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)))
451 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
454 * We start only with one section of pages, more pages are added as
455 * needed until the rest of deferred pages are initialized.
458 if ((nr_initialised > PAGES_PER_SECTION) &&
459 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
460 NODE_DATA(nid)->first_deferred_pfn = pfn;
466 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
468 static inline bool early_page_uninitialised(unsigned long pfn)
473 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
479 /* Return a pointer to the bitmap storing bits affecting a block of pages */
480 static inline unsigned long *get_pageblock_bitmap(struct page *page,
483 #ifdef CONFIG_SPARSEMEM
484 return section_to_usemap(__pfn_to_section(pfn));
486 return page_zone(page)->pageblock_flags;
487 #endif /* CONFIG_SPARSEMEM */
490 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
492 #ifdef CONFIG_SPARSEMEM
493 pfn &= (PAGES_PER_SECTION-1);
495 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
496 #endif /* CONFIG_SPARSEMEM */
497 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
501 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
502 * @page: The page within the block of interest
503 * @pfn: The target page frame number
504 * @mask: mask of bits that the caller is interested in
506 * Return: pageblock_bits flags
508 static __always_inline
509 unsigned long __get_pfnblock_flags_mask(struct page *page,
513 unsigned long *bitmap;
514 unsigned long bitidx, word_bitidx;
517 bitmap = get_pageblock_bitmap(page, pfn);
518 bitidx = pfn_to_bitidx(page, pfn);
519 word_bitidx = bitidx / BITS_PER_LONG;
520 bitidx &= (BITS_PER_LONG-1);
522 word = bitmap[word_bitidx];
523 return (word >> bitidx) & mask;
526 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
529 return __get_pfnblock_flags_mask(page, pfn, mask);
532 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
534 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
538 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
539 * @page: The page within the block of interest
540 * @flags: The flags to set
541 * @pfn: The target page frame number
542 * @mask: mask of bits that the caller is interested in
544 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
548 unsigned long *bitmap;
549 unsigned long bitidx, word_bitidx;
550 unsigned long old_word, word;
552 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
553 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
555 bitmap = get_pageblock_bitmap(page, pfn);
556 bitidx = pfn_to_bitidx(page, pfn);
557 word_bitidx = bitidx / BITS_PER_LONG;
558 bitidx &= (BITS_PER_LONG-1);
560 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
565 word = READ_ONCE(bitmap[word_bitidx]);
567 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
568 if (word == old_word)
574 void set_pageblock_migratetype(struct page *page, int migratetype)
576 if (unlikely(page_group_by_mobility_disabled &&
577 migratetype < MIGRATE_PCPTYPES))
578 migratetype = MIGRATE_UNMOVABLE;
580 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
581 page_to_pfn(page), MIGRATETYPE_MASK);
584 #ifdef CONFIG_DEBUG_VM
585 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
589 unsigned long pfn = page_to_pfn(page);
590 unsigned long sp, start_pfn;
593 seq = zone_span_seqbegin(zone);
594 start_pfn = zone->zone_start_pfn;
595 sp = zone->spanned_pages;
596 if (!zone_spans_pfn(zone, pfn))
598 } while (zone_span_seqretry(zone, seq));
601 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
602 pfn, zone_to_nid(zone), zone->name,
603 start_pfn, start_pfn + sp);
608 static int page_is_consistent(struct zone *zone, struct page *page)
610 if (!pfn_valid_within(page_to_pfn(page)))
612 if (zone != page_zone(page))
618 * Temporary debugging check for pages not lying within a given zone.
620 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
622 if (page_outside_zone_boundaries(zone, page))
624 if (!page_is_consistent(zone, page))
630 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
636 static void bad_page(struct page *page, const char *reason)
638 static unsigned long resume;
639 static unsigned long nr_shown;
640 static unsigned long nr_unshown;
643 * Allow a burst of 60 reports, then keep quiet for that minute;
644 * or allow a steady drip of one report per second.
646 if (nr_shown == 60) {
647 if (time_before(jiffies, resume)) {
653 "BUG: Bad page state: %lu messages suppressed\n",
660 resume = jiffies + 60 * HZ;
662 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
663 current->comm, page_to_pfn(page));
664 __dump_page(page, reason);
665 dump_page_owner(page);
670 /* Leave bad fields for debug, except PageBuddy could make trouble */
671 page_mapcount_reset(page); /* remove PageBuddy */
672 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
676 * Higher-order pages are called "compound pages". They are structured thusly:
678 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
680 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
681 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
683 * The first tail page's ->compound_dtor holds the offset in array of compound
684 * page destructors. See compound_page_dtors.
686 * The first tail page's ->compound_order holds the order of allocation.
687 * This usage means that zero-order pages may not be compound.
690 void free_compound_page(struct page *page)
692 mem_cgroup_uncharge(page);
693 __free_pages_ok(page, compound_order(page), FPI_NONE);
696 void prep_compound_page(struct page *page, unsigned int order)
699 int nr_pages = 1 << order;
702 for (i = 1; i < nr_pages; i++) {
703 struct page *p = page + i;
704 set_page_count(p, 0);
705 p->mapping = TAIL_MAPPING;
706 set_compound_head(p, page);
709 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
710 set_compound_order(page, order);
711 atomic_set(compound_mapcount_ptr(page), -1);
712 if (hpage_pincount_available(page))
713 atomic_set(compound_pincount_ptr(page), 0);
716 #ifdef CONFIG_DEBUG_PAGEALLOC
717 unsigned int _debug_guardpage_minorder;
719 bool _debug_pagealloc_enabled_early __read_mostly
720 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
721 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
722 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
723 EXPORT_SYMBOL(_debug_pagealloc_enabled);
725 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
727 static int __init early_debug_pagealloc(char *buf)
729 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
731 early_param("debug_pagealloc", early_debug_pagealloc);
733 void init_debug_pagealloc(void)
735 if (!debug_pagealloc_enabled())
738 static_branch_enable(&_debug_pagealloc_enabled);
740 if (!debug_guardpage_minorder())
743 static_branch_enable(&_debug_guardpage_enabled);
746 static int __init debug_guardpage_minorder_setup(char *buf)
750 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
751 pr_err("Bad debug_guardpage_minorder value\n");
754 _debug_guardpage_minorder = res;
755 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
758 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
760 static inline bool set_page_guard(struct zone *zone, struct page *page,
761 unsigned int order, int migratetype)
763 if (!debug_guardpage_enabled())
766 if (order >= debug_guardpage_minorder())
769 __SetPageGuard(page);
770 INIT_LIST_HEAD(&page->lru);
771 set_page_private(page, order);
772 /* Guard pages are not available for any usage */
773 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
778 static inline void clear_page_guard(struct zone *zone, struct page *page,
779 unsigned int order, int migratetype)
781 if (!debug_guardpage_enabled())
784 __ClearPageGuard(page);
786 set_page_private(page, 0);
787 if (!is_migrate_isolate(migratetype))
788 __mod_zone_freepage_state(zone, (1 << order), migratetype);
791 static inline bool set_page_guard(struct zone *zone, struct page *page,
792 unsigned int order, int migratetype) { return false; }
793 static inline void clear_page_guard(struct zone *zone, struct page *page,
794 unsigned int order, int migratetype) {}
797 static inline void set_buddy_order(struct page *page, unsigned int order)
799 set_page_private(page, order);
800 __SetPageBuddy(page);
804 * This function checks whether a page is free && is the buddy
805 * we can coalesce a page and its buddy if
806 * (a) the buddy is not in a hole (check before calling!) &&
807 * (b) the buddy is in the buddy system &&
808 * (c) a page and its buddy have the same order &&
809 * (d) a page and its buddy are in the same zone.
811 * For recording whether a page is in the buddy system, we set PageBuddy.
812 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
814 * For recording page's order, we use page_private(page).
816 static inline bool page_is_buddy(struct page *page, struct page *buddy,
819 if (!page_is_guard(buddy) && !PageBuddy(buddy))
822 if (buddy_order(buddy) != order)
826 * zone check is done late to avoid uselessly calculating
827 * zone/node ids for pages that could never merge.
829 if (page_zone_id(page) != page_zone_id(buddy))
832 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
837 #ifdef CONFIG_COMPACTION
838 static inline struct capture_control *task_capc(struct zone *zone)
840 struct capture_control *capc = current->capture_control;
842 return unlikely(capc) &&
843 !(current->flags & PF_KTHREAD) &&
845 capc->cc->zone == zone ? capc : NULL;
849 compaction_capture(struct capture_control *capc, struct page *page,
850 int order, int migratetype)
852 if (!capc || order != capc->cc->order)
855 /* Do not accidentally pollute CMA or isolated regions*/
856 if (is_migrate_cma(migratetype) ||
857 is_migrate_isolate(migratetype))
861 * Do not let lower order allocations polluate a movable pageblock.
862 * This might let an unmovable request use a reclaimable pageblock
863 * and vice-versa but no more than normal fallback logic which can
864 * have trouble finding a high-order free page.
866 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
874 static inline struct capture_control *task_capc(struct zone *zone)
880 compaction_capture(struct capture_control *capc, struct page *page,
881 int order, int migratetype)
885 #endif /* CONFIG_COMPACTION */
887 /* Used for pages not on another list */
888 static inline void add_to_free_list(struct page *page, struct zone *zone,
889 unsigned int order, int migratetype)
891 struct free_area *area = &zone->free_area[order];
893 list_add(&page->lru, &area->free_list[migratetype]);
897 /* Used for pages not on another list */
898 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
899 unsigned int order, int migratetype)
901 struct free_area *area = &zone->free_area[order];
903 list_add_tail(&page->lru, &area->free_list[migratetype]);
908 * Used for pages which are on another list. Move the pages to the tail
909 * of the list - so the moved pages won't immediately be considered for
910 * allocation again (e.g., optimization for memory onlining).
912 static inline void move_to_free_list(struct page *page, struct zone *zone,
913 unsigned int order, int migratetype)
915 struct free_area *area = &zone->free_area[order];
917 list_move_tail(&page->lru, &area->free_list[migratetype]);
920 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
923 /* clear reported state and update reported page count */
924 if (page_reported(page))
925 __ClearPageReported(page);
927 list_del(&page->lru);
928 __ClearPageBuddy(page);
929 set_page_private(page, 0);
930 zone->free_area[order].nr_free--;
934 * If this is not the largest possible page, check if the buddy
935 * of the next-highest order is free. If it is, it's possible
936 * that pages are being freed that will coalesce soon. In case,
937 * that is happening, add the free page to the tail of the list
938 * so it's less likely to be used soon and more likely to be merged
939 * as a higher order page
942 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
943 struct page *page, unsigned int order)
945 struct page *higher_page, *higher_buddy;
946 unsigned long combined_pfn;
948 if (order >= MAX_ORDER - 2)
951 if (!pfn_valid_within(buddy_pfn))
954 combined_pfn = buddy_pfn & pfn;
955 higher_page = page + (combined_pfn - pfn);
956 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
957 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
959 return pfn_valid_within(buddy_pfn) &&
960 page_is_buddy(higher_page, higher_buddy, order + 1);
964 * Freeing function for a buddy system allocator.
966 * The concept of a buddy system is to maintain direct-mapped table
967 * (containing bit values) for memory blocks of various "orders".
968 * The bottom level table contains the map for the smallest allocatable
969 * units of memory (here, pages), and each level above it describes
970 * pairs of units from the levels below, hence, "buddies".
971 * At a high level, all that happens here is marking the table entry
972 * at the bottom level available, and propagating the changes upward
973 * as necessary, plus some accounting needed to play nicely with other
974 * parts of the VM system.
975 * At each level, we keep a list of pages, which are heads of continuous
976 * free pages of length of (1 << order) and marked with PageBuddy.
977 * Page's order is recorded in page_private(page) field.
978 * So when we are allocating or freeing one, we can derive the state of the
979 * other. That is, if we allocate a small block, and both were
980 * free, the remainder of the region must be split into blocks.
981 * If a block is freed, and its buddy is also free, then this
982 * triggers coalescing into a block of larger size.
987 static inline void __free_one_page(struct page *page,
989 struct zone *zone, unsigned int order,
990 int migratetype, fpi_t fpi_flags)
992 struct capture_control *capc = task_capc(zone);
993 unsigned long buddy_pfn;
994 unsigned long combined_pfn;
995 unsigned int max_order;
999 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
1001 VM_BUG_ON(!zone_is_initialized(zone));
1002 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1004 VM_BUG_ON(migratetype == -1);
1005 if (likely(!is_migrate_isolate(migratetype)))
1006 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1008 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1009 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1012 while (order < max_order - 1) {
1013 if (compaction_capture(capc, page, order, migratetype)) {
1014 __mod_zone_freepage_state(zone, -(1 << order),
1018 buddy_pfn = __find_buddy_pfn(pfn, order);
1019 buddy = page + (buddy_pfn - pfn);
1021 if (!pfn_valid_within(buddy_pfn))
1023 if (!page_is_buddy(page, buddy, order))
1026 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1027 * merge with it and move up one order.
1029 if (page_is_guard(buddy))
1030 clear_page_guard(zone, buddy, order, migratetype);
1032 del_page_from_free_list(buddy, zone, order);
1033 combined_pfn = buddy_pfn & pfn;
1034 page = page + (combined_pfn - pfn);
1038 if (max_order < MAX_ORDER) {
1039 /* If we are here, it means order is >= pageblock_order.
1040 * We want to prevent merge between freepages on isolate
1041 * pageblock and normal pageblock. Without this, pageblock
1042 * isolation could cause incorrect freepage or CMA accounting.
1044 * We don't want to hit this code for the more frequent
1045 * low-order merging.
1047 if (unlikely(has_isolate_pageblock(zone))) {
1050 buddy_pfn = __find_buddy_pfn(pfn, order);
1051 buddy = page + (buddy_pfn - pfn);
1052 buddy_mt = get_pageblock_migratetype(buddy);
1054 if (migratetype != buddy_mt
1055 && (is_migrate_isolate(migratetype) ||
1056 is_migrate_isolate(buddy_mt)))
1060 goto continue_merging;
1064 set_buddy_order(page, order);
1066 if (fpi_flags & FPI_TO_TAIL)
1068 else if (is_shuffle_order(order))
1069 to_tail = shuffle_pick_tail();
1071 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1074 add_to_free_list_tail(page, zone, order, migratetype);
1076 add_to_free_list(page, zone, order, migratetype);
1078 /* Notify page reporting subsystem of freed page */
1079 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1080 page_reporting_notify_free(order);
1084 * A bad page could be due to a number of fields. Instead of multiple branches,
1085 * try and check multiple fields with one check. The caller must do a detailed
1086 * check if necessary.
1088 static inline bool page_expected_state(struct page *page,
1089 unsigned long check_flags)
1091 if (unlikely(atomic_read(&page->_mapcount) != -1))
1094 if (unlikely((unsigned long)page->mapping |
1095 page_ref_count(page) |
1097 (unsigned long)page->mem_cgroup |
1099 (page->flags & check_flags)))
1105 static const char *page_bad_reason(struct page *page, unsigned long flags)
1107 const char *bad_reason = NULL;
1109 if (unlikely(atomic_read(&page->_mapcount) != -1))
1110 bad_reason = "nonzero mapcount";
1111 if (unlikely(page->mapping != NULL))
1112 bad_reason = "non-NULL mapping";
1113 if (unlikely(page_ref_count(page) != 0))
1114 bad_reason = "nonzero _refcount";
1115 if (unlikely(page->flags & flags)) {
1116 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1117 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1119 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1122 if (unlikely(page->mem_cgroup))
1123 bad_reason = "page still charged to cgroup";
1128 static void check_free_page_bad(struct page *page)
1131 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1134 static inline int check_free_page(struct page *page)
1136 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1139 /* Something has gone sideways, find it */
1140 check_free_page_bad(page);
1144 static int free_tail_pages_check(struct page *head_page, struct page *page)
1149 * We rely page->lru.next never has bit 0 set, unless the page
1150 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1152 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1154 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1158 switch (page - head_page) {
1160 /* the first tail page: ->mapping may be compound_mapcount() */
1161 if (unlikely(compound_mapcount(page))) {
1162 bad_page(page, "nonzero compound_mapcount");
1168 * the second tail page: ->mapping is
1169 * deferred_list.next -- ignore value.
1173 if (page->mapping != TAIL_MAPPING) {
1174 bad_page(page, "corrupted mapping in tail page");
1179 if (unlikely(!PageTail(page))) {
1180 bad_page(page, "PageTail not set");
1183 if (unlikely(compound_head(page) != head_page)) {
1184 bad_page(page, "compound_head not consistent");
1189 page->mapping = NULL;
1190 clear_compound_head(page);
1194 static void kernel_init_free_pages(struct page *page, int numpages)
1198 /* s390's use of memset() could override KASAN redzones. */
1199 kasan_disable_current();
1200 for (i = 0; i < numpages; i++)
1201 clear_highpage(page + i);
1202 kasan_enable_current();
1205 static __always_inline bool free_pages_prepare(struct page *page,
1206 unsigned int order, bool check_free)
1210 VM_BUG_ON_PAGE(PageTail(page), page);
1212 trace_mm_page_free(page, order);
1214 if (unlikely(PageHWPoison(page)) && !order) {
1216 * Do not let hwpoison pages hit pcplists/buddy
1217 * Untie memcg state and reset page's owner
1219 if (memcg_kmem_enabled() && PageKmemcg(page))
1220 __memcg_kmem_uncharge_page(page, order);
1221 reset_page_owner(page, order);
1226 * Check tail pages before head page information is cleared to
1227 * avoid checking PageCompound for order-0 pages.
1229 if (unlikely(order)) {
1230 bool compound = PageCompound(page);
1233 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1236 ClearPageDoubleMap(page);
1237 for (i = 1; i < (1 << order); i++) {
1239 bad += free_tail_pages_check(page, page + i);
1240 if (unlikely(check_free_page(page + i))) {
1244 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1247 if (PageMappingFlags(page))
1248 page->mapping = NULL;
1249 if (memcg_kmem_enabled() && PageKmemcg(page))
1250 __memcg_kmem_uncharge_page(page, order);
1252 bad += check_free_page(page);
1256 page_cpupid_reset_last(page);
1257 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1258 reset_page_owner(page, order);
1260 if (!PageHighMem(page)) {
1261 debug_check_no_locks_freed(page_address(page),
1262 PAGE_SIZE << order);
1263 debug_check_no_obj_freed(page_address(page),
1264 PAGE_SIZE << order);
1266 if (want_init_on_free())
1267 kernel_init_free_pages(page, 1 << order);
1269 kernel_poison_pages(page, 1 << order, 0);
1271 * arch_free_page() can make the page's contents inaccessible. s390
1272 * does this. So nothing which can access the page's contents should
1273 * happen after this.
1275 arch_free_page(page, order);
1277 if (debug_pagealloc_enabled_static())
1278 kernel_map_pages(page, 1 << order, 0);
1280 kasan_free_nondeferred_pages(page, order);
1285 #ifdef CONFIG_DEBUG_VM
1287 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1288 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1289 * moved from pcp lists to free lists.
1291 static bool free_pcp_prepare(struct page *page)
1293 return free_pages_prepare(page, 0, true);
1296 static bool bulkfree_pcp_prepare(struct page *page)
1298 if (debug_pagealloc_enabled_static())
1299 return check_free_page(page);
1305 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1306 * moving from pcp lists to free list in order to reduce overhead. With
1307 * debug_pagealloc enabled, they are checked also immediately when being freed
1310 static bool free_pcp_prepare(struct page *page)
1312 if (debug_pagealloc_enabled_static())
1313 return free_pages_prepare(page, 0, true);
1315 return free_pages_prepare(page, 0, false);
1318 static bool bulkfree_pcp_prepare(struct page *page)
1320 return check_free_page(page);
1322 #endif /* CONFIG_DEBUG_VM */
1324 static inline void prefetch_buddy(struct page *page)
1326 unsigned long pfn = page_to_pfn(page);
1327 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1328 struct page *buddy = page + (buddy_pfn - pfn);
1334 * Frees a number of pages from the PCP lists
1335 * Assumes all pages on list are in same zone, and of same order.
1336 * count is the number of pages to free.
1338 * If the zone was previously in an "all pages pinned" state then look to
1339 * see if this freeing clears that state.
1341 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1342 * pinned" detection logic.
1344 static void free_pcppages_bulk(struct zone *zone, int count,
1345 struct per_cpu_pages *pcp)
1347 int migratetype = 0;
1349 int prefetch_nr = 0;
1350 bool isolated_pageblocks;
1351 struct page *page, *tmp;
1355 * Ensure proper count is passed which otherwise would stuck in the
1356 * below while (list_empty(list)) loop.
1358 count = min(pcp->count, count);
1360 struct list_head *list;
1363 * Remove pages from lists in a round-robin fashion. A
1364 * batch_free count is maintained that is incremented when an
1365 * empty list is encountered. This is so more pages are freed
1366 * off fuller lists instead of spinning excessively around empty
1371 if (++migratetype == MIGRATE_PCPTYPES)
1373 list = &pcp->lists[migratetype];
1374 } while (list_empty(list));
1376 /* This is the only non-empty list. Free them all. */
1377 if (batch_free == MIGRATE_PCPTYPES)
1381 page = list_last_entry(list, struct page, lru);
1382 /* must delete to avoid corrupting pcp list */
1383 list_del(&page->lru);
1386 if (bulkfree_pcp_prepare(page))
1389 list_add_tail(&page->lru, &head);
1392 * We are going to put the page back to the global
1393 * pool, prefetch its buddy to speed up later access
1394 * under zone->lock. It is believed the overhead of
1395 * an additional test and calculating buddy_pfn here
1396 * can be offset by reduced memory latency later. To
1397 * avoid excessive prefetching due to large count, only
1398 * prefetch buddy for the first pcp->batch nr of pages.
1400 if (prefetch_nr++ < pcp->batch)
1401 prefetch_buddy(page);
1402 } while (--count && --batch_free && !list_empty(list));
1405 spin_lock(&zone->lock);
1406 isolated_pageblocks = has_isolate_pageblock(zone);
1409 * Use safe version since after __free_one_page(),
1410 * page->lru.next will not point to original list.
1412 list_for_each_entry_safe(page, tmp, &head, lru) {
1413 int mt = get_pcppage_migratetype(page);
1414 /* MIGRATE_ISOLATE page should not go to pcplists */
1415 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1416 /* Pageblock could have been isolated meanwhile */
1417 if (unlikely(isolated_pageblocks))
1418 mt = get_pageblock_migratetype(page);
1420 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1421 trace_mm_page_pcpu_drain(page, 0, mt);
1423 spin_unlock(&zone->lock);
1426 static void free_one_page(struct zone *zone,
1427 struct page *page, unsigned long pfn,
1429 int migratetype, fpi_t fpi_flags)
1431 spin_lock(&zone->lock);
1432 if (unlikely(has_isolate_pageblock(zone) ||
1433 is_migrate_isolate(migratetype))) {
1434 migratetype = get_pfnblock_migratetype(page, pfn);
1436 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1437 spin_unlock(&zone->lock);
1440 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1441 unsigned long zone, int nid)
1443 mm_zero_struct_page(page);
1444 set_page_links(page, zone, nid, pfn);
1445 init_page_count(page);
1446 page_mapcount_reset(page);
1447 page_cpupid_reset_last(page);
1448 page_kasan_tag_reset(page);
1450 INIT_LIST_HEAD(&page->lru);
1451 #ifdef WANT_PAGE_VIRTUAL
1452 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1453 if (!is_highmem_idx(zone))
1454 set_page_address(page, __va(pfn << PAGE_SHIFT));
1458 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1459 static void __meminit init_reserved_page(unsigned long pfn)
1464 if (!early_page_uninitialised(pfn))
1467 nid = early_pfn_to_nid(pfn);
1468 pgdat = NODE_DATA(nid);
1470 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1471 struct zone *zone = &pgdat->node_zones[zid];
1473 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1476 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1479 static inline void init_reserved_page(unsigned long pfn)
1482 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1485 * Initialised pages do not have PageReserved set. This function is
1486 * called for each range allocated by the bootmem allocator and
1487 * marks the pages PageReserved. The remaining valid pages are later
1488 * sent to the buddy page allocator.
1490 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1492 unsigned long start_pfn = PFN_DOWN(start);
1493 unsigned long end_pfn = PFN_UP(end);
1495 for (; start_pfn < end_pfn; start_pfn++) {
1496 if (pfn_valid(start_pfn)) {
1497 struct page *page = pfn_to_page(start_pfn);
1499 init_reserved_page(start_pfn);
1501 /* Avoid false-positive PageTail() */
1502 INIT_LIST_HEAD(&page->lru);
1505 * no need for atomic set_bit because the struct
1506 * page is not visible yet so nobody should
1509 __SetPageReserved(page);
1514 static void __free_pages_ok(struct page *page, unsigned int order,
1517 unsigned long flags;
1519 unsigned long pfn = page_to_pfn(page);
1521 if (!free_pages_prepare(page, order, true))
1524 migratetype = get_pfnblock_migratetype(page, pfn);
1525 local_irq_save(flags);
1526 __count_vm_events(PGFREE, 1 << order);
1527 free_one_page(page_zone(page), page, pfn, order, migratetype,
1529 local_irq_restore(flags);
1532 void __free_pages_core(struct page *page, unsigned int order)
1534 unsigned int nr_pages = 1 << order;
1535 struct page *p = page;
1539 * When initializing the memmap, __init_single_page() sets the refcount
1540 * of all pages to 1 ("allocated"/"not free"). We have to set the
1541 * refcount of all involved pages to 0.
1544 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1546 __ClearPageReserved(p);
1547 set_page_count(p, 0);
1549 __ClearPageReserved(p);
1550 set_page_count(p, 0);
1552 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1555 * Bypass PCP and place fresh pages right to the tail, primarily
1556 * relevant for memory onlining.
1558 __free_pages_ok(page, order, FPI_TO_TAIL);
1561 #ifdef CONFIG_NEED_MULTIPLE_NODES
1563 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1565 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1568 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1570 int __meminit __early_pfn_to_nid(unsigned long pfn,
1571 struct mminit_pfnnid_cache *state)
1573 unsigned long start_pfn, end_pfn;
1576 if (state->last_start <= pfn && pfn < state->last_end)
1577 return state->last_nid;
1579 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1580 if (nid != NUMA_NO_NODE) {
1581 state->last_start = start_pfn;
1582 state->last_end = end_pfn;
1583 state->last_nid = nid;
1588 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1590 int __meminit early_pfn_to_nid(unsigned long pfn)
1592 static DEFINE_SPINLOCK(early_pfn_lock);
1595 spin_lock(&early_pfn_lock);
1596 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1598 nid = first_online_node;
1599 spin_unlock(&early_pfn_lock);
1603 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1605 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1608 if (early_page_uninitialised(pfn))
1610 __free_pages_core(page, order);
1614 * Check that the whole (or subset of) a pageblock given by the interval of
1615 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1616 * with the migration of free compaction scanner. The scanners then need to
1617 * use only pfn_valid_within() check for arches that allow holes within
1620 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1622 * It's possible on some configurations to have a setup like node0 node1 node0
1623 * i.e. it's possible that all pages within a zones range of pages do not
1624 * belong to a single zone. We assume that a border between node0 and node1
1625 * can occur within a single pageblock, but not a node0 node1 node0
1626 * interleaving within a single pageblock. It is therefore sufficient to check
1627 * the first and last page of a pageblock and avoid checking each individual
1628 * page in a pageblock.
1630 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1631 unsigned long end_pfn, struct zone *zone)
1633 struct page *start_page;
1634 struct page *end_page;
1636 /* end_pfn is one past the range we are checking */
1639 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1642 start_page = pfn_to_online_page(start_pfn);
1646 if (page_zone(start_page) != zone)
1649 end_page = pfn_to_page(end_pfn);
1651 /* This gives a shorter code than deriving page_zone(end_page) */
1652 if (page_zone_id(start_page) != page_zone_id(end_page))
1658 void set_zone_contiguous(struct zone *zone)
1660 unsigned long block_start_pfn = zone->zone_start_pfn;
1661 unsigned long block_end_pfn;
1663 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1664 for (; block_start_pfn < zone_end_pfn(zone);
1665 block_start_pfn = block_end_pfn,
1666 block_end_pfn += pageblock_nr_pages) {
1668 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1670 if (!__pageblock_pfn_to_page(block_start_pfn,
1671 block_end_pfn, zone))
1676 /* We confirm that there is no hole */
1677 zone->contiguous = true;
1680 void clear_zone_contiguous(struct zone *zone)
1682 zone->contiguous = false;
1685 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1686 static void __init deferred_free_range(unsigned long pfn,
1687 unsigned long nr_pages)
1695 page = pfn_to_page(pfn);
1697 /* Free a large naturally-aligned chunk if possible */
1698 if (nr_pages == pageblock_nr_pages &&
1699 (pfn & (pageblock_nr_pages - 1)) == 0) {
1700 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1701 __free_pages_core(page, pageblock_order);
1705 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1706 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1707 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1708 __free_pages_core(page, 0);
1712 /* Completion tracking for deferred_init_memmap() threads */
1713 static atomic_t pgdat_init_n_undone __initdata;
1714 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1716 static inline void __init pgdat_init_report_one_done(void)
1718 if (atomic_dec_and_test(&pgdat_init_n_undone))
1719 complete(&pgdat_init_all_done_comp);
1723 * Returns true if page needs to be initialized or freed to buddy allocator.
1725 * First we check if pfn is valid on architectures where it is possible to have
1726 * holes within pageblock_nr_pages. On systems where it is not possible, this
1727 * function is optimized out.
1729 * Then, we check if a current large page is valid by only checking the validity
1732 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1734 if (!pfn_valid_within(pfn))
1736 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1742 * Free pages to buddy allocator. Try to free aligned pages in
1743 * pageblock_nr_pages sizes.
1745 static void __init deferred_free_pages(unsigned long pfn,
1746 unsigned long end_pfn)
1748 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1749 unsigned long nr_free = 0;
1751 for (; pfn < end_pfn; pfn++) {
1752 if (!deferred_pfn_valid(pfn)) {
1753 deferred_free_range(pfn - nr_free, nr_free);
1755 } else if (!(pfn & nr_pgmask)) {
1756 deferred_free_range(pfn - nr_free, nr_free);
1762 /* Free the last block of pages to allocator */
1763 deferred_free_range(pfn - nr_free, nr_free);
1767 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1768 * by performing it only once every pageblock_nr_pages.
1769 * Return number of pages initialized.
1771 static unsigned long __init deferred_init_pages(struct zone *zone,
1773 unsigned long end_pfn)
1775 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1776 int nid = zone_to_nid(zone);
1777 unsigned long nr_pages = 0;
1778 int zid = zone_idx(zone);
1779 struct page *page = NULL;
1781 for (; pfn < end_pfn; pfn++) {
1782 if (!deferred_pfn_valid(pfn)) {
1785 } else if (!page || !(pfn & nr_pgmask)) {
1786 page = pfn_to_page(pfn);
1790 __init_single_page(page, pfn, zid, nid);
1797 * This function is meant to pre-load the iterator for the zone init.
1798 * Specifically it walks through the ranges until we are caught up to the
1799 * first_init_pfn value and exits there. If we never encounter the value we
1800 * return false indicating there are no valid ranges left.
1803 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1804 unsigned long *spfn, unsigned long *epfn,
1805 unsigned long first_init_pfn)
1810 * Start out by walking through the ranges in this zone that have
1811 * already been initialized. We don't need to do anything with them
1812 * so we just need to flush them out of the system.
1814 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1815 if (*epfn <= first_init_pfn)
1817 if (*spfn < first_init_pfn)
1818 *spfn = first_init_pfn;
1827 * Initialize and free pages. We do it in two loops: first we initialize
1828 * struct page, then free to buddy allocator, because while we are
1829 * freeing pages we can access pages that are ahead (computing buddy
1830 * page in __free_one_page()).
1832 * In order to try and keep some memory in the cache we have the loop
1833 * broken along max page order boundaries. This way we will not cause
1834 * any issues with the buddy page computation.
1836 static unsigned long __init
1837 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1838 unsigned long *end_pfn)
1840 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1841 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1842 unsigned long nr_pages = 0;
1845 /* First we loop through and initialize the page values */
1846 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1849 if (mo_pfn <= *start_pfn)
1852 t = min(mo_pfn, *end_pfn);
1853 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1855 if (mo_pfn < *end_pfn) {
1856 *start_pfn = mo_pfn;
1861 /* Reset values and now loop through freeing pages as needed */
1864 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1870 t = min(mo_pfn, epfn);
1871 deferred_free_pages(spfn, t);
1881 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1884 unsigned long spfn, epfn;
1885 struct zone *zone = arg;
1888 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1891 * Initialize and free pages in MAX_ORDER sized increments so that we
1892 * can avoid introducing any issues with the buddy allocator.
1894 while (spfn < end_pfn) {
1895 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1900 /* An arch may override for more concurrency. */
1902 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1907 /* Initialise remaining memory on a node */
1908 static int __init deferred_init_memmap(void *data)
1910 pg_data_t *pgdat = data;
1911 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1912 unsigned long spfn = 0, epfn = 0;
1913 unsigned long first_init_pfn, flags;
1914 unsigned long start = jiffies;
1916 int zid, max_threads;
1919 /* Bind memory initialisation thread to a local node if possible */
1920 if (!cpumask_empty(cpumask))
1921 set_cpus_allowed_ptr(current, cpumask);
1923 pgdat_resize_lock(pgdat, &flags);
1924 first_init_pfn = pgdat->first_deferred_pfn;
1925 if (first_init_pfn == ULONG_MAX) {
1926 pgdat_resize_unlock(pgdat, &flags);
1927 pgdat_init_report_one_done();
1931 /* Sanity check boundaries */
1932 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1933 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1934 pgdat->first_deferred_pfn = ULONG_MAX;
1937 * Once we unlock here, the zone cannot be grown anymore, thus if an
1938 * interrupt thread must allocate this early in boot, zone must be
1939 * pre-grown prior to start of deferred page initialization.
1941 pgdat_resize_unlock(pgdat, &flags);
1943 /* Only the highest zone is deferred so find it */
1944 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1945 zone = pgdat->node_zones + zid;
1946 if (first_init_pfn < zone_end_pfn(zone))
1950 /* If the zone is empty somebody else may have cleared out the zone */
1951 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1955 max_threads = deferred_page_init_max_threads(cpumask);
1957 while (spfn < epfn) {
1958 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1959 struct padata_mt_job job = {
1960 .thread_fn = deferred_init_memmap_chunk,
1963 .size = epfn_align - spfn,
1964 .align = PAGES_PER_SECTION,
1965 .min_chunk = PAGES_PER_SECTION,
1966 .max_threads = max_threads,
1969 padata_do_multithreaded(&job);
1970 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1974 /* Sanity check that the next zone really is unpopulated */
1975 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1977 pr_info("node %d deferred pages initialised in %ums\n",
1978 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1980 pgdat_init_report_one_done();
1985 * If this zone has deferred pages, try to grow it by initializing enough
1986 * deferred pages to satisfy the allocation specified by order, rounded up to
1987 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1988 * of SECTION_SIZE bytes by initializing struct pages in increments of
1989 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1991 * Return true when zone was grown, otherwise return false. We return true even
1992 * when we grow less than requested, to let the caller decide if there are
1993 * enough pages to satisfy the allocation.
1995 * Note: We use noinline because this function is needed only during boot, and
1996 * it is called from a __ref function _deferred_grow_zone. This way we are
1997 * making sure that it is not inlined into permanent text section.
1999 static noinline bool __init
2000 deferred_grow_zone(struct zone *zone, unsigned int order)
2002 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2003 pg_data_t *pgdat = zone->zone_pgdat;
2004 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2005 unsigned long spfn, epfn, flags;
2006 unsigned long nr_pages = 0;
2009 /* Only the last zone may have deferred pages */
2010 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2013 pgdat_resize_lock(pgdat, &flags);
2016 * If someone grew this zone while we were waiting for spinlock, return
2017 * true, as there might be enough pages already.
2019 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2020 pgdat_resize_unlock(pgdat, &flags);
2024 /* If the zone is empty somebody else may have cleared out the zone */
2025 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2026 first_deferred_pfn)) {
2027 pgdat->first_deferred_pfn = ULONG_MAX;
2028 pgdat_resize_unlock(pgdat, &flags);
2029 /* Retry only once. */
2030 return first_deferred_pfn != ULONG_MAX;
2034 * Initialize and free pages in MAX_ORDER sized increments so
2035 * that we can avoid introducing any issues with the buddy
2038 while (spfn < epfn) {
2039 /* update our first deferred PFN for this section */
2040 first_deferred_pfn = spfn;
2042 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2043 touch_nmi_watchdog();
2045 /* We should only stop along section boundaries */
2046 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2049 /* If our quota has been met we can stop here */
2050 if (nr_pages >= nr_pages_needed)
2054 pgdat->first_deferred_pfn = spfn;
2055 pgdat_resize_unlock(pgdat, &flags);
2057 return nr_pages > 0;
2061 * deferred_grow_zone() is __init, but it is called from
2062 * get_page_from_freelist() during early boot until deferred_pages permanently
2063 * disables this call. This is why we have refdata wrapper to avoid warning,
2064 * and to ensure that the function body gets unloaded.
2067 _deferred_grow_zone(struct zone *zone, unsigned int order)
2069 return deferred_grow_zone(zone, order);
2072 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2074 void __init page_alloc_init_late(void)
2079 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2081 /* There will be num_node_state(N_MEMORY) threads */
2082 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2083 for_each_node_state(nid, N_MEMORY) {
2084 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2087 /* Block until all are initialised */
2088 wait_for_completion(&pgdat_init_all_done_comp);
2091 * The number of managed pages has changed due to the initialisation
2092 * so the pcpu batch and high limits needs to be updated or the limits
2093 * will be artificially small.
2095 for_each_populated_zone(zone)
2096 zone_pcp_update(zone);
2099 * We initialized the rest of the deferred pages. Permanently disable
2100 * on-demand struct page initialization.
2102 static_branch_disable(&deferred_pages);
2104 /* Reinit limits that are based on free pages after the kernel is up */
2105 files_maxfiles_init();
2108 /* Discard memblock private memory */
2111 for_each_node_state(nid, N_MEMORY)
2112 shuffle_free_memory(NODE_DATA(nid));
2114 for_each_populated_zone(zone)
2115 set_zone_contiguous(zone);
2119 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2120 void __init init_cma_reserved_pageblock(struct page *page)
2122 unsigned i = pageblock_nr_pages;
2123 struct page *p = page;
2126 __ClearPageReserved(p);
2127 set_page_count(p, 0);
2130 set_pageblock_migratetype(page, MIGRATE_CMA);
2132 if (pageblock_order >= MAX_ORDER) {
2133 i = pageblock_nr_pages;
2136 set_page_refcounted(p);
2137 __free_pages(p, MAX_ORDER - 1);
2138 p += MAX_ORDER_NR_PAGES;
2139 } while (i -= MAX_ORDER_NR_PAGES);
2141 set_page_refcounted(page);
2142 __free_pages(page, pageblock_order);
2145 adjust_managed_page_count(page, pageblock_nr_pages);
2150 * The order of subdivision here is critical for the IO subsystem.
2151 * Please do not alter this order without good reasons and regression
2152 * testing. Specifically, as large blocks of memory are subdivided,
2153 * the order in which smaller blocks are delivered depends on the order
2154 * they're subdivided in this function. This is the primary factor
2155 * influencing the order in which pages are delivered to the IO
2156 * subsystem according to empirical testing, and this is also justified
2157 * by considering the behavior of a buddy system containing a single
2158 * large block of memory acted on by a series of small allocations.
2159 * This behavior is a critical factor in sglist merging's success.
2163 static inline void expand(struct zone *zone, struct page *page,
2164 int low, int high, int migratetype)
2166 unsigned long size = 1 << high;
2168 while (high > low) {
2171 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2174 * Mark as guard pages (or page), that will allow to
2175 * merge back to allocator when buddy will be freed.
2176 * Corresponding page table entries will not be touched,
2177 * pages will stay not present in virtual address space
2179 if (set_page_guard(zone, &page[size], high, migratetype))
2182 add_to_free_list(&page[size], zone, high, migratetype);
2183 set_buddy_order(&page[size], high);
2187 static void check_new_page_bad(struct page *page)
2189 if (unlikely(page->flags & __PG_HWPOISON)) {
2190 /* Don't complain about hwpoisoned pages */
2191 page_mapcount_reset(page); /* remove PageBuddy */
2196 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2200 * This page is about to be returned from the page allocator
2202 static inline int check_new_page(struct page *page)
2204 if (likely(page_expected_state(page,
2205 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2208 check_new_page_bad(page);
2212 static inline bool free_pages_prezeroed(void)
2214 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2215 page_poisoning_enabled()) || want_init_on_free();
2218 #ifdef CONFIG_DEBUG_VM
2220 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2221 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2222 * also checked when pcp lists are refilled from the free lists.
2224 static inline bool check_pcp_refill(struct page *page)
2226 if (debug_pagealloc_enabled_static())
2227 return check_new_page(page);
2232 static inline bool check_new_pcp(struct page *page)
2234 return check_new_page(page);
2238 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2239 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2240 * enabled, they are also checked when being allocated from the pcp lists.
2242 static inline bool check_pcp_refill(struct page *page)
2244 return check_new_page(page);
2246 static inline bool check_new_pcp(struct page *page)
2248 if (debug_pagealloc_enabled_static())
2249 return check_new_page(page);
2253 #endif /* CONFIG_DEBUG_VM */
2255 static bool check_new_pages(struct page *page, unsigned int order)
2258 for (i = 0; i < (1 << order); i++) {
2259 struct page *p = page + i;
2261 if (unlikely(check_new_page(p)))
2268 inline void post_alloc_hook(struct page *page, unsigned int order,
2271 set_page_private(page, 0);
2272 set_page_refcounted(page);
2274 arch_alloc_page(page, order);
2275 if (debug_pagealloc_enabled_static())
2276 kernel_map_pages(page, 1 << order, 1);
2277 kasan_alloc_pages(page, order);
2278 kernel_poison_pages(page, 1 << order, 1);
2279 set_page_owner(page, order, gfp_flags);
2282 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2283 unsigned int alloc_flags)
2285 post_alloc_hook(page, order, gfp_flags);
2287 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2288 kernel_init_free_pages(page, 1 << order);
2290 if (order && (gfp_flags & __GFP_COMP))
2291 prep_compound_page(page, order);
2294 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2295 * allocate the page. The expectation is that the caller is taking
2296 * steps that will free more memory. The caller should avoid the page
2297 * being used for !PFMEMALLOC purposes.
2299 if (alloc_flags & ALLOC_NO_WATERMARKS)
2300 set_page_pfmemalloc(page);
2302 clear_page_pfmemalloc(page);
2306 * Go through the free lists for the given migratetype and remove
2307 * the smallest available page from the freelists
2309 static __always_inline
2310 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2313 unsigned int current_order;
2314 struct free_area *area;
2317 /* Find a page of the appropriate size in the preferred list */
2318 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2319 area = &(zone->free_area[current_order]);
2320 page = get_page_from_free_area(area, migratetype);
2323 del_page_from_free_list(page, zone, current_order);
2324 expand(zone, page, order, current_order, migratetype);
2325 set_pcppage_migratetype(page, migratetype);
2334 * This array describes the order lists are fallen back to when
2335 * the free lists for the desirable migrate type are depleted
2337 static int fallbacks[MIGRATE_TYPES][3] = {
2338 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2339 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2340 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2342 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2344 #ifdef CONFIG_MEMORY_ISOLATION
2345 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2350 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2353 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2356 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2357 unsigned int order) { return NULL; }
2361 * Move the free pages in a range to the freelist tail of the requested type.
2362 * Note that start_page and end_pages are not aligned on a pageblock
2363 * boundary. If alignment is required, use move_freepages_block()
2365 static int move_freepages(struct zone *zone,
2366 struct page *start_page, struct page *end_page,
2367 int migratetype, int *num_movable)
2371 int pages_moved = 0;
2373 for (page = start_page; page <= end_page;) {
2374 if (!pfn_valid_within(page_to_pfn(page))) {
2379 if (!PageBuddy(page)) {
2381 * We assume that pages that could be isolated for
2382 * migration are movable. But we don't actually try
2383 * isolating, as that would be expensive.
2386 (PageLRU(page) || __PageMovable(page)))
2393 /* Make sure we are not inadvertently changing nodes */
2394 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2395 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2397 order = buddy_order(page);
2398 move_to_free_list(page, zone, order, migratetype);
2400 pages_moved += 1 << order;
2406 int move_freepages_block(struct zone *zone, struct page *page,
2407 int migratetype, int *num_movable)
2409 unsigned long start_pfn, end_pfn;
2410 struct page *start_page, *end_page;
2415 start_pfn = page_to_pfn(page);
2416 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2417 start_page = pfn_to_page(start_pfn);
2418 end_page = start_page + pageblock_nr_pages - 1;
2419 end_pfn = start_pfn + pageblock_nr_pages - 1;
2421 /* Do not cross zone boundaries */
2422 if (!zone_spans_pfn(zone, start_pfn))
2424 if (!zone_spans_pfn(zone, end_pfn))
2427 return move_freepages(zone, start_page, end_page, migratetype,
2431 static void change_pageblock_range(struct page *pageblock_page,
2432 int start_order, int migratetype)
2434 int nr_pageblocks = 1 << (start_order - pageblock_order);
2436 while (nr_pageblocks--) {
2437 set_pageblock_migratetype(pageblock_page, migratetype);
2438 pageblock_page += pageblock_nr_pages;
2443 * When we are falling back to another migratetype during allocation, try to
2444 * steal extra free pages from the same pageblocks to satisfy further
2445 * allocations, instead of polluting multiple pageblocks.
2447 * If we are stealing a relatively large buddy page, it is likely there will
2448 * be more free pages in the pageblock, so try to steal them all. For
2449 * reclaimable and unmovable allocations, we steal regardless of page size,
2450 * as fragmentation caused by those allocations polluting movable pageblocks
2451 * is worse than movable allocations stealing from unmovable and reclaimable
2454 static bool can_steal_fallback(unsigned int order, int start_mt)
2457 * Leaving this order check is intended, although there is
2458 * relaxed order check in next check. The reason is that
2459 * we can actually steal whole pageblock if this condition met,
2460 * but, below check doesn't guarantee it and that is just heuristic
2461 * so could be changed anytime.
2463 if (order >= pageblock_order)
2466 if (order >= pageblock_order / 2 ||
2467 start_mt == MIGRATE_RECLAIMABLE ||
2468 start_mt == MIGRATE_UNMOVABLE ||
2469 page_group_by_mobility_disabled)
2475 static inline bool boost_watermark(struct zone *zone)
2477 unsigned long max_boost;
2479 if (!watermark_boost_factor)
2482 * Don't bother in zones that are unlikely to produce results.
2483 * On small machines, including kdump capture kernels running
2484 * in a small area, boosting the watermark can cause an out of
2485 * memory situation immediately.
2487 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2490 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2491 watermark_boost_factor, 10000);
2494 * high watermark may be uninitialised if fragmentation occurs
2495 * very early in boot so do not boost. We do not fall
2496 * through and boost by pageblock_nr_pages as failing
2497 * allocations that early means that reclaim is not going
2498 * to help and it may even be impossible to reclaim the
2499 * boosted watermark resulting in a hang.
2504 max_boost = max(pageblock_nr_pages, max_boost);
2506 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2513 * This function implements actual steal behaviour. If order is large enough,
2514 * we can steal whole pageblock. If not, we first move freepages in this
2515 * pageblock to our migratetype and determine how many already-allocated pages
2516 * are there in the pageblock with a compatible migratetype. If at least half
2517 * of pages are free or compatible, we can change migratetype of the pageblock
2518 * itself, so pages freed in the future will be put on the correct free list.
2520 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2521 unsigned int alloc_flags, int start_type, bool whole_block)
2523 unsigned int current_order = buddy_order(page);
2524 int free_pages, movable_pages, alike_pages;
2527 old_block_type = get_pageblock_migratetype(page);
2530 * This can happen due to races and we want to prevent broken
2531 * highatomic accounting.
2533 if (is_migrate_highatomic(old_block_type))
2536 /* Take ownership for orders >= pageblock_order */
2537 if (current_order >= pageblock_order) {
2538 change_pageblock_range(page, current_order, start_type);
2543 * Boost watermarks to increase reclaim pressure to reduce the
2544 * likelihood of future fallbacks. Wake kswapd now as the node
2545 * may be balanced overall and kswapd will not wake naturally.
2547 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2548 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2550 /* We are not allowed to try stealing from the whole block */
2554 free_pages = move_freepages_block(zone, page, start_type,
2557 * Determine how many pages are compatible with our allocation.
2558 * For movable allocation, it's the number of movable pages which
2559 * we just obtained. For other types it's a bit more tricky.
2561 if (start_type == MIGRATE_MOVABLE) {
2562 alike_pages = movable_pages;
2565 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2566 * to MOVABLE pageblock, consider all non-movable pages as
2567 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2568 * vice versa, be conservative since we can't distinguish the
2569 * exact migratetype of non-movable pages.
2571 if (old_block_type == MIGRATE_MOVABLE)
2572 alike_pages = pageblock_nr_pages
2573 - (free_pages + movable_pages);
2578 /* moving whole block can fail due to zone boundary conditions */
2583 * If a sufficient number of pages in the block are either free or of
2584 * comparable migratability as our allocation, claim the whole block.
2586 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2587 page_group_by_mobility_disabled)
2588 set_pageblock_migratetype(page, start_type);
2593 move_to_free_list(page, zone, current_order, start_type);
2597 * Check whether there is a suitable fallback freepage with requested order.
2598 * If only_stealable is true, this function returns fallback_mt only if
2599 * we can steal other freepages all together. This would help to reduce
2600 * fragmentation due to mixed migratetype pages in one pageblock.
2602 int find_suitable_fallback(struct free_area *area, unsigned int order,
2603 int migratetype, bool only_stealable, bool *can_steal)
2608 if (area->nr_free == 0)
2613 fallback_mt = fallbacks[migratetype][i];
2614 if (fallback_mt == MIGRATE_TYPES)
2617 if (free_area_empty(area, fallback_mt))
2620 if (can_steal_fallback(order, migratetype))
2623 if (!only_stealable)
2634 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2635 * there are no empty page blocks that contain a page with a suitable order
2637 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2638 unsigned int alloc_order)
2641 unsigned long max_managed, flags;
2644 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2645 * Check is race-prone but harmless.
2647 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2648 if (zone->nr_reserved_highatomic >= max_managed)
2651 spin_lock_irqsave(&zone->lock, flags);
2653 /* Recheck the nr_reserved_highatomic limit under the lock */
2654 if (zone->nr_reserved_highatomic >= max_managed)
2658 mt = get_pageblock_migratetype(page);
2659 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2660 && !is_migrate_cma(mt)) {
2661 zone->nr_reserved_highatomic += pageblock_nr_pages;
2662 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2663 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2667 spin_unlock_irqrestore(&zone->lock, flags);
2671 * Used when an allocation is about to fail under memory pressure. This
2672 * potentially hurts the reliability of high-order allocations when under
2673 * intense memory pressure but failed atomic allocations should be easier
2674 * to recover from than an OOM.
2676 * If @force is true, try to unreserve a pageblock even though highatomic
2677 * pageblock is exhausted.
2679 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2682 struct zonelist *zonelist = ac->zonelist;
2683 unsigned long flags;
2690 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2693 * Preserve at least one pageblock unless memory pressure
2696 if (!force && zone->nr_reserved_highatomic <=
2700 spin_lock_irqsave(&zone->lock, flags);
2701 for (order = 0; order < MAX_ORDER; order++) {
2702 struct free_area *area = &(zone->free_area[order]);
2704 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2709 * In page freeing path, migratetype change is racy so
2710 * we can counter several free pages in a pageblock
2711 * in this loop althoug we changed the pageblock type
2712 * from highatomic to ac->migratetype. So we should
2713 * adjust the count once.
2715 if (is_migrate_highatomic_page(page)) {
2717 * It should never happen but changes to
2718 * locking could inadvertently allow a per-cpu
2719 * drain to add pages to MIGRATE_HIGHATOMIC
2720 * while unreserving so be safe and watch for
2723 zone->nr_reserved_highatomic -= min(
2725 zone->nr_reserved_highatomic);
2729 * Convert to ac->migratetype and avoid the normal
2730 * pageblock stealing heuristics. Minimally, the caller
2731 * is doing the work and needs the pages. More
2732 * importantly, if the block was always converted to
2733 * MIGRATE_UNMOVABLE or another type then the number
2734 * of pageblocks that cannot be completely freed
2737 set_pageblock_migratetype(page, ac->migratetype);
2738 ret = move_freepages_block(zone, page, ac->migratetype,
2741 spin_unlock_irqrestore(&zone->lock, flags);
2745 spin_unlock_irqrestore(&zone->lock, flags);
2752 * Try finding a free buddy page on the fallback list and put it on the free
2753 * list of requested migratetype, possibly along with other pages from the same
2754 * block, depending on fragmentation avoidance heuristics. Returns true if
2755 * fallback was found so that __rmqueue_smallest() can grab it.
2757 * The use of signed ints for order and current_order is a deliberate
2758 * deviation from the rest of this file, to make the for loop
2759 * condition simpler.
2761 static __always_inline bool
2762 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2763 unsigned int alloc_flags)
2765 struct free_area *area;
2767 int min_order = order;
2773 * Do not steal pages from freelists belonging to other pageblocks
2774 * i.e. orders < pageblock_order. If there are no local zones free,
2775 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2777 if (alloc_flags & ALLOC_NOFRAGMENT)
2778 min_order = pageblock_order;
2781 * Find the largest available free page in the other list. This roughly
2782 * approximates finding the pageblock with the most free pages, which
2783 * would be too costly to do exactly.
2785 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2787 area = &(zone->free_area[current_order]);
2788 fallback_mt = find_suitable_fallback(area, current_order,
2789 start_migratetype, false, &can_steal);
2790 if (fallback_mt == -1)
2794 * We cannot steal all free pages from the pageblock and the
2795 * requested migratetype is movable. In that case it's better to
2796 * steal and split the smallest available page instead of the
2797 * largest available page, because even if the next movable
2798 * allocation falls back into a different pageblock than this
2799 * one, it won't cause permanent fragmentation.
2801 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2802 && current_order > order)
2811 for (current_order = order; current_order < MAX_ORDER;
2813 area = &(zone->free_area[current_order]);
2814 fallback_mt = find_suitable_fallback(area, current_order,
2815 start_migratetype, false, &can_steal);
2816 if (fallback_mt != -1)
2821 * This should not happen - we already found a suitable fallback
2822 * when looking for the largest page.
2824 VM_BUG_ON(current_order == MAX_ORDER);
2827 page = get_page_from_free_area(area, fallback_mt);
2829 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2832 trace_mm_page_alloc_extfrag(page, order, current_order,
2833 start_migratetype, fallback_mt);
2840 * Do the hard work of removing an element from the buddy allocator.
2841 * Call me with the zone->lock already held.
2843 static __always_inline struct page *
2844 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2845 unsigned int alloc_flags)
2849 if (IS_ENABLED(CONFIG_CMA)) {
2851 * Balance movable allocations between regular and CMA areas by
2852 * allocating from CMA when over half of the zone's free memory
2853 * is in the CMA area.
2855 if (alloc_flags & ALLOC_CMA &&
2856 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2857 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2858 page = __rmqueue_cma_fallback(zone, order);
2864 page = __rmqueue_smallest(zone, order, migratetype);
2865 if (unlikely(!page)) {
2866 if (alloc_flags & ALLOC_CMA)
2867 page = __rmqueue_cma_fallback(zone, order);
2869 if (!page && __rmqueue_fallback(zone, order, migratetype,
2875 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2880 * Obtain a specified number of elements from the buddy allocator, all under
2881 * a single hold of the lock, for efficiency. Add them to the supplied list.
2882 * Returns the number of new pages which were placed at *list.
2884 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2885 unsigned long count, struct list_head *list,
2886 int migratetype, unsigned int alloc_flags)
2890 spin_lock(&zone->lock);
2891 for (i = 0; i < count; ++i) {
2892 struct page *page = __rmqueue(zone, order, migratetype,
2894 if (unlikely(page == NULL))
2897 if (unlikely(check_pcp_refill(page)))
2901 * Split buddy pages returned by expand() are received here in
2902 * physical page order. The page is added to the tail of
2903 * caller's list. From the callers perspective, the linked list
2904 * is ordered by page number under some conditions. This is
2905 * useful for IO devices that can forward direction from the
2906 * head, thus also in the physical page order. This is useful
2907 * for IO devices that can merge IO requests if the physical
2908 * pages are ordered properly.
2910 list_add_tail(&page->lru, list);
2912 if (is_migrate_cma(get_pcppage_migratetype(page)))
2913 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2918 * i pages were removed from the buddy list even if some leak due
2919 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2920 * on i. Do not confuse with 'alloced' which is the number of
2921 * pages added to the pcp list.
2923 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2924 spin_unlock(&zone->lock);
2930 * Called from the vmstat counter updater to drain pagesets of this
2931 * currently executing processor on remote nodes after they have
2934 * Note that this function must be called with the thread pinned to
2935 * a single processor.
2937 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2939 unsigned long flags;
2940 int to_drain, batch;
2942 local_irq_save(flags);
2943 batch = READ_ONCE(pcp->batch);
2944 to_drain = min(pcp->count, batch);
2946 free_pcppages_bulk(zone, to_drain, pcp);
2947 local_irq_restore(flags);
2952 * Drain pcplists of the indicated processor and zone.
2954 * The processor must either be the current processor and the
2955 * thread pinned to the current processor or a processor that
2958 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2960 unsigned long flags;
2961 struct per_cpu_pageset *pset;
2962 struct per_cpu_pages *pcp;
2964 local_irq_save(flags);
2965 pset = per_cpu_ptr(zone->pageset, cpu);
2969 free_pcppages_bulk(zone, pcp->count, pcp);
2970 local_irq_restore(flags);
2974 * Drain pcplists of all zones on the indicated processor.
2976 * The processor must either be the current processor and the
2977 * thread pinned to the current processor or a processor that
2980 static void drain_pages(unsigned int cpu)
2984 for_each_populated_zone(zone) {
2985 drain_pages_zone(cpu, zone);
2990 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2992 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2993 * the single zone's pages.
2995 void drain_local_pages(struct zone *zone)
2997 int cpu = smp_processor_id();
3000 drain_pages_zone(cpu, zone);
3005 static void drain_local_pages_wq(struct work_struct *work)
3007 struct pcpu_drain *drain;
3009 drain = container_of(work, struct pcpu_drain, work);
3012 * drain_all_pages doesn't use proper cpu hotplug protection so
3013 * we can race with cpu offline when the WQ can move this from
3014 * a cpu pinned worker to an unbound one. We can operate on a different
3015 * cpu which is allright but we also have to make sure to not move to
3019 drain_local_pages(drain->zone);
3024 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3026 * When zone parameter is non-NULL, spill just the single zone's pages.
3028 * Note that this can be extremely slow as the draining happens in a workqueue.
3030 void drain_all_pages(struct zone *zone)
3035 * Allocate in the BSS so we wont require allocation in
3036 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3038 static cpumask_t cpus_with_pcps;
3041 * Make sure nobody triggers this path before mm_percpu_wq is fully
3044 if (WARN_ON_ONCE(!mm_percpu_wq))
3048 * Do not drain if one is already in progress unless it's specific to
3049 * a zone. Such callers are primarily CMA and memory hotplug and need
3050 * the drain to be complete when the call returns.
3052 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3055 mutex_lock(&pcpu_drain_mutex);
3059 * We don't care about racing with CPU hotplug event
3060 * as offline notification will cause the notified
3061 * cpu to drain that CPU pcps and on_each_cpu_mask
3062 * disables preemption as part of its processing
3064 for_each_online_cpu(cpu) {
3065 struct per_cpu_pageset *pcp;
3067 bool has_pcps = false;
3070 pcp = per_cpu_ptr(zone->pageset, cpu);
3074 for_each_populated_zone(z) {
3075 pcp = per_cpu_ptr(z->pageset, cpu);
3076 if (pcp->pcp.count) {
3084 cpumask_set_cpu(cpu, &cpus_with_pcps);
3086 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3089 for_each_cpu(cpu, &cpus_with_pcps) {
3090 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3093 INIT_WORK(&drain->work, drain_local_pages_wq);
3094 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3096 for_each_cpu(cpu, &cpus_with_pcps)
3097 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3099 mutex_unlock(&pcpu_drain_mutex);
3102 #ifdef CONFIG_HIBERNATION
3105 * Touch the watchdog for every WD_PAGE_COUNT pages.
3107 #define WD_PAGE_COUNT (128*1024)
3109 void mark_free_pages(struct zone *zone)
3111 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3112 unsigned long flags;
3113 unsigned int order, t;
3116 if (zone_is_empty(zone))
3119 spin_lock_irqsave(&zone->lock, flags);
3121 max_zone_pfn = zone_end_pfn(zone);
3122 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3123 if (pfn_valid(pfn)) {
3124 page = pfn_to_page(pfn);
3126 if (!--page_count) {
3127 touch_nmi_watchdog();
3128 page_count = WD_PAGE_COUNT;
3131 if (page_zone(page) != zone)
3134 if (!swsusp_page_is_forbidden(page))
3135 swsusp_unset_page_free(page);
3138 for_each_migratetype_order(order, t) {
3139 list_for_each_entry(page,
3140 &zone->free_area[order].free_list[t], lru) {
3143 pfn = page_to_pfn(page);
3144 for (i = 0; i < (1UL << order); i++) {
3145 if (!--page_count) {
3146 touch_nmi_watchdog();
3147 page_count = WD_PAGE_COUNT;
3149 swsusp_set_page_free(pfn_to_page(pfn + i));
3153 spin_unlock_irqrestore(&zone->lock, flags);
3155 #endif /* CONFIG_PM */
3157 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3161 if (!free_pcp_prepare(page))
3164 migratetype = get_pfnblock_migratetype(page, pfn);
3165 set_pcppage_migratetype(page, migratetype);
3169 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3171 struct zone *zone = page_zone(page);
3172 struct per_cpu_pages *pcp;
3175 migratetype = get_pcppage_migratetype(page);
3176 __count_vm_event(PGFREE);
3179 * We only track unmovable, reclaimable and movable on pcp lists.
3180 * Free ISOLATE pages back to the allocator because they are being
3181 * offlined but treat HIGHATOMIC as movable pages so we can get those
3182 * areas back if necessary. Otherwise, we may have to free
3183 * excessively into the page allocator
3185 if (migratetype >= MIGRATE_PCPTYPES) {
3186 if (unlikely(is_migrate_isolate(migratetype))) {
3187 free_one_page(zone, page, pfn, 0, migratetype,
3191 migratetype = MIGRATE_MOVABLE;
3194 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3195 list_add(&page->lru, &pcp->lists[migratetype]);
3197 if (pcp->count >= pcp->high) {
3198 unsigned long batch = READ_ONCE(pcp->batch);
3199 free_pcppages_bulk(zone, batch, pcp);
3204 * Free a 0-order page
3206 void free_unref_page(struct page *page)
3208 unsigned long flags;
3209 unsigned long pfn = page_to_pfn(page);
3211 if (!free_unref_page_prepare(page, pfn))
3214 local_irq_save(flags);
3215 free_unref_page_commit(page, pfn);
3216 local_irq_restore(flags);
3220 * Free a list of 0-order pages
3222 void free_unref_page_list(struct list_head *list)
3224 struct page *page, *next;
3225 unsigned long flags, pfn;
3226 int batch_count = 0;
3228 /* Prepare pages for freeing */
3229 list_for_each_entry_safe(page, next, list, lru) {
3230 pfn = page_to_pfn(page);
3231 if (!free_unref_page_prepare(page, pfn))
3232 list_del(&page->lru);
3233 set_page_private(page, pfn);
3236 local_irq_save(flags);
3237 list_for_each_entry_safe(page, next, list, lru) {
3238 unsigned long pfn = page_private(page);
3240 set_page_private(page, 0);
3241 trace_mm_page_free_batched(page);
3242 free_unref_page_commit(page, pfn);
3245 * Guard against excessive IRQ disabled times when we get
3246 * a large list of pages to free.
3248 if (++batch_count == SWAP_CLUSTER_MAX) {
3249 local_irq_restore(flags);
3251 local_irq_save(flags);
3254 local_irq_restore(flags);
3258 * split_page takes a non-compound higher-order page, and splits it into
3259 * n (1<<order) sub-pages: page[0..n]
3260 * Each sub-page must be freed individually.
3262 * Note: this is probably too low level an operation for use in drivers.
3263 * Please consult with lkml before using this in your driver.
3265 void split_page(struct page *page, unsigned int order)
3269 VM_BUG_ON_PAGE(PageCompound(page), page);
3270 VM_BUG_ON_PAGE(!page_count(page), page);
3272 for (i = 1; i < (1 << order); i++)
3273 set_page_refcounted(page + i);
3274 split_page_owner(page, 1 << order);
3276 EXPORT_SYMBOL_GPL(split_page);
3278 int __isolate_free_page(struct page *page, unsigned int order)
3280 unsigned long watermark;
3284 BUG_ON(!PageBuddy(page));
3286 zone = page_zone(page);
3287 mt = get_pageblock_migratetype(page);
3289 if (!is_migrate_isolate(mt)) {
3291 * Obey watermarks as if the page was being allocated. We can
3292 * emulate a high-order watermark check with a raised order-0
3293 * watermark, because we already know our high-order page
3296 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3297 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3300 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3303 /* Remove page from free list */
3305 del_page_from_free_list(page, zone, order);
3308 * Set the pageblock if the isolated page is at least half of a
3311 if (order >= pageblock_order - 1) {
3312 struct page *endpage = page + (1 << order) - 1;
3313 for (; page < endpage; page += pageblock_nr_pages) {
3314 int mt = get_pageblock_migratetype(page);
3315 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3316 && !is_migrate_highatomic(mt))
3317 set_pageblock_migratetype(page,
3323 return 1UL << order;
3327 * __putback_isolated_page - Return a now-isolated page back where we got it
3328 * @page: Page that was isolated
3329 * @order: Order of the isolated page
3330 * @mt: The page's pageblock's migratetype
3332 * This function is meant to return a page pulled from the free lists via
3333 * __isolate_free_page back to the free lists they were pulled from.
3335 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3337 struct zone *zone = page_zone(page);
3339 /* zone lock should be held when this function is called */
3340 lockdep_assert_held(&zone->lock);
3342 /* Return isolated page to tail of freelist. */
3343 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3344 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3348 * Update NUMA hit/miss statistics
3350 * Must be called with interrupts disabled.
3352 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3355 enum numa_stat_item local_stat = NUMA_LOCAL;
3357 /* skip numa counters update if numa stats is disabled */
3358 if (!static_branch_likely(&vm_numa_stat_key))
3361 if (zone_to_nid(z) != numa_node_id())
3362 local_stat = NUMA_OTHER;
3364 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3365 __inc_numa_state(z, NUMA_HIT);
3367 __inc_numa_state(z, NUMA_MISS);
3368 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3370 __inc_numa_state(z, local_stat);
3374 /* Remove page from the per-cpu list, caller must protect the list */
3375 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3376 unsigned int alloc_flags,
3377 struct per_cpu_pages *pcp,
3378 struct list_head *list)
3383 if (list_empty(list)) {
3384 pcp->count += rmqueue_bulk(zone, 0,
3386 migratetype, alloc_flags);
3387 if (unlikely(list_empty(list)))
3391 page = list_first_entry(list, struct page, lru);
3392 list_del(&page->lru);
3394 } while (check_new_pcp(page));
3399 /* Lock and remove page from the per-cpu list */
3400 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3401 struct zone *zone, gfp_t gfp_flags,
3402 int migratetype, unsigned int alloc_flags)
3404 struct per_cpu_pages *pcp;
3405 struct list_head *list;
3407 unsigned long flags;
3409 local_irq_save(flags);
3410 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3411 list = &pcp->lists[migratetype];
3412 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3414 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3415 zone_statistics(preferred_zone, zone);
3417 local_irq_restore(flags);
3422 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3425 struct page *rmqueue(struct zone *preferred_zone,
3426 struct zone *zone, unsigned int order,
3427 gfp_t gfp_flags, unsigned int alloc_flags,
3430 unsigned long flags;
3433 if (likely(order == 0)) {
3435 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3436 * we need to skip it when CMA area isn't allowed.
3438 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3439 migratetype != MIGRATE_MOVABLE) {
3440 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3441 migratetype, alloc_flags);
3447 * We most definitely don't want callers attempting to
3448 * allocate greater than order-1 page units with __GFP_NOFAIL.
3450 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3451 spin_lock_irqsave(&zone->lock, flags);
3456 * order-0 request can reach here when the pcplist is skipped
3457 * due to non-CMA allocation context. HIGHATOMIC area is
3458 * reserved for high-order atomic allocation, so order-0
3459 * request should skip it.
3461 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3462 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3464 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3467 page = __rmqueue(zone, order, migratetype, alloc_flags);
3468 } while (page && check_new_pages(page, order));
3469 spin_unlock(&zone->lock);
3472 __mod_zone_freepage_state(zone, -(1 << order),
3473 get_pcppage_migratetype(page));
3475 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3476 zone_statistics(preferred_zone, zone);
3477 local_irq_restore(flags);
3480 /* Separate test+clear to avoid unnecessary atomics */
3481 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3482 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3483 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3486 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3490 local_irq_restore(flags);
3494 #ifdef CONFIG_FAIL_PAGE_ALLOC
3497 struct fault_attr attr;
3499 bool ignore_gfp_highmem;
3500 bool ignore_gfp_reclaim;
3502 } fail_page_alloc = {
3503 .attr = FAULT_ATTR_INITIALIZER,
3504 .ignore_gfp_reclaim = true,
3505 .ignore_gfp_highmem = true,
3509 static int __init setup_fail_page_alloc(char *str)
3511 return setup_fault_attr(&fail_page_alloc.attr, str);
3513 __setup("fail_page_alloc=", setup_fail_page_alloc);
3515 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3517 if (order < fail_page_alloc.min_order)
3519 if (gfp_mask & __GFP_NOFAIL)
3521 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3523 if (fail_page_alloc.ignore_gfp_reclaim &&
3524 (gfp_mask & __GFP_DIRECT_RECLAIM))
3527 return should_fail(&fail_page_alloc.attr, 1 << order);
3530 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3532 static int __init fail_page_alloc_debugfs(void)
3534 umode_t mode = S_IFREG | 0600;
3537 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3538 &fail_page_alloc.attr);
3540 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3541 &fail_page_alloc.ignore_gfp_reclaim);
3542 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3543 &fail_page_alloc.ignore_gfp_highmem);
3544 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3549 late_initcall(fail_page_alloc_debugfs);
3551 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3553 #else /* CONFIG_FAIL_PAGE_ALLOC */
3555 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3560 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3562 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3564 return __should_fail_alloc_page(gfp_mask, order);
3566 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3568 static inline long __zone_watermark_unusable_free(struct zone *z,
3569 unsigned int order, unsigned int alloc_flags)
3571 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3572 long unusable_free = (1 << order) - 1;
3575 * If the caller does not have rights to ALLOC_HARDER then subtract
3576 * the high-atomic reserves. This will over-estimate the size of the
3577 * atomic reserve but it avoids a search.
3579 if (likely(!alloc_harder))
3580 unusable_free += z->nr_reserved_highatomic;
3583 /* If allocation can't use CMA areas don't use free CMA pages */
3584 if (!(alloc_flags & ALLOC_CMA))
3585 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3588 return unusable_free;
3592 * Return true if free base pages are above 'mark'. For high-order checks it
3593 * will return true of the order-0 watermark is reached and there is at least
3594 * one free page of a suitable size. Checking now avoids taking the zone lock
3595 * to check in the allocation paths if no pages are free.
3597 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3598 int highest_zoneidx, unsigned int alloc_flags,
3603 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3605 /* free_pages may go negative - that's OK */
3606 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3608 if (alloc_flags & ALLOC_HIGH)
3611 if (unlikely(alloc_harder)) {
3613 * OOM victims can try even harder than normal ALLOC_HARDER
3614 * users on the grounds that it's definitely going to be in
3615 * the exit path shortly and free memory. Any allocation it
3616 * makes during the free path will be small and short-lived.
3618 if (alloc_flags & ALLOC_OOM)
3625 * Check watermarks for an order-0 allocation request. If these
3626 * are not met, then a high-order request also cannot go ahead
3627 * even if a suitable page happened to be free.
3629 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3632 /* If this is an order-0 request then the watermark is fine */
3636 /* For a high-order request, check at least one suitable page is free */
3637 for (o = order; o < MAX_ORDER; o++) {
3638 struct free_area *area = &z->free_area[o];
3644 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3645 if (!free_area_empty(area, mt))
3650 if ((alloc_flags & ALLOC_CMA) &&
3651 !free_area_empty(area, MIGRATE_CMA)) {
3655 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3661 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3662 int highest_zoneidx, unsigned int alloc_flags)
3664 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3665 zone_page_state(z, NR_FREE_PAGES));
3668 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3669 unsigned long mark, int highest_zoneidx,
3670 unsigned int alloc_flags, gfp_t gfp_mask)
3674 free_pages = zone_page_state(z, NR_FREE_PAGES);
3677 * Fast check for order-0 only. If this fails then the reserves
3678 * need to be calculated.
3683 fast_free = free_pages;
3684 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3685 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3689 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3693 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3694 * when checking the min watermark. The min watermark is the
3695 * point where boosting is ignored so that kswapd is woken up
3696 * when below the low watermark.
3698 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3699 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3700 mark = z->_watermark[WMARK_MIN];
3701 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3702 alloc_flags, free_pages);
3708 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3709 unsigned long mark, int highest_zoneidx)
3711 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3713 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3714 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3716 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3721 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3723 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3724 node_reclaim_distance;
3726 #else /* CONFIG_NUMA */
3727 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3731 #endif /* CONFIG_NUMA */
3734 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3735 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3736 * premature use of a lower zone may cause lowmem pressure problems that
3737 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3738 * probably too small. It only makes sense to spread allocations to avoid
3739 * fragmentation between the Normal and DMA32 zones.
3741 static inline unsigned int
3742 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3744 unsigned int alloc_flags;
3747 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3750 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3752 #ifdef CONFIG_ZONE_DMA32
3756 if (zone_idx(zone) != ZONE_NORMAL)
3760 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3761 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3762 * on UMA that if Normal is populated then so is DMA32.
3764 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3765 if (nr_online_nodes > 1 && !populated_zone(--zone))
3768 alloc_flags |= ALLOC_NOFRAGMENT;
3769 #endif /* CONFIG_ZONE_DMA32 */
3773 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3774 unsigned int alloc_flags)
3777 unsigned int pflags = current->flags;
3779 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3780 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3781 alloc_flags |= ALLOC_CMA;
3788 * get_page_from_freelist goes through the zonelist trying to allocate
3791 static struct page *
3792 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3793 const struct alloc_context *ac)
3797 struct pglist_data *last_pgdat_dirty_limit = NULL;
3802 * Scan zonelist, looking for a zone with enough free.
3803 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3805 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3806 z = ac->preferred_zoneref;
3807 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3812 if (cpusets_enabled() &&
3813 (alloc_flags & ALLOC_CPUSET) &&
3814 !__cpuset_zone_allowed(zone, gfp_mask))
3817 * When allocating a page cache page for writing, we
3818 * want to get it from a node that is within its dirty
3819 * limit, such that no single node holds more than its
3820 * proportional share of globally allowed dirty pages.
3821 * The dirty limits take into account the node's
3822 * lowmem reserves and high watermark so that kswapd
3823 * should be able to balance it without having to
3824 * write pages from its LRU list.
3826 * XXX: For now, allow allocations to potentially
3827 * exceed the per-node dirty limit in the slowpath
3828 * (spread_dirty_pages unset) before going into reclaim,
3829 * which is important when on a NUMA setup the allowed
3830 * nodes are together not big enough to reach the
3831 * global limit. The proper fix for these situations
3832 * will require awareness of nodes in the
3833 * dirty-throttling and the flusher threads.
3835 if (ac->spread_dirty_pages) {
3836 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3839 if (!node_dirty_ok(zone->zone_pgdat)) {
3840 last_pgdat_dirty_limit = zone->zone_pgdat;
3845 if (no_fallback && nr_online_nodes > 1 &&
3846 zone != ac->preferred_zoneref->zone) {
3850 * If moving to a remote node, retry but allow
3851 * fragmenting fallbacks. Locality is more important
3852 * than fragmentation avoidance.
3854 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3855 if (zone_to_nid(zone) != local_nid) {
3856 alloc_flags &= ~ALLOC_NOFRAGMENT;
3861 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3862 if (!zone_watermark_fast(zone, order, mark,
3863 ac->highest_zoneidx, alloc_flags,
3867 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3869 * Watermark failed for this zone, but see if we can
3870 * grow this zone if it contains deferred pages.
3872 if (static_branch_unlikely(&deferred_pages)) {
3873 if (_deferred_grow_zone(zone, order))
3877 /* Checked here to keep the fast path fast */
3878 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3879 if (alloc_flags & ALLOC_NO_WATERMARKS)
3882 if (node_reclaim_mode == 0 ||
3883 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3886 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3888 case NODE_RECLAIM_NOSCAN:
3891 case NODE_RECLAIM_FULL:
3892 /* scanned but unreclaimable */
3895 /* did we reclaim enough */
3896 if (zone_watermark_ok(zone, order, mark,
3897 ac->highest_zoneidx, alloc_flags))
3905 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3906 gfp_mask, alloc_flags, ac->migratetype);
3908 prep_new_page(page, order, gfp_mask, alloc_flags);
3911 * If this is a high-order atomic allocation then check
3912 * if the pageblock should be reserved for the future
3914 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3915 reserve_highatomic_pageblock(page, zone, order);
3919 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3920 /* Try again if zone has deferred pages */
3921 if (static_branch_unlikely(&deferred_pages)) {
3922 if (_deferred_grow_zone(zone, order))
3930 * It's possible on a UMA machine to get through all zones that are
3931 * fragmented. If avoiding fragmentation, reset and try again.
3934 alloc_flags &= ~ALLOC_NOFRAGMENT;
3941 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3943 unsigned int filter = SHOW_MEM_FILTER_NODES;
3946 * This documents exceptions given to allocations in certain
3947 * contexts that are allowed to allocate outside current's set
3950 if (!(gfp_mask & __GFP_NOMEMALLOC))
3951 if (tsk_is_oom_victim(current) ||
3952 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3953 filter &= ~SHOW_MEM_FILTER_NODES;
3954 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3955 filter &= ~SHOW_MEM_FILTER_NODES;
3957 show_mem(filter, nodemask);
3960 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3962 struct va_format vaf;
3964 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3966 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3969 va_start(args, fmt);
3972 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3973 current->comm, &vaf, gfp_mask, &gfp_mask,
3974 nodemask_pr_args(nodemask));
3977 cpuset_print_current_mems_allowed();
3980 warn_alloc_show_mem(gfp_mask, nodemask);
3983 static inline struct page *
3984 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3985 unsigned int alloc_flags,
3986 const struct alloc_context *ac)
3990 page = get_page_from_freelist(gfp_mask, order,
3991 alloc_flags|ALLOC_CPUSET, ac);
3993 * fallback to ignore cpuset restriction if our nodes
3997 page = get_page_from_freelist(gfp_mask, order,
4003 static inline struct page *
4004 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4005 const struct alloc_context *ac, unsigned long *did_some_progress)
4007 struct oom_control oc = {
4008 .zonelist = ac->zonelist,
4009 .nodemask = ac->nodemask,
4011 .gfp_mask = gfp_mask,
4016 *did_some_progress = 0;
4019 * Acquire the oom lock. If that fails, somebody else is
4020 * making progress for us.
4022 if (!mutex_trylock(&oom_lock)) {
4023 *did_some_progress = 1;
4024 schedule_timeout_uninterruptible(1);
4029 * Go through the zonelist yet one more time, keep very high watermark
4030 * here, this is only to catch a parallel oom killing, we must fail if
4031 * we're still under heavy pressure. But make sure that this reclaim
4032 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4033 * allocation which will never fail due to oom_lock already held.
4035 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4036 ~__GFP_DIRECT_RECLAIM, order,
4037 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4041 /* Coredumps can quickly deplete all memory reserves */
4042 if (current->flags & PF_DUMPCORE)
4044 /* The OOM killer will not help higher order allocs */
4045 if (order > PAGE_ALLOC_COSTLY_ORDER)
4048 * We have already exhausted all our reclaim opportunities without any
4049 * success so it is time to admit defeat. We will skip the OOM killer
4050 * because it is very likely that the caller has a more reasonable
4051 * fallback than shooting a random task.
4053 * The OOM killer may not free memory on a specific node.
4055 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4057 /* The OOM killer does not needlessly kill tasks for lowmem */
4058 if (ac->highest_zoneidx < ZONE_NORMAL)
4060 if (pm_suspended_storage())
4063 * XXX: GFP_NOFS allocations should rather fail than rely on
4064 * other request to make a forward progress.
4065 * We are in an unfortunate situation where out_of_memory cannot
4066 * do much for this context but let's try it to at least get
4067 * access to memory reserved if the current task is killed (see
4068 * out_of_memory). Once filesystems are ready to handle allocation
4069 * failures more gracefully we should just bail out here.
4072 /* Exhausted what can be done so it's blame time */
4073 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4074 *did_some_progress = 1;
4077 * Help non-failing allocations by giving them access to memory
4080 if (gfp_mask & __GFP_NOFAIL)
4081 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4082 ALLOC_NO_WATERMARKS, ac);
4085 mutex_unlock(&oom_lock);
4090 * Maximum number of compaction retries wit a progress before OOM
4091 * killer is consider as the only way to move forward.
4093 #define MAX_COMPACT_RETRIES 16
4095 #ifdef CONFIG_COMPACTION
4096 /* Try memory compaction for high-order allocations before reclaim */
4097 static struct page *
4098 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4099 unsigned int alloc_flags, const struct alloc_context *ac,
4100 enum compact_priority prio, enum compact_result *compact_result)
4102 struct page *page = NULL;
4103 unsigned long pflags;
4104 unsigned int noreclaim_flag;
4109 psi_memstall_enter(&pflags);
4110 noreclaim_flag = memalloc_noreclaim_save();
4112 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4115 memalloc_noreclaim_restore(noreclaim_flag);
4116 psi_memstall_leave(&pflags);
4119 * At least in one zone compaction wasn't deferred or skipped, so let's
4120 * count a compaction stall
4122 count_vm_event(COMPACTSTALL);
4124 /* Prep a captured page if available */
4126 prep_new_page(page, order, gfp_mask, alloc_flags);
4128 /* Try get a page from the freelist if available */
4130 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4133 struct zone *zone = page_zone(page);
4135 zone->compact_blockskip_flush = false;
4136 compaction_defer_reset(zone, order, true);
4137 count_vm_event(COMPACTSUCCESS);
4142 * It's bad if compaction run occurs and fails. The most likely reason
4143 * is that pages exist, but not enough to satisfy watermarks.
4145 count_vm_event(COMPACTFAIL);
4153 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4154 enum compact_result compact_result,
4155 enum compact_priority *compact_priority,
4156 int *compaction_retries)
4158 int max_retries = MAX_COMPACT_RETRIES;
4161 int retries = *compaction_retries;
4162 enum compact_priority priority = *compact_priority;
4167 if (compaction_made_progress(compact_result))
4168 (*compaction_retries)++;
4171 * compaction considers all the zone as desperately out of memory
4172 * so it doesn't really make much sense to retry except when the
4173 * failure could be caused by insufficient priority
4175 if (compaction_failed(compact_result))
4176 goto check_priority;
4179 * compaction was skipped because there are not enough order-0 pages
4180 * to work with, so we retry only if it looks like reclaim can help.
4182 if (compaction_needs_reclaim(compact_result)) {
4183 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4188 * make sure the compaction wasn't deferred or didn't bail out early
4189 * due to locks contention before we declare that we should give up.
4190 * But the next retry should use a higher priority if allowed, so
4191 * we don't just keep bailing out endlessly.
4193 if (compaction_withdrawn(compact_result)) {
4194 goto check_priority;
4198 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4199 * costly ones because they are de facto nofail and invoke OOM
4200 * killer to move on while costly can fail and users are ready
4201 * to cope with that. 1/4 retries is rather arbitrary but we
4202 * would need much more detailed feedback from compaction to
4203 * make a better decision.
4205 if (order > PAGE_ALLOC_COSTLY_ORDER)
4207 if (*compaction_retries <= max_retries) {
4213 * Make sure there are attempts at the highest priority if we exhausted
4214 * all retries or failed at the lower priorities.
4217 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4218 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4220 if (*compact_priority > min_priority) {
4221 (*compact_priority)--;
4222 *compaction_retries = 0;
4226 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4230 static inline struct page *
4231 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4232 unsigned int alloc_flags, const struct alloc_context *ac,
4233 enum compact_priority prio, enum compact_result *compact_result)
4235 *compact_result = COMPACT_SKIPPED;
4240 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4241 enum compact_result compact_result,
4242 enum compact_priority *compact_priority,
4243 int *compaction_retries)
4248 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4252 * There are setups with compaction disabled which would prefer to loop
4253 * inside the allocator rather than hit the oom killer prematurely.
4254 * Let's give them a good hope and keep retrying while the order-0
4255 * watermarks are OK.
4257 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4258 ac->highest_zoneidx, ac->nodemask) {
4259 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4260 ac->highest_zoneidx, alloc_flags))
4265 #endif /* CONFIG_COMPACTION */
4267 #ifdef CONFIG_LOCKDEP
4268 static struct lockdep_map __fs_reclaim_map =
4269 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4271 static bool __need_fs_reclaim(gfp_t gfp_mask)
4273 gfp_mask = current_gfp_context(gfp_mask);
4275 /* no reclaim without waiting on it */
4276 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4279 /* this guy won't enter reclaim */
4280 if (current->flags & PF_MEMALLOC)
4283 /* We're only interested __GFP_FS allocations for now */
4284 if (!(gfp_mask & __GFP_FS))
4287 if (gfp_mask & __GFP_NOLOCKDEP)
4293 void __fs_reclaim_acquire(void)
4295 lock_map_acquire(&__fs_reclaim_map);
4298 void __fs_reclaim_release(void)
4300 lock_map_release(&__fs_reclaim_map);
4303 void fs_reclaim_acquire(gfp_t gfp_mask)
4305 if (__need_fs_reclaim(gfp_mask))
4306 __fs_reclaim_acquire();
4308 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4310 void fs_reclaim_release(gfp_t gfp_mask)
4312 if (__need_fs_reclaim(gfp_mask))
4313 __fs_reclaim_release();
4315 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4318 /* Perform direct synchronous page reclaim */
4319 static unsigned long
4320 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4321 const struct alloc_context *ac)
4323 unsigned int noreclaim_flag;
4324 unsigned long pflags, progress;
4328 /* We now go into synchronous reclaim */
4329 cpuset_memory_pressure_bump();
4330 psi_memstall_enter(&pflags);
4331 fs_reclaim_acquire(gfp_mask);
4332 noreclaim_flag = memalloc_noreclaim_save();
4334 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4337 memalloc_noreclaim_restore(noreclaim_flag);
4338 fs_reclaim_release(gfp_mask);
4339 psi_memstall_leave(&pflags);
4346 /* The really slow allocator path where we enter direct reclaim */
4347 static inline struct page *
4348 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4349 unsigned int alloc_flags, const struct alloc_context *ac,
4350 unsigned long *did_some_progress)
4352 struct page *page = NULL;
4353 bool drained = false;
4355 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4356 if (unlikely(!(*did_some_progress)))
4360 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4363 * If an allocation failed after direct reclaim, it could be because
4364 * pages are pinned on the per-cpu lists or in high alloc reserves.
4365 * Shrink them and try again
4367 if (!page && !drained) {
4368 unreserve_highatomic_pageblock(ac, false);
4369 drain_all_pages(NULL);
4377 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4378 const struct alloc_context *ac)
4382 pg_data_t *last_pgdat = NULL;
4383 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4385 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4387 if (last_pgdat != zone->zone_pgdat)
4388 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4389 last_pgdat = zone->zone_pgdat;
4393 static inline unsigned int
4394 gfp_to_alloc_flags(gfp_t gfp_mask)
4396 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4399 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4400 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4401 * to save two branches.
4403 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4404 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4407 * The caller may dip into page reserves a bit more if the caller
4408 * cannot run direct reclaim, or if the caller has realtime scheduling
4409 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4410 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4412 alloc_flags |= (__force int)
4413 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4415 if (gfp_mask & __GFP_ATOMIC) {
4417 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4418 * if it can't schedule.
4420 if (!(gfp_mask & __GFP_NOMEMALLOC))
4421 alloc_flags |= ALLOC_HARDER;
4423 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4424 * comment for __cpuset_node_allowed().
4426 alloc_flags &= ~ALLOC_CPUSET;
4427 } else if (unlikely(rt_task(current)) && !in_interrupt())
4428 alloc_flags |= ALLOC_HARDER;
4430 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4435 static bool oom_reserves_allowed(struct task_struct *tsk)
4437 if (!tsk_is_oom_victim(tsk))
4441 * !MMU doesn't have oom reaper so give access to memory reserves
4442 * only to the thread with TIF_MEMDIE set
4444 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4451 * Distinguish requests which really need access to full memory
4452 * reserves from oom victims which can live with a portion of it
4454 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4456 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4458 if (gfp_mask & __GFP_MEMALLOC)
4459 return ALLOC_NO_WATERMARKS;
4460 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4461 return ALLOC_NO_WATERMARKS;
4462 if (!in_interrupt()) {
4463 if (current->flags & PF_MEMALLOC)
4464 return ALLOC_NO_WATERMARKS;
4465 else if (oom_reserves_allowed(current))
4472 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4474 return !!__gfp_pfmemalloc_flags(gfp_mask);
4478 * Checks whether it makes sense to retry the reclaim to make a forward progress
4479 * for the given allocation request.
4481 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4482 * without success, or when we couldn't even meet the watermark if we
4483 * reclaimed all remaining pages on the LRU lists.
4485 * Returns true if a retry is viable or false to enter the oom path.
4488 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4489 struct alloc_context *ac, int alloc_flags,
4490 bool did_some_progress, int *no_progress_loops)
4497 * Costly allocations might have made a progress but this doesn't mean
4498 * their order will become available due to high fragmentation so
4499 * always increment the no progress counter for them
4501 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4502 *no_progress_loops = 0;
4504 (*no_progress_loops)++;
4507 * Make sure we converge to OOM if we cannot make any progress
4508 * several times in the row.
4510 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4511 /* Before OOM, exhaust highatomic_reserve */
4512 return unreserve_highatomic_pageblock(ac, true);
4516 * Keep reclaiming pages while there is a chance this will lead
4517 * somewhere. If none of the target zones can satisfy our allocation
4518 * request even if all reclaimable pages are considered then we are
4519 * screwed and have to go OOM.
4521 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4522 ac->highest_zoneidx, ac->nodemask) {
4523 unsigned long available;
4524 unsigned long reclaimable;
4525 unsigned long min_wmark = min_wmark_pages(zone);
4528 available = reclaimable = zone_reclaimable_pages(zone);
4529 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4532 * Would the allocation succeed if we reclaimed all
4533 * reclaimable pages?
4535 wmark = __zone_watermark_ok(zone, order, min_wmark,
4536 ac->highest_zoneidx, alloc_flags, available);
4537 trace_reclaim_retry_zone(z, order, reclaimable,
4538 available, min_wmark, *no_progress_loops, wmark);
4541 * If we didn't make any progress and have a lot of
4542 * dirty + writeback pages then we should wait for
4543 * an IO to complete to slow down the reclaim and
4544 * prevent from pre mature OOM
4546 if (!did_some_progress) {
4547 unsigned long write_pending;
4549 write_pending = zone_page_state_snapshot(zone,
4550 NR_ZONE_WRITE_PENDING);
4552 if (2 * write_pending > reclaimable) {
4553 congestion_wait(BLK_RW_ASYNC, HZ/10);
4565 * Memory allocation/reclaim might be called from a WQ context and the
4566 * current implementation of the WQ concurrency control doesn't
4567 * recognize that a particular WQ is congested if the worker thread is
4568 * looping without ever sleeping. Therefore we have to do a short sleep
4569 * here rather than calling cond_resched().
4571 if (current->flags & PF_WQ_WORKER)
4572 schedule_timeout_uninterruptible(1);
4579 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4582 * It's possible that cpuset's mems_allowed and the nodemask from
4583 * mempolicy don't intersect. This should be normally dealt with by
4584 * policy_nodemask(), but it's possible to race with cpuset update in
4585 * such a way the check therein was true, and then it became false
4586 * before we got our cpuset_mems_cookie here.
4587 * This assumes that for all allocations, ac->nodemask can come only
4588 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4589 * when it does not intersect with the cpuset restrictions) or the
4590 * caller can deal with a violated nodemask.
4592 if (cpusets_enabled() && ac->nodemask &&
4593 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4594 ac->nodemask = NULL;
4599 * When updating a task's mems_allowed or mempolicy nodemask, it is
4600 * possible to race with parallel threads in such a way that our
4601 * allocation can fail while the mask is being updated. If we are about
4602 * to fail, check if the cpuset changed during allocation and if so,
4605 if (read_mems_allowed_retry(cpuset_mems_cookie))
4611 static inline struct page *
4612 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4613 struct alloc_context *ac)
4615 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4616 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4617 struct page *page = NULL;
4618 unsigned int alloc_flags;
4619 unsigned long did_some_progress;
4620 enum compact_priority compact_priority;
4621 enum compact_result compact_result;
4622 int compaction_retries;
4623 int no_progress_loops;
4624 unsigned int cpuset_mems_cookie;
4628 * We also sanity check to catch abuse of atomic reserves being used by
4629 * callers that are not in atomic context.
4631 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4632 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4633 gfp_mask &= ~__GFP_ATOMIC;
4636 compaction_retries = 0;
4637 no_progress_loops = 0;
4638 compact_priority = DEF_COMPACT_PRIORITY;
4639 cpuset_mems_cookie = read_mems_allowed_begin();
4642 * The fast path uses conservative alloc_flags to succeed only until
4643 * kswapd needs to be woken up, and to avoid the cost of setting up
4644 * alloc_flags precisely. So we do that now.
4646 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4649 * We need to recalculate the starting point for the zonelist iterator
4650 * because we might have used different nodemask in the fast path, or
4651 * there was a cpuset modification and we are retrying - otherwise we
4652 * could end up iterating over non-eligible zones endlessly.
4654 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4655 ac->highest_zoneidx, ac->nodemask);
4656 if (!ac->preferred_zoneref->zone)
4659 if (alloc_flags & ALLOC_KSWAPD)
4660 wake_all_kswapds(order, gfp_mask, ac);
4663 * The adjusted alloc_flags might result in immediate success, so try
4666 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4671 * For costly allocations, try direct compaction first, as it's likely
4672 * that we have enough base pages and don't need to reclaim. For non-
4673 * movable high-order allocations, do that as well, as compaction will
4674 * try prevent permanent fragmentation by migrating from blocks of the
4676 * Don't try this for allocations that are allowed to ignore
4677 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4679 if (can_direct_reclaim &&
4681 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4682 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4683 page = __alloc_pages_direct_compact(gfp_mask, order,
4685 INIT_COMPACT_PRIORITY,
4691 * Checks for costly allocations with __GFP_NORETRY, which
4692 * includes some THP page fault allocations
4694 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4696 * If allocating entire pageblock(s) and compaction
4697 * failed because all zones are below low watermarks
4698 * or is prohibited because it recently failed at this
4699 * order, fail immediately unless the allocator has
4700 * requested compaction and reclaim retry.
4703 * - potentially very expensive because zones are far
4704 * below their low watermarks or this is part of very
4705 * bursty high order allocations,
4706 * - not guaranteed to help because isolate_freepages()
4707 * may not iterate over freed pages as part of its
4709 * - unlikely to make entire pageblocks free on its
4712 if (compact_result == COMPACT_SKIPPED ||
4713 compact_result == COMPACT_DEFERRED)
4717 * Looks like reclaim/compaction is worth trying, but
4718 * sync compaction could be very expensive, so keep
4719 * using async compaction.
4721 compact_priority = INIT_COMPACT_PRIORITY;
4726 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4727 if (alloc_flags & ALLOC_KSWAPD)
4728 wake_all_kswapds(order, gfp_mask, ac);
4730 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4732 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4735 * Reset the nodemask and zonelist iterators if memory policies can be
4736 * ignored. These allocations are high priority and system rather than
4739 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4740 ac->nodemask = NULL;
4741 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4742 ac->highest_zoneidx, ac->nodemask);
4745 /* Attempt with potentially adjusted zonelist and alloc_flags */
4746 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4750 /* Caller is not willing to reclaim, we can't balance anything */
4751 if (!can_direct_reclaim)
4754 /* Avoid recursion of direct reclaim */
4755 if (current->flags & PF_MEMALLOC)
4758 /* Try direct reclaim and then allocating */
4759 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4760 &did_some_progress);
4764 /* Try direct compaction and then allocating */
4765 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4766 compact_priority, &compact_result);
4770 /* Do not loop if specifically requested */
4771 if (gfp_mask & __GFP_NORETRY)
4775 * Do not retry costly high order allocations unless they are
4776 * __GFP_RETRY_MAYFAIL
4778 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4781 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4782 did_some_progress > 0, &no_progress_loops))
4786 * It doesn't make any sense to retry for the compaction if the order-0
4787 * reclaim is not able to make any progress because the current
4788 * implementation of the compaction depends on the sufficient amount
4789 * of free memory (see __compaction_suitable)
4791 if (did_some_progress > 0 &&
4792 should_compact_retry(ac, order, alloc_flags,
4793 compact_result, &compact_priority,
4794 &compaction_retries))
4798 /* Deal with possible cpuset update races before we start OOM killing */
4799 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4802 /* Reclaim has failed us, start killing things */
4803 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4807 /* Avoid allocations with no watermarks from looping endlessly */
4808 if (tsk_is_oom_victim(current) &&
4809 (alloc_flags & ALLOC_OOM ||
4810 (gfp_mask & __GFP_NOMEMALLOC)))
4813 /* Retry as long as the OOM killer is making progress */
4814 if (did_some_progress) {
4815 no_progress_loops = 0;
4820 /* Deal with possible cpuset update races before we fail */
4821 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4825 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4828 if (gfp_mask & __GFP_NOFAIL) {
4830 * All existing users of the __GFP_NOFAIL are blockable, so warn
4831 * of any new users that actually require GFP_NOWAIT
4833 if (WARN_ON_ONCE(!can_direct_reclaim))
4837 * PF_MEMALLOC request from this context is rather bizarre
4838 * because we cannot reclaim anything and only can loop waiting
4839 * for somebody to do a work for us
4841 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4844 * non failing costly orders are a hard requirement which we
4845 * are not prepared for much so let's warn about these users
4846 * so that we can identify them and convert them to something
4849 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4852 * Help non-failing allocations by giving them access to memory
4853 * reserves but do not use ALLOC_NO_WATERMARKS because this
4854 * could deplete whole memory reserves which would just make
4855 * the situation worse
4857 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4865 warn_alloc(gfp_mask, ac->nodemask,
4866 "page allocation failure: order:%u", order);
4871 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4872 int preferred_nid, nodemask_t *nodemask,
4873 struct alloc_context *ac, gfp_t *alloc_mask,
4874 unsigned int *alloc_flags)
4876 ac->highest_zoneidx = gfp_zone(gfp_mask);
4877 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4878 ac->nodemask = nodemask;
4879 ac->migratetype = gfp_migratetype(gfp_mask);
4881 if (cpusets_enabled()) {
4882 *alloc_mask |= __GFP_HARDWALL;
4884 * When we are in the interrupt context, it is irrelevant
4885 * to the current task context. It means that any node ok.
4887 if (!in_interrupt() && !ac->nodemask)
4888 ac->nodemask = &cpuset_current_mems_allowed;
4890 *alloc_flags |= ALLOC_CPUSET;
4893 fs_reclaim_acquire(gfp_mask);
4894 fs_reclaim_release(gfp_mask);
4896 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4898 if (should_fail_alloc_page(gfp_mask, order))
4901 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4903 /* Dirty zone balancing only done in the fast path */
4904 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4907 * The preferred zone is used for statistics but crucially it is
4908 * also used as the starting point for the zonelist iterator. It
4909 * may get reset for allocations that ignore memory policies.
4911 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4912 ac->highest_zoneidx, ac->nodemask);
4918 * This is the 'heart' of the zoned buddy allocator.
4921 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4922 nodemask_t *nodemask)
4925 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4926 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4927 struct alloc_context ac = { };
4930 * There are several places where we assume that the order value is sane
4931 * so bail out early if the request is out of bound.
4933 if (unlikely(order >= MAX_ORDER)) {
4934 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4938 gfp_mask &= gfp_allowed_mask;
4939 alloc_mask = gfp_mask;
4940 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4944 * Forbid the first pass from falling back to types that fragment
4945 * memory until all local zones are considered.
4947 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4949 /* First allocation attempt */
4950 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4955 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4956 * resp. GFP_NOIO which has to be inherited for all allocation requests
4957 * from a particular context which has been marked by
4958 * memalloc_no{fs,io}_{save,restore}.
4960 alloc_mask = current_gfp_context(gfp_mask);
4961 ac.spread_dirty_pages = false;
4964 * Restore the original nodemask if it was potentially replaced with
4965 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4967 ac.nodemask = nodemask;
4969 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4972 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4973 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4974 __free_pages(page, order);
4978 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4982 EXPORT_SYMBOL(__alloc_pages_nodemask);
4985 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4986 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4987 * you need to access high mem.
4989 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4993 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4996 return (unsigned long) page_address(page);
4998 EXPORT_SYMBOL(__get_free_pages);
5000 unsigned long get_zeroed_page(gfp_t gfp_mask)
5002 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5004 EXPORT_SYMBOL(get_zeroed_page);
5006 static inline void free_the_page(struct page *page, unsigned int order)
5008 if (order == 0) /* Via pcp? */
5009 free_unref_page(page);
5011 __free_pages_ok(page, order, FPI_NONE);
5014 void __free_pages(struct page *page, unsigned int order)
5016 if (put_page_testzero(page))
5017 free_the_page(page, order);
5018 else if (!PageHead(page))
5020 free_the_page(page + (1 << order), order);
5022 EXPORT_SYMBOL(__free_pages);
5024 void free_pages(unsigned long addr, unsigned int order)
5027 VM_BUG_ON(!virt_addr_valid((void *)addr));
5028 __free_pages(virt_to_page((void *)addr), order);
5032 EXPORT_SYMBOL(free_pages);
5036 * An arbitrary-length arbitrary-offset area of memory which resides
5037 * within a 0 or higher order page. Multiple fragments within that page
5038 * are individually refcounted, in the page's reference counter.
5040 * The page_frag functions below provide a simple allocation framework for
5041 * page fragments. This is used by the network stack and network device
5042 * drivers to provide a backing region of memory for use as either an
5043 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5045 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5048 struct page *page = NULL;
5049 gfp_t gfp = gfp_mask;
5051 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5052 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5054 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5055 PAGE_FRAG_CACHE_MAX_ORDER);
5056 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5058 if (unlikely(!page))
5059 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5061 nc->va = page ? page_address(page) : NULL;
5066 void __page_frag_cache_drain(struct page *page, unsigned int count)
5068 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5070 if (page_ref_sub_and_test(page, count))
5071 free_the_page(page, compound_order(page));
5073 EXPORT_SYMBOL(__page_frag_cache_drain);
5075 void *page_frag_alloc(struct page_frag_cache *nc,
5076 unsigned int fragsz, gfp_t gfp_mask)
5078 unsigned int size = PAGE_SIZE;
5082 if (unlikely(!nc->va)) {
5084 page = __page_frag_cache_refill(nc, gfp_mask);
5088 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5089 /* if size can vary use size else just use PAGE_SIZE */
5092 /* Even if we own the page, we do not use atomic_set().
5093 * This would break get_page_unless_zero() users.
5095 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5097 /* reset page count bias and offset to start of new frag */
5098 nc->pfmemalloc = page_is_pfmemalloc(page);
5099 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5103 offset = nc->offset - fragsz;
5104 if (unlikely(offset < 0)) {
5105 page = virt_to_page(nc->va);
5107 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5110 if (unlikely(nc->pfmemalloc)) {
5111 free_the_page(page, compound_order(page));
5115 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5116 /* if size can vary use size else just use PAGE_SIZE */
5119 /* OK, page count is 0, we can safely set it */
5120 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5122 /* reset page count bias and offset to start of new frag */
5123 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5124 offset = size - fragsz;
5128 nc->offset = offset;
5130 return nc->va + offset;
5132 EXPORT_SYMBOL(page_frag_alloc);
5135 * Frees a page fragment allocated out of either a compound or order 0 page.
5137 void page_frag_free(void *addr)
5139 struct page *page = virt_to_head_page(addr);
5141 if (unlikely(put_page_testzero(page)))
5142 free_the_page(page, compound_order(page));
5144 EXPORT_SYMBOL(page_frag_free);
5146 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5150 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5151 unsigned long used = addr + PAGE_ALIGN(size);
5153 split_page(virt_to_page((void *)addr), order);
5154 while (used < alloc_end) {
5159 return (void *)addr;
5163 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5164 * @size: the number of bytes to allocate
5165 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5167 * This function is similar to alloc_pages(), except that it allocates the
5168 * minimum number of pages to satisfy the request. alloc_pages() can only
5169 * allocate memory in power-of-two pages.
5171 * This function is also limited by MAX_ORDER.
5173 * Memory allocated by this function must be released by free_pages_exact().
5175 * Return: pointer to the allocated area or %NULL in case of error.
5177 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5179 unsigned int order = get_order(size);
5182 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5183 gfp_mask &= ~__GFP_COMP;
5185 addr = __get_free_pages(gfp_mask, order);
5186 return make_alloc_exact(addr, order, size);
5188 EXPORT_SYMBOL(alloc_pages_exact);
5191 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5193 * @nid: the preferred node ID where memory should be allocated
5194 * @size: the number of bytes to allocate
5195 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5197 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5200 * Return: pointer to the allocated area or %NULL in case of error.
5202 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5204 unsigned int order = get_order(size);
5207 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5208 gfp_mask &= ~__GFP_COMP;
5210 p = alloc_pages_node(nid, gfp_mask, order);
5213 return make_alloc_exact((unsigned long)page_address(p), order, size);
5217 * free_pages_exact - release memory allocated via alloc_pages_exact()
5218 * @virt: the value returned by alloc_pages_exact.
5219 * @size: size of allocation, same value as passed to alloc_pages_exact().
5221 * Release the memory allocated by a previous call to alloc_pages_exact.
5223 void free_pages_exact(void *virt, size_t size)
5225 unsigned long addr = (unsigned long)virt;
5226 unsigned long end = addr + PAGE_ALIGN(size);
5228 while (addr < end) {
5233 EXPORT_SYMBOL(free_pages_exact);
5236 * nr_free_zone_pages - count number of pages beyond high watermark
5237 * @offset: The zone index of the highest zone
5239 * nr_free_zone_pages() counts the number of pages which are beyond the
5240 * high watermark within all zones at or below a given zone index. For each
5241 * zone, the number of pages is calculated as:
5243 * nr_free_zone_pages = managed_pages - high_pages
5245 * Return: number of pages beyond high watermark.
5247 static unsigned long nr_free_zone_pages(int offset)
5252 /* Just pick one node, since fallback list is circular */
5253 unsigned long sum = 0;
5255 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5257 for_each_zone_zonelist(zone, z, zonelist, offset) {
5258 unsigned long size = zone_managed_pages(zone);
5259 unsigned long high = high_wmark_pages(zone);
5268 * nr_free_buffer_pages - count number of pages beyond high watermark
5270 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5271 * watermark within ZONE_DMA and ZONE_NORMAL.
5273 * Return: number of pages beyond high watermark within ZONE_DMA and
5276 unsigned long nr_free_buffer_pages(void)
5278 return nr_free_zone_pages(gfp_zone(GFP_USER));
5280 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5282 static inline void show_node(struct zone *zone)
5284 if (IS_ENABLED(CONFIG_NUMA))
5285 printk("Node %d ", zone_to_nid(zone));
5288 long si_mem_available(void)
5291 unsigned long pagecache;
5292 unsigned long wmark_low = 0;
5293 unsigned long pages[NR_LRU_LISTS];
5294 unsigned long reclaimable;
5298 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5299 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5302 wmark_low += low_wmark_pages(zone);
5305 * Estimate the amount of memory available for userspace allocations,
5306 * without causing swapping.
5308 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5311 * Not all the page cache can be freed, otherwise the system will
5312 * start swapping. Assume at least half of the page cache, or the
5313 * low watermark worth of cache, needs to stay.
5315 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5316 pagecache -= min(pagecache / 2, wmark_low);
5317 available += pagecache;
5320 * Part of the reclaimable slab and other kernel memory consists of
5321 * items that are in use, and cannot be freed. Cap this estimate at the
5324 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5325 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5326 available += reclaimable - min(reclaimable / 2, wmark_low);
5332 EXPORT_SYMBOL_GPL(si_mem_available);
5334 void si_meminfo(struct sysinfo *val)
5336 val->totalram = totalram_pages();
5337 val->sharedram = global_node_page_state(NR_SHMEM);
5338 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5339 val->bufferram = nr_blockdev_pages();
5340 val->totalhigh = totalhigh_pages();
5341 val->freehigh = nr_free_highpages();
5342 val->mem_unit = PAGE_SIZE;
5345 EXPORT_SYMBOL(si_meminfo);
5348 void si_meminfo_node(struct sysinfo *val, int nid)
5350 int zone_type; /* needs to be signed */
5351 unsigned long managed_pages = 0;
5352 unsigned long managed_highpages = 0;
5353 unsigned long free_highpages = 0;
5354 pg_data_t *pgdat = NODE_DATA(nid);
5356 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5357 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5358 val->totalram = managed_pages;
5359 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5360 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5361 #ifdef CONFIG_HIGHMEM
5362 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5363 struct zone *zone = &pgdat->node_zones[zone_type];
5365 if (is_highmem(zone)) {
5366 managed_highpages += zone_managed_pages(zone);
5367 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5370 val->totalhigh = managed_highpages;
5371 val->freehigh = free_highpages;
5373 val->totalhigh = managed_highpages;
5374 val->freehigh = free_highpages;
5376 val->mem_unit = PAGE_SIZE;
5381 * Determine whether the node should be displayed or not, depending on whether
5382 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5384 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5386 if (!(flags & SHOW_MEM_FILTER_NODES))
5390 * no node mask - aka implicit memory numa policy. Do not bother with
5391 * the synchronization - read_mems_allowed_begin - because we do not
5392 * have to be precise here.
5395 nodemask = &cpuset_current_mems_allowed;
5397 return !node_isset(nid, *nodemask);
5400 #define K(x) ((x) << (PAGE_SHIFT-10))
5402 static void show_migration_types(unsigned char type)
5404 static const char types[MIGRATE_TYPES] = {
5405 [MIGRATE_UNMOVABLE] = 'U',
5406 [MIGRATE_MOVABLE] = 'M',
5407 [MIGRATE_RECLAIMABLE] = 'E',
5408 [MIGRATE_HIGHATOMIC] = 'H',
5410 [MIGRATE_CMA] = 'C',
5412 #ifdef CONFIG_MEMORY_ISOLATION
5413 [MIGRATE_ISOLATE] = 'I',
5416 char tmp[MIGRATE_TYPES + 1];
5420 for (i = 0; i < MIGRATE_TYPES; i++) {
5421 if (type & (1 << i))
5426 printk(KERN_CONT "(%s) ", tmp);
5430 * Show free area list (used inside shift_scroll-lock stuff)
5431 * We also calculate the percentage fragmentation. We do this by counting the
5432 * memory on each free list with the exception of the first item on the list.
5435 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5438 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5440 unsigned long free_pcp = 0;
5445 for_each_populated_zone(zone) {
5446 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5449 for_each_online_cpu(cpu)
5450 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5453 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5454 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5455 " unevictable:%lu dirty:%lu writeback:%lu\n"
5456 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5457 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5458 " free:%lu free_pcp:%lu free_cma:%lu\n",
5459 global_node_page_state(NR_ACTIVE_ANON),
5460 global_node_page_state(NR_INACTIVE_ANON),
5461 global_node_page_state(NR_ISOLATED_ANON),
5462 global_node_page_state(NR_ACTIVE_FILE),
5463 global_node_page_state(NR_INACTIVE_FILE),
5464 global_node_page_state(NR_ISOLATED_FILE),
5465 global_node_page_state(NR_UNEVICTABLE),
5466 global_node_page_state(NR_FILE_DIRTY),
5467 global_node_page_state(NR_WRITEBACK),
5468 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5469 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5470 global_node_page_state(NR_FILE_MAPPED),
5471 global_node_page_state(NR_SHMEM),
5472 global_zone_page_state(NR_PAGETABLE),
5473 global_zone_page_state(NR_BOUNCE),
5474 global_zone_page_state(NR_FREE_PAGES),
5476 global_zone_page_state(NR_FREE_CMA_PAGES));
5478 for_each_online_pgdat(pgdat) {
5479 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5483 " active_anon:%lukB"
5484 " inactive_anon:%lukB"
5485 " active_file:%lukB"
5486 " inactive_file:%lukB"
5487 " unevictable:%lukB"
5488 " isolated(anon):%lukB"
5489 " isolated(file):%lukB"
5494 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5496 " shmem_pmdmapped: %lukB"
5499 " writeback_tmp:%lukB"
5500 " kernel_stack:%lukB"
5501 #ifdef CONFIG_SHADOW_CALL_STACK
5502 " shadow_call_stack:%lukB"
5504 " all_unreclaimable? %s"
5507 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5508 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5509 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5510 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5511 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5512 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5513 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5514 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5515 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5516 K(node_page_state(pgdat, NR_WRITEBACK)),
5517 K(node_page_state(pgdat, NR_SHMEM)),
5518 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5519 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5520 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5522 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5524 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5525 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5526 #ifdef CONFIG_SHADOW_CALL_STACK
5527 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5529 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5533 for_each_populated_zone(zone) {
5536 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5540 for_each_online_cpu(cpu)
5541 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5550 " reserved_highatomic:%luKB"
5551 " active_anon:%lukB"
5552 " inactive_anon:%lukB"
5553 " active_file:%lukB"
5554 " inactive_file:%lukB"
5555 " unevictable:%lukB"
5556 " writepending:%lukB"
5567 K(zone_page_state(zone, NR_FREE_PAGES)),
5568 K(min_wmark_pages(zone)),
5569 K(low_wmark_pages(zone)),
5570 K(high_wmark_pages(zone)),
5571 K(zone->nr_reserved_highatomic),
5572 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5573 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5574 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5575 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5576 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5577 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5578 K(zone->present_pages),
5579 K(zone_managed_pages(zone)),
5580 K(zone_page_state(zone, NR_MLOCK)),
5581 K(zone_page_state(zone, NR_PAGETABLE)),
5582 K(zone_page_state(zone, NR_BOUNCE)),
5584 K(this_cpu_read(zone->pageset->pcp.count)),
5585 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5586 printk("lowmem_reserve[]:");
5587 for (i = 0; i < MAX_NR_ZONES; i++)
5588 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5589 printk(KERN_CONT "\n");
5592 for_each_populated_zone(zone) {
5594 unsigned long nr[MAX_ORDER], flags, total = 0;
5595 unsigned char types[MAX_ORDER];
5597 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5600 printk(KERN_CONT "%s: ", zone->name);
5602 spin_lock_irqsave(&zone->lock, flags);
5603 for (order = 0; order < MAX_ORDER; order++) {
5604 struct free_area *area = &zone->free_area[order];
5607 nr[order] = area->nr_free;
5608 total += nr[order] << order;
5611 for (type = 0; type < MIGRATE_TYPES; type++) {
5612 if (!free_area_empty(area, type))
5613 types[order] |= 1 << type;
5616 spin_unlock_irqrestore(&zone->lock, flags);
5617 for (order = 0; order < MAX_ORDER; order++) {
5618 printk(KERN_CONT "%lu*%lukB ",
5619 nr[order], K(1UL) << order);
5621 show_migration_types(types[order]);
5623 printk(KERN_CONT "= %lukB\n", K(total));
5626 hugetlb_show_meminfo();
5628 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5630 show_swap_cache_info();
5633 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5635 zoneref->zone = zone;
5636 zoneref->zone_idx = zone_idx(zone);
5640 * Builds allocation fallback zone lists.
5642 * Add all populated zones of a node to the zonelist.
5644 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5647 enum zone_type zone_type = MAX_NR_ZONES;
5652 zone = pgdat->node_zones + zone_type;
5653 if (managed_zone(zone)) {
5654 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5655 check_highest_zone(zone_type);
5657 } while (zone_type);
5664 static int __parse_numa_zonelist_order(char *s)
5667 * We used to support different zonlists modes but they turned
5668 * out to be just not useful. Let's keep the warning in place
5669 * if somebody still use the cmd line parameter so that we do
5670 * not fail it silently
5672 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5673 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5679 char numa_zonelist_order[] = "Node";
5682 * sysctl handler for numa_zonelist_order
5684 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5685 void *buffer, size_t *length, loff_t *ppos)
5688 return __parse_numa_zonelist_order(buffer);
5689 return proc_dostring(table, write, buffer, length, ppos);
5693 #define MAX_NODE_LOAD (nr_online_nodes)
5694 static int node_load[MAX_NUMNODES];
5697 * find_next_best_node - find the next node that should appear in a given node's fallback list
5698 * @node: node whose fallback list we're appending
5699 * @used_node_mask: nodemask_t of already used nodes
5701 * We use a number of factors to determine which is the next node that should
5702 * appear on a given node's fallback list. The node should not have appeared
5703 * already in @node's fallback list, and it should be the next closest node
5704 * according to the distance array (which contains arbitrary distance values
5705 * from each node to each node in the system), and should also prefer nodes
5706 * with no CPUs, since presumably they'll have very little allocation pressure
5707 * on them otherwise.
5709 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5711 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5714 int min_val = INT_MAX;
5715 int best_node = NUMA_NO_NODE;
5717 /* Use the local node if we haven't already */
5718 if (!node_isset(node, *used_node_mask)) {
5719 node_set(node, *used_node_mask);
5723 for_each_node_state(n, N_MEMORY) {
5725 /* Don't want a node to appear more than once */
5726 if (node_isset(n, *used_node_mask))
5729 /* Use the distance array to find the distance */
5730 val = node_distance(node, n);
5732 /* Penalize nodes under us ("prefer the next node") */
5735 /* Give preference to headless and unused nodes */
5736 if (!cpumask_empty(cpumask_of_node(n)))
5737 val += PENALTY_FOR_NODE_WITH_CPUS;
5739 /* Slight preference for less loaded node */
5740 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5741 val += node_load[n];
5743 if (val < min_val) {
5750 node_set(best_node, *used_node_mask);
5757 * Build zonelists ordered by node and zones within node.
5758 * This results in maximum locality--normal zone overflows into local
5759 * DMA zone, if any--but risks exhausting DMA zone.
5761 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5764 struct zoneref *zonerefs;
5767 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5769 for (i = 0; i < nr_nodes; i++) {
5772 pg_data_t *node = NODE_DATA(node_order[i]);
5774 nr_zones = build_zonerefs_node(node, zonerefs);
5775 zonerefs += nr_zones;
5777 zonerefs->zone = NULL;
5778 zonerefs->zone_idx = 0;
5782 * Build gfp_thisnode zonelists
5784 static void build_thisnode_zonelists(pg_data_t *pgdat)
5786 struct zoneref *zonerefs;
5789 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5790 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5791 zonerefs += nr_zones;
5792 zonerefs->zone = NULL;
5793 zonerefs->zone_idx = 0;
5797 * Build zonelists ordered by zone and nodes within zones.
5798 * This results in conserving DMA zone[s] until all Normal memory is
5799 * exhausted, but results in overflowing to remote node while memory
5800 * may still exist in local DMA zone.
5803 static void build_zonelists(pg_data_t *pgdat)
5805 static int node_order[MAX_NUMNODES];
5806 int node, load, nr_nodes = 0;
5807 nodemask_t used_mask = NODE_MASK_NONE;
5808 int local_node, prev_node;
5810 /* NUMA-aware ordering of nodes */
5811 local_node = pgdat->node_id;
5812 load = nr_online_nodes;
5813 prev_node = local_node;
5815 memset(node_order, 0, sizeof(node_order));
5816 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5818 * We don't want to pressure a particular node.
5819 * So adding penalty to the first node in same
5820 * distance group to make it round-robin.
5822 if (node_distance(local_node, node) !=
5823 node_distance(local_node, prev_node))
5824 node_load[node] = load;
5826 node_order[nr_nodes++] = node;
5831 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5832 build_thisnode_zonelists(pgdat);
5835 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5837 * Return node id of node used for "local" allocations.
5838 * I.e., first node id of first zone in arg node's generic zonelist.
5839 * Used for initializing percpu 'numa_mem', which is used primarily
5840 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5842 int local_memory_node(int node)
5846 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5847 gfp_zone(GFP_KERNEL),
5849 return zone_to_nid(z->zone);
5853 static void setup_min_unmapped_ratio(void);
5854 static void setup_min_slab_ratio(void);
5855 #else /* CONFIG_NUMA */
5857 static void build_zonelists(pg_data_t *pgdat)
5859 int node, local_node;
5860 struct zoneref *zonerefs;
5863 local_node = pgdat->node_id;
5865 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5866 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5867 zonerefs += nr_zones;
5870 * Now we build the zonelist so that it contains the zones
5871 * of all the other nodes.
5872 * We don't want to pressure a particular node, so when
5873 * building the zones for node N, we make sure that the
5874 * zones coming right after the local ones are those from
5875 * node N+1 (modulo N)
5877 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5878 if (!node_online(node))
5880 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5881 zonerefs += nr_zones;
5883 for (node = 0; node < local_node; node++) {
5884 if (!node_online(node))
5886 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5887 zonerefs += nr_zones;
5890 zonerefs->zone = NULL;
5891 zonerefs->zone_idx = 0;
5894 #endif /* CONFIG_NUMA */
5897 * Boot pageset table. One per cpu which is going to be used for all
5898 * zones and all nodes. The parameters will be set in such a way
5899 * that an item put on a list will immediately be handed over to
5900 * the buddy list. This is safe since pageset manipulation is done
5901 * with interrupts disabled.
5903 * The boot_pagesets must be kept even after bootup is complete for
5904 * unused processors and/or zones. They do play a role for bootstrapping
5905 * hotplugged processors.
5907 * zoneinfo_show() and maybe other functions do
5908 * not check if the processor is online before following the pageset pointer.
5909 * Other parts of the kernel may not check if the zone is available.
5911 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5912 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5913 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5915 static void __build_all_zonelists(void *data)
5918 int __maybe_unused cpu;
5919 pg_data_t *self = data;
5920 static DEFINE_SPINLOCK(lock);
5925 memset(node_load, 0, sizeof(node_load));
5929 * This node is hotadded and no memory is yet present. So just
5930 * building zonelists is fine - no need to touch other nodes.
5932 if (self && !node_online(self->node_id)) {
5933 build_zonelists(self);
5935 for_each_online_node(nid) {
5936 pg_data_t *pgdat = NODE_DATA(nid);
5938 build_zonelists(pgdat);
5941 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5943 * We now know the "local memory node" for each node--
5944 * i.e., the node of the first zone in the generic zonelist.
5945 * Set up numa_mem percpu variable for on-line cpus. During
5946 * boot, only the boot cpu should be on-line; we'll init the
5947 * secondary cpus' numa_mem as they come on-line. During
5948 * node/memory hotplug, we'll fixup all on-line cpus.
5950 for_each_online_cpu(cpu)
5951 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5958 static noinline void __init
5959 build_all_zonelists_init(void)
5963 __build_all_zonelists(NULL);
5966 * Initialize the boot_pagesets that are going to be used
5967 * for bootstrapping processors. The real pagesets for
5968 * each zone will be allocated later when the per cpu
5969 * allocator is available.
5971 * boot_pagesets are used also for bootstrapping offline
5972 * cpus if the system is already booted because the pagesets
5973 * are needed to initialize allocators on a specific cpu too.
5974 * F.e. the percpu allocator needs the page allocator which
5975 * needs the percpu allocator in order to allocate its pagesets
5976 * (a chicken-egg dilemma).
5978 for_each_possible_cpu(cpu)
5979 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5981 mminit_verify_zonelist();
5982 cpuset_init_current_mems_allowed();
5986 * unless system_state == SYSTEM_BOOTING.
5988 * __ref due to call of __init annotated helper build_all_zonelists_init
5989 * [protected by SYSTEM_BOOTING].
5991 void __ref build_all_zonelists(pg_data_t *pgdat)
5993 unsigned long vm_total_pages;
5995 if (system_state == SYSTEM_BOOTING) {
5996 build_all_zonelists_init();
5998 __build_all_zonelists(pgdat);
5999 /* cpuset refresh routine should be here */
6001 /* Get the number of free pages beyond high watermark in all zones. */
6002 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6004 * Disable grouping by mobility if the number of pages in the
6005 * system is too low to allow the mechanism to work. It would be
6006 * more accurate, but expensive to check per-zone. This check is
6007 * made on memory-hotadd so a system can start with mobility
6008 * disabled and enable it later
6010 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6011 page_group_by_mobility_disabled = 1;
6013 page_group_by_mobility_disabled = 0;
6015 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6017 page_group_by_mobility_disabled ? "off" : "on",
6020 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6024 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6025 static bool __meminit
6026 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6028 static struct memblock_region *r;
6030 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6031 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6032 for_each_mem_region(r) {
6033 if (*pfn < memblock_region_memory_end_pfn(r))
6037 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6038 memblock_is_mirror(r)) {
6039 *pfn = memblock_region_memory_end_pfn(r);
6047 * Initially all pages are reserved - free ones are freed
6048 * up by memblock_free_all() once the early boot process is
6049 * done. Non-atomic initialization, single-pass.
6051 * All aligned pageblocks are initialized to the specified migratetype
6052 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6053 * zone stats (e.g., nr_isolate_pageblock) are touched.
6055 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
6056 unsigned long start_pfn, unsigned long zone_end_pfn,
6057 enum meminit_context context,
6058 struct vmem_altmap *altmap, int migratetype)
6060 unsigned long pfn, end_pfn = start_pfn + size;
6063 if (highest_memmap_pfn < end_pfn - 1)
6064 highest_memmap_pfn = end_pfn - 1;
6066 #ifdef CONFIG_ZONE_DEVICE
6068 * Honor reservation requested by the driver for this ZONE_DEVICE
6069 * memory. We limit the total number of pages to initialize to just
6070 * those that might contain the memory mapping. We will defer the
6071 * ZONE_DEVICE page initialization until after we have released
6074 if (zone == ZONE_DEVICE) {
6078 if (start_pfn == altmap->base_pfn)
6079 start_pfn += altmap->reserve;
6080 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6084 for (pfn = start_pfn; pfn < end_pfn; ) {
6086 * There can be holes in boot-time mem_map[]s handed to this
6087 * function. They do not exist on hotplugged memory.
6089 if (context == MEMINIT_EARLY) {
6090 if (overlap_memmap_init(zone, &pfn))
6092 if (defer_init(nid, pfn, zone_end_pfn))
6096 page = pfn_to_page(pfn);
6097 __init_single_page(page, pfn, zone, nid);
6098 if (context == MEMINIT_HOTPLUG)
6099 __SetPageReserved(page);
6102 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6103 * such that unmovable allocations won't be scattered all
6104 * over the place during system boot.
6106 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6107 set_pageblock_migratetype(page, migratetype);
6114 #ifdef CONFIG_ZONE_DEVICE
6115 void __ref memmap_init_zone_device(struct zone *zone,
6116 unsigned long start_pfn,
6117 unsigned long nr_pages,
6118 struct dev_pagemap *pgmap)
6120 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6121 struct pglist_data *pgdat = zone->zone_pgdat;
6122 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6123 unsigned long zone_idx = zone_idx(zone);
6124 unsigned long start = jiffies;
6125 int nid = pgdat->node_id;
6127 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6131 * The call to memmap_init_zone should have already taken care
6132 * of the pages reserved for the memmap, so we can just jump to
6133 * the end of that region and start processing the device pages.
6136 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6137 nr_pages = end_pfn - start_pfn;
6140 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6141 struct page *page = pfn_to_page(pfn);
6143 __init_single_page(page, pfn, zone_idx, nid);
6146 * Mark page reserved as it will need to wait for onlining
6147 * phase for it to be fully associated with a zone.
6149 * We can use the non-atomic __set_bit operation for setting
6150 * the flag as we are still initializing the pages.
6152 __SetPageReserved(page);
6155 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6156 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6157 * ever freed or placed on a driver-private list.
6159 page->pgmap = pgmap;
6160 page->zone_device_data = NULL;
6163 * Mark the block movable so that blocks are reserved for
6164 * movable at startup. This will force kernel allocations
6165 * to reserve their blocks rather than leaking throughout
6166 * the address space during boot when many long-lived
6167 * kernel allocations are made.
6169 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6170 * because this is done early in section_activate()
6172 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6173 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6178 pr_info("%s initialised %lu pages in %ums\n", __func__,
6179 nr_pages, jiffies_to_msecs(jiffies - start));
6183 static void __meminit zone_init_free_lists(struct zone *zone)
6185 unsigned int order, t;
6186 for_each_migratetype_order(order, t) {
6187 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6188 zone->free_area[order].nr_free = 0;
6192 void __meminit __weak memmap_init(unsigned long size, int nid,
6194 unsigned long range_start_pfn)
6196 unsigned long start_pfn, end_pfn;
6197 unsigned long range_end_pfn = range_start_pfn + size;
6200 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6201 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6202 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6204 if (end_pfn > start_pfn) {
6205 size = end_pfn - start_pfn;
6206 memmap_init_zone(size, nid, zone, start_pfn, range_end_pfn,
6207 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6212 static int zone_batchsize(struct zone *zone)
6218 * The per-cpu-pages pools are set to around 1000th of the
6221 batch = zone_managed_pages(zone) / 1024;
6222 /* But no more than a meg. */
6223 if (batch * PAGE_SIZE > 1024 * 1024)
6224 batch = (1024 * 1024) / PAGE_SIZE;
6225 batch /= 4; /* We effectively *= 4 below */
6230 * Clamp the batch to a 2^n - 1 value. Having a power
6231 * of 2 value was found to be more likely to have
6232 * suboptimal cache aliasing properties in some cases.
6234 * For example if 2 tasks are alternately allocating
6235 * batches of pages, one task can end up with a lot
6236 * of pages of one half of the possible page colors
6237 * and the other with pages of the other colors.
6239 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6244 /* The deferral and batching of frees should be suppressed under NOMMU
6247 * The problem is that NOMMU needs to be able to allocate large chunks
6248 * of contiguous memory as there's no hardware page translation to
6249 * assemble apparent contiguous memory from discontiguous pages.
6251 * Queueing large contiguous runs of pages for batching, however,
6252 * causes the pages to actually be freed in smaller chunks. As there
6253 * can be a significant delay between the individual batches being
6254 * recycled, this leads to the once large chunks of space being
6255 * fragmented and becoming unavailable for high-order allocations.
6262 * pcp->high and pcp->batch values are related and dependent on one another:
6263 * ->batch must never be higher then ->high.
6264 * The following function updates them in a safe manner without read side
6267 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6268 * those fields changing asynchronously (acording to the above rule).
6270 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6271 * outside of boot time (or some other assurance that no concurrent updaters
6274 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6275 unsigned long batch)
6277 /* start with a fail safe value for batch */
6281 /* Update high, then batch, in order */
6288 /* a companion to pageset_set_high() */
6289 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6291 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6294 static void pageset_init(struct per_cpu_pageset *p)
6296 struct per_cpu_pages *pcp;
6299 memset(p, 0, sizeof(*p));
6302 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6303 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6306 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6309 pageset_set_batch(p, batch);
6313 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6314 * to the value high for the pageset p.
6316 static void pageset_set_high(struct per_cpu_pageset *p,
6319 unsigned long batch = max(1UL, high / 4);
6320 if ((high / 4) > (PAGE_SHIFT * 8))
6321 batch = PAGE_SHIFT * 8;
6323 pageset_update(&p->pcp, high, batch);
6326 static void pageset_set_high_and_batch(struct zone *zone,
6327 struct per_cpu_pageset *pcp)
6329 if (percpu_pagelist_fraction)
6330 pageset_set_high(pcp,
6331 (zone_managed_pages(zone) /
6332 percpu_pagelist_fraction));
6334 pageset_set_batch(pcp, zone_batchsize(zone));
6337 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6339 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6342 pageset_set_high_and_batch(zone, pcp);
6345 void __meminit setup_zone_pageset(struct zone *zone)
6348 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6349 for_each_possible_cpu(cpu)
6350 zone_pageset_init(zone, cpu);
6354 * Allocate per cpu pagesets and initialize them.
6355 * Before this call only boot pagesets were available.
6357 void __init setup_per_cpu_pageset(void)
6359 struct pglist_data *pgdat;
6361 int __maybe_unused cpu;
6363 for_each_populated_zone(zone)
6364 setup_zone_pageset(zone);
6368 * Unpopulated zones continue using the boot pagesets.
6369 * The numa stats for these pagesets need to be reset.
6370 * Otherwise, they will end up skewing the stats of
6371 * the nodes these zones are associated with.
6373 for_each_possible_cpu(cpu) {
6374 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6375 memset(pcp->vm_numa_stat_diff, 0,
6376 sizeof(pcp->vm_numa_stat_diff));
6380 for_each_online_pgdat(pgdat)
6381 pgdat->per_cpu_nodestats =
6382 alloc_percpu(struct per_cpu_nodestat);
6385 static __meminit void zone_pcp_init(struct zone *zone)
6388 * per cpu subsystem is not up at this point. The following code
6389 * relies on the ability of the linker to provide the
6390 * offset of a (static) per cpu variable into the per cpu area.
6392 zone->pageset = &boot_pageset;
6394 if (populated_zone(zone))
6395 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6396 zone->name, zone->present_pages,
6397 zone_batchsize(zone));
6400 void __meminit init_currently_empty_zone(struct zone *zone,
6401 unsigned long zone_start_pfn,
6404 struct pglist_data *pgdat = zone->zone_pgdat;
6405 int zone_idx = zone_idx(zone) + 1;
6407 if (zone_idx > pgdat->nr_zones)
6408 pgdat->nr_zones = zone_idx;
6410 zone->zone_start_pfn = zone_start_pfn;
6412 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6413 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6415 (unsigned long)zone_idx(zone),
6416 zone_start_pfn, (zone_start_pfn + size));
6418 zone_init_free_lists(zone);
6419 zone->initialized = 1;
6423 * get_pfn_range_for_nid - Return the start and end page frames for a node
6424 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6425 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6426 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6428 * It returns the start and end page frame of a node based on information
6429 * provided by memblock_set_node(). If called for a node
6430 * with no available memory, a warning is printed and the start and end
6433 void __init get_pfn_range_for_nid(unsigned int nid,
6434 unsigned long *start_pfn, unsigned long *end_pfn)
6436 unsigned long this_start_pfn, this_end_pfn;
6442 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6443 *start_pfn = min(*start_pfn, this_start_pfn);
6444 *end_pfn = max(*end_pfn, this_end_pfn);
6447 if (*start_pfn == -1UL)
6452 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6453 * assumption is made that zones within a node are ordered in monotonic
6454 * increasing memory addresses so that the "highest" populated zone is used
6456 static void __init find_usable_zone_for_movable(void)
6459 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6460 if (zone_index == ZONE_MOVABLE)
6463 if (arch_zone_highest_possible_pfn[zone_index] >
6464 arch_zone_lowest_possible_pfn[zone_index])
6468 VM_BUG_ON(zone_index == -1);
6469 movable_zone = zone_index;
6473 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6474 * because it is sized independent of architecture. Unlike the other zones,
6475 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6476 * in each node depending on the size of each node and how evenly kernelcore
6477 * is distributed. This helper function adjusts the zone ranges
6478 * provided by the architecture for a given node by using the end of the
6479 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6480 * zones within a node are in order of monotonic increases memory addresses
6482 static void __init adjust_zone_range_for_zone_movable(int nid,
6483 unsigned long zone_type,
6484 unsigned long node_start_pfn,
6485 unsigned long node_end_pfn,
6486 unsigned long *zone_start_pfn,
6487 unsigned long *zone_end_pfn)
6489 /* Only adjust if ZONE_MOVABLE is on this node */
6490 if (zone_movable_pfn[nid]) {
6491 /* Size ZONE_MOVABLE */
6492 if (zone_type == ZONE_MOVABLE) {
6493 *zone_start_pfn = zone_movable_pfn[nid];
6494 *zone_end_pfn = min(node_end_pfn,
6495 arch_zone_highest_possible_pfn[movable_zone]);
6497 /* Adjust for ZONE_MOVABLE starting within this range */
6498 } else if (!mirrored_kernelcore &&
6499 *zone_start_pfn < zone_movable_pfn[nid] &&
6500 *zone_end_pfn > zone_movable_pfn[nid]) {
6501 *zone_end_pfn = zone_movable_pfn[nid];
6503 /* Check if this whole range is within ZONE_MOVABLE */
6504 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6505 *zone_start_pfn = *zone_end_pfn;
6510 * Return the number of pages a zone spans in a node, including holes
6511 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6513 static unsigned long __init zone_spanned_pages_in_node(int nid,
6514 unsigned long zone_type,
6515 unsigned long node_start_pfn,
6516 unsigned long node_end_pfn,
6517 unsigned long *zone_start_pfn,
6518 unsigned long *zone_end_pfn)
6520 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6521 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6522 /* When hotadd a new node from cpu_up(), the node should be empty */
6523 if (!node_start_pfn && !node_end_pfn)
6526 /* Get the start and end of the zone */
6527 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6528 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6529 adjust_zone_range_for_zone_movable(nid, zone_type,
6530 node_start_pfn, node_end_pfn,
6531 zone_start_pfn, zone_end_pfn);
6533 /* Check that this node has pages within the zone's required range */
6534 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6537 /* Move the zone boundaries inside the node if necessary */
6538 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6539 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6541 /* Return the spanned pages */
6542 return *zone_end_pfn - *zone_start_pfn;
6546 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6547 * then all holes in the requested range will be accounted for.
6549 unsigned long __init __absent_pages_in_range(int nid,
6550 unsigned long range_start_pfn,
6551 unsigned long range_end_pfn)
6553 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6554 unsigned long start_pfn, end_pfn;
6557 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6558 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6559 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6560 nr_absent -= end_pfn - start_pfn;
6566 * absent_pages_in_range - Return number of page frames in holes within a range
6567 * @start_pfn: The start PFN to start searching for holes
6568 * @end_pfn: The end PFN to stop searching for holes
6570 * Return: the number of pages frames in memory holes within a range.
6572 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6573 unsigned long end_pfn)
6575 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6578 /* Return the number of page frames in holes in a zone on a node */
6579 static unsigned long __init zone_absent_pages_in_node(int nid,
6580 unsigned long zone_type,
6581 unsigned long node_start_pfn,
6582 unsigned long node_end_pfn)
6584 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6585 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6586 unsigned long zone_start_pfn, zone_end_pfn;
6587 unsigned long nr_absent;
6589 /* When hotadd a new node from cpu_up(), the node should be empty */
6590 if (!node_start_pfn && !node_end_pfn)
6593 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6594 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6596 adjust_zone_range_for_zone_movable(nid, zone_type,
6597 node_start_pfn, node_end_pfn,
6598 &zone_start_pfn, &zone_end_pfn);
6599 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6602 * ZONE_MOVABLE handling.
6603 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6606 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6607 unsigned long start_pfn, end_pfn;
6608 struct memblock_region *r;
6610 for_each_mem_region(r) {
6611 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6612 zone_start_pfn, zone_end_pfn);
6613 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6614 zone_start_pfn, zone_end_pfn);
6616 if (zone_type == ZONE_MOVABLE &&
6617 memblock_is_mirror(r))
6618 nr_absent += end_pfn - start_pfn;
6620 if (zone_type == ZONE_NORMAL &&
6621 !memblock_is_mirror(r))
6622 nr_absent += end_pfn - start_pfn;
6629 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6630 unsigned long node_start_pfn,
6631 unsigned long node_end_pfn)
6633 unsigned long realtotalpages = 0, totalpages = 0;
6636 for (i = 0; i < MAX_NR_ZONES; i++) {
6637 struct zone *zone = pgdat->node_zones + i;
6638 unsigned long zone_start_pfn, zone_end_pfn;
6639 unsigned long spanned, absent;
6640 unsigned long size, real_size;
6642 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6647 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6652 real_size = size - absent;
6655 zone->zone_start_pfn = zone_start_pfn;
6657 zone->zone_start_pfn = 0;
6658 zone->spanned_pages = size;
6659 zone->present_pages = real_size;
6662 realtotalpages += real_size;
6665 pgdat->node_spanned_pages = totalpages;
6666 pgdat->node_present_pages = realtotalpages;
6667 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6671 #ifndef CONFIG_SPARSEMEM
6673 * Calculate the size of the zone->blockflags rounded to an unsigned long
6674 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6675 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6676 * round what is now in bits to nearest long in bits, then return it in
6679 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6681 unsigned long usemapsize;
6683 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6684 usemapsize = roundup(zonesize, pageblock_nr_pages);
6685 usemapsize = usemapsize >> pageblock_order;
6686 usemapsize *= NR_PAGEBLOCK_BITS;
6687 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6689 return usemapsize / 8;
6692 static void __ref setup_usemap(struct pglist_data *pgdat,
6694 unsigned long zone_start_pfn,
6695 unsigned long zonesize)
6697 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6698 zone->pageblock_flags = NULL;
6700 zone->pageblock_flags =
6701 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6703 if (!zone->pageblock_flags)
6704 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6705 usemapsize, zone->name, pgdat->node_id);
6709 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6710 unsigned long zone_start_pfn, unsigned long zonesize) {}
6711 #endif /* CONFIG_SPARSEMEM */
6713 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6715 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6716 void __init set_pageblock_order(void)
6720 /* Check that pageblock_nr_pages has not already been setup */
6721 if (pageblock_order)
6724 if (HPAGE_SHIFT > PAGE_SHIFT)
6725 order = HUGETLB_PAGE_ORDER;
6727 order = MAX_ORDER - 1;
6730 * Assume the largest contiguous order of interest is a huge page.
6731 * This value may be variable depending on boot parameters on IA64 and
6734 pageblock_order = order;
6736 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6739 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6740 * is unused as pageblock_order is set at compile-time. See
6741 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6744 void __init set_pageblock_order(void)
6748 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6750 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6751 unsigned long present_pages)
6753 unsigned long pages = spanned_pages;
6756 * Provide a more accurate estimation if there are holes within
6757 * the zone and SPARSEMEM is in use. If there are holes within the
6758 * zone, each populated memory region may cost us one or two extra
6759 * memmap pages due to alignment because memmap pages for each
6760 * populated regions may not be naturally aligned on page boundary.
6761 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6763 if (spanned_pages > present_pages + (present_pages >> 4) &&
6764 IS_ENABLED(CONFIG_SPARSEMEM))
6765 pages = present_pages;
6767 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6770 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6771 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6773 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6775 spin_lock_init(&ds_queue->split_queue_lock);
6776 INIT_LIST_HEAD(&ds_queue->split_queue);
6777 ds_queue->split_queue_len = 0;
6780 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6783 #ifdef CONFIG_COMPACTION
6784 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6786 init_waitqueue_head(&pgdat->kcompactd_wait);
6789 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6792 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6794 pgdat_resize_init(pgdat);
6796 pgdat_init_split_queue(pgdat);
6797 pgdat_init_kcompactd(pgdat);
6799 init_waitqueue_head(&pgdat->kswapd_wait);
6800 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6802 pgdat_page_ext_init(pgdat);
6803 spin_lock_init(&pgdat->lru_lock);
6804 lruvec_init(&pgdat->__lruvec);
6807 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6808 unsigned long remaining_pages)
6810 atomic_long_set(&zone->managed_pages, remaining_pages);
6811 zone_set_nid(zone, nid);
6812 zone->name = zone_names[idx];
6813 zone->zone_pgdat = NODE_DATA(nid);
6814 spin_lock_init(&zone->lock);
6815 zone_seqlock_init(zone);
6816 zone_pcp_init(zone);
6820 * Set up the zone data structures
6821 * - init pgdat internals
6822 * - init all zones belonging to this node
6824 * NOTE: this function is only called during memory hotplug
6826 #ifdef CONFIG_MEMORY_HOTPLUG
6827 void __ref free_area_init_core_hotplug(int nid)
6830 pg_data_t *pgdat = NODE_DATA(nid);
6832 pgdat_init_internals(pgdat);
6833 for (z = 0; z < MAX_NR_ZONES; z++)
6834 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6839 * Set up the zone data structures:
6840 * - mark all pages reserved
6841 * - mark all memory queues empty
6842 * - clear the memory bitmaps
6844 * NOTE: pgdat should get zeroed by caller.
6845 * NOTE: this function is only called during early init.
6847 static void __init free_area_init_core(struct pglist_data *pgdat)
6850 int nid = pgdat->node_id;
6852 pgdat_init_internals(pgdat);
6853 pgdat->per_cpu_nodestats = &boot_nodestats;
6855 for (j = 0; j < MAX_NR_ZONES; j++) {
6856 struct zone *zone = pgdat->node_zones + j;
6857 unsigned long size, freesize, memmap_pages;
6858 unsigned long zone_start_pfn = zone->zone_start_pfn;
6860 size = zone->spanned_pages;
6861 freesize = zone->present_pages;
6864 * Adjust freesize so that it accounts for how much memory
6865 * is used by this zone for memmap. This affects the watermark
6866 * and per-cpu initialisations
6868 memmap_pages = calc_memmap_size(size, freesize);
6869 if (!is_highmem_idx(j)) {
6870 if (freesize >= memmap_pages) {
6871 freesize -= memmap_pages;
6874 " %s zone: %lu pages used for memmap\n",
6875 zone_names[j], memmap_pages);
6877 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6878 zone_names[j], memmap_pages, freesize);
6881 /* Account for reserved pages */
6882 if (j == 0 && freesize > dma_reserve) {
6883 freesize -= dma_reserve;
6884 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6885 zone_names[0], dma_reserve);
6888 if (!is_highmem_idx(j))
6889 nr_kernel_pages += freesize;
6890 /* Charge for highmem memmap if there are enough kernel pages */
6891 else if (nr_kernel_pages > memmap_pages * 2)
6892 nr_kernel_pages -= memmap_pages;
6893 nr_all_pages += freesize;
6896 * Set an approximate value for lowmem here, it will be adjusted
6897 * when the bootmem allocator frees pages into the buddy system.
6898 * And all highmem pages will be managed by the buddy system.
6900 zone_init_internals(zone, j, nid, freesize);
6905 set_pageblock_order();
6906 setup_usemap(pgdat, zone, zone_start_pfn, size);
6907 init_currently_empty_zone(zone, zone_start_pfn, size);
6908 memmap_init(size, nid, j, zone_start_pfn);
6912 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6913 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6915 unsigned long __maybe_unused start = 0;
6916 unsigned long __maybe_unused offset = 0;
6918 /* Skip empty nodes */
6919 if (!pgdat->node_spanned_pages)
6922 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6923 offset = pgdat->node_start_pfn - start;
6924 /* ia64 gets its own node_mem_map, before this, without bootmem */
6925 if (!pgdat->node_mem_map) {
6926 unsigned long size, end;
6930 * The zone's endpoints aren't required to be MAX_ORDER
6931 * aligned but the node_mem_map endpoints must be in order
6932 * for the buddy allocator to function correctly.
6934 end = pgdat_end_pfn(pgdat);
6935 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6936 size = (end - start) * sizeof(struct page);
6937 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6940 panic("Failed to allocate %ld bytes for node %d memory map\n",
6941 size, pgdat->node_id);
6942 pgdat->node_mem_map = map + offset;
6944 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6945 __func__, pgdat->node_id, (unsigned long)pgdat,
6946 (unsigned long)pgdat->node_mem_map);
6947 #ifndef CONFIG_NEED_MULTIPLE_NODES
6949 * With no DISCONTIG, the global mem_map is just set as node 0's
6951 if (pgdat == NODE_DATA(0)) {
6952 mem_map = NODE_DATA(0)->node_mem_map;
6953 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6959 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6960 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6962 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6963 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6965 pgdat->first_deferred_pfn = ULONG_MAX;
6968 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6971 static void __init free_area_init_node(int nid)
6973 pg_data_t *pgdat = NODE_DATA(nid);
6974 unsigned long start_pfn = 0;
6975 unsigned long end_pfn = 0;
6977 /* pg_data_t should be reset to zero when it's allocated */
6978 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6980 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6982 pgdat->node_id = nid;
6983 pgdat->node_start_pfn = start_pfn;
6984 pgdat->per_cpu_nodestats = NULL;
6986 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6987 (u64)start_pfn << PAGE_SHIFT,
6988 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6989 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6991 alloc_node_mem_map(pgdat);
6992 pgdat_set_deferred_range(pgdat);
6994 free_area_init_core(pgdat);
6997 void __init free_area_init_memoryless_node(int nid)
6999 free_area_init_node(nid);
7002 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
7004 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
7005 * PageReserved(). Return the number of struct pages that were initialized.
7007 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
7012 for (pfn = spfn; pfn < epfn; pfn++) {
7013 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
7014 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
7015 + pageblock_nr_pages - 1;
7019 * Use a fake node/zone (0) for now. Some of these pages
7020 * (in memblock.reserved but not in memblock.memory) will
7021 * get re-initialized via reserve_bootmem_region() later.
7023 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7024 __SetPageReserved(pfn_to_page(pfn));
7032 * Only struct pages that are backed by physical memory are zeroed and
7033 * initialized by going through __init_single_page(). But, there are some
7034 * struct pages which are reserved in memblock allocator and their fields
7035 * may be accessed (for example page_to_pfn() on some configuration accesses
7036 * flags). We must explicitly initialize those struct pages.
7038 * This function also addresses a similar issue where struct pages are left
7039 * uninitialized because the physical address range is not covered by
7040 * memblock.memory or memblock.reserved. That could happen when memblock
7041 * layout is manually configured via memmap=, or when the highest physical
7042 * address (max_pfn) does not end on a section boundary.
7044 static void __init init_unavailable_mem(void)
7046 phys_addr_t start, end;
7048 phys_addr_t next = 0;
7051 * Loop through unavailable ranges not covered by memblock.memory.
7054 for_each_mem_range(i, &start, &end) {
7056 pgcnt += init_unavailable_range(PFN_DOWN(next),
7062 * Early sections always have a fully populated memmap for the whole
7063 * section - see pfn_valid(). If the last section has holes at the
7064 * end and that section is marked "online", the memmap will be
7065 * considered initialized. Make sure that memmap has a well defined
7068 pgcnt += init_unavailable_range(PFN_DOWN(next),
7069 round_up(max_pfn, PAGES_PER_SECTION));
7072 * Struct pages that do not have backing memory. This could be because
7073 * firmware is using some of this memory, or for some other reasons.
7076 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7079 static inline void __init init_unavailable_mem(void)
7082 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7084 #if MAX_NUMNODES > 1
7086 * Figure out the number of possible node ids.
7088 void __init setup_nr_node_ids(void)
7090 unsigned int highest;
7092 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7093 nr_node_ids = highest + 1;
7098 * node_map_pfn_alignment - determine the maximum internode alignment
7100 * This function should be called after node map is populated and sorted.
7101 * It calculates the maximum power of two alignment which can distinguish
7104 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7105 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7106 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7107 * shifted, 1GiB is enough and this function will indicate so.
7109 * This is used to test whether pfn -> nid mapping of the chosen memory
7110 * model has fine enough granularity to avoid incorrect mapping for the
7111 * populated node map.
7113 * Return: the determined alignment in pfn's. 0 if there is no alignment
7114 * requirement (single node).
7116 unsigned long __init node_map_pfn_alignment(void)
7118 unsigned long accl_mask = 0, last_end = 0;
7119 unsigned long start, end, mask;
7120 int last_nid = NUMA_NO_NODE;
7123 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7124 if (!start || last_nid < 0 || last_nid == nid) {
7131 * Start with a mask granular enough to pin-point to the
7132 * start pfn and tick off bits one-by-one until it becomes
7133 * too coarse to separate the current node from the last.
7135 mask = ~((1 << __ffs(start)) - 1);
7136 while (mask && last_end <= (start & (mask << 1)))
7139 /* accumulate all internode masks */
7143 /* convert mask to number of pages */
7144 return ~accl_mask + 1;
7148 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7150 * Return: the minimum PFN based on information provided via
7151 * memblock_set_node().
7153 unsigned long __init find_min_pfn_with_active_regions(void)
7155 return PHYS_PFN(memblock_start_of_DRAM());
7159 * early_calculate_totalpages()
7160 * Sum pages in active regions for movable zone.
7161 * Populate N_MEMORY for calculating usable_nodes.
7163 static unsigned long __init early_calculate_totalpages(void)
7165 unsigned long totalpages = 0;
7166 unsigned long start_pfn, end_pfn;
7169 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7170 unsigned long pages = end_pfn - start_pfn;
7172 totalpages += pages;
7174 node_set_state(nid, N_MEMORY);
7180 * Find the PFN the Movable zone begins in each node. Kernel memory
7181 * is spread evenly between nodes as long as the nodes have enough
7182 * memory. When they don't, some nodes will have more kernelcore than
7185 static void __init find_zone_movable_pfns_for_nodes(void)
7188 unsigned long usable_startpfn;
7189 unsigned long kernelcore_node, kernelcore_remaining;
7190 /* save the state before borrow the nodemask */
7191 nodemask_t saved_node_state = node_states[N_MEMORY];
7192 unsigned long totalpages = early_calculate_totalpages();
7193 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7194 struct memblock_region *r;
7196 /* Need to find movable_zone earlier when movable_node is specified. */
7197 find_usable_zone_for_movable();
7200 * If movable_node is specified, ignore kernelcore and movablecore
7203 if (movable_node_is_enabled()) {
7204 for_each_mem_region(r) {
7205 if (!memblock_is_hotpluggable(r))
7208 nid = memblock_get_region_node(r);
7210 usable_startpfn = PFN_DOWN(r->base);
7211 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7212 min(usable_startpfn, zone_movable_pfn[nid]) :
7220 * If kernelcore=mirror is specified, ignore movablecore option
7222 if (mirrored_kernelcore) {
7223 bool mem_below_4gb_not_mirrored = false;
7225 for_each_mem_region(r) {
7226 if (memblock_is_mirror(r))
7229 nid = memblock_get_region_node(r);
7231 usable_startpfn = memblock_region_memory_base_pfn(r);
7233 if (usable_startpfn < 0x100000) {
7234 mem_below_4gb_not_mirrored = true;
7238 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7239 min(usable_startpfn, zone_movable_pfn[nid]) :
7243 if (mem_below_4gb_not_mirrored)
7244 pr_warn("This configuration results in unmirrored kernel memory.\n");
7250 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7251 * amount of necessary memory.
7253 if (required_kernelcore_percent)
7254 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7256 if (required_movablecore_percent)
7257 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7261 * If movablecore= was specified, calculate what size of
7262 * kernelcore that corresponds so that memory usable for
7263 * any allocation type is evenly spread. If both kernelcore
7264 * and movablecore are specified, then the value of kernelcore
7265 * will be used for required_kernelcore if it's greater than
7266 * what movablecore would have allowed.
7268 if (required_movablecore) {
7269 unsigned long corepages;
7272 * Round-up so that ZONE_MOVABLE is at least as large as what
7273 * was requested by the user
7275 required_movablecore =
7276 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7277 required_movablecore = min(totalpages, required_movablecore);
7278 corepages = totalpages - required_movablecore;
7280 required_kernelcore = max(required_kernelcore, corepages);
7284 * If kernelcore was not specified or kernelcore size is larger
7285 * than totalpages, there is no ZONE_MOVABLE.
7287 if (!required_kernelcore || required_kernelcore >= totalpages)
7290 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7291 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7294 /* Spread kernelcore memory as evenly as possible throughout nodes */
7295 kernelcore_node = required_kernelcore / usable_nodes;
7296 for_each_node_state(nid, N_MEMORY) {
7297 unsigned long start_pfn, end_pfn;
7300 * Recalculate kernelcore_node if the division per node
7301 * now exceeds what is necessary to satisfy the requested
7302 * amount of memory for the kernel
7304 if (required_kernelcore < kernelcore_node)
7305 kernelcore_node = required_kernelcore / usable_nodes;
7308 * As the map is walked, we track how much memory is usable
7309 * by the kernel using kernelcore_remaining. When it is
7310 * 0, the rest of the node is usable by ZONE_MOVABLE
7312 kernelcore_remaining = kernelcore_node;
7314 /* Go through each range of PFNs within this node */
7315 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7316 unsigned long size_pages;
7318 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7319 if (start_pfn >= end_pfn)
7322 /* Account for what is only usable for kernelcore */
7323 if (start_pfn < usable_startpfn) {
7324 unsigned long kernel_pages;
7325 kernel_pages = min(end_pfn, usable_startpfn)
7328 kernelcore_remaining -= min(kernel_pages,
7329 kernelcore_remaining);
7330 required_kernelcore -= min(kernel_pages,
7331 required_kernelcore);
7333 /* Continue if range is now fully accounted */
7334 if (end_pfn <= usable_startpfn) {
7337 * Push zone_movable_pfn to the end so
7338 * that if we have to rebalance
7339 * kernelcore across nodes, we will
7340 * not double account here
7342 zone_movable_pfn[nid] = end_pfn;
7345 start_pfn = usable_startpfn;
7349 * The usable PFN range for ZONE_MOVABLE is from
7350 * start_pfn->end_pfn. Calculate size_pages as the
7351 * number of pages used as kernelcore
7353 size_pages = end_pfn - start_pfn;
7354 if (size_pages > kernelcore_remaining)
7355 size_pages = kernelcore_remaining;
7356 zone_movable_pfn[nid] = start_pfn + size_pages;
7359 * Some kernelcore has been met, update counts and
7360 * break if the kernelcore for this node has been
7363 required_kernelcore -= min(required_kernelcore,
7365 kernelcore_remaining -= size_pages;
7366 if (!kernelcore_remaining)
7372 * If there is still required_kernelcore, we do another pass with one
7373 * less node in the count. This will push zone_movable_pfn[nid] further
7374 * along on the nodes that still have memory until kernelcore is
7378 if (usable_nodes && required_kernelcore > usable_nodes)
7382 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7383 for (nid = 0; nid < MAX_NUMNODES; nid++)
7384 zone_movable_pfn[nid] =
7385 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7388 /* restore the node_state */
7389 node_states[N_MEMORY] = saved_node_state;
7392 /* Any regular or high memory on that node ? */
7393 static void check_for_memory(pg_data_t *pgdat, int nid)
7395 enum zone_type zone_type;
7397 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7398 struct zone *zone = &pgdat->node_zones[zone_type];
7399 if (populated_zone(zone)) {
7400 if (IS_ENABLED(CONFIG_HIGHMEM))
7401 node_set_state(nid, N_HIGH_MEMORY);
7402 if (zone_type <= ZONE_NORMAL)
7403 node_set_state(nid, N_NORMAL_MEMORY);
7410 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7411 * such cases we allow max_zone_pfn sorted in the descending order
7413 bool __weak arch_has_descending_max_zone_pfns(void)
7419 * free_area_init - Initialise all pg_data_t and zone data
7420 * @max_zone_pfn: an array of max PFNs for each zone
7422 * This will call free_area_init_node() for each active node in the system.
7423 * Using the page ranges provided by memblock_set_node(), the size of each
7424 * zone in each node and their holes is calculated. If the maximum PFN
7425 * between two adjacent zones match, it is assumed that the zone is empty.
7426 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7427 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7428 * starts where the previous one ended. For example, ZONE_DMA32 starts
7429 * at arch_max_dma_pfn.
7431 void __init free_area_init(unsigned long *max_zone_pfn)
7433 unsigned long start_pfn, end_pfn;
7437 /* Record where the zone boundaries are */
7438 memset(arch_zone_lowest_possible_pfn, 0,
7439 sizeof(arch_zone_lowest_possible_pfn));
7440 memset(arch_zone_highest_possible_pfn, 0,
7441 sizeof(arch_zone_highest_possible_pfn));
7443 start_pfn = find_min_pfn_with_active_regions();
7444 descending = arch_has_descending_max_zone_pfns();
7446 for (i = 0; i < MAX_NR_ZONES; i++) {
7448 zone = MAX_NR_ZONES - i - 1;
7452 if (zone == ZONE_MOVABLE)
7455 end_pfn = max(max_zone_pfn[zone], start_pfn);
7456 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7457 arch_zone_highest_possible_pfn[zone] = end_pfn;
7459 start_pfn = end_pfn;
7462 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7463 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7464 find_zone_movable_pfns_for_nodes();
7466 /* Print out the zone ranges */
7467 pr_info("Zone ranges:\n");
7468 for (i = 0; i < MAX_NR_ZONES; i++) {
7469 if (i == ZONE_MOVABLE)
7471 pr_info(" %-8s ", zone_names[i]);
7472 if (arch_zone_lowest_possible_pfn[i] ==
7473 arch_zone_highest_possible_pfn[i])
7476 pr_cont("[mem %#018Lx-%#018Lx]\n",
7477 (u64)arch_zone_lowest_possible_pfn[i]
7479 ((u64)arch_zone_highest_possible_pfn[i]
7480 << PAGE_SHIFT) - 1);
7483 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7484 pr_info("Movable zone start for each node\n");
7485 for (i = 0; i < MAX_NUMNODES; i++) {
7486 if (zone_movable_pfn[i])
7487 pr_info(" Node %d: %#018Lx\n", i,
7488 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7492 * Print out the early node map, and initialize the
7493 * subsection-map relative to active online memory ranges to
7494 * enable future "sub-section" extensions of the memory map.
7496 pr_info("Early memory node ranges\n");
7497 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7498 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7499 (u64)start_pfn << PAGE_SHIFT,
7500 ((u64)end_pfn << PAGE_SHIFT) - 1);
7501 subsection_map_init(start_pfn, end_pfn - start_pfn);
7504 /* Initialise every node */
7505 mminit_verify_pageflags_layout();
7506 setup_nr_node_ids();
7507 init_unavailable_mem();
7508 for_each_online_node(nid) {
7509 pg_data_t *pgdat = NODE_DATA(nid);
7510 free_area_init_node(nid);
7512 /* Any memory on that node */
7513 if (pgdat->node_present_pages)
7514 node_set_state(nid, N_MEMORY);
7515 check_for_memory(pgdat, nid);
7519 static int __init cmdline_parse_core(char *p, unsigned long *core,
7520 unsigned long *percent)
7522 unsigned long long coremem;
7528 /* Value may be a percentage of total memory, otherwise bytes */
7529 coremem = simple_strtoull(p, &endptr, 0);
7530 if (*endptr == '%') {
7531 /* Paranoid check for percent values greater than 100 */
7532 WARN_ON(coremem > 100);
7536 coremem = memparse(p, &p);
7537 /* Paranoid check that UL is enough for the coremem value */
7538 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7540 *core = coremem >> PAGE_SHIFT;
7547 * kernelcore=size sets the amount of memory for use for allocations that
7548 * cannot be reclaimed or migrated.
7550 static int __init cmdline_parse_kernelcore(char *p)
7552 /* parse kernelcore=mirror */
7553 if (parse_option_str(p, "mirror")) {
7554 mirrored_kernelcore = true;
7558 return cmdline_parse_core(p, &required_kernelcore,
7559 &required_kernelcore_percent);
7563 * movablecore=size sets the amount of memory for use for allocations that
7564 * can be reclaimed or migrated.
7566 static int __init cmdline_parse_movablecore(char *p)
7568 return cmdline_parse_core(p, &required_movablecore,
7569 &required_movablecore_percent);
7572 early_param("kernelcore", cmdline_parse_kernelcore);
7573 early_param("movablecore", cmdline_parse_movablecore);
7575 void adjust_managed_page_count(struct page *page, long count)
7577 atomic_long_add(count, &page_zone(page)->managed_pages);
7578 totalram_pages_add(count);
7579 #ifdef CONFIG_HIGHMEM
7580 if (PageHighMem(page))
7581 totalhigh_pages_add(count);
7584 EXPORT_SYMBOL(adjust_managed_page_count);
7586 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7589 unsigned long pages = 0;
7591 start = (void *)PAGE_ALIGN((unsigned long)start);
7592 end = (void *)((unsigned long)end & PAGE_MASK);
7593 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7594 struct page *page = virt_to_page(pos);
7595 void *direct_map_addr;
7598 * 'direct_map_addr' might be different from 'pos'
7599 * because some architectures' virt_to_page()
7600 * work with aliases. Getting the direct map
7601 * address ensures that we get a _writeable_
7602 * alias for the memset().
7604 direct_map_addr = page_address(page);
7605 if ((unsigned int)poison <= 0xFF)
7606 memset(direct_map_addr, poison, PAGE_SIZE);
7608 free_reserved_page(page);
7612 pr_info("Freeing %s memory: %ldK\n",
7613 s, pages << (PAGE_SHIFT - 10));
7618 #ifdef CONFIG_HIGHMEM
7619 void free_highmem_page(struct page *page)
7621 __free_reserved_page(page);
7622 totalram_pages_inc();
7623 atomic_long_inc(&page_zone(page)->managed_pages);
7624 totalhigh_pages_inc();
7629 void __init mem_init_print_info(const char *str)
7631 unsigned long physpages, codesize, datasize, rosize, bss_size;
7632 unsigned long init_code_size, init_data_size;
7634 physpages = get_num_physpages();
7635 codesize = _etext - _stext;
7636 datasize = _edata - _sdata;
7637 rosize = __end_rodata - __start_rodata;
7638 bss_size = __bss_stop - __bss_start;
7639 init_data_size = __init_end - __init_begin;
7640 init_code_size = _einittext - _sinittext;
7643 * Detect special cases and adjust section sizes accordingly:
7644 * 1) .init.* may be embedded into .data sections
7645 * 2) .init.text.* may be out of [__init_begin, __init_end],
7646 * please refer to arch/tile/kernel/vmlinux.lds.S.
7647 * 3) .rodata.* may be embedded into .text or .data sections.
7649 #define adj_init_size(start, end, size, pos, adj) \
7651 if (start <= pos && pos < end && size > adj) \
7655 adj_init_size(__init_begin, __init_end, init_data_size,
7656 _sinittext, init_code_size);
7657 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7658 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7659 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7660 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7662 #undef adj_init_size
7664 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7665 #ifdef CONFIG_HIGHMEM
7669 nr_free_pages() << (PAGE_SHIFT - 10),
7670 physpages << (PAGE_SHIFT - 10),
7671 codesize >> 10, datasize >> 10, rosize >> 10,
7672 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7673 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7674 totalcma_pages << (PAGE_SHIFT - 10),
7675 #ifdef CONFIG_HIGHMEM
7676 totalhigh_pages() << (PAGE_SHIFT - 10),
7678 str ? ", " : "", str ? str : "");
7682 * set_dma_reserve - set the specified number of pages reserved in the first zone
7683 * @new_dma_reserve: The number of pages to mark reserved
7685 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7686 * In the DMA zone, a significant percentage may be consumed by kernel image
7687 * and other unfreeable allocations which can skew the watermarks badly. This
7688 * function may optionally be used to account for unfreeable pages in the
7689 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7690 * smaller per-cpu batchsize.
7692 void __init set_dma_reserve(unsigned long new_dma_reserve)
7694 dma_reserve = new_dma_reserve;
7697 static int page_alloc_cpu_dead(unsigned int cpu)
7700 lru_add_drain_cpu(cpu);
7704 * Spill the event counters of the dead processor
7705 * into the current processors event counters.
7706 * This artificially elevates the count of the current
7709 vm_events_fold_cpu(cpu);
7712 * Zero the differential counters of the dead processor
7713 * so that the vm statistics are consistent.
7715 * This is only okay since the processor is dead and cannot
7716 * race with what we are doing.
7718 cpu_vm_stats_fold(cpu);
7723 int hashdist = HASHDIST_DEFAULT;
7725 static int __init set_hashdist(char *str)
7729 hashdist = simple_strtoul(str, &str, 0);
7732 __setup("hashdist=", set_hashdist);
7735 void __init page_alloc_init(void)
7740 if (num_node_state(N_MEMORY) == 1)
7744 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7745 "mm/page_alloc:dead", NULL,
7746 page_alloc_cpu_dead);
7751 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7752 * or min_free_kbytes changes.
7754 static void calculate_totalreserve_pages(void)
7756 struct pglist_data *pgdat;
7757 unsigned long reserve_pages = 0;
7758 enum zone_type i, j;
7760 for_each_online_pgdat(pgdat) {
7762 pgdat->totalreserve_pages = 0;
7764 for (i = 0; i < MAX_NR_ZONES; i++) {
7765 struct zone *zone = pgdat->node_zones + i;
7767 unsigned long managed_pages = zone_managed_pages(zone);
7769 /* Find valid and maximum lowmem_reserve in the zone */
7770 for (j = i; j < MAX_NR_ZONES; j++) {
7771 if (zone->lowmem_reserve[j] > max)
7772 max = zone->lowmem_reserve[j];
7775 /* we treat the high watermark as reserved pages. */
7776 max += high_wmark_pages(zone);
7778 if (max > managed_pages)
7779 max = managed_pages;
7781 pgdat->totalreserve_pages += max;
7783 reserve_pages += max;
7786 totalreserve_pages = reserve_pages;
7790 * setup_per_zone_lowmem_reserve - called whenever
7791 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7792 * has a correct pages reserved value, so an adequate number of
7793 * pages are left in the zone after a successful __alloc_pages().
7795 static void setup_per_zone_lowmem_reserve(void)
7797 struct pglist_data *pgdat;
7798 enum zone_type j, idx;
7800 for_each_online_pgdat(pgdat) {
7801 for (j = 0; j < MAX_NR_ZONES; j++) {
7802 struct zone *zone = pgdat->node_zones + j;
7803 unsigned long managed_pages = zone_managed_pages(zone);
7805 zone->lowmem_reserve[j] = 0;
7809 struct zone *lower_zone;
7812 lower_zone = pgdat->node_zones + idx;
7814 if (!sysctl_lowmem_reserve_ratio[idx] ||
7815 !zone_managed_pages(lower_zone)) {
7816 lower_zone->lowmem_reserve[j] = 0;
7819 lower_zone->lowmem_reserve[j] =
7820 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7822 managed_pages += zone_managed_pages(lower_zone);
7827 /* update totalreserve_pages */
7828 calculate_totalreserve_pages();
7831 static void __setup_per_zone_wmarks(void)
7833 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7834 unsigned long lowmem_pages = 0;
7836 unsigned long flags;
7838 /* Calculate total number of !ZONE_HIGHMEM pages */
7839 for_each_zone(zone) {
7840 if (!is_highmem(zone))
7841 lowmem_pages += zone_managed_pages(zone);
7844 for_each_zone(zone) {
7847 spin_lock_irqsave(&zone->lock, flags);
7848 tmp = (u64)pages_min * zone_managed_pages(zone);
7849 do_div(tmp, lowmem_pages);
7850 if (is_highmem(zone)) {
7852 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7853 * need highmem pages, so cap pages_min to a small
7856 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7857 * deltas control async page reclaim, and so should
7858 * not be capped for highmem.
7860 unsigned long min_pages;
7862 min_pages = zone_managed_pages(zone) / 1024;
7863 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7864 zone->_watermark[WMARK_MIN] = min_pages;
7867 * If it's a lowmem zone, reserve a number of pages
7868 * proportionate to the zone's size.
7870 zone->_watermark[WMARK_MIN] = tmp;
7874 * Set the kswapd watermarks distance according to the
7875 * scale factor in proportion to available memory, but
7876 * ensure a minimum size on small systems.
7878 tmp = max_t(u64, tmp >> 2,
7879 mult_frac(zone_managed_pages(zone),
7880 watermark_scale_factor, 10000));
7882 zone->watermark_boost = 0;
7883 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7884 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7886 spin_unlock_irqrestore(&zone->lock, flags);
7889 /* update totalreserve_pages */
7890 calculate_totalreserve_pages();
7894 * setup_per_zone_wmarks - called when min_free_kbytes changes
7895 * or when memory is hot-{added|removed}
7897 * Ensures that the watermark[min,low,high] values for each zone are set
7898 * correctly with respect to min_free_kbytes.
7900 void setup_per_zone_wmarks(void)
7902 static DEFINE_SPINLOCK(lock);
7905 __setup_per_zone_wmarks();
7910 * Initialise min_free_kbytes.
7912 * For small machines we want it small (128k min). For large machines
7913 * we want it large (256MB max). But it is not linear, because network
7914 * bandwidth does not increase linearly with machine size. We use
7916 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7917 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7933 int __meminit init_per_zone_wmark_min(void)
7935 unsigned long lowmem_kbytes;
7936 int new_min_free_kbytes;
7938 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7939 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7941 if (new_min_free_kbytes > user_min_free_kbytes) {
7942 min_free_kbytes = new_min_free_kbytes;
7943 if (min_free_kbytes < 128)
7944 min_free_kbytes = 128;
7945 if (min_free_kbytes > 262144)
7946 min_free_kbytes = 262144;
7948 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7949 new_min_free_kbytes, user_min_free_kbytes);
7951 setup_per_zone_wmarks();
7952 refresh_zone_stat_thresholds();
7953 setup_per_zone_lowmem_reserve();
7956 setup_min_unmapped_ratio();
7957 setup_min_slab_ratio();
7960 khugepaged_min_free_kbytes_update();
7964 postcore_initcall(init_per_zone_wmark_min)
7967 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7968 * that we can call two helper functions whenever min_free_kbytes
7971 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7972 void *buffer, size_t *length, loff_t *ppos)
7976 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7981 user_min_free_kbytes = min_free_kbytes;
7982 setup_per_zone_wmarks();
7987 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7988 void *buffer, size_t *length, loff_t *ppos)
7992 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7997 setup_per_zone_wmarks();
8003 static void setup_min_unmapped_ratio(void)
8008 for_each_online_pgdat(pgdat)
8009 pgdat->min_unmapped_pages = 0;
8012 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8013 sysctl_min_unmapped_ratio) / 100;
8017 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8018 void *buffer, size_t *length, loff_t *ppos)
8022 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8026 setup_min_unmapped_ratio();
8031 static void setup_min_slab_ratio(void)
8036 for_each_online_pgdat(pgdat)
8037 pgdat->min_slab_pages = 0;
8040 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8041 sysctl_min_slab_ratio) / 100;
8044 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8045 void *buffer, size_t *length, loff_t *ppos)
8049 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8053 setup_min_slab_ratio();
8060 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8061 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8062 * whenever sysctl_lowmem_reserve_ratio changes.
8064 * The reserve ratio obviously has absolutely no relation with the
8065 * minimum watermarks. The lowmem reserve ratio can only make sense
8066 * if in function of the boot time zone sizes.
8068 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8069 void *buffer, size_t *length, loff_t *ppos)
8073 proc_dointvec_minmax(table, write, buffer, length, ppos);
8075 for (i = 0; i < MAX_NR_ZONES; i++) {
8076 if (sysctl_lowmem_reserve_ratio[i] < 1)
8077 sysctl_lowmem_reserve_ratio[i] = 0;
8080 setup_per_zone_lowmem_reserve();
8084 static void __zone_pcp_update(struct zone *zone)
8088 for_each_possible_cpu(cpu)
8089 pageset_set_high_and_batch(zone,
8090 per_cpu_ptr(zone->pageset, cpu));
8094 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8095 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8096 * pagelist can have before it gets flushed back to buddy allocator.
8098 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8099 void *buffer, size_t *length, loff_t *ppos)
8102 int old_percpu_pagelist_fraction;
8105 mutex_lock(&pcp_batch_high_lock);
8106 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8108 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8109 if (!write || ret < 0)
8112 /* Sanity checking to avoid pcp imbalance */
8113 if (percpu_pagelist_fraction &&
8114 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8115 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8121 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8124 for_each_populated_zone(zone)
8125 __zone_pcp_update(zone);
8127 mutex_unlock(&pcp_batch_high_lock);
8131 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8133 * Returns the number of pages that arch has reserved but
8134 * is not known to alloc_large_system_hash().
8136 static unsigned long __init arch_reserved_kernel_pages(void)
8143 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8144 * machines. As memory size is increased the scale is also increased but at
8145 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8146 * quadruples the scale is increased by one, which means the size of hash table
8147 * only doubles, instead of quadrupling as well.
8148 * Because 32-bit systems cannot have large physical memory, where this scaling
8149 * makes sense, it is disabled on such platforms.
8151 #if __BITS_PER_LONG > 32
8152 #define ADAPT_SCALE_BASE (64ul << 30)
8153 #define ADAPT_SCALE_SHIFT 2
8154 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8158 * allocate a large system hash table from bootmem
8159 * - it is assumed that the hash table must contain an exact power-of-2
8160 * quantity of entries
8161 * - limit is the number of hash buckets, not the total allocation size
8163 void *__init alloc_large_system_hash(const char *tablename,
8164 unsigned long bucketsize,
8165 unsigned long numentries,
8168 unsigned int *_hash_shift,
8169 unsigned int *_hash_mask,
8170 unsigned long low_limit,
8171 unsigned long high_limit)
8173 unsigned long long max = high_limit;
8174 unsigned long log2qty, size;
8179 /* allow the kernel cmdline to have a say */
8181 /* round applicable memory size up to nearest megabyte */
8182 numentries = nr_kernel_pages;
8183 numentries -= arch_reserved_kernel_pages();
8185 /* It isn't necessary when PAGE_SIZE >= 1MB */
8186 if (PAGE_SHIFT < 20)
8187 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8189 #if __BITS_PER_LONG > 32
8191 unsigned long adapt;
8193 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8194 adapt <<= ADAPT_SCALE_SHIFT)
8199 /* limit to 1 bucket per 2^scale bytes of low memory */
8200 if (scale > PAGE_SHIFT)
8201 numentries >>= (scale - PAGE_SHIFT);
8203 numentries <<= (PAGE_SHIFT - scale);
8205 /* Make sure we've got at least a 0-order allocation.. */
8206 if (unlikely(flags & HASH_SMALL)) {
8207 /* Makes no sense without HASH_EARLY */
8208 WARN_ON(!(flags & HASH_EARLY));
8209 if (!(numentries >> *_hash_shift)) {
8210 numentries = 1UL << *_hash_shift;
8211 BUG_ON(!numentries);
8213 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8214 numentries = PAGE_SIZE / bucketsize;
8216 numentries = roundup_pow_of_two(numentries);
8218 /* limit allocation size to 1/16 total memory by default */
8220 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8221 do_div(max, bucketsize);
8223 max = min(max, 0x80000000ULL);
8225 if (numentries < low_limit)
8226 numentries = low_limit;
8227 if (numentries > max)
8230 log2qty = ilog2(numentries);
8232 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8235 size = bucketsize << log2qty;
8236 if (flags & HASH_EARLY) {
8237 if (flags & HASH_ZERO)
8238 table = memblock_alloc(size, SMP_CACHE_BYTES);
8240 table = memblock_alloc_raw(size,
8242 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8243 table = __vmalloc(size, gfp_flags);
8247 * If bucketsize is not a power-of-two, we may free
8248 * some pages at the end of hash table which
8249 * alloc_pages_exact() automatically does
8251 table = alloc_pages_exact(size, gfp_flags);
8252 kmemleak_alloc(table, size, 1, gfp_flags);
8254 } while (!table && size > PAGE_SIZE && --log2qty);
8257 panic("Failed to allocate %s hash table\n", tablename);
8259 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8260 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8261 virt ? "vmalloc" : "linear");
8264 *_hash_shift = log2qty;
8266 *_hash_mask = (1 << log2qty) - 1;
8272 * This function checks whether pageblock includes unmovable pages or not.
8274 * PageLRU check without isolation or lru_lock could race so that
8275 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8276 * check without lock_page also may miss some movable non-lru pages at
8277 * race condition. So you can't expect this function should be exact.
8279 * Returns a page without holding a reference. If the caller wants to
8280 * dereference that page (e.g., dumping), it has to make sure that it
8281 * cannot get removed (e.g., via memory unplug) concurrently.
8284 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8285 int migratetype, int flags)
8287 unsigned long iter = 0;
8288 unsigned long pfn = page_to_pfn(page);
8289 unsigned long offset = pfn % pageblock_nr_pages;
8291 if (is_migrate_cma_page(page)) {
8293 * CMA allocations (alloc_contig_range) really need to mark
8294 * isolate CMA pageblocks even when they are not movable in fact
8295 * so consider them movable here.
8297 if (is_migrate_cma(migratetype))
8303 for (; iter < pageblock_nr_pages - offset; iter++) {
8304 if (!pfn_valid_within(pfn + iter))
8307 page = pfn_to_page(pfn + iter);
8310 * Both, bootmem allocations and memory holes are marked
8311 * PG_reserved and are unmovable. We can even have unmovable
8312 * allocations inside ZONE_MOVABLE, for example when
8313 * specifying "movablecore".
8315 if (PageReserved(page))
8319 * If the zone is movable and we have ruled out all reserved
8320 * pages then it should be reasonably safe to assume the rest
8323 if (zone_idx(zone) == ZONE_MOVABLE)
8327 * Hugepages are not in LRU lists, but they're movable.
8328 * THPs are on the LRU, but need to be counted as #small pages.
8329 * We need not scan over tail pages because we don't
8330 * handle each tail page individually in migration.
8332 if (PageHuge(page) || PageTransCompound(page)) {
8333 struct page *head = compound_head(page);
8334 unsigned int skip_pages;
8336 if (PageHuge(page)) {
8337 if (!hugepage_migration_supported(page_hstate(head)))
8339 } else if (!PageLRU(head) && !__PageMovable(head)) {
8343 skip_pages = compound_nr(head) - (page - head);
8344 iter += skip_pages - 1;
8349 * We can't use page_count without pin a page
8350 * because another CPU can free compound page.
8351 * This check already skips compound tails of THP
8352 * because their page->_refcount is zero at all time.
8354 if (!page_ref_count(page)) {
8355 if (PageBuddy(page))
8356 iter += (1 << buddy_order(page)) - 1;
8361 * The HWPoisoned page may be not in buddy system, and
8362 * page_count() is not 0.
8364 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8368 * We treat all PageOffline() pages as movable when offlining
8369 * to give drivers a chance to decrement their reference count
8370 * in MEM_GOING_OFFLINE in order to indicate that these pages
8371 * can be offlined as there are no direct references anymore.
8372 * For actually unmovable PageOffline() where the driver does
8373 * not support this, we will fail later when trying to actually
8374 * move these pages that still have a reference count > 0.
8375 * (false negatives in this function only)
8377 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8380 if (__PageMovable(page) || PageLRU(page))
8384 * If there are RECLAIMABLE pages, we need to check
8385 * it. But now, memory offline itself doesn't call
8386 * shrink_node_slabs() and it still to be fixed.
8393 #ifdef CONFIG_CONTIG_ALLOC
8394 static unsigned long pfn_max_align_down(unsigned long pfn)
8396 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8397 pageblock_nr_pages) - 1);
8400 static unsigned long pfn_max_align_up(unsigned long pfn)
8402 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8403 pageblock_nr_pages));
8406 /* [start, end) must belong to a single zone. */
8407 static int __alloc_contig_migrate_range(struct compact_control *cc,
8408 unsigned long start, unsigned long end)
8410 /* This function is based on compact_zone() from compaction.c. */
8411 unsigned int nr_reclaimed;
8412 unsigned long pfn = start;
8413 unsigned int tries = 0;
8415 struct migration_target_control mtc = {
8416 .nid = zone_to_nid(cc->zone),
8417 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8422 while (pfn < end || !list_empty(&cc->migratepages)) {
8423 if (fatal_signal_pending(current)) {
8428 if (list_empty(&cc->migratepages)) {
8429 cc->nr_migratepages = 0;
8430 pfn = isolate_migratepages_range(cc, pfn, end);
8436 } else if (++tries == 5) {
8437 ret = ret < 0 ? ret : -EBUSY;
8441 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8443 cc->nr_migratepages -= nr_reclaimed;
8445 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8446 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8449 putback_movable_pages(&cc->migratepages);
8456 * alloc_contig_range() -- tries to allocate given range of pages
8457 * @start: start PFN to allocate
8458 * @end: one-past-the-last PFN to allocate
8459 * @migratetype: migratetype of the underlaying pageblocks (either
8460 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8461 * in range must have the same migratetype and it must
8462 * be either of the two.
8463 * @gfp_mask: GFP mask to use during compaction
8465 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8466 * aligned. The PFN range must belong to a single zone.
8468 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8469 * pageblocks in the range. Once isolated, the pageblocks should not
8470 * be modified by others.
8472 * Return: zero on success or negative error code. On success all
8473 * pages which PFN is in [start, end) are allocated for the caller and
8474 * need to be freed with free_contig_range().
8476 int alloc_contig_range(unsigned long start, unsigned long end,
8477 unsigned migratetype, gfp_t gfp_mask)
8479 unsigned long outer_start, outer_end;
8483 struct compact_control cc = {
8484 .nr_migratepages = 0,
8486 .zone = page_zone(pfn_to_page(start)),
8487 .mode = MIGRATE_SYNC,
8488 .ignore_skip_hint = true,
8489 .no_set_skip_hint = true,
8490 .gfp_mask = current_gfp_context(gfp_mask),
8491 .alloc_contig = true,
8493 INIT_LIST_HEAD(&cc.migratepages);
8496 * What we do here is we mark all pageblocks in range as
8497 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8498 * have different sizes, and due to the way page allocator
8499 * work, we align the range to biggest of the two pages so
8500 * that page allocator won't try to merge buddies from
8501 * different pageblocks and change MIGRATE_ISOLATE to some
8502 * other migration type.
8504 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8505 * migrate the pages from an unaligned range (ie. pages that
8506 * we are interested in). This will put all the pages in
8507 * range back to page allocator as MIGRATE_ISOLATE.
8509 * When this is done, we take the pages in range from page
8510 * allocator removing them from the buddy system. This way
8511 * page allocator will never consider using them.
8513 * This lets us mark the pageblocks back as
8514 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8515 * aligned range but not in the unaligned, original range are
8516 * put back to page allocator so that buddy can use them.
8519 ret = start_isolate_page_range(pfn_max_align_down(start),
8520 pfn_max_align_up(end), migratetype, 0);
8525 * In case of -EBUSY, we'd like to know which page causes problem.
8526 * So, just fall through. test_pages_isolated() has a tracepoint
8527 * which will report the busy page.
8529 * It is possible that busy pages could become available before
8530 * the call to test_pages_isolated, and the range will actually be
8531 * allocated. So, if we fall through be sure to clear ret so that
8532 * -EBUSY is not accidentally used or returned to caller.
8534 ret = __alloc_contig_migrate_range(&cc, start, end);
8535 if (ret && ret != -EBUSY)
8540 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8541 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8542 * more, all pages in [start, end) are free in page allocator.
8543 * What we are going to do is to allocate all pages from
8544 * [start, end) (that is remove them from page allocator).
8546 * The only problem is that pages at the beginning and at the
8547 * end of interesting range may be not aligned with pages that
8548 * page allocator holds, ie. they can be part of higher order
8549 * pages. Because of this, we reserve the bigger range and
8550 * once this is done free the pages we are not interested in.
8552 * We don't have to hold zone->lock here because the pages are
8553 * isolated thus they won't get removed from buddy.
8556 lru_add_drain_all();
8559 outer_start = start;
8560 while (!PageBuddy(pfn_to_page(outer_start))) {
8561 if (++order >= MAX_ORDER) {
8562 outer_start = start;
8565 outer_start &= ~0UL << order;
8568 if (outer_start != start) {
8569 order = buddy_order(pfn_to_page(outer_start));
8572 * outer_start page could be small order buddy page and
8573 * it doesn't include start page. Adjust outer_start
8574 * in this case to report failed page properly
8575 * on tracepoint in test_pages_isolated()
8577 if (outer_start + (1UL << order) <= start)
8578 outer_start = start;
8581 /* Make sure the range is really isolated. */
8582 if (test_pages_isolated(outer_start, end, 0)) {
8587 /* Grab isolated pages from freelists. */
8588 outer_end = isolate_freepages_range(&cc, outer_start, end);
8594 /* Free head and tail (if any) */
8595 if (start != outer_start)
8596 free_contig_range(outer_start, start - outer_start);
8597 if (end != outer_end)
8598 free_contig_range(end, outer_end - end);
8601 undo_isolate_page_range(pfn_max_align_down(start),
8602 pfn_max_align_up(end), migratetype);
8605 EXPORT_SYMBOL(alloc_contig_range);
8607 static int __alloc_contig_pages(unsigned long start_pfn,
8608 unsigned long nr_pages, gfp_t gfp_mask)
8610 unsigned long end_pfn = start_pfn + nr_pages;
8612 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8616 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8617 unsigned long nr_pages)
8619 unsigned long i, end_pfn = start_pfn + nr_pages;
8622 for (i = start_pfn; i < end_pfn; i++) {
8623 page = pfn_to_online_page(i);
8627 if (page_zone(page) != z)
8630 if (PageReserved(page))
8633 if (page_count(page) > 0)
8642 static bool zone_spans_last_pfn(const struct zone *zone,
8643 unsigned long start_pfn, unsigned long nr_pages)
8645 unsigned long last_pfn = start_pfn + nr_pages - 1;
8647 return zone_spans_pfn(zone, last_pfn);
8651 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8652 * @nr_pages: Number of contiguous pages to allocate
8653 * @gfp_mask: GFP mask to limit search and used during compaction
8655 * @nodemask: Mask for other possible nodes
8657 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8658 * on an applicable zonelist to find a contiguous pfn range which can then be
8659 * tried for allocation with alloc_contig_range(). This routine is intended
8660 * for allocation requests which can not be fulfilled with the buddy allocator.
8662 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8663 * power of two then the alignment is guaranteed to be to the given nr_pages
8664 * (e.g. 1GB request would be aligned to 1GB).
8666 * Allocated pages can be freed with free_contig_range() or by manually calling
8667 * __free_page() on each allocated page.
8669 * Return: pointer to contiguous pages on success, or NULL if not successful.
8671 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8672 int nid, nodemask_t *nodemask)
8674 unsigned long ret, pfn, flags;
8675 struct zonelist *zonelist;
8679 zonelist = node_zonelist(nid, gfp_mask);
8680 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8681 gfp_zone(gfp_mask), nodemask) {
8682 spin_lock_irqsave(&zone->lock, flags);
8684 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8685 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8686 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8688 * We release the zone lock here because
8689 * alloc_contig_range() will also lock the zone
8690 * at some point. If there's an allocation
8691 * spinning on this lock, it may win the race
8692 * and cause alloc_contig_range() to fail...
8694 spin_unlock_irqrestore(&zone->lock, flags);
8695 ret = __alloc_contig_pages(pfn, nr_pages,
8698 return pfn_to_page(pfn);
8699 spin_lock_irqsave(&zone->lock, flags);
8703 spin_unlock_irqrestore(&zone->lock, flags);
8707 #endif /* CONFIG_CONTIG_ALLOC */
8709 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8711 unsigned int count = 0;
8713 for (; nr_pages--; pfn++) {
8714 struct page *page = pfn_to_page(pfn);
8716 count += page_count(page) != 1;
8719 WARN(count != 0, "%d pages are still in use!\n", count);
8721 EXPORT_SYMBOL(free_contig_range);
8724 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8725 * page high values need to be recalulated.
8727 void __meminit zone_pcp_update(struct zone *zone)
8729 mutex_lock(&pcp_batch_high_lock);
8730 __zone_pcp_update(zone);
8731 mutex_unlock(&pcp_batch_high_lock);
8734 void zone_pcp_reset(struct zone *zone)
8736 unsigned long flags;
8738 struct per_cpu_pageset *pset;
8740 /* avoid races with drain_pages() */
8741 local_irq_save(flags);
8742 if (zone->pageset != &boot_pageset) {
8743 for_each_online_cpu(cpu) {
8744 pset = per_cpu_ptr(zone->pageset, cpu);
8745 drain_zonestat(zone, pset);
8747 free_percpu(zone->pageset);
8748 zone->pageset = &boot_pageset;
8750 local_irq_restore(flags);
8753 #ifdef CONFIG_MEMORY_HOTREMOVE
8755 * All pages in the range must be in a single zone, must not contain holes,
8756 * must span full sections, and must be isolated before calling this function.
8758 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8760 unsigned long pfn = start_pfn;
8764 unsigned long flags;
8766 offline_mem_sections(pfn, end_pfn);
8767 zone = page_zone(pfn_to_page(pfn));
8768 spin_lock_irqsave(&zone->lock, flags);
8769 while (pfn < end_pfn) {
8770 page = pfn_to_page(pfn);
8772 * The HWPoisoned page may be not in buddy system, and
8773 * page_count() is not 0.
8775 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8780 * At this point all remaining PageOffline() pages have a
8781 * reference count of 0 and can simply be skipped.
8783 if (PageOffline(page)) {
8784 BUG_ON(page_count(page));
8785 BUG_ON(PageBuddy(page));
8790 BUG_ON(page_count(page));
8791 BUG_ON(!PageBuddy(page));
8792 order = buddy_order(page);
8793 del_page_from_free_list(page, zone, order);
8794 pfn += (1 << order);
8796 spin_unlock_irqrestore(&zone->lock, flags);
8800 bool is_free_buddy_page(struct page *page)
8802 struct zone *zone = page_zone(page);
8803 unsigned long pfn = page_to_pfn(page);
8804 unsigned long flags;
8807 spin_lock_irqsave(&zone->lock, flags);
8808 for (order = 0; order < MAX_ORDER; order++) {
8809 struct page *page_head = page - (pfn & ((1 << order) - 1));
8811 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
8814 spin_unlock_irqrestore(&zone->lock, flags);
8816 return order < MAX_ORDER;
8819 #ifdef CONFIG_MEMORY_FAILURE
8821 * Break down a higher-order page in sub-pages, and keep our target out of
8824 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8825 struct page *target, int low, int high,
8828 unsigned long size = 1 << high;
8829 struct page *current_buddy, *next_page;
8831 while (high > low) {
8835 if (target >= &page[size]) {
8836 next_page = page + size;
8837 current_buddy = page;
8840 current_buddy = page + size;
8843 if (set_page_guard(zone, current_buddy, high, migratetype))
8846 if (current_buddy != target) {
8847 add_to_free_list(current_buddy, zone, high, migratetype);
8848 set_buddy_order(current_buddy, high);
8855 * Take a page that will be marked as poisoned off the buddy allocator.
8857 bool take_page_off_buddy(struct page *page)
8859 struct zone *zone = page_zone(page);
8860 unsigned long pfn = page_to_pfn(page);
8861 unsigned long flags;
8865 spin_lock_irqsave(&zone->lock, flags);
8866 for (order = 0; order < MAX_ORDER; order++) {
8867 struct page *page_head = page - (pfn & ((1 << order) - 1));
8868 int page_order = buddy_order(page_head);
8870 if (PageBuddy(page_head) && page_order >= order) {
8871 unsigned long pfn_head = page_to_pfn(page_head);
8872 int migratetype = get_pfnblock_migratetype(page_head,
8875 del_page_from_free_list(page_head, zone, page_order);
8876 break_down_buddy_pages(zone, page_head, page, 0,
8877 page_order, migratetype);
8881 if (page_count(page_head) > 0)
8884 spin_unlock_irqrestore(&zone->lock, flags);