_startpfn = PFN_DOWN(r->base);
7140 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7141 min(usable_startpfn, zone_movable_pfn[nid]) :
7149 * If kernelcore=mirror is specified, ignore movablecore option
7151 if (mirrored_kernelcore) {
7152 bool mem_below_4gb_not_mirrored = false;
7154 for_each_memblock(memory, r) {
7155 if (memblock_is_mirror(r))
7158 nid = memblock_get_region_node(r);
7160 usable_startpfn = memblock_region_memory_base_pfn(r);
7162 if (usable_startpfn < 0x100000) {
7163 mem_below_4gb_not_mirrored = true;
7167 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7168 min(usable_startpfn, zone_movable_pfn[nid]) :
7172 if (mem_below_4gb_not_mirrored)
7173 pr_warn("This configuration results in unmirrored kernel memory.\n");
7179 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7180 * amount of necessary memory.
7182 if (required_kernelcore_percent)
7183 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7185 if (required_movablecore_percent)
7186 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7190 * If movablecore= was specified, calculate what size of
7191 * kernelcore that corresponds so that memory usable for
7192 * any allocation type is evenly spread. If both kernelcore
7193 * and movablecore are specified, then the value of kernelcore
7194 * will be used for required_kernelcore if it's greater than
7195 * what movablecore would have allowed.
7197 if (required_movablecore) {
7198 unsigned long corepages;
7201 * Round-up so that ZONE_MOVABLE is at least as large as what
7202 * was requested by the user
7204 required_movablecore =
7205 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7206 required_movablecore = min(totalpages, required_movablecore);
7207 corepages = totalpages - required_movablecore;
7209 required_kernelcore = max(required_kernelcore, corepages);
7213 * If kernelcore was not specified or kernelcore size is larger
7214 * than totalpages, there is no ZONE_MOVABLE.
7216 if (!required_kernelcore || required_kernelcore >= totalpages)
7219 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7220 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7223 /* Spread kernelcore memory as evenly as possible throughout nodes */
7224 kernelcore_node = required_kernelcore / usable_nodes;
7225 for_each_node_state(nid, N_MEMORY) {
7226 unsigned long start_pfn, end_pfn;
7229 * Recalculate kernelcore_node if the division per node
7230 * now exceeds what is necessary to satisfy the requested
7231 * amount of memory for the kernel
7233 if (required_kernelcore < kernelcore_node)
7234 kernelcore_node = required_kernelcore / usable_nodes;
7237 * As the map is walked, we track how much memory is usable
7238 * by the kernel using kernelcore_remaining. When it is
7239 * 0, the rest of the node is usable by ZONE_MOVABLE
7241 kernelcore_remaining = kernelcore_node;
7243 /* Go through each range of PFNs within this node */
7244 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7245 unsigned long size_pages;
7247 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7248 if (start_pfn >= end_pfn)
7251 /* Account for what is only usable for kernelcore */
7252 if (start_pfn < usable_startpfn) {
7253 unsigned long kernel_pages;
7254 kernel_pages = min(end_pfn, usable_startpfn)
7257 kernelcore_remaining -= min(kernel_pages,
7258 kernelcore_remaining);
7259 required_kernelcore -= min(kernel_pages,
7260 required_kernelcore);
7262 /* Continue if range is now fully accounted */
7263 if (end_pfn <= usable_startpfn) {
7266 * Push zone_movable_pfn to the end so
7267 * that if we have to rebalance
7268 * kernelcore across nodes, we will
7269 * not double account here
7271 zone_movable_pfn[nid] = end_pfn;
7274 start_pfn = usable_startpfn;
7278 * The usable PFN range for ZONE_MOVABLE is from
7279 * start_pfn->end_pfn. Calculate size_pages as the
7280 * number of pages used as kernelcore
7282 size_pages = end_pfn - start_pfn;
7283 if (size_pages > kernelcore_remaining)
7284 size_pages = kernelcore_remaining;
7285 zone_movable_pfn[nid] = start_pfn + size_pages;
7288 * Some kernelcore has been met, update counts and
7289 * break if the kernelcore for this node has been
7292 required_kernelcore -= min(required_kernelcore,
7294 kernelcore_remaining -= size_pages;
7295 if (!kernelcore_remaining)
7301 * If there is still required_kernelcore, we do another pass with one
7302 * less node in the count. This will push zone_movable_pfn[nid] further
7303 * along on the nodes that still have memory until kernelcore is
7307 if (usable_nodes && required_kernelcore > usable_nodes)
7311 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7312 for (nid = 0; nid < MAX_NUMNODES; nid++)
7313 zone_movable_pfn[nid] =
7314 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7317 /* restore the node_state */
7318 node_states[N_MEMORY] = saved_node_state;
7321 /* Any regular or high memory on that node ? */
7322 static void check_for_memory(pg_data_t *pgdat, int nid)
7324 enum zone_type zone_type;
7326 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7327 struct zone *zone = &pgdat->node_zones[zone_type];
7328 if (populated_zone(zone)) {
7329 if (IS_ENABLED(CONFIG_HIGHMEM))
7330 node_set_state(nid, N_HIGH_MEMORY);
7331 if (zone_type <= ZONE_NORMAL)
7332 node_set_state(nid, N_NORMAL_MEMORY);
7339 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7340 * such cases we allow max_zone_pfn sorted in the descending order
7342 bool __weak arch_has_descending_max_zone_pfns(void)
7348 * free_area_init - Initialise all pg_data_t and zone data
7349 * @max_zone_pfn: an array of max PFNs for each zone
7351 * This will call free_area_init_node() for each active node in the system.
7352 * Using the page ranges provided by memblock_set_node(), the size of each
7353 * zone in each node and their holes is calculated. If the maximum PFN
7354 * between two adjacent zones match, it is assumed that the zone is empty.
7355 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7356 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7357 * starts where the previous one ended. For example, ZONE_DMA32 starts
7358 * at arch_max_dma_pfn.
7360 void __init free_area_init(unsigned long *max_zone_pfn)
7362 unsigned long start_pfn, end_pfn;
7366 /* Record where the zone boundaries are */
7367 memset(arch_zone_lowest_possible_pfn, 0,
7368 sizeof(arch_zone_lowest_possible_pfn));
7369 memset(arch_zone_highest_possible_pfn, 0,
7370 sizeof(arch_zone_highest_possible_pfn));
7372 start_pfn = find_min_pfn_with_active_regions();
7373 descending = arch_has_descending_max_zone_pfns();
7375 for (i = 0; i < MAX_NR_ZONES; i++) {
7377 zone = MAX_NR_ZONES - i - 1;
7381 if (zone == ZONE_MOVABLE)
7384 end_pfn = max(max_zone_pfn[zone], start_pfn);
7385 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7386 arch_zone_highest_possible_pfn[zone] = end_pfn;
7388 start_pfn = end_pfn;
7391 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7392 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7393 find_zone_movable_pfns_for_nodes();
7395 /* Print out the zone ranges */
7396 pr_info("Zone ranges:\n");
7397 for (i = 0; i < MAX_NR_ZONES; i++) {
7398 if (i == ZONE_MOVABLE)
7400 pr_info(" %-8s ", zone_names[i]);
7401 if (arch_zone_lowest_possible_pfn[i] ==
7402 arch_zone_highest_possible_pfn[i])
7405 pr_cont("[mem %#018Lx-%#018Lx]\n",
7406 (u64)arch_zone_lowest_possible_pfn[i]
7408 ((u64)arch_zone_highest_possible_pfn[i]
7409 << PAGE_SHIFT) - 1);
7412 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7413 pr_info("Movable zone start for each node\n");
7414 for (i = 0; i < MAX_NUMNODES; i++) {
7415 if (zone_movable_pfn[i])
7416 pr_info(" Node %d: %#018Lx\n", i,
7417 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7421 * Print out the early node map, and initialize the
7422 * subsection-map relative to active online memory ranges to
7423 * enable future "sub-section" extensions of the memory map.
7425 pr_info("Early memory node ranges\n");
7426 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7427 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7428 (u64)start_pfn << PAGE_SHIFT,
7429 ((u64)end_pfn << PAGE_SHIFT) - 1);
7430 subsection_map_init(start_pfn, end_pfn - start_pfn);
7433 /* Initialise every node */
7434 mminit_verify_pageflags_layout();
7435 setup_nr_node_ids();
7436 init_unavailable_mem();
7437 for_each_online_node(nid) {
7438 pg_data_t *pgdat = NODE_DATA(nid);
7439 free_area_init_node(nid);
7441 /* Any memory on that node */
7442 if (pgdat->node_present_pages)
7443 node_set_state(nid, N_MEMORY);
7444 check_for_memory(pgdat, nid);
7448 static int __init cmdline_parse_core(char *p, unsigned long *core,
7449 unsigned long *percent)
7451 unsigned long long coremem;
7457 /* Value may be a percentage of total memory, otherwise bytes */
7458 coremem = simple_strtoull(p, &endptr, 0);
7459 if (*endptr == '%') {
7460 /* Paranoid check for percent values greater than 100 */
7461 WARN_ON(coremem > 100);
7465 coremem = memparse(p, &p);
7466 /* Paranoid check that UL is enough for the coremem value */
7467 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7469 *core = coremem >> PAGE_SHIFT;
7476 * kernelcore=size sets the amount of memory for use for allocations that
7477 * cannot be reclaimed or migrated.
7479 static int __init cmdline_parse_kernelcore(char *p)
7481 /* parse kernelcore=mirror */
7482 if (parse_option_str(p, "mirror")) {
7483 mirrored_kernelcore = true;
7487 return cmdline_parse_core(p, &required_kernelcore,
7488 &required_kernelcore_percent);
7492 * movablecore=size sets the amount of memory for use for allocations that
7493 * can be reclaimed or migrated.
7495 static int __init cmdline_parse_movablecore(char *p)
7497 return cmdline_parse_core(p, &required_movablecore,
7498 &required_movablecore_percent);
7501 early_param("kernelcore", cmdline_parse_kernelcore);
7502 early_param("movablecore", cmdline_parse_movablecore);
7504 void adjust_managed_page_count(struct page *page, long count)
7506 atomic_long_add(count, &page_zone(page)->managed_pages);
7507 totalram_pages_add(count);
7508 #ifdef CONFIG_HIGHMEM
7509 if (PageHighMem(page))
7510 totalhigh_pages_add(count);
7513 EXPORT_SYMBOL(adjust_managed_page_count);
7515 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7518 unsigned long pages = 0;
7520 start = (void *)PAGE_ALIGN((unsigned long)start);
7521 end = (void *)((unsigned long)end & PAGE_MASK);
7522 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7523 struct page *page = virt_to_page(pos);
7524 void *direct_map_addr;
7527 * 'direct_map_addr' might be different from 'pos'
7528 * because some architectures' virt_to_page()
7529 * work with aliases. Getting the direct map
7530 * address ensures that we get a _writeable_
7531 * alias for the memset().
7533 direct_map_addr = page_address(page);
7534 if ((unsigned int)poison <= 0xFF)
7535 memset(direct_map_addr, poison, PAGE_SIZE);
7537 free_reserved_page(page);
7541 pr_info("Freeing %s memory: %ldK\n",
7542 s, pages << (PAGE_SHIFT - 10));
7547 #ifdef CONFIG_HIGHMEM
7548 void free_highmem_page(struct page *page)
7550 __free_reserved_page(page);
7551 totalram_pages_inc();
7552 atomic_long_inc(&page_zone(page)->managed_pages);
7553 totalhigh_pages_inc();
7558 void __init mem_init_print_info(const char *str)
7560 unsigned long physpages, codesize, datasize, rosize, bss_size;
7561 unsigned long init_code_size, init_data_size;
7563 physpages = get_num_physpages();
7564 codesize = _etext - _stext;
7565 datasize = _edata - _sdata;
7566 rosize = __end_rodata - __start_rodata;
7567 bss_size = __bss_stop - __bss_start;
7568 init_data_size = __init_end - __init_begin;
7569 init_code_size = _einittext - _sinittext;
7572 * Detect special cases and adjust section sizes accordingly:
7573 * 1) .init.* may be embedded into .data sections
7574 * 2) .init.text.* may be out of [__init_begin, __init_end],
7575 * please refer to arch/tile/kernel/vmlinux.lds.S.
7576 * 3) .rodata.* may be embedded into .text or .data sections.
7578 #define adj_init_size(start, end, size, pos, adj) \
7580 if (start <= pos && pos < end && size > adj) \
7584 adj_init_size(__init_begin, __init_end, init_data_size,
7585 _sinittext, init_code_size);
7586 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7587 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7588 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7589 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7591 #undef adj_init_size
7593 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7594 #ifdef CONFIG_HIGHMEM
7598 nr_free_pages() << (PAGE_SHIFT - 10),
7599 physpages << (PAGE_SHIFT - 10),
7600 codesize >> 10, datasize >> 10, rosize >> 10,
7601 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7602 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7603 totalcma_pages << (PAGE_SHIFT - 10),
7604 #ifdef CONFIG_HIGHMEM
7605 totalhigh_pages() << (PAGE_SHIFT - 10),
7607 str ? ", " : "", str ? str : "");
7611 * set_dma_reserve - set the specified number of pages reserved in the first zone
7612 * @new_dma_reserve: The number of pages to mark reserved
7614 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7615 * In the DMA zone, a significant percentage may be consumed by kernel image
7616 * and other unfreeable allocations which can skew the watermarks badly. This
7617 * function may optionally be used to account for unfreeable pages in the
7618 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7619 * smaller per-cpu batchsize.
7621 void __init set_dma_reserve(unsigned long new_dma_reserve)
7623 dma_reserve = new_dma_reserve;
7626 static int page_alloc_cpu_dead(unsigned int cpu)
7629 lru_add_drain_cpu(cpu);
7633 * Spill the event counters of the dead processor
7634 * into the current processors event counters.
7635 * This artificially elevates the count of the current
7638 vm_events_fold_cpu(cpu);
7641 * Zero the differential counters of the dead processor
7642 * so that the vm statistics are consistent.
7644 * This is only okay since the processor is dead and cannot
7645 * race with what we are doing.
7647 cpu_vm_stats_fold(cpu);
7652 int hashdist = HASHDIST_DEFAULT;
7654 static int __init set_hashdist(char *str)
7658 hashdist = simple_strtoul(str, &str, 0);
7661 __setup("hashdist=", set_hashdist);
7664 void __init page_alloc_init(void)
7669 if (num_node_state(N_MEMORY) == 1)
7673 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7674 "mm/page_alloc:dead", NULL,
7675 page_alloc_cpu_dead);
7680 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7681 * or min_free_kbytes changes.
7683 static void calculate_totalreserve_pages(void)
7685 struct pglist_data *pgdat;
7686 unsigned long reserve_pages = 0;
7687 enum zone_type i, j;
7689 for_each_online_pgdat(pgdat) {
7691 pgdat->totalreserve_pages = 0;
7693 for (i = 0; i < MAX_NR_ZONES; i++) {
7694 struct zone *zone = pgdat->node_zones + i;
7696 unsigned long managed_pages = zone_managed_pages(zone);
7698 /* Find valid and maximum lowmem_reserve in the zone */
7699 for (j = i; j < MAX_NR_ZONES; j++) {
7700 if (zone->lowmem_reserve[j] > max)
7701 max = zone->lowmem_reserve[j];
7704 /* we treat the high watermark as reserved pages. */
7705 max += high_wmark_pages(zone);
7707 if (max > managed_pages)
7708 max = managed_pages;
7710 pgdat->totalreserve_pages += max;
7712 reserve_pages += max;
7715 totalreserve_pages = reserve_pages;
7719 * setup_per_zone_lowmem_reserve - called whenever
7720 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7721 * has a correct pages reserved value, so an adequate number of
7722 * pages are left in the zone after a successful __alloc_pages().
7724 static void setup_per_zone_lowmem_reserve(void)
7726 struct pglist_data *pgdat;
7727 enum zone_type j, idx;
7729 for_each_online_pgdat(pgdat) {
7730 for (j = 0; j < MAX_NR_ZONES; j++) {
7731 struct zone *zone = pgdat->node_zones + j;
7732 unsigned long managed_pages = zone_managed_pages(zone);
7734 zone->lowmem_reserve[j] = 0;
7738 struct zone *lower_zone;
7741 lower_zone = pgdat->node_zones + idx;
7743 if (!sysctl_lowmem_reserve_ratio[idx] ||
7744 !zone_managed_pages(lower_zone)) {
7745 lower_zone->lowmem_reserve[j] = 0;
7748 lower_zone->lowmem_reserve[j] =
7749 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7751 managed_pages += zone_managed_pages(lower_zone);
7756 /* update totalreserve_pages */
7757 calculate_totalreserve_pages();
7760 static void __setup_per_zone_wmarks(void)
7762 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7763 unsigned long lowmem_pages = 0;
7765 unsigned long flags;
7767 /* Calculate total number of !ZONE_HIGHMEM pages */
7768 for_each_zone(zone) {
7769 if (!is_highmem(zone))
7770 lowmem_pages += zone_managed_pages(zone);
7773 for_each_zone(zone) {
7776 spin_lock_irqsave(&zone->lock, flags);
7777 tmp = (u64)pages_min * zone_managed_pages(zone);
7778 do_div(tmp, lowmem_pages);
7779 if (is_highmem(zone)) {
7781 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7782 * need highmem pages, so cap pages_min to a small
7785 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7786 * deltas control async page reclaim, and so should
7787 * not be capped for highmem.
7789 unsigned long min_pages;
7791 min_pages = zone_managed_pages(zone) / 1024;
7792 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7793 zone->_watermark[WMARK_MIN] = min_pages;
7796 * If it's a lowmem zone, reserve a number of pages
7797 * proportionate to the zone's size.
7799 zone->_watermark[WMARK_MIN] = tmp;
7803 * Set the kswapd watermarks distance according to the
7804 * scale factor in proportion to available memory, but
7805 * ensure a minimum size on small systems.
7807 tmp = max_t(u64, tmp >> 2,
7808 mult_frac(zone_managed_pages(zone),
7809 watermark_scale_factor, 10000));
7811 zone->watermark_boost = 0;
7812 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7813 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7815 spin_unlock_irqrestore(&zone->lock, flags);
7818 /* update totalreserve_pages */
7819 calculate_totalreserve_pages();
7823 * setup_per_zone_wmarks - called when min_free_kbytes changes
7824 * or when memory is hot-{added|removed}
7826 * Ensures that the watermark[min,low,high] values for each zone are set
7827 * correctly with respect to min_free_kbytes.
7829 void setup_per_zone_wmarks(void)
7831 static DEFINE_SPINLOCK(lock);
7834 __setup_per_zone_wmarks();
7839 * Initialise min_free_kbytes.
7841 * For small machines we want it small (128k min). For large machines
7842 * we want it large (256MB max). But it is not linear, because network
7843 * bandwidth does not increase linearly with machine size. We use
7845 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7846 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7862 int __meminit init_per_zone_wmark_min(void)
7864 unsigned long lowmem_kbytes;
7865 int new_min_free_kbytes;
7867 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7868 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7870 if (new_min_free_kbytes > user_min_free_kbytes) {
7871 min_free_kbytes = new_min_free_kbytes;
7872 if (min_free_kbytes < 128)
7873 min_free_kbytes = 128;
7874 if (min_free_kbytes > 262144)
7875 min_free_kbytes = 262144;
7877 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7878 new_min_free_kbytes, user_min_free_kbytes);
7880 setup_per_zone_wmarks();
7881 refresh_zone_stat_thresholds();
7882 setup_per_zone_lowmem_reserve();
7885 setup_min_unmapped_ratio();
7886 setup_min_slab_ratio();
7891 core_initcall(init_per_zone_wmark_min)
7894 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7895 * that we can call two helper functions whenever min_free_kbytes
7898 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7899 void *buffer, size_t *length, loff_t *ppos)
7903 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7908 user_min_free_kbytes = min_free_kbytes;
7909 setup_per_zone_wmarks();
7914 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7915 void *buffer, size_t *length, loff_t *ppos)
7919 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7924 setup_per_zone_wmarks();
7930 static void setup_min_unmapped_ratio(void)
7935 for_each_online_pgdat(pgdat)
7936 pgdat->min_unmapped_pages = 0;
7939 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7940 sysctl_min_unmapped_ratio) / 100;
7944 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7945 void *buffer, size_t *length, loff_t *ppos)
7949 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7953 setup_min_unmapped_ratio();
7958 static void setup_min_slab_ratio(void)
7963 for_each_online_pgdat(pgdat)
7964 pgdat->min_slab_pages = 0;
7967 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7968 sysctl_min_slab_ratio) / 100;
7971 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7972 void *buffer, size_t *length, loff_t *ppos)
7976 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7980 setup_min_slab_ratio();
7987 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7988 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7989 * whenever sysctl_lowmem_reserve_ratio changes.
7991 * The reserve ratio obviously has absolutely no relation with the
7992 * minimum watermarks. The lowmem reserve ratio can only make sense
7993 * if in function of the boot time zone sizes.
7995 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7996 void *buffer, size_t *length, loff_t *ppos)
8000 proc_dointvec_minmax(table, write, buffer, length, ppos);
8002 for (i = 0; i < MAX_NR_ZONES; i++) {
8003 if (sysctl_lowmem_reserve_ratio[i] < 1)
8004 sysctl_lowmem_reserve_ratio[i] = 0;
8007 setup_per_zone_lowmem_reserve();
8011 static void __zone_pcp_update(struct zone *zone)
8015 for_each_possible_cpu(cpu)
8016 pageset_set_high_and_batch(zone,
8017 per_cpu_ptr(zone->pageset, cpu));
8021 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8022 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8023 * pagelist can have before it gets flushed back to buddy allocator.
8025 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8026 void *buffer, size_t *length, loff_t *ppos)
8029 int old_percpu_pagelist_fraction;
8032 mutex_lock(&pcp_batch_high_lock);
8033 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8035 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8036 if (!write || ret < 0)
8039 /* Sanity checking to avoid pcp imbalance */
8040 if (percpu_pagelist_fraction &&
8041 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8042 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8048 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8051 for_each_populated_zone(zone)
8052 __zone_pcp_update(zone);
8054 mutex_unlock(&pcp_batch_high_lock);
8058 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8060 * Returns the number of pages that arch has reserved but
8061 * is not known to alloc_large_system_hash().
8063 static unsigned long __init arch_reserved_kernel_pages(void)
8070 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8071 * machines. As memory size is increased the scale is also increased but at
8072 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8073 * quadruples the scale is increased by one, which means the size of hash table
8074 * only doubles, instead of quadrupling as well.
8075 * Because 32-bit systems cannot have large physical memory, where this scaling
8076 * makes sense, it is disabled on such platforms.
8078 #if __BITS_PER_LONG > 32
8079 #define ADAPT_SCALE_BASE (64ul << 30)
8080 #define ADAPT_SCALE_SHIFT 2
8081 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8085 * allocate a large system hash table from bootmem
8086 * - it is assumed that the hash table must contain an exact power-of-2
8087 * quantity of entries
8088 * - limit is the number of hash buckets, not the total allocation size
8090 void *__init alloc_large_system_hash(const char *tablename,
8091 unsigned long bucketsize,
8092 unsigned long numentries,
8095 unsigned int *_hash_shift,
8096 unsigned int *_hash_mask,
8097 unsigned long low_limit,
8098 unsigned long high_limit)
8100 unsigned long long max = high_limit;
8101 unsigned long log2qty, size;
8106 /* allow the kernel cmdline to have a say */
8108 /* round applicable memory size up to nearest megabyte */
8109 numentries = nr_kernel_pages;
8110 numentries -= arch_reserved_kernel_pages();
8112 /* It isn't necessary when PAGE_SIZE >= 1MB */
8113 if (PAGE_SHIFT < 20)
8114 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8116 #if __BITS_PER_LONG > 32
8118 unsigned long adapt;
8120 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8121 adapt <<= ADAPT_SCALE_SHIFT)
8126 /* limit to 1 bucket per 2^scale bytes of low memory */
8127 if (scale > PAGE_SHIFT)
8128 numentries >>= (scale - PAGE_SHIFT);
8130 numentries <<= (PAGE_SHIFT - scale);
8132 /* Make sure we've got at least a 0-order allocation.. */
8133 if (unlikely(flags & HASH_SMALL)) {
8134 /* Makes no sense without HASH_EARLY */
8135 WARN_ON(!(flags & HASH_EARLY));
8136 if (!(numentries >> *_hash_shift)) {
8137 numentries = 1UL << *_hash_shift;
8138 BUG_ON(!numentries);
8140 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8141 numentries = PAGE_SIZE / bucketsize;
8143 numentries = roundup_pow_of_two(numentries);
8145 /* limit allocation size to 1/16 total memory by default */
8147 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8148 do_div(max, bucketsize);
8150 max = min(max, 0x80000000ULL);
8152 if (numentries < low_limit)
8153 numentries = low_limit;
8154 if (numentries > max)
8157 log2qty = ilog2(numentries);
8159 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8162 size = bucketsize << log2qty;
8163 if (flags & HASH_EARLY) {
8164 if (flags & HASH_ZERO)
8165 table = memblock_alloc(size, SMP_CACHE_BYTES);
8167 table = memblock_alloc_raw(size,
8169 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8170 table = __vmalloc(size, gfp_flags);
8174 * If bucketsize is not a power-of-two, we may free
8175 * some pages at the end of hash table which
8176 * alloc_pages_exact() automatically does
8178 table = alloc_pages_exact(size, gfp_flags);
8179 kmemleak_alloc(table, size, 1, gfp_flags);
8181 } while (!table && size > PAGE_SIZE && --log2qty);
8184 panic("Failed to allocate %s hash table\n", tablename);
8186 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8187 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8188 virt ? "vmalloc" : "linear");
8191 *_hash_shift = log2qty;
8193 *_hash_mask = (1 << log2qty) - 1;
8199 * This function checks whether pageblock includes unmovable pages or not.
8201 * PageLRU check without isolation or lru_lock could race so that
8202 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8203 * check without lock_page also may miss some movable non-lru pages at
8204 * race condition. So you can't expect this function should be exact.
8206 * Returns a page without holding a reference. If the caller wants to
8207 * dereference that page (e.g., dumping), it has to make sure that it
8208 * cannot get removed (e.g., via memory unplug) concurrently.
8211 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8212 int migratetype, int flags)
8214 unsigned long iter = 0;
8215 unsigned long pfn = page_to_pfn(page);
8218 * TODO we could make this much more efficient by not checking every
8219 * page in the range if we know all of them are in MOVABLE_ZONE and
8220 * that the movable zone guarantees that pages are migratable but
8221 * the later is not the case right now unfortunatelly. E.g. movablecore
8222 * can still lead to having bootmem allocations in zone_movable.
8225 if (is_migrate_cma_page(page)) {
8227 * CMA allocations (alloc_contig_range) really need to mark
8228 * isolate CMA pageblocks even when they are not movable in fact
8229 * so consider them movable here.
8231 if (is_migrate_cma(migratetype))
8237 for (; iter < pageblock_nr_pages; iter++) {
8238 if (!pfn_valid_within(pfn + iter))
8241 page = pfn_to_page(pfn + iter);
8243 if (PageReserved(page))
8247 * If the zone is movable and we have ruled out all reserved
8248 * pages then it should be reasonably safe to assume the rest
8251 if (zone_idx(zone) == ZONE_MOVABLE)
8255 * Hugepages are not in LRU lists, but they're movable.
8256 * THPs are on the LRU, but need to be counted as #small pages.
8257 * We need not scan over tail pages because we don't
8258 * handle each tail page individually in migration.
8260 if (PageHuge(page) || PageTransCompound(page)) {
8261 struct page *head = compound_head(page);
8262 unsigned int skip_pages;
8264 if (PageHuge(page)) {
8265 if (!hugepage_migration_supported(page_hstate(head)))
8267 } else if (!PageLRU(head) && !__PageMovable(head)) {
8271 skip_pages = compound_nr(head) - (page - head);
8272 iter += skip_pages - 1;
8277 * We can't use page_count without pin a page
8278 * because another CPU can free compound page.
8279 * This check already skips compound tails of THP
8280 * because their page->_refcount is zero at all time.
8282 if (!page_ref_count(page)) {
8283 if (PageBuddy(page))
8284 iter += (1 << page_order(page)) - 1;
8289 * The HWPoisoned page may be not in buddy system, and
8290 * page_count() is not 0.
8292 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8296 * We treat all PageOffline() pages as movable when offlining
8297 * to give drivers a chance to decrement their reference count
8298 * in MEM_GOING_OFFLINE in order to indicate that these pages
8299 * can be offlined as there are no direct references anymore.
8300 * For actually unmovable PageOffline() where the driver does
8301 * not support this, we will fail later when trying to actually
8302 * move these pages that still have a reference count > 0.
8303 * (false negatives in this function only)
8305 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8308 if (__PageMovable(page) || PageLRU(page))
8312 * If there are RECLAIMABLE pages, we need to check
8313 * it. But now, memory offline itself doesn't call
8314 * shrink_node_slabs() and it still to be fixed.
8317 * If the page is not RAM, page_count()should be 0.
8318 * we don't need more check. This is an _used_ not-movable page.
8320 * The problematic thing here is PG_reserved pages. PG_reserved
8321 * is set to both of a memory hole page and a _used_ kernel
8329 #ifdef CONFIG_CONTIG_ALLOC
8330 static unsigned long pfn_max_align_down(unsigned long pfn)
8332 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8333 pageblock_nr_pages) - 1);
8336 static unsigned long pfn_max_align_up(unsigned long pfn)
8338 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8339 pageblock_nr_pages));
8342 /* [start, end) must belong to a single zone. */
8343 static int __alloc_contig_migrate_range(struct compact_control *cc,
8344 unsigned long start, unsigned long end)
8346 /* This function is based on compact_zone() from compaction.c. */
8347 unsigned int nr_reclaimed;
8348 unsigned long pfn = start;
8349 unsigned int tries = 0;
8351 struct migration_target_control mtc = {
8352 .nid = zone_to_nid(cc->zone),
8353 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8358 while (pfn < end || !list_empty(&cc->migratepages)) {
8359 if (fatal_signal_pending(current)) {
8364 if (list_empty(&cc->migratepages)) {
8365 cc->nr_migratepages = 0;
8366 pfn = isolate_migratepages_range(cc, pfn, end);
8372 } else if (++tries == 5) {
8373 ret = ret < 0 ? ret : -EBUSY;
8377 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8379 cc->nr_migratepages -= nr_reclaimed;
8381 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8382 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8385 putback_movable_pages(&cc->migratepages);
8392 * alloc_contig_range() -- tries to allocate given range of pages
8393 * @start: start PFN to allocate
8394 * @end: one-past-the-last PFN to allocate
8395 * @migratetype: migratetype of the underlaying pageblocks (either
8396 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8397 * in range must have the same migratetype and it must
8398 * be either of the two.
8399 * @gfp_mask: GFP mask to use during compaction
8401 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8402 * aligned. The PFN range must belong to a single zone.
8404 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8405 * pageblocks in the range. Once isolated, the pageblocks should not
8406 * be modified by others.
8408 * Return: zero on success or negative error code. On success all
8409 * pages which PFN is in [start, end) are allocated for the caller and
8410 * need to be freed with free_contig_range().
8412 int alloc_contig_range(unsigned long start, unsigned long end,
8413 unsigned migratetype, gfp_t gfp_mask)
8415 unsigned long outer_start, outer_end;
8419 struct compact_control cc = {
8420 .nr_migratepages = 0,
8422 .zone = page_zone(pfn_to_page(start)),
8423 .mode = MIGRATE_SYNC,
8424 .ignore_skip_hint = true,
8425 .no_set_skip_hint = true,
8426 .gfp_mask = current_gfp_context(gfp_mask),
8427 .alloc_contig = true,
8429 INIT_LIST_HEAD(&cc.migratepages);
8432 * What we do here is we mark all pageblocks in range as
8433 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8434 * have different sizes, and due to the way page allocator
8435 * work, we align the range to biggest of the two pages so
8436 * that page allocator won't try to merge buddies from
8437 * different pageblocks and change MIGRATE_ISOLATE to some
8438 * other migration type.
8440 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8441 * migrate the pages from an unaligned range (ie. pages that
8442 * we are interested in). This will put all the pages in
8443 * range back to page allocator as MIGRATE_ISOLATE.
8445 * When this is done, we take the pages in range from page
8446 * allocator removing them from the buddy system. This way
8447 * page allocator will never consider using them.
8449 * This lets us mark the pageblocks back as
8450 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8451 * aligned range but not in the unaligned, original range are
8452 * put back to page allocator so that buddy can use them.
8455 ret = start_isolate_page_range(pfn_max_align_down(start),
8456 pfn_max_align_up(end), migratetype, 0);
8461 * In case of -EBUSY, we'd like to know which page causes problem.
8462 * So, just fall through. test_pages_isolated() has a tracepoint
8463 * which will report the busy page.
8465 * It is possible that busy pages could become available before
8466 * the call to test_pages_isolated, and the range will actually be
8467 * allocated. So, if we fall through be sure to clear ret so that
8468 * -EBUSY is not accidentally used or returned to caller.
8470 ret = __alloc_contig_migrate_range(&cc, start, end);
8471 if (ret && ret != -EBUSY)
8476 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8477 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8478 * more, all pages in [start, end) are free in page allocator.
8479 * What we are going to do is to allocate all pages from
8480 * [start, end) (that is remove them from page allocator).
8482 * The only problem is that pages at the beginning and at the
8483 * end of interesting range may be not aligned with pages that
8484 * page allocator holds, ie. they can be part of higher order
8485 * pages. Because of this, we reserve the bigger range and
8486 * once this is done free the pages we are not interested in.
8488 * We don't have to hold zone->lock here because the pages are
8489 * isolated thus they won't get removed from buddy.
8492 lru_add_drain_all();
8495 outer_start = start;
8496 while (!PageBuddy(pfn_to_page(outer_start))) {
8497 if (++order >= MAX_ORDER) {
8498 outer_start = start;
8501 outer_start &= ~0UL << order;
8504 if (outer_start != start) {
8505 order = page_order(pfn_to_page(outer_start));
8508 * outer_start page could be small order buddy page and
8509 * it doesn't include start page. Adjust outer_start
8510 * in this case to report failed page properly
8511 * on tracepoint in test_pages_isolated()
8513 if (outer_start + (1UL << order) <= start)
8514 outer_start = start;
8517 /* Make sure the range is really isolated. */
8518 if (test_pages_isolated(outer_start, end, 0)) {
8519 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8520 __func__, outer_start, end);
8525 /* Grab isolated pages from freelists. */
8526 outer_end = isolate_freepages_range(&cc, outer_start, end);
8532 /* Free head and tail (if any) */
8533 if (start != outer_start)
8534 free_contig_range(outer_start, start - outer_start);
8535 if (end != outer_end)
8536 free_contig_range(end, outer_end - end);
8539 undo_isolate_page_range(pfn_max_align_down(start),
8540 pfn_max_align_up(end), migratetype);
8543 EXPORT_SYMBOL(alloc_contig_range);
8545 static int __alloc_contig_pages(unsigned long start_pfn,
8546 unsigned long nr_pages, gfp_t gfp_mask)
8548 unsigned long end_pfn = start_pfn + nr_pages;
8550 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8554 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8555 unsigned long nr_pages)
8557 unsigned long i, end_pfn = start_pfn + nr_pages;
8560 for (i = start_pfn; i < end_pfn; i++) {
8561 page = pfn_to_online_page(i);
8565 if (page_zone(page) != z)
8568 if (PageReserved(page))
8571 if (page_count(page) > 0)
8580 static bool zone_spans_last_pfn(const struct zone *zone,
8581 unsigned long start_pfn, unsigned long nr_pages)
8583 unsigned long last_pfn = start_pfn + nr_pages - 1;
8585 return zone_spans_pfn(zone, last_pfn);
8589 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8590 * @nr_pages: Number of contiguous pages to allocate
8591 * @gfp_mask: GFP mask to limit search and used during compaction
8593 * @nodemask: Mask for other possible nodes
8595 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8596 * on an applicable zonelist to find a contiguous pfn range which can then be
8597 * tried for allocation with alloc_contig_range(). This routine is intended
8598 * for allocation requests which can not be fulfilled with the buddy allocator.
8600 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8601 * power of two then the alignment is guaranteed to be to the given nr_pages
8602 * (e.g. 1GB request would be aligned to 1GB).
8604 * Allocated pages can be freed with free_contig_range() or by manually calling
8605 * __free_page() on each allocated page.
8607 * Return: pointer to contiguous pages on success, or NULL if not successful.
8609 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8610 int nid, nodemask_t *nodemask)
8612 unsigned long ret, pfn, flags;
8613 struct zonelist *zonelist;
8617 zonelist = node_zonelist(nid, gfp_mask);
8618 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8619 gfp_zone(gfp_mask), nodemask) {
8620 spin_lock_irqsave(&zone->lock, flags);
8622 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8623 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8624 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8626 * We release the zone lock here because
8627 * alloc_contig_range() will also lock the zone
8628 * at some point. If there's an allocation
8629 * spinning on this lock, it may win the race
8630 * and cause alloc_contig_range() to fail...
8632 spin_unlock_irqrestore(&zone->lock, flags);
8633 ret = __alloc_contig_pages(pfn, nr_pages,
8636 return pfn_to_page(pfn);
8637 spin_lock_irqsave(&zone->lock, flags);
8641 spin_unlock_irqrestore(&zone->lock, flags);
8645 #endif /* CONFIG_CONTIG_ALLOC */
8647 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8649 unsigned int count = 0;
8651 for (; nr_pages--; pfn++) {
8652 struct page *page = pfn_to_page(pfn);
8654 count += page_count(page) != 1;
8657 WARN(count != 0, "%d pages are still in use!\n", count);
8659 EXPORT_SYMBOL(free_contig_range);
8662 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8663 * page high values need to be recalulated.
8665 void __meminit zone_pcp_update(struct zone *zone)
8667 mutex_lock(&pcp_batch_high_lock);
8668 __zone_pcp_update(zone);
8669 mutex_unlock(&pcp_batch_high_lock);
8672 void zone_pcp_reset(struct zone *zone)
8674 unsigned long flags;
8676 struct per_cpu_pageset *pset;
8678 /* avoid races with drain_pages() */
8679 local_irq_save(flags);
8680 if (zone->pageset != &boot_pageset) {
8681 for_each_online_cpu(cpu) {
8682 pset = per_cpu_ptr(zone->pageset, cpu);
8683 drain_zonestat(zone, pset);
8685 free_percpu(zone->pageset);
8686 zone->pageset = &boot_pageset;
8688 local_irq_restore(flags);
8691 #ifdef CONFIG_MEMORY_HOTREMOVE
8693 * All pages in the range must be in a single zone and isolated
8694 * before calling this.
8697 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8703 unsigned long flags;
8704 unsigned long offlined_pages = 0;
8706 /* find the first valid pfn */
8707 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8711 return offlined_pages;
8713 offline_mem_sections(pfn, end_pfn);
8714 zone = page_zone(pfn_to_page(pfn));
8715 spin_lock_irqsave(&zone->lock, flags);
8717 while (pfn < end_pfn) {
8718 if (!pfn_valid(pfn)) {
8722 page = pfn_to_page(pfn);
8724 * The HWPoisoned page may be not in buddy system, and
8725 * page_count() is not 0.
8727 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8733 * At this point all remaining PageOffline() pages have a
8734 * reference count of 0 and can simply be skipped.
8736 if (PageOffline(page)) {
8737 BUG_ON(page_count(page));
8738 BUG_ON(PageBuddy(page));
8744 BUG_ON(page_count(page));
8745 BUG_ON(!PageBuddy(page));
8746 order = page_order(page);
8747 offlined_pages += 1 << order;
8748 del_page_from_free_list(page, zone, order);
8749 pfn += (1 << order);
8751 spin_unlock_irqrestore(&zone->lock, flags);
8753 return offlined_pages;
8757 bool is_free_buddy_page(struct page *page)
8759 struct zone *zone = page_zone(page);
8760 unsigned long pfn = page_to_pfn(page);
8761 unsigned long flags;
8764 spin_lock_irqsave(&zone->lock, flags);
8765 for (order = 0; order < MAX_ORDER; order++) {
8766 struct page *page_head = page - (pfn & ((1 << order) - 1));
8768 if (PageBuddy(page_head) && page_order(page_head) >= order)
8771 spin_unlock_irqrestore(&zone->lock, flags);
8773 return order < MAX_ORDER;
8776 #ifdef CONFIG_MEMORY_FAILURE
8778 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8779 * test is performed under the zone lock to prevent a race against page
8782 bool set_hwpoison_free_buddy_page(struct page *page)
8784 struct zone *zone = page_zone(page);
8785 unsigned long pfn = page_to_pfn(page);
8786 unsigned long flags;
8788 bool hwpoisoned = false;
8790 spin_lock_irqsave(&zone->lock, flags);
8791 for (order = 0; order < MAX_ORDER; order++) {
8792 struct page *page_head = page - (pfn & ((1 << order) - 1));
8794 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8795 if (!TestSetPageHWPoison(page))
8800 spin_unlock_irqrestore(&zone->lock, flags);