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>
72 #include <linux/khugepaged.h>
74 #include <asm/sections.h>
75 #include <asm/tlbflush.h>
76 #include <asm/div64.h>
79 #include "page_reporting.h"
81 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
82 typedef int __bitwise fpi_t;
84 /* No special request */
85 #define FPI_NONE ((__force fpi_t)0)
88 * Skip free page reporting notification for the (possibly merged) page.
89 * This does not hinder free page reporting from grabbing the page,
90 * reporting it and marking it "reported" - it only skips notifying
91 * the free page reporting infrastructure about a newly freed page. For
92 * example, used when temporarily pulling a page from a freelist and
93 * putting it back unmodified.
95 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
98 * Place the (possibly merged) page to the tail of the freelist. Will ignore
99 * page shuffling (relevant code - e.g., memory onlining - is expected to
100 * shuffle the whole zone).
102 * Note: No code should rely on this flag for correctness - it's purely
103 * to allow for optimizations when handing back either fresh pages
104 * (memory onlining) or untouched pages (page isolation, free page
107 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
109 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
110 static DEFINE_MUTEX(pcp_batch_high_lock);
111 #define MIN_PERCPU_PAGELIST_FRACTION (8)
113 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
114 DEFINE_PER_CPU(int, numa_node);
115 EXPORT_PER_CPU_SYMBOL(numa_node);
118 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
120 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
122 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
123 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
124 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
125 * defined in <linux/topology.h>.
127 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
128 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
131 /* work_structs for global per-cpu drains */
134 struct work_struct work;
136 static DEFINE_MUTEX(pcpu_drain_mutex);
137 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
139 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
140 volatile unsigned long latent_entropy __latent_entropy;
141 EXPORT_SYMBOL(latent_entropy);
145 * Array of node states.
147 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
148 [N_POSSIBLE] = NODE_MASK_ALL,
149 [N_ONLINE] = { { [0] = 1UL } },
151 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
152 #ifdef CONFIG_HIGHMEM
153 [N_HIGH_MEMORY] = { { [0] = 1UL } },
155 [N_MEMORY] = { { [0] = 1UL } },
156 [N_CPU] = { { [0] = 1UL } },
159 EXPORT_SYMBOL(node_states);
161 atomic_long_t _totalram_pages __read_mostly;
162 EXPORT_SYMBOL(_totalram_pages);
163 unsigned long totalreserve_pages __read_mostly;
164 unsigned long totalcma_pages __read_mostly;
166 int percpu_pagelist_fraction;
167 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
168 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
169 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
171 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
173 EXPORT_SYMBOL(init_on_alloc);
175 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
176 DEFINE_STATIC_KEY_TRUE(init_on_free);
178 DEFINE_STATIC_KEY_FALSE(init_on_free);
180 EXPORT_SYMBOL(init_on_free);
182 static int __init early_init_on_alloc(char *buf)
187 ret = kstrtobool(buf, &bool_result);
190 if (bool_result && page_poisoning_enabled())
191 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
193 static_branch_enable(&init_on_alloc);
195 static_branch_disable(&init_on_alloc);
198 early_param("init_on_alloc", early_init_on_alloc);
200 static int __init early_init_on_free(char *buf)
205 ret = kstrtobool(buf, &bool_result);
208 if (bool_result && page_poisoning_enabled())
209 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
211 static_branch_enable(&init_on_free);
213 static_branch_disable(&init_on_free);
216 early_param("init_on_free", early_init_on_free);
219 * A cached value of the page's pageblock's migratetype, used when the page is
220 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
221 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
222 * Also the migratetype set in the page does not necessarily match the pcplist
223 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
224 * other index - this ensures that it will be put on the correct CMA freelist.
226 static inline int get_pcppage_migratetype(struct page *page)
231 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
233 page->index = migratetype;
236 #ifdef CONFIG_PM_SLEEP
238 * The following functions are used by the suspend/hibernate code to temporarily
239 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
240 * while devices are suspended. To avoid races with the suspend/hibernate code,
241 * they should always be called with system_transition_mutex held
242 * (gfp_allowed_mask also should only be modified with system_transition_mutex
243 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
244 * with that modification).
247 static gfp_t saved_gfp_mask;
249 void pm_restore_gfp_mask(void)
251 WARN_ON(!mutex_is_locked(&system_transition_mutex));
252 if (saved_gfp_mask) {
253 gfp_allowed_mask = saved_gfp_mask;
258 void pm_restrict_gfp_mask(void)
260 WARN_ON(!mutex_is_locked(&system_transition_mutex));
261 WARN_ON(saved_gfp_mask);
262 saved_gfp_mask = gfp_allowed_mask;
263 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
266 bool pm_suspended_storage(void)
268 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
272 #endif /* CONFIG_PM_SLEEP */
274 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
275 unsigned int pageblock_order __read_mostly;
278 static void __free_pages_ok(struct page *page, unsigned int order,
282 * results with 256, 32 in the lowmem_reserve sysctl:
283 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
284 * 1G machine -> (16M dma, 784M normal, 224M high)
285 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
286 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
287 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
289 * TBD: should special case ZONE_DMA32 machines here - in those we normally
290 * don't need any ZONE_NORMAL reservation
292 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
293 #ifdef CONFIG_ZONE_DMA
296 #ifdef CONFIG_ZONE_DMA32
300 #ifdef CONFIG_HIGHMEM
306 static char * const zone_names[MAX_NR_ZONES] = {
307 #ifdef CONFIG_ZONE_DMA
310 #ifdef CONFIG_ZONE_DMA32
314 #ifdef CONFIG_HIGHMEM
318 #ifdef CONFIG_ZONE_DEVICE
323 const char * const migratetype_names[MIGRATE_TYPES] = {
331 #ifdef CONFIG_MEMORY_ISOLATION
336 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
337 [NULL_COMPOUND_DTOR] = NULL,
338 [COMPOUND_PAGE_DTOR] = free_compound_page,
339 #ifdef CONFIG_HUGETLB_PAGE
340 [HUGETLB_PAGE_DTOR] = free_huge_page,
342 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
343 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
347 int min_free_kbytes = 1024;
348 int user_min_free_kbytes = -1;
349 #ifdef CONFIG_DISCONTIGMEM
351 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
352 * are not on separate NUMA nodes. Functionally this works but with
353 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
354 * quite small. By default, do not boost watermarks on discontigmem as in
355 * many cases very high-order allocations like THP are likely to be
356 * unsupported and the premature reclaim offsets the advantage of long-term
357 * fragmentation avoidance.
359 int watermark_boost_factor __read_mostly;
361 int watermark_boost_factor __read_mostly = 15000;
363 int watermark_scale_factor = 10;
365 static unsigned long nr_kernel_pages __initdata;
366 static unsigned long nr_all_pages __initdata;
367 static unsigned long dma_reserve __initdata;
369 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
370 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
371 static unsigned long required_kernelcore __initdata;
372 static unsigned long required_kernelcore_percent __initdata;
373 static unsigned long required_movablecore __initdata;
374 static unsigned long required_movablecore_percent __initdata;
375 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
376 static bool mirrored_kernelcore __meminitdata;
378 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
380 EXPORT_SYMBOL(movable_zone);
383 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
384 unsigned int nr_online_nodes __read_mostly = 1;
385 EXPORT_SYMBOL(nr_node_ids);
386 EXPORT_SYMBOL(nr_online_nodes);
389 int page_group_by_mobility_disabled __read_mostly;
391 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
393 * During boot we initialize deferred pages on-demand, as needed, but once
394 * page_alloc_init_late() has finished, the deferred pages are all initialized,
395 * and we can permanently disable that path.
397 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
400 * Calling kasan_free_pages() only after deferred memory initialization
401 * has completed. Poisoning pages during deferred memory init will greatly
402 * lengthen the process and cause problem in large memory systems as the
403 * deferred pages initialization is done with interrupt disabled.
405 * Assuming that there will be no reference to those newly initialized
406 * pages before they are ever allocated, this should have no effect on
407 * KASAN memory tracking as the poison will be properly inserted at page
408 * allocation time. The only corner case is when pages are allocated by
409 * on-demand allocation and then freed again before the deferred pages
410 * initialization is done, but this is not likely to happen.
412 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
414 if (!static_branch_unlikely(&deferred_pages))
415 kasan_free_pages(page, order);
418 /* Returns true if the struct page for the pfn is uninitialised */
419 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
421 int nid = early_pfn_to_nid(pfn);
423 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
430 * Returns true when the remaining initialisation should be deferred until
431 * later in the boot cycle when it can be parallelised.
433 static bool __meminit
434 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
436 static unsigned long prev_end_pfn, nr_initialised;
439 * prev_end_pfn static that contains the end of previous zone
440 * No need to protect because called very early in boot before smp_init.
442 if (prev_end_pfn != end_pfn) {
443 prev_end_pfn = end_pfn;
447 /* Always populate low zones for address-constrained allocations */
448 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
452 * We start only with one section of pages, more pages are added as
453 * needed until the rest of deferred pages are initialized.
456 if ((nr_initialised > PAGES_PER_SECTION) &&
457 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
458 NODE_DATA(nid)->first_deferred_pfn = pfn;
464 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
466 static inline bool early_page_uninitialised(unsigned long pfn)
471 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
477 /* Return a pointer to the bitmap storing bits affecting a block of pages */
478 static inline unsigned long *get_pageblock_bitmap(struct page *page,
481 #ifdef CONFIG_SPARSEMEM
482 return section_to_usemap(__pfn_to_section(pfn));
484 return page_zone(page)->pageblock_flags;
485 #endif /* CONFIG_SPARSEMEM */
488 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
490 #ifdef CONFIG_SPARSEMEM
491 pfn &= (PAGES_PER_SECTION-1);
493 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
494 #endif /* CONFIG_SPARSEMEM */
495 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
499 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
500 * @page: The page within the block of interest
501 * @pfn: The target page frame number
502 * @mask: mask of bits that the caller is interested in
504 * Return: pageblock_bits flags
506 static __always_inline
507 unsigned long __get_pfnblock_flags_mask(struct page *page,
511 unsigned long *bitmap;
512 unsigned long bitidx, word_bitidx;
515 bitmap = get_pageblock_bitmap(page, pfn);
516 bitidx = pfn_to_bitidx(page, pfn);
517 word_bitidx = bitidx / BITS_PER_LONG;
518 bitidx &= (BITS_PER_LONG-1);
520 word = bitmap[word_bitidx];
521 return (word >> bitidx) & mask;
524 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
527 return __get_pfnblock_flags_mask(page, pfn, mask);
530 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
532 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
536 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
537 * @page: The page within the block of interest
538 * @flags: The flags to set
539 * @pfn: The target page frame number
540 * @mask: mask of bits that the caller is interested in
542 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
546 unsigned long *bitmap;
547 unsigned long bitidx, word_bitidx;
548 unsigned long old_word, word;
550 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
551 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
553 bitmap = get_pageblock_bitmap(page, pfn);
554 bitidx = pfn_to_bitidx(page, pfn);
555 word_bitidx = bitidx / BITS_PER_LONG;
556 bitidx &= (BITS_PER_LONG-1);
558 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
563 word = READ_ONCE(bitmap[word_bitidx]);
565 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
566 if (word == old_word)
572 void set_pageblock_migratetype(struct page *page, int migratetype)
574 if (unlikely(page_group_by_mobility_disabled &&
575 migratetype < MIGRATE_PCPTYPES))
576 migratetype = MIGRATE_UNMOVABLE;
578 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
579 page_to_pfn(page), MIGRATETYPE_MASK);
582 #ifdef CONFIG_DEBUG_VM
583 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
587 unsigned long pfn = page_to_pfn(page);
588 unsigned long sp, start_pfn;
591 seq = zone_span_seqbegin(zone);
592 start_pfn = zone->zone_start_pfn;
593 sp = zone->spanned_pages;
594 if (!zone_spans_pfn(zone, pfn))
596 } while (zone_span_seqretry(zone, seq));
599 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
600 pfn, zone_to_nid(zone), zone->name,
601 start_pfn, start_pfn + sp);
606 static int page_is_consistent(struct zone *zone, struct page *page)
608 if (!pfn_valid_within(page_to_pfn(page)))
610 if (zone != page_zone(page))
616 * Temporary debugging check for pages not lying within a given zone.
618 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
620 if (page_outside_zone_boundaries(zone, page))
622 if (!page_is_consistent(zone, page))
628 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
634 static void bad_page(struct page *page, const char *reason)
636 static unsigned long resume;
637 static unsigned long nr_shown;
638 static unsigned long nr_unshown;
641 * Allow a burst of 60 reports, then keep quiet for that minute;
642 * or allow a steady drip of one report per second.
644 if (nr_shown == 60) {
645 if (time_before(jiffies, resume)) {
651 "BUG: Bad page state: %lu messages suppressed\n",
658 resume = jiffies + 60 * HZ;
660 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
661 current->comm, page_to_pfn(page));
662 __dump_page(page, reason);
663 dump_page_owner(page);
668 /* Leave bad fields for debug, except PageBuddy could make trouble */
669 page_mapcount_reset(page); /* remove PageBuddy */
670 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
674 * Higher-order pages are called "compound pages". They are structured thusly:
676 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
678 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
679 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
681 * The first tail page's ->compound_dtor holds the offset in array of compound
682 * page destructors. See compound_page_dtors.
684 * The first tail page's ->compound_order holds the order of allocation.
685 * This usage means that zero-order pages may not be compound.
688 void free_compound_page(struct page *page)
690 mem_cgroup_uncharge(page);
691 __free_pages_ok(page, compound_order(page), FPI_NONE);
694 void prep_compound_page(struct page *page, unsigned int order)
697 int nr_pages = 1 << order;
700 for (i = 1; i < nr_pages; i++) {
701 struct page *p = page + i;
702 set_page_count(p, 0);
703 p->mapping = TAIL_MAPPING;
704 set_compound_head(p, page);
707 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
708 set_compound_order(page, order);
709 atomic_set(compound_mapcount_ptr(page), -1);
710 if (hpage_pincount_available(page))
711 atomic_set(compound_pincount_ptr(page), 0);
714 #ifdef CONFIG_DEBUG_PAGEALLOC
715 unsigned int _debug_guardpage_minorder;
717 bool _debug_pagealloc_enabled_early __read_mostly
718 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
719 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
720 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
721 EXPORT_SYMBOL(_debug_pagealloc_enabled);
723 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
725 static int __init early_debug_pagealloc(char *buf)
727 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
729 early_param("debug_pagealloc", early_debug_pagealloc);
731 void init_debug_pagealloc(void)
733 if (!debug_pagealloc_enabled())
736 static_branch_enable(&_debug_pagealloc_enabled);
738 if (!debug_guardpage_minorder())
741 static_branch_enable(&_debug_guardpage_enabled);
744 static int __init debug_guardpage_minorder_setup(char *buf)
748 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
749 pr_err("Bad debug_guardpage_minorder value\n");
752 _debug_guardpage_minorder = res;
753 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
756 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
758 static inline bool set_page_guard(struct zone *zone, struct page *page,
759 unsigned int order, int migratetype)
761 if (!debug_guardpage_enabled())
764 if (order >= debug_guardpage_minorder())
767 __SetPageGuard(page);
768 INIT_LIST_HEAD(&page->lru);
769 set_page_private(page, order);
770 /* Guard pages are not available for any usage */
771 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
776 static inline void clear_page_guard(struct zone *zone, struct page *page,
777 unsigned int order, int migratetype)
779 if (!debug_guardpage_enabled())
782 __ClearPageGuard(page);
784 set_page_private(page, 0);
785 if (!is_migrate_isolate(migratetype))
786 __mod_zone_freepage_state(zone, (1 << order), migratetype);
789 static inline bool set_page_guard(struct zone *zone, struct page *page,
790 unsigned int order, int migratetype) { return false; }
791 static inline void clear_page_guard(struct zone *zone, struct page *page,
792 unsigned int order, int migratetype) {}
795 static inline void set_buddy_order(struct page *page, unsigned int order)
797 set_page_private(page, order);
798 __SetPageBuddy(page);
802 * This function checks whether a page is free && is the buddy
803 * we can coalesce a page and its buddy if
804 * (a) the buddy is not in a hole (check before calling!) &&
805 * (b) the buddy is in the buddy system &&
806 * (c) a page and its buddy have the same order &&
807 * (d) a page and its buddy are in the same zone.
809 * For recording whether a page is in the buddy system, we set PageBuddy.
810 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
812 * For recording page's order, we use page_private(page).
814 static inline bool page_is_buddy(struct page *page, struct page *buddy,
817 if (!page_is_guard(buddy) && !PageBuddy(buddy))
820 if (buddy_order(buddy) != order)
824 * zone check is done late to avoid uselessly calculating
825 * zone/node ids for pages that could never merge.
827 if (page_zone_id(page) != page_zone_id(buddy))
830 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
835 #ifdef CONFIG_COMPACTION
836 static inline struct capture_control *task_capc(struct zone *zone)
838 struct capture_control *capc = current->capture_control;
840 return unlikely(capc) &&
841 !(current->flags & PF_KTHREAD) &&
843 capc->cc->zone == zone ? capc : NULL;
847 compaction_capture(struct capture_control *capc, struct page *page,
848 int order, int migratetype)
850 if (!capc || order != capc->cc->order)
853 /* Do not accidentally pollute CMA or isolated regions*/
854 if (is_migrate_cma(migratetype) ||
855 is_migrate_isolate(migratetype))
859 * Do not let lower order allocations polluate a movable pageblock.
860 * This might let an unmovable request use a reclaimable pageblock
861 * and vice-versa but no more than normal fallback logic which can
862 * have trouble finding a high-order free page.
864 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
872 static inline struct capture_control *task_capc(struct zone *zone)
878 compaction_capture(struct capture_control *capc, struct page *page,
879 int order, int migratetype)
883 #endif /* CONFIG_COMPACTION */
885 /* Used for pages not on another list */
886 static inline void add_to_free_list(struct page *page, struct zone *zone,
887 unsigned int order, int migratetype)
889 struct free_area *area = &zone->free_area[order];
891 list_add(&page->lru, &area->free_list[migratetype]);
895 /* Used for pages not on another list */
896 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
897 unsigned int order, int migratetype)
899 struct free_area *area = &zone->free_area[order];
901 list_add_tail(&page->lru, &area->free_list[migratetype]);
906 * Used for pages which are on another list. Move the pages to the tail
907 * of the list - so the moved pages won't immediately be considered for
908 * allocation again (e.g., optimization for memory onlining).
910 static inline void move_to_free_list(struct page *page, struct zone *zone,
911 unsigned int order, int migratetype)
913 struct free_area *area = &zone->free_area[order];
915 list_move_tail(&page->lru, &area->free_list[migratetype]);
918 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
921 /* clear reported state and update reported page count */
922 if (page_reported(page))
923 __ClearPageReported(page);
925 list_del(&page->lru);
926 __ClearPageBuddy(page);
927 set_page_private(page, 0);
928 zone->free_area[order].nr_free--;
932 * If this is not the largest possible page, check if the buddy
933 * of the next-highest order is free. If it is, it's possible
934 * that pages are being freed that will coalesce soon. In case,
935 * that is happening, add the free page to the tail of the list
936 * so it's less likely to be used soon and more likely to be merged
937 * as a higher order page
940 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
941 struct page *page, unsigned int order)
943 struct page *higher_page, *higher_buddy;
944 unsigned long combined_pfn;
946 if (order >= MAX_ORDER - 2)
949 if (!pfn_valid_within(buddy_pfn))
952 combined_pfn = buddy_pfn & pfn;
953 higher_page = page + (combined_pfn - pfn);
954 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
955 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
957 return pfn_valid_within(buddy_pfn) &&
958 page_is_buddy(higher_page, higher_buddy, order + 1);
962 * Freeing function for a buddy system allocator.
964 * The concept of a buddy system is to maintain direct-mapped table
965 * (containing bit values) for memory blocks of various "orders".
966 * The bottom level table contains the map for the smallest allocatable
967 * units of memory (here, pages), and each level above it describes
968 * pairs of units from the levels below, hence, "buddies".
969 * At a high level, all that happens here is marking the table entry
970 * at the bottom level available, and propagating the changes upward
971 * as necessary, plus some accounting needed to play nicely with other
972 * parts of the VM system.
973 * At each level, we keep a list of pages, which are heads of continuous
974 * free pages of length of (1 << order) and marked with PageBuddy.
975 * Page's order is recorded in page_private(page) field.
976 * So when we are allocating or freeing one, we can derive the state of the
977 * other. That is, if we allocate a small block, and both were
978 * free, the remainder of the region must be split into blocks.
979 * If a block is freed, and its buddy is also free, then this
980 * triggers coalescing into a block of larger size.
985 static inline void __free_one_page(struct page *page,
987 struct zone *zone, unsigned int order,
988 int migratetype, fpi_t fpi_flags)
990 struct capture_control *capc = task_capc(zone);
991 unsigned long buddy_pfn;
992 unsigned long combined_pfn;
993 unsigned int max_order;
997 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
999 VM_BUG_ON(!zone_is_initialized(zone));
1000 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1002 VM_BUG_ON(migratetype == -1);
1003 if (likely(!is_migrate_isolate(migratetype)))
1004 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1006 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1007 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1010 while (order < max_order - 1) {
1011 if (compaction_capture(capc, page, order, migratetype)) {
1012 __mod_zone_freepage_state(zone, -(1 << order),
1016 buddy_pfn = __find_buddy_pfn(pfn, order);
1017 buddy = page + (buddy_pfn - pfn);
1019 if (!pfn_valid_within(buddy_pfn))
1021 if (!page_is_buddy(page, buddy, order))
1024 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1025 * merge with it and move up one order.
1027 if (page_is_guard(buddy))
1028 clear_page_guard(zone, buddy, order, migratetype);
1030 del_page_from_free_list(buddy, zone, order);
1031 combined_pfn = buddy_pfn & pfn;
1032 page = page + (combined_pfn - pfn);
1036 if (max_order < MAX_ORDER) {
1037 /* If we are here, it means order is >= pageblock_order.
1038 * We want to prevent merge between freepages on isolate
1039 * pageblock and normal pageblock. Without this, pageblock
1040 * isolation could cause incorrect freepage or CMA accounting.
1042 * We don't want to hit this code for the more frequent
1043 * low-order merging.
1045 if (unlikely(has_isolate_pageblock(zone))) {
1048 buddy_pfn = __find_buddy_pfn(pfn, order);
1049 buddy = page + (buddy_pfn - pfn);
1050 buddy_mt = get_pageblock_migratetype(buddy);
1052 if (migratetype != buddy_mt
1053 && (is_migrate_isolate(migratetype) ||
1054 is_migrate_isolate(buddy_mt)))
1058 goto continue_merging;
1062 set_buddy_order(page, order);
1064 if (fpi_flags & FPI_TO_TAIL)
1066 else if (is_shuffle_order(order))
1067 to_tail = shuffle_pick_tail();
1069 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1072 add_to_free_list_tail(page, zone, order, migratetype);
1074 add_to_free_list(page, zone, order, migratetype);
1076 /* Notify page reporting subsystem of freed page */
1077 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1078 page_reporting_notify_free(order);
1082 * A bad page could be due to a number of fields. Instead of multiple branches,
1083 * try and check multiple fields with one check. The caller must do a detailed
1084 * check if necessary.
1086 static inline bool page_expected_state(struct page *page,
1087 unsigned long check_flags)
1089 if (unlikely(atomic_read(&page->_mapcount) != -1))
1092 if (unlikely((unsigned long)page->mapping |
1093 page_ref_count(page) |
1095 (unsigned long)page->mem_cgroup |
1097 (page->flags & check_flags)))
1103 static const char *page_bad_reason(struct page *page, unsigned long flags)
1105 const char *bad_reason = NULL;
1107 if (unlikely(atomic_read(&page->_mapcount) != -1))
1108 bad_reason = "nonzero mapcount";
1109 if (unlikely(page->mapping != NULL))
1110 bad_reason = "non-NULL mapping";
1111 if (unlikely(page_ref_count(page) != 0))
1112 bad_reason = "nonzero _refcount";
1113 if (unlikely(page->flags & flags)) {
1114 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1115 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1117 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1120 if (unlikely(page->mem_cgroup))
1121 bad_reason = "page still charged to cgroup";
1126 static void check_free_page_bad(struct page *page)
1129 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1132 static inline int check_free_page(struct page *page)
1134 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1137 /* Something has gone sideways, find it */
1138 check_free_page_bad(page);
1142 static int free_tail_pages_check(struct page *head_page, struct page *page)
1147 * We rely page->lru.next never has bit 0 set, unless the page
1148 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1150 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1152 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1156 switch (page - head_page) {
1158 /* the first tail page: ->mapping may be compound_mapcount() */
1159 if (unlikely(compound_mapcount(page))) {
1160 bad_page(page, "nonzero compound_mapcount");
1166 * the second tail page: ->mapping is
1167 * deferred_list.next -- ignore value.
1171 if (page->mapping != TAIL_MAPPING) {
1172 bad_page(page, "corrupted mapping in tail page");
1177 if (unlikely(!PageTail(page))) {
1178 bad_page(page, "PageTail not set");
1181 if (unlikely(compound_head(page) != head_page)) {
1182 bad_page(page, "compound_head not consistent");
1187 page->mapping = NULL;
1188 clear_compound_head(page);
1192 static void kernel_init_free_pages(struct page *page, int numpages)
1196 /* s390's use of memset() could override KASAN redzones. */
1197 kasan_disable_current();
1198 for (i = 0; i < numpages; i++)
1199 clear_highpage(page + i);
1200 kasan_enable_current();
1203 static __always_inline bool free_pages_prepare(struct page *page,
1204 unsigned int order, bool check_free)
1208 VM_BUG_ON_PAGE(PageTail(page), page);
1210 trace_mm_page_free(page, order);
1212 if (unlikely(PageHWPoison(page)) && !order) {
1214 * Do not let hwpoison pages hit pcplists/buddy
1215 * Untie memcg state and reset page's owner
1217 if (memcg_kmem_enabled() && PageKmemcg(page))
1218 __memcg_kmem_uncharge_page(page, order);
1219 reset_page_owner(page, order);
1224 * Check tail pages before head page information is cleared to
1225 * avoid checking PageCompound for order-0 pages.
1227 if (unlikely(order)) {
1228 bool compound = PageCompound(page);
1231 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1234 ClearPageDoubleMap(page);
1235 for (i = 1; i < (1 << order); i++) {
1237 bad += free_tail_pages_check(page, page + i);
1238 if (unlikely(check_free_page(page + i))) {
1242 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1245 if (PageMappingFlags(page))
1246 page->mapping = NULL;
1247 if (memcg_kmem_enabled() && PageKmemcg(page))
1248 __memcg_kmem_uncharge_page(page, order);
1250 bad += check_free_page(page);
1254 page_cpupid_reset_last(page);
1255 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1256 reset_page_owner(page, order);
1258 if (!PageHighMem(page)) {
1259 debug_check_no_locks_freed(page_address(page),
1260 PAGE_SIZE << order);
1261 debug_check_no_obj_freed(page_address(page),
1262 PAGE_SIZE << order);
1264 if (want_init_on_free())
1265 kernel_init_free_pages(page, 1 << order);
1267 kernel_poison_pages(page, 1 << order, 0);
1269 * arch_free_page() can make the page's contents inaccessible. s390
1270 * does this. So nothing which can access the page's contents should
1271 * happen after this.
1273 arch_free_page(page, order);
1275 if (debug_pagealloc_enabled_static())
1276 kernel_map_pages(page, 1 << order, 0);
1278 kasan_free_nondeferred_pages(page, order);
1283 #ifdef CONFIG_DEBUG_VM
1285 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1286 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1287 * moved from pcp lists to free lists.
1289 static bool free_pcp_prepare(struct page *page)
1291 return free_pages_prepare(page, 0, true);
1294 static bool bulkfree_pcp_prepare(struct page *page)
1296 if (debug_pagealloc_enabled_static())
1297 return check_free_page(page);
1303 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1304 * moving from pcp lists to free list in order to reduce overhead. With
1305 * debug_pagealloc enabled, they are checked also immediately when being freed
1308 static bool free_pcp_prepare(struct page *page)
1310 if (debug_pagealloc_enabled_static())
1311 return free_pages_prepare(page, 0, true);
1313 return free_pages_prepare(page, 0, false);
1316 static bool bulkfree_pcp_prepare(struct page *page)
1318 return check_free_page(page);
1320 #endif /* CONFIG_DEBUG_VM */
1322 static inline void prefetch_buddy(struct page *page)
1324 unsigned long pfn = page_to_pfn(page);
1325 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1326 struct page *buddy = page + (buddy_pfn - pfn);
1332 * Frees a number of pages from the PCP lists
1333 * Assumes all pages on list are in same zone, and of same order.
1334 * count is the number of pages to free.
1336 * If the zone was previously in an "all pages pinned" state then look to
1337 * see if this freeing clears that state.
1339 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1340 * pinned" detection logic.
1342 static void free_pcppages_bulk(struct zone *zone, int count,
1343 struct per_cpu_pages *pcp)
1345 int migratetype = 0;
1347 int prefetch_nr = 0;
1348 bool isolated_pageblocks;
1349 struct page *page, *tmp;
1353 * Ensure proper count is passed which otherwise would stuck in the
1354 * below while (list_empty(list)) loop.
1356 count = min(pcp->count, count);
1358 struct list_head *list;
1361 * Remove pages from lists in a round-robin fashion. A
1362 * batch_free count is maintained that is incremented when an
1363 * empty list is encountered. This is so more pages are freed
1364 * off fuller lists instead of spinning excessively around empty
1369 if (++migratetype == MIGRATE_PCPTYPES)
1371 list = &pcp->lists[migratetype];
1372 } while (list_empty(list));
1374 /* This is the only non-empty list. Free them all. */
1375 if (batch_free == MIGRATE_PCPTYPES)
1379 page = list_last_entry(list, struct page, lru);
1380 /* must delete to avoid corrupting pcp list */
1381 list_del(&page->lru);
1384 if (bulkfree_pcp_prepare(page))
1387 list_add_tail(&page->lru, &head);
1390 * We are going to put the page back to the global
1391 * pool, prefetch its buddy to speed up later access
1392 * under zone->lock. It is believed the overhead of
1393 * an additional test and calculating buddy_pfn here
1394 * can be offset by reduced memory latency later. To
1395 * avoid excessive prefetching due to large count, only
1396 * prefetch buddy for the first pcp->batch nr of pages.
1398 if (prefetch_nr++ < pcp->batch)
1399 prefetch_buddy(page);
1400 } while (--count && --batch_free && !list_empty(list));
1403 spin_lock(&zone->lock);
1404 isolated_pageblocks = has_isolate_pageblock(zone);
1407 * Use safe version since after __free_one_page(),
1408 * page->lru.next will not point to original list.
1410 list_for_each_entry_safe(page, tmp, &head, lru) {
1411 int mt = get_pcppage_migratetype(page);
1412 /* MIGRATE_ISOLATE page should not go to pcplists */
1413 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1414 /* Pageblock could have been isolated meanwhile */
1415 if (unlikely(isolated_pageblocks))
1416 mt = get_pageblock_migratetype(page);
1418 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1419 trace_mm_page_pcpu_drain(page, 0, mt);
1421 spin_unlock(&zone->lock);
1424 static void free_one_page(struct zone *zone,
1425 struct page *page, unsigned long pfn,
1427 int migratetype, fpi_t fpi_flags)
1429 spin_lock(&zone->lock);
1430 if (unlikely(has_isolate_pageblock(zone) ||
1431 is_migrate_isolate(migratetype))) {
1432 migratetype = get_pfnblock_migratetype(page, pfn);
1434 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1435 spin_unlock(&zone->lock);
1438 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1439 unsigned long zone, int nid)
1441 mm_zero_struct_page(page);
1442 set_page_links(page, zone, nid, pfn);
1443 init_page_count(page);
1444 page_mapcount_reset(page);
1445 page_cpupid_reset_last(page);
1446 page_kasan_tag_reset(page);
1448 INIT_LIST_HEAD(&page->lru);
1449 #ifdef WANT_PAGE_VIRTUAL
1450 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1451 if (!is_highmem_idx(zone))
1452 set_page_address(page, __va(pfn << PAGE_SHIFT));
1456 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1457 static void __meminit init_reserved_page(unsigned long pfn)
1462 if (!early_page_uninitialised(pfn))
1465 nid = early_pfn_to_nid(pfn);
1466 pgdat = NODE_DATA(nid);
1468 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1469 struct zone *zone = &pgdat->node_zones[zid];
1471 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1474 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1477 static inline void init_reserved_page(unsigned long pfn)
1480 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1483 * Initialised pages do not have PageReserved set. This function is
1484 * called for each range allocated by the bootmem allocator and
1485 * marks the pages PageReserved. The remaining valid pages are later
1486 * sent to the buddy page allocator.
1488 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1490 unsigned long start_pfn = PFN_DOWN(start);
1491 unsigned long end_pfn = PFN_UP(end);
1493 for (; start_pfn < end_pfn; start_pfn++) {
1494 if (pfn_valid(start_pfn)) {
1495 struct page *page = pfn_to_page(start_pfn);
1497 init_reserved_page(start_pfn);
1499 /* Avoid false-positive PageTail() */
1500 INIT_LIST_HEAD(&page->lru);
1503 * no need for atomic set_bit because the struct
1504 * page is not visible yet so nobody should
1507 __SetPageReserved(page);
1512 static void __free_pages_ok(struct page *page, unsigned int order,
1515 unsigned long flags;
1517 unsigned long pfn = page_to_pfn(page);
1519 if (!free_pages_prepare(page, order, true))
1522 migratetype = get_pfnblock_migratetype(page, pfn);
1523 local_irq_save(flags);
1524 __count_vm_events(PGFREE, 1 << order);
1525 free_one_page(page_zone(page), page, pfn, order, migratetype,
1527 local_irq_restore(flags);
1530 void __free_pages_core(struct page *page, unsigned int order)
1532 unsigned int nr_pages = 1 << order;
1533 struct page *p = page;
1537 * When initializing the memmap, __init_single_page() sets the refcount
1538 * of all pages to 1 ("allocated"/"not free"). We have to set the
1539 * refcount of all involved pages to 0.
1542 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1544 __ClearPageReserved(p);
1545 set_page_count(p, 0);
1547 __ClearPageReserved(p);
1548 set_page_count(p, 0);
1550 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1553 * Bypass PCP and place fresh pages right to the tail, primarily
1554 * relevant for memory onlining.
1556 __free_pages_ok(page, order, FPI_TO_TAIL);
1559 #ifdef CONFIG_NEED_MULTIPLE_NODES
1561 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1563 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1566 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1568 int __meminit __early_pfn_to_nid(unsigned long pfn,
1569 struct mminit_pfnnid_cache *state)
1571 unsigned long start_pfn, end_pfn;
1574 if (state->last_start <= pfn && pfn < state->last_end)
1575 return state->last_nid;
1577 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1578 if (nid != NUMA_NO_NODE) {
1579 state->last_start = start_pfn;
1580 state->last_end = end_pfn;
1581 state->last_nid = nid;
1586 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1588 int __meminit early_pfn_to_nid(unsigned long pfn)
1590 static DEFINE_SPINLOCK(early_pfn_lock);
1593 spin_lock(&early_pfn_lock);
1594 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1596 nid = first_online_node;
1597 spin_unlock(&early_pfn_lock);
1601 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1603 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1606 if (early_page_uninitialised(pfn))
1608 __free_pages_core(page, order);
1612 * Check that the whole (or subset of) a pageblock given by the interval of
1613 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1614 * with the migration of free compaction scanner. The scanners then need to
1615 * use only pfn_valid_within() check for arches that allow holes within
1618 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1620 * It's possible on some configurations to have a setup like node0 node1 node0
1621 * i.e. it's possible that all pages within a zones range of pages do not
1622 * belong to a single zone. We assume that a border between node0 and node1
1623 * can occur within a single pageblock, but not a node0 node1 node0
1624 * interleaving within a single pageblock. It is therefore sufficient to check
1625 * the first and last page of a pageblock and avoid checking each individual
1626 * page in a pageblock.
1628 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1629 unsigned long end_pfn, struct zone *zone)
1631 struct page *start_page;
1632 struct page *end_page;
1634 /* end_pfn is one past the range we are checking */
1637 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1640 start_page = pfn_to_online_page(start_pfn);
1644 if (page_zone(start_page) != zone)
1647 end_page = pfn_to_page(end_pfn);
1649 /* This gives a shorter code than deriving page_zone(end_page) */
1650 if (page_zone_id(start_page) != page_zone_id(end_page))
1656 void set_zone_contiguous(struct zone *zone)
1658 unsigned long block_start_pfn = zone->zone_start_pfn;
1659 unsigned long block_end_pfn;
1661 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1662 for (; block_start_pfn < zone_end_pfn(zone);
1663 block_start_pfn = block_end_pfn,
1664 block_end_pfn += pageblock_nr_pages) {
1666 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1668 if (!__pageblock_pfn_to_page(block_start_pfn,
1669 block_end_pfn, zone))
1674 /* We confirm that there is no hole */
1675 zone->contiguous = true;
1678 void clear_zone_contiguous(struct zone *zone)
1680 zone->contiguous = false;
1683 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1684 static void __init deferred_free_range(unsigned long pfn,
1685 unsigned long nr_pages)
1693 page = pfn_to_page(pfn);
1695 /* Free a large naturally-aligned chunk if possible */
1696 if (nr_pages == pageblock_nr_pages &&
1697 (pfn & (pageblock_nr_pages - 1)) == 0) {
1698 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1699 __free_pages_core(page, pageblock_order);
1703 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1704 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1705 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1706 __free_pages_core(page, 0);
1710 /* Completion tracking for deferred_init_memmap() threads */
1711 static atomic_t pgdat_init_n_undone __initdata;
1712 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1714 static inline void __init pgdat_init_report_one_done(void)
1716 if (atomic_dec_and_test(&pgdat_init_n_undone))
1717 complete(&pgdat_init_all_done_comp);
1721 * Returns true if page needs to be initialized or freed to buddy allocator.
1723 * First we check if pfn is valid on architectures where it is possible to have
1724 * holes within pageblock_nr_pages. On systems where it is not possible, this
1725 * function is optimized out.
1727 * Then, we check if a current large page is valid by only checking the validity
1730 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1732 if (!pfn_valid_within(pfn))
1734 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1740 * Free pages to buddy allocator. Try to free aligned pages in
1741 * pageblock_nr_pages sizes.
1743 static void __init deferred_free_pages(unsigned long pfn,
1744 unsigned long end_pfn)
1746 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1747 unsigned long nr_free = 0;
1749 for (; pfn < end_pfn; pfn++) {
1750 if (!deferred_pfn_valid(pfn)) {
1751 deferred_free_range(pfn - nr_free, nr_free);
1753 } else if (!(pfn & nr_pgmask)) {
1754 deferred_free_range(pfn - nr_free, nr_free);
1760 /* Free the last block of pages to allocator */
1761 deferred_free_range(pfn - nr_free, nr_free);
1765 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1766 * by performing it only once every pageblock_nr_pages.
1767 * Return number of pages initialized.
1769 static unsigned long __init deferred_init_pages(struct zone *zone,
1771 unsigned long end_pfn)
1773 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1774 int nid = zone_to_nid(zone);
1775 unsigned long nr_pages = 0;
1776 int zid = zone_idx(zone);
1777 struct page *page = NULL;
1779 for (; pfn < end_pfn; pfn++) {
1780 if (!deferred_pfn_valid(pfn)) {
1783 } else if (!page || !(pfn & nr_pgmask)) {
1784 page = pfn_to_page(pfn);
1788 __init_single_page(page, pfn, zid, nid);
1795 * This function is meant to pre-load the iterator for the zone init.
1796 * Specifically it walks through the ranges until we are caught up to the
1797 * first_init_pfn value and exits there. If we never encounter the value we
1798 * return false indicating there are no valid ranges left.
1801 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1802 unsigned long *spfn, unsigned long *epfn,
1803 unsigned long first_init_pfn)
1808 * Start out by walking through the ranges in this zone that have
1809 * already been initialized. We don't need to do anything with them
1810 * so we just need to flush them out of the system.
1812 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1813 if (*epfn <= first_init_pfn)
1815 if (*spfn < first_init_pfn)
1816 *spfn = first_init_pfn;
1825 * Initialize and free pages. We do it in two loops: first we initialize
1826 * struct page, then free to buddy allocator, because while we are
1827 * freeing pages we can access pages that are ahead (computing buddy
1828 * page in __free_one_page()).
1830 * In order to try and keep some memory in the cache we have the loop
1831 * broken along max page order boundaries. This way we will not cause
1832 * any issues with the buddy page computation.
1834 static unsigned long __init
1835 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1836 unsigned long *end_pfn)
1838 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1839 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1840 unsigned long nr_pages = 0;
1843 /* First we loop through and initialize the page values */
1844 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1847 if (mo_pfn <= *start_pfn)
1850 t = min(mo_pfn, *end_pfn);
1851 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1853 if (mo_pfn < *end_pfn) {
1854 *start_pfn = mo_pfn;
1859 /* Reset values and now loop through freeing pages as needed */
1862 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1868 t = min(mo_pfn, epfn);
1869 deferred_free_pages(spfn, t);
1879 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1882 unsigned long spfn, epfn;
1883 struct zone *zone = arg;
1886 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1889 * Initialize and free pages in MAX_ORDER sized increments so that we
1890 * can avoid introducing any issues with the buddy allocator.
1892 while (spfn < end_pfn) {
1893 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1898 /* An arch may override for more concurrency. */
1900 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1905 /* Initialise remaining memory on a node */
1906 static int __init deferred_init_memmap(void *data)
1908 pg_data_t *pgdat = data;
1909 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1910 unsigned long spfn = 0, epfn = 0;
1911 unsigned long first_init_pfn, flags;
1912 unsigned long start = jiffies;
1914 int zid, max_threads;
1917 /* Bind memory initialisation thread to a local node if possible */
1918 if (!cpumask_empty(cpumask))
1919 set_cpus_allowed_ptr(current, cpumask);
1921 pgdat_resize_lock(pgdat, &flags);
1922 first_init_pfn = pgdat->first_deferred_pfn;
1923 if (first_init_pfn == ULONG_MAX) {
1924 pgdat_resize_unlock(pgdat, &flags);
1925 pgdat_init_report_one_done();
1929 /* Sanity check boundaries */
1930 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1931 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1932 pgdat->first_deferred_pfn = ULONG_MAX;
1935 * Once we unlock here, the zone cannot be grown anymore, thus if an
1936 * interrupt thread must allocate this early in boot, zone must be
1937 * pre-grown prior to start of deferred page initialization.
1939 pgdat_resize_unlock(pgdat, &flags);
1941 /* Only the highest zone is deferred so find it */
1942 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1943 zone = pgdat->node_zones + zid;
1944 if (first_init_pfn < zone_end_pfn(zone))
1948 /* If the zone is empty somebody else may have cleared out the zone */
1949 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1953 max_threads = deferred_page_init_max_threads(cpumask);
1955 while (spfn < epfn) {
1956 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1957 struct padata_mt_job job = {
1958 .thread_fn = deferred_init_memmap_chunk,
1961 .size = epfn_align - spfn,
1962 .align = PAGES_PER_SECTION,
1963 .min_chunk = PAGES_PER_SECTION,
1964 .max_threads = max_threads,
1967 padata_do_multithreaded(&job);
1968 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1972 /* Sanity check that the next zone really is unpopulated */
1973 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1975 pr_info("node %d deferred pages initialised in %ums\n",
1976 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1978 pgdat_init_report_one_done();
1983 * If this zone has deferred pages, try to grow it by initializing enough
1984 * deferred pages to satisfy the allocation specified by order, rounded up to
1985 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1986 * of SECTION_SIZE bytes by initializing struct pages in increments of
1987 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1989 * Return true when zone was grown, otherwise return false. We return true even
1990 * when we grow less than requested, to let the caller decide if there are
1991 * enough pages to satisfy the allocation.
1993 * Note: We use noinline because this function is needed only during boot, and
1994 * it is called from a __ref function _deferred_grow_zone. This way we are
1995 * making sure that it is not inlined into permanent text section.
1997 static noinline bool __init
1998 deferred_grow_zone(struct zone *zone, unsigned int order)
2000 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2001 pg_data_t *pgdat = zone->zone_pgdat;
2002 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2003 unsigned long spfn, epfn, flags;
2004 unsigned long nr_pages = 0;
2007 /* Only the last zone may have deferred pages */
2008 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2011 pgdat_resize_lock(pgdat, &flags);
2014 * If someone grew this zone while we were waiting for spinlock, return
2015 * true, as there might be enough pages already.
2017 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2018 pgdat_resize_unlock(pgdat, &flags);
2022 /* If the zone is empty somebody else may have cleared out the zone */
2023 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2024 first_deferred_pfn)) {
2025 pgdat->first_deferred_pfn = ULONG_MAX;
2026 pgdat_resize_unlock(pgdat, &flags);
2027 /* Retry only once. */
2028 return first_deferred_pfn != ULONG_MAX;
2032 * Initialize and free pages in MAX_ORDER sized increments so
2033 * that we can avoid introducing any issues with the buddy
2036 while (spfn < epfn) {
2037 /* update our first deferred PFN for this section */
2038 first_deferred_pfn = spfn;
2040 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2041 touch_nmi_watchdog();
2043 /* We should only stop along section boundaries */
2044 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2047 /* If our quota has been met we can stop here */
2048 if (nr_pages >= nr_pages_needed)
2052 pgdat->first_deferred_pfn = spfn;
2053 pgdat_resize_unlock(pgdat, &flags);
2055 return nr_pages > 0;
2059 * deferred_grow_zone() is __init, but it is called from
2060 * get_page_from_freelist() during early boot until deferred_pages permanently
2061 * disables this call. This is why we have refdata wrapper to avoid warning,
2062 * and to ensure that the function body gets unloaded.
2065 _deferred_grow_zone(struct zone *zone, unsigned int order)
2067 return deferred_grow_zone(zone, order);
2070 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2072 void __init page_alloc_init_late(void)
2077 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2079 /* There will be num_node_state(N_MEMORY) threads */
2080 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2081 for_each_node_state(nid, N_MEMORY) {
2082 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2085 /* Block until all are initialised */
2086 wait_for_completion(&pgdat_init_all_done_comp);
2089 * The number of managed pages has changed due to the initialisation
2090 * so the pcpu batch and high limits needs to be updated or the limits
2091 * will be artificially small.
2093 for_each_populated_zone(zone)
2094 zone_pcp_update(zone);
2097 * We initialized the rest of the deferred pages. Permanently disable
2098 * on-demand struct page initialization.
2100 static_branch_disable(&deferred_pages);
2102 /* Reinit limits that are based on free pages after the kernel is up */
2103 files_maxfiles_init();
2106 /* Discard memblock private memory */
2109 for_each_node_state(nid, N_MEMORY)
2110 shuffle_free_memory(NODE_DATA(nid));
2112 for_each_populated_zone(zone)
2113 set_zone_contiguous(zone);
2117 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2118 void __init init_cma_reserved_pageblock(struct page *page)
2120 unsigned i = pageblock_nr_pages;
2121 struct page *p = page;
2124 __ClearPageReserved(p);
2125 set_page_count(p, 0);
2128 set_pageblock_migratetype(page, MIGRATE_CMA);
2130 if (pageblock_order >= MAX_ORDER) {
2131 i = pageblock_nr_pages;
2134 set_page_refcounted(p);
2135 __free_pages(p, MAX_ORDER - 1);
2136 p += MAX_ORDER_NR_PAGES;
2137 } while (i -= MAX_ORDER_NR_PAGES);
2139 set_page_refcounted(page);
2140 __free_pages(page, pageblock_order);
2143 adjust_managed_page_count(page, pageblock_nr_pages);
2148 * The order of subdivision here is critical for the IO subsystem.
2149 * Please do not alter this order without good reasons and regression
2150 * testing. Specifically, as large blocks of memory are subdivided,
2151 * the order in which smaller blocks are delivered depends on the order
2152 * they're subdivided in this function. This is the primary factor
2153 * influencing the order in which pages are delivered to the IO
2154 * subsystem according to empirical testing, and this is also justified
2155 * by considering the behavior of a buddy system containing a single
2156 * large block of memory acted on by a series of small allocations.
2157 * This behavior is a critical factor in sglist merging's success.
2161 static inline void expand(struct zone *zone, struct page *page,
2162 int low, int high, int migratetype)
2164 unsigned long size = 1 << high;
2166 while (high > low) {
2169 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2172 * Mark as guard pages (or page), that will allow to
2173 * merge back to allocator when buddy will be freed.
2174 * Corresponding page table entries will not be touched,
2175 * pages will stay not present in virtual address space
2177 if (set_page_guard(zone, &page[size], high, migratetype))
2180 add_to_free_list(&page[size], zone, high, migratetype);
2181 set_buddy_order(&page[size], high);
2185 static void check_new_page_bad(struct page *page)
2187 if (unlikely(page->flags & __PG_HWPOISON)) {
2188 /* Don't complain about hwpoisoned pages */
2189 page_mapcount_reset(page); /* remove PageBuddy */
2194 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2198 * This page is about to be returned from the page allocator
2200 static inline int check_new_page(struct page *page)
2202 if (likely(page_expected_state(page,
2203 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2206 check_new_page_bad(page);
2210 static inline bool free_pages_prezeroed(void)
2212 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2213 page_poisoning_enabled()) || want_init_on_free();
2216 #ifdef CONFIG_DEBUG_VM
2218 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2219 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2220 * also checked when pcp lists are refilled from the free lists.
2222 static inline bool check_pcp_refill(struct page *page)
2224 if (debug_pagealloc_enabled_static())
2225 return check_new_page(page);
2230 static inline bool check_new_pcp(struct page *page)
2232 return check_new_page(page);
2236 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2237 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2238 * enabled, they are also checked when being allocated from the pcp lists.
2240 static inline bool check_pcp_refill(struct page *page)
2242 return check_new_page(page);
2244 static inline bool check_new_pcp(struct page *page)
2246 if (debug_pagealloc_enabled_static())
2247 return check_new_page(page);
2251 #endif /* CONFIG_DEBUG_VM */
2253 static bool check_new_pages(struct page *page, unsigned int order)
2256 for (i = 0; i < (1 << order); i++) {
2257 struct page *p = page + i;
2259 if (unlikely(check_new_page(p)))
2266 inline void post_alloc_hook(struct page *page, unsigned int order,
2269 set_page_private(page, 0);
2270 set_page_refcounted(page);
2272 arch_alloc_page(page, order);
2273 if (debug_pagealloc_enabled_static())
2274 kernel_map_pages(page, 1 << order, 1);
2275 kasan_alloc_pages(page, order);
2276 kernel_poison_pages(page, 1 << order, 1);
2277 set_page_owner(page, order, gfp_flags);
2280 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2281 unsigned int alloc_flags)
2283 post_alloc_hook(page, order, gfp_flags);
2285 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2286 kernel_init_free_pages(page, 1 << order);
2288 if (order && (gfp_flags & __GFP_COMP))
2289 prep_compound_page(page, order);
2292 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2293 * allocate the page. The expectation is that the caller is taking
2294 * steps that will free more memory. The caller should avoid the page
2295 * being used for !PFMEMALLOC purposes.
2297 if (alloc_flags & ALLOC_NO_WATERMARKS)
2298 set_page_pfmemalloc(page);
2300 clear_page_pfmemalloc(page);
2304 * Go through the free lists for the given migratetype and remove
2305 * the smallest available page from the freelists
2307 static __always_inline
2308 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2311 unsigned int current_order;
2312 struct free_area *area;
2315 /* Find a page of the appropriate size in the preferred list */
2316 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2317 area = &(zone->free_area[current_order]);
2318 page = get_page_from_free_area(area, migratetype);
2321 del_page_from_free_list(page, zone, current_order);
2322 expand(zone, page, order, current_order, migratetype);
2323 set_pcppage_migratetype(page, migratetype);
2332 * This array describes the order lists are fallen back to when
2333 * the free lists for the desirable migrate type are depleted
2335 static int fallbacks[MIGRATE_TYPES][3] = {
2336 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2337 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2338 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2340 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2342 #ifdef CONFIG_MEMORY_ISOLATION
2343 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2348 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2351 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2354 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2355 unsigned int order) { return NULL; }
2359 * Move the free pages in a range to the freelist tail of the requested type.
2360 * Note that start_page and end_pages are not aligned on a pageblock
2361 * boundary. If alignment is required, use move_freepages_block()
2363 static int move_freepages(struct zone *zone,
2364 struct page *start_page, struct page *end_page,
2365 int migratetype, int *num_movable)
2369 int pages_moved = 0;
2371 for (page = start_page; page <= end_page;) {
2372 if (!pfn_valid_within(page_to_pfn(page))) {
2377 if (!PageBuddy(page)) {
2379 * We assume that pages that could be isolated for
2380 * migration are movable. But we don't actually try
2381 * isolating, as that would be expensive.
2384 (PageLRU(page) || __PageMovable(page)))
2391 /* Make sure we are not inadvertently changing nodes */
2392 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2393 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2395 order = buddy_order(page);
2396 move_to_free_list(page, zone, order, migratetype);
2398 pages_moved += 1 << order;
2404 int move_freepages_block(struct zone *zone, struct page *page,
2405 int migratetype, int *num_movable)
2407 unsigned long start_pfn, end_pfn;
2408 struct page *start_page, *end_page;
2413 start_pfn = page_to_pfn(page);
2414 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2415 start_page = pfn_to_page(start_pfn);
2416 end_page = start_page + pageblock_nr_pages - 1;
2417 end_pfn = start_pfn + pageblock_nr_pages - 1;
2419 /* Do not cross zone boundaries */
2420 if (!zone_spans_pfn(zone, start_pfn))
2422 if (!zone_spans_pfn(zone, end_pfn))
2425 return move_freepages(zone, start_page, end_page, migratetype,
2429 static void change_pageblock_range(struct page *pageblock_page,
2430 int start_order, int migratetype)
2432 int nr_pageblocks = 1 << (start_order - pageblock_order);
2434 while (nr_pageblocks--) {
2435 set_pageblock_migratetype(pageblock_page, migratetype);
2436 pageblock_page += pageblock_nr_pages;
2441 * When we are falling back to another migratetype during allocation, try to
2442 * steal extra free pages from the same pageblocks to satisfy further
2443 * allocations, instead of polluting multiple pageblocks.
2445 * If we are stealing a relatively large buddy page, it is likely there will
2446 * be more free pages in the pageblock, so try to steal them all. For
2447 * reclaimable and unmovable allocations, we steal regardless of page size,
2448 * as fragmentation caused by those allocations polluting movable pageblocks
2449 * is worse than movable allocations stealing from unmovable and reclaimable
2452 static bool can_steal_fallback(unsigned int order, int start_mt)
2455 * Leaving this order check is intended, although there is
2456 * relaxed order check in next check. The reason is that
2457 * we can actually steal whole pageblock if this condition met,
2458 * but, below check doesn't guarantee it and that is just heuristic
2459 * so could be changed anytime.
2461 if (order >= pageblock_order)
2464 if (order >= pageblock_order / 2 ||
2465 start_mt == MIGRATE_RECLAIMABLE ||
2466 start_mt == MIGRATE_UNMOVABLE ||
2467 page_group_by_mobility_disabled)
2473 static inline void boost_watermark(struct zone *zone)
2475 unsigned long max_boost;
2477 if (!watermark_boost_factor)
2480 * Don't bother in zones that are unlikely to produce results.
2481 * On small machines, including kdump capture kernels running
2482 * in a small area, boosting the watermark can cause an out of
2483 * memory situation immediately.
2485 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2488 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2489 watermark_boost_factor, 10000);
2492 * high watermark may be uninitialised if fragmentation occurs
2493 * very early in boot so do not boost. We do not fall
2494 * through and boost by pageblock_nr_pages as failing
2495 * allocations that early means that reclaim is not going
2496 * to help and it may even be impossible to reclaim the
2497 * boosted watermark resulting in a hang.
2502 max_boost = max(pageblock_nr_pages, max_boost);
2504 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2509 * This function implements actual steal behaviour. If order is large enough,
2510 * we can steal whole pageblock. If not, we first move freepages in this
2511 * pageblock to our migratetype and determine how many already-allocated pages
2512 * are there in the pageblock with a compatible migratetype. If at least half
2513 * of pages are free or compatible, we can change migratetype of the pageblock
2514 * itself, so pages freed in the future will be put on the correct free list.
2516 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2517 unsigned int alloc_flags, int start_type, bool whole_block)
2519 unsigned int current_order = buddy_order(page);
2520 int free_pages, movable_pages, alike_pages;
2523 old_block_type = get_pageblock_migratetype(page);
2526 * This can happen due to races and we want to prevent broken
2527 * highatomic accounting.
2529 if (is_migrate_highatomic(old_block_type))
2532 /* Take ownership for orders >= pageblock_order */
2533 if (current_order >= pageblock_order) {
2534 change_pageblock_range(page, current_order, start_type);
2539 * Boost watermarks to increase reclaim pressure to reduce the
2540 * likelihood of future fallbacks. Wake kswapd now as the node
2541 * may be balanced overall and kswapd will not wake naturally.
2543 boost_watermark(zone);
2544 if (alloc_flags & ALLOC_KSWAPD)
2545 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2547 /* We are not allowed to try stealing from the whole block */
2551 free_pages = move_freepages_block(zone, page, start_type,
2554 * Determine how many pages are compatible with our allocation.
2555 * For movable allocation, it's the number of movable pages which
2556 * we just obtained. For other types it's a bit more tricky.
2558 if (start_type == MIGRATE_MOVABLE) {
2559 alike_pages = movable_pages;
2562 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2563 * to MOVABLE pageblock, consider all non-movable pages as
2564 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2565 * vice versa, be conservative since we can't distinguish the
2566 * exact migratetype of non-movable pages.
2568 if (old_block_type == MIGRATE_MOVABLE)
2569 alike_pages = pageblock_nr_pages
2570 - (free_pages + movable_pages);
2575 /* moving whole block can fail due to zone boundary conditions */
2580 * If a sufficient number of pages in the block are either free or of
2581 * comparable migratability as our allocation, claim the whole block.
2583 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2584 page_group_by_mobility_disabled)
2585 set_pageblock_migratetype(page, start_type);
2590 move_to_free_list(page, zone, current_order, start_type);
2594 * Check whether there is a suitable fallback freepage with requested order.
2595 * If only_stealable is true, this function returns fallback_mt only if
2596 * we can steal other freepages all together. This would help to reduce
2597 * fragmentation due to mixed migratetype pages in one pageblock.
2599 int find_suitable_fallback(struct free_area *area, unsigned int order,
2600 int migratetype, bool only_stealable, bool *can_steal)
2605 if (area->nr_free == 0)
2610 fallback_mt = fallbacks[migratetype][i];
2611 if (fallback_mt == MIGRATE_TYPES)
2614 if (free_area_empty(area, fallback_mt))
2617 if (can_steal_fallback(order, migratetype))
2620 if (!only_stealable)
2631 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2632 * there are no empty page blocks that contain a page with a suitable order
2634 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2635 unsigned int alloc_order)
2638 unsigned long max_managed, flags;
2641 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2642 * Check is race-prone but harmless.
2644 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2645 if (zone->nr_reserved_highatomic >= max_managed)
2648 spin_lock_irqsave(&zone->lock, flags);
2650 /* Recheck the nr_reserved_highatomic limit under the lock */
2651 if (zone->nr_reserved_highatomic >= max_managed)
2655 mt = get_pageblock_migratetype(page);
2656 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2657 && !is_migrate_cma(mt)) {
2658 zone->nr_reserved_highatomic += pageblock_nr_pages;
2659 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2660 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2664 spin_unlock_irqrestore(&zone->lock, flags);
2668 * Used when an allocation is about to fail under memory pressure. This
2669 * potentially hurts the reliability of high-order allocations when under
2670 * intense memory pressure but failed atomic allocations should be easier
2671 * to recover from than an OOM.
2673 * If @force is true, try to unreserve a pageblock even though highatomic
2674 * pageblock is exhausted.
2676 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2679 struct zonelist *zonelist = ac->zonelist;
2680 unsigned long flags;
2687 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2690 * Preserve at least one pageblock unless memory pressure
2693 if (!force && zone->nr_reserved_highatomic <=
2697 spin_lock_irqsave(&zone->lock, flags);
2698 for (order = 0; order < MAX_ORDER; order++) {
2699 struct free_area *area = &(zone->free_area[order]);
2701 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2706 * In page freeing path, migratetype change is racy so
2707 * we can counter several free pages in a pageblock
2708 * in this loop althoug we changed the pageblock type
2709 * from highatomic to ac->migratetype. So we should
2710 * adjust the count once.
2712 if (is_migrate_highatomic_page(page)) {
2714 * It should never happen but changes to
2715 * locking could inadvertently allow a per-cpu
2716 * drain to add pages to MIGRATE_HIGHATOMIC
2717 * while unreserving so be safe and watch for
2720 zone->nr_reserved_highatomic -= min(
2722 zone->nr_reserved_highatomic);
2726 * Convert to ac->migratetype and avoid the normal
2727 * pageblock stealing heuristics. Minimally, the caller
2728 * is doing the work and needs the pages. More
2729 * importantly, if the block was always converted to
2730 * MIGRATE_UNMOVABLE or another type then the number
2731 * of pageblocks that cannot be completely freed
2734 set_pageblock_migratetype(page, ac->migratetype);
2735 ret = move_freepages_block(zone, page, ac->migratetype,
2738 spin_unlock_irqrestore(&zone->lock, flags);
2742 spin_unlock_irqrestore(&zone->lock, flags);
2749 * Try finding a free buddy page on the fallback list and put it on the free
2750 * list of requested migratetype, possibly along with other pages from the same
2751 * block, depending on fragmentation avoidance heuristics. Returns true if
2752 * fallback was found so that __rmqueue_smallest() can grab it.
2754 * The use of signed ints for order and current_order is a deliberate
2755 * deviation from the rest of this file, to make the for loop
2756 * condition simpler.
2758 static __always_inline bool
2759 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2760 unsigned int alloc_flags)
2762 struct free_area *area;
2764 int min_order = order;
2770 * Do not steal pages from freelists belonging to other pageblocks
2771 * i.e. orders < pageblock_order. If there are no local zones free,
2772 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2774 if (alloc_flags & ALLOC_NOFRAGMENT)
2775 min_order = pageblock_order;
2778 * Find the largest available free page in the other list. This roughly
2779 * approximates finding the pageblock with the most free pages, which
2780 * would be too costly to do exactly.
2782 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2784 area = &(zone->free_area[current_order]);
2785 fallback_mt = find_suitable_fallback(area, current_order,
2786 start_migratetype, false, &can_steal);
2787 if (fallback_mt == -1)
2791 * We cannot steal all free pages from the pageblock and the
2792 * requested migratetype is movable. In that case it's better to
2793 * steal and split the smallest available page instead of the
2794 * largest available page, because even if the next movable
2795 * allocation falls back into a different pageblock than this
2796 * one, it won't cause permanent fragmentation.
2798 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2799 && current_order > order)
2808 for (current_order = order; current_order < MAX_ORDER;
2810 area = &(zone->free_area[current_order]);
2811 fallback_mt = find_suitable_fallback(area, current_order,
2812 start_migratetype, false, &can_steal);
2813 if (fallback_mt != -1)
2818 * This should not happen - we already found a suitable fallback
2819 * when looking for the largest page.
2821 VM_BUG_ON(current_order == MAX_ORDER);
2824 page = get_page_from_free_area(area, fallback_mt);
2826 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2829 trace_mm_page_alloc_extfrag(page, order, current_order,
2830 start_migratetype, fallback_mt);
2837 * Do the hard work of removing an element from the buddy allocator.
2838 * Call me with the zone->lock already held.
2840 static __always_inline struct page *
2841 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2842 unsigned int alloc_flags)
2848 * Balance movable allocations between regular and CMA areas by
2849 * allocating from CMA when over half of the zone's free memory
2850 * is in the CMA area.
2852 if (alloc_flags & ALLOC_CMA &&
2853 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2854 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2855 page = __rmqueue_cma_fallback(zone, order);
2861 page = __rmqueue_smallest(zone, order, migratetype);
2862 if (unlikely(!page)) {
2863 if (alloc_flags & ALLOC_CMA)
2864 page = __rmqueue_cma_fallback(zone, order);
2866 if (!page && __rmqueue_fallback(zone, order, migratetype,
2871 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2876 * Obtain a specified number of elements from the buddy allocator, all under
2877 * a single hold of the lock, for efficiency. Add them to the supplied list.
2878 * Returns the number of new pages which were placed at *list.
2880 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2881 unsigned long count, struct list_head *list,
2882 int migratetype, unsigned int alloc_flags)
2886 spin_lock(&zone->lock);
2887 for (i = 0; i < count; ++i) {
2888 struct page *page = __rmqueue(zone, order, migratetype,
2890 if (unlikely(page == NULL))
2893 if (unlikely(check_pcp_refill(page)))
2897 * Split buddy pages returned by expand() are received here in
2898 * physical page order. The page is added to the tail of
2899 * caller's list. From the callers perspective, the linked list
2900 * is ordered by page number under some conditions. This is
2901 * useful for IO devices that can forward direction from the
2902 * head, thus also in the physical page order. This is useful
2903 * for IO devices that can merge IO requests if the physical
2904 * pages are ordered properly.
2906 list_add_tail(&page->lru, list);
2908 if (is_migrate_cma(get_pcppage_migratetype(page)))
2909 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2914 * i pages were removed from the buddy list even if some leak due
2915 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2916 * on i. Do not confuse with 'alloced' which is the number of
2917 * pages added to the pcp list.
2919 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2920 spin_unlock(&zone->lock);
2926 * Called from the vmstat counter updater to drain pagesets of this
2927 * currently executing processor on remote nodes after they have
2930 * Note that this function must be called with the thread pinned to
2931 * a single processor.
2933 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2935 unsigned long flags;
2936 int to_drain, batch;
2938 local_irq_save(flags);
2939 batch = READ_ONCE(pcp->batch);
2940 to_drain = min(pcp->count, batch);
2942 free_pcppages_bulk(zone, to_drain, pcp);
2943 local_irq_restore(flags);
2948 * Drain pcplists of the indicated processor and zone.
2950 * The processor must either be the current processor and the
2951 * thread pinned to the current processor or a processor that
2954 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2956 unsigned long flags;
2957 struct per_cpu_pageset *pset;
2958 struct per_cpu_pages *pcp;
2960 local_irq_save(flags);
2961 pset = per_cpu_ptr(zone->pageset, cpu);
2965 free_pcppages_bulk(zone, pcp->count, pcp);
2966 local_irq_restore(flags);
2970 * Drain pcplists of all zones on the indicated processor.
2972 * The processor must either be the current processor and the
2973 * thread pinned to the current processor or a processor that
2976 static void drain_pages(unsigned int cpu)
2980 for_each_populated_zone(zone) {
2981 drain_pages_zone(cpu, zone);
2986 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2988 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2989 * the single zone's pages.
2991 void drain_local_pages(struct zone *zone)
2993 int cpu = smp_processor_id();
2996 drain_pages_zone(cpu, zone);
3001 static void drain_local_pages_wq(struct work_struct *work)
3003 struct pcpu_drain *drain;
3005 drain = container_of(work, struct pcpu_drain, work);
3008 * drain_all_pages doesn't use proper cpu hotplug protection so
3009 * we can race with cpu offline when the WQ can move this from
3010 * a cpu pinned worker to an unbound one. We can operate on a different
3011 * cpu which is allright but we also have to make sure to not move to
3015 drain_local_pages(drain->zone);
3020 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3022 * When zone parameter is non-NULL, spill just the single zone's pages.
3024 * Note that this can be extremely slow as the draining happens in a workqueue.
3026 void drain_all_pages(struct zone *zone)
3031 * Allocate in the BSS so we wont require allocation in
3032 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3034 static cpumask_t cpus_with_pcps;
3037 * Make sure nobody triggers this path before mm_percpu_wq is fully
3040 if (WARN_ON_ONCE(!mm_percpu_wq))
3044 * Do not drain if one is already in progress unless it's specific to
3045 * a zone. Such callers are primarily CMA and memory hotplug and need
3046 * the drain to be complete when the call returns.
3048 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3051 mutex_lock(&pcpu_drain_mutex);
3055 * We don't care about racing with CPU hotplug event
3056 * as offline notification will cause the notified
3057 * cpu to drain that CPU pcps and on_each_cpu_mask
3058 * disables preemption as part of its processing
3060 for_each_online_cpu(cpu) {
3061 struct per_cpu_pageset *pcp;
3063 bool has_pcps = false;
3066 pcp = per_cpu_ptr(zone->pageset, cpu);
3070 for_each_populated_zone(z) {
3071 pcp = per_cpu_ptr(z->pageset, cpu);
3072 if (pcp->pcp.count) {
3080 cpumask_set_cpu(cpu, &cpus_with_pcps);
3082 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3085 for_each_cpu(cpu, &cpus_with_pcps) {
3086 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3089 INIT_WORK(&drain->work, drain_local_pages_wq);
3090 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3092 for_each_cpu(cpu, &cpus_with_pcps)
3093 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3095 mutex_unlock(&pcpu_drain_mutex);
3098 #ifdef CONFIG_HIBERNATION
3101 * Touch the watchdog for every WD_PAGE_COUNT pages.
3103 #define WD_PAGE_COUNT (128*1024)
3105 void mark_free_pages(struct zone *zone)
3107 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3108 unsigned long flags;
3109 unsigned int order, t;
3112 if (zone_is_empty(zone))
3115 spin_lock_irqsave(&zone->lock, flags);
3117 max_zone_pfn = zone_end_pfn(zone);
3118 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3119 if (pfn_valid(pfn)) {
3120 page = pfn_to_page(pfn);
3122 if (!--page_count) {
3123 touch_nmi_watchdog();
3124 page_count = WD_PAGE_COUNT;
3127 if (page_zone(page) != zone)
3130 if (!swsusp_page_is_forbidden(page))
3131 swsusp_unset_page_free(page);
3134 for_each_migratetype_order(order, t) {
3135 list_for_each_entry(page,
3136 &zone->free_area[order].free_list[t], lru) {
3139 pfn = page_to_pfn(page);
3140 for (i = 0; i < (1UL << order); i++) {
3141 if (!--page_count) {
3142 touch_nmi_watchdog();
3143 page_count = WD_PAGE_COUNT;
3145 swsusp_set_page_free(pfn_to_page(pfn + i));
3149 spin_unlock_irqrestore(&zone->lock, flags);
3151 #endif /* CONFIG_PM */
3153 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3157 if (!free_pcp_prepare(page))
3160 migratetype = get_pfnblock_migratetype(page, pfn);
3161 set_pcppage_migratetype(page, migratetype);
3165 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3167 struct zone *zone = page_zone(page);
3168 struct per_cpu_pages *pcp;
3171 migratetype = get_pcppage_migratetype(page);
3172 __count_vm_event(PGFREE);
3175 * We only track unmovable, reclaimable and movable on pcp lists.
3176 * Free ISOLATE pages back to the allocator because they are being
3177 * offlined but treat HIGHATOMIC as movable pages so we can get those
3178 * areas back if necessary. Otherwise, we may have to free
3179 * excessively into the page allocator
3181 if (migratetype >= MIGRATE_PCPTYPES) {
3182 if (unlikely(is_migrate_isolate(migratetype))) {
3183 free_one_page(zone, page, pfn, 0, migratetype,
3187 migratetype = MIGRATE_MOVABLE;
3190 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3191 list_add(&page->lru, &pcp->lists[migratetype]);
3193 if (pcp->count >= pcp->high) {
3194 unsigned long batch = READ_ONCE(pcp->batch);
3195 free_pcppages_bulk(zone, batch, pcp);
3200 * Free a 0-order page
3202 void free_unref_page(struct page *page)
3204 unsigned long flags;
3205 unsigned long pfn = page_to_pfn(page);
3207 if (!free_unref_page_prepare(page, pfn))
3210 local_irq_save(flags);
3211 free_unref_page_commit(page, pfn);
3212 local_irq_restore(flags);
3216 * Free a list of 0-order pages
3218 void free_unref_page_list(struct list_head *list)
3220 struct page *page, *next;
3221 unsigned long flags, pfn;
3222 int batch_count = 0;
3224 /* Prepare pages for freeing */
3225 list_for_each_entry_safe(page, next, list, lru) {
3226 pfn = page_to_pfn(page);
3227 if (!free_unref_page_prepare(page, pfn))
3228 list_del(&page->lru);
3229 set_page_private(page, pfn);
3232 local_irq_save(flags);
3233 list_for_each_entry_safe(page, next, list, lru) {
3234 unsigned long pfn = page_private(page);
3236 set_page_private(page, 0);
3237 trace_mm_page_free_batched(page);
3238 free_unref_page_commit(page, pfn);
3241 * Guard against excessive IRQ disabled times when we get
3242 * a large list of pages to free.
3244 if (++batch_count == SWAP_CLUSTER_MAX) {
3245 local_irq_restore(flags);
3247 local_irq_save(flags);
3250 local_irq_restore(flags);
3254 * split_page takes a non-compound higher-order page, and splits it into
3255 * n (1<<order) sub-pages: page[0..n]
3256 * Each sub-page must be freed individually.
3258 * Note: this is probably too low level an operation for use in drivers.
3259 * Please consult with lkml before using this in your driver.
3261 void split_page(struct page *page, unsigned int order)
3265 VM_BUG_ON_PAGE(PageCompound(page), page);
3266 VM_BUG_ON_PAGE(!page_count(page), page);
3268 for (i = 1; i < (1 << order); i++)
3269 set_page_refcounted(page + i);
3270 split_page_owner(page, 1 << order);
3272 EXPORT_SYMBOL_GPL(split_page);
3274 int __isolate_free_page(struct page *page, unsigned int order)
3276 unsigned long watermark;
3280 BUG_ON(!PageBuddy(page));
3282 zone = page_zone(page);
3283 mt = get_pageblock_migratetype(page);
3285 if (!is_migrate_isolate(mt)) {
3287 * Obey watermarks as if the page was being allocated. We can
3288 * emulate a high-order watermark check with a raised order-0
3289 * watermark, because we already know our high-order page
3292 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3293 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3296 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3299 /* Remove page from free list */
3301 del_page_from_free_list(page, zone, order);
3304 * Set the pageblock if the isolated page is at least half of a
3307 if (order >= pageblock_order - 1) {
3308 struct page *endpage = page + (1 << order) - 1;
3309 for (; page < endpage; page += pageblock_nr_pages) {
3310 int mt = get_pageblock_migratetype(page);
3311 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3312 && !is_migrate_highatomic(mt))
3313 set_pageblock_migratetype(page,
3319 return 1UL << order;
3323 * __putback_isolated_page - Return a now-isolated page back where we got it
3324 * @page: Page that was isolated
3325 * @order: Order of the isolated page
3326 * @mt: The page's pageblock's migratetype
3328 * This function is meant to return a page pulled from the free lists via
3329 * __isolate_free_page back to the free lists they were pulled from.
3331 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3333 struct zone *zone = page_zone(page);
3335 /* zone lock should be held when this function is called */
3336 lockdep_assert_held(&zone->lock);
3338 /* Return isolated page to tail of freelist. */
3339 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3340 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3344 * Update NUMA hit/miss statistics
3346 * Must be called with interrupts disabled.
3348 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3351 enum numa_stat_item local_stat = NUMA_LOCAL;
3353 /* skip numa counters update if numa stats is disabled */
3354 if (!static_branch_likely(&vm_numa_stat_key))
3357 if (zone_to_nid(z) != numa_node_id())
3358 local_stat = NUMA_OTHER;
3360 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3361 __inc_numa_state(z, NUMA_HIT);
3363 __inc_numa_state(z, NUMA_MISS);
3364 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3366 __inc_numa_state(z, local_stat);
3370 /* Remove page from the per-cpu list, caller must protect the list */
3371 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3372 unsigned int alloc_flags,
3373 struct per_cpu_pages *pcp,
3374 struct list_head *list)
3379 if (list_empty(list)) {
3380 pcp->count += rmqueue_bulk(zone, 0,
3382 migratetype, alloc_flags);
3383 if (unlikely(list_empty(list)))
3387 page = list_first_entry(list, struct page, lru);
3388 list_del(&page->lru);
3390 } while (check_new_pcp(page));
3395 /* Lock and remove page from the per-cpu list */
3396 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3397 struct zone *zone, gfp_t gfp_flags,
3398 int migratetype, unsigned int alloc_flags)
3400 struct per_cpu_pages *pcp;
3401 struct list_head *list;
3403 unsigned long flags;
3405 local_irq_save(flags);
3406 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3407 list = &pcp->lists[migratetype];
3408 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3410 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3411 zone_statistics(preferred_zone, zone);
3413 local_irq_restore(flags);
3418 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3421 struct page *rmqueue(struct zone *preferred_zone,
3422 struct zone *zone, unsigned int order,
3423 gfp_t gfp_flags, unsigned int alloc_flags,
3426 unsigned long flags;
3429 if (likely(order == 0)) {
3431 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3432 * we need to skip it when CMA area isn't allowed.
3434 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3435 migratetype != MIGRATE_MOVABLE) {
3436 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3437 migratetype, alloc_flags);
3443 * We most definitely don't want callers attempting to
3444 * allocate greater than order-1 page units with __GFP_NOFAIL.
3446 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3447 spin_lock_irqsave(&zone->lock, flags);
3452 * order-0 request can reach here when the pcplist is skipped
3453 * due to non-CMA allocation context. HIGHATOMIC area is
3454 * reserved for high-order atomic allocation, so order-0
3455 * request should skip it.
3457 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3458 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3460 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3463 page = __rmqueue(zone, order, migratetype, alloc_flags);
3464 } while (page && check_new_pages(page, order));
3465 spin_unlock(&zone->lock);
3468 __mod_zone_freepage_state(zone, -(1 << order),
3469 get_pcppage_migratetype(page));
3471 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3472 zone_statistics(preferred_zone, zone);
3473 local_irq_restore(flags);
3476 /* Separate test+clear to avoid unnecessary atomics */
3477 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3478 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3479 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3482 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3486 local_irq_restore(flags);
3490 #ifdef CONFIG_FAIL_PAGE_ALLOC
3493 struct fault_attr attr;
3495 bool ignore_gfp_highmem;
3496 bool ignore_gfp_reclaim;
3498 } fail_page_alloc = {
3499 .attr = FAULT_ATTR_INITIALIZER,
3500 .ignore_gfp_reclaim = true,
3501 .ignore_gfp_highmem = true,
3505 static int __init setup_fail_page_alloc(char *str)
3507 return setup_fault_attr(&fail_page_alloc.attr, str);
3509 __setup("fail_page_alloc=", setup_fail_page_alloc);
3511 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3513 if (order < fail_page_alloc.min_order)
3515 if (gfp_mask & __GFP_NOFAIL)
3517 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3519 if (fail_page_alloc.ignore_gfp_reclaim &&
3520 (gfp_mask & __GFP_DIRECT_RECLAIM))
3523 return should_fail(&fail_page_alloc.attr, 1 << order);
3526 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3528 static int __init fail_page_alloc_debugfs(void)
3530 umode_t mode = S_IFREG | 0600;
3533 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3534 &fail_page_alloc.attr);
3536 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3537 &fail_page_alloc.ignore_gfp_reclaim);
3538 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3539 &fail_page_alloc.ignore_gfp_highmem);
3540 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3545 late_initcall(fail_page_alloc_debugfs);
3547 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3549 #else /* CONFIG_FAIL_PAGE_ALLOC */
3551 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3556 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3558 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3560 return __should_fail_alloc_page(gfp_mask, order);
3562 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3564 static inline long __zone_watermark_unusable_free(struct zone *z,
3565 unsigned int order, unsigned int alloc_flags)
3567 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3568 long unusable_free = (1 << order) - 1;
3571 * If the caller does not have rights to ALLOC_HARDER then subtract
3572 * the high-atomic reserves. This will over-estimate the size of the
3573 * atomic reserve but it avoids a search.
3575 if (likely(!alloc_harder))
3576 unusable_free += z->nr_reserved_highatomic;
3579 /* If allocation can't use CMA areas don't use free CMA pages */
3580 if (!(alloc_flags & ALLOC_CMA))
3581 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3584 return unusable_free;
3588 * Return true if free base pages are above 'mark'. For high-order checks it
3589 * will return true of the order-0 watermark is reached and there is at least
3590 * one free page of a suitable size. Checking now avoids taking the zone lock
3591 * to check in the allocation paths if no pages are free.
3593 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3594 int highest_zoneidx, unsigned int alloc_flags,
3599 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3601 /* free_pages may go negative - that's OK */
3602 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3604 if (alloc_flags & ALLOC_HIGH)
3607 if (unlikely(alloc_harder)) {
3609 * OOM victims can try even harder than normal ALLOC_HARDER
3610 * users on the grounds that it's definitely going to be in
3611 * the exit path shortly and free memory. Any allocation it
3612 * makes during the free path will be small and short-lived.
3614 if (alloc_flags & ALLOC_OOM)
3621 * Check watermarks for an order-0 allocation request. If these
3622 * are not met, then a high-order request also cannot go ahead
3623 * even if a suitable page happened to be free.
3625 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3628 /* If this is an order-0 request then the watermark is fine */
3632 /* For a high-order request, check at least one suitable page is free */
3633 for (o = order; o < MAX_ORDER; o++) {
3634 struct free_area *area = &z->free_area[o];
3640 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3641 if (!free_area_empty(area, mt))
3646 if ((alloc_flags & ALLOC_CMA) &&
3647 !free_area_empty(area, MIGRATE_CMA)) {
3651 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3657 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3658 int highest_zoneidx, unsigned int alloc_flags)
3660 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3661 zone_page_state(z, NR_FREE_PAGES));
3664 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3665 unsigned long mark, int highest_zoneidx,
3666 unsigned int alloc_flags, gfp_t gfp_mask)
3670 free_pages = zone_page_state(z, NR_FREE_PAGES);
3673 * Fast check for order-0 only. If this fails then the reserves
3674 * need to be calculated.
3679 fast_free = free_pages;
3680 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3681 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3685 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3689 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3690 * when checking the min watermark. The min watermark is the
3691 * point where boosting is ignored so that kswapd is woken up
3692 * when below the low watermark.
3694 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3695 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3696 mark = z->_watermark[WMARK_MIN];
3697 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3698 alloc_flags, free_pages);
3704 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3705 unsigned long mark, int highest_zoneidx)
3707 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3709 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3710 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3712 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3717 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3719 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3720 node_reclaim_distance;
3722 #else /* CONFIG_NUMA */
3723 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3727 #endif /* CONFIG_NUMA */
3730 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3731 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3732 * premature use of a lower zone may cause lowmem pressure problems that
3733 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3734 * probably too small. It only makes sense to spread allocations to avoid
3735 * fragmentation between the Normal and DMA32 zones.
3737 static inline unsigned int
3738 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3740 unsigned int alloc_flags;
3743 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3746 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3748 #ifdef CONFIG_ZONE_DMA32
3752 if (zone_idx(zone) != ZONE_NORMAL)
3756 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3757 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3758 * on UMA that if Normal is populated then so is DMA32.
3760 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3761 if (nr_online_nodes > 1 && !populated_zone(--zone))
3764 alloc_flags |= ALLOC_NOFRAGMENT;
3765 #endif /* CONFIG_ZONE_DMA32 */
3769 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3770 unsigned int alloc_flags)
3773 unsigned int pflags = current->flags;
3775 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3776 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3777 alloc_flags |= ALLOC_CMA;
3784 * get_page_from_freelist goes through the zonelist trying to allocate
3787 static struct page *
3788 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3789 const struct alloc_context *ac)
3793 struct pglist_data *last_pgdat_dirty_limit = NULL;
3798 * Scan zonelist, looking for a zone with enough free.
3799 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3801 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3802 z = ac->preferred_zoneref;
3803 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3808 if (cpusets_enabled() &&
3809 (alloc_flags & ALLOC_CPUSET) &&
3810 !__cpuset_zone_allowed(zone, gfp_mask))
3813 * When allocating a page cache page for writing, we
3814 * want to get it from a node that is within its dirty
3815 * limit, such that no single node holds more than its
3816 * proportional share of globally allowed dirty pages.
3817 * The dirty limits take into account the node's
3818 * lowmem reserves and high watermark so that kswapd
3819 * should be able to balance it without having to
3820 * write pages from its LRU list.
3822 * XXX: For now, allow allocations to potentially
3823 * exceed the per-node dirty limit in the slowpath
3824 * (spread_dirty_pages unset) before going into reclaim,
3825 * which is important when on a NUMA setup the allowed
3826 * nodes are together not big enough to reach the
3827 * global limit. The proper fix for these situations
3828 * will require awareness of nodes in the
3829 * dirty-throttling and the flusher threads.
3831 if (ac->spread_dirty_pages) {
3832 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3835 if (!node_dirty_ok(zone->zone_pgdat)) {
3836 last_pgdat_dirty_limit = zone->zone_pgdat;
3841 if (no_fallback && nr_online_nodes > 1 &&
3842 zone != ac->preferred_zoneref->zone) {
3846 * If moving to a remote node, retry but allow
3847 * fragmenting fallbacks. Locality is more important
3848 * than fragmentation avoidance.
3850 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3851 if (zone_to_nid(zone) != local_nid) {
3852 alloc_flags &= ~ALLOC_NOFRAGMENT;
3857 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3858 if (!zone_watermark_fast(zone, order, mark,
3859 ac->highest_zoneidx, alloc_flags,
3863 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3865 * Watermark failed for this zone, but see if we can
3866 * grow this zone if it contains deferred pages.
3868 if (static_branch_unlikely(&deferred_pages)) {
3869 if (_deferred_grow_zone(zone, order))
3873 /* Checked here to keep the fast path fast */
3874 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3875 if (alloc_flags & ALLOC_NO_WATERMARKS)
3878 if (node_reclaim_mode == 0 ||
3879 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3882 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3884 case NODE_RECLAIM_NOSCAN:
3887 case NODE_RECLAIM_FULL:
3888 /* scanned but unreclaimable */
3891 /* did we reclaim enough */
3892 if (zone_watermark_ok(zone, order, mark,
3893 ac->highest_zoneidx, alloc_flags))
3901 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3902 gfp_mask, alloc_flags, ac->migratetype);
3904 prep_new_page(page, order, gfp_mask, alloc_flags);
3907 * If this is a high-order atomic allocation then check
3908 * if the pageblock should be reserved for the future
3910 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3911 reserve_highatomic_pageblock(page, zone, order);
3915 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3916 /* Try again if zone has deferred pages */
3917 if (static_branch_unlikely(&deferred_pages)) {
3918 if (_deferred_grow_zone(zone, order))
3926 * It's possible on a UMA machine to get through all zones that are
3927 * fragmented. If avoiding fragmentation, reset and try again.
3930 alloc_flags &= ~ALLOC_NOFRAGMENT;
3937 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3939 unsigned int filter = SHOW_MEM_FILTER_NODES;
3942 * This documents exceptions given to allocations in certain
3943 * contexts that are allowed to allocate outside current's set
3946 if (!(gfp_mask & __GFP_NOMEMALLOC))
3947 if (tsk_is_oom_victim(current) ||
3948 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3949 filter &= ~SHOW_MEM_FILTER_NODES;
3950 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3951 filter &= ~SHOW_MEM_FILTER_NODES;
3953 show_mem(filter, nodemask);
3956 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3958 struct va_format vaf;
3960 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3962 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3965 va_start(args, fmt);
3968 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3969 current->comm, &vaf, gfp_mask, &gfp_mask,
3970 nodemask_pr_args(nodemask));
3973 cpuset_print_current_mems_allowed();
3976 warn_alloc_show_mem(gfp_mask, nodemask);
3979 static inline struct page *
3980 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3981 unsigned int alloc_flags,
3982 const struct alloc_context *ac)
3986 page = get_page_from_freelist(gfp_mask, order,
3987 alloc_flags|ALLOC_CPUSET, ac);
3989 * fallback to ignore cpuset restriction if our nodes
3993 page = get_page_from_freelist(gfp_mask, order,
3999 static inline struct page *
4000 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4001 const struct alloc_context *ac, unsigned long *did_some_progress)
4003 struct oom_control oc = {
4004 .zonelist = ac->zonelist,
4005 .nodemask = ac->nodemask,
4007 .gfp_mask = gfp_mask,
4012 *did_some_progress = 0;
4015 * Acquire the oom lock. If that fails, somebody else is
4016 * making progress for us.
4018 if (!mutex_trylock(&oom_lock)) {
4019 *did_some_progress = 1;
4020 schedule_timeout_uninterruptible(1);
4025 * Go through the zonelist yet one more time, keep very high watermark
4026 * here, this is only to catch a parallel oom killing, we must fail if
4027 * we're still under heavy pressure. But make sure that this reclaim
4028 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4029 * allocation which will never fail due to oom_lock already held.
4031 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4032 ~__GFP_DIRECT_RECLAIM, order,
4033 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4037 /* Coredumps can quickly deplete all memory reserves */
4038 if (current->flags & PF_DUMPCORE)
4040 /* The OOM killer will not help higher order allocs */
4041 if (order > PAGE_ALLOC_COSTLY_ORDER)
4044 * We have already exhausted all our reclaim opportunities without any
4045 * success so it is time to admit defeat. We will skip the OOM killer
4046 * because it is very likely that the caller has a more reasonable
4047 * fallback than shooting a random task.
4049 * The OOM killer may not free memory on a specific node.
4051 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4053 /* The OOM killer does not needlessly kill tasks for lowmem */
4054 if (ac->highest_zoneidx < ZONE_NORMAL)
4056 if (pm_suspended_storage())
4059 * XXX: GFP_NOFS allocations should rather fail than rely on
4060 * other request to make a forward progress.
4061 * We are in an unfortunate situation where out_of_memory cannot
4062 * do much for this context but let's try it to at least get
4063 * access to memory reserved if the current task is killed (see
4064 * out_of_memory). Once filesystems are ready to handle allocation
4065 * failures more gracefully we should just bail out here.
4068 /* Exhausted what can be done so it's blame time */
4069 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4070 *did_some_progress = 1;
4073 * Help non-failing allocations by giving them access to memory
4076 if (gfp_mask & __GFP_NOFAIL)
4077 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4078 ALLOC_NO_WATERMARKS, ac);
4081 mutex_unlock(&oom_lock);
4086 * Maximum number of compaction retries wit a progress before OOM
4087 * killer is consider as the only way to move forward.
4089 #define MAX_COMPACT_RETRIES 16
4091 #ifdef CONFIG_COMPACTION
4092 /* Try memory compaction for high-order allocations before reclaim */
4093 static struct page *
4094 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4095 unsigned int alloc_flags, const struct alloc_context *ac,
4096 enum compact_priority prio, enum compact_result *compact_result)
4098 struct page *page = NULL;
4099 unsigned long pflags;
4100 unsigned int noreclaim_flag;
4105 psi_memstall_enter(&pflags);
4106 noreclaim_flag = memalloc_noreclaim_save();
4108 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4111 memalloc_noreclaim_restore(noreclaim_flag);
4112 psi_memstall_leave(&pflags);
4115 * At least in one zone compaction wasn't deferred or skipped, so let's
4116 * count a compaction stall
4118 count_vm_event(COMPACTSTALL);
4120 /* Prep a captured page if available */
4122 prep_new_page(page, order, gfp_mask, alloc_flags);
4124 /* Try get a page from the freelist if available */
4126 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4129 struct zone *zone = page_zone(page);
4131 zone->compact_blockskip_flush = false;
4132 compaction_defer_reset(zone, order, true);
4133 count_vm_event(COMPACTSUCCESS);
4138 * It's bad if compaction run occurs and fails. The most likely reason
4139 * is that pages exist, but not enough to satisfy watermarks.
4141 count_vm_event(COMPACTFAIL);
4149 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4150 enum compact_result compact_result,
4151 enum compact_priority *compact_priority,
4152 int *compaction_retries)
4154 int max_retries = MAX_COMPACT_RETRIES;
4157 int retries = *compaction_retries;
4158 enum compact_priority priority = *compact_priority;
4163 if (compaction_made_progress(compact_result))
4164 (*compaction_retries)++;
4167 * compaction considers all the zone as desperately out of memory
4168 * so it doesn't really make much sense to retry except when the
4169 * failure could be caused by insufficient priority
4171 if (compaction_failed(compact_result))
4172 goto check_priority;
4175 * compaction was skipped because there are not enough order-0 pages
4176 * to work with, so we retry only if it looks like reclaim can help.
4178 if (compaction_needs_reclaim(compact_result)) {
4179 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4184 * make sure the compaction wasn't deferred or didn't bail out early
4185 * due to locks contention before we declare that we should give up.
4186 * But the next retry should use a higher priority if allowed, so
4187 * we don't just keep bailing out endlessly.
4189 if (compaction_withdrawn(compact_result)) {
4190 goto check_priority;
4194 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4195 * costly ones because they are de facto nofail and invoke OOM
4196 * killer to move on while costly can fail and users are ready
4197 * to cope with that. 1/4 retries is rather arbitrary but we
4198 * would need much more detailed feedback from compaction to
4199 * make a better decision.
4201 if (order > PAGE_ALLOC_COSTLY_ORDER)
4203 if (*compaction_retries <= max_retries) {
4209 * Make sure there are attempts at the highest priority if we exhausted
4210 * all retries or failed at the lower priorities.
4213 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4214 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4216 if (*compact_priority > min_priority) {
4217 (*compact_priority)--;
4218 *compaction_retries = 0;
4222 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4226 static inline struct page *
4227 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4228 unsigned int alloc_flags, const struct alloc_context *ac,
4229 enum compact_priority prio, enum compact_result *compact_result)
4231 *compact_result = COMPACT_SKIPPED;
4236 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4237 enum compact_result compact_result,
4238 enum compact_priority *compact_priority,
4239 int *compaction_retries)
4244 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4248 * There are setups with compaction disabled which would prefer to loop
4249 * inside the allocator rather than hit the oom killer prematurely.
4250 * Let's give them a good hope and keep retrying while the order-0
4251 * watermarks are OK.
4253 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4254 ac->highest_zoneidx, ac->nodemask) {
4255 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4256 ac->highest_zoneidx, alloc_flags))
4261 #endif /* CONFIG_COMPACTION */
4263 #ifdef CONFIG_LOCKDEP
4264 static struct lockdep_map __fs_reclaim_map =
4265 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4267 static bool __need_fs_reclaim(gfp_t gfp_mask)
4269 gfp_mask = current_gfp_context(gfp_mask);
4271 /* no reclaim without waiting on it */
4272 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4275 /* this guy won't enter reclaim */
4276 if (current->flags & PF_MEMALLOC)
4279 /* We're only interested __GFP_FS allocations for now */
4280 if (!(gfp_mask & __GFP_FS))
4283 if (gfp_mask & __GFP_NOLOCKDEP)
4289 void __fs_reclaim_acquire(void)
4291 lock_map_acquire(&__fs_reclaim_map);
4294 void __fs_reclaim_release(void)
4296 lock_map_release(&__fs_reclaim_map);
4299 void fs_reclaim_acquire(gfp_t gfp_mask)
4301 if (__need_fs_reclaim(gfp_mask))
4302 __fs_reclaim_acquire();
4304 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4306 void fs_reclaim_release(gfp_t gfp_mask)
4308 if (__need_fs_reclaim(gfp_mask))
4309 __fs_reclaim_release();
4311 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4314 /* Perform direct synchronous page reclaim */
4315 static unsigned long
4316 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4317 const struct alloc_context *ac)
4319 unsigned int noreclaim_flag;
4320 unsigned long pflags, progress;
4324 /* We now go into synchronous reclaim */
4325 cpuset_memory_pressure_bump();
4326 psi_memstall_enter(&pflags);
4327 fs_reclaim_acquire(gfp_mask);
4328 noreclaim_flag = memalloc_noreclaim_save();
4330 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4333 memalloc_noreclaim_restore(noreclaim_flag);
4334 fs_reclaim_release(gfp_mask);
4335 psi_memstall_leave(&pflags);
4342 /* The really slow allocator path where we enter direct reclaim */
4343 static inline struct page *
4344 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4345 unsigned int alloc_flags, const struct alloc_context *ac,
4346 unsigned long *did_some_progress)
4348 struct page *page = NULL;
4349 bool drained = false;
4351 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4352 if (unlikely(!(*did_some_progress)))
4356 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4359 * If an allocation failed after direct reclaim, it could be because
4360 * pages are pinned on the per-cpu lists or in high alloc reserves.
4361 * Shrink them and try again
4363 if (!page && !drained) {
4364 unreserve_highatomic_pageblock(ac, false);
4365 drain_all_pages(NULL);
4373 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4374 const struct alloc_context *ac)
4378 pg_data_t *last_pgdat = NULL;
4379 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4381 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4383 if (last_pgdat != zone->zone_pgdat)
4384 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4385 last_pgdat = zone->zone_pgdat;
4389 static inline unsigned int
4390 gfp_to_alloc_flags(gfp_t gfp_mask)
4392 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4395 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4396 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4397 * to save two branches.
4399 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4400 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4403 * The caller may dip into page reserves a bit more if the caller
4404 * cannot run direct reclaim, or if the caller has realtime scheduling
4405 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4406 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4408 alloc_flags |= (__force int)
4409 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4411 if (gfp_mask & __GFP_ATOMIC) {
4413 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4414 * if it can't schedule.
4416 if (!(gfp_mask & __GFP_NOMEMALLOC))
4417 alloc_flags |= ALLOC_HARDER;
4419 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4420 * comment for __cpuset_node_allowed().
4422 alloc_flags &= ~ALLOC_CPUSET;
4423 } else if (unlikely(rt_task(current)) && !in_interrupt())
4424 alloc_flags |= ALLOC_HARDER;
4426 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4431 static bool oom_reserves_allowed(struct task_struct *tsk)
4433 if (!tsk_is_oom_victim(tsk))
4437 * !MMU doesn't have oom reaper so give access to memory reserves
4438 * only to the thread with TIF_MEMDIE set
4440 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4447 * Distinguish requests which really need access to full memory
4448 * reserves from oom victims which can live with a portion of it
4450 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4452 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4454 if (gfp_mask & __GFP_MEMALLOC)
4455 return ALLOC_NO_WATERMARKS;
4456 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4457 return ALLOC_NO_WATERMARKS;
4458 if (!in_interrupt()) {
4459 if (current->flags & PF_MEMALLOC)
4460 return ALLOC_NO_WATERMARKS;
4461 else if (oom_reserves_allowed(current))
4468 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4470 return !!__gfp_pfmemalloc_flags(gfp_mask);
4474 * Checks whether it makes sense to retry the reclaim to make a forward progress
4475 * for the given allocation request.
4477 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4478 * without success, or when we couldn't even meet the watermark if we
4479 * reclaimed all remaining pages on the LRU lists.
4481 * Returns true if a retry is viable or false to enter the oom path.
4484 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4485 struct alloc_context *ac, int alloc_flags,
4486 bool did_some_progress, int *no_progress_loops)
4493 * Costly allocations might have made a progress but this doesn't mean
4494 * their order will become available due to high fragmentation so
4495 * always increment the no progress counter for them
4497 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4498 *no_progress_loops = 0;
4500 (*no_progress_loops)++;
4503 * Make sure we converge to OOM if we cannot make any progress
4504 * several times in the row.
4506 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4507 /* Before OOM, exhaust highatomic_reserve */
4508 return unreserve_highatomic_pageblock(ac, true);
4512 * Keep reclaiming pages while there is a chance this will lead
4513 * somewhere. If none of the target zones can satisfy our allocation
4514 * request even if all reclaimable pages are considered then we are
4515 * screwed and have to go OOM.
4517 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4518 ac->highest_zoneidx, ac->nodemask) {
4519 unsigned long available;
4520 unsigned long reclaimable;
4521 unsigned long min_wmark = min_wmark_pages(zone);
4524 available = reclaimable = zone_reclaimable_pages(zone);
4525 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4528 * Would the allocation succeed if we reclaimed all
4529 * reclaimable pages?
4531 wmark = __zone_watermark_ok(zone, order, min_wmark,
4532 ac->highest_zoneidx, alloc_flags, available);
4533 trace_reclaim_retry_zone(z, order, reclaimable,
4534 available, min_wmark, *no_progress_loops, wmark);
4537 * If we didn't make any progress and have a lot of
4538 * dirty + writeback pages then we should wait for
4539 * an IO to complete to slow down the reclaim and
4540 * prevent from pre mature OOM
4542 if (!did_some_progress) {
4543 unsigned long write_pending;
4545 write_pending = zone_page_state_snapshot(zone,
4546 NR_ZONE_WRITE_PENDING);
4548 if (2 * write_pending > reclaimable) {
4549 congestion_wait(BLK_RW_ASYNC, HZ/10);
4561 * Memory allocation/reclaim might be called from a WQ context and the
4562 * current implementation of the WQ concurrency control doesn't
4563 * recognize that a particular WQ is congested if the worker thread is
4564 * looping without ever sleeping. Therefore we have to do a short sleep
4565 * here rather than calling cond_resched().
4567 if (current->flags & PF_WQ_WORKER)
4568 schedule_timeout_uninterruptible(1);
4575 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4578 * It's possible that cpuset's mems_allowed and the nodemask from
4579 * mempolicy don't intersect. This should be normally dealt with by
4580 * policy_nodemask(), but it's possible to race with cpuset update in
4581 * such a way the check therein was true, and then it became false
4582 * before we got our cpuset_mems_cookie here.
4583 * This assumes that for all allocations, ac->nodemask can come only
4584 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4585 * when it does not intersect with the cpuset restrictions) or the
4586 * caller can deal with a violated nodemask.
4588 if (cpusets_enabled() && ac->nodemask &&
4589 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4590 ac->nodemask = NULL;
4595 * When updating a task's mems_allowed or mempolicy nodemask, it is
4596 * possible to race with parallel threads in such a way that our
4597 * allocation can fail while the mask is being updated. If we are about
4598 * to fail, check if the cpuset changed during allocation and if so,
4601 if (read_mems_allowed_retry(cpuset_mems_cookie))
4607 static inline struct page *
4608 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4609 struct alloc_context *ac)
4611 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4612 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4613 struct page *page = NULL;
4614 unsigned int alloc_flags;
4615 unsigned long did_some_progress;
4616 enum compact_priority compact_priority;
4617 enum compact_result compact_result;
4618 int compaction_retries;
4619 int no_progress_loops;
4620 unsigned int cpuset_mems_cookie;
4624 * We also sanity check to catch abuse of atomic reserves being used by
4625 * callers that are not in atomic context.
4627 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4628 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4629 gfp_mask &= ~__GFP_ATOMIC;
4632 compaction_retries = 0;
4633 no_progress_loops = 0;
4634 compact_priority = DEF_COMPACT_PRIORITY;
4635 cpuset_mems_cookie = read_mems_allowed_begin();
4638 * The fast path uses conservative alloc_flags to succeed only until
4639 * kswapd needs to be woken up, and to avoid the cost of setting up
4640 * alloc_flags precisely. So we do that now.
4642 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4645 * We need to recalculate the starting point for the zonelist iterator
4646 * because we might have used different nodemask in the fast path, or
4647 * there was a cpuset modification and we are retrying - otherwise we
4648 * could end up iterating over non-eligible zones endlessly.
4650 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4651 ac->highest_zoneidx, ac->nodemask);
4652 if (!ac->preferred_zoneref->zone)
4655 if (alloc_flags & ALLOC_KSWAPD)
4656 wake_all_kswapds(order, gfp_mask, ac);
4659 * The adjusted alloc_flags might result in immediate success, so try
4662 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4667 * For costly allocations, try direct compaction first, as it's likely
4668 * that we have enough base pages and don't need to reclaim. For non-
4669 * movable high-order allocations, do that as well, as compaction will
4670 * try prevent permanent fragmentation by migrating from blocks of the
4672 * Don't try this for allocations that are allowed to ignore
4673 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4675 if (can_direct_reclaim &&
4677 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4678 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4679 page = __alloc_pages_direct_compact(gfp_mask, order,
4681 INIT_COMPACT_PRIORITY,
4687 * Checks for costly allocations with __GFP_NORETRY, which
4688 * includes some THP page fault allocations
4690 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4692 * If allocating entire pageblock(s) and compaction
4693 * failed because all zones are below low watermarks
4694 * or is prohibited because it recently failed at this
4695 * order, fail immediately unless the allocator has
4696 * requested compaction and reclaim retry.
4699 * - potentially very expensive because zones are far
4700 * below their low watermarks or this is part of very
4701 * bursty high order allocations,
4702 * - not guaranteed to help because isolate_freepages()
4703 * may not iterate over freed pages as part of its
4705 * - unlikely to make entire pageblocks free on its
4708 if (compact_result == COMPACT_SKIPPED ||
4709 compact_result == COMPACT_DEFERRED)
4713 * Looks like reclaim/compaction is worth trying, but
4714 * sync compaction could be very expensive, so keep
4715 * using async compaction.
4717 compact_priority = INIT_COMPACT_PRIORITY;
4722 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4723 if (alloc_flags & ALLOC_KSWAPD)
4724 wake_all_kswapds(order, gfp_mask, ac);
4726 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4728 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4731 * Reset the nodemask and zonelist iterators if memory policies can be
4732 * ignored. These allocations are high priority and system rather than
4735 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4736 ac->nodemask = NULL;
4737 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4738 ac->highest_zoneidx, ac->nodemask);
4741 /* Attempt with potentially adjusted zonelist and alloc_flags */
4742 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4746 /* Caller is not willing to reclaim, we can't balance anything */
4747 if (!can_direct_reclaim)
4750 /* Avoid recursion of direct reclaim */
4751 if (current->flags & PF_MEMALLOC)
4754 /* Try direct reclaim and then allocating */
4755 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4756 &did_some_progress);
4760 /* Try direct compaction and then allocating */
4761 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4762 compact_priority, &compact_result);
4766 /* Do not loop if specifically requested */
4767 if (gfp_mask & __GFP_NORETRY)
4771 * Do not retry costly high order allocations unless they are
4772 * __GFP_RETRY_MAYFAIL
4774 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4777 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4778 did_some_progress > 0, &no_progress_loops))
4782 * It doesn't make any sense to retry for the compaction if the order-0
4783 * reclaim is not able to make any progress because the current
4784 * implementation of the compaction depends on the sufficient amount
4785 * of free memory (see __compaction_suitable)
4787 if (did_some_progress > 0 &&
4788 should_compact_retry(ac, order, alloc_flags,
4789 compact_result, &compact_priority,
4790 &compaction_retries))
4794 /* Deal with possible cpuset update races before we start OOM killing */
4795 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4798 /* Reclaim has failed us, start killing things */
4799 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4803 /* Avoid allocations with no watermarks from looping endlessly */
4804 if (tsk_is_oom_victim(current) &&
4805 (alloc_flags & ALLOC_OOM ||
4806 (gfp_mask & __GFP_NOMEMALLOC)))
4809 /* Retry as long as the OOM killer is making progress */
4810 if (did_some_progress) {
4811 no_progress_loops = 0;
4816 /* Deal with possible cpuset update races before we fail */
4817 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4821 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4824 if (gfp_mask & __GFP_NOFAIL) {
4826 * All existing users of the __GFP_NOFAIL are blockable, so warn
4827 * of any new users that actually require GFP_NOWAIT
4829 if (WARN_ON_ONCE(!can_direct_reclaim))
4833 * PF_MEMALLOC request from this context is rather bizarre
4834 * because we cannot reclaim anything and only can loop waiting
4835 * for somebody to do a work for us
4837 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4840 * non failing costly orders are a hard requirement which we
4841 * are not prepared for much so let's warn about these users
4842 * so that we can identify them and convert them to something
4845 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4848 * Help non-failing allocations by giving them access to memory
4849 * reserves but do not use ALLOC_NO_WATERMARKS because this
4850 * could deplete whole memory reserves which would just make
4851 * the situation worse
4853 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4861 warn_alloc(gfp_mask, ac->nodemask,
4862 "page allocation failure: order:%u", order);
4867 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4868 int preferred_nid, nodemask_t *nodemask,
4869 struct alloc_context *ac, gfp_t *alloc_mask,
4870 unsigned int *alloc_flags)
4872 ac->highest_zoneidx = gfp_zone(gfp_mask);
4873 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4874 ac->nodemask = nodemask;
4875 ac->migratetype = gfp_migratetype(gfp_mask);
4877 if (cpusets_enabled()) {
4878 *alloc_mask |= __GFP_HARDWALL;
4880 * When we are in the interrupt context, it is irrelevant
4881 * to the current task context. It means that any node ok.
4883 if (!in_interrupt() && !ac->nodemask)
4884 ac->nodemask = &cpuset_current_mems_allowed;
4886 *alloc_flags |= ALLOC_CPUSET;
4889 fs_reclaim_acquire(gfp_mask);
4890 fs_reclaim_release(gfp_mask);
4892 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4894 if (should_fail_alloc_page(gfp_mask, order))
4897 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4899 /* Dirty zone balancing only done in the fast path */
4900 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4903 * The preferred zone is used for statistics but crucially it is
4904 * also used as the starting point for the zonelist iterator. It
4905 * may get reset for allocations that ignore memory policies.
4907 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4908 ac->highest_zoneidx, ac->nodemask);
4914 * This is the 'heart' of the zoned buddy allocator.
4917 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4918 nodemask_t *nodemask)
4921 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4922 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4923 struct alloc_context ac = { };
4926 * There are several places where we assume that the order value is sane
4927 * so bail out early if the request is out of bound.
4929 if (unlikely(order >= MAX_ORDER)) {
4930 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4934 gfp_mask &= gfp_allowed_mask;
4935 alloc_mask = gfp_mask;
4936 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4940 * Forbid the first pass from falling back to types that fragment
4941 * memory until all local zones are considered.
4943 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4945 /* First allocation attempt */
4946 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4951 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4952 * resp. GFP_NOIO which has to be inherited for all allocation requests
4953 * from a particular context which has been marked by
4954 * memalloc_no{fs,io}_{save,restore}.
4956 alloc_mask = current_gfp_context(gfp_mask);
4957 ac.spread_dirty_pages = false;
4960 * Restore the original nodemask if it was potentially replaced with
4961 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4963 ac.nodemask = nodemask;
4965 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4968 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4969 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4970 __free_pages(page, order);
4974 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4978 EXPORT_SYMBOL(__alloc_pages_nodemask);
4981 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4982 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4983 * you need to access high mem.
4985 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4989 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4992 return (unsigned long) page_address(page);
4994 EXPORT_SYMBOL(__get_free_pages);
4996 unsigned long get_zeroed_page(gfp_t gfp_mask)
4998 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5000 EXPORT_SYMBOL(get_zeroed_page);
5002 static inline void free_the_page(struct page *page, unsigned int order)
5004 if (order == 0) /* Via pcp? */
5005 free_unref_page(page);
5007 __free_pages_ok(page, order, FPI_NONE);
5010 void __free_pages(struct page *page, unsigned int order)
5012 if (put_page_testzero(page))
5013 free_the_page(page, order);
5014 else if (!PageHead(page))
5016 free_the_page(page + (1 << order), order);
5018 EXPORT_SYMBOL(__free_pages);
5020 void free_pages(unsigned long addr, unsigned int order)
5023 VM_BUG_ON(!virt_addr_valid((void *)addr));
5024 __free_pages(virt_to_page((void *)addr), order);
5028 EXPORT_SYMBOL(free_pages);
5032 * An arbitrary-length arbitrary-offset area of memory which resides
5033 * within a 0 or higher order page. Multiple fragments within that page
5034 * are individually refcounted, in the page's reference counter.
5036 * The page_frag functions below provide a simple allocation framework for
5037 * page fragments. This is used by the network stack and network device
5038 * drivers to provide a backing region of memory for use as either an
5039 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5041 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5044 struct page *page = NULL;
5045 gfp_t gfp = gfp_mask;
5047 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5048 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5050 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5051 PAGE_FRAG_CACHE_MAX_ORDER);
5052 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5054 if (unlikely(!page))
5055 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5057 nc->va = page ? page_address(page) : NULL;
5062 void __page_frag_cache_drain(struct page *page, unsigned int count)
5064 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5066 if (page_ref_sub_and_test(page, count))
5067 free_the_page(page, compound_order(page));
5069 EXPORT_SYMBOL(__page_frag_cache_drain);
5071 void *page_frag_alloc(struct page_frag_cache *nc,
5072 unsigned int fragsz, gfp_t gfp_mask)
5074 unsigned int size = PAGE_SIZE;
5078 if (unlikely(!nc->va)) {
5080 page = __page_frag_cache_refill(nc, gfp_mask);
5084 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5085 /* if size can vary use size else just use PAGE_SIZE */
5088 /* Even if we own the page, we do not use atomic_set().
5089 * This would break get_page_unless_zero() users.
5091 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5093 /* reset page count bias and offset to start of new frag */
5094 nc->pfmemalloc = page_is_pfmemalloc(page);
5095 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5099 offset = nc->offset - fragsz;
5100 if (unlikely(offset < 0)) {
5101 page = virt_to_page(nc->va);
5103 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5106 if (unlikely(nc->pfmemalloc)) {
5107 free_the_page(page, compound_order(page));
5111 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5112 /* if size can vary use size else just use PAGE_SIZE */
5115 /* OK, page count is 0, we can safely set it */
5116 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5118 /* reset page count bias and offset to start of new frag */
5119 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5120 offset = size - fragsz;
5124 nc->offset = offset;
5126 return nc->va + offset;
5128 EXPORT_SYMBOL(page_frag_alloc);
5131 * Frees a page fragment allocated out of either a compound or order 0 page.
5133 void page_frag_free(void *addr)
5135 struct page *page = virt_to_head_page(addr);
5137 if (unlikely(put_page_testzero(page)))
5138 free_the_page(page, compound_order(page));
5140 EXPORT_SYMBOL(page_frag_free);
5142 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5146 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5147 unsigned long used = addr + PAGE_ALIGN(size);
5149 split_page(virt_to_page((void *)addr), order);
5150 while (used < alloc_end) {
5155 return (void *)addr;
5159 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5160 * @size: the number of bytes to allocate
5161 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5163 * This function is similar to alloc_pages(), except that it allocates the
5164 * minimum number of pages to satisfy the request. alloc_pages() can only
5165 * allocate memory in power-of-two pages.
5167 * This function is also limited by MAX_ORDER.
5169 * Memory allocated by this function must be released by free_pages_exact().
5171 * Return: pointer to the allocated area or %NULL in case of error.
5173 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5175 unsigned int order = get_order(size);
5178 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5179 gfp_mask &= ~__GFP_COMP;
5181 addr = __get_free_pages(gfp_mask, order);
5182 return make_alloc_exact(addr, order, size);
5184 EXPORT_SYMBOL(alloc_pages_exact);
5187 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5189 * @nid: the preferred node ID where memory should be allocated
5190 * @size: the number of bytes to allocate
5191 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5193 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5196 * Return: pointer to the allocated area or %NULL in case of error.
5198 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5200 unsigned int order = get_order(size);
5203 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5204 gfp_mask &= ~__GFP_COMP;
5206 p = alloc_pages_node(nid, gfp_mask, order);
5209 return make_alloc_exact((unsigned long)page_address(p), order, size);
5213 * free_pages_exact - release memory allocated via alloc_pages_exact()
5214 * @virt: the value returned by alloc_pages_exact.
5215 * @size: size of allocation, same value as passed to alloc_pages_exact().
5217 * Release the memory allocated by a previous call to alloc_pages_exact.
5219 void free_pages_exact(void *virt, size_t size)
5221 unsigned long addr = (unsigned long)virt;
5222 unsigned long end = addr + PAGE_ALIGN(size);
5224 while (addr < end) {
5229 EXPORT_SYMBOL(free_pages_exact);
5232 * nr_free_zone_pages - count number of pages beyond high watermark
5233 * @offset: The zone index of the highest zone
5235 * nr_free_zone_pages() counts the number of pages which are beyond the
5236 * high watermark within all zones at or below a given zone index. For each
5237 * zone, the number of pages is calculated as:
5239 * nr_free_zone_pages = managed_pages - high_pages
5241 * Return: number of pages beyond high watermark.
5243 static unsigned long nr_free_zone_pages(int offset)
5248 /* Just pick one node, since fallback list is circular */
5249 unsigned long sum = 0;
5251 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5253 for_each_zone_zonelist(zone, z, zonelist, offset) {
5254 unsigned long size = zone_managed_pages(zone);
5255 unsigned long high = high_wmark_pages(zone);
5264 * nr_free_buffer_pages - count number of pages beyond high watermark
5266 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5267 * watermark within ZONE_DMA and ZONE_NORMAL.
5269 * Return: number of pages beyond high watermark within ZONE_DMA and
5272 unsigned long nr_free_buffer_pages(void)
5274 return nr_free_zone_pages(gfp_zone(GFP_USER));
5276 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5278 static inline void show_node(struct zone *zone)
5280 if (IS_ENABLED(CONFIG_NUMA))
5281 printk("Node %d ", zone_to_nid(zone));
5284 long si_mem_available(void)
5287 unsigned long pagecache;
5288 unsigned long wmark_low = 0;
5289 unsigned long pages[NR_LRU_LISTS];
5290 unsigned long reclaimable;
5294 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5295 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5298 wmark_low += low_wmark_pages(zone);
5301 * Estimate the amount of memory available for userspace allocations,
5302 * without causing swapping.
5304 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5307 * Not all the page cache can be freed, otherwise the system will
5308 * start swapping. Assume at least half of the page cache, or the
5309 * low watermark worth of cache, needs to stay.
5311 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5312 pagecache -= min(pagecache / 2, wmark_low);
5313 available += pagecache;
5316 * Part of the reclaimable slab and other kernel memory consists of
5317 * items that are in use, and cannot be freed. Cap this estimate at the
5320 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5321 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5322 available += reclaimable - min(reclaimable / 2, wmark_low);
5328 EXPORT_SYMBOL_GPL(si_mem_available);
5330 void si_meminfo(struct sysinfo *val)
5332 val->totalram = totalram_pages();
5333 val->sharedram = global_node_page_state(NR_SHMEM);
5334 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5335 val->bufferram = nr_blockdev_pages();
5336 val->totalhigh = totalhigh_pages();
5337 val->freehigh = nr_free_highpages();
5338 val->mem_unit = PAGE_SIZE;
5341 EXPORT_SYMBOL(si_meminfo);
5344 void si_meminfo_node(struct sysinfo *val, int nid)
5346 int zone_type; /* needs to be signed */
5347 unsigned long managed_pages = 0;
5348 unsigned long managed_highpages = 0;
5349 unsigned long free_highpages = 0;
5350 pg_data_t *pgdat = NODE_DATA(nid);
5352 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5353 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5354 val->totalram = managed_pages;
5355 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5356 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5357 #ifdef CONFIG_HIGHMEM
5358 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5359 struct zone *zone = &pgdat->node_zones[zone_type];
5361 if (is_highmem(zone)) {
5362 managed_highpages += zone_managed_pages(zone);
5363 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5366 val->totalhigh = managed_highpages;
5367 val->freehigh = free_highpages;
5369 val->totalhigh = managed_highpages;
5370 val->freehigh = free_highpages;
5372 val->mem_unit = PAGE_SIZE;
5377 * Determine whether the node should be displayed or not, depending on whether
5378 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5380 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5382 if (!(flags & SHOW_MEM_FILTER_NODES))
5386 * no node mask - aka implicit memory numa policy. Do not bother with
5387 * the synchronization - read_mems_allowed_begin - because we do not
5388 * have to be precise here.
5391 nodemask = &cpuset_current_mems_allowed;
5393 return !node_isset(nid, *nodemask);
5396 #define K(x) ((x) << (PAGE_SHIFT-10))
5398 static void show_migration_types(unsigned char type)
5400 static const char types[MIGRATE_TYPES] = {
5401 [MIGRATE_UNMOVABLE] = 'U',
5402 [MIGRATE_MOVABLE] = 'M',
5403 [MIGRATE_RECLAIMABLE] = 'E',
5404 [MIGRATE_HIGHATOMIC] = 'H',
5406 [MIGRATE_CMA] = 'C',
5408 #ifdef CONFIG_MEMORY_ISOLATION
5409 [MIGRATE_ISOLATE] = 'I',
5412 char tmp[MIGRATE_TYPES + 1];
5416 for (i = 0; i < MIGRATE_TYPES; i++) {
5417 if (type & (1 << i))
5422 printk(KERN_CONT "(%s) ", tmp);
5426 * Show free area list (used inside shift_scroll-lock stuff)
5427 * We also calculate the percentage fragmentation. We do this by counting the
5428 * memory on each free list with the exception of the first item on the list.
5431 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5434 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5436 unsigned long free_pcp = 0;
5441 for_each_populated_zone(zone) {
5442 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5445 for_each_online_cpu(cpu)
5446 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5449 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5450 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5451 " unevictable:%lu dirty:%lu writeback:%lu\n"
5452 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5453 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5454 " free:%lu free_pcp:%lu free_cma:%lu\n",
5455 global_node_page_state(NR_ACTIVE_ANON),
5456 global_node_page_state(NR_INACTIVE_ANON),
5457 global_node_page_state(NR_ISOLATED_ANON),
5458 global_node_page_state(NR_ACTIVE_FILE),
5459 global_node_page_state(NR_INACTIVE_FILE),
5460 global_node_page_state(NR_ISOLATED_FILE),
5461 global_node_page_state(NR_UNEVICTABLE),
5462 global_node_page_state(NR_FILE_DIRTY),
5463 global_node_page_state(NR_WRITEBACK),
5464 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5465 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5466 global_node_page_state(NR_FILE_MAPPED),
5467 global_node_page_state(NR_SHMEM),
5468 global_zone_page_state(NR_PAGETABLE),
5469 global_zone_page_state(NR_BOUNCE),
5470 global_zone_page_state(NR_FREE_PAGES),
5472 global_zone_page_state(NR_FREE_CMA_PAGES));
5474 for_each_online_pgdat(pgdat) {
5475 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5479 " active_anon:%lukB"
5480 " inactive_anon:%lukB"
5481 " active_file:%lukB"
5482 " inactive_file:%lukB"
5483 " unevictable:%lukB"
5484 " isolated(anon):%lukB"
5485 " isolated(file):%lukB"
5490 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5492 " shmem_pmdmapped: %lukB"
5495 " writeback_tmp:%lukB"
5496 " kernel_stack:%lukB"
5497 #ifdef CONFIG_SHADOW_CALL_STACK
5498 " shadow_call_stack:%lukB"
5500 " all_unreclaimable? %s"
5503 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5504 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5505 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5506 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5507 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5508 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5509 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5510 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5511 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5512 K(node_page_state(pgdat, NR_WRITEBACK)),
5513 K(node_page_state(pgdat, NR_SHMEM)),
5514 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5515 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5516 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5518 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5520 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5521 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5522 #ifdef CONFIG_SHADOW_CALL_STACK
5523 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5525 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5529 for_each_populated_zone(zone) {
5532 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5536 for_each_online_cpu(cpu)
5537 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5546 " reserved_highatomic:%luKB"
5547 " active_anon:%lukB"
5548 " inactive_anon:%lukB"
5549 " active_file:%lukB"
5550 " inactive_file:%lukB"
5551 " unevictable:%lukB"
5552 " writepending:%lukB"
5563 K(zone_page_state(zone, NR_FREE_PAGES)),
5564 K(min_wmark_pages(zone)),
5565 K(low_wmark_pages(zone)),
5566 K(high_wmark_pages(zone)),
5567 K(zone->nr_reserved_highatomic),
5568 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5569 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5570 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5571 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5572 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5573 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5574 K(zone->present_pages),
5575 K(zone_managed_pages(zone)),
5576 K(zone_page_state(zone, NR_MLOCK)),
5577 K(zone_page_state(zone, NR_PAGETABLE)),
5578 K(zone_page_state(zone, NR_BOUNCE)),
5580 K(this_cpu_read(zone->pageset->pcp.count)),
5581 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5582 printk("lowmem_reserve[]:");
5583 for (i = 0; i < MAX_NR_ZONES; i++)
5584 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5585 printk(KERN_CONT "\n");
5588 for_each_populated_zone(zone) {
5590 unsigned long nr[MAX_ORDER], flags, total = 0;
5591 unsigned char types[MAX_ORDER];
5593 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5596 printk(KERN_CONT "%s: ", zone->name);
5598 spin_lock_irqsave(&zone->lock, flags);
5599 for (order = 0; order < MAX_ORDER; order++) {
5600 struct free_area *area = &zone->free_area[order];
5603 nr[order] = area->nr_free;
5604 total += nr[order] << order;
5607 for (type = 0; type < MIGRATE_TYPES; type++) {
5608 if (!free_area_empty(area, type))
5609 types[order] |= 1 << type;
5612 spin_unlock_irqrestore(&zone->lock, flags);
5613 for (order = 0; order < MAX_ORDER; order++) {
5614 printk(KERN_CONT "%lu*%lukB ",
5615 nr[order], K(1UL) << order);
5617 show_migration_types(types[order]);
5619 printk(KERN_CONT "= %lukB\n", K(total));
5622 hugetlb_show_meminfo();
5624 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5626 show_swap_cache_info();
5629 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5631 zoneref->zone = zone;
5632 zoneref->zone_idx = zone_idx(zone);
5636 * Builds allocation fallback zone lists.
5638 * Add all populated zones of a node to the zonelist.
5640 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5643 enum zone_type zone_type = MAX_NR_ZONES;
5648 zone = pgdat->node_zones + zone_type;
5649 if (managed_zone(zone)) {
5650 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5651 check_highest_zone(zone_type);
5653 } while (zone_type);
5660 static int __parse_numa_zonelist_order(char *s)
5663 * We used to support different zonlists modes but they turned
5664 * out to be just not useful. Let's keep the warning in place
5665 * if somebody still use the cmd line parameter so that we do
5666 * not fail it silently
5668 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5669 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5675 char numa_zonelist_order[] = "Node";
5678 * sysctl handler for numa_zonelist_order
5680 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5681 void *buffer, size_t *length, loff_t *ppos)
5684 return __parse_numa_zonelist_order(buffer);
5685 return proc_dostring(table, write, buffer, length, ppos);
5689 #define MAX_NODE_LOAD (nr_online_nodes)
5690 static int node_load[MAX_NUMNODES];
5693 * find_next_best_node - find the next node that should appear in a given node's fallback list
5694 * @node: node whose fallback list we're appending
5695 * @used_node_mask: nodemask_t of already used nodes
5697 * We use a number of factors to determine which is the next node that should
5698 * appear on a given node's fallback list. The node should not have appeared
5699 * already in @node's fallback list, and it should be the next closest node
5700 * according to the distance array (which contains arbitrary distance values
5701 * from each node to each node in the system), and should also prefer nodes
5702 * with no CPUs, since presumably they'll have very little allocation pressure
5703 * on them otherwise.
5705 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5707 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5710 int min_val = INT_MAX;
5711 int best_node = NUMA_NO_NODE;
5713 /* Use the local node if we haven't already */
5714 if (!node_isset(node, *used_node_mask)) {
5715 node_set(node, *used_node_mask);
5719 for_each_node_state(n, N_MEMORY) {
5721 /* Don't want a node to appear more than once */
5722 if (node_isset(n, *used_node_mask))
5725 /* Use the distance array to find the distance */
5726 val = node_distance(node, n);
5728 /* Penalize nodes under us ("prefer the next node") */
5731 /* Give preference to headless and unused nodes */
5732 if (!cpumask_empty(cpumask_of_node(n)))
5733 val += PENALTY_FOR_NODE_WITH_CPUS;
5735 /* Slight preference for less loaded node */
5736 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5737 val += node_load[n];
5739 if (val < min_val) {
5746 node_set(best_node, *used_node_mask);
5753 * Build zonelists ordered by node and zones within node.
5754 * This results in maximum locality--normal zone overflows into local
5755 * DMA zone, if any--but risks exhausting DMA zone.
5757 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5760 struct zoneref *zonerefs;
5763 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5765 for (i = 0; i < nr_nodes; i++) {
5768 pg_data_t *node = NODE_DATA(node_order[i]);
5770 nr_zones = build_zonerefs_node(node, zonerefs);
5771 zonerefs += nr_zones;
5773 zonerefs->zone = NULL;
5774 zonerefs->zone_idx = 0;
5778 * Build gfp_thisnode zonelists
5780 static void build_thisnode_zonelists(pg_data_t *pgdat)
5782 struct zoneref *zonerefs;
5785 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5786 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5787 zonerefs += nr_zones;
5788 zonerefs->zone = NULL;
5789 zonerefs->zone_idx = 0;
5793 * Build zonelists ordered by zone and nodes within zones.
5794 * This results in conserving DMA zone[s] until all Normal memory is
5795 * exhausted, but results in overflowing to remote node while memory
5796 * may still exist in local DMA zone.
5799 static void build_zonelists(pg_data_t *pgdat)
5801 static int node_order[MAX_NUMNODES];
5802 int node, load, nr_nodes = 0;
5803 nodemask_t used_mask = NODE_MASK_NONE;
5804 int local_node, prev_node;
5806 /* NUMA-aware ordering of nodes */
5807 local_node = pgdat->node_id;
5808 load = nr_online_nodes;
5809 prev_node = local_node;
5811 memset(node_order, 0, sizeof(node_order));
5812 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5814 * We don't want to pressure a particular node.
5815 * So adding penalty to the first node in same
5816 * distance group to make it round-robin.
5818 if (node_distance(local_node, node) !=
5819 node_distance(local_node, prev_node))
5820 node_load[node] = load;
5822 node_order[nr_nodes++] = node;
5827 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5828 build_thisnode_zonelists(pgdat);
5831 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5833 * Return node id of node used for "local" allocations.
5834 * I.e., first node id of first zone in arg node's generic zonelist.
5835 * Used for initializing percpu 'numa_mem', which is used primarily
5836 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5838 int local_memory_node(int node)
5842 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5843 gfp_zone(GFP_KERNEL),
5845 return zone_to_nid(z->zone);
5849 static void setup_min_unmapped_ratio(void);
5850 static void setup_min_slab_ratio(void);
5851 #else /* CONFIG_NUMA */
5853 static void build_zonelists(pg_data_t *pgdat)
5855 int node, local_node;
5856 struct zoneref *zonerefs;
5859 local_node = pgdat->node_id;
5861 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5862 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5863 zonerefs += nr_zones;
5866 * Now we build the zonelist so that it contains the zones
5867 * of all the other nodes.
5868 * We don't want to pressure a particular node, so when
5869 * building the zones for node N, we make sure that the
5870 * zones coming right after the local ones are those from
5871 * node N+1 (modulo N)
5873 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5874 if (!node_online(node))
5876 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5877 zonerefs += nr_zones;
5879 for (node = 0; node < local_node; node++) {
5880 if (!node_online(node))
5882 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5883 zonerefs += nr_zones;
5886 zonerefs->zone = NULL;
5887 zonerefs->zone_idx = 0;
5890 #endif /* CONFIG_NUMA */
5893 * Boot pageset table. One per cpu which is going to be used for all
5894 * zones and all nodes. The parameters will be set in such a way
5895 * that an item put on a list will immediately be handed over to
5896 * the buddy list. This is safe since pageset manipulation is done
5897 * with interrupts disabled.
5899 * The boot_pagesets must be kept even after bootup is complete for
5900 * unused processors and/or zones. They do play a role for bootstrapping
5901 * hotplugged processors.
5903 * zoneinfo_show() and maybe other functions do
5904 * not check if the processor is online before following the pageset pointer.
5905 * Other parts of the kernel may not check if the zone is available.
5907 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5908 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5909 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5911 static void __build_all_zonelists(void *data)
5914 int __maybe_unused cpu;
5915 pg_data_t *self = data;
5916 static DEFINE_SPINLOCK(lock);
5921 memset(node_load, 0, sizeof(node_load));
5925 * This node is hotadded and no memory is yet present. So just
5926 * building zonelists is fine - no need to touch other nodes.
5928 if (self && !node_online(self->node_id)) {
5929 build_zonelists(self);
5931 for_each_online_node(nid) {
5932 pg_data_t *pgdat = NODE_DATA(nid);
5934 build_zonelists(pgdat);
5937 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5939 * We now know the "local memory node" for each node--
5940 * i.e., the node of the first zone in the generic zonelist.
5941 * Set up numa_mem percpu variable for on-line cpus. During
5942 * boot, only the boot cpu should be on-line; we'll init the
5943 * secondary cpus' numa_mem as they come on-line. During
5944 * node/memory hotplug, we'll fixup all on-line cpus.
5946 for_each_online_cpu(cpu)
5947 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5954 static noinline void __init
5955 build_all_zonelists_init(void)
5959 __build_all_zonelists(NULL);
5962 * Initialize the boot_pagesets that are going to be used
5963 * for bootstrapping processors. The real pagesets for
5964 * each zone will be allocated later when the per cpu
5965 * allocator is available.
5967 * boot_pagesets are used also for bootstrapping offline
5968 * cpus if the system is already booted because the pagesets
5969 * are needed to initialize allocators on a specific cpu too.
5970 * F.e. the percpu allocator needs the page allocator which
5971 * needs the percpu allocator in order to allocate its pagesets
5972 * (a chicken-egg dilemma).
5974 for_each_possible_cpu(cpu)
5975 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5977 mminit_verify_zonelist();
5978 cpuset_init_current_mems_allowed();
5982 * unless system_state == SYSTEM_BOOTING.
5984 * __ref due to call of __init annotated helper build_all_zonelists_init
5985 * [protected by SYSTEM_BOOTING].
5987 void __ref build_all_zonelists(pg_data_t *pgdat)
5989 unsigned long vm_total_pages;
5991 if (system_state == SYSTEM_BOOTING) {
5992 build_all_zonelists_init();
5994 __build_all_zonelists(pgdat);
5995 /* cpuset refresh routine should be here */
5997 /* Get the number of free pages beyond high watermark in all zones. */
5998 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6000 * Disable grouping by mobility if the number of pages in the
6001 * system is too low to allow the mechanism to work. It would be
6002 * more accurate, but expensive to check per-zone. This check is
6003 * made on memory-hotadd so a system can start with mobility
6004 * disabled and enable it later
6006 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6007 page_group_by_mobility_disabled = 1;
6009 page_group_by_mobility_disabled = 0;
6011 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6013 page_group_by_mobility_disabled ? "off" : "on",
6016 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6020 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6021 static bool __meminit
6022 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6024 static struct memblock_region *r;
6026 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6027 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6028 for_each_mem_region(r) {
6029 if (*pfn < memblock_region_memory_end_pfn(r))
6033 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6034 memblock_is_mirror(r)) {
6035 *pfn = memblock_region_memory_end_pfn(r);
6043 * Initially all pages are reserved - free ones are freed
6044 * up by memblock_free_all() once the early boot process is
6045 * done. Non-atomic initialization, single-pass.
6047 * All aligned pageblocks are initialized to the specified migratetype
6048 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6049 * zone stats (e.g., nr_isolate_pageblock) are touched.
6051 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
6052 unsigned long start_pfn,
6053 enum meminit_context context,
6054 struct vmem_altmap *altmap, int migratetype)
6056 unsigned long pfn, end_pfn = start_pfn + size;
6059 if (highest_memmap_pfn < end_pfn - 1)
6060 highest_memmap_pfn = end_pfn - 1;
6062 #ifdef CONFIG_ZONE_DEVICE
6064 * Honor reservation requested by the driver for this ZONE_DEVICE
6065 * memory. We limit the total number of pages to initialize to just
6066 * those that might contain the memory mapping. We will defer the
6067 * ZONE_DEVICE page initialization until after we have released
6070 if (zone == ZONE_DEVICE) {
6074 if (start_pfn == altmap->base_pfn)
6075 start_pfn += altmap->reserve;
6076 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6080 for (pfn = start_pfn; pfn < end_pfn; ) {
6082 * There can be holes in boot-time mem_map[]s handed to this
6083 * function. They do not exist on hotplugged memory.
6085 if (context == MEMINIT_EARLY) {
6086 if (overlap_memmap_init(zone, &pfn))
6088 if (defer_init(nid, pfn, end_pfn))
6092 page = pfn_to_page(pfn);
6093 __init_single_page(page, pfn, zone, nid);
6094 if (context == MEMINIT_HOTPLUG)
6095 __SetPageReserved(page);
6098 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6099 * such that unmovable allocations won't be scattered all
6100 * over the place during system boot.
6102 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6103 set_pageblock_migratetype(page, migratetype);
6110 #ifdef CONFIG_ZONE_DEVICE
6111 void __ref memmap_init_zone_device(struct zone *zone,
6112 unsigned long start_pfn,
6113 unsigned long nr_pages,
6114 struct dev_pagemap *pgmap)
6116 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6117 struct pglist_data *pgdat = zone->zone_pgdat;
6118 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6119 unsigned long zone_idx = zone_idx(zone);
6120 unsigned long start = jiffies;
6121 int nid = pgdat->node_id;
6123 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6127 * The call to memmap_init_zone should have already taken care
6128 * of the pages reserved for the memmap, so we can just jump to
6129 * the end of that region and start processing the device pages.
6132 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6133 nr_pages = end_pfn - start_pfn;
6136 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6137 struct page *page = pfn_to_page(pfn);
6139 __init_single_page(page, pfn, zone_idx, nid);
6142 * Mark page reserved as it will need to wait for onlining
6143 * phase for it to be fully associated with a zone.
6145 * We can use the non-atomic __set_bit operation for setting
6146 * the flag as we are still initializing the pages.
6148 __SetPageReserved(page);
6151 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6152 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6153 * ever freed or placed on a driver-private list.
6155 page->pgmap = pgmap;
6156 page->zone_device_data = NULL;
6159 * Mark the block movable so that blocks are reserved for
6160 * movable at startup. This will force kernel allocations
6161 * to reserve their blocks rather than leaking throughout
6162 * the address space during boot when many long-lived
6163 * kernel allocations are made.
6165 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6166 * because this is done early in section_activate()
6168 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6169 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6174 pr_info("%s initialised %lu pages in %ums\n", __func__,
6175 nr_pages, jiffies_to_msecs(jiffies - start));
6179 static void __meminit zone_init_free_lists(struct zone *zone)
6181 unsigned int order, t;
6182 for_each_migratetype_order(order, t) {
6183 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6184 zone->free_area[order].nr_free = 0;
6188 void __meminit __weak memmap_init(unsigned long size, int nid,
6190 unsigned long range_start_pfn)
6192 unsigned long start_pfn, end_pfn;
6193 unsigned long range_end_pfn = range_start_pfn + size;
6196 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6197 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6198 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6200 if (end_pfn > start_pfn) {
6201 size = end_pfn - start_pfn;
6202 memmap_init_zone(size, nid, zone, start_pfn,
6203 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6208 static int zone_batchsize(struct zone *zone)
6214 * The per-cpu-pages pools are set to around 1000th of the
6217 batch = zone_managed_pages(zone) / 1024;
6218 /* But no more than a meg. */
6219 if (batch * PAGE_SIZE > 1024 * 1024)
6220 batch = (1024 * 1024) / PAGE_SIZE;
6221 batch /= 4; /* We effectively *= 4 below */
6226 * Clamp the batch to a 2^n - 1 value. Having a power
6227 * of 2 value was found to be more likely to have
6228 * suboptimal cache aliasing properties in some cases.
6230 * For example if 2 tasks are alternately allocating
6231 * batches of pages, one task can end up with a lot
6232 * of pages of one half of the possible page colors
6233 * and the other with pages of the other colors.
6235 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6240 /* The deferral and batching of frees should be suppressed under NOMMU
6243 * The problem is that NOMMU needs to be able to allocate large chunks
6244 * of contiguous memory as there's no hardware page translation to
6245 * assemble apparent contiguous memory from discontiguous pages.
6247 * Queueing large contiguous runs of pages for batching, however,
6248 * causes the pages to actually be freed in smaller chunks. As there
6249 * can be a significant delay between the individual batches being
6250 * recycled, this leads to the once large chunks of space being
6251 * fragmented and becoming unavailable for high-order allocations.
6258 * pcp->high and pcp->batch values are related and dependent on one another:
6259 * ->batch must never be higher then ->high.
6260 * The following function updates them in a safe manner without read side
6263 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6264 * those fields changing asynchronously (acording to the above rule).
6266 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6267 * outside of boot time (or some other assurance that no concurrent updaters
6270 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6271 unsigned long batch)
6273 /* start with a fail safe value for batch */
6277 /* Update high, then batch, in order */
6284 /* a companion to pageset_set_high() */
6285 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6287 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6290 static void pageset_init(struct per_cpu_pageset *p)
6292 struct per_cpu_pages *pcp;
6295 memset(p, 0, sizeof(*p));
6298 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6299 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6302 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6305 pageset_set_batch(p, batch);
6309 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6310 * to the value high for the pageset p.
6312 static void pageset_set_high(struct per_cpu_pageset *p,
6315 unsigned long batch = max(1UL, high / 4);
6316 if ((high / 4) > (PAGE_SHIFT * 8))
6317 batch = PAGE_SHIFT * 8;
6319 pageset_update(&p->pcp, high, batch);
6322 static void pageset_set_high_and_batch(struct zone *zone,
6323 struct per_cpu_pageset *pcp)
6325 if (percpu_pagelist_fraction)
6326 pageset_set_high(pcp,
6327 (zone_managed_pages(zone) /
6328 percpu_pagelist_fraction));
6330 pageset_set_batch(pcp, zone_batchsize(zone));
6333 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6335 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6338 pageset_set_high_and_batch(zone, pcp);
6341 void __meminit setup_zone_pageset(struct zone *zone)
6344 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6345 for_each_possible_cpu(cpu)
6346 zone_pageset_init(zone, cpu);
6350 * Allocate per cpu pagesets and initialize them.
6351 * Before this call only boot pagesets were available.
6353 void __init setup_per_cpu_pageset(void)
6355 struct pglist_data *pgdat;
6357 int __maybe_unused cpu;
6359 for_each_populated_zone(zone)
6360 setup_zone_pageset(zone);
6364 * Unpopulated zones continue using the boot pagesets.
6365 * The numa stats for these pagesets need to be reset.
6366 * Otherwise, they will end up skewing the stats of
6367 * the nodes these zones are associated with.
6369 for_each_possible_cpu(cpu) {
6370 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6371 memset(pcp->vm_numa_stat_diff, 0,
6372 sizeof(pcp->vm_numa_stat_diff));
6376 for_each_online_pgdat(pgdat)
6377 pgdat->per_cpu_nodestats =
6378 alloc_percpu(struct per_cpu_nodestat);
6381 static __meminit void zone_pcp_init(struct zone *zone)
6384 * per cpu subsystem is not up at this point. The following code
6385 * relies on the ability of the linker to provide the
6386 * offset of a (static) per cpu variable into the per cpu area.
6388 zone->pageset = &boot_pageset;
6390 if (populated_zone(zone))
6391 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6392 zone->name, zone->present_pages,
6393 zone_batchsize(zone));
6396 void __meminit init_currently_empty_zone(struct zone *zone,
6397 unsigned long zone_start_pfn,
6400 struct pglist_data *pgdat = zone->zone_pgdat;
6401 int zone_idx = zone_idx(zone) + 1;
6403 if (zone_idx > pgdat->nr_zones)
6404 pgdat->nr_zones = zone_idx;
6406 zone->zone_start_pfn = zone_start_pfn;
6408 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6409 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6411 (unsigned long)zone_idx(zone),
6412 zone_start_pfn, (zone_start_pfn + size));
6414 zone_init_free_lists(zone);
6415 zone->initialized = 1;
6419 * get_pfn_range_for_nid - Return the start and end page frames for a node
6420 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6421 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6422 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6424 * It returns the start and end page frame of a node based on information
6425 * provided by memblock_set_node(). If called for a node
6426 * with no available memory, a warning is printed and the start and end
6429 void __init get_pfn_range_for_nid(unsigned int nid,
6430 unsigned long *start_pfn, unsigned long *end_pfn)
6432 unsigned long this_start_pfn, this_end_pfn;
6438 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6439 *start_pfn = min(*start_pfn, this_start_pfn);
6440 *end_pfn = max(*end_pfn, this_end_pfn);
6443 if (*start_pfn == -1UL)
6448 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6449 * assumption is made that zones within a node are ordered in monotonic
6450 * increasing memory addresses so that the "highest" populated zone is used
6452 static void __init find_usable_zone_for_movable(void)
6455 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6456 if (zone_index == ZONE_MOVABLE)
6459 if (arch_zone_highest_possible_pfn[zone_index] >
6460 arch_zone_lowest_possible_pfn[zone_index])
6464 VM_BUG_ON(zone_index == -1);
6465 movable_zone = zone_index;
6469 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6470 * because it is sized independent of architecture. Unlike the other zones,
6471 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6472 * in each node depending on the size of each node and how evenly kernelcore
6473 * is distributed. This helper function adjusts the zone ranges
6474 * provided by the architecture for a given node by using the end of the
6475 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6476 * zones within a node are in order of monotonic increases memory addresses
6478 static void __init adjust_zone_range_for_zone_movable(int nid,
6479 unsigned long zone_type,
6480 unsigned long node_start_pfn,
6481 unsigned long node_end_pfn,
6482 unsigned long *zone_start_pfn,
6483 unsigned long *zone_end_pfn)
6485 /* Only adjust if ZONE_MOVABLE is on this node */
6486 if (zone_movable_pfn[nid]) {
6487 /* Size ZONE_MOVABLE */
6488 if (zone_type == ZONE_MOVABLE) {
6489 *zone_start_pfn = zone_movable_pfn[nid];
6490 *zone_end_pfn = min(node_end_pfn,
6491 arch_zone_highest_possible_pfn[movable_zone]);
6493 /* Adjust for ZONE_MOVABLE starting within this range */
6494 } else if (!mirrored_kernelcore &&
6495 *zone_start_pfn < zone_movable_pfn[nid] &&
6496 *zone_end_pfn > zone_movable_pfn[nid]) {
6497 *zone_end_pfn = zone_movable_pfn[nid];
6499 /* Check if this whole range is within ZONE_MOVABLE */
6500 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6501 *zone_start_pfn = *zone_end_pfn;
6506 * Return the number of pages a zone spans in a node, including holes
6507 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6509 static unsigned long __init zone_spanned_pages_in_node(int nid,
6510 unsigned long zone_type,
6511 unsigned long node_start_pfn,
6512 unsigned long node_end_pfn,
6513 unsigned long *zone_start_pfn,
6514 unsigned long *zone_end_pfn)
6516 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6517 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6518 /* When hotadd a new node from cpu_up(), the node should be empty */
6519 if (!node_start_pfn && !node_end_pfn)
6522 /* Get the start and end of the zone */
6523 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6524 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6525 adjust_zone_range_for_zone_movable(nid, zone_type,
6526 node_start_pfn, node_end_pfn,
6527 zone_start_pfn, zone_end_pfn);
6529 /* Check that this node has pages within the zone's required range */
6530 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6533 /* Move the zone boundaries inside the node if necessary */
6534 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6535 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6537 /* Return the spanned pages */
6538 return *zone_end_pfn - *zone_start_pfn;
6542 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6543 * then all holes in the requested range will be accounted for.
6545 unsigned long __init __absent_pages_in_range(int nid,
6546 unsigned long range_start_pfn,
6547 unsigned long range_end_pfn)
6549 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6550 unsigned long start_pfn, end_pfn;
6553 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6554 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6555 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6556 nr_absent -= end_pfn - start_pfn;
6562 * absent_pages_in_range - Return number of page frames in holes within a range
6563 * @start_pfn: The start PFN to start searching for holes
6564 * @end_pfn: The end PFN to stop searching for holes
6566 * Return: the number of pages frames in memory holes within a range.
6568 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6569 unsigned long end_pfn)
6571 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6574 /* Return the number of page frames in holes in a zone on a node */
6575 static unsigned long __init zone_absent_pages_in_node(int nid,
6576 unsigned long zone_type,
6577 unsigned long node_start_pfn,
6578 unsigned long node_end_pfn)
6580 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6581 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6582 unsigned long zone_start_pfn, zone_end_pfn;
6583 unsigned long nr_absent;
6585 /* When hotadd a new node from cpu_up(), the node should be empty */
6586 if (!node_start_pfn && !node_end_pfn)
6589 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6590 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6592 adjust_zone_range_for_zone_movable(nid, zone_type,
6593 node_start_pfn, node_end_pfn,
6594 &zone_start_pfn, &zone_end_pfn);
6595 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6598 * ZONE_MOVABLE handling.
6599 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6602 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6603 unsigned long start_pfn, end_pfn;
6604 struct memblock_region *r;
6606 for_each_mem_region(r) {
6607 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6608 zone_start_pfn, zone_end_pfn);
6609 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6610 zone_start_pfn, zone_end_pfn);
6612 if (zone_type == ZONE_MOVABLE &&
6613 memblock_is_mirror(r))
6614 nr_absent += end_pfn - start_pfn;
6616 if (zone_type == ZONE_NORMAL &&
6617 !memblock_is_mirror(r))
6618 nr_absent += end_pfn - start_pfn;
6625 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6626 unsigned long node_start_pfn,
6627 unsigned long node_end_pfn)
6629 unsigned long realtotalpages = 0, totalpages = 0;
6632 for (i = 0; i < MAX_NR_ZONES; i++) {
6633 struct zone *zone = pgdat->node_zones + i;
6634 unsigned long zone_start_pfn, zone_end_pfn;
6635 unsigned long spanned, absent;
6636 unsigned long size, real_size;
6638 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6643 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6648 real_size = size - absent;
6651 zone->zone_start_pfn = zone_start_pfn;
6653 zone->zone_start_pfn = 0;
6654 zone->spanned_pages = size;
6655 zone->present_pages = real_size;
6658 realtotalpages += real_size;
6661 pgdat->node_spanned_pages = totalpages;
6662 pgdat->node_present_pages = realtotalpages;
6663 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6667 #ifndef CONFIG_SPARSEMEM
6669 * Calculate the size of the zone->blockflags rounded to an unsigned long
6670 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6671 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6672 * round what is now in bits to nearest long in bits, then return it in
6675 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6677 unsigned long usemapsize;
6679 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6680 usemapsize = roundup(zonesize, pageblock_nr_pages);
6681 usemapsize = usemapsize >> pageblock_order;
6682 usemapsize *= NR_PAGEBLOCK_BITS;
6683 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6685 return usemapsize / 8;
6688 static void __ref setup_usemap(struct pglist_data *pgdat,
6690 unsigned long zone_start_pfn,
6691 unsigned long zonesize)
6693 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6694 zone->pageblock_flags = NULL;
6696 zone->pageblock_flags =
6697 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6699 if (!zone->pageblock_flags)
6700 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6701 usemapsize, zone->name, pgdat->node_id);
6705 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6706 unsigned long zone_start_pfn, unsigned long zonesize) {}
6707 #endif /* CONFIG_SPARSEMEM */
6709 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6711 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6712 void __init set_pageblock_order(void)
6716 /* Check that pageblock_nr_pages has not already been setup */
6717 if (pageblock_order)
6720 if (HPAGE_SHIFT > PAGE_SHIFT)
6721 order = HUGETLB_PAGE_ORDER;
6723 order = MAX_ORDER - 1;
6726 * Assume the largest contiguous order of interest is a huge page.
6727 * This value may be variable depending on boot parameters on IA64 and
6730 pageblock_order = order;
6732 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6735 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6736 * is unused as pageblock_order is set at compile-time. See
6737 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6740 void __init set_pageblock_order(void)
6744 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6746 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6747 unsigned long present_pages)
6749 unsigned long pages = spanned_pages;
6752 * Provide a more accurate estimation if there are holes within
6753 * the zone and SPARSEMEM is in use. If there are holes within the
6754 * zone, each populated memory region may cost us one or two extra
6755 * memmap pages due to alignment because memmap pages for each
6756 * populated regions may not be naturally aligned on page boundary.
6757 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6759 if (spanned_pages > present_pages + (present_pages >> 4) &&
6760 IS_ENABLED(CONFIG_SPARSEMEM))
6761 pages = present_pages;
6763 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6766 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6767 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6769 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6771 spin_lock_init(&ds_queue->split_queue_lock);
6772 INIT_LIST_HEAD(&ds_queue->split_queue);
6773 ds_queue->split_queue_len = 0;
6776 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6779 #ifdef CONFIG_COMPACTION
6780 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6782 init_waitqueue_head(&pgdat->kcompactd_wait);
6785 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6788 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6790 pgdat_resize_init(pgdat);
6792 pgdat_init_split_queue(pgdat);
6793 pgdat_init_kcompactd(pgdat);
6795 init_waitqueue_head(&pgdat->kswapd_wait);
6796 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6798 pgdat_page_ext_init(pgdat);
6799 spin_lock_init(&pgdat->lru_lock);
6800 lruvec_init(&pgdat->__lruvec);
6803 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6804 unsigned long remaining_pages)
6806 atomic_long_set(&zone->managed_pages, remaining_pages);
6807 zone_set_nid(zone, nid);
6808 zone->name = zone_names[idx];
6809 zone->zone_pgdat = NODE_DATA(nid);
6810 spin_lock_init(&zone->lock);
6811 zone_seqlock_init(zone);
6812 zone_pcp_init(zone);
6816 * Set up the zone data structures
6817 * - init pgdat internals
6818 * - init all zones belonging to this node
6820 * NOTE: this function is only called during memory hotplug
6822 #ifdef CONFIG_MEMORY_HOTPLUG
6823 void __ref free_area_init_core_hotplug(int nid)
6826 pg_data_t *pgdat = NODE_DATA(nid);
6828 pgdat_init_internals(pgdat);
6829 for (z = 0; z < MAX_NR_ZONES; z++)
6830 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6835 * Set up the zone data structures:
6836 * - mark all pages reserved
6837 * - mark all memory queues empty
6838 * - clear the memory bitmaps
6840 * NOTE: pgdat should get zeroed by caller.
6841 * NOTE: this function is only called during early init.
6843 static void __init free_area_init_core(struct pglist_data *pgdat)
6846 int nid = pgdat->node_id;
6848 pgdat_init_internals(pgdat);
6849 pgdat->per_cpu_nodestats = &boot_nodestats;
6851 for (j = 0; j < MAX_NR_ZONES; j++) {
6852 struct zone *zone = pgdat->node_zones + j;
6853 unsigned long size, freesize, memmap_pages;
6854 unsigned long zone_start_pfn = zone->zone_start_pfn;
6856 size = zone->spanned_pages;
6857 freesize = zone->present_pages;
6860 * Adjust freesize so that it accounts for how much memory
6861 * is used by this zone for memmap. This affects the watermark
6862 * and per-cpu initialisations
6864 memmap_pages = calc_memmap_size(size, freesize);
6865 if (!is_highmem_idx(j)) {
6866 if (freesize >= memmap_pages) {
6867 freesize -= memmap_pages;
6870 " %s zone: %lu pages used for memmap\n",
6871 zone_names[j], memmap_pages);
6873 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6874 zone_names[j], memmap_pages, freesize);
6877 /* Account for reserved pages */
6878 if (j == 0 && freesize > dma_reserve) {
6879 freesize -= dma_reserve;
6880 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6881 zone_names[0], dma_reserve);
6884 if (!is_highmem_idx(j))
6885 nr_kernel_pages += freesize;
6886 /* Charge for highmem memmap if there are enough kernel pages */
6887 else if (nr_kernel_pages > memmap_pages * 2)
6888 nr_kernel_pages -= memmap_pages;
6889 nr_all_pages += freesize;
6892 * Set an approximate value for lowmem here, it will be adjusted
6893 * when the bootmem allocator frees pages into the buddy system.
6894 * And all highmem pages will be managed by the buddy system.
6896 zone_init_internals(zone, j, nid, freesize);
6901 set_pageblock_order();
6902 setup_usemap(pgdat, zone, zone_start_pfn, size);
6903 init_currently_empty_zone(zone, zone_start_pfn, size);
6904 memmap_init(size, nid, j, zone_start_pfn);
6908 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6909 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6911 unsigned long __maybe_unused start = 0;
6912 unsigned long __maybe_unused offset = 0;
6914 /* Skip empty nodes */
6915 if (!pgdat->node_spanned_pages)
6918 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6919 offset = pgdat->node_start_pfn - start;
6920 /* ia64 gets its own node_mem_map, before this, without bootmem */
6921 if (!pgdat->node_mem_map) {
6922 unsigned long size, end;
6926 * The zone's endpoints aren't required to be MAX_ORDER
6927 * aligned but the node_mem_map endpoints must be in order
6928 * for the buddy allocator to function correctly.
6930 end = pgdat_end_pfn(pgdat);
6931 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6932 size = (end - start) * sizeof(struct page);
6933 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6936 panic("Failed to allocate %ld bytes for node %d memory map\n",
6937 size, pgdat->node_id);
6938 pgdat->node_mem_map = map + offset;
6940 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6941 __func__, pgdat->node_id, (unsigned long)pgdat,
6942 (unsigned long)pgdat->node_mem_map);
6943 #ifndef CONFIG_NEED_MULTIPLE_NODES
6945 * With no DISCONTIG, the global mem_map is just set as node 0's
6947 if (pgdat == NODE_DATA(0)) {
6948 mem_map = NODE_DATA(0)->node_mem_map;
6949 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6955 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6956 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6958 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6959 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6961 pgdat->first_deferred_pfn = ULONG_MAX;
6964 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6967 static void __init free_area_init_node(int nid)
6969 pg_data_t *pgdat = NODE_DATA(nid);
6970 unsigned long start_pfn = 0;
6971 unsigned long end_pfn = 0;
6973 /* pg_data_t should be reset to zero when it's allocated */
6974 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6976 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6978 pgdat->node_id = nid;
6979 pgdat->node_start_pfn = start_pfn;
6980 pgdat->per_cpu_nodestats = NULL;
6982 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6983 (u64)start_pfn << PAGE_SHIFT,
6984 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6985 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6987 alloc_node_mem_map(pgdat);
6988 pgdat_set_deferred_range(pgdat);
6990 free_area_init_core(pgdat);
6993 void __init free_area_init_memoryless_node(int nid)
6995 free_area_init_node(nid);
6998 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
7000 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
7001 * PageReserved(). Return the number of struct pages that were initialized.
7003 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
7008 for (pfn = spfn; pfn < epfn; pfn++) {
7009 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
7010 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
7011 + pageblock_nr_pages - 1;
7015 * Use a fake node/zone (0) for now. Some of these pages
7016 * (in memblock.reserved but not in memblock.memory) will
7017 * get re-initialized via reserve_bootmem_region() later.
7019 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7020 __SetPageReserved(pfn_to_page(pfn));
7028 * Only struct pages that are backed by physical memory are zeroed and
7029 * initialized by going through __init_single_page(). But, there are some
7030 * struct pages which are reserved in memblock allocator and their fields
7031 * may be accessed (for example page_to_pfn() on some configuration accesses
7032 * flags). We must explicitly initialize those struct pages.
7034 * This function also addresses a similar issue where struct pages are left
7035 * uninitialized because the physical address range is not covered by
7036 * memblock.memory or memblock.reserved. That could happen when memblock
7037 * layout is manually configured via memmap=, or when the highest physical
7038 * address (max_pfn) does not end on a section boundary.
7040 static void __init init_unavailable_mem(void)
7042 phys_addr_t start, end;
7044 phys_addr_t next = 0;
7047 * Loop through unavailable ranges not covered by memblock.memory.
7050 for_each_mem_range(i, &start, &end) {
7052 pgcnt += init_unavailable_range(PFN_DOWN(next),
7058 * Early sections always have a fully populated memmap for the whole
7059 * section - see pfn_valid(). If the last section has holes at the
7060 * end and that section is marked "online", the memmap will be
7061 * considered initialized. Make sure that memmap has a well defined
7064 pgcnt += init_unavailable_range(PFN_DOWN(next),
7065 round_up(max_pfn, PAGES_PER_SECTION));
7068 * Struct pages that do not have backing memory. This could be because
7069 * firmware is using some of this memory, or for some other reasons.
7072 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7075 static inline void __init init_unavailable_mem(void)
7078 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7080 #if MAX_NUMNODES > 1
7082 * Figure out the number of possible node ids.
7084 void __init setup_nr_node_ids(void)
7086 unsigned int highest;
7088 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7089 nr_node_ids = highest + 1;
7094 * node_map_pfn_alignment - determine the maximum internode alignment
7096 * This function should be called after node map is populated and sorted.
7097 * It calculates the maximum power of two alignment which can distinguish
7100 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7101 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7102 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7103 * shifted, 1GiB is enough and this function will indicate so.
7105 * This is used to test whether pfn -> nid mapping of the chosen memory
7106 * model has fine enough granularity to avoid incorrect mapping for the
7107 * populated node map.
7109 * Return: the determined alignment in pfn's. 0 if there is no alignment
7110 * requirement (single node).
7112 unsigned long __init node_map_pfn_alignment(void)
7114 unsigned long accl_mask = 0, last_end = 0;
7115 unsigned long start, end, mask;
7116 int last_nid = NUMA_NO_NODE;
7119 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7120 if (!start || last_nid < 0 || last_nid == nid) {
7127 * Start with a mask granular enough to pin-point to the
7128 * start pfn and tick off bits one-by-one until it becomes
7129 * too coarse to separate the current node from the last.
7131 mask = ~((1 << __ffs(start)) - 1);
7132 while (mask && last_end <= (start & (mask << 1)))
7135 /* accumulate all internode masks */
7139 /* convert mask to number of pages */
7140 return ~accl_mask + 1;
7144 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7146 * Return: the minimum PFN based on information provided via
7147 * memblock_set_node().
7149 unsigned long __init find_min_pfn_with_active_regions(void)
7151 return PHYS_PFN(memblock_start_of_DRAM());
7155 * early_calculate_totalpages()
7156 * Sum pages in active regions for movable zone.
7157 * Populate N_MEMORY for calculating usable_nodes.
7159 static unsigned long __init early_calculate_totalpages(void)
7161 unsigned long totalpages = 0;
7162 unsigned long start_pfn, end_pfn;
7165 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7166 unsigned long pages = end_pfn - start_pfn;
7168 totalpages += pages;
7170 node_set_state(nid, N_MEMORY);
7176 * Find the PFN the Movable zone begins in each node. Kernel memory
7177 * is spread evenly between nodes as long as the nodes have enough
7178 * memory. When they don't, some nodes will have more kernelcore than
7181 static void __init find_zone_movable_pfns_for_nodes(void)
7184 unsigned long usable_startpfn;
7185 unsigned long kernelcore_node, kernelcore_remaining;
7186 /* save the state before borrow the nodemask */
7187 nodemask_t saved_node_state = node_states[N_MEMORY];
7188 unsigned long totalpages = early_calculate_totalpages();
7189 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7190 struct memblock_region *r;
7192 /* Need to find movable_zone earlier when movable_node is specified. */
7193 find_usable_zone_for_movable();
7196 * If movable_node is specified, ignore kernelcore and movablecore
7199 if (movable_node_is_enabled()) {
7200 for_each_mem_region(r) {
7201 if (!memblock_is_hotpluggable(r))
7204 nid = memblock_get_region_node(r);
7206 usable_startpfn = PFN_DOWN(r->base);
7207 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7208 min(usable_startpfn, zone_movable_pfn[nid]) :
7216 * If kernelcore=mirror is specified, ignore movablecore option
7218 if (mirrored_kernelcore) {
7219 bool mem_below_4gb_not_mirrored = false;
7221 for_each_mem_region(r) {
7222 if (memblock_is_mirror(r))
7225 nid = memblock_get_region_node(r);
7227 usable_startpfn = memblock_region_memory_base_pfn(r);
7229 if (usable_startpfn < 0x100000) {
7230 mem_below_4gb_not_mirrored = true;
7234 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7235 min(usable_startpfn, zone_movable_pfn[nid]) :
7239 if (mem_below_4gb_not_mirrored)
7240 pr_warn("This configuration results in unmirrored kernel memory.\n");
7246 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7247 * amount of necessary memory.
7249 if (required_kernelcore_percent)
7250 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7252 if (required_movablecore_percent)
7253 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7257 * If movablecore= was specified, calculate what size of
7258 * kernelcore that corresponds so that memory usable for
7259 * any allocation type is evenly spread. If both kernelcore
7260 * and movablecore are specified, then the value of kernelcore
7261 * will be used for required_kernelcore if it's greater than
7262 * what movablecore would have allowed.
7264 if (required_movablecore) {
7265 unsigned long corepages;
7268 * Round-up so that ZONE_MOVABLE is at least as large as what
7269 * was requested by the user
7271 required_movablecore =
7272 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7273 required_movablecore = min(totalpages, required_movablecore);
7274 corepages = totalpages - required_movablecore;
7276 required_kernelcore = max(required_kernelcore, corepages);
7280 * If kernelcore was not specified or kernelcore size is larger
7281 * than totalpages, there is no ZONE_MOVABLE.
7283 if (!required_kernelcore || required_kernelcore >= totalpages)
7286 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7287 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7290 /* Spread kernelcore memory as evenly as possible throughout nodes */
7291 kernelcore_node = required_kernelcore / usable_nodes;
7292 for_each_node_state(nid, N_MEMORY) {
7293 unsigned long start_pfn, end_pfn;
7296 * Recalculate kernelcore_node if the division per node
7297 * now exceeds what is necessary to satisfy the requested
7298 * amount of memory for the kernel
7300 if (required_kernelcore < kernelcore_node)
7301 kernelcore_node = required_kernelcore / usable_nodes;
7304 * As the map is walked, we track how much memory is usable
7305 * by the kernel using kernelcore_remaining. When it is
7306 * 0, the rest of the node is usable by ZONE_MOVABLE
7308 kernelcore_remaining = kernelcore_node;
7310 /* Go through each range of PFNs within this node */
7311 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7312 unsigned long size_pages;
7314 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7315 if (start_pfn >= end_pfn)
7318 /* Account for what is only usable for kernelcore */
7319 if (start_pfn < usable_startpfn) {
7320 unsigned long kernel_pages;
7321 kernel_pages = min(end_pfn, usable_startpfn)
7324 kernelcore_remaining -= min(kernel_pages,
7325 kernelcore_remaining);
7326 required_kernelcore -= min(kernel_pages,
7327 required_kernelcore);
7329 /* Continue if range is now fully accounted */
7330 if (end_pfn <= usable_startpfn) {
7333 * Push zone_movable_pfn to the end so
7334 * that if we have to rebalance
7335 * kernelcore across nodes, we will
7336 * not double account here
7338 zone_movable_pfn[nid] = end_pfn;
7341 start_pfn = usable_startpfn;
7345 * The usable PFN range for ZONE_MOVABLE is from
7346 * start_pfn->end_pfn. Calculate size_pages as the
7347 * number of pages used as kernelcore
7349 size_pages = end_pfn - start_pfn;
7350 if (size_pages > kernelcore_remaining)
7351 size_pages = kernelcore_remaining;
7352 zone_movable_pfn[nid] = start_pfn + size_pages;
7355 * Some kernelcore has been met, update counts and
7356 * break if the kernelcore for this node has been
7359 required_kernelcore -= min(required_kernelcore,
7361 kernelcore_remaining -= size_pages;
7362 if (!kernelcore_remaining)
7368 * If there is still required_kernelcore, we do another pass with one
7369 * less node in the count. This will push zone_movable_pfn[nid] further
7370 * along on the nodes that still have memory until kernelcore is
7374 if (usable_nodes && required_kernelcore > usable_nodes)
7378 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7379 for (nid = 0; nid < MAX_NUMNODES; nid++)
7380 zone_movable_pfn[nid] =
7381 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7384 /* restore the node_state */
7385 node_states[N_MEMORY] = saved_node_state;
7388 /* Any regular or high memory on that node ? */
7389 static void check_for_memory(pg_data_t *pgdat, int nid)
7391 enum zone_type zone_type;
7393 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7394 struct zone *zone = &pgdat->node_zones[zone_type];
7395 if (populated_zone(zone)) {
7396 if (IS_ENABLED(CONFIG_HIGHMEM))
7397 node_set_state(nid, N_HIGH_MEMORY);
7398 if (zone_type <= ZONE_NORMAL)
7399 node_set_state(nid, N_NORMAL_MEMORY);
7406 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7407 * such cases we allow max_zone_pfn sorted in the descending order
7409 bool __weak arch_has_descending_max_zone_pfns(void)
7415 * free_area_init - Initialise all pg_data_t and zone data
7416 * @max_zone_pfn: an array of max PFNs for each zone
7418 * This will call free_area_init_node() for each active node in the system.
7419 * Using the page ranges provided by memblock_set_node(), the size of each
7420 * zone in each node and their holes is calculated. If the maximum PFN
7421 * between two adjacent zones match, it is assumed that the zone is empty.
7422 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7423 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7424 * starts where the previous one ended. For example, ZONE_DMA32 starts
7425 * at arch_max_dma_pfn.
7427 void __init free_area_init(unsigned long *max_zone_pfn)
7429 unsigned long start_pfn, end_pfn;
7433 /* Record where the zone boundaries are */
7434 memset(arch_zone_lowest_possible_pfn, 0,
7435 sizeof(arch_zone_lowest_possible_pfn));
7436 memset(arch_zone_highest_possible_pfn, 0,
7437 sizeof(arch_zone_highest_possible_pfn));
7439 start_pfn = find_min_pfn_with_active_regions();
7440 descending = arch_has_descending_max_zone_pfns();
7442 for (i = 0; i < MAX_NR_ZONES; i++) {
7444 zone = MAX_NR_ZONES - i - 1;
7448 if (zone == ZONE_MOVABLE)
7451 end_pfn = max(max_zone_pfn[zone], start_pfn);
7452 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7453 arch_zone_highest_possible_pfn[zone] = end_pfn;
7455 start_pfn = end_pfn;
7458 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7459 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7460 find_zone_movable_pfns_for_nodes();
7462 /* Print out the zone ranges */
7463 pr_info("Zone ranges:\n");
7464 for (i = 0; i < MAX_NR_ZONES; i++) {
7465 if (i == ZONE_MOVABLE)
7467 pr_info(" %-8s ", zone_names[i]);
7468 if (arch_zone_lowest_possible_pfn[i] ==
7469 arch_zone_highest_possible_pfn[i])
7472 pr_cont("[mem %#018Lx-%#018Lx]\n",
7473 (u64)arch_zone_lowest_possible_pfn[i]
7475 ((u64)arch_zone_highest_possible_pfn[i]
7476 << PAGE_SHIFT) - 1);
7479 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7480 pr_info("Movable zone start for each node\n");
7481 for (i = 0; i < MAX_NUMNODES; i++) {
7482 if (zone_movable_pfn[i])
7483 pr_info(" Node %d: %#018Lx\n", i,
7484 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7488 * Print out the early node map, and initialize the
7489 * subsection-map relative to active online memory ranges to
7490 * enable future "sub-section" extensions of the memory map.
7492 pr_info("Early memory node ranges\n");
7493 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7494 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7495 (u64)start_pfn << PAGE_SHIFT,
7496 ((u64)end_pfn << PAGE_SHIFT) - 1);
7497 subsection_map_init(start_pfn, end_pfn - start_pfn);
7500 /* Initialise every node */
7501 mminit_verify_pageflags_layout();
7502 setup_nr_node_ids();
7503 init_unavailable_mem();
7504 for_each_online_node(nid) {
7505 pg_data_t *pgdat = NODE_DATA(nid);
7506 free_area_init_node(nid);
7508 /* Any memory on that node */
7509 if (pgdat->node_present_pages)
7510 node_set_state(nid, N_MEMORY);
7511 check_for_memory(pgdat, nid);
7515 static int __init cmdline_parse_core(char *p, unsigned long *core,
7516 unsigned long *percent)
7518 unsigned long long coremem;
7524 /* Value may be a percentage of total memory, otherwise bytes */
7525 coremem = simple_strtoull(p, &endptr, 0);
7526 if (*endptr == '%') {
7527 /* Paranoid check for percent values greater than 100 */
7528 WARN_ON(coremem > 100);
7532 coremem = memparse(p, &p);
7533 /* Paranoid check that UL is enough for the coremem value */
7534 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7536 *core = coremem >> PAGE_SHIFT;
7543 * kernelcore=size sets the amount of memory for use for allocations that
7544 * cannot be reclaimed or migrated.
7546 static int __init cmdline_parse_kernelcore(char *p)
7548 /* parse kernelcore=mirror */
7549 if (parse_option_str(p, "mirror")) {
7550 mirrored_kernelcore = true;
7554 return cmdline_parse_core(p, &required_kernelcore,
7555 &required_kernelcore_percent);
7559 * movablecore=size sets the amount of memory for use for allocations that
7560 * can be reclaimed or migrated.
7562 static int __init cmdline_parse_movablecore(char *p)
7564 return cmdline_parse_core(p, &required_movablecore,
7565 &required_movablecore_percent);
7568 early_param("kernelcore", cmdline_parse_kernelcore);
7569 early_param("movablecore", cmdline_parse_movablecore);
7571 void adjust_managed_page_count(struct page *page, long count)
7573 atomic_long_add(count, &page_zone(page)->managed_pages);
7574 totalram_pages_add(count);
7575 #ifdef CONFIG_HIGHMEM
7576 if (PageHighMem(page))
7577 totalhigh_pages_add(count);
7580 EXPORT_SYMBOL(adjust_managed_page_count);
7582 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7585 unsigned long pages = 0;
7587 start = (void *)PAGE_ALIGN((unsigned long)start);
7588 end = (void *)((unsigned long)end & PAGE_MASK);
7589 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7590 struct page *page = virt_to_page(pos);
7591 void *direct_map_addr;
7594 * 'direct_map_addr' might be different from 'pos'
7595 * because some architectures' virt_to_page()
7596 * work with aliases. Getting the direct map
7597 * address ensures that we get a _writeable_
7598 * alias for the memset().
7600 direct_map_addr = page_address(page);
7601 if ((unsigned int)poison <= 0xFF)
7602 memset(direct_map_addr, poison, PAGE_SIZE);
7604 free_reserved_page(page);
7608 pr_info("Freeing %s memory: %ldK\n",
7609 s, pages << (PAGE_SHIFT - 10));
7614 #ifdef CONFIG_HIGHMEM
7615 void free_highmem_page(struct page *page)
7617 __free_reserved_page(page);
7618 totalram_pages_inc();
7619 atomic_long_inc(&page_zone(page)->managed_pages);
7620 totalhigh_pages_inc();
7625 void __init mem_init_print_info(const char *str)
7627 unsigned long physpages, codesize, datasize, rosize, bss_size;
7628 unsigned long init_code_size, init_data_size;
7630 physpages = get_num_physpages();
7631 codesize = _etext - _stext;
7632 datasize = _edata - _sdata;
7633 rosize = __end_rodata - __start_rodata;
7634 bss_size = __bss_stop - __bss_start;
7635 init_data_size = __init_end - __init_begin;
7636 init_code_size = _einittext - _sinittext;
7639 * Detect special cases and adjust section sizes accordingly:
7640 * 1) .init.* may be embedded into .data sections
7641 * 2) .init.text.* may be out of [__init_begin, __init_end],
7642 * please refer to arch/tile/kernel/vmlinux.lds.S.
7643 * 3) .rodata.* may be embedded into .text or .data sections.
7645 #define adj_init_size(start, end, size, pos, adj) \
7647 if (start <= pos && pos < end && size > adj) \
7651 adj_init_size(__init_begin, __init_end, init_data_size,
7652 _sinittext, init_code_size);
7653 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7654 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7655 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7656 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7658 #undef adj_init_size
7660 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7661 #ifdef CONFIG_HIGHMEM
7665 nr_free_pages() << (PAGE_SHIFT - 10),
7666 physpages << (PAGE_SHIFT - 10),
7667 codesize >> 10, datasize >> 10, rosize >> 10,
7668 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7669 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7670 totalcma_pages << (PAGE_SHIFT - 10),
7671 #ifdef CONFIG_HIGHMEM
7672 totalhigh_pages() << (PAGE_SHIFT - 10),
7674 str ? ", " : "", str ? str : "");
7678 * set_dma_reserve - set the specified number of pages reserved in the first zone
7679 * @new_dma_reserve: The number of pages to mark reserved
7681 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7682 * In the DMA zone, a significant percentage may be consumed by kernel image
7683 * and other unfreeable allocations which can skew the watermarks badly. This
7684 * function may optionally be used to account for unfreeable pages in the
7685 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7686 * smaller per-cpu batchsize.
7688 void __init set_dma_reserve(unsigned long new_dma_reserve)
7690 dma_reserve = new_dma_reserve;
7693 static int page_alloc_cpu_dead(unsigned int cpu)
7696 lru_add_drain_cpu(cpu);
7700 * Spill the event counters of the dead processor
7701 * into the current processors event counters.
7702 * This artificially elevates the count of the current
7705 vm_events_fold_cpu(cpu);
7708 * Zero the differential counters of the dead processor
7709 * so that the vm statistics are consistent.
7711 * This is only okay since the processor is dead and cannot
7712 * race with what we are doing.
7714 cpu_vm_stats_fold(cpu);
7719 int hashdist = HASHDIST_DEFAULT;
7721 static int __init set_hashdist(char *str)
7725 hashdist = simple_strtoul(str, &str, 0);
7728 __setup("hashdist=", set_hashdist);
7731 void __init page_alloc_init(void)
7736 if (num_node_state(N_MEMORY) == 1)
7740 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7741 "mm/page_alloc:dead", NULL,
7742 page_alloc_cpu_dead);
7747 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7748 * or min_free_kbytes changes.
7750 static void calculate_totalreserve_pages(void)
7752 struct pglist_data *pgdat;
7753 unsigned long reserve_pages = 0;
7754 enum zone_type i, j;
7756 for_each_online_pgdat(pgdat) {
7758 pgdat->totalreserve_pages = 0;
7760 for (i = 0; i < MAX_NR_ZONES; i++) {
7761 struct zone *zone = pgdat->node_zones + i;
7763 unsigned long managed_pages = zone_managed_pages(zone);
7765 /* Find valid and maximum lowmem_reserve in the zone */
7766 for (j = i; j < MAX_NR_ZONES; j++) {
7767 if (zone->lowmem_reserve[j] > max)
7768 max = zone->lowmem_reserve[j];
7771 /* we treat the high watermark as reserved pages. */
7772 max += high_wmark_pages(zone);
7774 if (max > managed_pages)
7775 max = managed_pages;
7777 pgdat->totalreserve_pages += max;
7779 reserve_pages += max;
7782 totalreserve_pages = reserve_pages;
7786 * setup_per_zone_lowmem_reserve - called whenever
7787 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7788 * has a correct pages reserved value, so an adequate number of
7789 * pages are left in the zone after a successful __alloc_pages().
7791 static void setup_per_zone_lowmem_reserve(void)
7793 struct pglist_data *pgdat;
7794 enum zone_type j, idx;
7796 for_each_online_pgdat(pgdat) {
7797 for (j = 0; j < MAX_NR_ZONES; j++) {
7798 struct zone *zone = pgdat->node_zones + j;
7799 unsigned long managed_pages = zone_managed_pages(zone);
7801 zone->lowmem_reserve[j] = 0;
7805 struct zone *lower_zone;
7808 lower_zone = pgdat->node_zones + idx;
7810 if (!sysctl_lowmem_reserve_ratio[idx] ||
7811 !zone_managed_pages(lower_zone)) {
7812 lower_zone->lowmem_reserve[j] = 0;
7815 lower_zone->lowmem_reserve[j] =
7816 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7818 managed_pages += zone_managed_pages(lower_zone);
7823 /* update totalreserve_pages */
7824 calculate_totalreserve_pages();
7827 static void __setup_per_zone_wmarks(void)
7829 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7830 unsigned long lowmem_pages = 0;
7832 unsigned long flags;
7834 /* Calculate total number of !ZONE_HIGHMEM pages */
7835 for_each_zone(zone) {
7836 if (!is_highmem(zone))
7837 lowmem_pages += zone_managed_pages(zone);
7840 for_each_zone(zone) {
7843 spin_lock_irqsave(&zone->lock, flags);
7844 tmp = (u64)pages_min * zone_managed_pages(zone);
7845 do_div(tmp, lowmem_pages);
7846 if (is_highmem(zone)) {
7848 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7849 * need highmem pages, so cap pages_min to a small
7852 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7853 * deltas control async page reclaim, and so should
7854 * not be capped for highmem.
7856 unsigned long min_pages;
7858 min_pages = zone_managed_pages(zone) / 1024;
7859 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7860 zone->_watermark[WMARK_MIN] = min_pages;
7863 * If it's a lowmem zone, reserve a number of pages
7864 * proportionate to the zone's size.
7866 zone->_watermark[WMARK_MIN] = tmp;
7870 * Set the kswapd watermarks distance according to the
7871 * scale factor in proportion to available memory, but
7872 * ensure a minimum size on small systems.
7874 tmp = max_t(u64, tmp >> 2,
7875 mult_frac(zone_managed_pages(zone),
7876 watermark_scale_factor, 10000));
7878 zone->watermark_boost = 0;
7879 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7880 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7882 spin_unlock_irqrestore(&zone->lock, flags);
7885 /* update totalreserve_pages */
7886 calculate_totalreserve_pages();
7890 * setup_per_zone_wmarks - called when min_free_kbytes changes
7891 * or when memory is hot-{added|removed}
7893 * Ensures that the watermark[min,low,high] values for each zone are set
7894 * correctly with respect to min_free_kbytes.
7896 void setup_per_zone_wmarks(void)
7898 static DEFINE_SPINLOCK(lock);
7901 __setup_per_zone_wmarks();
7906 * Initialise min_free_kbytes.
7908 * For small machines we want it small (128k min). For large machines
7909 * we want it large (256MB max). But it is not linear, because network
7910 * bandwidth does not increase linearly with machine size. We use
7912 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7913 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7929 int __meminit init_per_zone_wmark_min(void)
7931 unsigned long lowmem_kbytes;
7932 int new_min_free_kbytes;
7934 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7935 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7937 if (new_min_free_kbytes > user_min_free_kbytes) {
7938 min_free_kbytes = new_min_free_kbytes;
7939 if (min_free_kbytes < 128)
7940 min_free_kbytes = 128;
7941 if (min_free_kbytes > 262144)
7942 min_free_kbytes = 262144;
7944 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7945 new_min_free_kbytes, user_min_free_kbytes);
7947 setup_per_zone_wmarks();
7948 refresh_zone_stat_thresholds();
7949 setup_per_zone_lowmem_reserve();
7952 setup_min_unmapped_ratio();
7953 setup_min_slab_ratio();
7956 khugepaged_min_free_kbytes_update();
7960 postcore_initcall(init_per_zone_wmark_min)
7963 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7964 * that we can call two helper functions whenever min_free_kbytes
7967 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7968 void *buffer, size_t *length, loff_t *ppos)
7972 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7977 user_min_free_kbytes = min_free_kbytes;
7978 setup_per_zone_wmarks();
7983 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7984 void *buffer, size_t *length, loff_t *ppos)
7988 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7993 setup_per_zone_wmarks();
7999 static void setup_min_unmapped_ratio(void)
8004 for_each_online_pgdat(pgdat)
8005 pgdat->min_unmapped_pages = 0;
8008 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8009 sysctl_min_unmapped_ratio) / 100;
8013 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8014 void *buffer, size_t *length, loff_t *ppos)
8018 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8022 setup_min_unmapped_ratio();
8027 static void setup_min_slab_ratio(void)
8032 for_each_online_pgdat(pgdat)
8033 pgdat->min_slab_pages = 0;
8036 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8037 sysctl_min_slab_ratio) / 100;
8040 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8041 void *buffer, size_t *length, loff_t *ppos)
8045 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8049 setup_min_slab_ratio();
8056 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8057 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8058 * whenever sysctl_lowmem_reserve_ratio changes.
8060 * The reserve ratio obviously has absolutely no relation with the
8061 * minimum watermarks. The lowmem reserve ratio can only make sense
8062 * if in function of the boot time zone sizes.
8064 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8065 void *buffer, size_t *length, loff_t *ppos)
8069 proc_dointvec_minmax(table, write, buffer, length, ppos);
8071 for (i = 0; i < MAX_NR_ZONES; i++) {
8072 if (sysctl_lowmem_reserve_ratio[i] < 1)
8073 sysctl_lowmem_reserve_ratio[i] = 0;
8076 setup_per_zone_lowmem_reserve();
8080 static void __zone_pcp_update(struct zone *zone)
8084 for_each_possible_cpu(cpu)
8085 pageset_set_high_and_batch(zone,
8086 per_cpu_ptr(zone->pageset, cpu));
8090 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8091 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8092 * pagelist can have before it gets flushed back to buddy allocator.
8094 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8095 void *buffer, size_t *length, loff_t *ppos)
8098 int old_percpu_pagelist_fraction;
8101 mutex_lock(&pcp_batch_high_lock);
8102 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8104 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8105 if (!write || ret < 0)
8108 /* Sanity checking to avoid pcp imbalance */
8109 if (percpu_pagelist_fraction &&
8110 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8111 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8117 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8120 for_each_populated_zone(zone)
8121 __zone_pcp_update(zone);
8123 mutex_unlock(&pcp_batch_high_lock);
8127 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8129 * Returns the number of pages that arch has reserved but
8130 * is not known to alloc_large_system_hash().
8132 static unsigned long __init arch_reserved_kernel_pages(void)
8139 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8140 * machines. As memory size is increased the scale is also increased but at
8141 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8142 * quadruples the scale is increased by one, which means the size of hash table
8143 * only doubles, instead of quadrupling as well.
8144 * Because 32-bit systems cannot have large physical memory, where this scaling
8145 * makes sense, it is disabled on such platforms.
8147 #if __BITS_PER_LONG > 32
8148 #define ADAPT_SCALE_BASE (64ul << 30)
8149 #define ADAPT_SCALE_SHIFT 2
8150 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8154 * allocate a large system hash table from bootmem
8155 * - it is assumed that the hash table must contain an exact power-of-2
8156 * quantity of entries
8157 * - limit is the number of hash buckets, not the total allocation size
8159 void *__init alloc_large_system_hash(const char *tablename,
8160 unsigned long bucketsize,
8161 unsigned long numentries,
8164 unsigned int *_hash_shift,
8165 unsigned int *_hash_mask,
8166 unsigned long low_limit,
8167 unsigned long high_limit)
8169 unsigned long long max = high_limit;
8170 unsigned long log2qty, size;
8175 /* allow the kernel cmdline to have a say */
8177 /* round applicable memory size up to nearest megabyte */
8178 numentries = nr_kernel_pages;
8179 numentries -= arch_reserved_kernel_pages();
8181 /* It isn't necessary when PAGE_SIZE >= 1MB */
8182 if (PAGE_SHIFT < 20)
8183 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8185 #if __BITS_PER_LONG > 32
8187 unsigned long adapt;
8189 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8190 adapt <<= ADAPT_SCALE_SHIFT)
8195 /* limit to 1 bucket per 2^scale bytes of low memory */
8196 if (scale > PAGE_SHIFT)
8197 numentries >>= (scale - PAGE_SHIFT);
8199 numentries <<= (PAGE_SHIFT - scale);
8201 /* Make sure we've got at least a 0-order allocation.. */
8202 if (unlikely(flags & HASH_SMALL)) {
8203 /* Makes no sense without HASH_EARLY */
8204 WARN_ON(!(flags & HASH_EARLY));
8205 if (!(numentries >> *_hash_shift)) {
8206 numentries = 1UL << *_hash_shift;
8207 BUG_ON(!numentries);
8209 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8210 numentries = PAGE_SIZE / bucketsize;
8212 numentries = roundup_pow_of_two(numentries);
8214 /* limit allocation size to 1/16 total memory by default */
8216 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8217 do_div(max, bucketsize);
8219 max = min(max, 0x80000000ULL);
8221 if (numentries < low_limit)
8222 numentries = low_limit;
8223 if (numentries > max)
8226 log2qty = ilog2(numentries);
8228 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8231 size = bucketsize << log2qty;
8232 if (flags & HASH_EARLY) {
8233 if (flags & HASH_ZERO)
8234 table = memblock_alloc(size, SMP_CACHE_BYTES);
8236 table = memblock_alloc_raw(size,
8238 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8239 table = __vmalloc(size, gfp_flags);
8243 * If bucketsize is not a power-of-two, we may free
8244 * some pages at the end of hash table which
8245 * alloc_pages_exact() automatically does
8247 table = alloc_pages_exact(size, gfp_flags);
8248 kmemleak_alloc(table, size, 1, gfp_flags);
8250 } while (!table && size > PAGE_SIZE && --log2qty);
8253 panic("Failed to allocate %s hash table\n", tablename);
8255 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8256 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8257 virt ? "vmalloc" : "linear");
8260 *_hash_shift = log2qty;
8262 *_hash_mask = (1 << log2qty) - 1;
8268 * This function checks whether pageblock includes unmovable pages or not.
8270 * PageLRU check without isolation or lru_lock could race so that
8271 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8272 * check without lock_page also may miss some movable non-lru pages at
8273 * race condition. So you can't expect this function should be exact.
8275 * Returns a page without holding a reference. If the caller wants to
8276 * dereference that page (e.g., dumping), it has to make sure that it
8277 * cannot get removed (e.g., via memory unplug) concurrently.
8280 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8281 int migratetype, int flags)
8283 unsigned long iter = 0;
8284 unsigned long pfn = page_to_pfn(page);
8285 unsigned long offset = pfn % pageblock_nr_pages;
8287 if (is_migrate_cma_page(page)) {
8289 * CMA allocations (alloc_contig_range) really need to mark
8290 * isolate CMA pageblocks even when they are not movable in fact
8291 * so consider them movable here.
8293 if (is_migrate_cma(migratetype))
8299 for (; iter < pageblock_nr_pages - offset; iter++) {
8300 if (!pfn_valid_within(pfn + iter))
8303 page = pfn_to_page(pfn + iter);
8306 * Both, bootmem allocations and memory holes are marked
8307 * PG_reserved and are unmovable. We can even have unmovable
8308 * allocations inside ZONE_MOVABLE, for example when
8309 * specifying "movablecore".
8311 if (PageReserved(page))
8315 * If the zone is movable and we have ruled out all reserved
8316 * pages then it should be reasonably safe to assume the rest
8319 if (zone_idx(zone) == ZONE_MOVABLE)
8323 * Hugepages are not in LRU lists, but they're movable.
8324 * THPs are on the LRU, but need to be counted as #small pages.
8325 * We need not scan over tail pages because we don't
8326 * handle each tail page individually in migration.
8328 if (PageHuge(page) || PageTransCompound(page)) {
8329 struct page *head = compound_head(page);
8330 unsigned int skip_pages;
8332 if (PageHuge(page)) {
8333 if (!hugepage_migration_supported(page_hstate(head)))
8335 } else if (!PageLRU(head) && !__PageMovable(head)) {
8339 skip_pages = compound_nr(head) - (page - head);
8340 iter += skip_pages - 1;
8345 * We can't use page_count without pin a page
8346 * because another CPU can free compound page.
8347 * This check already skips compound tails of THP
8348 * because their page->_refcount is zero at all time.
8350 if (!page_ref_count(page)) {
8351 if (PageBuddy(page))
8352 iter += (1 << buddy_order(page)) - 1;
8357 * The HWPoisoned page may be not in buddy system, and
8358 * page_count() is not 0.
8360 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8364 * We treat all PageOffline() pages as movable when offlining
8365 * to give drivers a chance to decrement their reference count
8366 * in MEM_GOING_OFFLINE in order to indicate that these pages
8367 * can be offlined as there are no direct references anymore.
8368 * For actually unmovable PageOffline() where the driver does
8369 * not support this, we will fail later when trying to actually
8370 * move these pages that still have a reference count > 0.
8371 * (false negatives in this function only)
8373 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8376 if (__PageMovable(page) || PageLRU(page))
8380 * If there are RECLAIMABLE pages, we need to check
8381 * it. But now, memory offline itself doesn't call
8382 * shrink_node_slabs() and it still to be fixed.
8389 #ifdef CONFIG_CONTIG_ALLOC
8390 static unsigned long pfn_max_align_down(unsigned long pfn)
8392 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8393 pageblock_nr_pages) - 1);
8396 static unsigned long pfn_max_align_up(unsigned long pfn)
8398 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8399 pageblock_nr_pages));
8402 /* [start, end) must belong to a single zone. */
8403 static int __alloc_contig_migrate_range(struct compact_control *cc,
8404 unsigned long start, unsigned long end)
8406 /* This function is based on compact_zone() from compaction.c. */
8407 unsigned int nr_reclaimed;
8408 unsigned long pfn = start;
8409 unsigned int tries = 0;
8411 struct migration_target_control mtc = {
8412 .nid = zone_to_nid(cc->zone),
8413 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8418 while (pfn < end || !list_empty(&cc->migratepages)) {
8419 if (fatal_signal_pending(current)) {
8424 if (list_empty(&cc->migratepages)) {
8425 cc->nr_migratepages = 0;
8426 pfn = isolate_migratepages_range(cc, pfn, end);
8432 } else if (++tries == 5) {
8433 ret = ret < 0 ? ret : -EBUSY;
8437 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8439 cc->nr_migratepages -= nr_reclaimed;
8441 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8442 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8445 putback_movable_pages(&cc->migratepages);
8452 * alloc_contig_range() -- tries to allocate given range of pages
8453 * @start: start PFN to allocate
8454 * @end: one-past-the-last PFN to allocate
8455 * @migratetype: migratetype of the underlaying pageblocks (either
8456 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8457 * in range must have the same migratetype and it must
8458 * be either of the two.
8459 * @gfp_mask: GFP mask to use during compaction
8461 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8462 * aligned. The PFN range must belong to a single zone.
8464 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8465 * pageblocks in the range. Once isolated, the pageblocks should not
8466 * be modified by others.
8468 * Return: zero on success or negative error code. On success all
8469 * pages which PFN is in [start, end) are allocated for the caller and
8470 * need to be freed with free_contig_range().
8472 int alloc_contig_range(unsigned long start, unsigned long end,
8473 unsigned migratetype, gfp_t gfp_mask)
8475 unsigned long outer_start, outer_end;
8479 struct compact_control cc = {
8480 .nr_migratepages = 0,
8482 .zone = page_zone(pfn_to_page(start)),
8483 .mode = MIGRATE_SYNC,
8484 .ignore_skip_hint = true,
8485 .no_set_skip_hint = true,
8486 .gfp_mask = current_gfp_context(gfp_mask),
8487 .alloc_contig = true,
8489 INIT_LIST_HEAD(&cc.migratepages);
8492 * What we do here is we mark all pageblocks in range as
8493 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8494 * have different sizes, and due to the way page allocator
8495 * work, we align the range to biggest of the two pages so
8496 * that page allocator won't try to merge buddies from
8497 * different pageblocks and change MIGRATE_ISOLATE to some
8498 * other migration type.
8500 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8501 * migrate the pages from an unaligned range (ie. pages that
8502 * we are interested in). This will put all the pages in
8503 * range back to page allocator as MIGRATE_ISOLATE.
8505 * When this is done, we take the pages in range from page
8506 * allocator removing them from the buddy system. This way
8507 * page allocator will never consider using them.
8509 * This lets us mark the pageblocks back as
8510 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8511 * aligned range but not in the unaligned, original range are
8512 * put back to page allocator so that buddy can use them.
8515 ret = start_isolate_page_range(pfn_max_align_down(start),
8516 pfn_max_align_up(end), migratetype, 0);
8521 * In case of -EBUSY, we'd like to know which page causes problem.
8522 * So, just fall through. test_pages_isolated() has a tracepoint
8523 * which will report the busy page.
8525 * It is possible that busy pages could become available before
8526 * the call to test_pages_isolated, and the range will actually be
8527 * allocated. So, if we fall through be sure to clear ret so that
8528 * -EBUSY is not accidentally used or returned to caller.
8530 ret = __alloc_contig_migrate_range(&cc, start, end);
8531 if (ret && ret != -EBUSY)
8536 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8537 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8538 * more, all pages in [start, end) are free in page allocator.
8539 * What we are going to do is to allocate all pages from
8540 * [start, end) (that is remove them from page allocator).
8542 * The only problem is that pages at the beginning and at the
8543 * end of interesting range may be not aligned with pages that
8544 * page allocator holds, ie. they can be part of higher order
8545 * pages. Because of this, we reserve the bigger range and
8546 * once this is done free the pages we are not interested in.
8548 * We don't have to hold zone->lock here because the pages are
8549 * isolated thus they won't get removed from buddy.
8552 lru_add_drain_all();
8555 outer_start = start;
8556 while (!PageBuddy(pfn_to_page(outer_start))) {
8557 if (++order >= MAX_ORDER) {
8558 outer_start = start;
8561 outer_start &= ~0UL << order;
8564 if (outer_start != start) {
8565 order = buddy_order(pfn_to_page(outer_start));
8568 * outer_start page could be small order buddy page and
8569 * it doesn't include start page. Adjust outer_start
8570 * in this case to report failed page properly
8571 * on tracepoint in test_pages_isolated()
8573 if (outer_start + (1UL << order) <= start)
8574 outer_start = start;
8577 /* Make sure the range is really isolated. */
8578 if (test_pages_isolated(outer_start, end, 0)) {
8579 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8580 __func__, outer_start, end);
8585 /* Grab isolated pages from freelists. */
8586 outer_end = isolate_freepages_range(&cc, outer_start, end);
8592 /* Free head and tail (if any) */
8593 if (start != outer_start)
8594 free_contig_range(outer_start, start - outer_start);
8595 if (end != outer_end)
8596 free_contig_range(end, outer_end - end);
8599 undo_isolate_page_range(pfn_max_align_down(start),
8600 pfn_max_align_up(end), migratetype);
8603 EXPORT_SYMBOL(alloc_contig_range);
8605 static int __alloc_contig_pages(unsigned long start_pfn,
8606 unsigned long nr_pages, gfp_t gfp_mask)
8608 unsigned long end_pfn = start_pfn + nr_pages;
8610 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8614 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8615 unsigned long nr_pages)
8617 unsigned long i, end_pfn = start_pfn + nr_pages;
8620 for (i = start_pfn; i < end_pfn; i++) {
8621 page = pfn_to_online_page(i);
8625 if (page_zone(page) != z)
8628 if (PageReserved(page))
8631 if (page_count(page) > 0)
8640 static bool zone_spans_last_pfn(const struct zone *zone,
8641 unsigned long start_pfn, unsigned long nr_pages)
8643 unsigned long last_pfn = start_pfn + nr_pages - 1;
8645 return zone_spans_pfn(zone, last_pfn);
8649 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8650 * @nr_pages: Number of contiguous pages to allocate
8651 * @gfp_mask: GFP mask to limit search and used during compaction
8653 * @nodemask: Mask for other possible nodes
8655 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8656 * on an applicable zonelist to find a contiguous pfn range which can then be
8657 * tried for allocation with alloc_contig_range(). This routine is intended
8658 * for allocation requests which can not be fulfilled with the buddy allocator.
8660 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8661 * power of two then the alignment is guaranteed to be to the given nr_pages
8662 * (e.g. 1GB request would be aligned to 1GB).
8664 * Allocated pages can be freed with free_contig_range() or by manually calling
8665 * __free_page() on each allocated page.
8667 * Return: pointer to contiguous pages on success, or NULL if not successful.
8669 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8670 int nid, nodemask_t *nodemask)
8672 unsigned long ret, pfn, flags;
8673 struct zonelist *zonelist;
8677 zonelist = node_zonelist(nid, gfp_mask);
8678 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8679 gfp_zone(gfp_mask), nodemask) {
8680 spin_lock_irqsave(&zone->lock, flags);
8682 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8683 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8684 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8686 * We release the zone lock here because
8687 * alloc_contig_range() will also lock the zone
8688 * at some point. If there's an allocation
8689 * spinning on this lock, it may win the race
8690 * and cause alloc_contig_range() to fail...
8692 spin_unlock_irqrestore(&zone->lock, flags);
8693 ret = __alloc_contig_pages(pfn, nr_pages,
8696 return pfn_to_page(pfn);
8697 spin_lock_irqsave(&zone->lock, flags);
8701 spin_unlock_irqrestore(&zone->lock, flags);
8705 #endif /* CONFIG_CONTIG_ALLOC */
8707 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8709 unsigned int count = 0;
8711 for (; nr_pages--; pfn++) {
8712 struct page *page = pfn_to_page(pfn);
8714 count += page_count(page) != 1;
8717 WARN(count != 0, "%d pages are still in use!\n", count);
8719 EXPORT_SYMBOL(free_contig_range);
8722 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8723 * page high values need to be recalulated.
8725 void __meminit zone_pcp_update(struct zone *zone)
8727 mutex_lock(&pcp_batch_high_lock);
8728 __zone_pcp_update(zone);
8729 mutex_unlock(&pcp_batch_high_lock);
8732 void zone_pcp_reset(struct zone *zone)
8734 unsigned long flags;
8736 struct per_cpu_pageset *pset;
8738 /* avoid races with drain_pages() */
8739 local_irq_save(flags);
8740 if (zone->pageset != &boot_pageset) {
8741 for_each_online_cpu(cpu) {
8742 pset = per_cpu_ptr(zone->pageset, cpu);
8743 drain_zonestat(zone, pset);
8745 free_percpu(zone->pageset);
8746 zone->pageset = &boot_pageset;
8748 local_irq_restore(flags);
8751 #ifdef CONFIG_MEMORY_HOTREMOVE
8753 * All pages in the range must be in a single zone, must not contain holes,
8754 * must span full sections, and must be isolated before calling this function.
8756 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8758 unsigned long pfn = start_pfn;
8762 unsigned long flags;
8764 offline_mem_sections(pfn, end_pfn);
8765 zone = page_zone(pfn_to_page(pfn));
8766 spin_lock_irqsave(&zone->lock, flags);
8767 while (pfn < end_pfn) {
8768 page = pfn_to_page(pfn);
8770 * The HWPoisoned page may be not in buddy system, and
8771 * page_count() is not 0.
8773 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8778 * At this point all remaining PageOffline() pages have a
8779 * reference count of 0 and can simply be skipped.
8781 if (PageOffline(page)) {
8782 BUG_ON(page_count(page));
8783 BUG_ON(PageBuddy(page));
8788 BUG_ON(page_count(page));
8789 BUG_ON(!PageBuddy(page));
8790 order = buddy_order(page);
8791 del_page_from_free_list(page, zone, order);
8792 pfn += (1 << order);
8794 spin_unlock_irqrestore(&zone->lock, flags);
8798 bool is_free_buddy_page(struct page *page)
8800 struct zone *zone = page_zone(page);
8801 unsigned long pfn = page_to_pfn(page);
8802 unsigned long flags;
8805 spin_lock_irqsave(&zone->lock, flags);
8806 for (order = 0; order < MAX_ORDER; order++) {
8807 struct page *page_head = page - (pfn & ((1 << order) - 1));
8809 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
8812 spin_unlock_irqrestore(&zone->lock, flags);
8814 return order < MAX_ORDER;
8817 #ifdef CONFIG_MEMORY_FAILURE
8819 * Break down a higher-order page in sub-pages, and keep our target out of
8822 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8823 struct page *target, int low, int high,
8826 unsigned long size = 1 << high;
8827 struct page *current_buddy, *next_page;
8829 while (high > low) {
8833 if (target >= &page[size]) {
8834 next_page = page + size;
8835 current_buddy = page;
8838 current_buddy = page + size;
8841 if (set_page_guard(zone, current_buddy, high, migratetype))
8844 if (current_buddy != target) {
8845 add_to_free_list(current_buddy, zone, high, migratetype);
8846 set_buddy_order(current_buddy, high);
8853 * Take a page that will be marked as poisoned off the buddy allocator.
8855 bool take_page_off_buddy(struct page *page)
8857 struct zone *zone = page_zone(page);
8858 unsigned long pfn = page_to_pfn(page);
8859 unsigned long flags;
8863 spin_lock_irqsave(&zone->lock, flags);
8864 for (order = 0; order < MAX_ORDER; order++) {
8865 struct page *page_head = page - (pfn & ((1 << order) - 1));
8866 int page_order = buddy_order(page_head);
8868 if (PageBuddy(page_head) && page_order >= order) {
8869 unsigned long pfn_head = page_to_pfn(page_head);
8870 int migratetype = get_pfnblock_migratetype(page_head,
8873 del_page_from_free_list(page_head, zone, page_order);
8874 break_down_buddy_pages(zone, page_head, page, 0,
8875 page_order, migratetype);
8879 if (page_count(page_head) > 0)
8882 spin_unlock_irqrestore(&zone->lock, flags);