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/local_lock.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
71 #include <linux/psi.h>
72 #include <linux/padata.h>
73 #include <linux/khugepaged.h>
75 #include <asm/sections.h>
76 #include <asm/tlbflush.h>
77 #include <asm/div64.h>
80 #include "page_reporting.h"
82 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
83 typedef int __bitwise fpi_t;
85 /* No special request */
86 #define FPI_NONE ((__force fpi_t)0)
89 * Skip free page reporting notification for the (possibly merged) page.
90 * This does not hinder free page reporting from grabbing the page,
91 * reporting it and marking it "reported" - it only skips notifying
92 * the free page reporting infrastructure about a newly freed page. For
93 * example, used when temporarily pulling a page from a freelist and
94 * putting it back unmodified.
96 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
99 * Place the (possibly merged) page to the tail of the freelist. Will ignore
100 * page shuffling (relevant code - e.g., memory onlining - is expected to
101 * shuffle the whole zone).
103 * Note: No code should rely on this flag for correctness - it's purely
104 * to allow for optimizations when handing back either fresh pages
105 * (memory onlining) or untouched pages (page isolation, free page
108 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
110 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
111 static DEFINE_MUTEX(pcp_batch_high_lock);
112 #define MIN_PERCPU_PAGELIST_FRACTION (8)
114 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
115 DEFINE_PER_CPU(int, numa_node);
116 EXPORT_PER_CPU_SYMBOL(numa_node);
119 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
121 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
123 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
124 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
125 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
126 * defined in <linux/topology.h>.
128 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
129 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
132 /* work_structs for global per-cpu drains */
135 struct work_struct work;
137 static DEFINE_MUTEX(pcpu_drain_mutex);
138 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
140 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
141 volatile unsigned long latent_entropy __latent_entropy;
142 EXPORT_SYMBOL(latent_entropy);
146 * Array of node states.
148 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
149 [N_POSSIBLE] = NODE_MASK_ALL,
150 [N_ONLINE] = { { [0] = 1UL } },
152 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
153 #ifdef CONFIG_HIGHMEM
154 [N_HIGH_MEMORY] = { { [0] = 1UL } },
156 [N_MEMORY] = { { [0] = 1UL } },
157 [N_CPU] = { { [0] = 1UL } },
160 EXPORT_SYMBOL(node_states);
162 atomic_long_t _totalram_pages __read_mostly;
163 EXPORT_SYMBOL(_totalram_pages);
164 unsigned long totalreserve_pages __read_mostly;
165 unsigned long totalcma_pages __read_mostly;
167 int percpu_pagelist_fraction;
168 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
169 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
170 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
172 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
174 EXPORT_SYMBOL(init_on_alloc);
176 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
177 DEFINE_STATIC_KEY_TRUE(init_on_free);
179 DEFINE_STATIC_KEY_FALSE(init_on_free);
181 EXPORT_SYMBOL(init_on_free);
183 static int __init early_init_on_alloc(char *buf)
188 ret = kstrtobool(buf, &bool_result);
191 if (bool_result && page_poisoning_enabled())
192 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
194 static_branch_enable(&init_on_alloc);
196 static_branch_disable(&init_on_alloc);
199 early_param("init_on_alloc", early_init_on_alloc);
201 static int __init early_init_on_free(char *buf)
206 ret = kstrtobool(buf, &bool_result);
209 if (bool_result && page_poisoning_enabled())
210 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
212 static_branch_enable(&init_on_free);
214 static_branch_disable(&init_on_free);
217 early_param("init_on_free", early_init_on_free);
220 * A cached value of the page's pageblock's migratetype, used when the page is
221 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
222 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
223 * Also the migratetype set in the page does not necessarily match the pcplist
224 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
225 * other index - this ensures that it will be put on the correct CMA freelist.
227 static inline int get_pcppage_migratetype(struct page *page)
232 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
234 page->index = migratetype;
237 #ifdef CONFIG_PM_SLEEP
239 * The following functions are used by the suspend/hibernate code to temporarily
240 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
241 * while devices are suspended. To avoid races with the suspend/hibernate code,
242 * they should always be called with system_transition_mutex held
243 * (gfp_allowed_mask also should only be modified with system_transition_mutex
244 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
245 * with that modification).
248 static gfp_t saved_gfp_mask;
250 void pm_restore_gfp_mask(void)
252 WARN_ON(!mutex_is_locked(&system_transition_mutex));
253 if (saved_gfp_mask) {
254 gfp_allowed_mask = saved_gfp_mask;
259 void pm_restrict_gfp_mask(void)
261 WARN_ON(!mutex_is_locked(&system_transition_mutex));
262 WARN_ON(saved_gfp_mask);
263 saved_gfp_mask = gfp_allowed_mask;
264 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
267 bool pm_suspended_storage(void)
269 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
273 #endif /* CONFIG_PM_SLEEP */
275 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
276 unsigned int pageblock_order __read_mostly;
279 static void __free_pages_ok(struct page *page, unsigned int order,
283 * results with 256, 32 in the lowmem_reserve sysctl:
284 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
285 * 1G machine -> (16M dma, 784M normal, 224M high)
286 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
287 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
288 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
290 * TBD: should special case ZONE_DMA32 machines here - in those we normally
291 * don't need any ZONE_NORMAL reservation
293 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
294 #ifdef CONFIG_ZONE_DMA
297 #ifdef CONFIG_ZONE_DMA32
301 #ifdef CONFIG_HIGHMEM
307 static char * const zone_names[MAX_NR_ZONES] = {
308 #ifdef CONFIG_ZONE_DMA
311 #ifdef CONFIG_ZONE_DMA32
315 #ifdef CONFIG_HIGHMEM
319 #ifdef CONFIG_ZONE_DEVICE
324 const char * const migratetype_names[MIGRATE_TYPES] = {
332 #ifdef CONFIG_MEMORY_ISOLATION
337 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
338 [NULL_COMPOUND_DTOR] = NULL,
339 [COMPOUND_PAGE_DTOR] = free_compound_page,
340 #ifdef CONFIG_HUGETLB_PAGE
341 [HUGETLB_PAGE_DTOR] = free_huge_page,
343 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
344 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
348 int min_free_kbytes = 1024;
349 int user_min_free_kbytes = -1;
350 #ifdef CONFIG_DISCONTIGMEM
352 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
353 * are not on separate NUMA nodes. Functionally this works but with
354 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
355 * quite small. By default, do not boost watermarks on discontigmem as in
356 * many cases very high-order allocations like THP are likely to be
357 * unsupported and the premature reclaim offsets the advantage of long-term
358 * fragmentation avoidance.
360 int watermark_boost_factor __read_mostly;
362 int watermark_boost_factor __read_mostly = 15000;
364 int watermark_scale_factor = 10;
366 static unsigned long nr_kernel_pages __initdata;
367 static unsigned long nr_all_pages __initdata;
368 static unsigned long dma_reserve __initdata;
370 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
371 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
372 static unsigned long required_kernelcore __initdata;
373 static unsigned long required_kernelcore_percent __initdata;
374 static unsigned long required_movablecore __initdata;
375 static unsigned long required_movablecore_percent __initdata;
376 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
377 static bool mirrored_kernelcore __meminitdata;
379 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
381 EXPORT_SYMBOL(movable_zone);
384 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
385 unsigned int nr_online_nodes __read_mostly = 1;
386 EXPORT_SYMBOL(nr_node_ids);
387 EXPORT_SYMBOL(nr_online_nodes);
393 static DEFINE_PER_CPU(struct pa_lock, pa_lock) = {
394 .l = INIT_LOCAL_LOCK(l),
397 int page_group_by_mobility_disabled __read_mostly;
399 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
401 * During boot we initialize deferred pages on-demand, as needed, but once
402 * page_alloc_init_late() has finished, the deferred pages are all initialized,
403 * and we can permanently disable that path.
405 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
408 * Calling kasan_free_pages() only after deferred memory initialization
409 * has completed. Poisoning pages during deferred memory init will greatly
410 * lengthen the process and cause problem in large memory systems as the
411 * deferred pages initialization is done with interrupt disabled.
413 * Assuming that there will be no reference to those newly initialized
414 * pages before they are ever allocated, this should have no effect on
415 * KASAN memory tracking as the poison will be properly inserted at page
416 * allocation time. The only corner case is when pages are allocated by
417 * on-demand allocation and then freed again before the deferred pages
418 * initialization is done, but this is not likely to happen.
420 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
422 if (!static_branch_unlikely(&deferred_pages))
423 kasan_free_pages(page, order);
426 /* Returns true if the struct page for the pfn is uninitialised */
427 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
429 int nid = early_pfn_to_nid(pfn);
431 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
438 * Returns true when the remaining initialisation should be deferred until
439 * later in the boot cycle when it can be parallelised.
441 static bool __meminit
442 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
444 static unsigned long prev_end_pfn, nr_initialised;
447 * prev_end_pfn static that contains the end of previous zone
448 * No need to protect because called very early in boot before smp_init.
450 if (prev_end_pfn != end_pfn) {
451 prev_end_pfn = end_pfn;
455 /* Always populate low zones for address-constrained allocations */
456 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
459 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
462 * We start only with one section of pages, more pages are added as
463 * needed until the rest of deferred pages are initialized.
466 if ((nr_initialised > PAGES_PER_SECTION) &&
467 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
468 NODE_DATA(nid)->first_deferred_pfn = pfn;
474 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
476 static inline bool early_page_uninitialised(unsigned long pfn)
481 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
487 /* Return a pointer to the bitmap storing bits affecting a block of pages */
488 static inline unsigned long *get_pageblock_bitmap(struct page *page,
491 #ifdef CONFIG_SPARSEMEM
492 return section_to_usemap(__pfn_to_section(pfn));
494 return page_zone(page)->pageblock_flags;
495 #endif /* CONFIG_SPARSEMEM */
498 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
500 #ifdef CONFIG_SPARSEMEM
501 pfn &= (PAGES_PER_SECTION-1);
503 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
504 #endif /* CONFIG_SPARSEMEM */
505 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
509 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
510 * @page: The page within the block of interest
511 * @pfn: The target page frame number
512 * @mask: mask of bits that the caller is interested in
514 * Return: pageblock_bits flags
516 static __always_inline
517 unsigned long __get_pfnblock_flags_mask(struct page *page,
521 unsigned long *bitmap;
522 unsigned long bitidx, word_bitidx;
525 bitmap = get_pageblock_bitmap(page, pfn);
526 bitidx = pfn_to_bitidx(page, pfn);
527 word_bitidx = bitidx / BITS_PER_LONG;
528 bitidx &= (BITS_PER_LONG-1);
530 word = bitmap[word_bitidx];
531 return (word >> bitidx) & mask;
534 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
537 return __get_pfnblock_flags_mask(page, pfn, mask);
540 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
542 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
546 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
547 * @page: The page within the block of interest
548 * @flags: The flags to set
549 * @pfn: The target page frame number
550 * @mask: mask of bits that the caller is interested in
552 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
556 unsigned long *bitmap;
557 unsigned long bitidx, word_bitidx;
558 unsigned long old_word, word;
560 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
561 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
563 bitmap = get_pageblock_bitmap(page, pfn);
564 bitidx = pfn_to_bitidx(page, pfn);
565 word_bitidx = bitidx / BITS_PER_LONG;
566 bitidx &= (BITS_PER_LONG-1);
568 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
573 word = READ_ONCE(bitmap[word_bitidx]);
575 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
576 if (word == old_word)
582 void set_pageblock_migratetype(struct page *page, int migratetype)
584 if (unlikely(page_group_by_mobility_disabled &&
585 migratetype < MIGRATE_PCPTYPES))
586 migratetype = MIGRATE_UNMOVABLE;
588 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
589 page_to_pfn(page), MIGRATETYPE_MASK);
592 #ifdef CONFIG_DEBUG_VM
593 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
597 unsigned long pfn = page_to_pfn(page);
598 unsigned long sp, start_pfn;
601 seq = zone_span_seqbegin(zone);
602 start_pfn = zone->zone_start_pfn;
603 sp = zone->spanned_pages;
604 if (!zone_spans_pfn(zone, pfn))
606 } while (zone_span_seqretry(zone, seq));
609 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
610 pfn, zone_to_nid(zone), zone->name,
611 start_pfn, start_pfn + sp);
616 static int page_is_consistent(struct zone *zone, struct page *page)
618 if (!pfn_valid_within(page_to_pfn(page)))
620 if (zone != page_zone(page))
626 * Temporary debugging check for pages not lying within a given zone.
628 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
630 if (page_outside_zone_boundaries(zone, page))
632 if (!page_is_consistent(zone, page))
638 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
644 static void bad_page(struct page *page, const char *reason)
646 static unsigned long resume;
647 static unsigned long nr_shown;
648 static unsigned long nr_unshown;
651 * Allow a burst of 60 reports, then keep quiet for that minute;
652 * or allow a steady drip of one report per second.
654 if (nr_shown == 60) {
655 if (time_before(jiffies, resume)) {
661 "BUG: Bad page state: %lu messages suppressed\n",
668 resume = jiffies + 60 * HZ;
670 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
671 current->comm, page_to_pfn(page));
672 __dump_page(page, reason);
673 dump_page_owner(page);
678 /* Leave bad fields for debug, except PageBuddy could make trouble */
679 page_mapcount_reset(page); /* remove PageBuddy */
680 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
684 * Higher-order pages are called "compound pages". They are structured thusly:
686 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
688 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
689 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
691 * The first tail page's ->compound_dtor holds the offset in array of compound
692 * page destructors. See compound_page_dtors.
694 * The first tail page's ->compound_order holds the order of allocation.
695 * This usage means that zero-order pages may not be compound.
698 void free_compound_page(struct page *page)
700 mem_cgroup_uncharge(page);
701 __free_pages_ok(page, compound_order(page), FPI_NONE);
704 void prep_compound_page(struct page *page, unsigned int order)
707 int nr_pages = 1 << order;
710 for (i = 1; i < nr_pages; i++) {
711 struct page *p = page + i;
712 set_page_count(p, 0);
713 p->mapping = TAIL_MAPPING;
714 set_compound_head(p, page);
717 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
718 set_compound_order(page, order);
719 atomic_set(compound_mapcount_ptr(page), -1);
720 if (hpage_pincount_available(page))
721 atomic_set(compound_pincount_ptr(page), 0);
724 #ifdef CONFIG_DEBUG_PAGEALLOC
725 unsigned int _debug_guardpage_minorder;
727 bool _debug_pagealloc_enabled_early __read_mostly
728 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
729 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
730 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
731 EXPORT_SYMBOL(_debug_pagealloc_enabled);
733 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
735 static int __init early_debug_pagealloc(char *buf)
737 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
739 early_param("debug_pagealloc", early_debug_pagealloc);
741 void init_debug_pagealloc(void)
743 if (!debug_pagealloc_enabled())
746 static_branch_enable(&_debug_pagealloc_enabled);
748 if (!debug_guardpage_minorder())
751 static_branch_enable(&_debug_guardpage_enabled);
754 static int __init debug_guardpage_minorder_setup(char *buf)
758 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
759 pr_err("Bad debug_guardpage_minorder value\n");
762 _debug_guardpage_minorder = res;
763 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
766 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
768 static inline bool set_page_guard(struct zone *zone, struct page *page,
769 unsigned int order, int migratetype)
771 if (!debug_guardpage_enabled())
774 if (order >= debug_guardpage_minorder())
777 __SetPageGuard(page);
778 INIT_LIST_HEAD(&page->lru);
779 set_page_private(page, order);
780 /* Guard pages are not available for any usage */
781 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
786 static inline void clear_page_guard(struct zone *zone, struct page *page,
787 unsigned int order, int migratetype)
789 if (!debug_guardpage_enabled())
792 __ClearPageGuard(page);
794 set_page_private(page, 0);
795 if (!is_migrate_isolate(migratetype))
796 __mod_zone_freepage_state(zone, (1 << order), migratetype);
799 static inline bool set_page_guard(struct zone *zone, struct page *page,
800 unsigned int order, int migratetype) { return false; }
801 static inline void clear_page_guard(struct zone *zone, struct page *page,
802 unsigned int order, int migratetype) {}
805 static inline void set_buddy_order(struct page *page, unsigned int order)
807 set_page_private(page, order);
808 __SetPageBuddy(page);
812 * This function checks whether a page is free && is the buddy
813 * we can coalesce a page and its buddy if
814 * (a) the buddy is not in a hole (check before calling!) &&
815 * (b) the buddy is in the buddy system &&
816 * (c) a page and its buddy have the same order &&
817 * (d) a page and its buddy are in the same zone.
819 * For recording whether a page is in the buddy system, we set PageBuddy.
820 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
822 * For recording page's order, we use page_private(page).
824 static inline bool page_is_buddy(struct page *page, struct page *buddy,
827 if (!page_is_guard(buddy) && !PageBuddy(buddy))
830 if (buddy_order(buddy) != order)
834 * zone check is done late to avoid uselessly calculating
835 * zone/node ids for pages that could never merge.
837 if (page_zone_id(page) != page_zone_id(buddy))
840 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
845 #ifdef CONFIG_COMPACTION
846 static inline struct capture_control *task_capc(struct zone *zone)
848 struct capture_control *capc = current->capture_control;
850 return unlikely(capc) &&
851 !(current->flags & PF_KTHREAD) &&
853 capc->cc->zone == zone ? capc : NULL;
857 compaction_capture(struct capture_control *capc, struct page *page,
858 int order, int migratetype)
860 if (!capc || order != capc->cc->order)
863 /* Do not accidentally pollute CMA or isolated regions*/
864 if (is_migrate_cma(migratetype) ||
865 is_migrate_isolate(migratetype))
869 * Do not let lower order allocations polluate a movable pageblock.
870 * This might let an unmovable request use a reclaimable pageblock
871 * and vice-versa but no more than normal fallback logic which can
872 * have trouble finding a high-order free page.
874 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
882 static inline struct capture_control *task_capc(struct zone *zone)
888 compaction_capture(struct capture_control *capc, struct page *page,
889 int order, int migratetype)
893 #endif /* CONFIG_COMPACTION */
895 /* Used for pages not on another list */
896 static inline void add_to_free_list(struct page *page, struct zone *zone,
897 unsigned int order, int migratetype)
899 struct free_area *area = &zone->free_area[order];
901 list_add(&page->lru, &area->free_list[migratetype]);
905 /* Used for pages not on another list */
906 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
907 unsigned int order, int migratetype)
909 struct free_area *area = &zone->free_area[order];
911 list_add_tail(&page->lru, &area->free_list[migratetype]);
916 * Used for pages which are on another list. Move the pages to the tail
917 * of the list - so the moved pages won't immediately be considered for
918 * allocation again (e.g., optimization for memory onlining).
920 static inline void move_to_free_list(struct page *page, struct zone *zone,
921 unsigned int order, int migratetype)
923 struct free_area *area = &zone->free_area[order];
925 list_move_tail(&page->lru, &area->free_list[migratetype]);
928 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
931 /* clear reported state and update reported page count */
932 if (page_reported(page))
933 __ClearPageReported(page);
935 list_del(&page->lru);
936 __ClearPageBuddy(page);
937 set_page_private(page, 0);
938 zone->free_area[order].nr_free--;
942 * If this is not the largest possible page, check if the buddy
943 * of the next-highest order is free. If it is, it's possible
944 * that pages are being freed that will coalesce soon. In case,
945 * that is happening, add the free page to the tail of the list
946 * so it's less likely to be used soon and more likely to be merged
947 * as a higher order page
950 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
951 struct page *page, unsigned int order)
953 struct page *higher_page, *higher_buddy;
954 unsigned long combined_pfn;
956 if (order >= MAX_ORDER - 2)
959 if (!pfn_valid_within(buddy_pfn))
962 combined_pfn = buddy_pfn & pfn;
963 higher_page = page + (combined_pfn - pfn);
964 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
965 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
967 return pfn_valid_within(buddy_pfn) &&
968 page_is_buddy(higher_page, higher_buddy, order + 1);
972 * Freeing function for a buddy system allocator.
974 * The concept of a buddy system is to maintain direct-mapped table
975 * (containing bit values) for memory blocks of various "orders".
976 * The bottom level table contains the map for the smallest allocatable
977 * units of memory (here, pages), and each level above it describes
978 * pairs of units from the levels below, hence, "buddies".
979 * At a high level, all that happens here is marking the table entry
980 * at the bottom level available, and propagating the changes upward
981 * as necessary, plus some accounting needed to play nicely with other
982 * parts of the VM system.
983 * At each level, we keep a list of pages, which are heads of continuous
984 * free pages of length of (1 << order) and marked with PageBuddy.
985 * Page's order is recorded in page_private(page) field.
986 * So when we are allocating or freeing one, we can derive the state of the
987 * other. That is, if we allocate a small block, and both were
988 * free, the remainder of the region must be split into blocks.
989 * If a block is freed, and its buddy is also free, then this
990 * triggers coalescing into a block of larger size.
995 static inline void __free_one_page(struct page *page,
997 struct zone *zone, unsigned int order,
998 int migratetype, fpi_t fpi_flags)
1000 struct capture_control *capc = task_capc(zone);
1001 unsigned long buddy_pfn;
1002 unsigned long combined_pfn;
1003 unsigned int max_order;
1007 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1009 VM_BUG_ON(!zone_is_initialized(zone));
1010 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1012 VM_BUG_ON(migratetype == -1);
1013 if (likely(!is_migrate_isolate(migratetype)))
1014 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1016 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1017 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1020 while (order < max_order) {
1021 if (compaction_capture(capc, page, order, migratetype)) {
1022 __mod_zone_freepage_state(zone, -(1 << order),
1026 buddy_pfn = __find_buddy_pfn(pfn, order);
1027 buddy = page + (buddy_pfn - pfn);
1029 if (!pfn_valid_within(buddy_pfn))
1031 if (!page_is_buddy(page, buddy, order))
1034 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1035 * merge with it and move up one order.
1037 if (page_is_guard(buddy))
1038 clear_page_guard(zone, buddy, order, migratetype);
1040 del_page_from_free_list(buddy, zone, order);
1041 combined_pfn = buddy_pfn & pfn;
1042 page = page + (combined_pfn - pfn);
1046 if (order < MAX_ORDER - 1) {
1047 /* If we are here, it means order is >= pageblock_order.
1048 * We want to prevent merge between freepages on isolate
1049 * pageblock and normal pageblock. Without this, pageblock
1050 * isolation could cause incorrect freepage or CMA accounting.
1052 * We don't want to hit this code for the more frequent
1053 * low-order merging.
1055 if (unlikely(has_isolate_pageblock(zone))) {
1058 buddy_pfn = __find_buddy_pfn(pfn, order);
1059 buddy = page + (buddy_pfn - pfn);
1060 buddy_mt = get_pageblock_migratetype(buddy);
1062 if (migratetype != buddy_mt
1063 && (is_migrate_isolate(migratetype) ||
1064 is_migrate_isolate(buddy_mt)))
1067 max_order = order + 1;
1068 goto continue_merging;
1072 set_buddy_order(page, order);
1074 if (fpi_flags & FPI_TO_TAIL)
1076 else if (is_shuffle_order(order))
1077 to_tail = shuffle_pick_tail();
1079 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1082 add_to_free_list_tail(page, zone, order, migratetype);
1084 add_to_free_list(page, zone, order, migratetype);
1086 /* Notify page reporting subsystem of freed page */
1087 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1088 page_reporting_notify_free(order);
1092 * A bad page could be due to a number of fields. Instead of multiple branches,
1093 * try and check multiple fields with one check. The caller must do a detailed
1094 * check if necessary.
1096 static inline bool page_expected_state(struct page *page,
1097 unsigned long check_flags)
1099 if (unlikely(atomic_read(&page->_mapcount) != -1))
1102 if (unlikely((unsigned long)page->mapping |
1103 page_ref_count(page) |
1105 (unsigned long)page->mem_cgroup |
1107 (page->flags & check_flags)))
1113 static const char *page_bad_reason(struct page *page, unsigned long flags)
1115 const char *bad_reason = NULL;
1117 if (unlikely(atomic_read(&page->_mapcount) != -1))
1118 bad_reason = "nonzero mapcount";
1119 if (unlikely(page->mapping != NULL))
1120 bad_reason = "non-NULL mapping";
1121 if (unlikely(page_ref_count(page) != 0))
1122 bad_reason = "nonzero _refcount";
1123 if (unlikely(page->flags & flags)) {
1124 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1125 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1127 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1130 if (unlikely(page->mem_cgroup))
1131 bad_reason = "page still charged to cgroup";
1136 static void check_free_page_bad(struct page *page)
1139 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1142 static inline int check_free_page(struct page *page)
1144 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1147 /* Something has gone sideways, find it */
1148 check_free_page_bad(page);
1152 static int free_tail_pages_check(struct page *head_page, struct page *page)
1157 * We rely page->lru.next never has bit 0 set, unless the page
1158 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1160 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1162 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1166 switch (page - head_page) {
1168 /* the first tail page: ->mapping may be compound_mapcount() */
1169 if (unlikely(compound_mapcount(page))) {
1170 bad_page(page, "nonzero compound_mapcount");
1176 * the second tail page: ->mapping is
1177 * deferred_list.next -- ignore value.
1181 if (page->mapping != TAIL_MAPPING) {
1182 bad_page(page, "corrupted mapping in tail page");
1187 if (unlikely(!PageTail(page))) {
1188 bad_page(page, "PageTail not set");
1191 if (unlikely(compound_head(page) != head_page)) {
1192 bad_page(page, "compound_head not consistent");
1197 page->mapping = NULL;
1198 clear_compound_head(page);
1202 static void kernel_init_free_pages(struct page *page, int numpages)
1206 /* s390's use of memset() could override KASAN redzones. */
1207 kasan_disable_current();
1208 for (i = 0; i < numpages; i++)
1209 clear_highpage(page + i);
1210 kasan_enable_current();
1213 static __always_inline bool free_pages_prepare(struct page *page,
1214 unsigned int order, bool check_free)
1218 VM_BUG_ON_PAGE(PageTail(page), page);
1220 trace_mm_page_free(page, order);
1222 if (unlikely(PageHWPoison(page)) && !order) {
1224 * Do not let hwpoison pages hit pcplists/buddy
1225 * Untie memcg state and reset page's owner
1227 if (memcg_kmem_enabled() && PageKmemcg(page))
1228 __memcg_kmem_uncharge_page(page, order);
1229 reset_page_owner(page, order);
1234 * Check tail pages before head page information is cleared to
1235 * avoid checking PageCompound for order-0 pages.
1237 if (unlikely(order)) {
1238 bool compound = PageCompound(page);
1241 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1244 ClearPageDoubleMap(page);
1245 for (i = 1; i < (1 << order); i++) {
1247 bad += free_tail_pages_check(page, page + i);
1248 if (unlikely(check_free_page(page + i))) {
1252 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1255 if (PageMappingFlags(page))
1256 page->mapping = NULL;
1257 if (memcg_kmem_enabled() && PageKmemcg(page))
1258 __memcg_kmem_uncharge_page(page, order);
1260 bad += check_free_page(page);
1264 page_cpupid_reset_last(page);
1265 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1266 reset_page_owner(page, order);
1268 if (!PageHighMem(page)) {
1269 debug_check_no_locks_freed(page_address(page),
1270 PAGE_SIZE << order);
1271 debug_check_no_obj_freed(page_address(page),
1272 PAGE_SIZE << order);
1274 if (want_init_on_free())
1275 kernel_init_free_pages(page, 1 << order);
1277 kernel_poison_pages(page, 1 << order, 0);
1279 * arch_free_page() can make the page's contents inaccessible. s390
1280 * does this. So nothing which can access the page's contents should
1281 * happen after this.
1283 arch_free_page(page, order);
1285 if (debug_pagealloc_enabled_static())
1286 kernel_map_pages(page, 1 << order, 0);
1288 kasan_free_nondeferred_pages(page, order);
1293 #ifdef CONFIG_DEBUG_VM
1295 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1296 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1297 * moved from pcp lists to free lists.
1299 static bool free_pcp_prepare(struct page *page)
1301 return free_pages_prepare(page, 0, true);
1304 static bool bulkfree_pcp_prepare(struct page *page)
1306 if (debug_pagealloc_enabled_static())
1307 return check_free_page(page);
1313 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1314 * moving from pcp lists to free list in order to reduce overhead. With
1315 * debug_pagealloc enabled, they are checked also immediately when being freed
1318 static bool free_pcp_prepare(struct page *page)
1320 if (debug_pagealloc_enabled_static())
1321 return free_pages_prepare(page, 0, true);
1323 return free_pages_prepare(page, 0, false);
1326 static bool bulkfree_pcp_prepare(struct page *page)
1328 return check_free_page(page);
1330 #endif /* CONFIG_DEBUG_VM */
1332 static inline void prefetch_buddy(struct page *page)
1334 unsigned long pfn = page_to_pfn(page);
1335 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1336 struct page *buddy = page + (buddy_pfn - pfn);
1342 * Frees a number of pages which have been collected from the pcp lists.
1343 * Assumes all pages on list are in same zone, and of same order.
1344 * count is the number of pages to free.
1346 * If the zone was previously in an "all pages pinned" state then look to
1347 * see if this freeing clears that state.
1349 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1350 * pinned" detection logic.
1352 static void free_pcppages_bulk(struct zone *zone, struct list_head *head,
1355 bool isolated_pageblocks;
1356 struct page *page, *tmp;
1357 unsigned long flags;
1359 spin_lock_irqsave(&zone->lock, flags);
1360 isolated_pageblocks = has_isolate_pageblock(zone);
1363 * Use safe version since after __free_one_page(),
1364 * page->lru.next will not point to original list.
1366 list_for_each_entry_safe(page, tmp, head, lru) {
1367 int mt = get_pcppage_migratetype(page);
1369 if (page_zone(page) != zone) {
1371 * free_unref_page_list() sorts pages by zone. If we end
1372 * up with pages from a different NUMA nodes belonging
1373 * to the same ZONE index then we need to redo with the
1374 * correct ZONE pointer. Skip the page for now, redo it
1375 * on the next iteration.
1377 WARN_ON_ONCE(zone_retry == false);
1382 /* MIGRATE_ISOLATE page should not go to pcplists */
1383 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1384 /* Pageblock could have been isolated meanwhile */
1385 if (unlikely(isolated_pageblocks))
1386 mt = get_pageblock_migratetype(page);
1388 list_del(&page->lru);
1389 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1390 trace_mm_page_pcpu_drain(page, 0, mt);
1392 spin_unlock_irqrestore(&zone->lock, flags);
1395 static void isolate_pcp_pages(int count, struct per_cpu_pages *pcp,
1396 struct list_head *dst)
1398 int migratetype = 0;
1400 int prefetch_nr = 0;
1404 * Ensure proper count is passed which otherwise would stuck in the
1405 * below while (list_empty(list)) loop.
1407 count = min(pcp->count, count);
1409 struct list_head *list;
1412 * Remove pages from lists in a round-robin fashion. A
1413 * batch_free count is maintained that is incremented when an
1414 * empty list is encountered. This is so more pages are freed
1415 * off fuller lists instead of spinning excessively around empty
1420 if (++migratetype == MIGRATE_PCPTYPES)
1422 list = &pcp->lists[migratetype];
1423 } while (list_empty(list));
1425 /* This is the only non-empty list. Free them all. */
1426 if (batch_free == MIGRATE_PCPTYPES)
1430 page = list_last_entry(list, struct page, lru);
1431 /* must delete to avoid corrupting pcp list */
1432 list_del(&page->lru);
1435 if (bulkfree_pcp_prepare(page))
1438 list_add_tail(&page->lru, dst);
1441 * We are going to put the page back to the global
1442 * pool, prefetch its buddy to speed up later access
1443 * under zone->lock. It is believed the overhead of
1444 * an additional test and calculating buddy_pfn here
1445 * can be offset by reduced memory latency later. To
1446 * avoid excessive prefetching due to large count, only
1447 * prefetch buddy for the first pcp->batch nr of pages.
1449 if (prefetch_nr++ < pcp->batch)
1450 prefetch_buddy(page);
1451 } while (--count && --batch_free && !list_empty(list));
1455 static void free_one_page(struct zone *zone,
1456 struct page *page, unsigned long pfn,
1458 int migratetype, fpi_t fpi_flags)
1460 spin_lock(&zone->lock);
1461 if (unlikely(has_isolate_pageblock(zone) ||
1462 is_migrate_isolate(migratetype))) {
1463 migratetype = get_pfnblock_migratetype(page, pfn);
1465 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1466 spin_unlock(&zone->lock);
1469 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1470 unsigned long zone, int nid)
1472 mm_zero_struct_page(page);
1473 set_page_links(page, zone, nid, pfn);
1474 init_page_count(page);
1475 page_mapcount_reset(page);
1476 page_cpupid_reset_last(page);
1477 page_kasan_tag_reset(page);
1479 INIT_LIST_HEAD(&page->lru);
1480 #ifdef WANT_PAGE_VIRTUAL
1481 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1482 if (!is_highmem_idx(zone))
1483 set_page_address(page, __va(pfn << PAGE_SHIFT));
1487 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1488 static void __meminit init_reserved_page(unsigned long pfn)
1493 if (!early_page_uninitialised(pfn))
1496 nid = early_pfn_to_nid(pfn);
1497 pgdat = NODE_DATA(nid);
1499 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1500 struct zone *zone = &pgdat->node_zones[zid];
1502 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1505 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1508 static inline void init_reserved_page(unsigned long pfn)
1511 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1514 * Initialised pages do not have PageReserved set. This function is
1515 * called for each range allocated by the bootmem allocator and
1516 * marks the pages PageReserved. The remaining valid pages are later
1517 * sent to the buddy page allocator.
1519 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1521 unsigned long start_pfn = PFN_DOWN(start);
1522 unsigned long end_pfn = PFN_UP(end);
1524 for (; start_pfn < end_pfn; start_pfn++) {
1525 if (pfn_valid(start_pfn)) {
1526 struct page *page = pfn_to_page(start_pfn);
1528 init_reserved_page(start_pfn);
1530 /* Avoid false-positive PageTail() */
1531 INIT_LIST_HEAD(&page->lru);
1534 * no need for atomic set_bit because the struct
1535 * page is not visible yet so nobody should
1538 __SetPageReserved(page);
1543 static void __free_pages_ok(struct page *page, unsigned int order,
1546 unsigned long flags;
1548 unsigned long pfn = page_to_pfn(page);
1550 if (!free_pages_prepare(page, order, true))
1553 migratetype = get_pfnblock_migratetype(page, pfn);
1554 local_lock_irqsave(&pa_lock.l, flags);
1555 __count_vm_events(PGFREE, 1 << order);
1556 free_one_page(page_zone(page), page, pfn, order, migratetype,
1558 local_unlock_irqrestore(&pa_lock.l, flags);
1561 void __free_pages_core(struct page *page, unsigned int order)
1563 unsigned int nr_pages = 1 << order;
1564 struct page *p = page;
1568 * When initializing the memmap, __init_single_page() sets the refcount
1569 * of all pages to 1 ("allocated"/"not free"). We have to set the
1570 * refcount of all involved pages to 0.
1573 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1575 __ClearPageReserved(p);
1576 set_page_count(p, 0);
1578 __ClearPageReserved(p);
1579 set_page_count(p, 0);
1581 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1584 * Bypass PCP and place fresh pages right to the tail, primarily
1585 * relevant for memory onlining.
1587 __free_pages_ok(page, order, FPI_TO_TAIL);
1590 #ifdef CONFIG_NEED_MULTIPLE_NODES
1592 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1594 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1597 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1599 int __meminit __early_pfn_to_nid(unsigned long pfn,
1600 struct mminit_pfnnid_cache *state)
1602 unsigned long start_pfn, end_pfn;
1605 if (state->last_start <= pfn && pfn < state->last_end)
1606 return state->last_nid;
1608 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1609 if (nid != NUMA_NO_NODE) {
1610 state->last_start = start_pfn;
1611 state->last_end = end_pfn;
1612 state->last_nid = nid;
1617 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1619 int __meminit early_pfn_to_nid(unsigned long pfn)
1621 static DEFINE_SPINLOCK(early_pfn_lock);
1624 spin_lock(&early_pfn_lock);
1625 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1627 nid = first_online_node;
1628 spin_unlock(&early_pfn_lock);
1632 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1634 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1637 if (early_page_uninitialised(pfn))
1639 __free_pages_core(page, order);
1643 * Check that the whole (or subset of) a pageblock given by the interval of
1644 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1645 * with the migration of free compaction scanner. The scanners then need to
1646 * use only pfn_valid_within() check for arches that allow holes within
1649 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1651 * It's possible on some configurations to have a setup like node0 node1 node0
1652 * i.e. it's possible that all pages within a zones range of pages do not
1653 * belong to a single zone. We assume that a border between node0 and node1
1654 * can occur within a single pageblock, but not a node0 node1 node0
1655 * interleaving within a single pageblock. It is therefore sufficient to check
1656 * the first and last page of a pageblock and avoid checking each individual
1657 * page in a pageblock.
1659 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1660 unsigned long end_pfn, struct zone *zone)
1662 struct page *start_page;
1663 struct page *end_page;
1665 /* end_pfn is one past the range we are checking */
1668 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1671 start_page = pfn_to_online_page(start_pfn);
1675 if (page_zone(start_page) != zone)
1678 end_page = pfn_to_page(end_pfn);
1680 /* This gives a shorter code than deriving page_zone(end_page) */
1681 if (page_zone_id(start_page) != page_zone_id(end_page))
1687 void set_zone_contiguous(struct zone *zone)
1689 unsigned long block_start_pfn = zone->zone_start_pfn;
1690 unsigned long block_end_pfn;
1692 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1693 for (; block_start_pfn < zone_end_pfn(zone);
1694 block_start_pfn = block_end_pfn,
1695 block_end_pfn += pageblock_nr_pages) {
1697 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1699 if (!__pageblock_pfn_to_page(block_start_pfn,
1700 block_end_pfn, zone))
1705 /* We confirm that there is no hole */
1706 zone->contiguous = true;
1709 void clear_zone_contiguous(struct zone *zone)
1711 zone->contiguous = false;
1714 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1715 static void __init deferred_free_range(unsigned long pfn,
1716 unsigned long nr_pages)
1724 page = pfn_to_page(pfn);
1726 /* Free a large naturally-aligned chunk if possible */
1727 if (nr_pages == pageblock_nr_pages &&
1728 (pfn & (pageblock_nr_pages - 1)) == 0) {
1729 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1730 __free_pages_core(page, pageblock_order);
1734 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1735 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1736 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1737 __free_pages_core(page, 0);
1741 /* Completion tracking for deferred_init_memmap() threads */
1742 static atomic_t pgdat_init_n_undone __initdata;
1743 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1745 static inline void __init pgdat_init_report_one_done(void)
1747 if (atomic_dec_and_test(&pgdat_init_n_undone))
1748 complete(&pgdat_init_all_done_comp);
1752 * Returns true if page needs to be initialized or freed to buddy allocator.
1754 * First we check if pfn is valid on architectures where it is possible to have
1755 * holes within pageblock_nr_pages. On systems where it is not possible, this
1756 * function is optimized out.
1758 * Then, we check if a current large page is valid by only checking the validity
1761 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1763 if (!pfn_valid_within(pfn))
1765 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1771 * Free pages to buddy allocator. Try to free aligned pages in
1772 * pageblock_nr_pages sizes.
1774 static void __init deferred_free_pages(unsigned long pfn,
1775 unsigned long end_pfn)
1777 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1778 unsigned long nr_free = 0;
1780 for (; pfn < end_pfn; pfn++) {
1781 if (!deferred_pfn_valid(pfn)) {
1782 deferred_free_range(pfn - nr_free, nr_free);
1784 } else if (!(pfn & nr_pgmask)) {
1785 deferred_free_range(pfn - nr_free, nr_free);
1791 /* Free the last block of pages to allocator */
1792 deferred_free_range(pfn - nr_free, nr_free);
1796 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1797 * by performing it only once every pageblock_nr_pages.
1798 * Return number of pages initialized.
1800 static unsigned long __init deferred_init_pages(struct zone *zone,
1802 unsigned long end_pfn)
1804 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1805 int nid = zone_to_nid(zone);
1806 unsigned long nr_pages = 0;
1807 int zid = zone_idx(zone);
1808 struct page *page = NULL;
1810 for (; pfn < end_pfn; pfn++) {
1811 if (!deferred_pfn_valid(pfn)) {
1814 } else if (!page || !(pfn & nr_pgmask)) {
1815 page = pfn_to_page(pfn);
1819 __init_single_page(page, pfn, zid, nid);
1826 * This function is meant to pre-load the iterator for the zone init.
1827 * Specifically it walks through the ranges until we are caught up to the
1828 * first_init_pfn value and exits there. If we never encounter the value we
1829 * return false indicating there are no valid ranges left.
1832 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1833 unsigned long *spfn, unsigned long *epfn,
1834 unsigned long first_init_pfn)
1839 * Start out by walking through the ranges in this zone that have
1840 * already been initialized. We don't need to do anything with them
1841 * so we just need to flush them out of the system.
1843 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1844 if (*epfn <= first_init_pfn)
1846 if (*spfn < first_init_pfn)
1847 *spfn = first_init_pfn;
1856 * Initialize and free pages. We do it in two loops: first we initialize
1857 * struct page, then free to buddy allocator, because while we are
1858 * freeing pages we can access pages that are ahead (computing buddy
1859 * page in __free_one_page()).
1861 * In order to try and keep some memory in the cache we have the loop
1862 * broken along max page order boundaries. This way we will not cause
1863 * any issues with the buddy page computation.
1865 static unsigned long __init
1866 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1867 unsigned long *end_pfn)
1869 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1870 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1871 unsigned long nr_pages = 0;
1874 /* First we loop through and initialize the page values */
1875 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1878 if (mo_pfn <= *start_pfn)
1881 t = min(mo_pfn, *end_pfn);
1882 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1884 if (mo_pfn < *end_pfn) {
1885 *start_pfn = mo_pfn;
1890 /* Reset values and now loop through freeing pages as needed */
1893 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1899 t = min(mo_pfn, epfn);
1900 deferred_free_pages(spfn, t);
1910 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1913 unsigned long spfn, epfn;
1914 struct zone *zone = arg;
1917 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1920 * Initialize and free pages in MAX_ORDER sized increments so that we
1921 * can avoid introducing any issues with the buddy allocator.
1923 while (spfn < end_pfn) {
1924 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1929 /* An arch may override for more concurrency. */
1931 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1936 /* Initialise remaining memory on a node */
1937 static int __init deferred_init_memmap(void *data)
1939 pg_data_t *pgdat = data;
1940 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1941 unsigned long spfn = 0, epfn = 0;
1942 unsigned long first_init_pfn, flags;
1943 unsigned long start = jiffies;
1945 int zid, max_threads;
1948 /* Bind memory initialisation thread to a local node if possible */
1949 if (!cpumask_empty(cpumask))
1950 set_cpus_allowed_ptr(current, cpumask);
1952 pgdat_resize_lock(pgdat, &flags);
1953 first_init_pfn = pgdat->first_deferred_pfn;
1954 if (first_init_pfn == ULONG_MAX) {
1955 pgdat_resize_unlock(pgdat, &flags);
1956 pgdat_init_report_one_done();
1960 /* Sanity check boundaries */
1961 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1962 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1963 pgdat->first_deferred_pfn = ULONG_MAX;
1966 * Once we unlock here, the zone cannot be grown anymore, thus if an
1967 * interrupt thread must allocate this early in boot, zone must be
1968 * pre-grown prior to start of deferred page initialization.
1970 pgdat_resize_unlock(pgdat, &flags);
1972 /* Only the highest zone is deferred so find it */
1973 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1974 zone = pgdat->node_zones + zid;
1975 if (first_init_pfn < zone_end_pfn(zone))
1979 /* If the zone is empty somebody else may have cleared out the zone */
1980 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1984 max_threads = deferred_page_init_max_threads(cpumask);
1986 while (spfn < epfn) {
1987 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1988 struct padata_mt_job job = {
1989 .thread_fn = deferred_init_memmap_chunk,
1992 .size = epfn_align - spfn,
1993 .align = PAGES_PER_SECTION,
1994 .min_chunk = PAGES_PER_SECTION,
1995 .max_threads = max_threads,
1998 padata_do_multithreaded(&job);
1999 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2003 /* Sanity check that the next zone really is unpopulated */
2004 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2006 pr_info("node %d deferred pages initialised in %ums\n",
2007 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2009 pgdat_init_report_one_done();
2014 * If this zone has deferred pages, try to grow it by initializing enough
2015 * deferred pages to satisfy the allocation specified by order, rounded up to
2016 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2017 * of SECTION_SIZE bytes by initializing struct pages in increments of
2018 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2020 * Return true when zone was grown, otherwise return false. We return true even
2021 * when we grow less than requested, to let the caller decide if there are
2022 * enough pages to satisfy the allocation.
2024 * Note: We use noinline because this function is needed only during boot, and
2025 * it is called from a __ref function _deferred_grow_zone. This way we are
2026 * making sure that it is not inlined into permanent text section.
2028 static noinline bool __init
2029 deferred_grow_zone(struct zone *zone, unsigned int order)
2031 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2032 pg_data_t *pgdat = zone->zone_pgdat;
2033 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2034 unsigned long spfn, epfn, flags;
2035 unsigned long nr_pages = 0;
2038 /* Only the last zone may have deferred pages */
2039 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2042 pgdat_resize_lock(pgdat, &flags);
2045 * If someone grew this zone while we were waiting for spinlock, return
2046 * true, as there might be enough pages already.
2048 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2049 pgdat_resize_unlock(pgdat, &flags);
2053 /* If the zone is empty somebody else may have cleared out the zone */
2054 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2055 first_deferred_pfn)) {
2056 pgdat->first_deferred_pfn = ULONG_MAX;
2057 pgdat_resize_unlock(pgdat, &flags);
2058 /* Retry only once. */
2059 return first_deferred_pfn != ULONG_MAX;
2063 * Initialize and free pages in MAX_ORDER sized increments so
2064 * that we can avoid introducing any issues with the buddy
2067 while (spfn < epfn) {
2068 /* update our first deferred PFN for this section */
2069 first_deferred_pfn = spfn;
2071 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2072 touch_nmi_watchdog();
2074 /* We should only stop along section boundaries */
2075 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2078 /* If our quota has been met we can stop here */
2079 if (nr_pages >= nr_pages_needed)
2083 pgdat->first_deferred_pfn = spfn;
2084 pgdat_resize_unlock(pgdat, &flags);
2086 return nr_pages > 0;
2090 * deferred_grow_zone() is __init, but it is called from
2091 * get_page_from_freelist() during early boot until deferred_pages permanently
2092 * disables this call. This is why we have refdata wrapper to avoid warning,
2093 * and to ensure that the function body gets unloaded.
2096 _deferred_grow_zone(struct zone *zone, unsigned int order)
2098 return deferred_grow_zone(zone, order);
2101 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2103 void __init page_alloc_init_late(void)
2108 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2110 /* There will be num_node_state(N_MEMORY) threads */
2111 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2112 for_each_node_state(nid, N_MEMORY) {
2113 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2116 /* Block until all are initialised */
2117 wait_for_completion(&pgdat_init_all_done_comp);
2120 * The number of managed pages has changed due to the initialisation
2121 * so the pcpu batch and high limits needs to be updated or the limits
2122 * will be artificially small.
2124 for_each_populated_zone(zone)
2125 zone_pcp_update(zone);
2128 * We initialized the rest of the deferred pages. Permanently disable
2129 * on-demand struct page initialization.
2131 static_branch_disable(&deferred_pages);
2133 /* Reinit limits that are based on free pages after the kernel is up */
2134 files_maxfiles_init();
2137 /* Discard memblock private memory */
2140 for_each_node_state(nid, N_MEMORY)
2141 shuffle_free_memory(NODE_DATA(nid));
2143 for_each_populated_zone(zone)
2144 set_zone_contiguous(zone);
2148 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2149 void __init init_cma_reserved_pageblock(struct page *page)
2151 unsigned i = pageblock_nr_pages;
2152 struct page *p = page;
2155 __ClearPageReserved(p);
2156 set_page_count(p, 0);
2159 set_pageblock_migratetype(page, MIGRATE_CMA);
2161 if (pageblock_order >= MAX_ORDER) {
2162 i = pageblock_nr_pages;
2165 set_page_refcounted(p);
2166 __free_pages(p, MAX_ORDER - 1);
2167 p += MAX_ORDER_NR_PAGES;
2168 } while (i -= MAX_ORDER_NR_PAGES);
2170 set_page_refcounted(page);
2171 __free_pages(page, pageblock_order);
2174 adjust_managed_page_count(page, pageblock_nr_pages);
2179 * The order of subdivision here is critical for the IO subsystem.
2180 * Please do not alter this order without good reasons and regression
2181 * testing. Specifically, as large blocks of memory are subdivided,
2182 * the order in which smaller blocks are delivered depends on the order
2183 * they're subdivided in this function. This is the primary factor
2184 * influencing the order in which pages are delivered to the IO
2185 * subsystem according to empirical testing, and this is also justified
2186 * by considering the behavior of a buddy system containing a single
2187 * large block of memory acted on by a series of small allocations.
2188 * This behavior is a critical factor in sglist merging's success.
2192 static inline void expand(struct zone *zone, struct page *page,
2193 int low, int high, int migratetype)
2195 unsigned long size = 1 << high;
2197 while (high > low) {
2200 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2203 * Mark as guard pages (or page), that will allow to
2204 * merge back to allocator when buddy will be freed.
2205 * Corresponding page table entries will not be touched,
2206 * pages will stay not present in virtual address space
2208 if (set_page_guard(zone, &page[size], high, migratetype))
2211 add_to_free_list(&page[size], zone, high, migratetype);
2212 set_buddy_order(&page[size], high);
2216 static void check_new_page_bad(struct page *page)
2218 if (unlikely(page->flags & __PG_HWPOISON)) {
2219 /* Don't complain about hwpoisoned pages */
2220 page_mapcount_reset(page); /* remove PageBuddy */
2225 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2229 * This page is about to be returned from the page allocator
2231 static inline int check_new_page(struct page *page)
2233 if (likely(page_expected_state(page,
2234 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2237 check_new_page_bad(page);
2241 static inline bool free_pages_prezeroed(void)
2243 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2244 page_poisoning_enabled()) || want_init_on_free();
2247 #ifdef CONFIG_DEBUG_VM
2249 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2250 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2251 * also checked when pcp lists are refilled from the free lists.
2253 static inline bool check_pcp_refill(struct page *page)
2255 if (debug_pagealloc_enabled_static())
2256 return check_new_page(page);
2261 static inline bool check_new_pcp(struct page *page)
2263 return check_new_page(page);
2267 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2268 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2269 * enabled, they are also checked when being allocated from the pcp lists.
2271 static inline bool check_pcp_refill(struct page *page)
2273 return check_new_page(page);
2275 static inline bool check_new_pcp(struct page *page)
2277 if (debug_pagealloc_enabled_static())
2278 return check_new_page(page);
2282 #endif /* CONFIG_DEBUG_VM */
2284 static bool check_new_pages(struct page *page, unsigned int order)
2287 for (i = 0; i < (1 << order); i++) {
2288 struct page *p = page + i;
2290 if (unlikely(check_new_page(p)))
2297 inline void post_alloc_hook(struct page *page, unsigned int order,
2300 set_page_private(page, 0);
2301 set_page_refcounted(page);
2303 arch_alloc_page(page, order);
2304 if (debug_pagealloc_enabled_static())
2305 kernel_map_pages(page, 1 << order, 1);
2306 kasan_alloc_pages(page, order);
2307 kernel_poison_pages(page, 1 << order, 1);
2308 set_page_owner(page, order, gfp_flags);
2311 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2312 unsigned int alloc_flags)
2314 post_alloc_hook(page, order, gfp_flags);
2316 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2317 kernel_init_free_pages(page, 1 << order);
2319 if (order && (gfp_flags & __GFP_COMP))
2320 prep_compound_page(page, order);
2323 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2324 * allocate the page. The expectation is that the caller is taking
2325 * steps that will free more memory. The caller should avoid the page
2326 * being used for !PFMEMALLOC purposes.
2328 if (alloc_flags & ALLOC_NO_WATERMARKS)
2329 set_page_pfmemalloc(page);
2331 clear_page_pfmemalloc(page);
2335 * Go through the free lists for the given migratetype and remove
2336 * the smallest available page from the freelists
2338 static __always_inline
2339 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2342 unsigned int current_order;
2343 struct free_area *area;
2346 /* Find a page of the appropriate size in the preferred list */
2347 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2348 area = &(zone->free_area[current_order]);
2349 page = get_page_from_free_area(area, migratetype);
2352 del_page_from_free_list(page, zone, current_order);
2353 expand(zone, page, order, current_order, migratetype);
2354 set_pcppage_migratetype(page, migratetype);
2363 * This array describes the order lists are fallen back to when
2364 * the free lists for the desirable migrate type are depleted
2366 static int fallbacks[MIGRATE_TYPES][3] = {
2367 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2368 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2369 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2371 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2373 #ifdef CONFIG_MEMORY_ISOLATION
2374 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2379 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2382 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2385 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2386 unsigned int order) { return NULL; }
2390 * Move the free pages in a range to the freelist tail of the requested type.
2391 * Note that start_page and end_pages are not aligned on a pageblock
2392 * boundary. If alignment is required, use move_freepages_block()
2394 static int move_freepages(struct zone *zone,
2395 struct page *start_page, struct page *end_page,
2396 int migratetype, int *num_movable)
2400 int pages_moved = 0;
2402 for (page = start_page; page <= end_page;) {
2403 if (!pfn_valid_within(page_to_pfn(page))) {
2408 if (!PageBuddy(page)) {
2410 * We assume that pages that could be isolated for
2411 * migration are movable. But we don't actually try
2412 * isolating, as that would be expensive.
2415 (PageLRU(page) || __PageMovable(page)))
2422 /* Make sure we are not inadvertently changing nodes */
2423 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2424 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2426 order = buddy_order(page);
2427 move_to_free_list(page, zone, order, migratetype);
2429 pages_moved += 1 << order;
2435 int move_freepages_block(struct zone *zone, struct page *page,
2436 int migratetype, int *num_movable)
2438 unsigned long start_pfn, end_pfn;
2439 struct page *start_page, *end_page;
2444 start_pfn = page_to_pfn(page);
2445 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2446 start_page = pfn_to_page(start_pfn);
2447 end_page = start_page + pageblock_nr_pages - 1;
2448 end_pfn = start_pfn + pageblock_nr_pages - 1;
2450 /* Do not cross zone boundaries */
2451 if (!zone_spans_pfn(zone, start_pfn))
2453 if (!zone_spans_pfn(zone, end_pfn))
2456 return move_freepages(zone, start_page, end_page, migratetype,
2460 static void change_pageblock_range(struct page *pageblock_page,
2461 int start_order, int migratetype)
2463 int nr_pageblocks = 1 << (start_order - pageblock_order);
2465 while (nr_pageblocks--) {
2466 set_pageblock_migratetype(pageblock_page, migratetype);
2467 pageblock_page += pageblock_nr_pages;
2472 * When we are falling back to another migratetype during allocation, try to
2473 * steal extra free pages from the same pageblocks to satisfy further
2474 * allocations, instead of polluting multiple pageblocks.
2476 * If we are stealing a relatively large buddy page, it is likely there will
2477 * be more free pages in the pageblock, so try to steal them all. For
2478 * reclaimable and unmovable allocations, we steal regardless of page size,
2479 * as fragmentation caused by those allocations polluting movable pageblocks
2480 * is worse than movable allocations stealing from unmovable and reclaimable
2483 static bool can_steal_fallback(unsigned int order, int start_mt)
2486 * Leaving this order check is intended, although there is
2487 * relaxed order check in next check. The reason is that
2488 * we can actually steal whole pageblock if this condition met,
2489 * but, below check doesn't guarantee it and that is just heuristic
2490 * so could be changed anytime.
2492 if (order >= pageblock_order)
2495 if (order >= pageblock_order / 2 ||
2496 start_mt == MIGRATE_RECLAIMABLE ||
2497 start_mt == MIGRATE_UNMOVABLE ||
2498 page_group_by_mobility_disabled)
2504 static inline bool boost_watermark(struct zone *zone)
2506 unsigned long max_boost;
2508 if (!watermark_boost_factor)
2511 * Don't bother in zones that are unlikely to produce results.
2512 * On small machines, including kdump capture kernels running
2513 * in a small area, boosting the watermark can cause an out of
2514 * memory situation immediately.
2516 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2519 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2520 watermark_boost_factor, 10000);
2523 * high watermark may be uninitialised if fragmentation occurs
2524 * very early in boot so do not boost. We do not fall
2525 * through and boost by pageblock_nr_pages as failing
2526 * allocations that early means that reclaim is not going
2527 * to help and it may even be impossible to reclaim the
2528 * boosted watermark resulting in a hang.
2533 max_boost = max(pageblock_nr_pages, max_boost);
2535 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2542 * This function implements actual steal behaviour. If order is large enough,
2543 * we can steal whole pageblock. If not, we first move freepages in this
2544 * pageblock to our migratetype and determine how many already-allocated pages
2545 * are there in the pageblock with a compatible migratetype. If at least half
2546 * of pages are free or compatible, we can change migratetype of the pageblock
2547 * itself, so pages freed in the future will be put on the correct free list.
2549 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2550 unsigned int alloc_flags, int start_type, bool whole_block)
2552 unsigned int current_order = buddy_order(page);
2553 int free_pages, movable_pages, alike_pages;
2556 old_block_type = get_pageblock_migratetype(page);
2559 * This can happen due to races and we want to prevent broken
2560 * highatomic accounting.
2562 if (is_migrate_highatomic(old_block_type))
2565 /* Take ownership for orders >= pageblock_order */
2566 if (current_order >= pageblock_order) {
2567 change_pageblock_range(page, current_order, start_type);
2572 * Boost watermarks to increase reclaim pressure to reduce the
2573 * likelihood of future fallbacks. Wake kswapd now as the node
2574 * may be balanced overall and kswapd will not wake naturally.
2576 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2577 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2579 /* We are not allowed to try stealing from the whole block */
2583 free_pages = move_freepages_block(zone, page, start_type,
2586 * Determine how many pages are compatible with our allocation.
2587 * For movable allocation, it's the number of movable pages which
2588 * we just obtained. For other types it's a bit more tricky.
2590 if (start_type == MIGRATE_MOVABLE) {
2591 alike_pages = movable_pages;
2594 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2595 * to MOVABLE pageblock, consider all non-movable pages as
2596 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2597 * vice versa, be conservative since we can't distinguish the
2598 * exact migratetype of non-movable pages.
2600 if (old_block_type == MIGRATE_MOVABLE)
2601 alike_pages = pageblock_nr_pages
2602 - (free_pages + movable_pages);
2607 /* moving whole block can fail due to zone boundary conditions */
2612 * If a sufficient number of pages in the block are either free or of
2613 * comparable migratability as our allocation, claim the whole block.
2615 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2616 page_group_by_mobility_disabled)
2617 set_pageblock_migratetype(page, start_type);
2622 move_to_free_list(page, zone, current_order, start_type);
2626 * Check whether there is a suitable fallback freepage with requested order.
2627 * If only_stealable is true, this function returns fallback_mt only if
2628 * we can steal other freepages all together. This would help to reduce
2629 * fragmentation due to mixed migratetype pages in one pageblock.
2631 int find_suitable_fallback(struct free_area *area, unsigned int order,
2632 int migratetype, bool only_stealable, bool *can_steal)
2637 if (area->nr_free == 0)
2642 fallback_mt = fallbacks[migratetype][i];
2643 if (fallback_mt == MIGRATE_TYPES)
2646 if (free_area_empty(area, fallback_mt))
2649 if (can_steal_fallback(order, migratetype))
2652 if (!only_stealable)
2663 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2664 * there are no empty page blocks that contain a page with a suitable order
2666 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2667 unsigned int alloc_order)
2670 unsigned long max_managed, flags;
2673 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2674 * Check is race-prone but harmless.
2676 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2677 if (zone->nr_reserved_highatomic >= max_managed)
2680 spin_lock_irqsave(&zone->lock, flags);
2682 /* Recheck the nr_reserved_highatomic limit under the lock */
2683 if (zone->nr_reserved_highatomic >= max_managed)
2687 mt = get_pageblock_migratetype(page);
2688 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2689 && !is_migrate_cma(mt)) {
2690 zone->nr_reserved_highatomic += pageblock_nr_pages;
2691 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2692 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2696 spin_unlock_irqrestore(&zone->lock, flags);
2700 * Used when an allocation is about to fail under memory pressure. This
2701 * potentially hurts the reliability of high-order allocations when under
2702 * intense memory pressure but failed atomic allocations should be easier
2703 * to recover from than an OOM.
2705 * If @force is true, try to unreserve a pageblock even though highatomic
2706 * pageblock is exhausted.
2708 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2711 struct zonelist *zonelist = ac->zonelist;
2712 unsigned long flags;
2719 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2722 * Preserve at least one pageblock unless memory pressure
2725 if (!force && zone->nr_reserved_highatomic <=
2729 spin_lock_irqsave(&zone->lock, flags);
2730 for (order = 0; order < MAX_ORDER; order++) {
2731 struct free_area *area = &(zone->free_area[order]);
2733 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2738 * In page freeing path, migratetype change is racy so
2739 * we can counter several free pages in a pageblock
2740 * in this loop althoug we changed the pageblock type
2741 * from highatomic to ac->migratetype. So we should
2742 * adjust the count once.
2744 if (is_migrate_highatomic_page(page)) {
2746 * It should never happen but changes to
2747 * locking could inadvertently allow a per-cpu
2748 * drain to add pages to MIGRATE_HIGHATOMIC
2749 * while unreserving so be safe and watch for
2752 zone->nr_reserved_highatomic -= min(
2754 zone->nr_reserved_highatomic);
2758 * Convert to ac->migratetype and avoid the normal
2759 * pageblock stealing heuristics. Minimally, the caller
2760 * is doing the work and needs the pages. More
2761 * importantly, if the block was always converted to
2762 * MIGRATE_UNMOVABLE or another type then the number
2763 * of pageblocks that cannot be completely freed
2766 set_pageblock_migratetype(page, ac->migratetype);
2767 ret = move_freepages_block(zone, page, ac->migratetype,
2770 spin_unlock_irqrestore(&zone->lock, flags);
2774 spin_unlock_irqrestore(&zone->lock, flags);
2781 * Try finding a free buddy page on the fallback list and put it on the free
2782 * list of requested migratetype, possibly along with other pages from the same
2783 * block, depending on fragmentation avoidance heuristics. Returns true if
2784 * fallback was found so that __rmqueue_smallest() can grab it.
2786 * The use of signed ints for order and current_order is a deliberate
2787 * deviation from the rest of this file, to make the for loop
2788 * condition simpler.
2790 static __always_inline bool
2791 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2792 unsigned int alloc_flags)
2794 struct free_area *area;
2796 int min_order = order;
2802 * Do not steal pages from freelists belonging to other pageblocks
2803 * i.e. orders < pageblock_order. If there are no local zones free,
2804 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2806 if (alloc_flags & ALLOC_NOFRAGMENT)
2807 min_order = pageblock_order;
2810 * Find the largest available free page in the other list. This roughly
2811 * approximates finding the pageblock with the most free pages, which
2812 * would be too costly to do exactly.
2814 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2816 area = &(zone->free_area[current_order]);
2817 fallback_mt = find_suitable_fallback(area, current_order,
2818 start_migratetype, false, &can_steal);
2819 if (fallback_mt == -1)
2823 * We cannot steal all free pages from the pageblock and the
2824 * requested migratetype is movable. In that case it's better to
2825 * steal and split the smallest available page instead of the
2826 * largest available page, because even if the next movable
2827 * allocation falls back into a different pageblock than this
2828 * one, it won't cause permanent fragmentation.
2830 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2831 && current_order > order)
2840 for (current_order = order; current_order < MAX_ORDER;
2842 area = &(zone->free_area[current_order]);
2843 fallback_mt = find_suitable_fallback(area, current_order,
2844 start_migratetype, false, &can_steal);
2845 if (fallback_mt != -1)
2850 * This should not happen - we already found a suitable fallback
2851 * when looking for the largest page.
2853 VM_BUG_ON(current_order == MAX_ORDER);
2856 page = get_page_from_free_area(area, fallback_mt);
2858 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2861 trace_mm_page_alloc_extfrag(page, order, current_order,
2862 start_migratetype, fallback_mt);
2869 * Do the hard work of removing an element from the buddy allocator.
2870 * Call me with the zone->lock already held.
2872 static __always_inline struct page *
2873 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2874 unsigned int alloc_flags)
2878 if (IS_ENABLED(CONFIG_CMA)) {
2880 * Balance movable allocations between regular and CMA areas by
2881 * allocating from CMA when over half of the zone's free memory
2882 * is in the CMA area.
2884 if (alloc_flags & ALLOC_CMA &&
2885 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2886 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2887 page = __rmqueue_cma_fallback(zone, order);
2893 page = __rmqueue_smallest(zone, order, migratetype);
2894 if (unlikely(!page)) {
2895 if (alloc_flags & ALLOC_CMA)
2896 page = __rmqueue_cma_fallback(zone, order);
2898 if (!page && __rmqueue_fallback(zone, order, migratetype,
2904 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2909 * Obtain a specified number of elements from the buddy allocator, all under
2910 * a single hold of the lock, for efficiency. Add them to the supplied list.
2911 * Returns the number of new pages which were placed at *list.
2913 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2914 unsigned long count, struct list_head *list,
2915 int migratetype, unsigned int alloc_flags)
2919 spin_lock(&zone->lock);
2920 for (i = 0; i < count; ++i) {
2921 struct page *page = __rmqueue(zone, order, migratetype,
2923 if (unlikely(page == NULL))
2926 if (unlikely(check_pcp_refill(page)))
2930 * Split buddy pages returned by expand() are received here in
2931 * physical page order. The page is added to the tail of
2932 * caller's list. From the callers perspective, the linked list
2933 * is ordered by page number under some conditions. This is
2934 * useful for IO devices that can forward direction from the
2935 * head, thus also in the physical page order. This is useful
2936 * for IO devices that can merge IO requests if the physical
2937 * pages are ordered properly.
2939 list_add_tail(&page->lru, list);
2941 if (is_migrate_cma(get_pcppage_migratetype(page)))
2942 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2947 * i pages were removed from the buddy list even if some leak due
2948 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2949 * on i. Do not confuse with 'alloced' which is the number of
2950 * pages added to the pcp list.
2952 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2953 spin_unlock(&zone->lock);
2959 * Called from the vmstat counter updater to drain pagesets of this
2960 * currently executing processor on remote nodes after they have
2963 * Note that this function must be called with the thread pinned to
2964 * a single processor.
2966 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2968 unsigned long flags;
2969 int to_drain, batch;
2972 local_lock_irqsave(&pa_lock.l, flags);
2973 batch = READ_ONCE(pcp->batch);
2974 to_drain = min(pcp->count, batch);
2976 isolate_pcp_pages(to_drain, pcp, &dst);
2978 local_unlock_irqrestore(&pa_lock.l, flags);
2981 free_pcppages_bulk(zone, &dst, false);
2986 * Drain pcplists of the indicated processor and zone.
2988 * The processor must either be the current processor and the
2989 * thread pinned to the current processor or a processor that
2992 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2994 unsigned long flags;
2995 struct per_cpu_pageset *pset;
2996 struct per_cpu_pages *pcp;
3000 local_lock_irqsave(&pa_lock.l, flags);
3001 pset = per_cpu_ptr(zone->pageset, cpu);
3006 isolate_pcp_pages(count, pcp, &dst);
3008 local_unlock_irqrestore(&pa_lock.l, flags);
3011 free_pcppages_bulk(zone, &dst, false);
3015 * Drain pcplists of all zones on the indicated processor.
3017 * The processor must either be the current processor and the
3018 * thread pinned to the current processor or a processor that
3021 static void drain_pages(unsigned int cpu)
3025 for_each_populated_zone(zone) {
3026 drain_pages_zone(cpu, zone);
3031 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3033 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3034 * the single zone's pages.
3036 void drain_local_pages(struct zone *zone)
3038 int cpu = smp_processor_id();
3041 drain_pages_zone(cpu, zone);
3046 static void drain_local_pages_wq(struct work_struct *work)
3048 struct pcpu_drain *drain;
3050 drain = container_of(work, struct pcpu_drain, work);
3053 * drain_all_pages doesn't use proper cpu hotplug protection so
3054 * we can race with cpu offline when the WQ can move this from
3055 * a cpu pinned worker to an unbound one. We can operate on a different
3056 * cpu which is allright but we also have to make sure to not move to
3060 drain_local_pages(drain->zone);
3065 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3067 * When zone parameter is non-NULL, spill just the single zone's pages.
3069 * Note that this can be extremely slow as the draining happens in a workqueue.
3071 void drain_all_pages(struct zone *zone)
3076 * Allocate in the BSS so we wont require allocation in
3077 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3079 static cpumask_t cpus_with_pcps;
3082 * Make sure nobody triggers this path before mm_percpu_wq is fully
3085 if (WARN_ON_ONCE(!mm_percpu_wq))
3089 * Do not drain if one is already in progress unless it's specific to
3090 * a zone. Such callers are primarily CMA and memory hotplug and need
3091 * the drain to be complete when the call returns.
3093 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3096 mutex_lock(&pcpu_drain_mutex);
3100 * We don't care about racing with CPU hotplug event
3101 * as offline notification will cause the notified
3102 * cpu to drain that CPU pcps and on_each_cpu_mask
3103 * disables preemption as part of its processing
3105 for_each_online_cpu(cpu) {
3106 struct per_cpu_pageset *pcp;
3108 bool has_pcps = false;
3111 pcp = per_cpu_ptr(zone->pageset, cpu);
3115 for_each_populated_zone(z) {
3116 pcp = per_cpu_ptr(z->pageset, cpu);
3117 if (pcp->pcp.count) {
3125 cpumask_set_cpu(cpu, &cpus_with_pcps);
3127 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3130 for_each_cpu(cpu, &cpus_with_pcps) {
3131 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3134 INIT_WORK(&drain->work, drain_local_pages_wq);
3135 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3137 for_each_cpu(cpu, &cpus_with_pcps)
3138 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3140 mutex_unlock(&pcpu_drain_mutex);
3143 #ifdef CONFIG_HIBERNATION
3146 * Touch the watchdog for every WD_PAGE_COUNT pages.
3148 #define WD_PAGE_COUNT (128*1024)
3150 void mark_free_pages(struct zone *zone)
3152 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3153 unsigned long flags;
3154 unsigned int order, t;
3157 if (zone_is_empty(zone))
3160 spin_lock_irqsave(&zone->lock, flags);
3162 max_zone_pfn = zone_end_pfn(zone);
3163 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3164 if (pfn_valid(pfn)) {
3165 page = pfn_to_page(pfn);
3167 if (!--page_count) {
3168 touch_nmi_watchdog();
3169 page_count = WD_PAGE_COUNT;
3172 if (page_zone(page) != zone)
3175 if (!swsusp_page_is_forbidden(page))
3176 swsusp_unset_page_free(page);
3179 for_each_migratetype_order(order, t) {
3180 list_for_each_entry(page,
3181 &zone->free_area[order].free_list[t], lru) {
3184 pfn = page_to_pfn(page);
3185 for (i = 0; i < (1UL << order); i++) {
3186 if (!--page_count) {
3187 touch_nmi_watchdog();
3188 page_count = WD_PAGE_COUNT;
3190 swsusp_set_page_free(pfn_to_page(pfn + i));
3194 spin_unlock_irqrestore(&zone->lock, flags);
3196 #endif /* CONFIG_PM */
3198 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3202 if (!free_pcp_prepare(page))
3205 migratetype = get_pfnblock_migratetype(page, pfn);
3206 set_pcppage_migratetype(page, migratetype);
3210 static void free_unref_page_commit(struct page *page, unsigned long pfn,
3211 struct list_head *dst)
3213 struct zone *zone = page_zone(page);
3214 struct per_cpu_pages *pcp;
3217 migratetype = get_pcppage_migratetype(page);
3218 __count_vm_event(PGFREE);
3221 * We only track unmovable, reclaimable and movable on pcp lists.
3222 * Free ISOLATE pages back to the allocator because they are being
3223 * offlined but treat HIGHATOMIC as movable pages so we can get those
3224 * areas back if necessary. Otherwise, we may have to free
3225 * excessively into the page allocator
3227 if (migratetype >= MIGRATE_PCPTYPES) {
3228 if (unlikely(is_migrate_isolate(migratetype))) {
3229 free_one_page(zone, page, pfn, 0, migratetype,
3233 migratetype = MIGRATE_MOVABLE;
3236 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3237 list_add(&page->lru, &pcp->lists[migratetype]);
3239 if (pcp->count >= pcp->high) {
3240 unsigned long batch = READ_ONCE(pcp->batch);
3242 isolate_pcp_pages(batch, pcp, dst);
3247 * Free a 0-order page
3249 void free_unref_page(struct page *page)
3251 unsigned long flags;
3252 unsigned long pfn = page_to_pfn(page);
3253 struct zone *zone = page_zone(page);
3256 if (!free_unref_page_prepare(page, pfn))
3259 local_lock_irqsave(&pa_lock.l, flags);
3260 free_unref_page_commit(page, pfn, &dst);
3261 local_unlock_irqrestore(&pa_lock.l, flags);
3262 if (!list_empty(&dst))
3263 free_pcppages_bulk(zone, &dst, false);
3267 * Free a list of 0-order pages
3269 void free_unref_page_list(struct list_head *list)
3271 struct page *page, *next;
3272 unsigned long flags, pfn;
3273 int batch_count = 0;
3274 struct list_head dsts[__MAX_NR_ZONES];
3277 for (i = 0; i < __MAX_NR_ZONES; i++)
3278 INIT_LIST_HEAD(&dsts[i]);
3280 /* Prepare pages for freeing */
3281 list_for_each_entry_safe(page, next, list, lru) {
3282 pfn = page_to_pfn(page);
3283 if (!free_unref_page_prepare(page, pfn))
3284 list_del(&page->lru);
3285 set_page_private(page, pfn);
3288 local_lock_irqsave(&pa_lock.l, flags);
3289 list_for_each_entry_safe(page, next, list, lru) {
3290 unsigned long pfn = page_private(page);
3291 enum zone_type type;
3293 set_page_private(page, 0);
3294 trace_mm_page_free_batched(page);
3295 type = page_zonenum(page);
3296 free_unref_page_commit(page, pfn, &dsts[type]);
3299 * Guard against excessive IRQ disabled times when we get
3300 * a large list of pages to free.
3302 if (++batch_count == SWAP_CLUSTER_MAX) {
3303 local_unlock_irqrestore(&pa_lock.l, flags);
3305 local_lock_irqsave(&pa_lock.l, flags);
3308 local_unlock_irqrestore(&pa_lock.l, flags);
3310 for (i = 0; i < __MAX_NR_ZONES; ) {
3314 if (list_empty(&dsts[i])) {
3319 page = list_first_entry(&dsts[i], struct page, lru);
3320 zone = page_zone(page);
3322 free_pcppages_bulk(zone, &dsts[i], true);
3327 * split_page takes a non-compound higher-order page, and splits it into
3328 * n (1<<order) sub-pages: page[0..n]
3329 * Each sub-page must be freed individually.
3331 * Note: this is probably too low level an operation for use in drivers.
3332 * Please consult with lkml before using this in your driver.
3334 void split_page(struct page *page, unsigned int order)
3338 VM_BUG_ON_PAGE(PageCompound(page), page);
3339 VM_BUG_ON_PAGE(!page_count(page), page);
3341 for (i = 1; i < (1 << order); i++)
3342 set_page_refcounted(page + i);
3343 split_page_owner(page, 1 << order);
3344 split_page_memcg(page, 1 << order);
3346 EXPORT_SYMBOL_GPL(split_page);
3348 int __isolate_free_page(struct page *page, unsigned int order)
3350 unsigned long watermark;
3354 BUG_ON(!PageBuddy(page));
3356 zone = page_zone(page);
3357 mt = get_pageblock_migratetype(page);
3359 if (!is_migrate_isolate(mt)) {
3361 * Obey watermarks as if the page was being allocated. We can
3362 * emulate a high-order watermark check with a raised order-0
3363 * watermark, because we already know our high-order page
3366 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3367 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3370 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3373 /* Remove page from free list */
3375 del_page_from_free_list(page, zone, order);
3378 * Set the pageblock if the isolated page is at least half of a
3381 if (order >= pageblock_order - 1) {
3382 struct page *endpage = page + (1 << order) - 1;
3383 for (; page < endpage; page += pageblock_nr_pages) {
3384 int mt = get_pageblock_migratetype(page);
3385 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3386 && !is_migrate_highatomic(mt))
3387 set_pageblock_migratetype(page,
3393 return 1UL << order;
3397 * __putback_isolated_page - Return a now-isolated page back where we got it
3398 * @page: Page that was isolated
3399 * @order: Order of the isolated page
3400 * @mt: The page's pageblock's migratetype
3402 * This function is meant to return a page pulled from the free lists via
3403 * __isolate_free_page back to the free lists they were pulled from.
3405 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3407 struct zone *zone = page_zone(page);
3409 /* zone lock should be held when this function is called */
3410 lockdep_assert_held(&zone->lock);
3412 /* Return isolated page to tail of freelist. */
3413 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3414 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3418 * Update NUMA hit/miss statistics
3420 * Must be called with interrupts disabled.
3422 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3425 enum numa_stat_item local_stat = NUMA_LOCAL;
3427 /* skip numa counters update if numa stats is disabled */
3428 if (!static_branch_likely(&vm_numa_stat_key))
3431 if (zone_to_nid(z) != numa_node_id())
3432 local_stat = NUMA_OTHER;
3434 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3435 __inc_numa_state(z, NUMA_HIT);
3437 __inc_numa_state(z, NUMA_MISS);
3438 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3440 __inc_numa_state(z, local_stat);
3444 /* Remove page from the per-cpu list, caller must protect the list */
3445 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3446 unsigned int alloc_flags,
3447 struct per_cpu_pages *pcp,
3448 struct list_head *list)
3453 if (list_empty(list)) {
3454 pcp->count += rmqueue_bulk(zone, 0,
3456 migratetype, alloc_flags);
3457 if (unlikely(list_empty(list)))
3461 page = list_first_entry(list, struct page, lru);
3462 list_del(&page->lru);
3464 } while (check_new_pcp(page));
3469 /* Lock and remove page from the per-cpu list */
3470 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3471 struct zone *zone, gfp_t gfp_flags,
3472 int migratetype, unsigned int alloc_flags)
3474 struct per_cpu_pages *pcp;
3475 struct list_head *list;
3477 unsigned long flags;
3479 local_lock_irqsave(&pa_lock.l, flags);
3480 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3481 list = &pcp->lists[migratetype];
3482 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3484 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3485 zone_statistics(preferred_zone, zone);
3487 local_unlock_irqrestore(&pa_lock.l, flags);
3492 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3495 struct page *rmqueue(struct zone *preferred_zone,
3496 struct zone *zone, unsigned int order,
3497 gfp_t gfp_flags, unsigned int alloc_flags,
3500 unsigned long flags;
3503 if (likely(order == 0)) {
3505 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3506 * we need to skip it when CMA area isn't allowed.
3508 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3509 migratetype != MIGRATE_MOVABLE) {
3510 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3511 migratetype, alloc_flags);
3517 * We most definitely don't want callers attempting to
3518 * allocate greater than order-1 page units with __GFP_NOFAIL.
3520 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3521 local_lock_irqsave(&pa_lock.l, flags);
3522 spin_lock(&zone->lock);
3527 * order-0 request can reach here when the pcplist is skipped
3528 * due to non-CMA allocation context. HIGHATOMIC area is
3529 * reserved for high-order atomic allocation, so order-0
3530 * request should skip it.
3532 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3533 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3535 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3538 page = __rmqueue(zone, order, migratetype, alloc_flags);
3539 } while (page && check_new_pages(page, order));
3540 spin_unlock(&zone->lock);
3543 __mod_zone_freepage_state(zone, -(1 << order),
3544 get_pcppage_migratetype(page));
3546 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3547 zone_statistics(preferred_zone, zone);
3548 local_unlock_irqrestore(&pa_lock.l, flags);
3551 /* Separate test+clear to avoid unnecessary atomics */
3552 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3553 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3554 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3557 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3561 local_unlock_irqrestore(&pa_lock.l, flags);
3565 #ifdef CONFIG_FAIL_PAGE_ALLOC
3568 struct fault_attr attr;
3570 bool ignore_gfp_highmem;
3571 bool ignore_gfp_reclaim;
3573 } fail_page_alloc = {
3574 .attr = FAULT_ATTR_INITIALIZER,
3575 .ignore_gfp_reclaim = true,
3576 .ignore_gfp_highmem = true,
3580 static int __init setup_fail_page_alloc(char *str)
3582 return setup_fault_attr(&fail_page_alloc.attr, str);
3584 __setup("fail_page_alloc=", setup_fail_page_alloc);
3586 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3588 if (order < fail_page_alloc.min_order)
3590 if (gfp_mask & __GFP_NOFAIL)
3592 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3594 if (fail_page_alloc.ignore_gfp_reclaim &&
3595 (gfp_mask & __GFP_DIRECT_RECLAIM))
3598 return should_fail(&fail_page_alloc.attr, 1 << order);
3601 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3603 static int __init fail_page_alloc_debugfs(void)
3605 umode_t mode = S_IFREG | 0600;
3608 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3609 &fail_page_alloc.attr);
3611 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3612 &fail_page_alloc.ignore_gfp_reclaim);
3613 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3614 &fail_page_alloc.ignore_gfp_highmem);
3615 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3620 late_initcall(fail_page_alloc_debugfs);
3622 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3624 #else /* CONFIG_FAIL_PAGE_ALLOC */
3626 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3631 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3633 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3635 return __should_fail_alloc_page(gfp_mask, order);
3637 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3639 static inline long __zone_watermark_unusable_free(struct zone *z,
3640 unsigned int order, unsigned int alloc_flags)
3642 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3643 long unusable_free = (1 << order) - 1;
3646 * If the caller does not have rights to ALLOC_HARDER then subtract
3647 * the high-atomic reserves. This will over-estimate the size of the
3648 * atomic reserve but it avoids a search.
3650 if (likely(!alloc_harder))
3651 unusable_free += z->nr_reserved_highatomic;
3654 /* If allocation can't use CMA areas don't use free CMA pages */
3655 if (!(alloc_flags & ALLOC_CMA))
3656 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3659 return unusable_free;
3663 * Return true if free base pages are above 'mark'. For high-order checks it
3664 * will return true of the order-0 watermark is reached and there is at least
3665 * one free page of a suitable size. Checking now avoids taking the zone lock
3666 * to check in the allocation paths if no pages are free.
3668 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3669 int highest_zoneidx, unsigned int alloc_flags,
3674 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3676 /* free_pages may go negative - that's OK */
3677 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3679 if (alloc_flags & ALLOC_HIGH)
3682 if (unlikely(alloc_harder)) {
3684 * OOM victims can try even harder than normal ALLOC_HARDER
3685 * users on the grounds that it's definitely going to be in
3686 * the exit path shortly and free memory. Any allocation it
3687 * makes during the free path will be small and short-lived.
3689 if (alloc_flags & ALLOC_OOM)
3696 * Check watermarks for an order-0 allocation request. If these
3697 * are not met, then a high-order request also cannot go ahead
3698 * even if a suitable page happened to be free.
3700 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3703 /* If this is an order-0 request then the watermark is fine */
3707 /* For a high-order request, check at least one suitable page is free */
3708 for (o = order; o < MAX_ORDER; o++) {
3709 struct free_area *area = &z->free_area[o];
3715 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3716 if (!free_area_empty(area, mt))
3721 if ((alloc_flags & ALLOC_CMA) &&
3722 !free_area_empty(area, MIGRATE_CMA)) {
3726 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3732 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3733 int highest_zoneidx, unsigned int alloc_flags)
3735 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3736 zone_page_state(z, NR_FREE_PAGES));
3739 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3740 unsigned long mark, int highest_zoneidx,
3741 unsigned int alloc_flags, gfp_t gfp_mask)
3745 free_pages = zone_page_state(z, NR_FREE_PAGES);
3748 * Fast check for order-0 only. If this fails then the reserves
3749 * need to be calculated.
3754 fast_free = free_pages;
3755 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3756 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3760 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3764 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3765 * when checking the min watermark. The min watermark is the
3766 * point where boosting is ignored so that kswapd is woken up
3767 * when below the low watermark.
3769 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3770 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3771 mark = z->_watermark[WMARK_MIN];
3772 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3773 alloc_flags, free_pages);
3779 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3780 unsigned long mark, int highest_zoneidx)
3782 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3784 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3785 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3787 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3792 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3794 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3795 node_reclaim_distance;
3797 #else /* CONFIG_NUMA */
3798 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3802 #endif /* CONFIG_NUMA */
3805 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3806 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3807 * premature use of a lower zone may cause lowmem pressure problems that
3808 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3809 * probably too small. It only makes sense to spread allocations to avoid
3810 * fragmentation between the Normal and DMA32 zones.
3812 static inline unsigned int
3813 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3815 unsigned int alloc_flags;
3818 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3821 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3823 #ifdef CONFIG_ZONE_DMA32
3827 if (zone_idx(zone) != ZONE_NORMAL)
3831 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3832 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3833 * on UMA that if Normal is populated then so is DMA32.
3835 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3836 if (nr_online_nodes > 1 && !populated_zone(--zone))
3839 alloc_flags |= ALLOC_NOFRAGMENT;
3840 #endif /* CONFIG_ZONE_DMA32 */
3844 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3845 unsigned int alloc_flags)
3848 unsigned int pflags = current->flags;
3850 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3851 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3852 alloc_flags |= ALLOC_CMA;
3859 * get_page_from_freelist goes through the zonelist trying to allocate
3862 static struct page *
3863 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3864 const struct alloc_context *ac)
3868 struct pglist_data *last_pgdat_dirty_limit = NULL;
3873 * Scan zonelist, looking for a zone with enough free.
3874 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3876 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3877 z = ac->preferred_zoneref;
3878 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3883 if (cpusets_enabled() &&
3884 (alloc_flags & ALLOC_CPUSET) &&
3885 !__cpuset_zone_allowed(zone, gfp_mask))
3888 * When allocating a page cache page for writing, we
3889 * want to get it from a node that is within its dirty
3890 * limit, such that no single node holds more than its
3891 * proportional share of globally allowed dirty pages.
3892 * The dirty limits take into account the node's
3893 * lowmem reserves and high watermark so that kswapd
3894 * should be able to balance it without having to
3895 * write pages from its LRU list.
3897 * XXX: For now, allow allocations to potentially
3898 * exceed the per-node dirty limit in the slowpath
3899 * (spread_dirty_pages unset) before going into reclaim,
3900 * which is important when on a NUMA setup the allowed
3901 * nodes are together not big enough to reach the
3902 * global limit. The proper fix for these situations
3903 * will require awareness of nodes in the
3904 * dirty-throttling and the flusher threads.
3906 if (ac->spread_dirty_pages) {
3907 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3910 if (!node_dirty_ok(zone->zone_pgdat)) {
3911 last_pgdat_dirty_limit = zone->zone_pgdat;
3916 if (no_fallback && nr_online_nodes > 1 &&
3917 zone != ac->preferred_zoneref->zone) {
3921 * If moving to a remote node, retry but allow
3922 * fragmenting fallbacks. Locality is more important
3923 * than fragmentation avoidance.
3925 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3926 if (zone_to_nid(zone) != local_nid) {
3927 alloc_flags &= ~ALLOC_NOFRAGMENT;
3932 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3933 if (!zone_watermark_fast(zone, order, mark,
3934 ac->highest_zoneidx, alloc_flags,
3938 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3940 * Watermark failed for this zone, but see if we can
3941 * grow this zone if it contains deferred pages.
3943 if (static_branch_unlikely(&deferred_pages)) {
3944 if (_deferred_grow_zone(zone, order))
3948 /* Checked here to keep the fast path fast */
3949 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3950 if (alloc_flags & ALLOC_NO_WATERMARKS)
3953 if (node_reclaim_mode == 0 ||
3954 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3957 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3959 case NODE_RECLAIM_NOSCAN:
3962 case NODE_RECLAIM_FULL:
3963 /* scanned but unreclaimable */
3966 /* did we reclaim enough */
3967 if (zone_watermark_ok(zone, order, mark,
3968 ac->highest_zoneidx, alloc_flags))
3976 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3977 gfp_mask, alloc_flags, ac->migratetype);
3979 prep_new_page(page, order, gfp_mask, alloc_flags);
3982 * If this is a high-order atomic allocation then check
3983 * if the pageblock should be reserved for the future
3985 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3986 reserve_highatomic_pageblock(page, zone, order);
3990 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3991 /* Try again if zone has deferred pages */
3992 if (static_branch_unlikely(&deferred_pages)) {
3993 if (_deferred_grow_zone(zone, order))
4001 * It's possible on a UMA machine to get through all zones that are
4002 * fragmented. If avoiding fragmentation, reset and try again.
4005 alloc_flags &= ~ALLOC_NOFRAGMENT;
4012 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4014 unsigned int filter = SHOW_MEM_FILTER_NODES;
4017 * This documents exceptions given to allocations in certain
4018 * contexts that are allowed to allocate outside current's set
4021 if (!(gfp_mask & __GFP_NOMEMALLOC))
4022 if (tsk_is_oom_victim(current) ||
4023 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4024 filter &= ~SHOW_MEM_FILTER_NODES;
4025 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4026 filter &= ~SHOW_MEM_FILTER_NODES;
4028 show_mem(filter, nodemask);
4031 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4033 struct va_format vaf;
4035 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4037 if ((gfp_mask & __GFP_NOWARN) ||
4038 !__ratelimit(&nopage_rs) ||
4039 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4042 va_start(args, fmt);
4045 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4046 current->comm, &vaf, gfp_mask, &gfp_mask,
4047 nodemask_pr_args(nodemask));
4050 cpuset_print_current_mems_allowed();
4053 warn_alloc_show_mem(gfp_mask, nodemask);
4056 static inline struct page *
4057 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4058 unsigned int alloc_flags,
4059 const struct alloc_context *ac)
4063 page = get_page_from_freelist(gfp_mask, order,
4064 alloc_flags|ALLOC_CPUSET, ac);
4066 * fallback to ignore cpuset restriction if our nodes
4070 page = get_page_from_freelist(gfp_mask, order,
4076 static inline struct page *
4077 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4078 const struct alloc_context *ac, unsigned long *did_some_progress)
4080 struct oom_control oc = {
4081 .zonelist = ac->zonelist,
4082 .nodemask = ac->nodemask,
4084 .gfp_mask = gfp_mask,
4089 *did_some_progress = 0;
4092 * Acquire the oom lock. If that fails, somebody else is
4093 * making progress for us.
4095 if (!mutex_trylock(&oom_lock)) {
4096 *did_some_progress = 1;
4097 schedule_timeout_uninterruptible(1);
4102 * Go through the zonelist yet one more time, keep very high watermark
4103 * here, this is only to catch a parallel oom killing, we must fail if
4104 * we're still under heavy pressure. But make sure that this reclaim
4105 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4106 * allocation which will never fail due to oom_lock already held.
4108 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4109 ~__GFP_DIRECT_RECLAIM, order,
4110 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4114 /* Coredumps can quickly deplete all memory reserves */
4115 if (current->flags & PF_DUMPCORE)
4117 /* The OOM killer will not help higher order allocs */
4118 if (order > PAGE_ALLOC_COSTLY_ORDER)
4121 * We have already exhausted all our reclaim opportunities without any
4122 * success so it is time to admit defeat. We will skip the OOM killer
4123 * because it is very likely that the caller has a more reasonable
4124 * fallback than shooting a random task.
4126 * The OOM killer may not free memory on a specific node.
4128 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4130 /* The OOM killer does not needlessly kill tasks for lowmem */
4131 if (ac->highest_zoneidx < ZONE_NORMAL)
4133 if (pm_suspended_storage())
4136 * XXX: GFP_NOFS allocations should rather fail than rely on
4137 * other request to make a forward progress.
4138 * We are in an unfortunate situation where out_of_memory cannot
4139 * do much for this context but let's try it to at least get
4140 * access to memory reserved if the current task is killed (see
4141 * out_of_memory). Once filesystems are ready to handle allocation
4142 * failures more gracefully we should just bail out here.
4145 /* Exhausted what can be done so it's blame time */
4146 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4147 *did_some_progress = 1;
4150 * Help non-failing allocations by giving them access to memory
4153 if (gfp_mask & __GFP_NOFAIL)
4154 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4155 ALLOC_NO_WATERMARKS, ac);
4158 mutex_unlock(&oom_lock);
4163 * Maximum number of compaction retries wit a progress before OOM
4164 * killer is consider as the only way to move forward.
4166 #define MAX_COMPACT_RETRIES 16
4168 #ifdef CONFIG_COMPACTION
4169 /* Try memory compaction for high-order allocations before reclaim */
4170 static struct page *
4171 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4172 unsigned int alloc_flags, const struct alloc_context *ac,
4173 enum compact_priority prio, enum compact_result *compact_result)
4175 struct page *page = NULL;
4176 unsigned long pflags;
4177 unsigned int noreclaim_flag;
4182 psi_memstall_enter(&pflags);
4183 noreclaim_flag = memalloc_noreclaim_save();
4185 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4188 memalloc_noreclaim_restore(noreclaim_flag);
4189 psi_memstall_leave(&pflags);
4192 * At least in one zone compaction wasn't deferred or skipped, so let's
4193 * count a compaction stall
4195 count_vm_event(COMPACTSTALL);
4197 /* Prep a captured page if available */
4199 prep_new_page(page, order, gfp_mask, alloc_flags);
4201 /* Try get a page from the freelist if available */
4203 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4206 struct zone *zone = page_zone(page);
4208 zone->compact_blockskip_flush = false;
4209 compaction_defer_reset(zone, order, true);
4210 count_vm_event(COMPACTSUCCESS);
4215 * It's bad if compaction run occurs and fails. The most likely reason
4216 * is that pages exist, but not enough to satisfy watermarks.
4218 count_vm_event(COMPACTFAIL);
4226 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4227 enum compact_result compact_result,
4228 enum compact_priority *compact_priority,
4229 int *compaction_retries)
4231 int max_retries = MAX_COMPACT_RETRIES;
4234 int retries = *compaction_retries;
4235 enum compact_priority priority = *compact_priority;
4240 if (compaction_made_progress(compact_result))
4241 (*compaction_retries)++;
4244 * compaction considers all the zone as desperately out of memory
4245 * so it doesn't really make much sense to retry except when the
4246 * failure could be caused by insufficient priority
4248 if (compaction_failed(compact_result))
4249 goto check_priority;
4252 * compaction was skipped because there are not enough order-0 pages
4253 * to work with, so we retry only if it looks like reclaim can help.
4255 if (compaction_needs_reclaim(compact_result)) {
4256 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4261 * make sure the compaction wasn't deferred or didn't bail out early
4262 * due to locks contention before we declare that we should give up.
4263 * But the next retry should use a higher priority if allowed, so
4264 * we don't just keep bailing out endlessly.
4266 if (compaction_withdrawn(compact_result)) {
4267 goto check_priority;
4271 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4272 * costly ones because they are de facto nofail and invoke OOM
4273 * killer to move on while costly can fail and users are ready
4274 * to cope with that. 1/4 retries is rather arbitrary but we
4275 * would need much more detailed feedback from compaction to
4276 * make a better decision.
4278 if (order > PAGE_ALLOC_COSTLY_ORDER)
4280 if (*compaction_retries <= max_retries) {
4286 * Make sure there are attempts at the highest priority if we exhausted
4287 * all retries or failed at the lower priorities.
4290 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4291 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4293 if (*compact_priority > min_priority) {
4294 (*compact_priority)--;
4295 *compaction_retries = 0;
4299 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4303 static inline struct page *
4304 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4305 unsigned int alloc_flags, const struct alloc_context *ac,
4306 enum compact_priority prio, enum compact_result *compact_result)
4308 *compact_result = COMPACT_SKIPPED;
4313 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4314 enum compact_result compact_result,
4315 enum compact_priority *compact_priority,
4316 int *compaction_retries)
4321 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4325 * There are setups with compaction disabled which would prefer to loop
4326 * inside the allocator rather than hit the oom killer prematurely.
4327 * Let's give them a good hope and keep retrying while the order-0
4328 * watermarks are OK.
4330 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4331 ac->highest_zoneidx, ac->nodemask) {
4332 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4333 ac->highest_zoneidx, alloc_flags))
4338 #endif /* CONFIG_COMPACTION */
4340 #ifdef CONFIG_LOCKDEP
4341 static struct lockdep_map __fs_reclaim_map =
4342 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4344 static bool __need_fs_reclaim(gfp_t gfp_mask)
4346 gfp_mask = current_gfp_context(gfp_mask);
4348 /* no reclaim without waiting on it */
4349 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4352 /* this guy won't enter reclaim */
4353 if (current->flags & PF_MEMALLOC)
4356 /* We're only interested __GFP_FS allocations for now */
4357 if (!(gfp_mask & __GFP_FS))
4360 if (gfp_mask & __GFP_NOLOCKDEP)
4366 void __fs_reclaim_acquire(void)
4368 lock_map_acquire(&__fs_reclaim_map);
4371 void __fs_reclaim_release(void)
4373 lock_map_release(&__fs_reclaim_map);
4376 void fs_reclaim_acquire(gfp_t gfp_mask)
4378 if (__need_fs_reclaim(gfp_mask))
4379 __fs_reclaim_acquire();
4381 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4383 void fs_reclaim_release(gfp_t gfp_mask)
4385 if (__need_fs_reclaim(gfp_mask))
4386 __fs_reclaim_release();
4388 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4391 /* Perform direct synchronous page reclaim */
4392 static unsigned long
4393 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4394 const struct alloc_context *ac)
4396 unsigned int noreclaim_flag;
4397 unsigned long pflags, progress;
4401 /* We now go into synchronous reclaim */
4402 cpuset_memory_pressure_bump();
4403 psi_memstall_enter(&pflags);
4404 fs_reclaim_acquire(gfp_mask);
4405 noreclaim_flag = memalloc_noreclaim_save();
4407 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4410 memalloc_noreclaim_restore(noreclaim_flag);
4411 fs_reclaim_release(gfp_mask);
4412 psi_memstall_leave(&pflags);
4419 /* The really slow allocator path where we enter direct reclaim */
4420 static inline struct page *
4421 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4422 unsigned int alloc_flags, const struct alloc_context *ac,
4423 unsigned long *did_some_progress)
4425 struct page *page = NULL;
4426 bool drained = false;
4428 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4429 if (unlikely(!(*did_some_progress)))
4433 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4436 * If an allocation failed after direct reclaim, it could be because
4437 * pages are pinned on the per-cpu lists or in high alloc reserves.
4438 * Shrink them and try again
4440 if (!page && !drained) {
4441 unreserve_highatomic_pageblock(ac, false);
4442 drain_all_pages(NULL);
4450 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4451 const struct alloc_context *ac)
4455 pg_data_t *last_pgdat = NULL;
4456 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4458 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4460 if (last_pgdat != zone->zone_pgdat)
4461 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4462 last_pgdat = zone->zone_pgdat;
4466 static inline unsigned int
4467 gfp_to_alloc_flags(gfp_t gfp_mask)
4469 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4472 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4473 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4474 * to save two branches.
4476 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4477 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4480 * The caller may dip into page reserves a bit more if the caller
4481 * cannot run direct reclaim, or if the caller has realtime scheduling
4482 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4483 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4485 alloc_flags |= (__force int)
4486 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4488 if (gfp_mask & __GFP_ATOMIC) {
4490 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4491 * if it can't schedule.
4493 if (!(gfp_mask & __GFP_NOMEMALLOC))
4494 alloc_flags |= ALLOC_HARDER;
4496 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4497 * comment for __cpuset_node_allowed().
4499 alloc_flags &= ~ALLOC_CPUSET;
4500 } else if (unlikely(rt_task(current)) && !in_interrupt())
4501 alloc_flags |= ALLOC_HARDER;
4503 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4508 static bool oom_reserves_allowed(struct task_struct *tsk)
4510 if (!tsk_is_oom_victim(tsk))
4514 * !MMU doesn't have oom reaper so give access to memory reserves
4515 * only to the thread with TIF_MEMDIE set
4517 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4524 * Distinguish requests which really need access to full memory
4525 * reserves from oom victims which can live with a portion of it
4527 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4529 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4531 if (gfp_mask & __GFP_MEMALLOC)
4532 return ALLOC_NO_WATERMARKS;
4533 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4534 return ALLOC_NO_WATERMARKS;
4535 if (!in_interrupt()) {
4536 if (current->flags & PF_MEMALLOC)
4537 return ALLOC_NO_WATERMARKS;
4538 else if (oom_reserves_allowed(current))
4545 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4547 return !!__gfp_pfmemalloc_flags(gfp_mask);
4551 * Checks whether it makes sense to retry the reclaim to make a forward progress
4552 * for the given allocation request.
4554 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4555 * without success, or when we couldn't even meet the watermark if we
4556 * reclaimed all remaining pages on the LRU lists.
4558 * Returns true if a retry is viable or false to enter the oom path.
4561 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4562 struct alloc_context *ac, int alloc_flags,
4563 bool did_some_progress, int *no_progress_loops)
4570 * Costly allocations might have made a progress but this doesn't mean
4571 * their order will become available due to high fragmentation so
4572 * always increment the no progress counter for them
4574 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4575 *no_progress_loops = 0;
4577 (*no_progress_loops)++;
4580 * Make sure we converge to OOM if we cannot make any progress
4581 * several times in the row.
4583 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4584 /* Before OOM, exhaust highatomic_reserve */
4585 return unreserve_highatomic_pageblock(ac, true);
4589 * Keep reclaiming pages while there is a chance this will lead
4590 * somewhere. If none of the target zones can satisfy our allocation
4591 * request even if all reclaimable pages are considered then we are
4592 * screwed and have to go OOM.
4594 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4595 ac->highest_zoneidx, ac->nodemask) {
4596 unsigned long available;
4597 unsigned long reclaimable;
4598 unsigned long min_wmark = min_wmark_pages(zone);
4601 available = reclaimable = zone_reclaimable_pages(zone);
4602 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4605 * Would the allocation succeed if we reclaimed all
4606 * reclaimable pages?
4608 wmark = __zone_watermark_ok(zone, order, min_wmark,
4609 ac->highest_zoneidx, alloc_flags, available);
4610 trace_reclaim_retry_zone(z, order, reclaimable,
4611 available, min_wmark, *no_progress_loops, wmark);
4614 * If we didn't make any progress and have a lot of
4615 * dirty + writeback pages then we should wait for
4616 * an IO to complete to slow down the reclaim and
4617 * prevent from pre mature OOM
4619 if (!did_some_progress) {
4620 unsigned long write_pending;
4622 write_pending = zone_page_state_snapshot(zone,
4623 NR_ZONE_WRITE_PENDING);
4625 if (2 * write_pending > reclaimable) {
4626 congestion_wait(BLK_RW_ASYNC, HZ/10);
4638 * Memory allocation/reclaim might be called from a WQ context and the
4639 * current implementation of the WQ concurrency control doesn't
4640 * recognize that a particular WQ is congested if the worker thread is
4641 * looping without ever sleeping. Therefore we have to do a short sleep
4642 * here rather than calling cond_resched().
4644 if (current->flags & PF_WQ_WORKER)
4645 schedule_timeout_uninterruptible(1);
4652 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4655 * It's possible that cpuset's mems_allowed and the nodemask from
4656 * mempolicy don't intersect. This should be normally dealt with by
4657 * policy_nodemask(), but it's possible to race with cpuset update in
4658 * such a way the check therein was true, and then it became false
4659 * before we got our cpuset_mems_cookie here.
4660 * This assumes that for all allocations, ac->nodemask can come only
4661 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4662 * when it does not intersect with the cpuset restrictions) or the
4663 * caller can deal with a violated nodemask.
4665 if (cpusets_enabled() && ac->nodemask &&
4666 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4667 ac->nodemask = NULL;
4672 * When updating a task's mems_allowed or mempolicy nodemask, it is
4673 * possible to race with parallel threads in such a way that our
4674 * allocation can fail while the mask is being updated. If we are about
4675 * to fail, check if the cpuset changed during allocation and if so,
4678 if (read_mems_allowed_retry(cpuset_mems_cookie))
4684 static inline struct page *
4685 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4686 struct alloc_context *ac)
4688 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4689 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4690 struct page *page = NULL;
4691 unsigned int alloc_flags;
4692 unsigned long did_some_progress;
4693 enum compact_priority compact_priority;
4694 enum compact_result compact_result;
4695 int compaction_retries;
4696 int no_progress_loops;
4697 unsigned int cpuset_mems_cookie;
4701 * We also sanity check to catch abuse of atomic reserves being used by
4702 * callers that are not in atomic context.
4704 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4705 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4706 gfp_mask &= ~__GFP_ATOMIC;
4709 compaction_retries = 0;
4710 no_progress_loops = 0;
4711 compact_priority = DEF_COMPACT_PRIORITY;
4712 cpuset_mems_cookie = read_mems_allowed_begin();
4715 * The fast path uses conservative alloc_flags to succeed only until
4716 * kswapd needs to be woken up, and to avoid the cost of setting up
4717 * alloc_flags precisely. So we do that now.
4719 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4722 * We need to recalculate the starting point for the zonelist iterator
4723 * because we might have used different nodemask in the fast path, or
4724 * there was a cpuset modification and we are retrying - otherwise we
4725 * could end up iterating over non-eligible zones endlessly.
4727 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4728 ac->highest_zoneidx, ac->nodemask);
4729 if (!ac->preferred_zoneref->zone)
4732 if (alloc_flags & ALLOC_KSWAPD)
4733 wake_all_kswapds(order, gfp_mask, ac);
4736 * The adjusted alloc_flags might result in immediate success, so try
4739 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4744 * For costly allocations, try direct compaction first, as it's likely
4745 * that we have enough base pages and don't need to reclaim. For non-
4746 * movable high-order allocations, do that as well, as compaction will
4747 * try prevent permanent fragmentation by migrating from blocks of the
4749 * Don't try this for allocations that are allowed to ignore
4750 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4752 if (can_direct_reclaim &&
4754 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4755 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4756 page = __alloc_pages_direct_compact(gfp_mask, order,
4758 INIT_COMPACT_PRIORITY,
4764 * Checks for costly allocations with __GFP_NORETRY, which
4765 * includes some THP page fault allocations
4767 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4769 * If allocating entire pageblock(s) and compaction
4770 * failed because all zones are below low watermarks
4771 * or is prohibited because it recently failed at this
4772 * order, fail immediately unless the allocator has
4773 * requested compaction and reclaim retry.
4776 * - potentially very expensive because zones are far
4777 * below their low watermarks or this is part of very
4778 * bursty high order allocations,
4779 * - not guaranteed to help because isolate_freepages()
4780 * may not iterate over freed pages as part of its
4782 * - unlikely to make entire pageblocks free on its
4785 if (compact_result == COMPACT_SKIPPED ||
4786 compact_result == COMPACT_DEFERRED)
4790 * Looks like reclaim/compaction is worth trying, but
4791 * sync compaction could be very expensive, so keep
4792 * using async compaction.
4794 compact_priority = INIT_COMPACT_PRIORITY;
4799 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4800 if (alloc_flags & ALLOC_KSWAPD)
4801 wake_all_kswapds(order, gfp_mask, ac);
4803 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4805 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4808 * Reset the nodemask and zonelist iterators if memory policies can be
4809 * ignored. These allocations are high priority and system rather than
4812 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4813 ac->nodemask = NULL;
4814 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4815 ac->highest_zoneidx, ac->nodemask);
4818 /* Attempt with potentially adjusted zonelist and alloc_flags */
4819 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4823 /* Caller is not willing to reclaim, we can't balance anything */
4824 if (!can_direct_reclaim)
4827 /* Avoid recursion of direct reclaim */
4828 if (current->flags & PF_MEMALLOC)
4831 /* Try direct reclaim and then allocating */
4832 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4833 &did_some_progress);
4837 /* Try direct compaction and then allocating */
4838 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4839 compact_priority, &compact_result);
4843 /* Do not loop if specifically requested */
4844 if (gfp_mask & __GFP_NORETRY)
4848 * Do not retry costly high order allocations unless they are
4849 * __GFP_RETRY_MAYFAIL
4851 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4854 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4855 did_some_progress > 0, &no_progress_loops))
4859 * It doesn't make any sense to retry for the compaction if the order-0
4860 * reclaim is not able to make any progress because the current
4861 * implementation of the compaction depends on the sufficient amount
4862 * of free memory (see __compaction_suitable)
4864 if (did_some_progress > 0 &&
4865 should_compact_retry(ac, order, alloc_flags,
4866 compact_result, &compact_priority,
4867 &compaction_retries))
4871 /* Deal with possible cpuset update races before we start OOM killing */
4872 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4875 /* Reclaim has failed us, start killing things */
4876 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4880 /* Avoid allocations with no watermarks from looping endlessly */
4881 if (tsk_is_oom_victim(current) &&
4882 (alloc_flags & ALLOC_OOM ||
4883 (gfp_mask & __GFP_NOMEMALLOC)))
4886 /* Retry as long as the OOM killer is making progress */
4887 if (did_some_progress) {
4888 no_progress_loops = 0;
4893 /* Deal with possible cpuset update races before we fail */
4894 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4898 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4901 if (gfp_mask & __GFP_NOFAIL) {
4903 * All existing users of the __GFP_NOFAIL are blockable, so warn
4904 * of any new users that actually require GFP_NOWAIT
4906 if (WARN_ON_ONCE(!can_direct_reclaim))
4910 * PF_MEMALLOC request from this context is rather bizarre
4911 * because we cannot reclaim anything and only can loop waiting
4912 * for somebody to do a work for us
4914 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4917 * non failing costly orders are a hard requirement which we
4918 * are not prepared for much so let's warn about these users
4919 * so that we can identify them and convert them to something
4922 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4925 * Help non-failing allocations by giving them access to memory
4926 * reserves but do not use ALLOC_NO_WATERMARKS because this
4927 * could deplete whole memory reserves which would just make
4928 * the situation worse
4930 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4938 warn_alloc(gfp_mask, ac->nodemask,
4939 "page allocation failure: order:%u", order);
4944 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4945 int preferred_nid, nodemask_t *nodemask,
4946 struct alloc_context *ac, gfp_t *alloc_mask,
4947 unsigned int *alloc_flags)
4949 ac->highest_zoneidx = gfp_zone(gfp_mask);
4950 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4951 ac->nodemask = nodemask;
4952 ac->migratetype = gfp_migratetype(gfp_mask);
4954 if (cpusets_enabled()) {
4955 *alloc_mask |= __GFP_HARDWALL;
4957 * When we are in the interrupt context, it is irrelevant
4958 * to the current task context. It means that any node ok.
4960 if (!in_interrupt() && !ac->nodemask)
4961 ac->nodemask = &cpuset_current_mems_allowed;
4963 *alloc_flags |= ALLOC_CPUSET;
4966 fs_reclaim_acquire(gfp_mask);
4967 fs_reclaim_release(gfp_mask);
4969 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4971 if (should_fail_alloc_page(gfp_mask, order))
4974 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4976 /* Dirty zone balancing only done in the fast path */
4977 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4980 * The preferred zone is used for statistics but crucially it is
4981 * also used as the starting point for the zonelist iterator. It
4982 * may get reset for allocations that ignore memory policies.
4984 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4985 ac->highest_zoneidx, ac->nodemask);
4991 * This is the 'heart' of the zoned buddy allocator.
4994 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4995 nodemask_t *nodemask)
4998 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4999 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
5000 struct alloc_context ac = { };
5003 * There are several places where we assume that the order value is sane
5004 * so bail out early if the request is out of bound.
5006 if (unlikely(order >= MAX_ORDER)) {
5007 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
5011 gfp_mask &= gfp_allowed_mask;
5012 alloc_mask = gfp_mask;
5013 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
5017 * Forbid the first pass from falling back to types that fragment
5018 * memory until all local zones are considered.
5020 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
5022 /* First allocation attempt */
5023 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
5028 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5029 * resp. GFP_NOIO which has to be inherited for all allocation requests
5030 * from a particular context which has been marked by
5031 * memalloc_no{fs,io}_{save,restore}.
5033 alloc_mask = current_gfp_context(gfp_mask);
5034 ac.spread_dirty_pages = false;
5037 * Restore the original nodemask if it was potentially replaced with
5038 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5040 ac.nodemask = nodemask;
5042 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
5045 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
5046 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
5047 __free_pages(page, order);
5051 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
5055 EXPORT_SYMBOL(__alloc_pages_nodemask);
5058 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5059 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5060 * you need to access high mem.
5062 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5066 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5069 return (unsigned long) page_address(page);
5071 EXPORT_SYMBOL(__get_free_pages);
5073 unsigned long get_zeroed_page(gfp_t gfp_mask)
5075 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5077 EXPORT_SYMBOL(get_zeroed_page);
5079 static inline void free_the_page(struct page *page, unsigned int order)
5081 if (order == 0) /* Via pcp? */
5082 free_unref_page(page);
5084 __free_pages_ok(page, order, FPI_NONE);
5087 void __free_pages(struct page *page, unsigned int order)
5089 if (put_page_testzero(page))
5090 free_the_page(page, order);
5091 else if (!PageHead(page))
5093 free_the_page(page + (1 << order), order);
5095 EXPORT_SYMBOL(__free_pages);
5097 void free_pages(unsigned long addr, unsigned int order)
5100 VM_BUG_ON(!virt_addr_valid((void *)addr));
5101 __free_pages(virt_to_page((void *)addr), order);
5105 EXPORT_SYMBOL(free_pages);
5109 * An arbitrary-length arbitrary-offset area of memory which resides
5110 * within a 0 or higher order page. Multiple fragments within that page
5111 * are individually refcounted, in the page's reference counter.
5113 * The page_frag functions below provide a simple allocation framework for
5114 * page fragments. This is used by the network stack and network device
5115 * drivers to provide a backing region of memory for use as either an
5116 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5118 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5121 struct page *page = NULL;
5122 gfp_t gfp = gfp_mask;
5124 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5125 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5127 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5128 PAGE_FRAG_CACHE_MAX_ORDER);
5129 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5131 if (unlikely(!page))
5132 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5134 nc->va = page ? page_address(page) : NULL;
5139 void __page_frag_cache_drain(struct page *page, unsigned int count)
5141 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5143 if (page_ref_sub_and_test(page, count))
5144 free_the_page(page, compound_order(page));
5146 EXPORT_SYMBOL(__page_frag_cache_drain);
5148 void *page_frag_alloc(struct page_frag_cache *nc,
5149 unsigned int fragsz, gfp_t gfp_mask)
5151 unsigned int size = PAGE_SIZE;
5155 if (unlikely(!nc->va)) {
5157 page = __page_frag_cache_refill(nc, gfp_mask);
5161 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5162 /* if size can vary use size else just use PAGE_SIZE */
5165 /* Even if we own the page, we do not use atomic_set().
5166 * This would break get_page_unless_zero() users.
5168 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5170 /* reset page count bias and offset to start of new frag */
5171 nc->pfmemalloc = page_is_pfmemalloc(page);
5172 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5176 offset = nc->offset - fragsz;
5177 if (unlikely(offset < 0)) {
5178 page = virt_to_page(nc->va);
5180 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5183 if (unlikely(nc->pfmemalloc)) {
5184 free_the_page(page, compound_order(page));
5188 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5189 /* if size can vary use size else just use PAGE_SIZE */
5192 /* OK, page count is 0, we can safely set it */
5193 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5195 /* reset page count bias and offset to start of new frag */
5196 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5197 offset = size - fragsz;
5201 nc->offset = offset;
5203 return nc->va + offset;
5205 EXPORT_SYMBOL(page_frag_alloc);
5208 * Frees a page fragment allocated out of either a compound or order 0 page.
5210 void page_frag_free(void *addr)
5212 struct page *page = virt_to_head_page(addr);
5214 if (unlikely(put_page_testzero(page)))
5215 free_the_page(page, compound_order(page));
5217 EXPORT_SYMBOL(page_frag_free);
5219 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5223 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5224 unsigned long used = addr + PAGE_ALIGN(size);
5226 split_page(virt_to_page((void *)addr), order);
5227 while (used < alloc_end) {
5232 return (void *)addr;
5236 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5237 * @size: the number of bytes to allocate
5238 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5240 * This function is similar to alloc_pages(), except that it allocates the
5241 * minimum number of pages to satisfy the request. alloc_pages() can only
5242 * allocate memory in power-of-two pages.
5244 * This function is also limited by MAX_ORDER.
5246 * Memory allocated by this function must be released by free_pages_exact().
5248 * Return: pointer to the allocated area or %NULL in case of error.
5250 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5252 unsigned int order = get_order(size);
5255 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5256 gfp_mask &= ~__GFP_COMP;
5258 addr = __get_free_pages(gfp_mask, order);
5259 return make_alloc_exact(addr, order, size);
5261 EXPORT_SYMBOL(alloc_pages_exact);
5264 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5266 * @nid: the preferred node ID where memory should be allocated
5267 * @size: the number of bytes to allocate
5268 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5270 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5273 * Return: pointer to the allocated area or %NULL in case of error.
5275 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5277 unsigned int order = get_order(size);
5280 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5281 gfp_mask &= ~__GFP_COMP;
5283 p = alloc_pages_node(nid, gfp_mask, order);
5286 return make_alloc_exact((unsigned long)page_address(p), order, size);
5290 * free_pages_exact - release memory allocated via alloc_pages_exact()
5291 * @virt: the value returned by alloc_pages_exact.
5292 * @size: size of allocation, same value as passed to alloc_pages_exact().
5294 * Release the memory allocated by a previous call to alloc_pages_exact.
5296 void free_pages_exact(void *virt, size_t size)
5298 unsigned long addr = (unsigned long)virt;
5299 unsigned long end = addr + PAGE_ALIGN(size);
5301 while (addr < end) {
5306 EXPORT_SYMBOL(free_pages_exact);
5309 * nr_free_zone_pages - count number of pages beyond high watermark
5310 * @offset: The zone index of the highest zone
5312 * nr_free_zone_pages() counts the number of pages which are beyond the
5313 * high watermark within all zones at or below a given zone index. For each
5314 * zone, the number of pages is calculated as:
5316 * nr_free_zone_pages = managed_pages - high_pages
5318 * Return: number of pages beyond high watermark.
5320 static unsigned long nr_free_zone_pages(int offset)
5325 /* Just pick one node, since fallback list is circular */
5326 unsigned long sum = 0;
5328 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5330 for_each_zone_zonelist(zone, z, zonelist, offset) {
5331 unsigned long size = zone_managed_pages(zone);
5332 unsigned long high = high_wmark_pages(zone);
5341 * nr_free_buffer_pages - count number of pages beyond high watermark
5343 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5344 * watermark within ZONE_DMA and ZONE_NORMAL.
5346 * Return: number of pages beyond high watermark within ZONE_DMA and
5349 unsigned long nr_free_buffer_pages(void)
5351 return nr_free_zone_pages(gfp_zone(GFP_USER));
5353 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5355 static inline void show_node(struct zone *zone)
5357 if (IS_ENABLED(CONFIG_NUMA))
5358 printk("Node %d ", zone_to_nid(zone));
5361 long si_mem_available(void)
5364 unsigned long pagecache;
5365 unsigned long wmark_low = 0;
5366 unsigned long pages[NR_LRU_LISTS];
5367 unsigned long reclaimable;
5371 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5372 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5375 wmark_low += low_wmark_pages(zone);
5378 * Estimate the amount of memory available for userspace allocations,
5379 * without causing swapping.
5381 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5384 * Not all the page cache can be freed, otherwise the system will
5385 * start swapping. Assume at least half of the page cache, or the
5386 * low watermark worth of cache, needs to stay.
5388 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5389 pagecache -= min(pagecache / 2, wmark_low);
5390 available += pagecache;
5393 * Part of the reclaimable slab and other kernel memory consists of
5394 * items that are in use, and cannot be freed. Cap this estimate at the
5397 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5398 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5399 available += reclaimable - min(reclaimable / 2, wmark_low);
5405 EXPORT_SYMBOL_GPL(si_mem_available);
5407 void si_meminfo(struct sysinfo *val)
5409 val->totalram = totalram_pages();
5410 val->sharedram = global_node_page_state(NR_SHMEM);
5411 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5412 val->bufferram = nr_blockdev_pages();
5413 val->totalhigh = totalhigh_pages();
5414 val->freehigh = nr_free_highpages();
5415 val->mem_unit = PAGE_SIZE;
5418 EXPORT_SYMBOL(si_meminfo);
5421 void si_meminfo_node(struct sysinfo *val, int nid)
5423 int zone_type; /* needs to be signed */
5424 unsigned long managed_pages = 0;
5425 unsigned long managed_highpages = 0;
5426 unsigned long free_highpages = 0;
5427 pg_data_t *pgdat = NODE_DATA(nid);
5429 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5430 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5431 val->totalram = managed_pages;
5432 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5433 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5434 #ifdef CONFIG_HIGHMEM
5435 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5436 struct zone *zone = &pgdat->node_zones[zone_type];
5438 if (is_highmem(zone)) {
5439 managed_highpages += zone_managed_pages(zone);
5440 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5443 val->totalhigh = managed_highpages;
5444 val->freehigh = free_highpages;
5446 val->totalhigh = managed_highpages;
5447 val->freehigh = free_highpages;
5449 val->mem_unit = PAGE_SIZE;
5454 * Determine whether the node should be displayed or not, depending on whether
5455 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5457 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5459 if (!(flags & SHOW_MEM_FILTER_NODES))
5463 * no node mask - aka implicit memory numa policy. Do not bother with
5464 * the synchronization - read_mems_allowed_begin - because we do not
5465 * have to be precise here.
5468 nodemask = &cpuset_current_mems_allowed;
5470 return !node_isset(nid, *nodemask);
5473 #define K(x) ((x) << (PAGE_SHIFT-10))
5475 static void show_migration_types(unsigned char type)
5477 static const char types[MIGRATE_TYPES] = {
5478 [MIGRATE_UNMOVABLE] = 'U',
5479 [MIGRATE_MOVABLE] = 'M',
5480 [MIGRATE_RECLAIMABLE] = 'E',
5481 [MIGRATE_HIGHATOMIC] = 'H',
5483 [MIGRATE_CMA] = 'C',
5485 #ifdef CONFIG_MEMORY_ISOLATION
5486 [MIGRATE_ISOLATE] = 'I',
5489 char tmp[MIGRATE_TYPES + 1];
5493 for (i = 0; i < MIGRATE_TYPES; i++) {
5494 if (type & (1 << i))
5499 printk(KERN_CONT "(%s) ", tmp);
5503 * Show free area list (used inside shift_scroll-lock stuff)
5504 * We also calculate the percentage fragmentation. We do this by counting the
5505 * memory on each free list with the exception of the first item on the list.
5508 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5511 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5513 unsigned long free_pcp = 0;
5518 for_each_populated_zone(zone) {
5519 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5522 for_each_online_cpu(cpu)
5523 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5526 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5527 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5528 " unevictable:%lu dirty:%lu writeback:%lu\n"
5529 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5530 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5531 " free:%lu free_pcp:%lu free_cma:%lu\n",
5532 global_node_page_state(NR_ACTIVE_ANON),
5533 global_node_page_state(NR_INACTIVE_ANON),
5534 global_node_page_state(NR_ISOLATED_ANON),
5535 global_node_page_state(NR_ACTIVE_FILE),
5536 global_node_page_state(NR_INACTIVE_FILE),
5537 global_node_page_state(NR_ISOLATED_FILE),
5538 global_node_page_state(NR_UNEVICTABLE),
5539 global_node_page_state(NR_FILE_DIRTY),
5540 global_node_page_state(NR_WRITEBACK),
5541 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5542 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5543 global_node_page_state(NR_FILE_MAPPED),
5544 global_node_page_state(NR_SHMEM),
5545 global_zone_page_state(NR_PAGETABLE),
5546 global_zone_page_state(NR_BOUNCE),
5547 global_zone_page_state(NR_FREE_PAGES),
5549 global_zone_page_state(NR_FREE_CMA_PAGES));
5551 for_each_online_pgdat(pgdat) {
5552 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5556 " active_anon:%lukB"
5557 " inactive_anon:%lukB"
5558 " active_file:%lukB"
5559 " inactive_file:%lukB"
5560 " unevictable:%lukB"
5561 " isolated(anon):%lukB"
5562 " isolated(file):%lukB"
5567 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5569 #ifdef CONFIG_FINEGRAINED_THP
5570 " shmem_64kb_thp: %lukB"
5571 #endif /* CONFIG_FINEGRAINED_THP */
5572 " shmem_pmdmapped: %lukB"
5575 " writeback_tmp:%lukB"
5576 " kernel_stack:%lukB"
5577 #ifdef CONFIG_SHADOW_CALL_STACK
5578 " shadow_call_stack:%lukB"
5580 " all_unreclaimable? %s"
5583 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5584 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5585 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5586 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5587 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5588 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5589 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5590 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5591 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5592 K(node_page_state(pgdat, NR_WRITEBACK)),
5593 K(node_page_state(pgdat, NR_SHMEM)),
5594 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5595 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5596 #ifdef CONFIG_FINEGRAINED_THP
5597 K(node_page_state(pgdat, NR_SHMEM_64KB_THPS) * HPAGE_CONT_PTE_NR),
5598 #endif /* CONFIG_FINEGRAINED_THP */
5599 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5601 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5603 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5604 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5605 #ifdef CONFIG_SHADOW_CALL_STACK
5606 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5608 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5612 for_each_populated_zone(zone) {
5615 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5619 for_each_online_cpu(cpu)
5620 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5629 " reserved_highatomic:%luKB"
5630 " active_anon:%lukB"
5631 " inactive_anon:%lukB"
5632 " active_file:%lukB"
5633 " inactive_file:%lukB"
5634 " unevictable:%lukB"
5635 " writepending:%lukB"
5646 K(zone_page_state(zone, NR_FREE_PAGES)),
5647 K(min_wmark_pages(zone)),
5648 K(low_wmark_pages(zone)),
5649 K(high_wmark_pages(zone)),
5650 K(zone->nr_reserved_highatomic),
5651 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5652 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5653 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5654 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5655 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5656 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5657 K(zone->present_pages),
5658 K(zone_managed_pages(zone)),
5659 K(zone_page_state(zone, NR_MLOCK)),
5660 K(zone_page_state(zone, NR_PAGETABLE)),
5661 K(zone_page_state(zone, NR_BOUNCE)),
5663 K(this_cpu_read(zone->pageset->pcp.count)),
5664 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5665 printk("lowmem_reserve[]:");
5666 for (i = 0; i < MAX_NR_ZONES; i++)
5667 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5668 printk(KERN_CONT "\n");
5671 for_each_populated_zone(zone) {
5673 unsigned long nr[MAX_ORDER], flags, total = 0;
5674 unsigned char types[MAX_ORDER];
5676 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5679 printk(KERN_CONT "%s: ", zone->name);
5681 spin_lock_irqsave(&zone->lock, flags);
5682 for (order = 0; order < MAX_ORDER; order++) {
5683 struct free_area *area = &zone->free_area[order];
5686 nr[order] = area->nr_free;
5687 total += nr[order] << order;
5690 for (type = 0; type < MIGRATE_TYPES; type++) {
5691 if (!free_area_empty(area, type))
5692 types[order] |= 1 << type;
5695 spin_unlock_irqrestore(&zone->lock, flags);
5696 for (order = 0; order < MAX_ORDER; order++) {
5697 printk(KERN_CONT "%lu*%lukB ",
5698 nr[order], K(1UL) << order);
5700 show_migration_types(types[order]);
5702 printk(KERN_CONT "= %lukB\n", K(total));
5705 hugetlb_show_meminfo();
5707 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5709 show_swap_cache_info();
5712 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5714 zoneref->zone = zone;
5715 zoneref->zone_idx = zone_idx(zone);
5719 * Builds allocation fallback zone lists.
5721 * Add all populated zones of a node to the zonelist.
5723 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5726 enum zone_type zone_type = MAX_NR_ZONES;
5731 zone = pgdat->node_zones + zone_type;
5732 if (managed_zone(zone)) {
5733 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5734 check_highest_zone(zone_type);
5736 } while (zone_type);
5743 static int __parse_numa_zonelist_order(char *s)
5746 * We used to support different zonlists modes but they turned
5747 * out to be just not useful. Let's keep the warning in place
5748 * if somebody still use the cmd line parameter so that we do
5749 * not fail it silently
5751 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5752 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5758 char numa_zonelist_order[] = "Node";
5761 * sysctl handler for numa_zonelist_order
5763 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5764 void *buffer, size_t *length, loff_t *ppos)
5767 return __parse_numa_zonelist_order(buffer);
5768 return proc_dostring(table, write, buffer, length, ppos);
5772 #define MAX_NODE_LOAD (nr_online_nodes)
5773 static int node_load[MAX_NUMNODES];
5776 * find_next_best_node - find the next node that should appear in a given node's fallback list
5777 * @node: node whose fallback list we're appending
5778 * @used_node_mask: nodemask_t of already used nodes
5780 * We use a number of factors to determine which is the next node that should
5781 * appear on a given node's fallback list. The node should not have appeared
5782 * already in @node's fallback list, and it should be the next closest node
5783 * according to the distance array (which contains arbitrary distance values
5784 * from each node to each node in the system), and should also prefer nodes
5785 * with no CPUs, since presumably they'll have very little allocation pressure
5786 * on them otherwise.
5788 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5790 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5793 int min_val = INT_MAX;
5794 int best_node = NUMA_NO_NODE;
5796 /* Use the local node if we haven't already */
5797 if (!node_isset(node, *used_node_mask)) {
5798 node_set(node, *used_node_mask);
5802 for_each_node_state(n, N_MEMORY) {
5804 /* Don't want a node to appear more than once */
5805 if (node_isset(n, *used_node_mask))
5808 /* Use the distance array to find the distance */
5809 val = node_distance(node, n);
5811 /* Penalize nodes under us ("prefer the next node") */
5814 /* Give preference to headless and unused nodes */
5815 if (!cpumask_empty(cpumask_of_node(n)))
5816 val += PENALTY_FOR_NODE_WITH_CPUS;
5818 /* Slight preference for less loaded node */
5819 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5820 val += node_load[n];
5822 if (val < min_val) {
5829 node_set(best_node, *used_node_mask);
5836 * Build zonelists ordered by node and zones within node.
5837 * This results in maximum locality--normal zone overflows into local
5838 * DMA zone, if any--but risks exhausting DMA zone.
5840 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5843 struct zoneref *zonerefs;
5846 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5848 for (i = 0; i < nr_nodes; i++) {
5851 pg_data_t *node = NODE_DATA(node_order[i]);
5853 nr_zones = build_zonerefs_node(node, zonerefs);
5854 zonerefs += nr_zones;
5856 zonerefs->zone = NULL;
5857 zonerefs->zone_idx = 0;
5861 * Build gfp_thisnode zonelists
5863 static void build_thisnode_zonelists(pg_data_t *pgdat)
5865 struct zoneref *zonerefs;
5868 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5869 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5870 zonerefs += nr_zones;
5871 zonerefs->zone = NULL;
5872 zonerefs->zone_idx = 0;
5876 * Build zonelists ordered by zone and nodes within zones.
5877 * This results in conserving DMA zone[s] until all Normal memory is
5878 * exhausted, but results in overflowing to remote node while memory
5879 * may still exist in local DMA zone.
5882 static void build_zonelists(pg_data_t *pgdat)
5884 static int node_order[MAX_NUMNODES];
5885 int node, load, nr_nodes = 0;
5886 nodemask_t used_mask = NODE_MASK_NONE;
5887 int local_node, prev_node;
5889 /* NUMA-aware ordering of nodes */
5890 local_node = pgdat->node_id;
5891 load = nr_online_nodes;
5892 prev_node = local_node;
5894 memset(node_order, 0, sizeof(node_order));
5895 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5897 * We don't want to pressure a particular node.
5898 * So adding penalty to the first node in same
5899 * distance group to make it round-robin.
5901 if (node_distance(local_node, node) !=
5902 node_distance(local_node, prev_node))
5903 node_load[node] = load;
5905 node_order[nr_nodes++] = node;
5910 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5911 build_thisnode_zonelists(pgdat);
5914 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5916 * Return node id of node used for "local" allocations.
5917 * I.e., first node id of first zone in arg node's generic zonelist.
5918 * Used for initializing percpu 'numa_mem', which is used primarily
5919 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5921 int local_memory_node(int node)
5925 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5926 gfp_zone(GFP_KERNEL),
5928 return zone_to_nid(z->zone);
5932 static void setup_min_unmapped_ratio(void);
5933 static void setup_min_slab_ratio(void);
5934 #else /* CONFIG_NUMA */
5936 static void build_zonelists(pg_data_t *pgdat)
5938 int node, local_node;
5939 struct zoneref *zonerefs;
5942 local_node = pgdat->node_id;
5944 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5945 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5946 zonerefs += nr_zones;
5949 * Now we build the zonelist so that it contains the zones
5950 * of all the other nodes.
5951 * We don't want to pressure a particular node, so when
5952 * building the zones for node N, we make sure that the
5953 * zones coming right after the local ones are those from
5954 * node N+1 (modulo N)
5956 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5957 if (!node_online(node))
5959 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5960 zonerefs += nr_zones;
5962 for (node = 0; node < local_node; node++) {
5963 if (!node_online(node))
5965 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5966 zonerefs += nr_zones;
5969 zonerefs->zone = NULL;
5970 zonerefs->zone_idx = 0;
5973 #endif /* CONFIG_NUMA */
5976 * Boot pageset table. One per cpu which is going to be used for all
5977 * zones and all nodes. The parameters will be set in such a way
5978 * that an item put on a list will immediately be handed over to
5979 * the buddy list. This is safe since pageset manipulation is done
5980 * with interrupts disabled.
5982 * The boot_pagesets must be kept even after bootup is complete for
5983 * unused processors and/or zones. They do play a role for bootstrapping
5984 * hotplugged processors.
5986 * zoneinfo_show() and maybe other functions do
5987 * not check if the processor is online before following the pageset pointer.
5988 * Other parts of the kernel may not check if the zone is available.
5990 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5991 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5992 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5994 static void __build_all_zonelists(void *data)
5997 int __maybe_unused cpu;
5998 pg_data_t *self = data;
5999 static DEFINE_SPINLOCK(lock);
6004 memset(node_load, 0, sizeof(node_load));
6008 * This node is hotadded and no memory is yet present. So just
6009 * building zonelists is fine - no need to touch other nodes.
6011 if (self && !node_online(self->node_id)) {
6012 build_zonelists(self);
6014 for_each_online_node(nid) {
6015 pg_data_t *pgdat = NODE_DATA(nid);
6017 build_zonelists(pgdat);
6020 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6022 * We now know the "local memory node" for each node--
6023 * i.e., the node of the first zone in the generic zonelist.
6024 * Set up numa_mem percpu variable for on-line cpus. During
6025 * boot, only the boot cpu should be on-line; we'll init the
6026 * secondary cpus' numa_mem as they come on-line. During
6027 * node/memory hotplug, we'll fixup all on-line cpus.
6029 for_each_online_cpu(cpu)
6030 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6037 static noinline void __init
6038 build_all_zonelists_init(void)
6042 __build_all_zonelists(NULL);
6045 * Initialize the boot_pagesets that are going to be used
6046 * for bootstrapping processors. The real pagesets for
6047 * each zone will be allocated later when the per cpu
6048 * allocator is available.
6050 * boot_pagesets are used also for bootstrapping offline
6051 * cpus if the system is already booted because the pagesets
6052 * are needed to initialize allocators on a specific cpu too.
6053 * F.e. the percpu allocator needs the page allocator which
6054 * needs the percpu allocator in order to allocate its pagesets
6055 * (a chicken-egg dilemma).
6057 for_each_possible_cpu(cpu)
6058 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
6060 mminit_verify_zonelist();
6061 cpuset_init_current_mems_allowed();
6065 * unless system_state == SYSTEM_BOOTING.
6067 * __ref due to call of __init annotated helper build_all_zonelists_init
6068 * [protected by SYSTEM_BOOTING].
6070 void __ref build_all_zonelists(pg_data_t *pgdat)
6072 unsigned long vm_total_pages;
6074 if (system_state == SYSTEM_BOOTING) {
6075 build_all_zonelists_init();
6077 __build_all_zonelists(pgdat);
6078 /* cpuset refresh routine should be here */
6080 /* Get the number of free pages beyond high watermark in all zones. */
6081 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6083 * Disable grouping by mobility if the number of pages in the
6084 * system is too low to allow the mechanism to work. It would be
6085 * more accurate, but expensive to check per-zone. This check is
6086 * made on memory-hotadd so a system can start with mobility
6087 * disabled and enable it later
6089 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6090 page_group_by_mobility_disabled = 1;
6092 page_group_by_mobility_disabled = 0;
6094 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6096 page_group_by_mobility_disabled ? "off" : "on",
6099 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6103 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6104 static bool __meminit
6105 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6107 static struct memblock_region *r;
6109 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6110 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6111 for_each_mem_region(r) {
6112 if (*pfn < memblock_region_memory_end_pfn(r))
6116 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6117 memblock_is_mirror(r)) {
6118 *pfn = memblock_region_memory_end_pfn(r);
6126 * Initially all pages are reserved - free ones are freed
6127 * up by memblock_free_all() once the early boot process is
6128 * done. Non-atomic initialization, single-pass.
6130 * All aligned pageblocks are initialized to the specified migratetype
6131 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6132 * zone stats (e.g., nr_isolate_pageblock) are touched.
6134 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
6135 unsigned long start_pfn, unsigned long zone_end_pfn,
6136 enum meminit_context context,
6137 struct vmem_altmap *altmap, int migratetype)
6139 unsigned long pfn, end_pfn = start_pfn + size;
6142 if (highest_memmap_pfn < end_pfn - 1)
6143 highest_memmap_pfn = end_pfn - 1;
6145 #ifdef CONFIG_ZONE_DEVICE
6147 * Honor reservation requested by the driver for this ZONE_DEVICE
6148 * memory. We limit the total number of pages to initialize to just
6149 * those that might contain the memory mapping. We will defer the
6150 * ZONE_DEVICE page initialization until after we have released
6153 if (zone == ZONE_DEVICE) {
6157 if (start_pfn == altmap->base_pfn)
6158 start_pfn += altmap->reserve;
6159 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6163 for (pfn = start_pfn; pfn < end_pfn; ) {
6165 * There can be holes in boot-time mem_map[]s handed to this
6166 * function. They do not exist on hotplugged memory.
6168 if (context == MEMINIT_EARLY) {
6169 if (overlap_memmap_init(zone, &pfn))
6171 if (defer_init(nid, pfn, zone_end_pfn))
6175 page = pfn_to_page(pfn);
6176 __init_single_page(page, pfn, zone, nid);
6177 if (context == MEMINIT_HOTPLUG)
6178 __SetPageReserved(page);
6181 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6182 * such that unmovable allocations won't be scattered all
6183 * over the place during system boot.
6185 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6186 set_pageblock_migratetype(page, migratetype);
6193 #ifdef CONFIG_ZONE_DEVICE
6194 void __ref memmap_init_zone_device(struct zone *zone,
6195 unsigned long start_pfn,
6196 unsigned long nr_pages,
6197 struct dev_pagemap *pgmap)
6199 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6200 struct pglist_data *pgdat = zone->zone_pgdat;
6201 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6202 unsigned long zone_idx = zone_idx(zone);
6203 unsigned long start = jiffies;
6204 int nid = pgdat->node_id;
6206 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6210 * The call to memmap_init should have already taken care
6211 * of the pages reserved for the memmap, so we can just jump to
6212 * the end of that region and start processing the device pages.
6215 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6216 nr_pages = end_pfn - start_pfn;
6219 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6220 struct page *page = pfn_to_page(pfn);
6222 __init_single_page(page, pfn, zone_idx, nid);
6225 * Mark page reserved as it will need to wait for onlining
6226 * phase for it to be fully associated with a zone.
6228 * We can use the non-atomic __set_bit operation for setting
6229 * the flag as we are still initializing the pages.
6231 __SetPageReserved(page);
6234 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6235 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6236 * ever freed or placed on a driver-private list.
6238 page->pgmap = pgmap;
6239 page->zone_device_data = NULL;
6242 * Mark the block movable so that blocks are reserved for
6243 * movable at startup. This will force kernel allocations
6244 * to reserve their blocks rather than leaking throughout
6245 * the address space during boot when many long-lived
6246 * kernel allocations are made.
6248 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6249 * because this is done early in section_activate()
6251 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6252 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6257 pr_info("%s initialised %lu pages in %ums\n", __func__,
6258 nr_pages, jiffies_to_msecs(jiffies - start));
6262 static void __meminit zone_init_free_lists(struct zone *zone)
6264 unsigned int order, t;
6265 for_each_migratetype_order(order, t) {
6266 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6267 zone->free_area[order].nr_free = 0;
6271 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6273 * Only struct pages that correspond to ranges defined by memblock.memory
6274 * are zeroed and initialized by going through __init_single_page() during
6275 * memmap_init_zone_range().
6277 * But, there could be struct pages that correspond to holes in
6278 * memblock.memory. This can happen because of the following reasons:
6279 * - physical memory bank size is not necessarily the exact multiple of the
6280 * arbitrary section size
6281 * - early reserved memory may not be listed in memblock.memory
6282 * - memory layouts defined with memmap= kernel parameter may not align
6283 * nicely with memmap sections
6285 * Explicitly initialize those struct pages so that:
6286 * - PG_Reserved is set
6287 * - zone and node links point to zone and node that span the page if the
6288 * hole is in the middle of a zone
6289 * - zone and node links point to adjacent zone/node if the hole falls on
6290 * the zone boundary; the pages in such holes will be prepended to the
6291 * zone/node above the hole except for the trailing pages in the last
6292 * section that will be appended to the zone/node below.
6294 static void __init init_unavailable_range(unsigned long spfn,
6301 for (pfn = spfn; pfn < epfn; pfn++) {
6302 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6303 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6304 + pageblock_nr_pages - 1;
6307 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6308 __SetPageReserved(pfn_to_page(pfn));
6313 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6314 node, zone_names[zone], pgcnt);
6317 static inline void init_unavailable_range(unsigned long spfn,
6324 static void __init memmap_init_zone_range(struct zone *zone,
6325 unsigned long start_pfn,
6326 unsigned long end_pfn,
6327 unsigned long *hole_pfn)
6329 unsigned long zone_start_pfn = zone->zone_start_pfn;
6330 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6331 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6333 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6334 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6336 if (start_pfn >= end_pfn)
6339 memmap_init_zone(end_pfn - start_pfn, nid, zone_id, start_pfn,
6340 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6342 if (*hole_pfn < start_pfn)
6343 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6345 *hole_pfn = end_pfn;
6348 void __init __weak memmap_init(void)
6350 unsigned long start_pfn, end_pfn;
6351 unsigned long hole_pfn = 0;
6352 int i, j, zone_id, nid;
6354 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6355 struct pglist_data *node = NODE_DATA(nid);
6357 for (j = 0; j < MAX_NR_ZONES; j++) {
6358 struct zone *zone = node->node_zones + j;
6360 if (!populated_zone(zone))
6363 memmap_init_zone_range(zone, start_pfn, end_pfn,
6369 #ifdef CONFIG_SPARSEMEM
6371 * Initialize the memory map for hole in the range [memory_end,
6373 * Append the pages in this hole to the highest zone in the last
6375 * The call to init_unavailable_range() is outside the ifdef to
6376 * silence the compiler warining about zone_id set but not used;
6377 * for FLATMEM it is a nop anyway
6379 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6380 if (hole_pfn < end_pfn)
6382 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6385 /* A stub for backwards compatibility with custom implementatin on IA-64 */
6386 void __meminit __weak arch_memmap_init(unsigned long size, int nid,
6388 unsigned long range_start_pfn)
6392 static int zone_batchsize(struct zone *zone)
6398 * The per-cpu-pages pools are set to around 1000th of the
6401 batch = zone_managed_pages(zone) / 1024;
6402 /* But no more than a meg. */
6403 if (batch * PAGE_SIZE > 1024 * 1024)
6404 batch = (1024 * 1024) / PAGE_SIZE;
6405 batch /= 4; /* We effectively *= 4 below */
6410 * Clamp the batch to a 2^n - 1 value. Having a power
6411 * of 2 value was found to be more likely to have
6412 * suboptimal cache aliasing properties in some cases.
6414 * For example if 2 tasks are alternately allocating
6415 * batches of pages, one task can end up with a lot
6416 * of pages of one half of the possible page colors
6417 * and the other with pages of the other colors.
6419 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6424 /* The deferral and batching of frees should be suppressed under NOMMU
6427 * The problem is that NOMMU needs to be able to allocate large chunks
6428 * of contiguous memory as there's no hardware page translation to
6429 * assemble apparent contiguous memory from discontiguous pages.
6431 * Queueing large contiguous runs of pages for batching, however,
6432 * causes the pages to actually be freed in smaller chunks. As there
6433 * can be a significant delay between the individual batches being
6434 * recycled, this leads to the once large chunks of space being
6435 * fragmented and becoming unavailable for high-order allocations.
6442 * pcp->high and pcp->batch values are related and dependent on one another:
6443 * ->batch must never be higher then ->high.
6444 * The following function updates them in a safe manner without read side
6447 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6448 * those fields changing asynchronously (acording to the above rule).
6450 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6451 * outside of boot time (or some other assurance that no concurrent updaters
6454 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6455 unsigned long batch)
6457 /* start with a fail safe value for batch */
6461 /* Update high, then batch, in order */
6468 /* a companion to pageset_set_high() */
6469 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6471 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6474 static void pageset_init(struct per_cpu_pageset *p)
6476 struct per_cpu_pages *pcp;
6479 memset(p, 0, sizeof(*p));
6482 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6483 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6486 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6489 pageset_set_batch(p, batch);
6493 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6494 * to the value high for the pageset p.
6496 static void pageset_set_high(struct per_cpu_pageset *p,
6499 unsigned long batch = max(1UL, high / 4);
6500 if ((high / 4) > (PAGE_SHIFT * 8))
6501 batch = PAGE_SHIFT * 8;
6503 pageset_update(&p->pcp, high, batch);
6506 static void pageset_set_high_and_batch(struct zone *zone,
6507 struct per_cpu_pageset *pcp)
6509 if (percpu_pagelist_fraction)
6510 pageset_set_high(pcp,
6511 (zone_managed_pages(zone) /
6512 percpu_pagelist_fraction));
6514 pageset_set_batch(pcp, zone_batchsize(zone));
6517 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6519 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6522 pageset_set_high_and_batch(zone, pcp);
6525 void __meminit setup_zone_pageset(struct zone *zone)
6528 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6529 for_each_possible_cpu(cpu)
6530 zone_pageset_init(zone, cpu);
6534 * Allocate per cpu pagesets and initialize them.
6535 * Before this call only boot pagesets were available.
6537 void __init setup_per_cpu_pageset(void)
6539 struct pglist_data *pgdat;
6541 int __maybe_unused cpu;
6543 for_each_populated_zone(zone)
6544 setup_zone_pageset(zone);
6548 * Unpopulated zones continue using the boot pagesets.
6549 * The numa stats for these pagesets need to be reset.
6550 * Otherwise, they will end up skewing the stats of
6551 * the nodes these zones are associated with.
6553 for_each_possible_cpu(cpu) {
6554 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6555 memset(pcp->vm_numa_stat_diff, 0,
6556 sizeof(pcp->vm_numa_stat_diff));
6560 for_each_online_pgdat(pgdat)
6561 pgdat->per_cpu_nodestats =
6562 alloc_percpu(struct per_cpu_nodestat);
6565 static __meminit void zone_pcp_init(struct zone *zone)
6568 * per cpu subsystem is not up at this point. The following code
6569 * relies on the ability of the linker to provide the
6570 * offset of a (static) per cpu variable into the per cpu area.
6572 zone->pageset = &boot_pageset;
6574 if (populated_zone(zone))
6575 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6576 zone->name, zone->present_pages,
6577 zone_batchsize(zone));
6580 void __meminit init_currently_empty_zone(struct zone *zone,
6581 unsigned long zone_start_pfn,
6584 struct pglist_data *pgdat = zone->zone_pgdat;
6585 int zone_idx = zone_idx(zone) + 1;
6587 if (zone_idx > pgdat->nr_zones)
6588 pgdat->nr_zones = zone_idx;
6590 zone->zone_start_pfn = zone_start_pfn;
6592 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6593 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6595 (unsigned long)zone_idx(zone),
6596 zone_start_pfn, (zone_start_pfn + size));
6598 zone_init_free_lists(zone);
6599 zone->initialized = 1;
6603 * get_pfn_range_for_nid - Return the start and end page frames for a node
6604 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6605 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6606 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6608 * It returns the start and end page frame of a node based on information
6609 * provided by memblock_set_node(). If called for a node
6610 * with no available memory, a warning is printed and the start and end
6613 void __init get_pfn_range_for_nid(unsigned int nid,
6614 unsigned long *start_pfn, unsigned long *end_pfn)
6616 unsigned long this_start_pfn, this_end_pfn;
6622 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6623 *start_pfn = min(*start_pfn, this_start_pfn);
6624 *end_pfn = max(*end_pfn, this_end_pfn);
6627 if (*start_pfn == -1UL)
6632 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6633 * assumption is made that zones within a node are ordered in monotonic
6634 * increasing memory addresses so that the "highest" populated zone is used
6636 static void __init find_usable_zone_for_movable(void)
6639 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6640 if (zone_index == ZONE_MOVABLE)
6643 if (arch_zone_highest_possible_pfn[zone_index] >
6644 arch_zone_lowest_possible_pfn[zone_index])
6648 VM_BUG_ON(zone_index == -1);
6649 movable_zone = zone_index;
6653 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6654 * because it is sized independent of architecture. Unlike the other zones,
6655 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6656 * in each node depending on the size of each node and how evenly kernelcore
6657 * is distributed. This helper function adjusts the zone ranges
6658 * provided by the architecture for a given node by using the end of the
6659 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6660 * zones within a node are in order of monotonic increases memory addresses
6662 static void __init adjust_zone_range_for_zone_movable(int nid,
6663 unsigned long zone_type,
6664 unsigned long node_start_pfn,
6665 unsigned long node_end_pfn,
6666 unsigned long *zone_start_pfn,
6667 unsigned long *zone_end_pfn)
6669 /* Only adjust if ZONE_MOVABLE is on this node */
6670 if (zone_movable_pfn[nid]) {
6671 /* Size ZONE_MOVABLE */
6672 if (zone_type == ZONE_MOVABLE) {
6673 *zone_start_pfn = zone_movable_pfn[nid];
6674 *zone_end_pfn = min(node_end_pfn,
6675 arch_zone_highest_possible_pfn[movable_zone]);
6677 /* Adjust for ZONE_MOVABLE starting within this range */
6678 } else if (!mirrored_kernelcore &&
6679 *zone_start_pfn < zone_movable_pfn[nid] &&
6680 *zone_end_pfn > zone_movable_pfn[nid]) {
6681 *zone_end_pfn = zone_movable_pfn[nid];
6683 /* Check if this whole range is within ZONE_MOVABLE */
6684 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6685 *zone_start_pfn = *zone_end_pfn;
6690 * Return the number of pages a zone spans in a node, including holes
6691 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6693 static unsigned long __init zone_spanned_pages_in_node(int nid,
6694 unsigned long zone_type,
6695 unsigned long node_start_pfn,
6696 unsigned long node_end_pfn,
6697 unsigned long *zone_start_pfn,
6698 unsigned long *zone_end_pfn)
6700 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6701 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6702 /* When hotadd a new node from cpu_up(), the node should be empty */
6703 if (!node_start_pfn && !node_end_pfn)
6706 /* Get the start and end of the zone */
6707 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6708 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6709 adjust_zone_range_for_zone_movable(nid, zone_type,
6710 node_start_pfn, node_end_pfn,
6711 zone_start_pfn, zone_end_pfn);
6713 /* Check that this node has pages within the zone's required range */
6714 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6717 /* Move the zone boundaries inside the node if necessary */
6718 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6719 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6721 /* Return the spanned pages */
6722 return *zone_end_pfn - *zone_start_pfn;
6726 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6727 * then all holes in the requested range will be accounted for.
6729 unsigned long __init __absent_pages_in_range(int nid,
6730 unsigned long range_start_pfn,
6731 unsigned long range_end_pfn)
6733 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6734 unsigned long start_pfn, end_pfn;
6737 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6738 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6739 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6740 nr_absent -= end_pfn - start_pfn;
6746 * absent_pages_in_range - Return number of page frames in holes within a range
6747 * @start_pfn: The start PFN to start searching for holes
6748 * @end_pfn: The end PFN to stop searching for holes
6750 * Return: the number of pages frames in memory holes within a range.
6752 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6753 unsigned long end_pfn)
6755 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6758 /* Return the number of page frames in holes in a zone on a node */
6759 static unsigned long __init zone_absent_pages_in_node(int nid,
6760 unsigned long zone_type,
6761 unsigned long node_start_pfn,
6762 unsigned long node_end_pfn)
6764 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6765 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6766 unsigned long zone_start_pfn, zone_end_pfn;
6767 unsigned long nr_absent;
6769 /* When hotadd a new node from cpu_up(), the node should be empty */
6770 if (!node_start_pfn && !node_end_pfn)
6773 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6774 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6776 adjust_zone_range_for_zone_movable(nid, zone_type,
6777 node_start_pfn, node_end_pfn,
6778 &zone_start_pfn, &zone_end_pfn);
6779 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6782 * ZONE_MOVABLE handling.
6783 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6786 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6787 unsigned long start_pfn, end_pfn;
6788 struct memblock_region *r;
6790 for_each_mem_region(r) {
6791 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6792 zone_start_pfn, zone_end_pfn);
6793 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6794 zone_start_pfn, zone_end_pfn);
6796 if (zone_type == ZONE_MOVABLE &&
6797 memblock_is_mirror(r))
6798 nr_absent += end_pfn - start_pfn;
6800 if (zone_type == ZONE_NORMAL &&
6801 !memblock_is_mirror(r))
6802 nr_absent += end_pfn - start_pfn;
6809 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6810 unsigned long node_start_pfn,
6811 unsigned long node_end_pfn)
6813 unsigned long realtotalpages = 0, totalpages = 0;
6816 for (i = 0; i < MAX_NR_ZONES; i++) {
6817 struct zone *zone = pgdat->node_zones + i;
6818 unsigned long zone_start_pfn, zone_end_pfn;
6819 unsigned long spanned, absent;
6820 unsigned long size, real_size;
6822 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6827 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6832 real_size = size - absent;
6835 zone->zone_start_pfn = zone_start_pfn;
6837 zone->zone_start_pfn = 0;
6838 zone->spanned_pages = size;
6839 zone->present_pages = real_size;
6842 realtotalpages += real_size;
6845 pgdat->node_spanned_pages = totalpages;
6846 pgdat->node_present_pages = realtotalpages;
6847 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6851 #ifndef CONFIG_SPARSEMEM
6853 * Calculate the size of the zone->blockflags rounded to an unsigned long
6854 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6855 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6856 * round what is now in bits to nearest long in bits, then return it in
6859 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6861 unsigned long usemapsize;
6863 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6864 usemapsize = roundup(zonesize, pageblock_nr_pages);
6865 usemapsize = usemapsize >> pageblock_order;
6866 usemapsize *= NR_PAGEBLOCK_BITS;
6867 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6869 return usemapsize / 8;
6872 static void __ref setup_usemap(struct pglist_data *pgdat,
6874 unsigned long zone_start_pfn,
6875 unsigned long zonesize)
6877 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6878 zone->pageblock_flags = NULL;
6880 zone->pageblock_flags =
6881 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6883 if (!zone->pageblock_flags)
6884 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6885 usemapsize, zone->name, pgdat->node_id);
6889 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6890 unsigned long zone_start_pfn, unsigned long zonesize) {}
6891 #endif /* CONFIG_SPARSEMEM */
6893 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6895 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6896 void __init set_pageblock_order(void)
6900 /* Check that pageblock_nr_pages has not already been setup */
6901 if (pageblock_order)
6904 if (HPAGE_SHIFT > PAGE_SHIFT)
6905 order = HUGETLB_PAGE_ORDER;
6907 order = MAX_ORDER - 1;
6910 * Assume the largest contiguous order of interest is a huge page.
6911 * This value may be variable depending on boot parameters on IA64 and
6914 pageblock_order = order;
6916 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6919 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6920 * is unused as pageblock_order is set at compile-time. See
6921 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6924 void __init set_pageblock_order(void)
6928 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6930 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6931 unsigned long present_pages)
6933 unsigned long pages = spanned_pages;
6936 * Provide a more accurate estimation if there are holes within
6937 * the zone and SPARSEMEM is in use. If there are holes within the
6938 * zone, each populated memory region may cost us one or two extra
6939 * memmap pages due to alignment because memmap pages for each
6940 * populated regions may not be naturally aligned on page boundary.
6941 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6943 if (spanned_pages > present_pages + (present_pages >> 4) &&
6944 IS_ENABLED(CONFIG_SPARSEMEM))
6945 pages = present_pages;
6947 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6950 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6951 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6953 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6955 spin_lock_init(&ds_queue->split_queue_lock);
6956 INIT_LIST_HEAD(&ds_queue->split_queue);
6957 ds_queue->split_queue_len = 0;
6960 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6963 #ifdef CONFIG_COMPACTION
6964 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6966 init_waitqueue_head(&pgdat->kcompactd_wait);
6969 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6972 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6974 pgdat_resize_init(pgdat);
6976 pgdat_init_split_queue(pgdat);
6977 pgdat_init_kcompactd(pgdat);
6979 init_waitqueue_head(&pgdat->kswapd_wait);
6980 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6982 pgdat_page_ext_init(pgdat);
6983 spin_lock_init(&pgdat->lru_lock);
6984 lruvec_init(&pgdat->__lruvec);
6987 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6988 unsigned long remaining_pages)
6990 atomic_long_set(&zone->managed_pages, remaining_pages);
6991 zone_set_nid(zone, nid);
6992 zone->name = zone_names[idx];
6993 zone->zone_pgdat = NODE_DATA(nid);
6994 spin_lock_init(&zone->lock);
6995 zone_seqlock_init(zone);
6996 zone_pcp_init(zone);
7000 * Set up the zone data structures
7001 * - init pgdat internals
7002 * - init all zones belonging to this node
7004 * NOTE: this function is only called during memory hotplug
7006 #ifdef CONFIG_MEMORY_HOTPLUG
7007 void __ref free_area_init_core_hotplug(int nid)
7010 pg_data_t *pgdat = NODE_DATA(nid);
7012 pgdat_init_internals(pgdat);
7013 for (z = 0; z < MAX_NR_ZONES; z++)
7014 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7019 * Set up the zone data structures:
7020 * - mark all pages reserved
7021 * - mark all memory queues empty
7022 * - clear the memory bitmaps
7024 * NOTE: pgdat should get zeroed by caller.
7025 * NOTE: this function is only called during early init.
7027 static void __init free_area_init_core(struct pglist_data *pgdat)
7030 int nid = pgdat->node_id;
7032 pgdat_init_internals(pgdat);
7033 pgdat->per_cpu_nodestats = &boot_nodestats;
7035 for (j = 0; j < MAX_NR_ZONES; j++) {
7036 struct zone *zone = pgdat->node_zones + j;
7037 unsigned long size, freesize, memmap_pages;
7038 unsigned long zone_start_pfn = zone->zone_start_pfn;
7040 size = zone->spanned_pages;
7041 freesize = zone->present_pages;
7044 * Adjust freesize so that it accounts for how much memory
7045 * is used by this zone for memmap. This affects the watermark
7046 * and per-cpu initialisations
7048 memmap_pages = calc_memmap_size(size, freesize);
7049 if (!is_highmem_idx(j)) {
7050 if (freesize >= memmap_pages) {
7051 freesize -= memmap_pages;
7054 " %s zone: %lu pages used for memmap\n",
7055 zone_names[j], memmap_pages);
7057 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
7058 zone_names[j], memmap_pages, freesize);
7061 /* Account for reserved pages */
7062 if (j == 0 && freesize > dma_reserve) {
7063 freesize -= dma_reserve;
7064 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
7065 zone_names[0], dma_reserve);
7068 if (!is_highmem_idx(j))
7069 nr_kernel_pages += freesize;
7070 /* Charge for highmem memmap if there are enough kernel pages */
7071 else if (nr_kernel_pages > memmap_pages * 2)
7072 nr_kernel_pages -= memmap_pages;
7073 nr_all_pages += freesize;
7076 * Set an approximate value for lowmem here, it will be adjusted
7077 * when the bootmem allocator frees pages into the buddy system.
7078 * And all highmem pages will be managed by the buddy system.
7080 zone_init_internals(zone, j, nid, freesize);
7085 set_pageblock_order();
7086 setup_usemap(pgdat, zone, zone_start_pfn, size);
7087 init_currently_empty_zone(zone, zone_start_pfn, size);
7088 arch_memmap_init(size, nid, j, zone_start_pfn);
7092 #ifdef CONFIG_FLAT_NODE_MEM_MAP
7093 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7095 unsigned long __maybe_unused start = 0;
7096 unsigned long __maybe_unused offset = 0;
7098 /* Skip empty nodes */
7099 if (!pgdat->node_spanned_pages)
7102 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7103 offset = pgdat->node_start_pfn - start;
7104 /* ia64 gets its own node_mem_map, before this, without bootmem */
7105 if (!pgdat->node_mem_map) {
7106 unsigned long size, end;
7110 * The zone's endpoints aren't required to be MAX_ORDER
7111 * aligned but the node_mem_map endpoints must be in order
7112 * for the buddy allocator to function correctly.
7114 end = pgdat_end_pfn(pgdat);
7115 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7116 size = (end - start) * sizeof(struct page);
7117 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7120 panic("Failed to allocate %ld bytes for node %d memory map\n",
7121 size, pgdat->node_id);
7122 pgdat->node_mem_map = map + offset;
7124 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7125 __func__, pgdat->node_id, (unsigned long)pgdat,
7126 (unsigned long)pgdat->node_mem_map);
7127 #ifndef CONFIG_NEED_MULTIPLE_NODES
7129 * With no DISCONTIG, the global mem_map is just set as node 0's
7131 if (pgdat == NODE_DATA(0)) {
7132 mem_map = NODE_DATA(0)->node_mem_map;
7133 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7139 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7140 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7142 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7143 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7145 pgdat->first_deferred_pfn = ULONG_MAX;
7148 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7151 static void __init free_area_init_node(int nid)
7153 pg_data_t *pgdat = NODE_DATA(nid);
7154 unsigned long start_pfn = 0;
7155 unsigned long end_pfn = 0;
7157 /* pg_data_t should be reset to zero when it's allocated */
7158 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7160 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7162 pgdat->node_id = nid;
7163 pgdat->node_start_pfn = start_pfn;
7164 pgdat->per_cpu_nodestats = NULL;
7166 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7167 (u64)start_pfn << PAGE_SHIFT,
7168 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7169 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7171 alloc_node_mem_map(pgdat);
7172 pgdat_set_deferred_range(pgdat);
7174 free_area_init_core(pgdat);
7177 void __init free_area_init_memoryless_node(int nid)
7179 free_area_init_node(nid);
7182 #if MAX_NUMNODES > 1
7184 * Figure out the number of possible node ids.
7186 void __init setup_nr_node_ids(void)
7188 unsigned int highest;
7190 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7191 nr_node_ids = highest + 1;
7196 * node_map_pfn_alignment - determine the maximum internode alignment
7198 * This function should be called after node map is populated and sorted.
7199 * It calculates the maximum power of two alignment which can distinguish
7202 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7203 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7204 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7205 * shifted, 1GiB is enough and this function will indicate so.
7207 * This is used to test whether pfn -> nid mapping of the chosen memory
7208 * model has fine enough granularity to avoid incorrect mapping for the
7209 * populated node map.
7211 * Return: the determined alignment in pfn's. 0 if there is no alignment
7212 * requirement (single node).
7214 unsigned long __init node_map_pfn_alignment(void)
7216 unsigned long accl_mask = 0, last_end = 0;
7217 unsigned long start, end, mask;
7218 int last_nid = NUMA_NO_NODE;
7221 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7222 if (!start || last_nid < 0 || last_nid == nid) {
7229 * Start with a mask granular enough to pin-point to the
7230 * start pfn and tick off bits one-by-one until it becomes
7231 * too coarse to separate the current node from the last.
7233 mask = ~((1 << __ffs(start)) - 1);
7234 while (mask && last_end <= (start & (mask << 1)))
7237 /* accumulate all internode masks */
7241 /* convert mask to number of pages */
7242 return ~accl_mask + 1;
7246 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7248 * Return: the minimum PFN based on information provided via
7249 * memblock_set_node().
7251 unsigned long __init find_min_pfn_with_active_regions(void)
7253 return PHYS_PFN(memblock_start_of_DRAM());
7257 * early_calculate_totalpages()
7258 * Sum pages in active regions for movable zone.
7259 * Populate N_MEMORY for calculating usable_nodes.
7261 static unsigned long __init early_calculate_totalpages(void)
7263 unsigned long totalpages = 0;
7264 unsigned long start_pfn, end_pfn;
7267 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7268 unsigned long pages = end_pfn - start_pfn;
7270 totalpages += pages;
7272 node_set_state(nid, N_MEMORY);
7278 * Find the PFN the Movable zone begins in each node. Kernel memory
7279 * is spread evenly between nodes as long as the nodes have enough
7280 * memory. When they don't, some nodes will have more kernelcore than
7283 static void __init find_zone_movable_pfns_for_nodes(void)
7286 unsigned long usable_startpfn;
7287 unsigned long kernelcore_node, kernelcore_remaining;
7288 /* save the state before borrow the nodemask */
7289 nodemask_t saved_node_state = node_states[N_MEMORY];
7290 unsigned long totalpages = early_calculate_totalpages();
7291 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7292 struct memblock_region *r;
7294 /* Need to find movable_zone earlier when movable_node is specified. */
7295 find_usable_zone_for_movable();
7298 * If movable_node is specified, ignore kernelcore and movablecore
7301 if (movable_node_is_enabled()) {
7302 for_each_mem_region(r) {
7303 if (!memblock_is_hotpluggable(r))
7306 nid = memblock_get_region_node(r);
7308 usable_startpfn = PFN_DOWN(r->base);
7309 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7310 min(usable_startpfn, zone_movable_pfn[nid]) :
7318 * If kernelcore=mirror is specified, ignore movablecore option
7320 if (mirrored_kernelcore) {
7321 bool mem_below_4gb_not_mirrored = false;
7323 for_each_mem_region(r) {
7324 if (memblock_is_mirror(r))
7327 nid = memblock_get_region_node(r);
7329 usable_startpfn = memblock_region_memory_base_pfn(r);
7331 if (usable_startpfn < 0x100000) {
7332 mem_below_4gb_not_mirrored = true;
7336 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7337 min(usable_startpfn, zone_movable_pfn[nid]) :
7341 if (mem_below_4gb_not_mirrored)
7342 pr_warn("This configuration results in unmirrored kernel memory.\n");
7348 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7349 * amount of necessary memory.
7351 if (required_kernelcore_percent)
7352 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7354 if (required_movablecore_percent)
7355 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7359 * If movablecore= was specified, calculate what size of
7360 * kernelcore that corresponds so that memory usable for
7361 * any allocation type is evenly spread. If both kernelcore
7362 * and movablecore are specified, then the value of kernelcore
7363 * will be used for required_kernelcore if it's greater than
7364 * what movablecore would have allowed.
7366 if (required_movablecore) {
7367 unsigned long corepages;
7370 * Round-up so that ZONE_MOVABLE is at least as large as what
7371 * was requested by the user
7373 required_movablecore =
7374 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7375 required_movablecore = min(totalpages, required_movablecore);
7376 corepages = totalpages - required_movablecore;
7378 required_kernelcore = max(required_kernelcore, corepages);
7382 * If kernelcore was not specified or kernelcore size is larger
7383 * than totalpages, there is no ZONE_MOVABLE.
7385 if (!required_kernelcore || required_kernelcore >= totalpages)
7388 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7389 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7392 /* Spread kernelcore memory as evenly as possible throughout nodes */
7393 kernelcore_node = required_kernelcore / usable_nodes;
7394 for_each_node_state(nid, N_MEMORY) {
7395 unsigned long start_pfn, end_pfn;
7398 * Recalculate kernelcore_node if the division per node
7399 * now exceeds what is necessary to satisfy the requested
7400 * amount of memory for the kernel
7402 if (required_kernelcore < kernelcore_node)
7403 kernelcore_node = required_kernelcore / usable_nodes;
7406 * As the map is walked, we track how much memory is usable
7407 * by the kernel using kernelcore_remaining. When it is
7408 * 0, the rest of the node is usable by ZONE_MOVABLE
7410 kernelcore_remaining = kernelcore_node;
7412 /* Go through each range of PFNs within this node */
7413 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7414 unsigned long size_pages;
7416 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7417 if (start_pfn >= end_pfn)
7420 /* Account for what is only usable for kernelcore */
7421 if (start_pfn < usable_startpfn) {
7422 unsigned long kernel_pages;
7423 kernel_pages = min(end_pfn, usable_startpfn)
7426 kernelcore_remaining -= min(kernel_pages,
7427 kernelcore_remaining);
7428 required_kernelcore -= min(kernel_pages,
7429 required_kernelcore);
7431 /* Continue if range is now fully accounted */
7432 if (end_pfn <= usable_startpfn) {
7435 * Push zone_movable_pfn to the end so
7436 * that if we have to rebalance
7437 * kernelcore across nodes, we will
7438 * not double account here
7440 zone_movable_pfn[nid] = end_pfn;
7443 start_pfn = usable_startpfn;
7447 * The usable PFN range for ZONE_MOVABLE is from
7448 * start_pfn->end_pfn. Calculate size_pages as the
7449 * number of pages used as kernelcore
7451 size_pages = end_pfn - start_pfn;
7452 if (size_pages > kernelcore_remaining)
7453 size_pages = kernelcore_remaining;
7454 zone_movable_pfn[nid] = start_pfn + size_pages;
7457 * Some kernelcore has been met, update counts and
7458 * break if the kernelcore for this node has been
7461 required_kernelcore -= min(required_kernelcore,
7463 kernelcore_remaining -= size_pages;
7464 if (!kernelcore_remaining)
7470 * If there is still required_kernelcore, we do another pass with one
7471 * less node in the count. This will push zone_movable_pfn[nid] further
7472 * along on the nodes that still have memory until kernelcore is
7476 if (usable_nodes && required_kernelcore > usable_nodes)
7480 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7481 for (nid = 0; nid < MAX_NUMNODES; nid++)
7482 zone_movable_pfn[nid] =
7483 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7486 /* restore the node_state */
7487 node_states[N_MEMORY] = saved_node_state;
7490 /* Any regular or high memory on that node ? */
7491 static void check_for_memory(pg_data_t *pgdat, int nid)
7493 enum zone_type zone_type;
7495 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7496 struct zone *zone = &pgdat->node_zones[zone_type];
7497 if (populated_zone(zone)) {
7498 if (IS_ENABLED(CONFIG_HIGHMEM))
7499 node_set_state(nid, N_HIGH_MEMORY);
7500 if (zone_type <= ZONE_NORMAL)
7501 node_set_state(nid, N_NORMAL_MEMORY);
7508 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7509 * such cases we allow max_zone_pfn sorted in the descending order
7511 bool __weak arch_has_descending_max_zone_pfns(void)
7517 * free_area_init - Initialise all pg_data_t and zone data
7518 * @max_zone_pfn: an array of max PFNs for each zone
7520 * This will call free_area_init_node() for each active node in the system.
7521 * Using the page ranges provided by memblock_set_node(), the size of each
7522 * zone in each node and their holes is calculated. If the maximum PFN
7523 * between two adjacent zones match, it is assumed that the zone is empty.
7524 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7525 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7526 * starts where the previous one ended. For example, ZONE_DMA32 starts
7527 * at arch_max_dma_pfn.
7529 void __init free_area_init(unsigned long *max_zone_pfn)
7531 unsigned long start_pfn, end_pfn;
7535 /* Record where the zone boundaries are */
7536 memset(arch_zone_lowest_possible_pfn, 0,
7537 sizeof(arch_zone_lowest_possible_pfn));
7538 memset(arch_zone_highest_possible_pfn, 0,
7539 sizeof(arch_zone_highest_possible_pfn));
7541 start_pfn = find_min_pfn_with_active_regions();
7542 descending = arch_has_descending_max_zone_pfns();
7544 for (i = 0; i < MAX_NR_ZONES; i++) {
7546 zone = MAX_NR_ZONES - i - 1;
7550 if (zone == ZONE_MOVABLE)
7553 end_pfn = max(max_zone_pfn[zone], start_pfn);
7554 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7555 arch_zone_highest_possible_pfn[zone] = end_pfn;
7557 start_pfn = end_pfn;
7560 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7561 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7562 find_zone_movable_pfns_for_nodes();
7564 /* Print out the zone ranges */
7565 pr_info("Zone ranges:\n");
7566 for (i = 0; i < MAX_NR_ZONES; i++) {
7567 if (i == ZONE_MOVABLE)
7569 pr_info(" %-8s ", zone_names[i]);
7570 if (arch_zone_lowest_possible_pfn[i] ==
7571 arch_zone_highest_possible_pfn[i])
7574 pr_cont("[mem %#018Lx-%#018Lx]\n",
7575 (u64)arch_zone_lowest_possible_pfn[i]
7577 ((u64)arch_zone_highest_possible_pfn[i]
7578 << PAGE_SHIFT) - 1);
7581 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7582 pr_info("Movable zone start for each node\n");
7583 for (i = 0; i < MAX_NUMNODES; i++) {
7584 if (zone_movable_pfn[i])
7585 pr_info(" Node %d: %#018Lx\n", i,
7586 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7590 * Print out the early node map, and initialize the
7591 * subsection-map relative to active online memory ranges to
7592 * enable future "sub-section" extensions of the memory map.
7594 pr_info("Early memory node ranges\n");
7595 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7596 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7597 (u64)start_pfn << PAGE_SHIFT,
7598 ((u64)end_pfn << PAGE_SHIFT) - 1);
7599 subsection_map_init(start_pfn, end_pfn - start_pfn);
7602 /* Initialise every node */
7603 mminit_verify_pageflags_layout();
7604 setup_nr_node_ids();
7605 for_each_online_node(nid) {
7606 pg_data_t *pgdat = NODE_DATA(nid);
7607 free_area_init_node(nid);
7609 /* Any memory on that node */
7610 if (pgdat->node_present_pages)
7611 node_set_state(nid, N_MEMORY);
7612 check_for_memory(pgdat, nid);
7618 static int __init cmdline_parse_core(char *p, unsigned long *core,
7619 unsigned long *percent)
7621 unsigned long long coremem;
7627 /* Value may be a percentage of total memory, otherwise bytes */
7628 coremem = simple_strtoull(p, &endptr, 0);
7629 if (*endptr == '%') {
7630 /* Paranoid check for percent values greater than 100 */
7631 WARN_ON(coremem > 100);
7635 coremem = memparse(p, &p);
7636 /* Paranoid check that UL is enough for the coremem value */
7637 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7639 *core = coremem >> PAGE_SHIFT;
7646 * kernelcore=size sets the amount of memory for use for allocations that
7647 * cannot be reclaimed or migrated.
7649 static int __init cmdline_parse_kernelcore(char *p)
7651 /* parse kernelcore=mirror */
7652 if (parse_option_str(p, "mirror")) {
7653 mirrored_kernelcore = true;
7657 return cmdline_parse_core(p, &required_kernelcore,
7658 &required_kernelcore_percent);
7662 * movablecore=size sets the amount of memory for use for allocations that
7663 * can be reclaimed or migrated.
7665 static int __init cmdline_parse_movablecore(char *p)
7667 return cmdline_parse_core(p, &required_movablecore,
7668 &required_movablecore_percent);
7671 early_param("kernelcore", cmdline_parse_kernelcore);
7672 early_param("movablecore", cmdline_parse_movablecore);
7674 void adjust_managed_page_count(struct page *page, long count)
7676 atomic_long_add(count, &page_zone(page)->managed_pages);
7677 totalram_pages_add(count);
7678 #ifdef CONFIG_HIGHMEM
7679 if (PageHighMem(page))
7680 totalhigh_pages_add(count);
7683 EXPORT_SYMBOL(adjust_managed_page_count);
7685 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7688 unsigned long pages = 0;
7690 start = (void *)PAGE_ALIGN((unsigned long)start);
7691 end = (void *)((unsigned long)end & PAGE_MASK);
7692 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7693 struct page *page = virt_to_page(pos);
7694 void *direct_map_addr;
7697 * 'direct_map_addr' might be different from 'pos'
7698 * because some architectures' virt_to_page()
7699 * work with aliases. Getting the direct map
7700 * address ensures that we get a _writeable_
7701 * alias for the memset().
7703 direct_map_addr = page_address(page);
7704 if ((unsigned int)poison <= 0xFF)
7705 memset(direct_map_addr, poison, PAGE_SIZE);
7707 free_reserved_page(page);
7711 pr_info("Freeing %s memory: %ldK\n",
7712 s, pages << (PAGE_SHIFT - 10));
7717 #ifdef CONFIG_HIGHMEM
7718 void free_highmem_page(struct page *page)
7720 __free_reserved_page(page);
7721 totalram_pages_inc();
7722 atomic_long_inc(&page_zone(page)->managed_pages);
7723 totalhigh_pages_inc();
7728 void __init mem_init_print_info(const char *str)
7730 unsigned long physpages, codesize, datasize, rosize, bss_size;
7731 unsigned long init_code_size, init_data_size;
7733 physpages = get_num_physpages();
7734 codesize = _etext - _stext;
7735 datasize = _edata - _sdata;
7736 rosize = __end_rodata - __start_rodata;
7737 bss_size = __bss_stop - __bss_start;
7738 init_data_size = __init_end - __init_begin;
7739 init_code_size = _einittext - _sinittext;
7742 * Detect special cases and adjust section sizes accordingly:
7743 * 1) .init.* may be embedded into .data sections
7744 * 2) .init.text.* may be out of [__init_begin, __init_end],
7745 * please refer to arch/tile/kernel/vmlinux.lds.S.
7746 * 3) .rodata.* may be embedded into .text or .data sections.
7748 #define adj_init_size(start, end, size, pos, adj) \
7750 if (start <= pos && pos < end && size > adj) \
7754 adj_init_size(__init_begin, __init_end, init_data_size,
7755 _sinittext, init_code_size);
7756 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7757 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7758 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7759 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7761 #undef adj_init_size
7763 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7764 #ifdef CONFIG_HIGHMEM
7768 nr_free_pages() << (PAGE_SHIFT - 10),
7769 physpages << (PAGE_SHIFT - 10),
7770 codesize >> 10, datasize >> 10, rosize >> 10,
7771 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7772 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7773 totalcma_pages << (PAGE_SHIFT - 10),
7774 #ifdef CONFIG_HIGHMEM
7775 totalhigh_pages() << (PAGE_SHIFT - 10),
7777 str ? ", " : "", str ? str : "");
7781 * set_dma_reserve - set the specified number of pages reserved in the first zone
7782 * @new_dma_reserve: The number of pages to mark reserved
7784 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7785 * In the DMA zone, a significant percentage may be consumed by kernel image
7786 * and other unfreeable allocations which can skew the watermarks badly. This
7787 * function may optionally be used to account for unfreeable pages in the
7788 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7789 * smaller per-cpu batchsize.
7791 void __init set_dma_reserve(unsigned long new_dma_reserve)
7793 dma_reserve = new_dma_reserve;
7796 static int page_alloc_cpu_dead(unsigned int cpu)
7799 lru_add_drain_cpu(cpu);
7803 * Spill the event counters of the dead processor
7804 * into the current processors event counters.
7805 * This artificially elevates the count of the current
7808 vm_events_fold_cpu(cpu);
7811 * Zero the differential counters of the dead processor
7812 * so that the vm statistics are consistent.
7814 * This is only okay since the processor is dead and cannot
7815 * race with what we are doing.
7817 cpu_vm_stats_fold(cpu);
7822 int hashdist = HASHDIST_DEFAULT;
7824 static int __init set_hashdist(char *str)
7828 hashdist = simple_strtoul(str, &str, 0);
7831 __setup("hashdist=", set_hashdist);
7834 void __init page_alloc_init(void)
7839 if (num_node_state(N_MEMORY) == 1)
7843 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7844 "mm/page_alloc:dead", NULL,
7845 page_alloc_cpu_dead);
7850 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7851 * or min_free_kbytes changes.
7853 static void calculate_totalreserve_pages(void)
7855 struct pglist_data *pgdat;
7856 unsigned long reserve_pages = 0;
7857 enum zone_type i, j;
7859 for_each_online_pgdat(pgdat) {
7861 pgdat->totalreserve_pages = 0;
7863 for (i = 0; i < MAX_NR_ZONES; i++) {
7864 struct zone *zone = pgdat->node_zones + i;
7866 unsigned long managed_pages = zone_managed_pages(zone);
7868 /* Find valid and maximum lowmem_reserve in the zone */
7869 for (j = i; j < MAX_NR_ZONES; j++) {
7870 if (zone->lowmem_reserve[j] > max)
7871 max = zone->lowmem_reserve[j];
7874 /* we treat the high watermark as reserved pages. */
7875 max += high_wmark_pages(zone);
7877 if (max > managed_pages)
7878 max = managed_pages;
7880 pgdat->totalreserve_pages += max;
7882 reserve_pages += max;
7885 totalreserve_pages = reserve_pages;
7889 * setup_per_zone_lowmem_reserve - called whenever
7890 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7891 * has a correct pages reserved value, so an adequate number of
7892 * pages are left in the zone after a successful __alloc_pages().
7894 static void setup_per_zone_lowmem_reserve(void)
7896 struct pglist_data *pgdat;
7897 enum zone_type i, j;
7899 for_each_online_pgdat(pgdat) {
7900 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
7901 struct zone *zone = &pgdat->node_zones[i];
7902 int ratio = sysctl_lowmem_reserve_ratio[i];
7903 bool clear = !ratio || !zone_managed_pages(zone);
7904 unsigned long managed_pages = 0;
7906 for (j = i + 1; j < MAX_NR_ZONES; j++) {
7907 struct zone *upper_zone = &pgdat->node_zones[j];
7909 managed_pages += zone_managed_pages(upper_zone);
7912 zone->lowmem_reserve[j] = 0;
7914 zone->lowmem_reserve[j] = managed_pages / ratio;
7919 /* update totalreserve_pages */
7920 calculate_totalreserve_pages();
7923 static void __setup_per_zone_wmarks(void)
7925 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7926 unsigned long lowmem_pages = 0;
7928 unsigned long flags;
7930 /* Calculate total number of !ZONE_HIGHMEM pages */
7931 for_each_zone(zone) {
7932 if (!is_highmem(zone))
7933 lowmem_pages += zone_managed_pages(zone);
7936 for_each_zone(zone) {
7939 spin_lock_irqsave(&zone->lock, flags);
7940 tmp = (u64)pages_min * zone_managed_pages(zone);
7941 do_div(tmp, lowmem_pages);
7942 if (is_highmem(zone)) {
7944 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7945 * need highmem pages, so cap pages_min to a small
7948 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7949 * deltas control async page reclaim, and so should
7950 * not be capped for highmem.
7952 unsigned long min_pages;
7954 min_pages = zone_managed_pages(zone) / 1024;
7955 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7956 zone->_watermark[WMARK_MIN] = min_pages;
7959 * If it's a lowmem zone, reserve a number of pages
7960 * proportionate to the zone's size.
7962 zone->_watermark[WMARK_MIN] = tmp;
7966 * Set the kswapd watermarks distance according to the
7967 * scale factor in proportion to available memory, but
7968 * ensure a minimum size on small systems.
7970 tmp = max_t(u64, tmp >> 2,
7971 mult_frac(zone_managed_pages(zone),
7972 watermark_scale_factor, 10000));
7974 zone->watermark_boost = 0;
7975 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7976 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7978 spin_unlock_irqrestore(&zone->lock, flags);
7981 /* update totalreserve_pages */
7982 calculate_totalreserve_pages();
7986 * setup_per_zone_wmarks - called when min_free_kbytes changes
7987 * or when memory is hot-{added|removed}
7989 * Ensures that the watermark[min,low,high] values for each zone are set
7990 * correctly with respect to min_free_kbytes.
7992 void setup_per_zone_wmarks(void)
7994 static DEFINE_SPINLOCK(lock);
7997 __setup_per_zone_wmarks();
8002 * Initialise min_free_kbytes.
8004 * For small machines we want it small (128k min). For large machines
8005 * we want it large (256MB max). But it is not linear, because network
8006 * bandwidth does not increase linearly with machine size. We use
8008 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8009 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8025 int __meminit init_per_zone_wmark_min(void)
8027 unsigned long lowmem_kbytes;
8028 int new_min_free_kbytes;
8030 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8031 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8033 if (new_min_free_kbytes > user_min_free_kbytes) {
8034 min_free_kbytes = new_min_free_kbytes;
8035 if (min_free_kbytes < 128)
8036 min_free_kbytes = 128;
8037 if (min_free_kbytes > 262144)
8038 min_free_kbytes = 262144;
8040 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8041 new_min_free_kbytes, user_min_free_kbytes);
8043 setup_per_zone_wmarks();
8044 refresh_zone_stat_thresholds();
8045 setup_per_zone_lowmem_reserve();
8048 setup_min_unmapped_ratio();
8049 setup_min_slab_ratio();
8052 khugepaged_min_free_kbytes_update();
8056 postcore_initcall(init_per_zone_wmark_min)
8059 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8060 * that we can call two helper functions whenever min_free_kbytes
8063 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8064 void *buffer, size_t *length, loff_t *ppos)
8068 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8073 user_min_free_kbytes = min_free_kbytes;
8074 setup_per_zone_wmarks();
8079 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8080 void *buffer, size_t *length, loff_t *ppos)
8084 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8089 setup_per_zone_wmarks();
8095 static void setup_min_unmapped_ratio(void)
8100 for_each_online_pgdat(pgdat)
8101 pgdat->min_unmapped_pages = 0;
8104 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8105 sysctl_min_unmapped_ratio) / 100;
8109 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8110 void *buffer, size_t *length, loff_t *ppos)
8114 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8118 setup_min_unmapped_ratio();
8123 static void setup_min_slab_ratio(void)
8128 for_each_online_pgdat(pgdat)
8129 pgdat->min_slab_pages = 0;
8132 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8133 sysctl_min_slab_ratio) / 100;
8136 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8137 void *buffer, size_t *length, loff_t *ppos)
8141 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8145 setup_min_slab_ratio();
8152 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8153 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8154 * whenever sysctl_lowmem_reserve_ratio changes.
8156 * The reserve ratio obviously has absolutely no relation with the
8157 * minimum watermarks. The lowmem reserve ratio can only make sense
8158 * if in function of the boot time zone sizes.
8160 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8161 void *buffer, size_t *length, loff_t *ppos)
8165 proc_dointvec_minmax(table, write, buffer, length, ppos);
8167 for (i = 0; i < MAX_NR_ZONES; i++) {
8168 if (sysctl_lowmem_reserve_ratio[i] < 1)
8169 sysctl_lowmem_reserve_ratio[i] = 0;
8172 setup_per_zone_lowmem_reserve();
8176 static void __zone_pcp_update(struct zone *zone)
8180 for_each_possible_cpu(cpu)
8181 pageset_set_high_and_batch(zone,
8182 per_cpu_ptr(zone->pageset, cpu));
8186 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8187 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8188 * pagelist can have before it gets flushed back to buddy allocator.
8190 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8191 void *buffer, size_t *length, loff_t *ppos)
8194 int old_percpu_pagelist_fraction;
8197 mutex_lock(&pcp_batch_high_lock);
8198 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8200 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8201 if (!write || ret < 0)
8204 /* Sanity checking to avoid pcp imbalance */
8205 if (percpu_pagelist_fraction &&
8206 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8207 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8213 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8216 for_each_populated_zone(zone)
8217 __zone_pcp_update(zone);
8219 mutex_unlock(&pcp_batch_high_lock);
8223 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8225 * Returns the number of pages that arch has reserved but
8226 * is not known to alloc_large_system_hash().
8228 static unsigned long __init arch_reserved_kernel_pages(void)
8235 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8236 * machines. As memory size is increased the scale is also increased but at
8237 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8238 * quadruples the scale is increased by one, which means the size of hash table
8239 * only doubles, instead of quadrupling as well.
8240 * Because 32-bit systems cannot have large physical memory, where this scaling
8241 * makes sense, it is disabled on such platforms.
8243 #if __BITS_PER_LONG > 32
8244 #define ADAPT_SCALE_BASE (64ul << 30)
8245 #define ADAPT_SCALE_SHIFT 2
8246 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8250 * allocate a large system hash table from bootmem
8251 * - it is assumed that the hash table must contain an exact power-of-2
8252 * quantity of entries
8253 * - limit is the number of hash buckets, not the total allocation size
8255 void *__init alloc_large_system_hash(const char *tablename,
8256 unsigned long bucketsize,
8257 unsigned long numentries,
8260 unsigned int *_hash_shift,
8261 unsigned int *_hash_mask,
8262 unsigned long low_limit,
8263 unsigned long high_limit)
8265 unsigned long long max = high_limit;
8266 unsigned long log2qty, size;
8271 /* allow the kernel cmdline to have a say */
8273 /* round applicable memory size up to nearest megabyte */
8274 numentries = nr_kernel_pages;
8275 numentries -= arch_reserved_kernel_pages();
8277 /* It isn't necessary when PAGE_SIZE >= 1MB */
8278 if (PAGE_SHIFT < 20)
8279 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8281 #if __BITS_PER_LONG > 32
8283 unsigned long adapt;
8285 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8286 adapt <<= ADAPT_SCALE_SHIFT)
8291 /* limit to 1 bucket per 2^scale bytes of low memory */
8292 if (scale > PAGE_SHIFT)
8293 numentries >>= (scale - PAGE_SHIFT);
8295 numentries <<= (PAGE_SHIFT - scale);
8297 /* Make sure we've got at least a 0-order allocation.. */
8298 if (unlikely(flags & HASH_SMALL)) {
8299 /* Makes no sense without HASH_EARLY */
8300 WARN_ON(!(flags & HASH_EARLY));
8301 if (!(numentries >> *_hash_shift)) {
8302 numentries = 1UL << *_hash_shift;
8303 BUG_ON(!numentries);
8305 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8306 numentries = PAGE_SIZE / bucketsize;
8308 numentries = roundup_pow_of_two(numentries);
8310 /* limit allocation size to 1/16 total memory by default */
8312 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8313 do_div(max, bucketsize);
8315 max = min(max, 0x80000000ULL);
8317 if (numentries < low_limit)
8318 numentries = low_limit;
8319 if (numentries > max)
8322 log2qty = ilog2(numentries);
8324 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8327 size = bucketsize << log2qty;
8328 if (flags & HASH_EARLY) {
8329 if (flags & HASH_ZERO)
8330 table = memblock_alloc(size, SMP_CACHE_BYTES);
8332 table = memblock_alloc_raw(size,
8334 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8335 table = __vmalloc(size, gfp_flags);
8339 * If bucketsize is not a power-of-two, we may free
8340 * some pages at the end of hash table which
8341 * alloc_pages_exact() automatically does
8343 table = alloc_pages_exact(size, gfp_flags);
8344 kmemleak_alloc(table, size, 1, gfp_flags);
8346 } while (!table && size > PAGE_SIZE && --log2qty);
8349 panic("Failed to allocate %s hash table\n", tablename);
8351 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8352 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8353 virt ? "vmalloc" : "linear");
8356 *_hash_shift = log2qty;
8358 *_hash_mask = (1 << log2qty) - 1;
8364 * This function checks whether pageblock includes unmovable pages or not.
8366 * PageLRU check without isolation or lru_lock could race so that
8367 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8368 * check without lock_page also may miss some movable non-lru pages at
8369 * race condition. So you can't expect this function should be exact.
8371 * Returns a page without holding a reference. If the caller wants to
8372 * dereference that page (e.g., dumping), it has to make sure that it
8373 * cannot get removed (e.g., via memory unplug) concurrently.
8376 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8377 int migratetype, int flags)
8379 unsigned long iter = 0;
8380 unsigned long pfn = page_to_pfn(page);
8381 unsigned long offset = pfn % pageblock_nr_pages;
8383 if (is_migrate_cma_page(page)) {
8385 * CMA allocations (alloc_contig_range) really need to mark
8386 * isolate CMA pageblocks even when they are not movable in fact
8387 * so consider them movable here.
8389 if (is_migrate_cma(migratetype))
8395 for (; iter < pageblock_nr_pages - offset; iter++) {
8396 if (!pfn_valid_within(pfn + iter))
8399 page = pfn_to_page(pfn + iter);
8402 * Both, bootmem allocations and memory holes are marked
8403 * PG_reserved and are unmovable. We can even have unmovable
8404 * allocations inside ZONE_MOVABLE, for example when
8405 * specifying "movablecore".
8407 if (PageReserved(page))
8411 * If the zone is movable and we have ruled out all reserved
8412 * pages then it should be reasonably safe to assume the rest
8415 if (zone_idx(zone) == ZONE_MOVABLE)
8419 * Hugepages are not in LRU lists, but they're movable.
8420 * THPs are on the LRU, but need to be counted as #small pages.
8421 * We need not scan over tail pages because we don't
8422 * handle each tail page individually in migration.
8424 if (PageHuge(page) || PageTransCompound(page)) {
8425 struct page *head = compound_head(page);
8426 unsigned int skip_pages;
8428 if (PageHuge(page)) {
8429 if (!hugepage_migration_supported(page_hstate(head)))
8431 } else if (!PageLRU(head) && !__PageMovable(head)) {
8435 skip_pages = compound_nr(head) - (page - head);
8436 iter += skip_pages - 1;
8441 * We can't use page_count without pin a page
8442 * because another CPU can free compound page.
8443 * This check already skips compound tails of THP
8444 * because their page->_refcount is zero at all time.
8446 if (!page_ref_count(page)) {
8447 if (PageBuddy(page))
8448 iter += (1 << buddy_order(page)) - 1;
8453 * The HWPoisoned page may be not in buddy system, and
8454 * page_count() is not 0.
8456 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8460 * We treat all PageOffline() pages as movable when offlining
8461 * to give drivers a chance to decrement their reference count
8462 * in MEM_GOING_OFFLINE in order to indicate that these pages
8463 * can be offlined as there are no direct references anymore.
8464 * For actually unmovable PageOffline() where the driver does
8465 * not support this, we will fail later when trying to actually
8466 * move these pages that still have a reference count > 0.
8467 * (false negatives in this function only)
8469 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8472 if (__PageMovable(page) || PageLRU(page))
8476 * If there are RECLAIMABLE pages, we need to check
8477 * it. But now, memory offline itself doesn't call
8478 * shrink_node_slabs() and it still to be fixed.
8485 #ifdef CONFIG_CONTIG_ALLOC
8486 static unsigned long pfn_max_align_down(unsigned long pfn)
8488 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8489 pageblock_nr_pages) - 1);
8492 static unsigned long pfn_max_align_up(unsigned long pfn)
8494 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8495 pageblock_nr_pages));
8498 /* [start, end) must belong to a single zone. */
8499 static int __alloc_contig_migrate_range(struct compact_control *cc,
8500 unsigned long start, unsigned long end)
8502 /* This function is based on compact_zone() from compaction.c. */
8503 unsigned int nr_reclaimed;
8504 unsigned long pfn = start;
8505 unsigned int tries = 0;
8507 struct migration_target_control mtc = {
8508 .nid = zone_to_nid(cc->zone),
8509 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8514 while (pfn < end || !list_empty(&cc->migratepages)) {
8515 if (fatal_signal_pending(current)) {
8516 pr_err("Fatal signal pending task: %s(%d), %#lx\n",
8517 current->comm, current->pid,
8518 current->pending.signal.sig[0]);
8523 if (list_empty(&cc->migratepages)) {
8524 cc->nr_migratepages = 0;
8525 pfn = isolate_migratepages_range(cc, pfn, end);
8531 } else if (++tries == 5) {
8532 ret = ret < 0 ? ret : -EBUSY;
8536 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8538 cc->nr_migratepages -= nr_reclaimed;
8540 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8541 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8544 putback_movable_pages(&cc->migratepages);
8551 * alloc_contig_range() -- tries to allocate given range of pages
8552 * @start: start PFN to allocate
8553 * @end: one-past-the-last PFN to allocate
8554 * @migratetype: migratetype of the underlaying pageblocks (either
8555 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8556 * in range must have the same migratetype and it must
8557 * be either of the two.
8558 * @gfp_mask: GFP mask to use during compaction
8560 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8561 * aligned. The PFN range must belong to a single zone.
8563 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8564 * pageblocks in the range. Once isolated, the pageblocks should not
8565 * be modified by others.
8567 * Return: zero on success or negative error code. On success all
8568 * pages which PFN is in [start, end) are allocated for the caller and
8569 * need to be freed with free_contig_range().
8571 int alloc_contig_range(unsigned long start, unsigned long end,
8572 unsigned migratetype, gfp_t gfp_mask)
8574 unsigned long outer_start, outer_end;
8578 struct compact_control cc = {
8579 .nr_migratepages = 0,
8581 .zone = page_zone(pfn_to_page(start)),
8582 .mode = MIGRATE_SYNC,
8583 .ignore_skip_hint = true,
8584 .no_set_skip_hint = true,
8585 .gfp_mask = current_gfp_context(gfp_mask),
8586 .alloc_contig = true,
8588 INIT_LIST_HEAD(&cc.migratepages);
8591 * What we do here is we mark all pageblocks in range as
8592 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8593 * have different sizes, and due to the way page allocator
8594 * work, we align the range to biggest of the two pages so
8595 * that page allocator won't try to merge buddies from
8596 * different pageblocks and change MIGRATE_ISOLATE to some
8597 * other migration type.
8599 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8600 * migrate the pages from an unaligned range (ie. pages that
8601 * we are interested in). This will put all the pages in
8602 * range back to page allocator as MIGRATE_ISOLATE.
8604 * When this is done, we take the pages in range from page
8605 * allocator removing them from the buddy system. This way
8606 * page allocator will never consider using them.
8608 * This lets us mark the pageblocks back as
8609 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8610 * aligned range but not in the unaligned, original range are
8611 * put back to page allocator so that buddy can use them.
8614 ret = start_isolate_page_range(pfn_max_align_down(start),
8615 pfn_max_align_up(end), migratetype, 0);
8620 * In case of -EBUSY, we'd like to know which page causes problem.
8621 * So, just fall through. test_pages_isolated() has a tracepoint
8622 * which will report the busy page.
8624 * It is possible that busy pages could become available before
8625 * the call to test_pages_isolated, and the range will actually be
8626 * allocated. So, if we fall through be sure to clear ret so that
8627 * -EBUSY is not accidentally used or returned to caller.
8629 ret = __alloc_contig_migrate_range(&cc, start, end);
8630 if (ret && ret != -EBUSY)
8635 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8636 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8637 * more, all pages in [start, end) are free in page allocator.
8638 * What we are going to do is to allocate all pages from
8639 * [start, end) (that is remove them from page allocator).
8641 * The only problem is that pages at the beginning and at the
8642 * end of interesting range may be not aligned with pages that
8643 * page allocator holds, ie. they can be part of higher order
8644 * pages. Because of this, we reserve the bigger range and
8645 * once this is done free the pages we are not interested in.
8647 * We don't have to hold zone->lock here because the pages are
8648 * isolated thus they won't get removed from buddy.
8651 lru_add_drain_all();
8654 outer_start = start;
8655 while (!PageBuddy(pfn_to_page(outer_start))) {
8656 if (++order >= MAX_ORDER) {
8657 outer_start = start;
8660 outer_start &= ~0UL << order;
8663 if (outer_start != start) {
8664 order = buddy_order(pfn_to_page(outer_start));
8667 * outer_start page could be small order buddy page and
8668 * it doesn't include start page. Adjust outer_start
8669 * in this case to report failed page properly
8670 * on tracepoint in test_pages_isolated()
8672 if (outer_start + (1UL << order) <= start)
8673 outer_start = start;
8676 /* Make sure the range is really isolated. */
8677 if (test_pages_isolated(outer_start, end, 0)) {
8682 /* Grab isolated pages from freelists. */
8683 outer_end = isolate_freepages_range(&cc, outer_start, end);
8689 /* Free head and tail (if any) */
8690 if (start != outer_start)
8691 free_contig_range(outer_start, start - outer_start);
8692 if (end != outer_end)
8693 free_contig_range(end, outer_end - end);
8696 undo_isolate_page_range(pfn_max_align_down(start),
8697 pfn_max_align_up(end), migratetype);
8700 EXPORT_SYMBOL(alloc_contig_range);
8702 static int __alloc_contig_pages(unsigned long start_pfn,
8703 unsigned long nr_pages, gfp_t gfp_mask)
8705 unsigned long end_pfn = start_pfn + nr_pages;
8707 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8711 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8712 unsigned long nr_pages)
8714 unsigned long i, end_pfn = start_pfn + nr_pages;
8717 for (i = start_pfn; i < end_pfn; i++) {
8718 page = pfn_to_online_page(i);
8722 if (page_zone(page) != z)
8725 if (PageReserved(page))
8728 if (page_count(page) > 0)
8737 static bool zone_spans_last_pfn(const struct zone *zone,
8738 unsigned long start_pfn, unsigned long nr_pages)
8740 unsigned long last_pfn = start_pfn + nr_pages - 1;
8742 return zone_spans_pfn(zone, last_pfn);
8746 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8747 * @nr_pages: Number of contiguous pages to allocate
8748 * @gfp_mask: GFP mask to limit search and used during compaction
8750 * @nodemask: Mask for other possible nodes
8752 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8753 * on an applicable zonelist to find a contiguous pfn range which can then be
8754 * tried for allocation with alloc_contig_range(). This routine is intended
8755 * for allocation requests which can not be fulfilled with the buddy allocator.
8757 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8758 * power of two then the alignment is guaranteed to be to the given nr_pages
8759 * (e.g. 1GB request would be aligned to 1GB).
8761 * Allocated pages can be freed with free_contig_range() or by manually calling
8762 * __free_page() on each allocated page.
8764 * Return: pointer to contiguous pages on success, or NULL if not successful.
8766 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8767 int nid, nodemask_t *nodemask)
8769 unsigned long ret, pfn, flags;
8770 struct zonelist *zonelist;
8774 zonelist = node_zonelist(nid, gfp_mask);
8775 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8776 gfp_zone(gfp_mask), nodemask) {
8777 spin_lock_irqsave(&zone->lock, flags);
8779 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8780 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8781 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8783 * We release the zone lock here because
8784 * alloc_contig_range() will also lock the zone
8785 * at some point. If there's an allocation
8786 * spinning on this lock, it may win the race
8787 * and cause alloc_contig_range() to fail...
8789 spin_unlock_irqrestore(&zone->lock, flags);
8790 ret = __alloc_contig_pages(pfn, nr_pages,
8793 return pfn_to_page(pfn);
8794 spin_lock_irqsave(&zone->lock, flags);
8798 spin_unlock_irqrestore(&zone->lock, flags);
8802 #endif /* CONFIG_CONTIG_ALLOC */
8804 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
8806 unsigned long count = 0;
8808 for (; nr_pages--; pfn++) {
8809 struct page *page = pfn_to_page(pfn);
8811 count += page_count(page) != 1;
8814 WARN(count != 0, "%lu pages are still in use!\n", count);
8816 EXPORT_SYMBOL(free_contig_range);
8819 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8820 * page high values need to be recalulated.
8822 void __meminit zone_pcp_update(struct zone *zone)
8824 mutex_lock(&pcp_batch_high_lock);
8825 __zone_pcp_update(zone);
8826 mutex_unlock(&pcp_batch_high_lock);
8829 void zone_pcp_reset(struct zone *zone)
8831 unsigned long flags;
8833 struct per_cpu_pageset *pset;
8835 /* avoid races with drain_pages() */
8836 local_lock_irqsave(&pa_lock.l, flags);
8837 if (zone->pageset != &boot_pageset) {
8838 for_each_online_cpu(cpu) {
8839 pset = per_cpu_ptr(zone->pageset, cpu);
8840 drain_zonestat(zone, pset);
8842 free_percpu(zone->pageset);
8843 zone->pageset = &boot_pageset;
8845 local_unlock_irqrestore(&pa_lock.l, flags);
8848 #ifdef CONFIG_MEMORY_HOTREMOVE
8850 * All pages in the range must be in a single zone, must not contain holes,
8851 * must span full sections, and must be isolated before calling this function.
8853 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8855 unsigned long pfn = start_pfn;
8859 unsigned long flags;
8861 offline_mem_sections(pfn, end_pfn);
8862 zone = page_zone(pfn_to_page(pfn));
8863 spin_lock_irqsave(&zone->lock, flags);
8864 while (pfn < end_pfn) {
8865 page = pfn_to_page(pfn);
8867 * The HWPoisoned page may be not in buddy system, and
8868 * page_count() is not 0.
8870 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8875 * At this point all remaining PageOffline() pages have a
8876 * reference count of 0 and can simply be skipped.
8878 if (PageOffline(page)) {
8879 BUG_ON(page_count(page));
8880 BUG_ON(PageBuddy(page));
8885 BUG_ON(page_count(page));
8886 BUG_ON(!PageBuddy(page));
8887 order = buddy_order(page);
8888 del_page_from_free_list(page, zone, order);
8889 pfn += (1 << order);
8891 spin_unlock_irqrestore(&zone->lock, flags);
8895 bool is_free_buddy_page(struct page *page)
8897 struct zone *zone = page_zone(page);
8898 unsigned long pfn = page_to_pfn(page);
8899 unsigned long flags;
8902 spin_lock_irqsave(&zone->lock, flags);
8903 for (order = 0; order < MAX_ORDER; order++) {
8904 struct page *page_head = page - (pfn & ((1 << order) - 1));
8906 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
8909 spin_unlock_irqrestore(&zone->lock, flags);
8911 return order < MAX_ORDER;
8914 #ifdef CONFIG_MEMORY_FAILURE
8916 * Break down a higher-order page in sub-pages, and keep our target out of
8919 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8920 struct page *target, int low, int high,
8923 unsigned long size = 1 << high;
8924 struct page *current_buddy, *next_page;
8926 while (high > low) {
8930 if (target >= &page[size]) {
8931 next_page = page + size;
8932 current_buddy = page;
8935 current_buddy = page + size;
8938 if (set_page_guard(zone, current_buddy, high, migratetype))
8941 if (current_buddy != target) {
8942 add_to_free_list(current_buddy, zone, high, migratetype);
8943 set_buddy_order(current_buddy, high);
8950 * Take a page that will be marked as poisoned off the buddy allocator.
8952 bool take_page_off_buddy(struct page *page)
8954 struct zone *zone = page_zone(page);
8955 unsigned long pfn = page_to_pfn(page);
8956 unsigned long flags;
8960 spin_lock_irqsave(&zone->lock, flags);
8961 for (order = 0; order < MAX_ORDER; order++) {
8962 struct page *page_head = page - (pfn & ((1 << order) - 1));
8963 int page_order = buddy_order(page_head);
8965 if (PageBuddy(page_head) && page_order >= order) {
8966 unsigned long pfn_head = page_to_pfn(page_head);
8967 int migratetype = get_pfnblock_migratetype(page_head,
8970 del_page_from_free_list(page_head, zone, page_order);
8971 break_down_buddy_pages(zone, page_head, page, 0,
8972 page_order, migratetype);
8973 if (!is_migrate_isolate(migratetype))
8974 __mod_zone_freepage_state(zone, -1, migratetype);
8978 if (page_count(page_head) > 0)
8981 spin_unlock_irqrestore(&zone->lock, flags);
8986 #ifdef CONFIG_ZONE_DMA
8987 bool has_managed_dma(void)
8989 struct pglist_data *pgdat;
8991 for_each_online_pgdat(pgdat) {
8992 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
8994 if (managed_zone(zone))
8999 #endif /* CONFIG_ZONE_DMA */