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/mmu_notifier.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.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>
74 #include <linux/buffer_head.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))
111 * Don't poison memory with KASAN (only for the tag-based modes).
112 * During boot, all non-reserved memblock memory is exposed to page_alloc.
113 * Poisoning all that memory lengthens boot time, especially on systems with
114 * large amount of RAM. This flag is used to skip that poisoning.
115 * This is only done for the tag-based KASAN modes, as those are able to
116 * detect memory corruptions with the memory tags assigned by default.
117 * All memory allocated normally after boot gets poisoned as usual.
119 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
121 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
122 static DEFINE_MUTEX(pcp_batch_high_lock);
123 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
127 #if defined(CONFIG_DEBUG_INFO_BTF) && \
128 !defined(CONFIG_DEBUG_LOCK_ALLOC) && \
129 !defined(CONFIG_PAHOLE_HAS_ZEROSIZE_PERCPU_SUPPORT)
131 * pahole 1.21 and earlier gets confused by zero-sized per-CPU
132 * variables and produces invalid BTF. Ensure that
133 * sizeof(struct pagesets) != 0 for older versions of pahole.
136 #warning "pahole too old to support zero-sized struct pagesets"
139 static DEFINE_PER_CPU(struct pagesets, pagesets) = {
140 .lock = INIT_LOCAL_LOCK(lock),
143 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
144 DEFINE_PER_CPU(int, numa_node);
145 EXPORT_PER_CPU_SYMBOL(numa_node);
148 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
150 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
152 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
153 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
154 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
155 * defined in <linux/topology.h>.
157 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
158 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
161 /* work_structs for global per-cpu drains */
164 struct work_struct work;
166 static DEFINE_MUTEX(pcpu_drain_mutex);
167 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
169 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
170 volatile unsigned long latent_entropy __latent_entropy;
171 EXPORT_SYMBOL(latent_entropy);
175 * Array of node states.
177 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
178 [N_POSSIBLE] = NODE_MASK_ALL,
179 [N_ONLINE] = { { [0] = 1UL } },
181 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
182 #ifdef CONFIG_HIGHMEM
183 [N_HIGH_MEMORY] = { { [0] = 1UL } },
185 [N_MEMORY] = { { [0] = 1UL } },
186 [N_CPU] = { { [0] = 1UL } },
189 EXPORT_SYMBOL(node_states);
191 atomic_long_t _totalram_pages __read_mostly;
192 EXPORT_SYMBOL(_totalram_pages);
193 unsigned long totalreserve_pages __read_mostly;
194 unsigned long totalcma_pages __read_mostly;
196 int percpu_pagelist_high_fraction;
197 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
198 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
199 EXPORT_SYMBOL(init_on_alloc);
201 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
202 EXPORT_SYMBOL(init_on_free);
204 static bool _init_on_alloc_enabled_early __read_mostly
205 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
206 static int __init early_init_on_alloc(char *buf)
209 return kstrtobool(buf, &_init_on_alloc_enabled_early);
211 early_param("init_on_alloc", early_init_on_alloc);
213 static bool _init_on_free_enabled_early __read_mostly
214 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
215 static int __init early_init_on_free(char *buf)
217 return kstrtobool(buf, &_init_on_free_enabled_early);
219 early_param("init_on_free", early_init_on_free);
222 * A cached value of the page's pageblock's migratetype, used when the page is
223 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
224 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
225 * Also the migratetype set in the page does not necessarily match the pcplist
226 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
227 * other index - this ensures that it will be put on the correct CMA freelist.
229 static inline int get_pcppage_migratetype(struct page *page)
234 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
236 page->index = migratetype;
239 #ifdef CONFIG_PM_SLEEP
241 * The following functions are used by the suspend/hibernate code to temporarily
242 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
243 * while devices are suspended. To avoid races with the suspend/hibernate code,
244 * they should always be called with system_transition_mutex held
245 * (gfp_allowed_mask also should only be modified with system_transition_mutex
246 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
247 * with that modification).
250 static gfp_t saved_gfp_mask;
252 void pm_restore_gfp_mask(void)
254 WARN_ON(!mutex_is_locked(&system_transition_mutex));
255 if (saved_gfp_mask) {
256 gfp_allowed_mask = saved_gfp_mask;
261 void pm_restrict_gfp_mask(void)
263 WARN_ON(!mutex_is_locked(&system_transition_mutex));
264 WARN_ON(saved_gfp_mask);
265 saved_gfp_mask = gfp_allowed_mask;
266 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
269 bool pm_suspended_storage(void)
271 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
275 #endif /* CONFIG_PM_SLEEP */
277 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
278 unsigned int pageblock_order __read_mostly;
281 static void __free_pages_ok(struct page *page, unsigned int order,
285 * results with 256, 32 in the lowmem_reserve sysctl:
286 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
287 * 1G machine -> (16M dma, 784M normal, 224M high)
288 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
289 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
290 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
292 * TBD: should special case ZONE_DMA32 machines here - in those we normally
293 * don't need any ZONE_NORMAL reservation
295 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
296 #ifdef CONFIG_ZONE_DMA
299 #ifdef CONFIG_ZONE_DMA32
303 #ifdef CONFIG_HIGHMEM
309 static char * const zone_names[MAX_NR_ZONES] = {
310 #ifdef CONFIG_ZONE_DMA
313 #ifdef CONFIG_ZONE_DMA32
317 #ifdef CONFIG_HIGHMEM
321 #ifdef CONFIG_ZONE_DEVICE
326 const char * const migratetype_names[MIGRATE_TYPES] = {
334 #ifdef CONFIG_MEMORY_ISOLATION
339 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
340 [NULL_COMPOUND_DTOR] = NULL,
341 [COMPOUND_PAGE_DTOR] = free_compound_page,
342 #ifdef CONFIG_HUGETLB_PAGE
343 [HUGETLB_PAGE_DTOR] = free_huge_page,
345 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
346 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
350 int min_free_kbytes = 1024;
351 int user_min_free_kbytes = -1;
352 int watermark_boost_factor __read_mostly = 15000;
353 int watermark_scale_factor = 10;
355 static unsigned long nr_kernel_pages __initdata;
356 static unsigned long nr_all_pages __initdata;
357 static unsigned long dma_reserve __initdata;
359 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
360 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
361 static unsigned long required_kernelcore __initdata;
362 static unsigned long required_kernelcore_percent __initdata;
363 static unsigned long required_movablecore __initdata;
364 static unsigned long required_movablecore_percent __initdata;
365 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
366 static bool mirrored_kernelcore __meminitdata;
368 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
370 EXPORT_SYMBOL(movable_zone);
373 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
374 unsigned int nr_online_nodes __read_mostly = 1;
375 EXPORT_SYMBOL(nr_node_ids);
376 EXPORT_SYMBOL(nr_online_nodes);
379 int page_group_by_mobility_disabled __read_mostly;
381 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
383 * During boot we initialize deferred pages on-demand, as needed, but once
384 * page_alloc_init_late() has finished, the deferred pages are all initialized,
385 * and we can permanently disable that path.
387 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
390 * Calling kasan_free_pages() only after deferred memory initialization
391 * has completed. Poisoning pages during deferred memory init will greatly
392 * lengthen the process and cause problem in large memory systems as the
393 * deferred pages initialization is done with interrupt disabled.
395 * Assuming that there will be no reference to those newly initialized
396 * pages before they are ever allocated, this should have no effect on
397 * KASAN memory tracking as the poison will be properly inserted at page
398 * allocation time. The only corner case is when pages are allocated by
399 * on-demand allocation and then freed again before the deferred pages
400 * initialization is done, but this is not likely to happen.
402 static inline void kasan_free_nondeferred_pages(struct page *page, int order,
403 bool init, fpi_t fpi_flags)
405 if (static_branch_unlikely(&deferred_pages))
407 if (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
408 (fpi_flags & FPI_SKIP_KASAN_POISON))
410 kasan_free_pages(page, order, init);
413 /* Returns true if the struct page for the pfn is uninitialised */
414 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
416 int nid = early_pfn_to_nid(pfn);
418 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
425 * Returns true when the remaining initialisation should be deferred until
426 * later in the boot cycle when it can be parallelised.
428 static bool __meminit
429 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
431 static unsigned long prev_end_pfn, nr_initialised;
434 * prev_end_pfn static that contains the end of previous zone
435 * No need to protect because called very early in boot before smp_init.
437 if (prev_end_pfn != end_pfn) {
438 prev_end_pfn = end_pfn;
442 /* Always populate low zones for address-constrained allocations */
443 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
446 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
449 * We start only with one section of pages, more pages are added as
450 * needed until the rest of deferred pages are initialized.
453 if ((nr_initialised > PAGES_PER_SECTION) &&
454 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
455 NODE_DATA(nid)->first_deferred_pfn = pfn;
461 static inline void kasan_free_nondeferred_pages(struct page *page, int order,
462 bool init, fpi_t fpi_flags)
464 if (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
465 (fpi_flags & FPI_SKIP_KASAN_POISON))
467 kasan_free_pages(page, order, init);
470 static inline bool early_page_uninitialised(unsigned long pfn)
475 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
481 /* Return a pointer to the bitmap storing bits affecting a block of pages */
482 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
485 #ifdef CONFIG_SPARSEMEM
486 return section_to_usemap(__pfn_to_section(pfn));
488 return page_zone(page)->pageblock_flags;
489 #endif /* CONFIG_SPARSEMEM */
492 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
494 #ifdef CONFIG_SPARSEMEM
495 pfn &= (PAGES_PER_SECTION-1);
497 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
498 #endif /* CONFIG_SPARSEMEM */
499 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
502 static __always_inline
503 unsigned long __get_pfnblock_flags_mask(const struct page *page,
507 unsigned long *bitmap;
508 unsigned long bitidx, word_bitidx;
511 bitmap = get_pageblock_bitmap(page, pfn);
512 bitidx = pfn_to_bitidx(page, pfn);
513 word_bitidx = bitidx / BITS_PER_LONG;
514 bitidx &= (BITS_PER_LONG-1);
516 word = bitmap[word_bitidx];
517 return (word >> bitidx) & mask;
521 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
522 * @page: The page within the block of interest
523 * @pfn: The target page frame number
524 * @mask: mask of bits that the caller is interested in
526 * Return: pageblock_bits flags
528 unsigned long get_pfnblock_flags_mask(const struct page *page,
529 unsigned long pfn, unsigned long mask)
531 return __get_pfnblock_flags_mask(page, pfn, mask);
534 static __always_inline int get_pfnblock_migratetype(const struct page *page,
537 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
541 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
542 * @page: The page within the block of interest
543 * @flags: The flags to set
544 * @pfn: The target page frame number
545 * @mask: mask of bits that the caller is interested in
547 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
551 unsigned long *bitmap;
552 unsigned long bitidx, word_bitidx;
553 unsigned long old_word, word;
555 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
556 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
558 bitmap = get_pageblock_bitmap(page, pfn);
559 bitidx = pfn_to_bitidx(page, pfn);
560 word_bitidx = bitidx / BITS_PER_LONG;
561 bitidx &= (BITS_PER_LONG-1);
563 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
568 word = READ_ONCE(bitmap[word_bitidx]);
570 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
571 if (word == old_word)
577 void set_pageblock_migratetype(struct page *page, int migratetype)
579 if (unlikely(page_group_by_mobility_disabled &&
580 migratetype < MIGRATE_PCPTYPES))
581 migratetype = MIGRATE_UNMOVABLE;
583 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
584 page_to_pfn(page), MIGRATETYPE_MASK);
587 #ifdef CONFIG_DEBUG_VM
588 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
592 unsigned long pfn = page_to_pfn(page);
593 unsigned long sp, start_pfn;
596 seq = zone_span_seqbegin(zone);
597 start_pfn = zone->zone_start_pfn;
598 sp = zone->spanned_pages;
599 if (!zone_spans_pfn(zone, pfn))
601 } while (zone_span_seqretry(zone, seq));
604 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
605 pfn, zone_to_nid(zone), zone->name,
606 start_pfn, start_pfn + sp);
611 static int page_is_consistent(struct zone *zone, struct page *page)
613 if (!pfn_valid_within(page_to_pfn(page)))
615 if (zone != page_zone(page))
621 * Temporary debugging check for pages not lying within a given zone.
623 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
625 if (page_outside_zone_boundaries(zone, page))
627 if (!page_is_consistent(zone, page))
633 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
639 static void bad_page(struct page *page, const char *reason)
641 static unsigned long resume;
642 static unsigned long nr_shown;
643 static unsigned long nr_unshown;
646 * Allow a burst of 60 reports, then keep quiet for that minute;
647 * or allow a steady drip of one report per second.
649 if (nr_shown == 60) {
650 if (time_before(jiffies, resume)) {
656 "BUG: Bad page state: %lu messages suppressed\n",
663 resume = jiffies + 60 * HZ;
665 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
666 current->comm, page_to_pfn(page));
667 dump_page(page, reason);
672 /* Leave bad fields for debug, except PageBuddy could make trouble */
673 page_mapcount_reset(page); /* remove PageBuddy */
674 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
677 static inline void free_the_page(struct page *page, unsigned int order)
679 if (order == 0) /* Via pcp? */
680 free_unref_page(page);
682 __free_pages_ok(page, order, FPI_NONE);
686 * Higher-order pages are called "compound pages". They are structured thusly:
688 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
690 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
691 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
693 * The first tail page's ->compound_dtor holds the offset in array of compound
694 * page destructors. See compound_page_dtors.
696 * The first tail page's ->compound_order holds the order of allocation.
697 * This usage means that zero-order pages may not be compound.
700 void free_compound_page(struct page *page)
702 mem_cgroup_uncharge(page);
703 __free_pages_ok(page, compound_order(page), FPI_NONE);
706 void prep_compound_page(struct page *page, unsigned int order)
709 int nr_pages = 1 << order;
712 for (i = 1; i < nr_pages; i++) {
713 struct page *p = page + i;
714 set_page_count(p, 0);
715 p->mapping = TAIL_MAPPING;
716 set_compound_head(p, page);
719 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
720 set_compound_order(page, order);
721 atomic_set(compound_mapcount_ptr(page), -1);
722 if (hpage_pincount_available(page))
723 atomic_set(compound_pincount_ptr(page), 0);
726 #ifdef CONFIG_DEBUG_PAGEALLOC
727 unsigned int _debug_guardpage_minorder;
729 bool _debug_pagealloc_enabled_early __read_mostly
730 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
731 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
732 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
733 EXPORT_SYMBOL(_debug_pagealloc_enabled);
735 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
737 static int __init early_debug_pagealloc(char *buf)
739 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
741 early_param("debug_pagealloc", early_debug_pagealloc);
743 static int __init debug_guardpage_minorder_setup(char *buf)
747 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
748 pr_err("Bad debug_guardpage_minorder value\n");
751 _debug_guardpage_minorder = res;
752 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
755 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
757 static inline bool set_page_guard(struct zone *zone, struct page *page,
758 unsigned int order, int migratetype)
760 if (!debug_guardpage_enabled())
763 if (order >= debug_guardpage_minorder())
766 __SetPageGuard(page);
767 INIT_LIST_HEAD(&page->lru);
768 set_page_private(page, order);
769 /* Guard pages are not available for any usage */
770 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
775 static inline void clear_page_guard(struct zone *zone, struct page *page,
776 unsigned int order, int migratetype)
778 if (!debug_guardpage_enabled())
781 __ClearPageGuard(page);
783 set_page_private(page, 0);
784 if (!is_migrate_isolate(migratetype))
785 __mod_zone_freepage_state(zone, (1 << order), migratetype);
788 static inline bool set_page_guard(struct zone *zone, struct page *page,
789 unsigned int order, int migratetype) { return false; }
790 static inline void clear_page_guard(struct zone *zone, struct page *page,
791 unsigned int order, int migratetype) {}
795 * Enable static keys related to various memory debugging and hardening options.
796 * Some override others, and depend on early params that are evaluated in the
797 * order of appearance. So we need to first gather the full picture of what was
798 * enabled, and then make decisions.
800 void init_mem_debugging_and_hardening(void)
802 bool page_poisoning_requested = false;
804 #ifdef CONFIG_PAGE_POISONING
806 * Page poisoning is debug page alloc for some arches. If
807 * either of those options are enabled, enable poisoning.
809 if (page_poisoning_enabled() ||
810 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
811 debug_pagealloc_enabled())) {
812 static_branch_enable(&_page_poisoning_enabled);
813 page_poisoning_requested = true;
817 if (_init_on_alloc_enabled_early) {
818 if (page_poisoning_requested)
819 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
820 "will take precedence over init_on_alloc\n");
822 static_branch_enable(&init_on_alloc);
824 if (_init_on_free_enabled_early) {
825 if (page_poisoning_requested)
826 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
827 "will take precedence over init_on_free\n");
829 static_branch_enable(&init_on_free);
832 #ifdef CONFIG_DEBUG_PAGEALLOC
833 if (!debug_pagealloc_enabled())
836 static_branch_enable(&_debug_pagealloc_enabled);
838 if (!debug_guardpage_minorder())
841 static_branch_enable(&_debug_guardpage_enabled);
845 static inline void set_buddy_order(struct page *page, unsigned int order)
847 set_page_private(page, order);
848 __SetPageBuddy(page);
852 * This function checks whether a page is free && is the buddy
853 * we can coalesce a page and its buddy if
854 * (a) the buddy is not in a hole (check before calling!) &&
855 * (b) the buddy is in the buddy system &&
856 * (c) a page and its buddy have the same order &&
857 * (d) a page and its buddy are in the same zone.
859 * For recording whether a page is in the buddy system, we set PageBuddy.
860 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
862 * For recording page's order, we use page_private(page).
864 static inline bool page_is_buddy(struct page *page, struct page *buddy,
867 if (!page_is_guard(buddy) && !PageBuddy(buddy))
870 if (buddy_order(buddy) != order)
874 * zone check is done late to avoid uselessly calculating
875 * zone/node ids for pages that could never merge.
877 if (page_zone_id(page) != page_zone_id(buddy))
880 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
885 #ifdef CONFIG_COMPACTION
886 static inline struct capture_control *task_capc(struct zone *zone)
888 struct capture_control *capc = current->capture_control;
890 return unlikely(capc) &&
891 !(current->flags & PF_KTHREAD) &&
893 capc->cc->zone == zone ? capc : NULL;
897 compaction_capture(struct capture_control *capc, struct page *page,
898 int order, int migratetype)
900 if (!capc || order != capc->cc->order)
903 /* Do not accidentally pollute CMA or isolated regions*/
904 if (is_migrate_cma(migratetype) ||
905 is_migrate_isolate(migratetype))
909 * Do not let lower order allocations pollute a movable pageblock.
910 * This might let an unmovable request use a reclaimable pageblock
911 * and vice-versa but no more than normal fallback logic which can
912 * have trouble finding a high-order free page.
914 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
922 static inline struct capture_control *task_capc(struct zone *zone)
928 compaction_capture(struct capture_control *capc, struct page *page,
929 int order, int migratetype)
933 #endif /* CONFIG_COMPACTION */
935 /* Used for pages not on another list */
936 static inline void add_to_free_list(struct page *page, struct zone *zone,
937 unsigned int order, int migratetype)
939 struct free_area *area = &zone->free_area[order];
941 list_add(&page->lru, &area->free_list[migratetype]);
945 /* Used for pages not on another list */
946 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
947 unsigned int order, int migratetype)
949 struct free_area *area = &zone->free_area[order];
951 list_add_tail(&page->lru, &area->free_list[migratetype]);
956 * Used for pages which are on another list. Move the pages to the tail
957 * of the list - so the moved pages won't immediately be considered for
958 * allocation again (e.g., optimization for memory onlining).
960 static inline void move_to_free_list(struct page *page, struct zone *zone,
961 unsigned int order, int migratetype)
963 struct free_area *area = &zone->free_area[order];
965 list_move_tail(&page->lru, &area->free_list[migratetype]);
968 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
971 /* clear reported state and update reported page count */
972 if (page_reported(page))
973 __ClearPageReported(page);
975 list_del(&page->lru);
976 __ClearPageBuddy(page);
977 set_page_private(page, 0);
978 zone->free_area[order].nr_free--;
982 * If this is not the largest possible page, check if the buddy
983 * of the next-highest order is free. If it is, it's possible
984 * that pages are being freed that will coalesce soon. In case,
985 * that is happening, add the free page to the tail of the list
986 * so it's less likely to be used soon and more likely to be merged
987 * as a higher order page
990 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
991 struct page *page, unsigned int order)
993 struct page *higher_page, *higher_buddy;
994 unsigned long combined_pfn;
996 if (order >= MAX_ORDER - 2)
999 if (!pfn_valid_within(buddy_pfn))
1002 combined_pfn = buddy_pfn & pfn;
1003 higher_page = page + (combined_pfn - pfn);
1004 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1005 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1007 return pfn_valid_within(buddy_pfn) &&
1008 page_is_buddy(higher_page, higher_buddy, order + 1);
1012 * Freeing function for a buddy system allocator.
1014 * The concept of a buddy system is to maintain direct-mapped table
1015 * (containing bit values) for memory blocks of various "orders".
1016 * The bottom level table contains the map for the smallest allocatable
1017 * units of memory (here, pages), and each level above it describes
1018 * pairs of units from the levels below, hence, "buddies".
1019 * At a high level, all that happens here is marking the table entry
1020 * at the bottom level available, and propagating the changes upward
1021 * as necessary, plus some accounting needed to play nicely with other
1022 * parts of the VM system.
1023 * At each level, we keep a list of pages, which are heads of continuous
1024 * free pages of length of (1 << order) and marked with PageBuddy.
1025 * Page's order is recorded in page_private(page) field.
1026 * So when we are allocating or freeing one, we can derive the state of the
1027 * other. That is, if we allocate a small block, and both were
1028 * free, the remainder of the region must be split into blocks.
1029 * If a block is freed, and its buddy is also free, then this
1030 * triggers coalescing into a block of larger size.
1035 static inline void __free_one_page(struct page *page,
1037 struct zone *zone, unsigned int order,
1038 int migratetype, fpi_t fpi_flags)
1040 struct capture_control *capc = task_capc(zone);
1041 unsigned long buddy_pfn;
1042 unsigned long combined_pfn;
1043 unsigned int max_order;
1047 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1049 VM_BUG_ON(!zone_is_initialized(zone));
1050 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1052 VM_BUG_ON(migratetype == -1);
1053 if (likely(!is_migrate_isolate(migratetype)))
1054 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1056 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1057 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1060 while (order < max_order) {
1061 if (compaction_capture(capc, page, order, migratetype)) {
1062 __mod_zone_freepage_state(zone, -(1 << order),
1066 buddy_pfn = __find_buddy_pfn(pfn, order);
1067 buddy = page + (buddy_pfn - pfn);
1069 if (!pfn_valid_within(buddy_pfn))
1071 if (!page_is_buddy(page, buddy, order))
1074 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1075 * merge with it and move up one order.
1077 if (page_is_guard(buddy))
1078 clear_page_guard(zone, buddy, order, migratetype);
1080 del_page_from_free_list(buddy, zone, order);
1081 combined_pfn = buddy_pfn & pfn;
1082 page = page + (combined_pfn - pfn);
1086 if (order < MAX_ORDER - 1) {
1087 /* If we are here, it means order is >= pageblock_order.
1088 * We want to prevent merge between freepages on isolate
1089 * pageblock and normal pageblock. Without this, pageblock
1090 * isolation could cause incorrect freepage or CMA accounting.
1092 * We don't want to hit this code for the more frequent
1093 * low-order merging.
1095 if (unlikely(has_isolate_pageblock(zone))) {
1098 buddy_pfn = __find_buddy_pfn(pfn, order);
1099 buddy = page + (buddy_pfn - pfn);
1100 buddy_mt = get_pageblock_migratetype(buddy);
1102 if (migratetype != buddy_mt
1103 && (is_migrate_isolate(migratetype) ||
1104 is_migrate_isolate(buddy_mt)))
1107 max_order = order + 1;
1108 goto continue_merging;
1112 set_buddy_order(page, order);
1114 if (fpi_flags & FPI_TO_TAIL)
1116 else if (is_shuffle_order(order))
1117 to_tail = shuffle_pick_tail();
1119 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1122 add_to_free_list_tail(page, zone, order, migratetype);
1124 add_to_free_list(page, zone, order, migratetype);
1126 /* Notify page reporting subsystem of freed page */
1127 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1128 page_reporting_notify_free(order);
1132 * A bad page could be due to a number of fields. Instead of multiple branches,
1133 * try and check multiple fields with one check. The caller must do a detailed
1134 * check if necessary.
1136 static inline bool page_expected_state(struct page *page,
1137 unsigned long check_flags)
1139 if (unlikely(atomic_read(&page->_mapcount) != -1))
1142 if (unlikely((unsigned long)page->mapping |
1143 page_ref_count(page) |
1147 (page->flags & check_flags)))
1153 static const char *page_bad_reason(struct page *page, unsigned long flags)
1155 const char *bad_reason = NULL;
1157 if (unlikely(atomic_read(&page->_mapcount) != -1))
1158 bad_reason = "nonzero mapcount";
1159 if (unlikely(page->mapping != NULL))
1160 bad_reason = "non-NULL mapping";
1161 if (unlikely(page_ref_count(page) != 0))
1162 bad_reason = "nonzero _refcount";
1163 if (unlikely(page->flags & flags)) {
1164 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1165 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1167 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1170 if (unlikely(page->memcg_data))
1171 bad_reason = "page still charged to cgroup";
1176 static void check_free_page_bad(struct page *page)
1179 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1182 static inline int check_free_page(struct page *page)
1184 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1187 /* Something has gone sideways, find it */
1188 check_free_page_bad(page);
1192 static int free_tail_pages_check(struct page *head_page, struct page *page)
1197 * We rely page->lru.next never has bit 0 set, unless the page
1198 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1200 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1202 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1206 switch (page - head_page) {
1208 /* the first tail page: ->mapping may be compound_mapcount() */
1209 if (unlikely(compound_mapcount(page))) {
1210 bad_page(page, "nonzero compound_mapcount");
1216 * the second tail page: ->mapping is
1217 * deferred_list.next -- ignore value.
1221 if (page->mapping != TAIL_MAPPING) {
1222 bad_page(page, "corrupted mapping in tail page");
1227 if (unlikely(!PageTail(page))) {
1228 bad_page(page, "PageTail not set");
1231 if (unlikely(compound_head(page) != head_page)) {
1232 bad_page(page, "compound_head not consistent");
1237 page->mapping = NULL;
1238 clear_compound_head(page);
1242 static void kernel_init_free_pages(struct page *page, int numpages)
1246 /* s390's use of memset() could override KASAN redzones. */
1247 kasan_disable_current();
1248 for (i = 0; i < numpages; i++) {
1249 u8 tag = page_kasan_tag(page + i);
1250 page_kasan_tag_reset(page + i);
1251 clear_highpage(page + i);
1252 page_kasan_tag_set(page + i, tag);
1254 kasan_enable_current();
1257 static __always_inline bool free_pages_prepare(struct page *page,
1258 unsigned int order, bool check_free, fpi_t fpi_flags)
1263 VM_BUG_ON_PAGE(PageTail(page), page);
1265 trace_mm_page_free(page, order);
1267 if (unlikely(PageHWPoison(page)) && !order) {
1269 * Do not let hwpoison pages hit pcplists/buddy
1270 * Untie memcg state and reset page's owner
1272 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1273 __memcg_kmem_uncharge_page(page, order);
1274 reset_page_owner(page, order);
1279 * Check tail pages before head page information is cleared to
1280 * avoid checking PageCompound for order-0 pages.
1282 if (unlikely(order)) {
1283 bool compound = PageCompound(page);
1286 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1289 ClearPageDoubleMap(page);
1290 for (i = 1; i < (1 << order); i++) {
1292 bad += free_tail_pages_check(page, page + i);
1293 if (unlikely(check_free_page(page + i))) {
1297 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1300 if (PageMappingFlags(page))
1301 page->mapping = NULL;
1302 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1303 __memcg_kmem_uncharge_page(page, order);
1305 bad += check_free_page(page);
1309 page_cpupid_reset_last(page);
1310 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1311 reset_page_owner(page, order);
1313 if (!PageHighMem(page)) {
1314 debug_check_no_locks_freed(page_address(page),
1315 PAGE_SIZE << order);
1316 debug_check_no_obj_freed(page_address(page),
1317 PAGE_SIZE << order);
1320 kernel_poison_pages(page, 1 << order);
1323 * As memory initialization might be integrated into KASAN,
1324 * kasan_free_pages and kernel_init_free_pages must be
1325 * kept together to avoid discrepancies in behavior.
1327 * With hardware tag-based KASAN, memory tags must be set before the
1328 * page becomes unavailable via debug_pagealloc or arch_free_page.
1330 init = want_init_on_free();
1331 if (init && !kasan_has_integrated_init())
1332 kernel_init_free_pages(page, 1 << order);
1333 kasan_free_nondeferred_pages(page, order, init, fpi_flags);
1336 * arch_free_page() can make the page's contents inaccessible. s390
1337 * does this. So nothing which can access the page's contents should
1338 * happen after this.
1340 arch_free_page(page, order);
1342 debug_pagealloc_unmap_pages(page, 1 << order);
1347 #ifdef CONFIG_DEBUG_VM
1349 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1350 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1351 * moved from pcp lists to free lists.
1353 static bool free_pcp_prepare(struct page *page)
1355 return free_pages_prepare(page, 0, true, FPI_NONE);
1358 static bool bulkfree_pcp_prepare(struct page *page)
1360 if (debug_pagealloc_enabled_static())
1361 return check_free_page(page);
1367 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1368 * moving from pcp lists to free list in order to reduce overhead. With
1369 * debug_pagealloc enabled, they are checked also immediately when being freed
1372 static bool free_pcp_prepare(struct page *page)
1374 if (debug_pagealloc_enabled_static())
1375 return free_pages_prepare(page, 0, true, FPI_NONE);
1377 return free_pages_prepare(page, 0, false, FPI_NONE);
1380 static bool bulkfree_pcp_prepare(struct page *page)
1382 return check_free_page(page);
1384 #endif /* CONFIG_DEBUG_VM */
1386 static inline void prefetch_buddy(struct page *page)
1388 unsigned long pfn = page_to_pfn(page);
1389 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1390 struct page *buddy = page + (buddy_pfn - pfn);
1396 * Frees a number of pages from the PCP lists
1397 * Assumes all pages on list are in same zone, and of same order.
1398 * count is the number of pages to free.
1400 * If the zone was previously in an "all pages pinned" state then look to
1401 * see if this freeing clears that state.
1403 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1404 * pinned" detection logic.
1406 static void free_pcppages_bulk(struct zone *zone, int count,
1407 struct per_cpu_pages *pcp)
1409 int migratetype = 0;
1411 int prefetch_nr = READ_ONCE(pcp->batch);
1412 bool isolated_pageblocks;
1413 struct page *page, *tmp;
1417 * Ensure proper count is passed which otherwise would stuck in the
1418 * below while (list_empty(list)) loop.
1420 count = min(pcp->count, count);
1422 struct list_head *list;
1425 * Remove pages from lists in a round-robin fashion. A
1426 * batch_free count is maintained that is incremented when an
1427 * empty list is encountered. This is so more pages are freed
1428 * off fuller lists instead of spinning excessively around empty
1433 if (++migratetype == MIGRATE_PCPTYPES)
1435 list = &pcp->lists[migratetype];
1436 } while (list_empty(list));
1438 /* This is the only non-empty list. Free them all. */
1439 if (batch_free == MIGRATE_PCPTYPES)
1443 page = list_last_entry(list, struct page, lru);
1444 /* must delete to avoid corrupting pcp list */
1445 list_del(&page->lru);
1448 if (bulkfree_pcp_prepare(page))
1451 list_add_tail(&page->lru, &head);
1454 * We are going to put the page back to the global
1455 * pool, prefetch its buddy to speed up later access
1456 * under zone->lock. It is believed the overhead of
1457 * an additional test and calculating buddy_pfn here
1458 * can be offset by reduced memory latency later. To
1459 * avoid excessive prefetching due to large count, only
1460 * prefetch buddy for the first pcp->batch nr of pages.
1463 prefetch_buddy(page);
1466 } while (--count && --batch_free && !list_empty(list));
1470 * local_lock_irq held so equivalent to spin_lock_irqsave for
1471 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1473 spin_lock(&zone->lock);
1474 isolated_pageblocks = has_isolate_pageblock(zone);
1477 * Use safe version since after __free_one_page(),
1478 * page->lru.next will not point to original list.
1480 list_for_each_entry_safe(page, tmp, &head, lru) {
1481 int mt = get_pcppage_migratetype(page);
1482 /* MIGRATE_ISOLATE page should not go to pcplists */
1483 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1484 /* Pageblock could have been isolated meanwhile */
1485 if (unlikely(isolated_pageblocks))
1486 mt = get_pageblock_migratetype(page);
1488 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1489 trace_mm_page_pcpu_drain(page, 0, mt);
1491 spin_unlock(&zone->lock);
1494 static void free_one_page(struct zone *zone,
1495 struct page *page, unsigned long pfn,
1497 int migratetype, fpi_t fpi_flags)
1499 unsigned long flags;
1501 spin_lock_irqsave(&zone->lock, flags);
1502 if (unlikely(has_isolate_pageblock(zone) ||
1503 is_migrate_isolate(migratetype))) {
1504 migratetype = get_pfnblock_migratetype(page, pfn);
1506 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1507 spin_unlock_irqrestore(&zone->lock, flags);
1510 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1511 unsigned long zone, int nid)
1513 mm_zero_struct_page(page);
1514 set_page_links(page, zone, nid, pfn);
1515 init_page_count(page);
1516 page_mapcount_reset(page);
1517 page_cpupid_reset_last(page);
1518 page_kasan_tag_reset(page);
1520 INIT_LIST_HEAD(&page->lru);
1521 #ifdef WANT_PAGE_VIRTUAL
1522 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1523 if (!is_highmem_idx(zone))
1524 set_page_address(page, __va(pfn << PAGE_SHIFT));
1528 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1529 static void __meminit init_reserved_page(unsigned long pfn)
1534 if (!early_page_uninitialised(pfn))
1537 nid = early_pfn_to_nid(pfn);
1538 pgdat = NODE_DATA(nid);
1540 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1541 struct zone *zone = &pgdat->node_zones[zid];
1543 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1546 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1549 static inline void init_reserved_page(unsigned long pfn)
1552 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1555 * Initialised pages do not have PageReserved set. This function is
1556 * called for each range allocated by the bootmem allocator and
1557 * marks the pages PageReserved. The remaining valid pages are later
1558 * sent to the buddy page allocator.
1560 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1562 unsigned long start_pfn = PFN_DOWN(start);
1563 unsigned long end_pfn = PFN_UP(end);
1565 for (; start_pfn < end_pfn; start_pfn++) {
1566 if (pfn_valid(start_pfn)) {
1567 struct page *page = pfn_to_page(start_pfn);
1569 init_reserved_page(start_pfn);
1571 /* Avoid false-positive PageTail() */
1572 INIT_LIST_HEAD(&page->lru);
1575 * no need for atomic set_bit because the struct
1576 * page is not visible yet so nobody should
1579 __SetPageReserved(page);
1584 static void __free_pages_ok(struct page *page, unsigned int order,
1587 unsigned long flags;
1589 unsigned long pfn = page_to_pfn(page);
1590 struct zone *zone = page_zone(page);
1592 if (!free_pages_prepare(page, order, true, fpi_flags))
1595 migratetype = get_pfnblock_migratetype(page, pfn);
1597 spin_lock_irqsave(&zone->lock, flags);
1598 if (unlikely(has_isolate_pageblock(zone) ||
1599 is_migrate_isolate(migratetype))) {
1600 migratetype = get_pfnblock_migratetype(page, pfn);
1602 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1603 spin_unlock_irqrestore(&zone->lock, flags);
1605 __count_vm_events(PGFREE, 1 << order);
1608 void __free_pages_core(struct page *page, unsigned int order)
1610 unsigned int nr_pages = 1 << order;
1611 struct page *p = page;
1615 * When initializing the memmap, __init_single_page() sets the refcount
1616 * of all pages to 1 ("allocated"/"not free"). We have to set the
1617 * refcount of all involved pages to 0.
1620 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1622 __ClearPageReserved(p);
1623 set_page_count(p, 0);
1625 __ClearPageReserved(p);
1626 set_page_count(p, 0);
1628 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1631 * Bypass PCP and place fresh pages right to the tail, primarily
1632 * relevant for memory onlining.
1634 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1640 * During memory init memblocks map pfns to nids. The search is expensive and
1641 * this caches recent lookups. The implementation of __early_pfn_to_nid
1642 * treats start/end as pfns.
1644 struct mminit_pfnnid_cache {
1645 unsigned long last_start;
1646 unsigned long last_end;
1650 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1653 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1655 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1656 struct mminit_pfnnid_cache *state)
1658 unsigned long start_pfn, end_pfn;
1661 if (state->last_start <= pfn && pfn < state->last_end)
1662 return state->last_nid;
1664 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1665 if (nid != NUMA_NO_NODE) {
1666 state->last_start = start_pfn;
1667 state->last_end = end_pfn;
1668 state->last_nid = nid;
1674 int __meminit early_pfn_to_nid(unsigned long pfn)
1676 static DEFINE_SPINLOCK(early_pfn_lock);
1679 spin_lock(&early_pfn_lock);
1680 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1682 nid = first_online_node;
1683 spin_unlock(&early_pfn_lock);
1687 #endif /* CONFIG_NUMA */
1689 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1692 if (early_page_uninitialised(pfn))
1694 __free_pages_core(page, order);
1698 * Check that the whole (or subset of) a pageblock given by the interval of
1699 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1700 * with the migration of free compaction scanner. The scanners then need to
1701 * use only pfn_valid_within() check for arches that allow holes within
1704 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1706 * It's possible on some configurations to have a setup like node0 node1 node0
1707 * i.e. it's possible that all pages within a zones range of pages do not
1708 * belong to a single zone. We assume that a border between node0 and node1
1709 * can occur within a single pageblock, but not a node0 node1 node0
1710 * interleaving within a single pageblock. It is therefore sufficient to check
1711 * the first and last page of a pageblock and avoid checking each individual
1712 * page in a pageblock.
1714 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1715 unsigned long end_pfn, struct zone *zone)
1717 struct page *start_page;
1718 struct page *end_page;
1720 /* end_pfn is one past the range we are checking */
1723 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1726 start_page = pfn_to_online_page(start_pfn);
1730 if (page_zone(start_page) != zone)
1733 end_page = pfn_to_page(end_pfn);
1735 /* This gives a shorter code than deriving page_zone(end_page) */
1736 if (page_zone_id(start_page) != page_zone_id(end_page))
1742 void set_zone_contiguous(struct zone *zone)
1744 unsigned long block_start_pfn = zone->zone_start_pfn;
1745 unsigned long block_end_pfn;
1747 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1748 for (; block_start_pfn < zone_end_pfn(zone);
1749 block_start_pfn = block_end_pfn,
1750 block_end_pfn += pageblock_nr_pages) {
1752 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1754 if (!__pageblock_pfn_to_page(block_start_pfn,
1755 block_end_pfn, zone))
1760 /* We confirm that there is no hole */
1761 zone->contiguous = true;
1764 void clear_zone_contiguous(struct zone *zone)
1766 zone->contiguous = false;
1769 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1770 static void __init deferred_free_range(unsigned long pfn,
1771 unsigned long nr_pages)
1779 page = pfn_to_page(pfn);
1781 /* Free a large naturally-aligned chunk if possible */
1782 if (nr_pages == pageblock_nr_pages &&
1783 (pfn & (pageblock_nr_pages - 1)) == 0) {
1784 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1785 __free_pages_core(page, pageblock_order);
1789 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1790 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1791 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1792 __free_pages_core(page, 0);
1796 /* Completion tracking for deferred_init_memmap() threads */
1797 static atomic_t pgdat_init_n_undone __initdata;
1798 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1800 static inline void __init pgdat_init_report_one_done(void)
1802 if (atomic_dec_and_test(&pgdat_init_n_undone))
1803 complete(&pgdat_init_all_done_comp);
1807 * Returns true if page needs to be initialized or freed to buddy allocator.
1809 * First we check if pfn is valid on architectures where it is possible to have
1810 * holes within pageblock_nr_pages. On systems where it is not possible, this
1811 * function is optimized out.
1813 * Then, we check if a current large page is valid by only checking the validity
1816 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1818 if (!pfn_valid_within(pfn))
1820 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1826 * Free pages to buddy allocator. Try to free aligned pages in
1827 * pageblock_nr_pages sizes.
1829 static void __init deferred_free_pages(unsigned long pfn,
1830 unsigned long end_pfn)
1832 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1833 unsigned long nr_free = 0;
1835 for (; pfn < end_pfn; pfn++) {
1836 if (!deferred_pfn_valid(pfn)) {
1837 deferred_free_range(pfn - nr_free, nr_free);
1839 } else if (!(pfn & nr_pgmask)) {
1840 deferred_free_range(pfn - nr_free, nr_free);
1846 /* Free the last block of pages to allocator */
1847 deferred_free_range(pfn - nr_free, nr_free);
1851 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1852 * by performing it only once every pageblock_nr_pages.
1853 * Return number of pages initialized.
1855 static unsigned long __init deferred_init_pages(struct zone *zone,
1857 unsigned long end_pfn)
1859 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1860 int nid = zone_to_nid(zone);
1861 unsigned long nr_pages = 0;
1862 int zid = zone_idx(zone);
1863 struct page *page = NULL;
1865 for (; pfn < end_pfn; pfn++) {
1866 if (!deferred_pfn_valid(pfn)) {
1869 } else if (!page || !(pfn & nr_pgmask)) {
1870 page = pfn_to_page(pfn);
1874 __init_single_page(page, pfn, zid, nid);
1881 * This function is meant to pre-load the iterator for the zone init.
1882 * Specifically it walks through the ranges until we are caught up to the
1883 * first_init_pfn value and exits there. If we never encounter the value we
1884 * return false indicating there are no valid ranges left.
1887 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1888 unsigned long *spfn, unsigned long *epfn,
1889 unsigned long first_init_pfn)
1894 * Start out by walking through the ranges in this zone that have
1895 * already been initialized. We don't need to do anything with them
1896 * so we just need to flush them out of the system.
1898 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1899 if (*epfn <= first_init_pfn)
1901 if (*spfn < first_init_pfn)
1902 *spfn = first_init_pfn;
1911 * Initialize and free pages. We do it in two loops: first we initialize
1912 * struct page, then free to buddy allocator, because while we are
1913 * freeing pages we can access pages that are ahead (computing buddy
1914 * page in __free_one_page()).
1916 * In order to try and keep some memory in the cache we have the loop
1917 * broken along max page order boundaries. This way we will not cause
1918 * any issues with the buddy page computation.
1920 static unsigned long __init
1921 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1922 unsigned long *end_pfn)
1924 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1925 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1926 unsigned long nr_pages = 0;
1929 /* First we loop through and initialize the page values */
1930 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1933 if (mo_pfn <= *start_pfn)
1936 t = min(mo_pfn, *end_pfn);
1937 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1939 if (mo_pfn < *end_pfn) {
1940 *start_pfn = mo_pfn;
1945 /* Reset values and now loop through freeing pages as needed */
1948 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1954 t = min(mo_pfn, epfn);
1955 deferred_free_pages(spfn, t);
1965 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1968 unsigned long spfn, epfn;
1969 struct zone *zone = arg;
1972 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1975 * Initialize and free pages in MAX_ORDER sized increments so that we
1976 * can avoid introducing any issues with the buddy allocator.
1978 while (spfn < end_pfn) {
1979 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1984 /* An arch may override for more concurrency. */
1986 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1991 /* Initialise remaining memory on a node */
1992 static int __init deferred_init_memmap(void *data)
1994 pg_data_t *pgdat = data;
1995 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1996 unsigned long spfn = 0, epfn = 0;
1997 unsigned long first_init_pfn, flags;
1998 unsigned long start = jiffies;
2000 int zid, max_threads;
2003 /* Bind memory initialisation thread to a local node if possible */
2004 if (!cpumask_empty(cpumask))
2005 set_cpus_allowed_ptr(current, cpumask);
2007 pgdat_resize_lock(pgdat, &flags);
2008 first_init_pfn = pgdat->first_deferred_pfn;
2009 if (first_init_pfn == ULONG_MAX) {
2010 pgdat_resize_unlock(pgdat, &flags);
2011 pgdat_init_report_one_done();
2015 /* Sanity check boundaries */
2016 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2017 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2018 pgdat->first_deferred_pfn = ULONG_MAX;
2021 * Once we unlock here, the zone cannot be grown anymore, thus if an
2022 * interrupt thread must allocate this early in boot, zone must be
2023 * pre-grown prior to start of deferred page initialization.
2025 pgdat_resize_unlock(pgdat, &flags);
2027 /* Only the highest zone is deferred so find it */
2028 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2029 zone = pgdat->node_zones + zid;
2030 if (first_init_pfn < zone_end_pfn(zone))
2034 /* If the zone is empty somebody else may have cleared out the zone */
2035 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2039 max_threads = deferred_page_init_max_threads(cpumask);
2041 while (spfn < epfn) {
2042 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2043 struct padata_mt_job job = {
2044 .thread_fn = deferred_init_memmap_chunk,
2047 .size = epfn_align - spfn,
2048 .align = PAGES_PER_SECTION,
2049 .min_chunk = PAGES_PER_SECTION,
2050 .max_threads = max_threads,
2053 padata_do_multithreaded(&job);
2054 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2058 /* Sanity check that the next zone really is unpopulated */
2059 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2061 pr_info("node %d deferred pages initialised in %ums\n",
2062 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2064 pgdat_init_report_one_done();
2069 * If this zone has deferred pages, try to grow it by initializing enough
2070 * deferred pages to satisfy the allocation specified by order, rounded up to
2071 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2072 * of SECTION_SIZE bytes by initializing struct pages in increments of
2073 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2075 * Return true when zone was grown, otherwise return false. We return true even
2076 * when we grow less than requested, to let the caller decide if there are
2077 * enough pages to satisfy the allocation.
2079 * Note: We use noinline because this function is needed only during boot, and
2080 * it is called from a __ref function _deferred_grow_zone. This way we are
2081 * making sure that it is not inlined into permanent text section.
2083 static noinline bool __init
2084 deferred_grow_zone(struct zone *zone, unsigned int order)
2086 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2087 pg_data_t *pgdat = zone->zone_pgdat;
2088 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2089 unsigned long spfn, epfn, flags;
2090 unsigned long nr_pages = 0;
2093 /* Only the last zone may have deferred pages */
2094 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2097 pgdat_resize_lock(pgdat, &flags);
2100 * If someone grew this zone while we were waiting for spinlock, return
2101 * true, as there might be enough pages already.
2103 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2104 pgdat_resize_unlock(pgdat, &flags);
2108 /* If the zone is empty somebody else may have cleared out the zone */
2109 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2110 first_deferred_pfn)) {
2111 pgdat->first_deferred_pfn = ULONG_MAX;
2112 pgdat_resize_unlock(pgdat, &flags);
2113 /* Retry only once. */
2114 return first_deferred_pfn != ULONG_MAX;
2118 * Initialize and free pages in MAX_ORDER sized increments so
2119 * that we can avoid introducing any issues with the buddy
2122 while (spfn < epfn) {
2123 /* update our first deferred PFN for this section */
2124 first_deferred_pfn = spfn;
2126 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2127 touch_nmi_watchdog();
2129 /* We should only stop along section boundaries */
2130 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2133 /* If our quota has been met we can stop here */
2134 if (nr_pages >= nr_pages_needed)
2138 pgdat->first_deferred_pfn = spfn;
2139 pgdat_resize_unlock(pgdat, &flags);
2141 return nr_pages > 0;
2145 * deferred_grow_zone() is __init, but it is called from
2146 * get_page_from_freelist() during early boot until deferred_pages permanently
2147 * disables this call. This is why we have refdata wrapper to avoid warning,
2148 * and to ensure that the function body gets unloaded.
2151 _deferred_grow_zone(struct zone *zone, unsigned int order)
2153 return deferred_grow_zone(zone, order);
2156 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2158 void __init page_alloc_init_late(void)
2163 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2165 /* There will be num_node_state(N_MEMORY) threads */
2166 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2167 for_each_node_state(nid, N_MEMORY) {
2168 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2171 /* Block until all are initialised */
2172 wait_for_completion(&pgdat_init_all_done_comp);
2175 * We initialized the rest of the deferred pages. Permanently disable
2176 * on-demand struct page initialization.
2178 static_branch_disable(&deferred_pages);
2180 /* Reinit limits that are based on free pages after the kernel is up */
2181 files_maxfiles_init();
2186 /* Discard memblock private memory */
2189 for_each_node_state(nid, N_MEMORY)
2190 shuffle_free_memory(NODE_DATA(nid));
2192 for_each_populated_zone(zone)
2193 set_zone_contiguous(zone);
2197 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2198 void __init init_cma_reserved_pageblock(struct page *page)
2200 unsigned i = pageblock_nr_pages;
2201 struct page *p = page;
2204 __ClearPageReserved(p);
2205 set_page_count(p, 0);
2208 set_pageblock_migratetype(page, MIGRATE_CMA);
2210 if (pageblock_order >= MAX_ORDER) {
2211 i = pageblock_nr_pages;
2214 set_page_refcounted(p);
2215 __free_pages(p, MAX_ORDER - 1);
2216 p += MAX_ORDER_NR_PAGES;
2217 } while (i -= MAX_ORDER_NR_PAGES);
2219 set_page_refcounted(page);
2220 __free_pages(page, pageblock_order);
2223 adjust_managed_page_count(page, pageblock_nr_pages);
2224 page_zone(page)->cma_pages += pageblock_nr_pages;
2229 * The order of subdivision here is critical for the IO subsystem.
2230 * Please do not alter this order without good reasons and regression
2231 * testing. Specifically, as large blocks of memory are subdivided,
2232 * the order in which smaller blocks are delivered depends on the order
2233 * they're subdivided in this function. This is the primary factor
2234 * influencing the order in which pages are delivered to the IO
2235 * subsystem according to empirical testing, and this is also justified
2236 * by considering the behavior of a buddy system containing a single
2237 * large block of memory acted on by a series of small allocations.
2238 * This behavior is a critical factor in sglist merging's success.
2242 static inline void expand(struct zone *zone, struct page *page,
2243 int low, int high, int migratetype)
2245 unsigned long size = 1 << high;
2247 while (high > low) {
2250 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2253 * Mark as guard pages (or page), that will allow to
2254 * merge back to allocator when buddy will be freed.
2255 * Corresponding page table entries will not be touched,
2256 * pages will stay not present in virtual address space
2258 if (set_page_guard(zone, &page[size], high, migratetype))
2261 add_to_free_list(&page[size], zone, high, migratetype);
2262 set_buddy_order(&page[size], high);
2266 static void check_new_page_bad(struct page *page)
2268 if (unlikely(page->flags & __PG_HWPOISON)) {
2269 /* Don't complain about hwpoisoned pages */
2270 page_mapcount_reset(page); /* remove PageBuddy */
2275 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2279 * This page is about to be returned from the page allocator
2281 static inline int check_new_page(struct page *page)
2283 if (likely(page_expected_state(page,
2284 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2287 check_new_page_bad(page);
2291 #ifdef CONFIG_DEBUG_VM
2293 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2294 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2295 * also checked when pcp lists are refilled from the free lists.
2297 static inline bool check_pcp_refill(struct page *page)
2299 if (debug_pagealloc_enabled_static())
2300 return check_new_page(page);
2305 static inline bool check_new_pcp(struct page *page)
2307 return check_new_page(page);
2311 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2312 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2313 * enabled, they are also checked when being allocated from the pcp lists.
2315 static inline bool check_pcp_refill(struct page *page)
2317 return check_new_page(page);
2319 static inline bool check_new_pcp(struct page *page)
2321 if (debug_pagealloc_enabled_static())
2322 return check_new_page(page);
2326 #endif /* CONFIG_DEBUG_VM */
2328 static bool check_new_pages(struct page *page, unsigned int order)
2331 for (i = 0; i < (1 << order); i++) {
2332 struct page *p = page + i;
2334 if (unlikely(check_new_page(p)))
2341 inline void post_alloc_hook(struct page *page, unsigned int order,
2346 set_page_private(page, 0);
2347 set_page_refcounted(page);
2349 arch_alloc_page(page, order);
2350 debug_pagealloc_map_pages(page, 1 << order);
2353 * Page unpoisoning must happen before memory initialization.
2354 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2355 * allocations and the page unpoisoning code will complain.
2357 kernel_unpoison_pages(page, 1 << order);
2360 * As memory initialization might be integrated into KASAN,
2361 * kasan_alloc_pages and kernel_init_free_pages must be
2362 * kept together to avoid discrepancies in behavior.
2364 init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2365 kasan_alloc_pages(page, order, init);
2366 if (init && !kasan_has_integrated_init())
2367 kernel_init_free_pages(page, 1 << order);
2369 set_page_owner(page, order, gfp_flags);
2372 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2373 unsigned int alloc_flags)
2375 post_alloc_hook(page, order, gfp_flags);
2377 if (order && (gfp_flags & __GFP_COMP))
2378 prep_compound_page(page, order);
2381 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2382 * allocate the page. The expectation is that the caller is taking
2383 * steps that will free more memory. The caller should avoid the page
2384 * being used for !PFMEMALLOC purposes.
2386 if (alloc_flags & ALLOC_NO_WATERMARKS)
2387 set_page_pfmemalloc(page);
2389 clear_page_pfmemalloc(page);
2393 * Go through the free lists for the given migratetype and remove
2394 * the smallest available page from the freelists
2396 static __always_inline
2397 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2400 unsigned int current_order;
2401 struct free_area *area;
2404 /* Find a page of the appropriate size in the preferred list */
2405 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2406 area = &(zone->free_area[current_order]);
2407 page = get_page_from_free_area(area, migratetype);
2410 del_page_from_free_list(page, zone, current_order);
2411 expand(zone, page, order, current_order, migratetype);
2412 set_pcppage_migratetype(page, migratetype);
2421 * This array describes the order lists are fallen back to when
2422 * the free lists for the desirable migrate type are depleted
2424 static int fallbacks[MIGRATE_TYPES][3] = {
2425 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2426 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2427 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2429 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2431 #ifdef CONFIG_MEMORY_ISOLATION
2432 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2437 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2440 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2443 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2444 unsigned int order) { return NULL; }
2448 * Move the free pages in a range to the freelist tail of the requested type.
2449 * Note that start_page and end_pages are not aligned on a pageblock
2450 * boundary. If alignment is required, use move_freepages_block()
2452 static int move_freepages(struct zone *zone,
2453 unsigned long start_pfn, unsigned long end_pfn,
2454 int migratetype, int *num_movable)
2459 int pages_moved = 0;
2461 for (pfn = start_pfn; pfn <= end_pfn;) {
2462 if (!pfn_valid_within(pfn)) {
2467 page = pfn_to_page(pfn);
2468 if (!PageBuddy(page)) {
2470 * We assume that pages that could be isolated for
2471 * migration are movable. But we don't actually try
2472 * isolating, as that would be expensive.
2475 (PageLRU(page) || __PageMovable(page)))
2481 /* Make sure we are not inadvertently changing nodes */
2482 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2483 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2485 order = buddy_order(page);
2486 move_to_free_list(page, zone, order, migratetype);
2488 pages_moved += 1 << order;
2494 int move_freepages_block(struct zone *zone, struct page *page,
2495 int migratetype, int *num_movable)
2497 unsigned long start_pfn, end_pfn, pfn;
2502 pfn = page_to_pfn(page);
2503 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2504 end_pfn = start_pfn + pageblock_nr_pages - 1;
2506 /* Do not cross zone boundaries */
2507 if (!zone_spans_pfn(zone, start_pfn))
2509 if (!zone_spans_pfn(zone, end_pfn))
2512 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2516 static void change_pageblock_range(struct page *pageblock_page,
2517 int start_order, int migratetype)
2519 int nr_pageblocks = 1 << (start_order - pageblock_order);
2521 while (nr_pageblocks--) {
2522 set_pageblock_migratetype(pageblock_page, migratetype);
2523 pageblock_page += pageblock_nr_pages;
2528 * When we are falling back to another migratetype during allocation, try to
2529 * steal extra free pages from the same pageblocks to satisfy further
2530 * allocations, instead of polluting multiple pageblocks.
2532 * If we are stealing a relatively large buddy page, it is likely there will
2533 * be more free pages in the pageblock, so try to steal them all. For
2534 * reclaimable and unmovable allocations, we steal regardless of page size,
2535 * as fragmentation caused by those allocations polluting movable pageblocks
2536 * is worse than movable allocations stealing from unmovable and reclaimable
2539 static bool can_steal_fallback(unsigned int order, int start_mt)
2542 * Leaving this order check is intended, although there is
2543 * relaxed order check in next check. The reason is that
2544 * we can actually steal whole pageblock if this condition met,
2545 * but, below check doesn't guarantee it and that is just heuristic
2546 * so could be changed anytime.
2548 if (order >= pageblock_order)
2551 if (order >= pageblock_order / 2 ||
2552 start_mt == MIGRATE_RECLAIMABLE ||
2553 start_mt == MIGRATE_UNMOVABLE ||
2554 page_group_by_mobility_disabled)
2560 static inline bool boost_watermark(struct zone *zone)
2562 unsigned long max_boost;
2564 if (!watermark_boost_factor)
2567 * Don't bother in zones that are unlikely to produce results.
2568 * On small machines, including kdump capture kernels running
2569 * in a small area, boosting the watermark can cause an out of
2570 * memory situation immediately.
2572 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2575 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2576 watermark_boost_factor, 10000);
2579 * high watermark may be uninitialised if fragmentation occurs
2580 * very early in boot so do not boost. We do not fall
2581 * through and boost by pageblock_nr_pages as failing
2582 * allocations that early means that reclaim is not going
2583 * to help and it may even be impossible to reclaim the
2584 * boosted watermark resulting in a hang.
2589 max_boost = max(pageblock_nr_pages, max_boost);
2591 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2598 * This function implements actual steal behaviour. If order is large enough,
2599 * we can steal whole pageblock. If not, we first move freepages in this
2600 * pageblock to our migratetype and determine how many already-allocated pages
2601 * are there in the pageblock with a compatible migratetype. If at least half
2602 * of pages are free or compatible, we can change migratetype of the pageblock
2603 * itself, so pages freed in the future will be put on the correct free list.
2605 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2606 unsigned int alloc_flags, int start_type, bool whole_block)
2608 unsigned int current_order = buddy_order(page);
2609 int free_pages, movable_pages, alike_pages;
2612 old_block_type = get_pageblock_migratetype(page);
2615 * This can happen due to races and we want to prevent broken
2616 * highatomic accounting.
2618 if (is_migrate_highatomic(old_block_type))
2621 /* Take ownership for orders >= pageblock_order */
2622 if (current_order >= pageblock_order) {
2623 change_pageblock_range(page, current_order, start_type);
2628 * Boost watermarks to increase reclaim pressure to reduce the
2629 * likelihood of future fallbacks. Wake kswapd now as the node
2630 * may be balanced overall and kswapd will not wake naturally.
2632 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2633 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2635 /* We are not allowed to try stealing from the whole block */
2639 free_pages = move_freepages_block(zone, page, start_type,
2642 * Determine how many pages are compatible with our allocation.
2643 * For movable allocation, it's the number of movable pages which
2644 * we just obtained. For other types it's a bit more tricky.
2646 if (start_type == MIGRATE_MOVABLE) {
2647 alike_pages = movable_pages;
2650 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2651 * to MOVABLE pageblock, consider all non-movable pages as
2652 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2653 * vice versa, be conservative since we can't distinguish the
2654 * exact migratetype of non-movable pages.
2656 if (old_block_type == MIGRATE_MOVABLE)
2657 alike_pages = pageblock_nr_pages
2658 - (free_pages + movable_pages);
2663 /* moving whole block can fail due to zone boundary conditions */
2668 * If a sufficient number of pages in the block are either free or of
2669 * comparable migratability as our allocation, claim the whole block.
2671 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2672 page_group_by_mobility_disabled)
2673 set_pageblock_migratetype(page, start_type);
2678 move_to_free_list(page, zone, current_order, start_type);
2682 * Check whether there is a suitable fallback freepage with requested order.
2683 * If only_stealable is true, this function returns fallback_mt only if
2684 * we can steal other freepages all together. This would help to reduce
2685 * fragmentation due to mixed migratetype pages in one pageblock.
2687 int find_suitable_fallback(struct free_area *area, unsigned int order,
2688 int migratetype, bool only_stealable, bool *can_steal)
2693 if (area->nr_free == 0)
2698 fallback_mt = fallbacks[migratetype][i];
2699 if (fallback_mt == MIGRATE_TYPES)
2702 if (free_area_empty(area, fallback_mt))
2705 if (can_steal_fallback(order, migratetype))
2708 if (!only_stealable)
2719 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2720 * there are no empty page blocks that contain a page with a suitable order
2722 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2723 unsigned int alloc_order)
2726 unsigned long max_managed, flags;
2729 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2730 * Check is race-prone but harmless.
2732 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2733 if (zone->nr_reserved_highatomic >= max_managed)
2736 spin_lock_irqsave(&zone->lock, flags);
2738 /* Recheck the nr_reserved_highatomic limit under the lock */
2739 if (zone->nr_reserved_highatomic >= max_managed)
2743 mt = get_pageblock_migratetype(page);
2744 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2745 && !is_migrate_cma(mt)) {
2746 zone->nr_reserved_highatomic += pageblock_nr_pages;
2747 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2748 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2752 spin_unlock_irqrestore(&zone->lock, flags);
2756 * Used when an allocation is about to fail under memory pressure. This
2757 * potentially hurts the reliability of high-order allocations when under
2758 * intense memory pressure but failed atomic allocations should be easier
2759 * to recover from than an OOM.
2761 * If @force is true, try to unreserve a pageblock even though highatomic
2762 * pageblock is exhausted.
2764 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2767 struct zonelist *zonelist = ac->zonelist;
2768 unsigned long flags;
2775 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2778 * Preserve at least one pageblock unless memory pressure
2781 if (!force && zone->nr_reserved_highatomic <=
2785 spin_lock_irqsave(&zone->lock, flags);
2786 for (order = 0; order < MAX_ORDER; order++) {
2787 struct free_area *area = &(zone->free_area[order]);
2789 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2794 * In page freeing path, migratetype change is racy so
2795 * we can counter several free pages in a pageblock
2796 * in this loop although we changed the pageblock type
2797 * from highatomic to ac->migratetype. So we should
2798 * adjust the count once.
2800 if (is_migrate_highatomic_page(page)) {
2802 * It should never happen but changes to
2803 * locking could inadvertently allow a per-cpu
2804 * drain to add pages to MIGRATE_HIGHATOMIC
2805 * while unreserving so be safe and watch for
2808 zone->nr_reserved_highatomic -= min(
2810 zone->nr_reserved_highatomic);
2814 * Convert to ac->migratetype and avoid the normal
2815 * pageblock stealing heuristics. Minimally, the caller
2816 * is doing the work and needs the pages. More
2817 * importantly, if the block was always converted to
2818 * MIGRATE_UNMOVABLE or another type then the number
2819 * of pageblocks that cannot be completely freed
2822 set_pageblock_migratetype(page, ac->migratetype);
2823 ret = move_freepages_block(zone, page, ac->migratetype,
2826 spin_unlock_irqrestore(&zone->lock, flags);
2830 spin_unlock_irqrestore(&zone->lock, flags);
2837 * Try finding a free buddy page on the fallback list and put it on the free
2838 * list of requested migratetype, possibly along with other pages from the same
2839 * block, depending on fragmentation avoidance heuristics. Returns true if
2840 * fallback was found so that __rmqueue_smallest() can grab it.
2842 * The use of signed ints for order and current_order is a deliberate
2843 * deviation from the rest of this file, to make the for loop
2844 * condition simpler.
2846 static __always_inline bool
2847 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2848 unsigned int alloc_flags)
2850 struct free_area *area;
2852 int min_order = order;
2858 * Do not steal pages from freelists belonging to other pageblocks
2859 * i.e. orders < pageblock_order. If there are no local zones free,
2860 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2862 if (alloc_flags & ALLOC_NOFRAGMENT)
2863 min_order = pageblock_order;
2866 * Find the largest available free page in the other list. This roughly
2867 * approximates finding the pageblock with the most free pages, which
2868 * would be too costly to do exactly.
2870 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2872 area = &(zone->free_area[current_order]);
2873 fallback_mt = find_suitable_fallback(area, current_order,
2874 start_migratetype, false, &can_steal);
2875 if (fallback_mt == -1)
2879 * We cannot steal all free pages from the pageblock and the
2880 * requested migratetype is movable. In that case it's better to
2881 * steal and split the smallest available page instead of the
2882 * largest available page, because even if the next movable
2883 * allocation falls back into a different pageblock than this
2884 * one, it won't cause permanent fragmentation.
2886 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2887 && current_order > order)
2896 for (current_order = order; current_order < MAX_ORDER;
2898 area = &(zone->free_area[current_order]);
2899 fallback_mt = find_suitable_fallback(area, current_order,
2900 start_migratetype, false, &can_steal);
2901 if (fallback_mt != -1)
2906 * This should not happen - we already found a suitable fallback
2907 * when looking for the largest page.
2909 VM_BUG_ON(current_order == MAX_ORDER);
2912 page = get_page_from_free_area(area, fallback_mt);
2914 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2917 trace_mm_page_alloc_extfrag(page, order, current_order,
2918 start_migratetype, fallback_mt);
2925 * Do the hard work of removing an element from the buddy allocator.
2926 * Call me with the zone->lock already held.
2928 static __always_inline struct page *
2929 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2930 unsigned int alloc_flags)
2934 if (IS_ENABLED(CONFIG_CMA)) {
2936 * Balance movable allocations between regular and CMA areas by
2937 * allocating from CMA when over half of the zone's free memory
2938 * is in the CMA area.
2940 if (alloc_flags & ALLOC_CMA &&
2941 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2942 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2943 page = __rmqueue_cma_fallback(zone, order);
2949 page = __rmqueue_smallest(zone, order, migratetype);
2950 if (unlikely(!page)) {
2951 if (alloc_flags & ALLOC_CMA)
2952 page = __rmqueue_cma_fallback(zone, order);
2954 if (!page && __rmqueue_fallback(zone, order, migratetype,
2960 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2965 * Obtain a specified number of elements from the buddy allocator, all under
2966 * a single hold of the lock, for efficiency. Add them to the supplied list.
2967 * Returns the number of new pages which were placed at *list.
2969 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2970 unsigned long count, struct list_head *list,
2971 int migratetype, unsigned int alloc_flags)
2973 int i, allocated = 0;
2976 * local_lock_irq held so equivalent to spin_lock_irqsave for
2977 * both PREEMPT_RT and non-PREEMPT_RT configurations.
2979 spin_lock(&zone->lock);
2980 for (i = 0; i < count; ++i) {
2981 struct page *page = __rmqueue(zone, order, migratetype,
2983 if (unlikely(page == NULL))
2986 if (unlikely(check_pcp_refill(page)))
2990 * Split buddy pages returned by expand() are received here in
2991 * physical page order. The page is added to the tail of
2992 * caller's list. From the callers perspective, the linked list
2993 * is ordered by page number under some conditions. This is
2994 * useful for IO devices that can forward direction from the
2995 * head, thus also in the physical page order. This is useful
2996 * for IO devices that can merge IO requests if the physical
2997 * pages are ordered properly.
2999 list_add_tail(&page->lru, list);
3001 if (is_migrate_cma(get_pcppage_migratetype(page)))
3002 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3007 * i pages were removed from the buddy list even if some leak due
3008 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3009 * on i. Do not confuse with 'allocated' which is the number of
3010 * pages added to the pcp list.
3012 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3013 spin_unlock(&zone->lock);
3019 * Called from the vmstat counter updater to drain pagesets of this
3020 * currently executing processor on remote nodes after they have
3023 * Note that this function must be called with the thread pinned to
3024 * a single processor.
3026 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3028 unsigned long flags;
3029 int to_drain, batch;
3031 local_lock_irqsave(&pagesets.lock, flags);
3032 batch = READ_ONCE(pcp->batch);
3033 to_drain = min(pcp->count, batch);
3035 free_pcppages_bulk(zone, to_drain, pcp);
3036 local_unlock_irqrestore(&pagesets.lock, flags);
3041 * Drain pcplists of the indicated processor and zone.
3043 * The processor must either be the current processor and the
3044 * thread pinned to the current processor or a processor that
3047 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3049 unsigned long flags;
3050 struct per_cpu_pages *pcp;
3052 local_lock_irqsave(&pagesets.lock, flags);
3054 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3056 free_pcppages_bulk(zone, pcp->count, pcp);
3058 local_unlock_irqrestore(&pagesets.lock, flags);
3062 * Drain pcplists of all zones on the indicated processor.
3064 * The processor must either be the current processor and the
3065 * thread pinned to the current processor or a processor that
3068 static void drain_pages(unsigned int cpu)
3072 for_each_populated_zone(zone) {
3073 drain_pages_zone(cpu, zone);
3078 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3080 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3081 * the single zone's pages.
3083 void drain_local_pages(struct zone *zone)
3085 int cpu = smp_processor_id();
3088 drain_pages_zone(cpu, zone);
3093 static void drain_local_pages_wq(struct work_struct *work)
3095 struct pcpu_drain *drain;
3097 drain = container_of(work, struct pcpu_drain, work);
3100 * drain_all_pages doesn't use proper cpu hotplug protection so
3101 * we can race with cpu offline when the WQ can move this from
3102 * a cpu pinned worker to an unbound one. We can operate on a different
3103 * cpu which is alright but we also have to make sure to not move to
3107 drain_local_pages(drain->zone);
3112 * The implementation of drain_all_pages(), exposing an extra parameter to
3113 * drain on all cpus.
3115 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3116 * not empty. The check for non-emptiness can however race with a free to
3117 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3118 * that need the guarantee that every CPU has drained can disable the
3119 * optimizing racy check.
3121 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3126 * Allocate in the BSS so we wont require allocation in
3127 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3129 static cpumask_t cpus_with_pcps;
3132 * Make sure nobody triggers this path before mm_percpu_wq is fully
3135 if (WARN_ON_ONCE(!mm_percpu_wq))
3139 * Do not drain if one is already in progress unless it's specific to
3140 * a zone. Such callers are primarily CMA and memory hotplug and need
3141 * the drain to be complete when the call returns.
3143 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3146 mutex_lock(&pcpu_drain_mutex);
3150 * We don't care about racing with CPU hotplug event
3151 * as offline notification will cause the notified
3152 * cpu to drain that CPU pcps and on_each_cpu_mask
3153 * disables preemption as part of its processing
3155 for_each_online_cpu(cpu) {
3156 struct per_cpu_pages *pcp;
3158 bool has_pcps = false;
3160 if (force_all_cpus) {
3162 * The pcp.count check is racy, some callers need a
3163 * guarantee that no cpu is missed.
3167 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3171 for_each_populated_zone(z) {
3172 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3181 cpumask_set_cpu(cpu, &cpus_with_pcps);
3183 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3186 for_each_cpu(cpu, &cpus_with_pcps) {
3187 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3190 INIT_WORK(&drain->work, drain_local_pages_wq);
3191 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3193 for_each_cpu(cpu, &cpus_with_pcps)
3194 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3196 mutex_unlock(&pcpu_drain_mutex);
3200 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3202 * When zone parameter is non-NULL, spill just the single zone's pages.
3204 * Note that this can be extremely slow as the draining happens in a workqueue.
3206 void drain_all_pages(struct zone *zone)
3208 __drain_all_pages(zone, false);
3211 #ifdef CONFIG_HIBERNATION
3214 * Touch the watchdog for every WD_PAGE_COUNT pages.
3216 #define WD_PAGE_COUNT (128*1024)
3218 void mark_free_pages(struct zone *zone)
3220 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3221 unsigned long flags;
3222 unsigned int order, t;
3225 if (zone_is_empty(zone))
3228 spin_lock_irqsave(&zone->lock, flags);
3230 max_zone_pfn = zone_end_pfn(zone);
3231 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3232 if (pfn_valid(pfn)) {
3233 page = pfn_to_page(pfn);
3235 if (!--page_count) {
3236 touch_nmi_watchdog();
3237 page_count = WD_PAGE_COUNT;
3240 if (page_zone(page) != zone)
3243 if (!swsusp_page_is_forbidden(page))
3244 swsusp_unset_page_free(page);
3247 for_each_migratetype_order(order, t) {
3248 list_for_each_entry(page,
3249 &zone->free_area[order].free_list[t], lru) {
3252 pfn = page_to_pfn(page);
3253 for (i = 0; i < (1UL << order); i++) {
3254 if (!--page_count) {
3255 touch_nmi_watchdog();
3256 page_count = WD_PAGE_COUNT;
3258 swsusp_set_page_free(pfn_to_page(pfn + i));
3262 spin_unlock_irqrestore(&zone->lock, flags);
3264 #endif /* CONFIG_PM */
3266 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3270 if (!free_pcp_prepare(page))
3273 migratetype = get_pfnblock_migratetype(page, pfn);
3274 set_pcppage_migratetype(page, migratetype);
3278 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch)
3280 int min_nr_free, max_nr_free;
3282 /* Check for PCP disabled or boot pageset */
3283 if (unlikely(high < batch))
3286 /* Leave at least pcp->batch pages on the list */
3287 min_nr_free = batch;
3288 max_nr_free = high - batch;
3291 * Double the number of pages freed each time there is subsequent
3292 * freeing of pages without any allocation.
3294 batch <<= pcp->free_factor;
3295 if (batch < max_nr_free)
3297 batch = clamp(batch, min_nr_free, max_nr_free);
3302 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone)
3304 int high = READ_ONCE(pcp->high);
3306 if (unlikely(!high))
3309 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3313 * If reclaim is active, limit the number of pages that can be
3314 * stored on pcp lists
3316 return min(READ_ONCE(pcp->batch) << 2, high);
3319 static void free_unref_page_commit(struct page *page, unsigned long pfn,
3322 struct zone *zone = page_zone(page);
3323 struct per_cpu_pages *pcp;
3326 __count_vm_event(PGFREE);
3327 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3328 list_add(&page->lru, &pcp->lists[migratetype]);
3330 high = nr_pcp_high(pcp, zone);
3331 if (pcp->count >= high) {
3332 int batch = READ_ONCE(pcp->batch);
3334 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch), pcp);
3339 * Free a 0-order page
3341 void free_unref_page(struct page *page)
3343 unsigned long flags;
3344 unsigned long pfn = page_to_pfn(page);
3347 if (!free_unref_page_prepare(page, pfn))
3351 * We only track unmovable, reclaimable and movable on pcp lists.
3352 * Place ISOLATE pages on the isolated list because they are being
3353 * offlined but treat HIGHATOMIC as movable pages so we can get those
3354 * areas back if necessary. Otherwise, we may have to free
3355 * excessively into the page allocator
3357 migratetype = get_pcppage_migratetype(page);
3358 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3359 if (unlikely(is_migrate_isolate(migratetype))) {
3360 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3363 migratetype = MIGRATE_MOVABLE;
3366 local_lock_irqsave(&pagesets.lock, flags);
3367 free_unref_page_commit(page, pfn, migratetype);
3368 local_unlock_irqrestore(&pagesets.lock, flags);
3372 * Free a list of 0-order pages
3374 void free_unref_page_list(struct list_head *list)
3376 struct page *page, *next;
3377 unsigned long flags, pfn;
3378 int batch_count = 0;
3381 /* Prepare pages for freeing */
3382 list_for_each_entry_safe(page, next, list, lru) {
3383 pfn = page_to_pfn(page);
3384 if (!free_unref_page_prepare(page, pfn))
3385 list_del(&page->lru);
3388 * Free isolated pages directly to the allocator, see
3389 * comment in free_unref_page.
3391 migratetype = get_pcppage_migratetype(page);
3392 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3393 if (unlikely(is_migrate_isolate(migratetype))) {
3394 list_del(&page->lru);
3395 free_one_page(page_zone(page), page, pfn, 0,
3396 migratetype, FPI_NONE);
3401 * Non-isolated types over MIGRATE_PCPTYPES get added
3402 * to the MIGRATE_MOVABLE pcp list.
3404 set_pcppage_migratetype(page, MIGRATE_MOVABLE);
3407 set_page_private(page, pfn);
3410 local_lock_irqsave(&pagesets.lock, flags);
3411 list_for_each_entry_safe(page, next, list, lru) {
3412 pfn = page_private(page);
3413 set_page_private(page, 0);
3414 migratetype = get_pcppage_migratetype(page);
3415 trace_mm_page_free_batched(page);
3416 free_unref_page_commit(page, pfn, migratetype);
3419 * Guard against excessive IRQ disabled times when we get
3420 * a large list of pages to free.
3422 if (++batch_count == SWAP_CLUSTER_MAX) {
3423 local_unlock_irqrestore(&pagesets.lock, flags);
3425 local_lock_irqsave(&pagesets.lock, flags);
3428 local_unlock_irqrestore(&pagesets.lock, flags);
3432 * split_page takes a non-compound higher-order page, and splits it into
3433 * n (1<<order) sub-pages: page[0..n]
3434 * Each sub-page must be freed individually.
3436 * Note: this is probably too low level an operation for use in drivers.
3437 * Please consult with lkml before using this in your driver.
3439 void split_page(struct page *page, unsigned int order)
3443 VM_BUG_ON_PAGE(PageCompound(page), page);
3444 VM_BUG_ON_PAGE(!page_count(page), page);
3446 for (i = 1; i < (1 << order); i++)
3447 set_page_refcounted(page + i);
3448 split_page_owner(page, 1 << order);
3449 split_page_memcg(page, 1 << order);
3451 EXPORT_SYMBOL_GPL(split_page);
3453 int __isolate_free_page(struct page *page, unsigned int order)
3455 unsigned long watermark;
3459 BUG_ON(!PageBuddy(page));
3461 zone = page_zone(page);
3462 mt = get_pageblock_migratetype(page);
3464 if (!is_migrate_isolate(mt)) {
3466 * Obey watermarks as if the page was being allocated. We can
3467 * emulate a high-order watermark check with a raised order-0
3468 * watermark, because we already know our high-order page
3471 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3472 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3475 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3478 /* Remove page from free list */
3480 del_page_from_free_list(page, zone, order);
3483 * Set the pageblock if the isolated page is at least half of a
3486 if (order >= pageblock_order - 1) {
3487 struct page *endpage = page + (1 << order) - 1;
3488 for (; page < endpage; page += pageblock_nr_pages) {
3489 int mt = get_pageblock_migratetype(page);
3490 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3491 && !is_migrate_highatomic(mt))
3492 set_pageblock_migratetype(page,
3498 return 1UL << order;
3502 * __putback_isolated_page - Return a now-isolated page back where we got it
3503 * @page: Page that was isolated
3504 * @order: Order of the isolated page
3505 * @mt: The page's pageblock's migratetype
3507 * This function is meant to return a page pulled from the free lists via
3508 * __isolate_free_page back to the free lists they were pulled from.
3510 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3512 struct zone *zone = page_zone(page);
3514 /* zone lock should be held when this function is called */
3515 lockdep_assert_held(&zone->lock);
3517 /* Return isolated page to tail of freelist. */
3518 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3519 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3523 * Update NUMA hit/miss statistics
3525 * Must be called with interrupts disabled.
3527 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3531 enum numa_stat_item local_stat = NUMA_LOCAL;
3533 /* skip numa counters update if numa stats is disabled */
3534 if (!static_branch_likely(&vm_numa_stat_key))
3537 if (zone_to_nid(z) != numa_node_id())
3538 local_stat = NUMA_OTHER;
3540 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3541 __count_numa_events(z, NUMA_HIT, nr_account);
3543 __count_numa_events(z, NUMA_MISS, nr_account);
3544 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3546 __count_numa_events(z, local_stat, nr_account);
3550 /* Remove page from the per-cpu list, caller must protect the list */
3552 struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3553 unsigned int alloc_flags,
3554 struct per_cpu_pages *pcp,
3555 struct list_head *list)
3560 if (list_empty(list)) {
3561 pcp->count += rmqueue_bulk(zone, 0,
3562 READ_ONCE(pcp->batch), list,
3563 migratetype, alloc_flags);
3564 if (unlikely(list_empty(list)))
3568 page = list_first_entry(list, struct page, lru);
3569 list_del(&page->lru);
3571 } while (check_new_pcp(page));
3576 /* Lock and remove page from the per-cpu list */
3577 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3578 struct zone *zone, gfp_t gfp_flags,
3579 int migratetype, unsigned int alloc_flags)
3581 struct per_cpu_pages *pcp;
3582 struct list_head *list;
3584 unsigned long flags;
3586 local_lock_irqsave(&pagesets.lock, flags);
3589 * On allocation, reduce the number of pages that are batch freed.
3590 * See nr_pcp_free() where free_factor is increased for subsequent
3593 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3594 pcp->free_factor >>= 1;
3595 list = &pcp->lists[migratetype];
3596 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3597 local_unlock_irqrestore(&pagesets.lock, flags);
3599 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3600 zone_statistics(preferred_zone, zone, 1);
3606 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3609 struct page *rmqueue(struct zone *preferred_zone,
3610 struct zone *zone, unsigned int order,
3611 gfp_t gfp_flags, unsigned int alloc_flags,
3614 unsigned long flags;
3617 if (likely(order == 0)) {
3619 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3620 * we need to skip it when CMA area isn't allowed.
3622 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3623 migratetype != MIGRATE_MOVABLE) {
3624 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3625 migratetype, alloc_flags);
3631 * We most definitely don't want callers attempting to
3632 * allocate greater than order-1 page units with __GFP_NOFAIL.
3634 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3635 spin_lock_irqsave(&zone->lock, flags);
3640 * order-0 request can reach here when the pcplist is skipped
3641 * due to non-CMA allocation context. HIGHATOMIC area is
3642 * reserved for high-order atomic allocation, so order-0
3643 * request should skip it.
3645 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3646 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3648 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3651 page = __rmqueue(zone, order, migratetype, alloc_flags);
3652 } while (page && check_new_pages(page, order));
3656 __mod_zone_freepage_state(zone, -(1 << order),
3657 get_pcppage_migratetype(page));
3658 spin_unlock_irqrestore(&zone->lock, flags);
3660 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3661 zone_statistics(preferred_zone, zone, 1);
3664 /* Separate test+clear to avoid unnecessary atomics */
3665 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3666 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3667 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3670 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3674 spin_unlock_irqrestore(&zone->lock, flags);
3678 #ifdef CONFIG_FAIL_PAGE_ALLOC
3681 struct fault_attr attr;
3683 bool ignore_gfp_highmem;
3684 bool ignore_gfp_reclaim;
3686 } fail_page_alloc = {
3687 .attr = FAULT_ATTR_INITIALIZER,
3688 .ignore_gfp_reclaim = true,
3689 .ignore_gfp_highmem = true,
3693 static int __init setup_fail_page_alloc(char *str)
3695 return setup_fault_attr(&fail_page_alloc.attr, str);
3697 __setup("fail_page_alloc=", setup_fail_page_alloc);
3699 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3701 if (order < fail_page_alloc.min_order)
3703 if (gfp_mask & __GFP_NOFAIL)
3705 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3707 if (fail_page_alloc.ignore_gfp_reclaim &&
3708 (gfp_mask & __GFP_DIRECT_RECLAIM))
3711 return should_fail(&fail_page_alloc.attr, 1 << order);
3714 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3716 static int __init fail_page_alloc_debugfs(void)
3718 umode_t mode = S_IFREG | 0600;
3721 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3722 &fail_page_alloc.attr);
3724 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3725 &fail_page_alloc.ignore_gfp_reclaim);
3726 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3727 &fail_page_alloc.ignore_gfp_highmem);
3728 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3733 late_initcall(fail_page_alloc_debugfs);
3735 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3737 #else /* CONFIG_FAIL_PAGE_ALLOC */
3739 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3744 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3746 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3748 return __should_fail_alloc_page(gfp_mask, order);
3750 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3752 static inline long __zone_watermark_unusable_free(struct zone *z,
3753 unsigned int order, unsigned int alloc_flags)
3755 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3756 long unusable_free = (1 << order) - 1;
3759 * If the caller does not have rights to ALLOC_HARDER then subtract
3760 * the high-atomic reserves. This will over-estimate the size of the
3761 * atomic reserve but it avoids a search.
3763 if (likely(!alloc_harder))
3764 unusable_free += z->nr_reserved_highatomic;
3767 /* If allocation can't use CMA areas don't use free CMA pages */
3768 if (!(alloc_flags & ALLOC_CMA))
3769 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3772 return unusable_free;
3776 * Return true if free base pages are above 'mark'. For high-order checks it
3777 * will return true of the order-0 watermark is reached and there is at least
3778 * one free page of a suitable size. Checking now avoids taking the zone lock
3779 * to check in the allocation paths if no pages are free.
3781 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3782 int highest_zoneidx, unsigned int alloc_flags,
3787 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3789 /* free_pages may go negative - that's OK */
3790 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3792 if (alloc_flags & ALLOC_HIGH)
3795 if (unlikely(alloc_harder)) {
3797 * OOM victims can try even harder than normal ALLOC_HARDER
3798 * users on the grounds that it's definitely going to be in
3799 * the exit path shortly and free memory. Any allocation it
3800 * makes during the free path will be small and short-lived.
3802 if (alloc_flags & ALLOC_OOM)
3809 * Check watermarks for an order-0 allocation request. If these
3810 * are not met, then a high-order request also cannot go ahead
3811 * even if a suitable page happened to be free.
3813 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3816 /* If this is an order-0 request then the watermark is fine */
3820 /* For a high-order request, check at least one suitable page is free */
3821 for (o = order; o < MAX_ORDER; o++) {
3822 struct free_area *area = &z->free_area[o];
3828 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3829 if (!free_area_empty(area, mt))
3834 if ((alloc_flags & ALLOC_CMA) &&
3835 !free_area_empty(area, MIGRATE_CMA)) {
3839 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3845 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3846 int highest_zoneidx, unsigned int alloc_flags)
3848 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3849 zone_page_state(z, NR_FREE_PAGES));
3852 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3853 unsigned long mark, int highest_zoneidx,
3854 unsigned int alloc_flags, gfp_t gfp_mask)
3858 free_pages = zone_page_state(z, NR_FREE_PAGES);
3861 * Fast check for order-0 only. If this fails then the reserves
3862 * need to be calculated.
3867 fast_free = free_pages;
3868 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3869 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3873 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3877 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3878 * when checking the min watermark. The min watermark is the
3879 * point where boosting is ignored so that kswapd is woken up
3880 * when below the low watermark.
3882 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3883 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3884 mark = z->_watermark[WMARK_MIN];
3885 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3886 alloc_flags, free_pages);
3892 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3893 unsigned long mark, int highest_zoneidx)
3895 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3897 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3898 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3900 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3905 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3907 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3908 node_reclaim_distance;
3910 #else /* CONFIG_NUMA */
3911 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3915 #endif /* CONFIG_NUMA */
3918 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3919 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3920 * premature use of a lower zone may cause lowmem pressure problems that
3921 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3922 * probably too small. It only makes sense to spread allocations to avoid
3923 * fragmentation between the Normal and DMA32 zones.
3925 static inline unsigned int
3926 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3928 unsigned int alloc_flags;
3931 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3934 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3936 #ifdef CONFIG_ZONE_DMA32
3940 if (zone_idx(zone) != ZONE_NORMAL)
3944 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3945 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3946 * on UMA that if Normal is populated then so is DMA32.
3948 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3949 if (nr_online_nodes > 1 && !populated_zone(--zone))
3952 alloc_flags |= ALLOC_NOFRAGMENT;
3953 #endif /* CONFIG_ZONE_DMA32 */
3957 /* Must be called after current_gfp_context() which can change gfp_mask */
3958 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3959 unsigned int alloc_flags)
3962 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3963 alloc_flags |= ALLOC_CMA;
3969 * get_page_from_freelist goes through the zonelist trying to allocate
3972 static struct page *
3973 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3974 const struct alloc_context *ac)
3978 struct pglist_data *last_pgdat_dirty_limit = NULL;
3983 * Scan zonelist, looking for a zone with enough free.
3984 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3986 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3987 z = ac->preferred_zoneref;
3988 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3993 if (cpusets_enabled() &&
3994 (alloc_flags & ALLOC_CPUSET) &&
3995 !__cpuset_zone_allowed(zone, gfp_mask))
3998 * When allocating a page cache page for writing, we
3999 * want to get it from a node that is within its dirty
4000 * limit, such that no single node holds more than its
4001 * proportional share of globally allowed dirty pages.
4002 * The dirty limits take into account the node's
4003 * lowmem reserves and high watermark so that kswapd
4004 * should be able to balance it without having to
4005 * write pages from its LRU list.
4007 * XXX: For now, allow allocations to potentially
4008 * exceed the per-node dirty limit in the slowpath
4009 * (spread_dirty_pages unset) before going into reclaim,
4010 * which is important when on a NUMA setup the allowed
4011 * nodes are together not big enough to reach the
4012 * global limit. The proper fix for these situations
4013 * will require awareness of nodes in the
4014 * dirty-throttling and the flusher threads.
4016 if (ac->spread_dirty_pages) {
4017 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4020 if (!node_dirty_ok(zone->zone_pgdat)) {
4021 last_pgdat_dirty_limit = zone->zone_pgdat;
4026 if (no_fallback && nr_online_nodes > 1 &&
4027 zone != ac->preferred_zoneref->zone) {
4031 * If moving to a remote node, retry but allow
4032 * fragmenting fallbacks. Locality is more important
4033 * than fragmentation avoidance.
4035 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4036 if (zone_to_nid(zone) != local_nid) {
4037 alloc_flags &= ~ALLOC_NOFRAGMENT;
4042 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4043 if (!zone_watermark_fast(zone, order, mark,
4044 ac->highest_zoneidx, alloc_flags,
4048 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4050 * Watermark failed for this zone, but see if we can
4051 * grow this zone if it contains deferred pages.
4053 if (static_branch_unlikely(&deferred_pages)) {
4054 if (_deferred_grow_zone(zone, order))
4058 /* Checked here to keep the fast path fast */
4059 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4060 if (alloc_flags & ALLOC_NO_WATERMARKS)
4063 if (!node_reclaim_enabled() ||
4064 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4067 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4069 case NODE_RECLAIM_NOSCAN:
4072 case NODE_RECLAIM_FULL:
4073 /* scanned but unreclaimable */
4076 /* did we reclaim enough */
4077 if (zone_watermark_ok(zone, order, mark,
4078 ac->highest_zoneidx, alloc_flags))
4086 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4087 gfp_mask, alloc_flags, ac->migratetype);
4089 prep_new_page(page, order, gfp_mask, alloc_flags);
4092 * If this is a high-order atomic allocation then check
4093 * if the pageblock should be reserved for the future
4095 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4096 reserve_highatomic_pageblock(page, zone, order);
4100 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4101 /* Try again if zone has deferred pages */
4102 if (static_branch_unlikely(&deferred_pages)) {
4103 if (_deferred_grow_zone(zone, order))
4111 * It's possible on a UMA machine to get through all zones that are
4112 * fragmented. If avoiding fragmentation, reset and try again.
4115 alloc_flags &= ~ALLOC_NOFRAGMENT;
4122 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4124 unsigned int filter = SHOW_MEM_FILTER_NODES;
4127 * This documents exceptions given to allocations in certain
4128 * contexts that are allowed to allocate outside current's set
4131 if (!(gfp_mask & __GFP_NOMEMALLOC))
4132 if (tsk_is_oom_victim(current) ||
4133 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4134 filter &= ~SHOW_MEM_FILTER_NODES;
4135 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4136 filter &= ~SHOW_MEM_FILTER_NODES;
4138 show_mem(filter, nodemask);
4141 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4143 struct va_format vaf;
4145 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4147 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4150 va_start(args, fmt);
4153 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4154 current->comm, &vaf, gfp_mask, &gfp_mask,
4155 nodemask_pr_args(nodemask));
4158 cpuset_print_current_mems_allowed();
4161 warn_alloc_show_mem(gfp_mask, nodemask);
4164 static inline struct page *
4165 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4166 unsigned int alloc_flags,
4167 const struct alloc_context *ac)
4171 page = get_page_from_freelist(gfp_mask, order,
4172 alloc_flags|ALLOC_CPUSET, ac);
4174 * fallback to ignore cpuset restriction if our nodes
4178 page = get_page_from_freelist(gfp_mask, order,
4184 static inline struct page *
4185 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4186 const struct alloc_context *ac, unsigned long *did_some_progress)
4188 struct oom_control oc = {
4189 .zonelist = ac->zonelist,
4190 .nodemask = ac->nodemask,
4192 .gfp_mask = gfp_mask,
4197 *did_some_progress = 0;
4200 * Acquire the oom lock. If that fails, somebody else is
4201 * making progress for us.
4203 if (!mutex_trylock(&oom_lock)) {
4204 *did_some_progress = 1;
4205 schedule_timeout_uninterruptible(1);
4210 * Go through the zonelist yet one more time, keep very high watermark
4211 * here, this is only to catch a parallel oom killing, we must fail if
4212 * we're still under heavy pressure. But make sure that this reclaim
4213 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4214 * allocation which will never fail due to oom_lock already held.
4216 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4217 ~__GFP_DIRECT_RECLAIM, order,
4218 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4222 /* Coredumps can quickly deplete all memory reserves */
4223 if (current->flags & PF_DUMPCORE)
4225 /* The OOM killer will not help higher order allocs */
4226 if (order > PAGE_ALLOC_COSTLY_ORDER)
4229 * We have already exhausted all our reclaim opportunities without any
4230 * success so it is time to admit defeat. We will skip the OOM killer
4231 * because it is very likely that the caller has a more reasonable
4232 * fallback than shooting a random task.
4234 * The OOM killer may not free memory on a specific node.
4236 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4238 /* The OOM killer does not needlessly kill tasks for lowmem */
4239 if (ac->highest_zoneidx < ZONE_NORMAL)
4241 if (pm_suspended_storage())
4244 * XXX: GFP_NOFS allocations should rather fail than rely on
4245 * other request to make a forward progress.
4246 * We are in an unfortunate situation where out_of_memory cannot
4247 * do much for this context but let's try it to at least get
4248 * access to memory reserved if the current task is killed (see
4249 * out_of_memory). Once filesystems are ready to handle allocation
4250 * failures more gracefully we should just bail out here.
4253 /* Exhausted what can be done so it's blame time */
4254 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4255 *did_some_progress = 1;
4258 * Help non-failing allocations by giving them access to memory
4261 if (gfp_mask & __GFP_NOFAIL)
4262 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4263 ALLOC_NO_WATERMARKS, ac);
4266 mutex_unlock(&oom_lock);
4271 * Maximum number of compaction retries with a progress before OOM
4272 * killer is consider as the only way to move forward.
4274 #define MAX_COMPACT_RETRIES 16
4276 #ifdef CONFIG_COMPACTION
4277 /* Try memory compaction for high-order allocations before reclaim */
4278 static struct page *
4279 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4280 unsigned int alloc_flags, const struct alloc_context *ac,
4281 enum compact_priority prio, enum compact_result *compact_result)
4283 struct page *page = NULL;
4284 unsigned long pflags;
4285 unsigned int noreclaim_flag;
4290 psi_memstall_enter(&pflags);
4291 noreclaim_flag = memalloc_noreclaim_save();
4293 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4296 memalloc_noreclaim_restore(noreclaim_flag);
4297 psi_memstall_leave(&pflags);
4299 if (*compact_result == COMPACT_SKIPPED)
4302 * At least in one zone compaction wasn't deferred or skipped, so let's
4303 * count a compaction stall
4305 count_vm_event(COMPACTSTALL);
4307 /* Prep a captured page if available */
4309 prep_new_page(page, order, gfp_mask, alloc_flags);
4311 /* Try get a page from the freelist if available */
4313 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4316 struct zone *zone = page_zone(page);
4318 zone->compact_blockskip_flush = false;
4319 compaction_defer_reset(zone, order, true);
4320 count_vm_event(COMPACTSUCCESS);
4325 * It's bad if compaction run occurs and fails. The most likely reason
4326 * is that pages exist, but not enough to satisfy watermarks.
4328 count_vm_event(COMPACTFAIL);
4336 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4337 enum compact_result compact_result,
4338 enum compact_priority *compact_priority,
4339 int *compaction_retries)
4341 int max_retries = MAX_COMPACT_RETRIES;
4344 int retries = *compaction_retries;
4345 enum compact_priority priority = *compact_priority;
4350 if (fatal_signal_pending(current))
4353 if (compaction_made_progress(compact_result))
4354 (*compaction_retries)++;
4357 * compaction considers all the zone as desperately out of memory
4358 * so it doesn't really make much sense to retry except when the
4359 * failure could be caused by insufficient priority
4361 if (compaction_failed(compact_result))
4362 goto check_priority;
4365 * compaction was skipped because there are not enough order-0 pages
4366 * to work with, so we retry only if it looks like reclaim can help.
4368 if (compaction_needs_reclaim(compact_result)) {
4369 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4374 * make sure the compaction wasn't deferred or didn't bail out early
4375 * due to locks contention before we declare that we should give up.
4376 * But the next retry should use a higher priority if allowed, so
4377 * we don't just keep bailing out endlessly.
4379 if (compaction_withdrawn(compact_result)) {
4380 goto check_priority;
4384 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4385 * costly ones because they are de facto nofail and invoke OOM
4386 * killer to move on while costly can fail and users are ready
4387 * to cope with that. 1/4 retries is rather arbitrary but we
4388 * would need much more detailed feedback from compaction to
4389 * make a better decision.
4391 if (order > PAGE_ALLOC_COSTLY_ORDER)
4393 if (*compaction_retries <= max_retries) {
4399 * Make sure there are attempts at the highest priority if we exhausted
4400 * all retries or failed at the lower priorities.
4403 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4404 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4406 if (*compact_priority > min_priority) {
4407 (*compact_priority)--;
4408 *compaction_retries = 0;
4412 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4416 static inline struct page *
4417 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4418 unsigned int alloc_flags, const struct alloc_context *ac,
4419 enum compact_priority prio, enum compact_result *compact_result)
4421 *compact_result = COMPACT_SKIPPED;
4426 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4427 enum compact_result compact_result,
4428 enum compact_priority *compact_priority,
4429 int *compaction_retries)
4434 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4438 * There are setups with compaction disabled which would prefer to loop
4439 * inside the allocator rather than hit the oom killer prematurely.
4440 * Let's give them a good hope and keep retrying while the order-0
4441 * watermarks are OK.
4443 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4444 ac->highest_zoneidx, ac->nodemask) {
4445 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4446 ac->highest_zoneidx, alloc_flags))
4451 #endif /* CONFIG_COMPACTION */
4453 #ifdef CONFIG_LOCKDEP
4454 static struct lockdep_map __fs_reclaim_map =
4455 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4457 static bool __need_reclaim(gfp_t gfp_mask)
4459 /* no reclaim without waiting on it */
4460 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4463 /* this guy won't enter reclaim */
4464 if (current->flags & PF_MEMALLOC)
4467 if (gfp_mask & __GFP_NOLOCKDEP)
4473 void __fs_reclaim_acquire(void)
4475 lock_map_acquire(&__fs_reclaim_map);
4478 void __fs_reclaim_release(void)
4480 lock_map_release(&__fs_reclaim_map);
4483 void fs_reclaim_acquire(gfp_t gfp_mask)
4485 gfp_mask = current_gfp_context(gfp_mask);
4487 if (__need_reclaim(gfp_mask)) {
4488 if (gfp_mask & __GFP_FS)
4489 __fs_reclaim_acquire();
4491 #ifdef CONFIG_MMU_NOTIFIER
4492 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4493 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4498 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4500 void fs_reclaim_release(gfp_t gfp_mask)
4502 gfp_mask = current_gfp_context(gfp_mask);
4504 if (__need_reclaim(gfp_mask)) {
4505 if (gfp_mask & __GFP_FS)
4506 __fs_reclaim_release();
4509 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4512 /* Perform direct synchronous page reclaim */
4513 static unsigned long
4514 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4515 const struct alloc_context *ac)
4517 unsigned int noreclaim_flag;
4518 unsigned long pflags, progress;
4522 /* We now go into synchronous reclaim */
4523 cpuset_memory_pressure_bump();
4524 psi_memstall_enter(&pflags);
4525 fs_reclaim_acquire(gfp_mask);
4526 noreclaim_flag = memalloc_noreclaim_save();
4528 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4531 memalloc_noreclaim_restore(noreclaim_flag);
4532 fs_reclaim_release(gfp_mask);
4533 psi_memstall_leave(&pflags);
4540 /* The really slow allocator path where we enter direct reclaim */
4541 static inline struct page *
4542 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4543 unsigned int alloc_flags, const struct alloc_context *ac,
4544 unsigned long *did_some_progress)
4546 struct page *page = NULL;
4547 bool drained = false;
4549 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4550 if (unlikely(!(*did_some_progress)))
4554 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4557 * If an allocation failed after direct reclaim, it could be because
4558 * pages are pinned on the per-cpu lists or in high alloc reserves.
4559 * Shrink them and try again
4561 if (!page && !drained) {
4562 unreserve_highatomic_pageblock(ac, false);
4563 drain_all_pages(NULL);
4571 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4572 const struct alloc_context *ac)
4576 pg_data_t *last_pgdat = NULL;
4577 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4579 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4581 if (last_pgdat != zone->zone_pgdat)
4582 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4583 last_pgdat = zone->zone_pgdat;
4587 static inline unsigned int
4588 gfp_to_alloc_flags(gfp_t gfp_mask)
4590 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4593 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4594 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4595 * to save two branches.
4597 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4598 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4601 * The caller may dip into page reserves a bit more if the caller
4602 * cannot run direct reclaim, or if the caller has realtime scheduling
4603 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4604 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4606 alloc_flags |= (__force int)
4607 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4609 if (gfp_mask & __GFP_ATOMIC) {
4611 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4612 * if it can't schedule.
4614 if (!(gfp_mask & __GFP_NOMEMALLOC))
4615 alloc_flags |= ALLOC_HARDER;
4617 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4618 * comment for __cpuset_node_allowed().
4620 alloc_flags &= ~ALLOC_CPUSET;
4621 } else if (unlikely(rt_task(current)) && !in_interrupt())
4622 alloc_flags |= ALLOC_HARDER;
4624 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4629 static bool oom_reserves_allowed(struct task_struct *tsk)
4631 if (!tsk_is_oom_victim(tsk))
4635 * !MMU doesn't have oom reaper so give access to memory reserves
4636 * only to the thread with TIF_MEMDIE set
4638 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4645 * Distinguish requests which really need access to full memory
4646 * reserves from oom victims which can live with a portion of it
4648 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4650 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4652 if (gfp_mask & __GFP_MEMALLOC)
4653 return ALLOC_NO_WATERMARKS;
4654 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4655 return ALLOC_NO_WATERMARKS;
4656 if (!in_interrupt()) {
4657 if (current->flags & PF_MEMALLOC)
4658 return ALLOC_NO_WATERMARKS;
4659 else if (oom_reserves_allowed(current))
4666 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4668 return !!__gfp_pfmemalloc_flags(gfp_mask);
4672 * Checks whether it makes sense to retry the reclaim to make a forward progress
4673 * for the given allocation request.
4675 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4676 * without success, or when we couldn't even meet the watermark if we
4677 * reclaimed all remaining pages on the LRU lists.
4679 * Returns true if a retry is viable or false to enter the oom path.
4682 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4683 struct alloc_context *ac, int alloc_flags,
4684 bool did_some_progress, int *no_progress_loops)
4691 * Costly allocations might have made a progress but this doesn't mean
4692 * their order will become available due to high fragmentation so
4693 * always increment the no progress counter for them
4695 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4696 *no_progress_loops = 0;
4698 (*no_progress_loops)++;
4701 * Make sure we converge to OOM if we cannot make any progress
4702 * several times in the row.
4704 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4705 /* Before OOM, exhaust highatomic_reserve */
4706 return unreserve_highatomic_pageblock(ac, true);
4710 * Keep reclaiming pages while there is a chance this will lead
4711 * somewhere. If none of the target zones can satisfy our allocation
4712 * request even if all reclaimable pages are considered then we are
4713 * screwed and have to go OOM.
4715 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4716 ac->highest_zoneidx, ac->nodemask) {
4717 unsigned long available;
4718 unsigned long reclaimable;
4719 unsigned long min_wmark = min_wmark_pages(zone);
4722 available = reclaimable = zone_reclaimable_pages(zone);
4723 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4726 * Would the allocation succeed if we reclaimed all
4727 * reclaimable pages?
4729 wmark = __zone_watermark_ok(zone, order, min_wmark,
4730 ac->highest_zoneidx, alloc_flags, available);
4731 trace_reclaim_retry_zone(z, order, reclaimable,
4732 available, min_wmark, *no_progress_loops, wmark);
4735 * If we didn't make any progress and have a lot of
4736 * dirty + writeback pages then we should wait for
4737 * an IO to complete to slow down the reclaim and
4738 * prevent from pre mature OOM
4740 if (!did_some_progress) {
4741 unsigned long write_pending;
4743 write_pending = zone_page_state_snapshot(zone,
4744 NR_ZONE_WRITE_PENDING);
4746 if (2 * write_pending > reclaimable) {
4747 congestion_wait(BLK_RW_ASYNC, HZ/10);
4759 * Memory allocation/reclaim might be called from a WQ context and the
4760 * current implementation of the WQ concurrency control doesn't
4761 * recognize that a particular WQ is congested if the worker thread is
4762 * looping without ever sleeping. Therefore we have to do a short sleep
4763 * here rather than calling cond_resched().
4765 if (current->flags & PF_WQ_WORKER)
4766 schedule_timeout_uninterruptible(1);
4773 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4776 * It's possible that cpuset's mems_allowed and the nodemask from
4777 * mempolicy don't intersect. This should be normally dealt with by
4778 * policy_nodemask(), but it's possible to race with cpuset update in
4779 * such a way the check therein was true, and then it became false
4780 * before we got our cpuset_mems_cookie here.
4781 * This assumes that for all allocations, ac->nodemask can come only
4782 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4783 * when it does not intersect with the cpuset restrictions) or the
4784 * caller can deal with a violated nodemask.
4786 if (cpusets_enabled() && ac->nodemask &&
4787 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4788 ac->nodemask = NULL;
4793 * When updating a task's mems_allowed or mempolicy nodemask, it is
4794 * possible to race with parallel threads in such a way that our
4795 * allocation can fail while the mask is being updated. If we are about
4796 * to fail, check if the cpuset changed during allocation and if so,
4799 if (read_mems_allowed_retry(cpuset_mems_cookie))
4805 static inline struct page *
4806 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4807 struct alloc_context *ac)
4809 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4810 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4811 struct page *page = NULL;
4812 unsigned int alloc_flags;
4813 unsigned long did_some_progress;
4814 enum compact_priority compact_priority;
4815 enum compact_result compact_result;
4816 int compaction_retries;
4817 int no_progress_loops;
4818 unsigned int cpuset_mems_cookie;
4822 * We also sanity check to catch abuse of atomic reserves being used by
4823 * callers that are not in atomic context.
4825 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4826 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4827 gfp_mask &= ~__GFP_ATOMIC;
4830 compaction_retries = 0;
4831 no_progress_loops = 0;
4832 compact_priority = DEF_COMPACT_PRIORITY;
4833 cpuset_mems_cookie = read_mems_allowed_begin();
4836 * The fast path uses conservative alloc_flags to succeed only until
4837 * kswapd needs to be woken up, and to avoid the cost of setting up
4838 * alloc_flags precisely. So we do that now.
4840 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4843 * We need to recalculate the starting point for the zonelist iterator
4844 * because we might have used different nodemask in the fast path, or
4845 * there was a cpuset modification and we are retrying - otherwise we
4846 * could end up iterating over non-eligible zones endlessly.
4848 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4849 ac->highest_zoneidx, ac->nodemask);
4850 if (!ac->preferred_zoneref->zone)
4853 if (alloc_flags & ALLOC_KSWAPD)
4854 wake_all_kswapds(order, gfp_mask, ac);
4857 * The adjusted alloc_flags might result in immediate success, so try
4860 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4865 * For costly allocations, try direct compaction first, as it's likely
4866 * that we have enough base pages and don't need to reclaim. For non-
4867 * movable high-order allocations, do that as well, as compaction will
4868 * try prevent permanent fragmentation by migrating from blocks of the
4870 * Don't try this for allocations that are allowed to ignore
4871 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4873 if (can_direct_reclaim &&
4875 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4876 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4877 page = __alloc_pages_direct_compact(gfp_mask, order,
4879 INIT_COMPACT_PRIORITY,
4885 * Checks for costly allocations with __GFP_NORETRY, which
4886 * includes some THP page fault allocations
4888 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4890 * If allocating entire pageblock(s) and compaction
4891 * failed because all zones are below low watermarks
4892 * or is prohibited because it recently failed at this
4893 * order, fail immediately unless the allocator has
4894 * requested compaction and reclaim retry.
4897 * - potentially very expensive because zones are far
4898 * below their low watermarks or this is part of very
4899 * bursty high order allocations,
4900 * - not guaranteed to help because isolate_freepages()
4901 * may not iterate over freed pages as part of its
4903 * - unlikely to make entire pageblocks free on its
4906 if (compact_result == COMPACT_SKIPPED ||
4907 compact_result == COMPACT_DEFERRED)
4911 * Looks like reclaim/compaction is worth trying, but
4912 * sync compaction could be very expensive, so keep
4913 * using async compaction.
4915 compact_priority = INIT_COMPACT_PRIORITY;
4920 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4921 if (alloc_flags & ALLOC_KSWAPD)
4922 wake_all_kswapds(order, gfp_mask, ac);
4924 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4926 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
4929 * Reset the nodemask and zonelist iterators if memory policies can be
4930 * ignored. These allocations are high priority and system rather than
4933 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4934 ac->nodemask = NULL;
4935 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4936 ac->highest_zoneidx, ac->nodemask);
4939 /* Attempt with potentially adjusted zonelist and alloc_flags */
4940 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4944 /* Caller is not willing to reclaim, we can't balance anything */
4945 if (!can_direct_reclaim)
4948 /* Avoid recursion of direct reclaim */
4949 if (current->flags & PF_MEMALLOC)
4952 /* Try direct reclaim and then allocating */
4953 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4954 &did_some_progress);
4958 /* Try direct compaction and then allocating */
4959 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4960 compact_priority, &compact_result);
4964 /* Do not loop if specifically requested */
4965 if (gfp_mask & __GFP_NORETRY)
4969 * Do not retry costly high order allocations unless they are
4970 * __GFP_RETRY_MAYFAIL
4972 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4975 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4976 did_some_progress > 0, &no_progress_loops))
4980 * It doesn't make any sense to retry for the compaction if the order-0
4981 * reclaim is not able to make any progress because the current
4982 * implementation of the compaction depends on the sufficient amount
4983 * of free memory (see __compaction_suitable)
4985 if (did_some_progress > 0 &&
4986 should_compact_retry(ac, order, alloc_flags,
4987 compact_result, &compact_priority,
4988 &compaction_retries))
4992 /* Deal with possible cpuset update races before we start OOM killing */
4993 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4996 /* Reclaim has failed us, start killing things */
4997 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5001 /* Avoid allocations with no watermarks from looping endlessly */
5002 if (tsk_is_oom_victim(current) &&
5003 (alloc_flags & ALLOC_OOM ||
5004 (gfp_mask & __GFP_NOMEMALLOC)))
5007 /* Retry as long as the OOM killer is making progress */
5008 if (did_some_progress) {
5009 no_progress_loops = 0;
5014 /* Deal with possible cpuset update races before we fail */
5015 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5019 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5022 if (gfp_mask & __GFP_NOFAIL) {
5024 * All existing users of the __GFP_NOFAIL are blockable, so warn
5025 * of any new users that actually require GFP_NOWAIT
5027 if (WARN_ON_ONCE(!can_direct_reclaim))
5031 * PF_MEMALLOC request from this context is rather bizarre
5032 * because we cannot reclaim anything and only can loop waiting
5033 * for somebody to do a work for us
5035 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5038 * non failing costly orders are a hard requirement which we
5039 * are not prepared for much so let's warn about these users
5040 * so that we can identify them and convert them to something
5043 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5046 * Help non-failing allocations by giving them access to memory
5047 * reserves but do not use ALLOC_NO_WATERMARKS because this
5048 * could deplete whole memory reserves which would just make
5049 * the situation worse
5051 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5059 warn_alloc(gfp_mask, ac->nodemask,
5060 "page allocation failure: order:%u", order);
5065 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5066 int preferred_nid, nodemask_t *nodemask,
5067 struct alloc_context *ac, gfp_t *alloc_gfp,
5068 unsigned int *alloc_flags)
5070 ac->highest_zoneidx = gfp_zone(gfp_mask);
5071 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5072 ac->nodemask = nodemask;
5073 ac->migratetype = gfp_migratetype(gfp_mask);
5075 if (cpusets_enabled()) {
5076 *alloc_gfp |= __GFP_HARDWALL;
5078 * When we are in the interrupt context, it is irrelevant
5079 * to the current task context. It means that any node ok.
5081 if (!in_interrupt() && !ac->nodemask)
5082 ac->nodemask = &cpuset_current_mems_allowed;
5084 *alloc_flags |= ALLOC_CPUSET;
5087 fs_reclaim_acquire(gfp_mask);
5088 fs_reclaim_release(gfp_mask);
5090 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5092 if (should_fail_alloc_page(gfp_mask, order))
5095 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5097 /* Dirty zone balancing only done in the fast path */
5098 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5101 * The preferred zone is used for statistics but crucially it is
5102 * also used as the starting point for the zonelist iterator. It
5103 * may get reset for allocations that ignore memory policies.
5105 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5106 ac->highest_zoneidx, ac->nodemask);
5112 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5113 * @gfp: GFP flags for the allocation
5114 * @preferred_nid: The preferred NUMA node ID to allocate from
5115 * @nodemask: Set of nodes to allocate from, may be NULL
5116 * @nr_pages: The number of pages desired on the list or array
5117 * @page_list: Optional list to store the allocated pages
5118 * @page_array: Optional array to store the pages
5120 * This is a batched version of the page allocator that attempts to
5121 * allocate nr_pages quickly. Pages are added to page_list if page_list
5122 * is not NULL, otherwise it is assumed that the page_array is valid.
5124 * For lists, nr_pages is the number of pages that should be allocated.
5126 * For arrays, only NULL elements are populated with pages and nr_pages
5127 * is the maximum number of pages that will be stored in the array.
5129 * Returns the number of pages on the list or array.
5131 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5132 nodemask_t *nodemask, int nr_pages,
5133 struct list_head *page_list,
5134 struct page **page_array)
5137 unsigned long flags;
5140 struct per_cpu_pages *pcp;
5141 struct list_head *pcp_list;
5142 struct alloc_context ac;
5144 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5145 int nr_populated = 0, nr_account = 0;
5147 if (unlikely(nr_pages <= 0))
5151 * Skip populated array elements to determine if any pages need
5152 * to be allocated before disabling IRQs.
5154 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5157 /* Already populated array? */
5158 if (unlikely(page_array && nr_pages - nr_populated == 0))
5159 return nr_populated;
5161 /* Use the single page allocator for one page. */
5162 if (nr_pages - nr_populated == 1)
5165 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5166 gfp &= gfp_allowed_mask;
5168 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5172 /* Find an allowed local zone that meets the low watermark. */
5173 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5176 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5177 !__cpuset_zone_allowed(zone, gfp)) {
5181 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5182 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5186 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5187 if (zone_watermark_fast(zone, 0, mark,
5188 zonelist_zone_idx(ac.preferred_zoneref),
5189 alloc_flags, gfp)) {
5195 * If there are no allowed local zones that meets the watermarks then
5196 * try to allocate a single page and reclaim if necessary.
5198 if (unlikely(!zone))
5201 /* Attempt the batch allocation */
5202 local_lock_irqsave(&pagesets.lock, flags);
5203 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5204 pcp_list = &pcp->lists[ac.migratetype];
5206 while (nr_populated < nr_pages) {
5208 /* Skip existing pages */
5209 if (page_array && page_array[nr_populated]) {
5214 page = __rmqueue_pcplist(zone, ac.migratetype, alloc_flags,
5216 if (unlikely(!page)) {
5217 /* Try and get at least one page */
5224 prep_new_page(page, 0, gfp, 0);
5226 list_add(&page->lru, page_list);
5228 page_array[nr_populated] = page;
5232 local_unlock_irqrestore(&pagesets.lock, flags);
5234 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5235 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5237 return nr_populated;
5240 local_unlock_irqrestore(&pagesets.lock, flags);
5243 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5246 list_add(&page->lru, page_list);
5248 page_array[nr_populated] = page;
5252 return nr_populated;
5254 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5257 * This is the 'heart' of the zoned buddy allocator.
5259 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5260 nodemask_t *nodemask)
5263 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5264 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5265 struct alloc_context ac = { };
5268 * There are several places where we assume that the order value is sane
5269 * so bail out early if the request is out of bound.
5271 if (unlikely(order >= MAX_ORDER)) {
5272 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5276 gfp &= gfp_allowed_mask;
5278 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5279 * resp. GFP_NOIO which has to be inherited for all allocation requests
5280 * from a particular context which has been marked by
5281 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5282 * movable zones are not used during allocation.
5284 gfp = current_gfp_context(gfp);
5286 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5287 &alloc_gfp, &alloc_flags))
5291 * Forbid the first pass from falling back to types that fragment
5292 * memory until all local zones are considered.
5294 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5296 /* First allocation attempt */
5297 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5302 ac.spread_dirty_pages = false;
5305 * Restore the original nodemask if it was potentially replaced with
5306 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5308 ac.nodemask = nodemask;
5310 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5313 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5314 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5315 __free_pages(page, order);
5319 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5323 EXPORT_SYMBOL(__alloc_pages);
5326 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5327 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5328 * you need to access high mem.
5330 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5334 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5337 return (unsigned long) page_address(page);
5339 EXPORT_SYMBOL(__get_free_pages);
5341 unsigned long get_zeroed_page(gfp_t gfp_mask)
5343 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5345 EXPORT_SYMBOL(get_zeroed_page);
5348 * __free_pages - Free pages allocated with alloc_pages().
5349 * @page: The page pointer returned from alloc_pages().
5350 * @order: The order of the allocation.
5352 * This function can free multi-page allocations that are not compound
5353 * pages. It does not check that the @order passed in matches that of
5354 * the allocation, so it is easy to leak memory. Freeing more memory
5355 * than was allocated will probably emit a warning.
5357 * If the last reference to this page is speculative, it will be released
5358 * by put_page() which only frees the first page of a non-compound
5359 * allocation. To prevent the remaining pages from being leaked, we free
5360 * the subsequent pages here. If you want to use the page's reference
5361 * count to decide when to free the allocation, you should allocate a
5362 * compound page, and use put_page() instead of __free_pages().
5364 * Context: May be called in interrupt context or while holding a normal
5365 * spinlock, but not in NMI context or while holding a raw spinlock.
5367 void __free_pages(struct page *page, unsigned int order)
5369 if (put_page_testzero(page))
5370 free_the_page(page, order);
5371 else if (!PageHead(page))
5373 free_the_page(page + (1 << order), order);
5375 EXPORT_SYMBOL(__free_pages);
5377 void free_pages(unsigned long addr, unsigned int order)
5380 VM_BUG_ON(!virt_addr_valid((void *)addr));
5381 __free_pages(virt_to_page((void *)addr), order);
5385 EXPORT_SYMBOL(free_pages);
5389 * An arbitrary-length arbitrary-offset area of memory which resides
5390 * within a 0 or higher order page. Multiple fragments within that page
5391 * are individually refcounted, in the page's reference counter.
5393 * The page_frag functions below provide a simple allocation framework for
5394 * page fragments. This is used by the network stack and network device
5395 * drivers to provide a backing region of memory for use as either an
5396 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5398 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5401 struct page *page = NULL;
5402 gfp_t gfp = gfp_mask;
5404 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5405 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5407 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5408 PAGE_FRAG_CACHE_MAX_ORDER);
5409 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5411 if (unlikely(!page))
5412 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5414 nc->va = page ? page_address(page) : NULL;
5419 void __page_frag_cache_drain(struct page *page, unsigned int count)
5421 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5423 if (page_ref_sub_and_test(page, count))
5424 free_the_page(page, compound_order(page));
5426 EXPORT_SYMBOL(__page_frag_cache_drain);
5428 void *page_frag_alloc_align(struct page_frag_cache *nc,
5429 unsigned int fragsz, gfp_t gfp_mask,
5430 unsigned int align_mask)
5432 unsigned int size = PAGE_SIZE;
5436 if (unlikely(!nc->va)) {
5438 page = __page_frag_cache_refill(nc, gfp_mask);
5442 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5443 /* if size can vary use size else just use PAGE_SIZE */
5446 /* Even if we own the page, we do not use atomic_set().
5447 * This would break get_page_unless_zero() users.
5449 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5451 /* reset page count bias and offset to start of new frag */
5452 nc->pfmemalloc = page_is_pfmemalloc(page);
5453 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5457 offset = nc->offset - fragsz;
5458 if (unlikely(offset < 0)) {
5459 page = virt_to_page(nc->va);
5461 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5464 if (unlikely(nc->pfmemalloc)) {
5465 free_the_page(page, compound_order(page));
5469 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5470 /* if size can vary use size else just use PAGE_SIZE */
5473 /* OK, page count is 0, we can safely set it */
5474 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5476 /* reset page count bias and offset to start of new frag */
5477 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5478 offset = size - fragsz;
5482 offset &= align_mask;
5483 nc->offset = offset;
5485 return nc->va + offset;
5487 EXPORT_SYMBOL(page_frag_alloc_align);
5490 * Frees a page fragment allocated out of either a compound or order 0 page.
5492 void page_frag_free(void *addr)
5494 struct page *page = virt_to_head_page(addr);
5496 if (unlikely(put_page_testzero(page)))
5497 free_the_page(page, compound_order(page));
5499 EXPORT_SYMBOL(page_frag_free);
5501 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5505 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5506 unsigned long used = addr + PAGE_ALIGN(size);
5508 split_page(virt_to_page((void *)addr), order);
5509 while (used < alloc_end) {
5514 return (void *)addr;
5518 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5519 * @size: the number of bytes to allocate
5520 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5522 * This function is similar to alloc_pages(), except that it allocates the
5523 * minimum number of pages to satisfy the request. alloc_pages() can only
5524 * allocate memory in power-of-two pages.
5526 * This function is also limited by MAX_ORDER.
5528 * Memory allocated by this function must be released by free_pages_exact().
5530 * Return: pointer to the allocated area or %NULL in case of error.
5532 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5534 unsigned int order = get_order(size);
5537 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5538 gfp_mask &= ~__GFP_COMP;
5540 addr = __get_free_pages(gfp_mask, order);
5541 return make_alloc_exact(addr, order, size);
5543 EXPORT_SYMBOL(alloc_pages_exact);
5546 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5548 * @nid: the preferred node ID where memory should be allocated
5549 * @size: the number of bytes to allocate
5550 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5552 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5555 * Return: pointer to the allocated area or %NULL in case of error.
5557 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5559 unsigned int order = get_order(size);
5562 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5563 gfp_mask &= ~__GFP_COMP;
5565 p = alloc_pages_node(nid, gfp_mask, order);
5568 return make_alloc_exact((unsigned long)page_address(p), order, size);
5572 * free_pages_exact - release memory allocated via alloc_pages_exact()
5573 * @virt: the value returned by alloc_pages_exact.
5574 * @size: size of allocation, same value as passed to alloc_pages_exact().
5576 * Release the memory allocated by a previous call to alloc_pages_exact.
5578 void free_pages_exact(void *virt, size_t size)
5580 unsigned long addr = (unsigned long)virt;
5581 unsigned long end = addr + PAGE_ALIGN(size);
5583 while (addr < end) {
5588 EXPORT_SYMBOL(free_pages_exact);
5591 * nr_free_zone_pages - count number of pages beyond high watermark
5592 * @offset: The zone index of the highest zone
5594 * nr_free_zone_pages() counts the number of pages which are beyond the
5595 * high watermark within all zones at or below a given zone index. For each
5596 * zone, the number of pages is calculated as:
5598 * nr_free_zone_pages = managed_pages - high_pages
5600 * Return: number of pages beyond high watermark.
5602 static unsigned long nr_free_zone_pages(int offset)
5607 /* Just pick one node, since fallback list is circular */
5608 unsigned long sum = 0;
5610 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5612 for_each_zone_zonelist(zone, z, zonelist, offset) {
5613 unsigned long size = zone_managed_pages(zone);
5614 unsigned long high = high_wmark_pages(zone);
5623 * nr_free_buffer_pages - count number of pages beyond high watermark
5625 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5626 * watermark within ZONE_DMA and ZONE_NORMAL.
5628 * Return: number of pages beyond high watermark within ZONE_DMA and
5631 unsigned long nr_free_buffer_pages(void)
5633 return nr_free_zone_pages(gfp_zone(GFP_USER));
5635 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5637 static inline void show_node(struct zone *zone)
5639 if (IS_ENABLED(CONFIG_NUMA))
5640 printk("Node %d ", zone_to_nid(zone));
5643 long si_mem_available(void)
5646 unsigned long pagecache;
5647 unsigned long wmark_low = 0;
5648 unsigned long pages[NR_LRU_LISTS];
5649 unsigned long reclaimable;
5653 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5654 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5657 wmark_low += low_wmark_pages(zone);
5660 * Estimate the amount of memory available for userspace allocations,
5661 * without causing swapping.
5663 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5666 * Not all the page cache can be freed, otherwise the system will
5667 * start swapping. Assume at least half of the page cache, or the
5668 * low watermark worth of cache, needs to stay.
5670 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5671 pagecache -= min(pagecache / 2, wmark_low);
5672 available += pagecache;
5675 * Part of the reclaimable slab and other kernel memory consists of
5676 * items that are in use, and cannot be freed. Cap this estimate at the
5679 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5680 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5681 available += reclaimable - min(reclaimable / 2, wmark_low);
5687 EXPORT_SYMBOL_GPL(si_mem_available);
5689 void si_meminfo(struct sysinfo *val)
5691 val->totalram = totalram_pages();
5692 val->sharedram = global_node_page_state(NR_SHMEM);
5693 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5694 val->bufferram = nr_blockdev_pages();
5695 val->totalhigh = totalhigh_pages();
5696 val->freehigh = nr_free_highpages();
5697 val->mem_unit = PAGE_SIZE;
5700 EXPORT_SYMBOL(si_meminfo);
5703 void si_meminfo_node(struct sysinfo *val, int nid)
5705 int zone_type; /* needs to be signed */
5706 unsigned long managed_pages = 0;
5707 unsigned long managed_highpages = 0;
5708 unsigned long free_highpages = 0;
5709 pg_data_t *pgdat = NODE_DATA(nid);
5711 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5712 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5713 val->totalram = managed_pages;
5714 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5715 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5716 #ifdef CONFIG_HIGHMEM
5717 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5718 struct zone *zone = &pgdat->node_zones[zone_type];
5720 if (is_highmem(zone)) {
5721 managed_highpages += zone_managed_pages(zone);
5722 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5725 val->totalhigh = managed_highpages;
5726 val->freehigh = free_highpages;
5728 val->totalhigh = managed_highpages;
5729 val->freehigh = free_highpages;
5731 val->mem_unit = PAGE_SIZE;
5736 * Determine whether the node should be displayed or not, depending on whether
5737 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5739 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5741 if (!(flags & SHOW_MEM_FILTER_NODES))
5745 * no node mask - aka implicit memory numa policy. Do not bother with
5746 * the synchronization - read_mems_allowed_begin - because we do not
5747 * have to be precise here.
5750 nodemask = &cpuset_current_mems_allowed;
5752 return !node_isset(nid, *nodemask);
5755 #define K(x) ((x) << (PAGE_SHIFT-10))
5757 static void show_migration_types(unsigned char type)
5759 static const char types[MIGRATE_TYPES] = {
5760 [MIGRATE_UNMOVABLE] = 'U',
5761 [MIGRATE_MOVABLE] = 'M',
5762 [MIGRATE_RECLAIMABLE] = 'E',
5763 [MIGRATE_HIGHATOMIC] = 'H',
5765 [MIGRATE_CMA] = 'C',
5767 #ifdef CONFIG_MEMORY_ISOLATION
5768 [MIGRATE_ISOLATE] = 'I',
5771 char tmp[MIGRATE_TYPES + 1];
5775 for (i = 0; i < MIGRATE_TYPES; i++) {
5776 if (type & (1 << i))
5781 printk(KERN_CONT "(%s) ", tmp);
5785 * Show free area list (used inside shift_scroll-lock stuff)
5786 * We also calculate the percentage fragmentation. We do this by counting the
5787 * memory on each free list with the exception of the first item on the list.
5790 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5793 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5795 unsigned long free_pcp = 0;
5800 for_each_populated_zone(zone) {
5801 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5804 for_each_online_cpu(cpu)
5805 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5808 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5809 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5810 " unevictable:%lu dirty:%lu writeback:%lu\n"
5811 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5812 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5813 " free:%lu free_pcp:%lu free_cma:%lu\n",
5814 global_node_page_state(NR_ACTIVE_ANON),
5815 global_node_page_state(NR_INACTIVE_ANON),
5816 global_node_page_state(NR_ISOLATED_ANON),
5817 global_node_page_state(NR_ACTIVE_FILE),
5818 global_node_page_state(NR_INACTIVE_FILE),
5819 global_node_page_state(NR_ISOLATED_FILE),
5820 global_node_page_state(NR_UNEVICTABLE),
5821 global_node_page_state(NR_FILE_DIRTY),
5822 global_node_page_state(NR_WRITEBACK),
5823 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5824 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5825 global_node_page_state(NR_FILE_MAPPED),
5826 global_node_page_state(NR_SHMEM),
5827 global_node_page_state(NR_PAGETABLE),
5828 global_zone_page_state(NR_BOUNCE),
5829 global_zone_page_state(NR_FREE_PAGES),
5831 global_zone_page_state(NR_FREE_CMA_PAGES));
5833 for_each_online_pgdat(pgdat) {
5834 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5838 " active_anon:%lukB"
5839 " inactive_anon:%lukB"
5840 " active_file:%lukB"
5841 " inactive_file:%lukB"
5842 " unevictable:%lukB"
5843 " isolated(anon):%lukB"
5844 " isolated(file):%lukB"
5849 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5851 " shmem_pmdmapped: %lukB"
5854 " writeback_tmp:%lukB"
5855 " kernel_stack:%lukB"
5856 #ifdef CONFIG_SHADOW_CALL_STACK
5857 " shadow_call_stack:%lukB"
5860 " all_unreclaimable? %s"
5863 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5864 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5865 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5866 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5867 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5868 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5869 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5870 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5871 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5872 K(node_page_state(pgdat, NR_WRITEBACK)),
5873 K(node_page_state(pgdat, NR_SHMEM)),
5874 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5875 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5876 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5877 K(node_page_state(pgdat, NR_ANON_THPS)),
5879 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5880 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5881 #ifdef CONFIG_SHADOW_CALL_STACK
5882 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5884 K(node_page_state(pgdat, NR_PAGETABLE)),
5885 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5889 for_each_populated_zone(zone) {
5892 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5896 for_each_online_cpu(cpu)
5897 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5906 " reserved_highatomic:%luKB"
5907 " active_anon:%lukB"
5908 " inactive_anon:%lukB"
5909 " active_file:%lukB"
5910 " inactive_file:%lukB"
5911 " unevictable:%lukB"
5912 " writepending:%lukB"
5922 K(zone_page_state(zone, NR_FREE_PAGES)),
5923 K(min_wmark_pages(zone)),
5924 K(low_wmark_pages(zone)),
5925 K(high_wmark_pages(zone)),
5926 K(zone->nr_reserved_highatomic),
5927 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5928 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5929 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5930 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5931 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5932 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5933 K(zone->present_pages),
5934 K(zone_managed_pages(zone)),
5935 K(zone_page_state(zone, NR_MLOCK)),
5936 K(zone_page_state(zone, NR_BOUNCE)),
5938 K(this_cpu_read(zone->per_cpu_pageset->count)),
5939 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5940 printk("lowmem_reserve[]:");
5941 for (i = 0; i < MAX_NR_ZONES; i++)
5942 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5943 printk(KERN_CONT "\n");
5946 for_each_populated_zone(zone) {
5948 unsigned long nr[MAX_ORDER], flags, total = 0;
5949 unsigned char types[MAX_ORDER];
5951 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5954 printk(KERN_CONT "%s: ", zone->name);
5956 spin_lock_irqsave(&zone->lock, flags);
5957 for (order = 0; order < MAX_ORDER; order++) {
5958 struct free_area *area = &zone->free_area[order];
5961 nr[order] = area->nr_free;
5962 total += nr[order] << order;
5965 for (type = 0; type < MIGRATE_TYPES; type++) {
5966 if (!free_area_empty(area, type))
5967 types[order] |= 1 << type;
5970 spin_unlock_irqrestore(&zone->lock, flags);
5971 for (order = 0; order < MAX_ORDER; order++) {
5972 printk(KERN_CONT "%lu*%lukB ",
5973 nr[order], K(1UL) << order);
5975 show_migration_types(types[order]);
5977 printk(KERN_CONT "= %lukB\n", K(total));
5980 hugetlb_show_meminfo();
5982 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5984 show_swap_cache_info();
5987 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5989 zoneref->zone = zone;
5990 zoneref->zone_idx = zone_idx(zone);
5994 * Builds allocation fallback zone lists.
5996 * Add all populated zones of a node to the zonelist.
5998 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6001 enum zone_type zone_type = MAX_NR_ZONES;
6006 zone = pgdat->node_zones + zone_type;
6007 if (managed_zone(zone)) {
6008 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6009 check_highest_zone(zone_type);
6011 } while (zone_type);
6018 static int __parse_numa_zonelist_order(char *s)
6021 * We used to support different zonelists modes but they turned
6022 * out to be just not useful. Let's keep the warning in place
6023 * if somebody still use the cmd line parameter so that we do
6024 * not fail it silently
6026 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6027 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6033 char numa_zonelist_order[] = "Node";
6036 * sysctl handler for numa_zonelist_order
6038 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6039 void *buffer, size_t *length, loff_t *ppos)
6042 return __parse_numa_zonelist_order(buffer);
6043 return proc_dostring(table, write, buffer, length, ppos);
6047 #define MAX_NODE_LOAD (nr_online_nodes)
6048 static int node_load[MAX_NUMNODES];
6051 * find_next_best_node - find the next node that should appear in a given node's fallback list
6052 * @node: node whose fallback list we're appending
6053 * @used_node_mask: nodemask_t of already used nodes
6055 * We use a number of factors to determine which is the next node that should
6056 * appear on a given node's fallback list. The node should not have appeared
6057 * already in @node's fallback list, and it should be the next closest node
6058 * according to the distance array (which contains arbitrary distance values
6059 * from each node to each node in the system), and should also prefer nodes
6060 * with no CPUs, since presumably they'll have very little allocation pressure
6061 * on them otherwise.
6063 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6065 static int find_next_best_node(int node, nodemask_t *used_node_mask)
6068 int min_val = INT_MAX;
6069 int best_node = NUMA_NO_NODE;
6071 /* Use the local node if we haven't already */
6072 if (!node_isset(node, *used_node_mask)) {
6073 node_set(node, *used_node_mask);
6077 for_each_node_state(n, N_MEMORY) {
6079 /* Don't want a node to appear more than once */
6080 if (node_isset(n, *used_node_mask))
6083 /* Use the distance array to find the distance */
6084 val = node_distance(node, n);
6086 /* Penalize nodes under us ("prefer the next node") */
6089 /* Give preference to headless and unused nodes */
6090 if (!cpumask_empty(cpumask_of_node(n)))
6091 val += PENALTY_FOR_NODE_WITH_CPUS;
6093 /* Slight preference for less loaded node */
6094 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6095 val += node_load[n];
6097 if (val < min_val) {
6104 node_set(best_node, *used_node_mask);
6111 * Build zonelists ordered by node and zones within node.
6112 * This results in maximum locality--normal zone overflows into local
6113 * DMA zone, if any--but risks exhausting DMA zone.
6115 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6118 struct zoneref *zonerefs;
6121 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6123 for (i = 0; i < nr_nodes; i++) {
6126 pg_data_t *node = NODE_DATA(node_order[i]);
6128 nr_zones = build_zonerefs_node(node, zonerefs);
6129 zonerefs += nr_zones;
6131 zonerefs->zone = NULL;
6132 zonerefs->zone_idx = 0;
6136 * Build gfp_thisnode zonelists
6138 static void build_thisnode_zonelists(pg_data_t *pgdat)
6140 struct zoneref *zonerefs;
6143 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6144 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6145 zonerefs += nr_zones;
6146 zonerefs->zone = NULL;
6147 zonerefs->zone_idx = 0;
6151 * Build zonelists ordered by zone and nodes within zones.
6152 * This results in conserving DMA zone[s] until all Normal memory is
6153 * exhausted, but results in overflowing to remote node while memory
6154 * may still exist in local DMA zone.
6157 static void build_zonelists(pg_data_t *pgdat)
6159 static int node_order[MAX_NUMNODES];
6160 int node, load, nr_nodes = 0;
6161 nodemask_t used_mask = NODE_MASK_NONE;
6162 int local_node, prev_node;
6164 /* NUMA-aware ordering of nodes */
6165 local_node = pgdat->node_id;
6166 load = nr_online_nodes;
6167 prev_node = local_node;
6169 memset(node_order, 0, sizeof(node_order));
6170 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6172 * We don't want to pressure a particular node.
6173 * So adding penalty to the first node in same
6174 * distance group to make it round-robin.
6176 if (node_distance(local_node, node) !=
6177 node_distance(local_node, prev_node))
6178 node_load[node] = load;
6180 node_order[nr_nodes++] = node;
6185 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6186 build_thisnode_zonelists(pgdat);
6189 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6191 * Return node id of node used for "local" allocations.
6192 * I.e., first node id of first zone in arg node's generic zonelist.
6193 * Used for initializing percpu 'numa_mem', which is used primarily
6194 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6196 int local_memory_node(int node)
6200 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6201 gfp_zone(GFP_KERNEL),
6203 return zone_to_nid(z->zone);
6207 static void setup_min_unmapped_ratio(void);
6208 static void setup_min_slab_ratio(void);
6209 #else /* CONFIG_NUMA */
6211 static void build_zonelists(pg_data_t *pgdat)
6213 int node, local_node;
6214 struct zoneref *zonerefs;
6217 local_node = pgdat->node_id;
6219 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6220 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6221 zonerefs += nr_zones;
6224 * Now we build the zonelist so that it contains the zones
6225 * of all the other nodes.
6226 * We don't want to pressure a particular node, so when
6227 * building the zones for node N, we make sure that the
6228 * zones coming right after the local ones are those from
6229 * node N+1 (modulo N)
6231 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6232 if (!node_online(node))
6234 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6235 zonerefs += nr_zones;
6237 for (node = 0; node < local_node; node++) {
6238 if (!node_online(node))
6240 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6241 zonerefs += nr_zones;
6244 zonerefs->zone = NULL;
6245 zonerefs->zone_idx = 0;
6248 #endif /* CONFIG_NUMA */
6251 * Boot pageset table. One per cpu which is going to be used for all
6252 * zones and all nodes. The parameters will be set in such a way
6253 * that an item put on a list will immediately be handed over to
6254 * the buddy list. This is safe since pageset manipulation is done
6255 * with interrupts disabled.
6257 * The boot_pagesets must be kept even after bootup is complete for
6258 * unused processors and/or zones. They do play a role for bootstrapping
6259 * hotplugged processors.
6261 * zoneinfo_show() and maybe other functions do
6262 * not check if the processor is online before following the pageset pointer.
6263 * Other parts of the kernel may not check if the zone is available.
6265 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6266 /* These effectively disable the pcplists in the boot pageset completely */
6267 #define BOOT_PAGESET_HIGH 0
6268 #define BOOT_PAGESET_BATCH 1
6269 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6270 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6271 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6273 static void __build_all_zonelists(void *data)
6276 int __maybe_unused cpu;
6277 pg_data_t *self = data;
6278 static DEFINE_SPINLOCK(lock);
6283 memset(node_load, 0, sizeof(node_load));
6287 * This node is hotadded and no memory is yet present. So just
6288 * building zonelists is fine - no need to touch other nodes.
6290 if (self && !node_online(self->node_id)) {
6291 build_zonelists(self);
6293 for_each_online_node(nid) {
6294 pg_data_t *pgdat = NODE_DATA(nid);
6296 build_zonelists(pgdat);
6299 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6301 * We now know the "local memory node" for each node--
6302 * i.e., the node of the first zone in the generic zonelist.
6303 * Set up numa_mem percpu variable for on-line cpus. During
6304 * boot, only the boot cpu should be on-line; we'll init the
6305 * secondary cpus' numa_mem as they come on-line. During
6306 * node/memory hotplug, we'll fixup all on-line cpus.
6308 for_each_online_cpu(cpu)
6309 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6316 static noinline void __init
6317 build_all_zonelists_init(void)
6321 __build_all_zonelists(NULL);
6324 * Initialize the boot_pagesets that are going to be used
6325 * for bootstrapping processors. The real pagesets for
6326 * each zone will be allocated later when the per cpu
6327 * allocator is available.
6329 * boot_pagesets are used also for bootstrapping offline
6330 * cpus if the system is already booted because the pagesets
6331 * are needed to initialize allocators on a specific cpu too.
6332 * F.e. the percpu allocator needs the page allocator which
6333 * needs the percpu allocator in order to allocate its pagesets
6334 * (a chicken-egg dilemma).
6336 for_each_possible_cpu(cpu)
6337 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6339 mminit_verify_zonelist();
6340 cpuset_init_current_mems_allowed();
6344 * unless system_state == SYSTEM_BOOTING.
6346 * __ref due to call of __init annotated helper build_all_zonelists_init
6347 * [protected by SYSTEM_BOOTING].
6349 void __ref build_all_zonelists(pg_data_t *pgdat)
6351 unsigned long vm_total_pages;
6353 if (system_state == SYSTEM_BOOTING) {
6354 build_all_zonelists_init();
6356 __build_all_zonelists(pgdat);
6357 /* cpuset refresh routine should be here */
6359 /* Get the number of free pages beyond high watermark in all zones. */
6360 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6362 * Disable grouping by mobility if the number of pages in the
6363 * system is too low to allow the mechanism to work. It would be
6364 * more accurate, but expensive to check per-zone. This check is
6365 * made on memory-hotadd so a system can start with mobility
6366 * disabled and enable it later
6368 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6369 page_group_by_mobility_disabled = 1;
6371 page_group_by_mobility_disabled = 0;
6373 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6375 page_group_by_mobility_disabled ? "off" : "on",
6378 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6382 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6383 static bool __meminit
6384 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6386 static struct memblock_region *r;
6388 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6389 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6390 for_each_mem_region(r) {
6391 if (*pfn < memblock_region_memory_end_pfn(r))
6395 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6396 memblock_is_mirror(r)) {
6397 *pfn = memblock_region_memory_end_pfn(r);
6405 * Initially all pages are reserved - free ones are freed
6406 * up by memblock_free_all() once the early boot process is
6407 * done. Non-atomic initialization, single-pass.
6409 * All aligned pageblocks are initialized to the specified migratetype
6410 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6411 * zone stats (e.g., nr_isolate_pageblock) are touched.
6413 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6414 unsigned long start_pfn, unsigned long zone_end_pfn,
6415 enum meminit_context context,
6416 struct vmem_altmap *altmap, int migratetype)
6418 unsigned long pfn, end_pfn = start_pfn + size;
6421 if (highest_memmap_pfn < end_pfn - 1)
6422 highest_memmap_pfn = end_pfn - 1;
6424 #ifdef CONFIG_ZONE_DEVICE
6426 * Honor reservation requested by the driver for this ZONE_DEVICE
6427 * memory. We limit the total number of pages to initialize to just
6428 * those that might contain the memory mapping. We will defer the
6429 * ZONE_DEVICE page initialization until after we have released
6432 if (zone == ZONE_DEVICE) {
6436 if (start_pfn == altmap->base_pfn)
6437 start_pfn += altmap->reserve;
6438 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6442 for (pfn = start_pfn; pfn < end_pfn; ) {
6444 * There can be holes in boot-time mem_map[]s handed to this
6445 * function. They do not exist on hotplugged memory.
6447 if (context == MEMINIT_EARLY) {
6448 if (overlap_memmap_init(zone, &pfn))
6450 if (defer_init(nid, pfn, zone_end_pfn))
6454 page = pfn_to_page(pfn);
6455 __init_single_page(page, pfn, zone, nid);
6456 if (context == MEMINIT_HOTPLUG)
6457 __SetPageReserved(page);
6460 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6461 * such that unmovable allocations won't be scattered all
6462 * over the place during system boot.
6464 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6465 set_pageblock_migratetype(page, migratetype);
6472 #ifdef CONFIG_ZONE_DEVICE
6473 void __ref memmap_init_zone_device(struct zone *zone,
6474 unsigned long start_pfn,
6475 unsigned long nr_pages,
6476 struct dev_pagemap *pgmap)
6478 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6479 struct pglist_data *pgdat = zone->zone_pgdat;
6480 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6481 unsigned long zone_idx = zone_idx(zone);
6482 unsigned long start = jiffies;
6483 int nid = pgdat->node_id;
6485 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6489 * The call to memmap_init should have already taken care
6490 * of the pages reserved for the memmap, so we can just jump to
6491 * the end of that region and start processing the device pages.
6494 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6495 nr_pages = end_pfn - start_pfn;
6498 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6499 struct page *page = pfn_to_page(pfn);
6501 __init_single_page(page, pfn, zone_idx, nid);
6504 * Mark page reserved as it will need to wait for onlining
6505 * phase for it to be fully associated with a zone.
6507 * We can use the non-atomic __set_bit operation for setting
6508 * the flag as we are still initializing the pages.
6510 __SetPageReserved(page);
6513 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6514 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6515 * ever freed or placed on a driver-private list.
6517 page->pgmap = pgmap;
6518 page->zone_device_data = NULL;
6521 * Mark the block movable so that blocks are reserved for
6522 * movable at startup. This will force kernel allocations
6523 * to reserve their blocks rather than leaking throughout
6524 * the address space during boot when many long-lived
6525 * kernel allocations are made.
6527 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6528 * because this is done early in section_activate()
6530 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6531 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6536 pr_info("%s initialised %lu pages in %ums\n", __func__,
6537 nr_pages, jiffies_to_msecs(jiffies - start));
6541 static void __meminit zone_init_free_lists(struct zone *zone)
6543 unsigned int order, t;
6544 for_each_migratetype_order(order, t) {
6545 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6546 zone->free_area[order].nr_free = 0;
6550 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6552 * Only struct pages that correspond to ranges defined by memblock.memory
6553 * are zeroed and initialized by going through __init_single_page() during
6554 * memmap_init_zone_range().
6556 * But, there could be struct pages that correspond to holes in
6557 * memblock.memory. This can happen because of the following reasons:
6558 * - physical memory bank size is not necessarily the exact multiple of the
6559 * arbitrary section size
6560 * - early reserved memory may not be listed in memblock.memory
6561 * - memory layouts defined with memmap= kernel parameter may not align
6562 * nicely with memmap sections
6564 * Explicitly initialize those struct pages so that:
6565 * - PG_Reserved is set
6566 * - zone and node links point to zone and node that span the page if the
6567 * hole is in the middle of a zone
6568 * - zone and node links point to adjacent zone/node if the hole falls on
6569 * the zone boundary; the pages in such holes will be prepended to the
6570 * zone/node above the hole except for the trailing pages in the last
6571 * section that will be appended to the zone/node below.
6573 static void __init init_unavailable_range(unsigned long spfn,
6580 for (pfn = spfn; pfn < epfn; pfn++) {
6581 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6582 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6583 + pageblock_nr_pages - 1;
6586 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6587 __SetPageReserved(pfn_to_page(pfn));
6592 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6593 node, zone_names[zone], pgcnt);
6596 static inline void init_unavailable_range(unsigned long spfn,
6603 static void __init memmap_init_zone_range(struct zone *zone,
6604 unsigned long start_pfn,
6605 unsigned long end_pfn,
6606 unsigned long *hole_pfn)
6608 unsigned long zone_start_pfn = zone->zone_start_pfn;
6609 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6610 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6612 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6613 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6615 if (start_pfn >= end_pfn)
6618 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6619 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6621 if (*hole_pfn < start_pfn)
6622 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6624 *hole_pfn = end_pfn;
6627 static void __init memmap_init(void)
6629 unsigned long start_pfn, end_pfn;
6630 unsigned long hole_pfn = 0;
6631 int i, j, zone_id, nid;
6633 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6634 struct pglist_data *node = NODE_DATA(nid);
6636 for (j = 0; j < MAX_NR_ZONES; j++) {
6637 struct zone *zone = node->node_zones + j;
6639 if (!populated_zone(zone))
6642 memmap_init_zone_range(zone, start_pfn, end_pfn,
6648 #ifdef CONFIG_SPARSEMEM
6650 * Initialize the memory map for hole in the range [memory_end,
6652 * Append the pages in this hole to the highest zone in the last
6654 * The call to init_unavailable_range() is outside the ifdef to
6655 * silence the compiler warining about zone_id set but not used;
6656 * for FLATMEM it is a nop anyway
6658 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6659 if (hole_pfn < end_pfn)
6661 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6664 static int zone_batchsize(struct zone *zone)
6670 * The number of pages to batch allocate is either ~0.1%
6671 * of the zone or 1MB, whichever is smaller. The batch
6672 * size is striking a balance between allocation latency
6673 * and zone lock contention.
6675 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6676 batch /= 4; /* We effectively *= 4 below */
6681 * Clamp the batch to a 2^n - 1 value. Having a power
6682 * of 2 value was found to be more likely to have
6683 * suboptimal cache aliasing properties in some cases.
6685 * For example if 2 tasks are alternately allocating
6686 * batches of pages, one task can end up with a lot
6687 * of pages of one half of the possible page colors
6688 * and the other with pages of the other colors.
6690 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6695 /* The deferral and batching of frees should be suppressed under NOMMU
6698 * The problem is that NOMMU needs to be able to allocate large chunks
6699 * of contiguous memory as there's no hardware page translation to
6700 * assemble apparent contiguous memory from discontiguous pages.
6702 * Queueing large contiguous runs of pages for batching, however,
6703 * causes the pages to actually be freed in smaller chunks. As there
6704 * can be a significant delay between the individual batches being
6705 * recycled, this leads to the once large chunks of space being
6706 * fragmented and becoming unavailable for high-order allocations.
6712 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6717 unsigned long total_pages;
6719 if (!percpu_pagelist_high_fraction) {
6721 * By default, the high value of the pcp is based on the zone
6722 * low watermark so that if they are full then background
6723 * reclaim will not be started prematurely.
6725 total_pages = low_wmark_pages(zone);
6728 * If percpu_pagelist_high_fraction is configured, the high
6729 * value is based on a fraction of the managed pages in the
6732 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6736 * Split the high value across all online CPUs local to the zone. Note
6737 * that early in boot that CPUs may not be online yet and that during
6738 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6741 nr_local_cpus = max(1U, cpumask_weight(cpumask_of_node(zone_to_nid(zone)))) + cpu_online;
6742 high = total_pages / nr_local_cpus;
6745 * Ensure high is at least batch*4. The multiple is based on the
6746 * historical relationship between high and batch.
6748 high = max(high, batch << 2);
6757 * pcp->high and pcp->batch values are related and generally batch is lower
6758 * than high. They are also related to pcp->count such that count is lower
6759 * than high, and as soon as it reaches high, the pcplist is flushed.
6761 * However, guaranteeing these relations at all times would require e.g. write
6762 * barriers here but also careful usage of read barriers at the read side, and
6763 * thus be prone to error and bad for performance. Thus the update only prevents
6764 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6765 * can cope with those fields changing asynchronously, and fully trust only the
6766 * pcp->count field on the local CPU with interrupts disabled.
6768 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6769 * outside of boot time (or some other assurance that no concurrent updaters
6772 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6773 unsigned long batch)
6775 WRITE_ONCE(pcp->batch, batch);
6776 WRITE_ONCE(pcp->high, high);
6779 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6783 memset(pcp, 0, sizeof(*pcp));
6784 memset(pzstats, 0, sizeof(*pzstats));
6786 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6787 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6790 * Set batch and high values safe for a boot pageset. A true percpu
6791 * pageset's initialization will update them subsequently. Here we don't
6792 * need to be as careful as pageset_update() as nobody can access the
6795 pcp->high = BOOT_PAGESET_HIGH;
6796 pcp->batch = BOOT_PAGESET_BATCH;
6797 pcp->free_factor = 0;
6800 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6801 unsigned long batch)
6803 struct per_cpu_pages *pcp;
6806 for_each_possible_cpu(cpu) {
6807 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6808 pageset_update(pcp, high, batch);
6813 * Calculate and set new high and batch values for all per-cpu pagesets of a
6814 * zone based on the zone's size.
6816 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6818 int new_high, new_batch;
6820 new_batch = max(1, zone_batchsize(zone));
6821 new_high = zone_highsize(zone, new_batch, cpu_online);
6823 if (zone->pageset_high == new_high &&
6824 zone->pageset_batch == new_batch)
6827 zone->pageset_high = new_high;
6828 zone->pageset_batch = new_batch;
6830 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6833 void __meminit setup_zone_pageset(struct zone *zone)
6837 /* Size may be 0 on !SMP && !NUMA */
6838 if (sizeof(struct per_cpu_zonestat) > 0)
6839 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6841 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6842 for_each_possible_cpu(cpu) {
6843 struct per_cpu_pages *pcp;
6844 struct per_cpu_zonestat *pzstats;
6846 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6847 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6848 per_cpu_pages_init(pcp, pzstats);
6851 zone_set_pageset_high_and_batch(zone, 0);
6855 * Allocate per cpu pagesets and initialize them.
6856 * Before this call only boot pagesets were available.
6858 void __init setup_per_cpu_pageset(void)
6860 struct pglist_data *pgdat;
6862 int __maybe_unused cpu;
6864 for_each_populated_zone(zone)
6865 setup_zone_pageset(zone);
6869 * Unpopulated zones continue using the boot pagesets.
6870 * The numa stats for these pagesets need to be reset.
6871 * Otherwise, they will end up skewing the stats of
6872 * the nodes these zones are associated with.
6874 for_each_possible_cpu(cpu) {
6875 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
6876 memset(pzstats->vm_numa_event, 0,
6877 sizeof(pzstats->vm_numa_event));
6881 for_each_online_pgdat(pgdat)
6882 pgdat->per_cpu_nodestats =
6883 alloc_percpu(struct per_cpu_nodestat);
6886 static __meminit void zone_pcp_init(struct zone *zone)
6889 * per cpu subsystem is not up at this point. The following code
6890 * relies on the ability of the linker to provide the
6891 * offset of a (static) per cpu variable into the per cpu area.
6893 zone->per_cpu_pageset = &boot_pageset;
6894 zone->per_cpu_zonestats = &boot_zonestats;
6895 zone->pageset_high = BOOT_PAGESET_HIGH;
6896 zone->pageset_batch = BOOT_PAGESET_BATCH;
6898 if (populated_zone(zone))
6899 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
6900 zone->present_pages, zone_batchsize(zone));
6903 void __meminit init_currently_empty_zone(struct zone *zone,
6904 unsigned long zone_start_pfn,
6907 struct pglist_data *pgdat = zone->zone_pgdat;
6908 int zone_idx = zone_idx(zone) + 1;
6910 if (zone_idx > pgdat->nr_zones)
6911 pgdat->nr_zones = zone_idx;
6913 zone->zone_start_pfn = zone_start_pfn;
6915 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6916 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6918 (unsigned long)zone_idx(zone),
6919 zone_start_pfn, (zone_start_pfn + size));
6921 zone_init_free_lists(zone);
6922 zone->initialized = 1;
6926 * get_pfn_range_for_nid - Return the start and end page frames for a node
6927 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6928 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6929 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6931 * It returns the start and end page frame of a node based on information
6932 * provided by memblock_set_node(). If called for a node
6933 * with no available memory, a warning is printed and the start and end
6936 void __init get_pfn_range_for_nid(unsigned int nid,
6937 unsigned long *start_pfn, unsigned long *end_pfn)
6939 unsigned long this_start_pfn, this_end_pfn;
6945 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6946 *start_pfn = min(*start_pfn, this_start_pfn);
6947 *end_pfn = max(*end_pfn, this_end_pfn);
6950 if (*start_pfn == -1UL)
6955 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6956 * assumption is made that zones within a node are ordered in monotonic
6957 * increasing memory addresses so that the "highest" populated zone is used
6959 static void __init find_usable_zone_for_movable(void)
6962 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6963 if (zone_index == ZONE_MOVABLE)
6966 if (arch_zone_highest_possible_pfn[zone_index] >
6967 arch_zone_lowest_possible_pfn[zone_index])
6971 VM_BUG_ON(zone_index == -1);
6972 movable_zone = zone_index;
6976 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6977 * because it is sized independent of architecture. Unlike the other zones,
6978 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6979 * in each node depending on the size of each node and how evenly kernelcore
6980 * is distributed. This helper function adjusts the zone ranges
6981 * provided by the architecture for a given node by using the end of the
6982 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6983 * zones within a node are in order of monotonic increases memory addresses
6985 static void __init adjust_zone_range_for_zone_movable(int nid,
6986 unsigned long zone_type,
6987 unsigned long node_start_pfn,
6988 unsigned long node_end_pfn,
6989 unsigned long *zone_start_pfn,
6990 unsigned long *zone_end_pfn)
6992 /* Only adjust if ZONE_MOVABLE is on this node */
6993 if (zone_movable_pfn[nid]) {
6994 /* Size ZONE_MOVABLE */
6995 if (zone_type == ZONE_MOVABLE) {
6996 *zone_start_pfn = zone_movable_pfn[nid];
6997 *zone_end_pfn = min(node_end_pfn,
6998 arch_zone_highest_possible_pfn[movable_zone]);
7000 /* Adjust for ZONE_MOVABLE starting within this range */
7001 } else if (!mirrored_kernelcore &&
7002 *zone_start_pfn < zone_movable_pfn[nid] &&
7003 *zone_end_pfn > zone_movable_pfn[nid]) {
7004 *zone_end_pfn = zone_movable_pfn[nid];
7006 /* Check if this whole range is within ZONE_MOVABLE */
7007 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7008 *zone_start_pfn = *zone_end_pfn;
7013 * Return the number of pages a zone spans in a node, including holes
7014 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7016 static unsigned long __init zone_spanned_pages_in_node(int nid,
7017 unsigned long zone_type,
7018 unsigned long node_start_pfn,
7019 unsigned long node_end_pfn,
7020 unsigned long *zone_start_pfn,
7021 unsigned long *zone_end_pfn)
7023 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7024 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7025 /* When hotadd a new node from cpu_up(), the node should be empty */
7026 if (!node_start_pfn && !node_end_pfn)
7029 /* Get the start and end of the zone */
7030 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7031 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7032 adjust_zone_range_for_zone_movable(nid, zone_type,
7033 node_start_pfn, node_end_pfn,
7034 zone_start_pfn, zone_end_pfn);
7036 /* Check that this node has pages within the zone's required range */
7037 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7040 /* Move the zone boundaries inside the node if necessary */
7041 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7042 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7044 /* Return the spanned pages */
7045 return *zone_end_pfn - *zone_start_pfn;
7049 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7050 * then all holes in the requested range will be accounted for.
7052 unsigned long __init __absent_pages_in_range(int nid,
7053 unsigned long range_start_pfn,
7054 unsigned long range_end_pfn)
7056 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7057 unsigned long start_pfn, end_pfn;
7060 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7061 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7062 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7063 nr_absent -= end_pfn - start_pfn;
7069 * absent_pages_in_range - Return number of page frames in holes within a range
7070 * @start_pfn: The start PFN to start searching for holes
7071 * @end_pfn: The end PFN to stop searching for holes
7073 * Return: the number of pages frames in memory holes within a range.
7075 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7076 unsigned long end_pfn)
7078 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7081 /* Return the number of page frames in holes in a zone on a node */
7082 static unsigned long __init zone_absent_pages_in_node(int nid,
7083 unsigned long zone_type,
7084 unsigned long node_start_pfn,
7085 unsigned long node_end_pfn)
7087 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7088 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7089 unsigned long zone_start_pfn, zone_end_pfn;
7090 unsigned long nr_absent;
7092 /* When hotadd a new node from cpu_up(), the node should be empty */
7093 if (!node_start_pfn && !node_end_pfn)
7096 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7097 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7099 adjust_zone_range_for_zone_movable(nid, zone_type,
7100 node_start_pfn, node_end_pfn,
7101 &zone_start_pfn, &zone_end_pfn);
7102 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7105 * ZONE_MOVABLE handling.
7106 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7109 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7110 unsigned long start_pfn, end_pfn;
7111 struct memblock_region *r;
7113 for_each_mem_region(r) {
7114 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7115 zone_start_pfn, zone_end_pfn);
7116 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7117 zone_start_pfn, zone_end_pfn);
7119 if (zone_type == ZONE_MOVABLE &&
7120 memblock_is_mirror(r))
7121 nr_absent += end_pfn - start_pfn;
7123 if (zone_type == ZONE_NORMAL &&
7124 !memblock_is_mirror(r))
7125 nr_absent += end_pfn - start_pfn;
7132 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7133 unsigned long node_start_pfn,
7134 unsigned long node_end_pfn)
7136 unsigned long realtotalpages = 0, totalpages = 0;
7139 for (i = 0; i < MAX_NR_ZONES; i++) {
7140 struct zone *zone = pgdat->node_zones + i;
7141 unsigned long zone_start_pfn, zone_end_pfn;
7142 unsigned long spanned, absent;
7143 unsigned long size, real_size;
7145 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7150 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7155 real_size = size - absent;
7158 zone->zone_start_pfn = zone_start_pfn;
7160 zone->zone_start_pfn = 0;
7161 zone->spanned_pages = size;
7162 zone->present_pages = real_size;
7165 realtotalpages += real_size;
7168 pgdat->node_spanned_pages = totalpages;
7169 pgdat->node_present_pages = realtotalpages;
7170 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7173 #ifndef CONFIG_SPARSEMEM
7175 * Calculate the size of the zone->blockflags rounded to an unsigned long
7176 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7177 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7178 * round what is now in bits to nearest long in bits, then return it in
7181 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7183 unsigned long usemapsize;
7185 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7186 usemapsize = roundup(zonesize, pageblock_nr_pages);
7187 usemapsize = usemapsize >> pageblock_order;
7188 usemapsize *= NR_PAGEBLOCK_BITS;
7189 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7191 return usemapsize / 8;
7194 static void __ref setup_usemap(struct zone *zone)
7196 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7197 zone->spanned_pages);
7198 zone->pageblock_flags = NULL;
7200 zone->pageblock_flags =
7201 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7203 if (!zone->pageblock_flags)
7204 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7205 usemapsize, zone->name, zone_to_nid(zone));
7209 static inline void setup_usemap(struct zone *zone) {}
7210 #endif /* CONFIG_SPARSEMEM */
7212 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7214 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7215 void __init set_pageblock_order(void)
7219 /* Check that pageblock_nr_pages has not already been setup */
7220 if (pageblock_order)
7223 if (HPAGE_SHIFT > PAGE_SHIFT)
7224 order = HUGETLB_PAGE_ORDER;
7226 order = MAX_ORDER - 1;
7229 * Assume the largest contiguous order of interest is a huge page.
7230 * This value may be variable depending on boot parameters on IA64 and
7233 pageblock_order = order;
7235 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7238 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7239 * is unused as pageblock_order is set at compile-time. See
7240 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7243 void __init set_pageblock_order(void)
7247 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7249 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7250 unsigned long present_pages)
7252 unsigned long pages = spanned_pages;
7255 * Provide a more accurate estimation if there are holes within
7256 * the zone and SPARSEMEM is in use. If there are holes within the
7257 * zone, each populated memory region may cost us one or two extra
7258 * memmap pages due to alignment because memmap pages for each
7259 * populated regions may not be naturally aligned on page boundary.
7260 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7262 if (spanned_pages > present_pages + (present_pages >> 4) &&
7263 IS_ENABLED(CONFIG_SPARSEMEM))
7264 pages = present_pages;
7266 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7269 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7270 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7272 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7274 spin_lock_init(&ds_queue->split_queue_lock);
7275 INIT_LIST_HEAD(&ds_queue->split_queue);
7276 ds_queue->split_queue_len = 0;
7279 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7282 #ifdef CONFIG_COMPACTION
7283 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7285 init_waitqueue_head(&pgdat->kcompactd_wait);
7288 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7291 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7293 pgdat_resize_init(pgdat);
7295 pgdat_init_split_queue(pgdat);
7296 pgdat_init_kcompactd(pgdat);
7298 init_waitqueue_head(&pgdat->kswapd_wait);
7299 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7301 pgdat_page_ext_init(pgdat);
7302 lruvec_init(&pgdat->__lruvec);
7305 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7306 unsigned long remaining_pages)
7308 atomic_long_set(&zone->managed_pages, remaining_pages);
7309 zone_set_nid(zone, nid);
7310 zone->name = zone_names[idx];
7311 zone->zone_pgdat = NODE_DATA(nid);
7312 spin_lock_init(&zone->lock);
7313 zone_seqlock_init(zone);
7314 zone_pcp_init(zone);
7318 * Set up the zone data structures
7319 * - init pgdat internals
7320 * - init all zones belonging to this node
7322 * NOTE: this function is only called during memory hotplug
7324 #ifdef CONFIG_MEMORY_HOTPLUG
7325 void __ref free_area_init_core_hotplug(int nid)
7328 pg_data_t *pgdat = NODE_DATA(nid);
7330 pgdat_init_internals(pgdat);
7331 for (z = 0; z < MAX_NR_ZONES; z++)
7332 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7337 * Set up the zone data structures:
7338 * - mark all pages reserved
7339 * - mark all memory queues empty
7340 * - clear the memory bitmaps
7342 * NOTE: pgdat should get zeroed by caller.
7343 * NOTE: this function is only called during early init.
7345 static void __init free_area_init_core(struct pglist_data *pgdat)
7348 int nid = pgdat->node_id;
7350 pgdat_init_internals(pgdat);
7351 pgdat->per_cpu_nodestats = &boot_nodestats;
7353 for (j = 0; j < MAX_NR_ZONES; j++) {
7354 struct zone *zone = pgdat->node_zones + j;
7355 unsigned long size, freesize, memmap_pages;
7357 size = zone->spanned_pages;
7358 freesize = zone->present_pages;
7361 * Adjust freesize so that it accounts for how much memory
7362 * is used by this zone for memmap. This affects the watermark
7363 * and per-cpu initialisations
7365 memmap_pages = calc_memmap_size(size, freesize);
7366 if (!is_highmem_idx(j)) {
7367 if (freesize >= memmap_pages) {
7368 freesize -= memmap_pages;
7370 pr_debug(" %s zone: %lu pages used for memmap\n",
7371 zone_names[j], memmap_pages);
7373 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7374 zone_names[j], memmap_pages, freesize);
7377 /* Account for reserved pages */
7378 if (j == 0 && freesize > dma_reserve) {
7379 freesize -= dma_reserve;
7380 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7383 if (!is_highmem_idx(j))
7384 nr_kernel_pages += freesize;
7385 /* Charge for highmem memmap if there are enough kernel pages */
7386 else if (nr_kernel_pages > memmap_pages * 2)
7387 nr_kernel_pages -= memmap_pages;
7388 nr_all_pages += freesize;
7391 * Set an approximate value for lowmem here, it will be adjusted
7392 * when the bootmem allocator frees pages into the buddy system.
7393 * And all highmem pages will be managed by the buddy system.
7395 zone_init_internals(zone, j, nid, freesize);
7400 set_pageblock_order();
7402 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7406 #ifdef CONFIG_FLAT_NODE_MEM_MAP
7407 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7409 unsigned long __maybe_unused start = 0;
7410 unsigned long __maybe_unused offset = 0;
7412 /* Skip empty nodes */
7413 if (!pgdat->node_spanned_pages)
7416 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7417 offset = pgdat->node_start_pfn - start;
7418 /* ia64 gets its own node_mem_map, before this, without bootmem */
7419 if (!pgdat->node_mem_map) {
7420 unsigned long size, end;
7424 * The zone's endpoints aren't required to be MAX_ORDER
7425 * aligned but the node_mem_map endpoints must be in order
7426 * for the buddy allocator to function correctly.
7428 end = pgdat_end_pfn(pgdat);
7429 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7430 size = (end - start) * sizeof(struct page);
7431 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7434 panic("Failed to allocate %ld bytes for node %d memory map\n",
7435 size, pgdat->node_id);
7436 pgdat->node_mem_map = map + offset;
7438 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7439 __func__, pgdat->node_id, (unsigned long)pgdat,
7440 (unsigned long)pgdat->node_mem_map);
7443 * With no DISCONTIG, the global mem_map is just set as node 0's
7445 if (pgdat == NODE_DATA(0)) {
7446 mem_map = NODE_DATA(0)->node_mem_map;
7447 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7453 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7454 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7456 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7457 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7459 pgdat->first_deferred_pfn = ULONG_MAX;
7462 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7465 static void __init free_area_init_node(int nid)
7467 pg_data_t *pgdat = NODE_DATA(nid);
7468 unsigned long start_pfn = 0;
7469 unsigned long end_pfn = 0;
7471 /* pg_data_t should be reset to zero when it's allocated */
7472 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7474 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7476 pgdat->node_id = nid;
7477 pgdat->node_start_pfn = start_pfn;
7478 pgdat->per_cpu_nodestats = NULL;
7480 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7481 (u64)start_pfn << PAGE_SHIFT,
7482 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7483 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7485 alloc_node_mem_map(pgdat);
7486 pgdat_set_deferred_range(pgdat);
7488 free_area_init_core(pgdat);
7491 void __init free_area_init_memoryless_node(int nid)
7493 free_area_init_node(nid);
7496 #if MAX_NUMNODES > 1
7498 * Figure out the number of possible node ids.
7500 void __init setup_nr_node_ids(void)
7502 unsigned int highest;
7504 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7505 nr_node_ids = highest + 1;
7510 * node_map_pfn_alignment - determine the maximum internode alignment
7512 * This function should be called after node map is populated and sorted.
7513 * It calculates the maximum power of two alignment which can distinguish
7516 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7517 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7518 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7519 * shifted, 1GiB is enough and this function will indicate so.
7521 * This is used to test whether pfn -> nid mapping of the chosen memory
7522 * model has fine enough granularity to avoid incorrect mapping for the
7523 * populated node map.
7525 * Return: the determined alignment in pfn's. 0 if there is no alignment
7526 * requirement (single node).
7528 unsigned long __init node_map_pfn_alignment(void)
7530 unsigned long accl_mask = 0, last_end = 0;
7531 unsigned long start, end, mask;
7532 int last_nid = NUMA_NO_NODE;
7535 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7536 if (!start || last_nid < 0 || last_nid == nid) {
7543 * Start with a mask granular enough to pin-point to the
7544 * start pfn and tick off bits one-by-one until it becomes
7545 * too coarse to separate the current node from the last.
7547 mask = ~((1 << __ffs(start)) - 1);
7548 while (mask && last_end <= (start & (mask << 1)))
7551 /* accumulate all internode masks */
7555 /* convert mask to number of pages */
7556 return ~accl_mask + 1;
7560 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7562 * Return: the minimum PFN based on information provided via
7563 * memblock_set_node().
7565 unsigned long __init find_min_pfn_with_active_regions(void)
7567 return PHYS_PFN(memblock_start_of_DRAM());
7571 * early_calculate_totalpages()
7572 * Sum pages in active regions for movable zone.
7573 * Populate N_MEMORY for calculating usable_nodes.
7575 static unsigned long __init early_calculate_totalpages(void)
7577 unsigned long totalpages = 0;
7578 unsigned long start_pfn, end_pfn;
7581 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7582 unsigned long pages = end_pfn - start_pfn;
7584 totalpages += pages;
7586 node_set_state(nid, N_MEMORY);
7592 * Find the PFN the Movable zone begins in each node. Kernel memory
7593 * is spread evenly between nodes as long as the nodes have enough
7594 * memory. When they don't, some nodes will have more kernelcore than
7597 static void __init find_zone_movable_pfns_for_nodes(void)
7600 unsigned long usable_startpfn;
7601 unsigned long kernelcore_node, kernelcore_remaining;
7602 /* save the state before borrow the nodemask */
7603 nodemask_t saved_node_state = node_states[N_MEMORY];
7604 unsigned long totalpages = early_calculate_totalpages();
7605 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7606 struct memblock_region *r;
7608 /* Need to find movable_zone earlier when movable_node is specified. */
7609 find_usable_zone_for_movable();
7612 * If movable_node is specified, ignore kernelcore and movablecore
7615 if (movable_node_is_enabled()) {
7616 for_each_mem_region(r) {
7617 if (!memblock_is_hotpluggable(r))
7620 nid = memblock_get_region_node(r);
7622 usable_startpfn = PFN_DOWN(r->base);
7623 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7624 min(usable_startpfn, zone_movable_pfn[nid]) :
7632 * If kernelcore=mirror is specified, ignore movablecore option
7634 if (mirrored_kernelcore) {
7635 bool mem_below_4gb_not_mirrored = false;
7637 for_each_mem_region(r) {
7638 if (memblock_is_mirror(r))
7641 nid = memblock_get_region_node(r);
7643 usable_startpfn = memblock_region_memory_base_pfn(r);
7645 if (usable_startpfn < 0x100000) {
7646 mem_below_4gb_not_mirrored = true;
7650 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7651 min(usable_startpfn, zone_movable_pfn[nid]) :
7655 if (mem_below_4gb_not_mirrored)
7656 pr_warn("This configuration results in unmirrored kernel memory.\n");
7662 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7663 * amount of necessary memory.
7665 if (required_kernelcore_percent)
7666 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7668 if (required_movablecore_percent)
7669 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7673 * If movablecore= was specified, calculate what size of
7674 * kernelcore that corresponds so that memory usable for
7675 * any allocation type is evenly spread. If both kernelcore
7676 * and movablecore are specified, then the value of kernelcore
7677 * will be used for required_kernelcore if it's greater than
7678 * what movablecore would have allowed.
7680 if (required_movablecore) {
7681 unsigned long corepages;
7684 * Round-up so that ZONE_MOVABLE is at least as large as what
7685 * was requested by the user
7687 required_movablecore =
7688 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7689 required_movablecore = min(totalpages, required_movablecore);
7690 corepages = totalpages - required_movablecore;
7692 required_kernelcore = max(required_kernelcore, corepages);
7696 * If kernelcore was not specified or kernelcore size is larger
7697 * than totalpages, there is no ZONE_MOVABLE.
7699 if (!required_kernelcore || required_kernelcore >= totalpages)
7702 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7703 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7706 /* Spread kernelcore memory as evenly as possible throughout nodes */
7707 kernelcore_node = required_kernelcore / usable_nodes;
7708 for_each_node_state(nid, N_MEMORY) {
7709 unsigned long start_pfn, end_pfn;
7712 * Recalculate kernelcore_node if the division per node
7713 * now exceeds what is necessary to satisfy the requested
7714 * amount of memory for the kernel
7716 if (required_kernelcore < kernelcore_node)
7717 kernelcore_node = required_kernelcore / usable_nodes;
7720 * As the map is walked, we track how much memory is usable
7721 * by the kernel using kernelcore_remaining. When it is
7722 * 0, the rest of the node is usable by ZONE_MOVABLE
7724 kernelcore_remaining = kernelcore_node;
7726 /* Go through each range of PFNs within this node */
7727 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7728 unsigned long size_pages;
7730 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7731 if (start_pfn >= end_pfn)
7734 /* Account for what is only usable for kernelcore */
7735 if (start_pfn < usable_startpfn) {
7736 unsigned long kernel_pages;
7737 kernel_pages = min(end_pfn, usable_startpfn)
7740 kernelcore_remaining -= min(kernel_pages,
7741 kernelcore_remaining);
7742 required_kernelcore -= min(kernel_pages,
7743 required_kernelcore);
7745 /* Continue if range is now fully accounted */
7746 if (end_pfn <= usable_startpfn) {
7749 * Push zone_movable_pfn to the end so
7750 * that if we have to rebalance
7751 * kernelcore across nodes, we will
7752 * not double account here
7754 zone_movable_pfn[nid] = end_pfn;
7757 start_pfn = usable_startpfn;
7761 * The usable PFN range for ZONE_MOVABLE is from
7762 * start_pfn->end_pfn. Calculate size_pages as the
7763 * number of pages used as kernelcore
7765 size_pages = end_pfn - start_pfn;
7766 if (size_pages > kernelcore_remaining)
7767 size_pages = kernelcore_remaining;
7768 zone_movable_pfn[nid] = start_pfn + size_pages;
7771 * Some kernelcore has been met, update counts and
7772 * break if the kernelcore for this node has been
7775 required_kernelcore -= min(required_kernelcore,
7777 kernelcore_remaining -= size_pages;
7778 if (!kernelcore_remaining)
7784 * If there is still required_kernelcore, we do another pass with one
7785 * less node in the count. This will push zone_movable_pfn[nid] further
7786 * along on the nodes that still have memory until kernelcore is
7790 if (usable_nodes && required_kernelcore > usable_nodes)
7794 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7795 for (nid = 0; nid < MAX_NUMNODES; nid++)
7796 zone_movable_pfn[nid] =
7797 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7800 /* restore the node_state */
7801 node_states[N_MEMORY] = saved_node_state;
7804 /* Any regular or high memory on that node ? */
7805 static void check_for_memory(pg_data_t *pgdat, int nid)
7807 enum zone_type zone_type;
7809 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7810 struct zone *zone = &pgdat->node_zones[zone_type];
7811 if (populated_zone(zone)) {
7812 if (IS_ENABLED(CONFIG_HIGHMEM))
7813 node_set_state(nid, N_HIGH_MEMORY);
7814 if (zone_type <= ZONE_NORMAL)
7815 node_set_state(nid, N_NORMAL_MEMORY);
7822 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7823 * such cases we allow max_zone_pfn sorted in the descending order
7825 bool __weak arch_has_descending_max_zone_pfns(void)
7831 * free_area_init - Initialise all pg_data_t and zone data
7832 * @max_zone_pfn: an array of max PFNs for each zone
7834 * This will call free_area_init_node() for each active node in the system.
7835 * Using the page ranges provided by memblock_set_node(), the size of each
7836 * zone in each node and their holes is calculated. If the maximum PFN
7837 * between two adjacent zones match, it is assumed that the zone is empty.
7838 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7839 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7840 * starts where the previous one ended. For example, ZONE_DMA32 starts
7841 * at arch_max_dma_pfn.
7843 void __init free_area_init(unsigned long *max_zone_pfn)
7845 unsigned long start_pfn, end_pfn;
7849 /* Record where the zone boundaries are */
7850 memset(arch_zone_lowest_possible_pfn, 0,
7851 sizeof(arch_zone_lowest_possible_pfn));
7852 memset(arch_zone_highest_possible_pfn, 0,
7853 sizeof(arch_zone_highest_possible_pfn));
7855 start_pfn = find_min_pfn_with_active_regions();
7856 descending = arch_has_descending_max_zone_pfns();
7858 for (i = 0; i < MAX_NR_ZONES; i++) {
7860 zone = MAX_NR_ZONES - i - 1;
7864 if (zone == ZONE_MOVABLE)
7867 end_pfn = max(max_zone_pfn[zone], start_pfn);
7868 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7869 arch_zone_highest_possible_pfn[zone] = end_pfn;
7871 start_pfn = end_pfn;
7874 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7875 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7876 find_zone_movable_pfns_for_nodes();
7878 /* Print out the zone ranges */
7879 pr_info("Zone ranges:\n");
7880 for (i = 0; i < MAX_NR_ZONES; i++) {
7881 if (i == ZONE_MOVABLE)
7883 pr_info(" %-8s ", zone_names[i]);
7884 if (arch_zone_lowest_possible_pfn[i] ==
7885 arch_zone_highest_possible_pfn[i])
7888 pr_cont("[mem %#018Lx-%#018Lx]\n",
7889 (u64)arch_zone_lowest_possible_pfn[i]
7891 ((u64)arch_zone_highest_possible_pfn[i]
7892 << PAGE_SHIFT) - 1);
7895 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7896 pr_info("Movable zone start for each node\n");
7897 for (i = 0; i < MAX_NUMNODES; i++) {
7898 if (zone_movable_pfn[i])
7899 pr_info(" Node %d: %#018Lx\n", i,
7900 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7904 * Print out the early node map, and initialize the
7905 * subsection-map relative to active online memory ranges to
7906 * enable future "sub-section" extensions of the memory map.
7908 pr_info("Early memory node ranges\n");
7909 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7910 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7911 (u64)start_pfn << PAGE_SHIFT,
7912 ((u64)end_pfn << PAGE_SHIFT) - 1);
7913 subsection_map_init(start_pfn, end_pfn - start_pfn);
7916 /* Initialise every node */
7917 mminit_verify_pageflags_layout();
7918 setup_nr_node_ids();
7919 for_each_online_node(nid) {
7920 pg_data_t *pgdat = NODE_DATA(nid);
7921 free_area_init_node(nid);
7923 /* Any memory on that node */
7924 if (pgdat->node_present_pages)
7925 node_set_state(nid, N_MEMORY);
7926 check_for_memory(pgdat, nid);
7932 static int __init cmdline_parse_core(char *p, unsigned long *core,
7933 unsigned long *percent)
7935 unsigned long long coremem;
7941 /* Value may be a percentage of total memory, otherwise bytes */
7942 coremem = simple_strtoull(p, &endptr, 0);
7943 if (*endptr == '%') {
7944 /* Paranoid check for percent values greater than 100 */
7945 WARN_ON(coremem > 100);
7949 coremem = memparse(p, &p);
7950 /* Paranoid check that UL is enough for the coremem value */
7951 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7953 *core = coremem >> PAGE_SHIFT;
7960 * kernelcore=size sets the amount of memory for use for allocations that
7961 * cannot be reclaimed or migrated.
7963 static int __init cmdline_parse_kernelcore(char *p)
7965 /* parse kernelcore=mirror */
7966 if (parse_option_str(p, "mirror")) {
7967 mirrored_kernelcore = true;
7971 return cmdline_parse_core(p, &required_kernelcore,
7972 &required_kernelcore_percent);
7976 * movablecore=size sets the amount of memory for use for allocations that
7977 * can be reclaimed or migrated.
7979 static int __init cmdline_parse_movablecore(char *p)
7981 return cmdline_parse_core(p, &required_movablecore,
7982 &required_movablecore_percent);
7985 early_param("kernelcore", cmdline_parse_kernelcore);
7986 early_param("movablecore", cmdline_parse_movablecore);
7988 void adjust_managed_page_count(struct page *page, long count)
7990 atomic_long_add(count, &page_zone(page)->managed_pages);
7991 totalram_pages_add(count);
7992 #ifdef CONFIG_HIGHMEM
7993 if (PageHighMem(page))
7994 totalhigh_pages_add(count);
7997 EXPORT_SYMBOL(adjust_managed_page_count);
7999 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8002 unsigned long pages = 0;
8004 start = (void *)PAGE_ALIGN((unsigned long)start);
8005 end = (void *)((unsigned long)end & PAGE_MASK);
8006 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8007 struct page *page = virt_to_page(pos);
8008 void *direct_map_addr;
8011 * 'direct_map_addr' might be different from 'pos'
8012 * because some architectures' virt_to_page()
8013 * work with aliases. Getting the direct map
8014 * address ensures that we get a _writeable_
8015 * alias for the memset().
8017 direct_map_addr = page_address(page);
8019 * Perform a kasan-unchecked memset() since this memory
8020 * has not been initialized.
8022 direct_map_addr = kasan_reset_tag(direct_map_addr);
8023 if ((unsigned int)poison <= 0xFF)
8024 memset(direct_map_addr, poison, PAGE_SIZE);
8026 free_reserved_page(page);
8030 pr_info("Freeing %s memory: %ldK\n",
8031 s, pages << (PAGE_SHIFT - 10));
8036 void __init mem_init_print_info(void)
8038 unsigned long physpages, codesize, datasize, rosize, bss_size;
8039 unsigned long init_code_size, init_data_size;
8041 physpages = get_num_physpages();
8042 codesize = _etext - _stext;
8043 datasize = _edata - _sdata;
8044 rosize = __end_rodata - __start_rodata;
8045 bss_size = __bss_stop - __bss_start;
8046 init_data_size = __init_end - __init_begin;
8047 init_code_size = _einittext - _sinittext;
8050 * Detect special cases and adjust section sizes accordingly:
8051 * 1) .init.* may be embedded into .data sections
8052 * 2) .init.text.* may be out of [__init_begin, __init_end],
8053 * please refer to arch/tile/kernel/vmlinux.lds.S.
8054 * 3) .rodata.* may be embedded into .text or .data sections.
8056 #define adj_init_size(start, end, size, pos, adj) \
8058 if (start <= pos && pos < end && size > adj) \
8062 adj_init_size(__init_begin, __init_end, init_data_size,
8063 _sinittext, init_code_size);
8064 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8065 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8066 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8067 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8069 #undef adj_init_size
8071 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8072 #ifdef CONFIG_HIGHMEM
8076 nr_free_pages() << (PAGE_SHIFT - 10),
8077 physpages << (PAGE_SHIFT - 10),
8078 codesize >> 10, datasize >> 10, rosize >> 10,
8079 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8080 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
8081 totalcma_pages << (PAGE_SHIFT - 10)
8082 #ifdef CONFIG_HIGHMEM
8083 , totalhigh_pages() << (PAGE_SHIFT - 10)
8089 * set_dma_reserve - set the specified number of pages reserved in the first zone
8090 * @new_dma_reserve: The number of pages to mark reserved
8092 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8093 * In the DMA zone, a significant percentage may be consumed by kernel image
8094 * and other unfreeable allocations which can skew the watermarks badly. This
8095 * function may optionally be used to account for unfreeable pages in the
8096 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8097 * smaller per-cpu batchsize.
8099 void __init set_dma_reserve(unsigned long new_dma_reserve)
8101 dma_reserve = new_dma_reserve;
8104 static int page_alloc_cpu_dead(unsigned int cpu)
8108 lru_add_drain_cpu(cpu);
8112 * Spill the event counters of the dead processor
8113 * into the current processors event counters.
8114 * This artificially elevates the count of the current
8117 vm_events_fold_cpu(cpu);
8120 * Zero the differential counters of the dead processor
8121 * so that the vm statistics are consistent.
8123 * This is only okay since the processor is dead and cannot
8124 * race with what we are doing.
8126 cpu_vm_stats_fold(cpu);
8128 for_each_populated_zone(zone)
8129 zone_pcp_update(zone, 0);
8134 static int page_alloc_cpu_online(unsigned int cpu)
8138 for_each_populated_zone(zone)
8139 zone_pcp_update(zone, 1);
8144 int hashdist = HASHDIST_DEFAULT;
8146 static int __init set_hashdist(char *str)
8150 hashdist = simple_strtoul(str, &str, 0);
8153 __setup("hashdist=", set_hashdist);
8156 void __init page_alloc_init(void)
8161 if (num_node_state(N_MEMORY) == 1)
8165 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8166 "mm/page_alloc:pcp",
8167 page_alloc_cpu_online,
8168 page_alloc_cpu_dead);
8173 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8174 * or min_free_kbytes changes.
8176 static void calculate_totalreserve_pages(void)
8178 struct pglist_data *pgdat;
8179 unsigned long reserve_pages = 0;
8180 enum zone_type i, j;
8182 for_each_online_pgdat(pgdat) {
8184 pgdat->totalreserve_pages = 0;
8186 for (i = 0; i < MAX_NR_ZONES; i++) {
8187 struct zone *zone = pgdat->node_zones + i;
8189 unsigned long managed_pages = zone_managed_pages(zone);
8191 /* Find valid and maximum lowmem_reserve in the zone */
8192 for (j = i; j < MAX_NR_ZONES; j++) {
8193 if (zone->lowmem_reserve[j] > max)
8194 max = zone->lowmem_reserve[j];
8197 /* we treat the high watermark as reserved pages. */
8198 max += high_wmark_pages(zone);
8200 if (max > managed_pages)
8201 max = managed_pages;
8203 pgdat->totalreserve_pages += max;
8205 reserve_pages += max;
8208 totalreserve_pages = reserve_pages;
8212 * setup_per_zone_lowmem_reserve - called whenever
8213 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8214 * has a correct pages reserved value, so an adequate number of
8215 * pages are left in the zone after a successful __alloc_pages().
8217 static void setup_per_zone_lowmem_reserve(void)
8219 struct pglist_data *pgdat;
8220 enum zone_type i, j;
8222 for_each_online_pgdat(pgdat) {
8223 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8224 struct zone *zone = &pgdat->node_zones[i];
8225 int ratio = sysctl_lowmem_reserve_ratio[i];
8226 bool clear = !ratio || !zone_managed_pages(zone);
8227 unsigned long managed_pages = 0;
8229 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8230 struct zone *upper_zone = &pgdat->node_zones[j];
8232 managed_pages += zone_managed_pages(upper_zone);
8235 zone->lowmem_reserve[j] = 0;
8237 zone->lowmem_reserve[j] = managed_pages / ratio;
8242 /* update totalreserve_pages */
8243 calculate_totalreserve_pages();
8246 static void __setup_per_zone_wmarks(void)
8248 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8249 unsigned long lowmem_pages = 0;
8251 unsigned long flags;
8253 /* Calculate total number of !ZONE_HIGHMEM pages */
8254 for_each_zone(zone) {
8255 if (!is_highmem(zone))
8256 lowmem_pages += zone_managed_pages(zone);
8259 for_each_zone(zone) {
8262 spin_lock_irqsave(&zone->lock, flags);
8263 tmp = (u64)pages_min * zone_managed_pages(zone);
8264 do_div(tmp, lowmem_pages);
8265 if (is_highmem(zone)) {
8267 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8268 * need highmem pages, so cap pages_min to a small
8271 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8272 * deltas control async page reclaim, and so should
8273 * not be capped for highmem.
8275 unsigned long min_pages;
8277 min_pages = zone_managed_pages(zone) / 1024;
8278 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8279 zone->_watermark[WMARK_MIN] = min_pages;
8282 * If it's a lowmem zone, reserve a number of pages
8283 * proportionate to the zone's size.
8285 zone->_watermark[WMARK_MIN] = tmp;
8289 * Set the kswapd watermarks distance according to the
8290 * scale factor in proportion to available memory, but
8291 * ensure a minimum size on small systems.
8293 tmp = max_t(u64, tmp >> 2,
8294 mult_frac(zone_managed_pages(zone),
8295 watermark_scale_factor, 10000));
8297 zone->watermark_boost = 0;
8298 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8299 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8301 spin_unlock_irqrestore(&zone->lock, flags);
8304 /* update totalreserve_pages */
8305 calculate_totalreserve_pages();
8309 * setup_per_zone_wmarks - called when min_free_kbytes changes
8310 * or when memory is hot-{added|removed}
8312 * Ensures that the watermark[min,low,high] values for each zone are set
8313 * correctly with respect to min_free_kbytes.
8315 void setup_per_zone_wmarks(void)
8318 static DEFINE_SPINLOCK(lock);
8321 __setup_per_zone_wmarks();
8325 * The watermark size have changed so update the pcpu batch
8326 * and high limits or the limits may be inappropriate.
8329 zone_pcp_update(zone, 0);
8333 * Initialise min_free_kbytes.
8335 * For small machines we want it small (128k min). For large machines
8336 * we want it large (256MB max). But it is not linear, because network
8337 * bandwidth does not increase linearly with machine size. We use
8339 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8340 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8356 int __meminit init_per_zone_wmark_min(void)
8358 unsigned long lowmem_kbytes;
8359 int new_min_free_kbytes;
8361 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8362 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8364 if (new_min_free_kbytes > user_min_free_kbytes) {
8365 min_free_kbytes = new_min_free_kbytes;
8366 if (min_free_kbytes < 128)
8367 min_free_kbytes = 128;
8368 if (min_free_kbytes > 262144)
8369 min_free_kbytes = 262144;
8371 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8372 new_min_free_kbytes, user_min_free_kbytes);
8374 setup_per_zone_wmarks();
8375 refresh_zone_stat_thresholds();
8376 setup_per_zone_lowmem_reserve();
8379 setup_min_unmapped_ratio();
8380 setup_min_slab_ratio();
8383 khugepaged_min_free_kbytes_update();
8387 postcore_initcall(init_per_zone_wmark_min)
8390 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8391 * that we can call two helper functions whenever min_free_kbytes
8394 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8395 void *buffer, size_t *length, loff_t *ppos)
8399 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8404 user_min_free_kbytes = min_free_kbytes;
8405 setup_per_zone_wmarks();
8410 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8411 void *buffer, size_t *length, loff_t *ppos)
8415 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8420 setup_per_zone_wmarks();
8426 static void setup_min_unmapped_ratio(void)
8431 for_each_online_pgdat(pgdat)
8432 pgdat->min_unmapped_pages = 0;
8435 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8436 sysctl_min_unmapped_ratio) / 100;
8440 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8441 void *buffer, size_t *length, loff_t *ppos)
8445 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8449 setup_min_unmapped_ratio();
8454 static void setup_min_slab_ratio(void)
8459 for_each_online_pgdat(pgdat)
8460 pgdat->min_slab_pages = 0;
8463 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8464 sysctl_min_slab_ratio) / 100;
8467 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8468 void *buffer, size_t *length, loff_t *ppos)
8472 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8476 setup_min_slab_ratio();
8483 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8484 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8485 * whenever sysctl_lowmem_reserve_ratio changes.
8487 * The reserve ratio obviously has absolutely no relation with the
8488 * minimum watermarks. The lowmem reserve ratio can only make sense
8489 * if in function of the boot time zone sizes.
8491 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8492 void *buffer, size_t *length, loff_t *ppos)
8496 proc_dointvec_minmax(table, write, buffer, length, ppos);
8498 for (i = 0; i < MAX_NR_ZONES; i++) {
8499 if (sysctl_lowmem_reserve_ratio[i] < 1)
8500 sysctl_lowmem_reserve_ratio[i] = 0;
8503 setup_per_zone_lowmem_reserve();
8508 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8509 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8510 * pagelist can have before it gets flushed back to buddy allocator.
8512 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8513 int write, void *buffer, size_t *length, loff_t *ppos)
8516 int old_percpu_pagelist_high_fraction;
8519 mutex_lock(&pcp_batch_high_lock);
8520 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8522 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8523 if (!write || ret < 0)
8526 /* Sanity checking to avoid pcp imbalance */
8527 if (percpu_pagelist_high_fraction &&
8528 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8529 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8535 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8538 for_each_populated_zone(zone)
8539 zone_set_pageset_high_and_batch(zone, 0);
8541 mutex_unlock(&pcp_batch_high_lock);
8545 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8547 * Returns the number of pages that arch has reserved but
8548 * is not known to alloc_large_system_hash().
8550 static unsigned long __init arch_reserved_kernel_pages(void)
8557 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8558 * machines. As memory size is increased the scale is also increased but at
8559 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8560 * quadruples the scale is increased by one, which means the size of hash table
8561 * only doubles, instead of quadrupling as well.
8562 * Because 32-bit systems cannot have large physical memory, where this scaling
8563 * makes sense, it is disabled on such platforms.
8565 #if __BITS_PER_LONG > 32
8566 #define ADAPT_SCALE_BASE (64ul << 30)
8567 #define ADAPT_SCALE_SHIFT 2
8568 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8572 * allocate a large system hash table from bootmem
8573 * - it is assumed that the hash table must contain an exact power-of-2
8574 * quantity of entries
8575 * - limit is the number of hash buckets, not the total allocation size
8577 void *__init alloc_large_system_hash(const char *tablename,
8578 unsigned long bucketsize,
8579 unsigned long numentries,
8582 unsigned int *_hash_shift,
8583 unsigned int *_hash_mask,
8584 unsigned long low_limit,
8585 unsigned long high_limit)
8587 unsigned long long max = high_limit;
8588 unsigned long log2qty, size;
8594 /* allow the kernel cmdline to have a say */
8596 /* round applicable memory size up to nearest megabyte */
8597 numentries = nr_kernel_pages;
8598 numentries -= arch_reserved_kernel_pages();
8600 /* It isn't necessary when PAGE_SIZE >= 1MB */
8601 if (PAGE_SHIFT < 20)
8602 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8604 #if __BITS_PER_LONG > 32
8606 unsigned long adapt;
8608 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8609 adapt <<= ADAPT_SCALE_SHIFT)
8614 /* limit to 1 bucket per 2^scale bytes of low memory */
8615 if (scale > PAGE_SHIFT)
8616 numentries >>= (scale - PAGE_SHIFT);
8618 numentries <<= (PAGE_SHIFT - scale);
8620 /* Make sure we've got at least a 0-order allocation.. */
8621 if (unlikely(flags & HASH_SMALL)) {
8622 /* Makes no sense without HASH_EARLY */
8623 WARN_ON(!(flags & HASH_EARLY));
8624 if (!(numentries >> *_hash_shift)) {
8625 numentries = 1UL << *_hash_shift;
8626 BUG_ON(!numentries);
8628 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8629 numentries = PAGE_SIZE / bucketsize;
8631 numentries = roundup_pow_of_two(numentries);
8633 /* limit allocation size to 1/16 total memory by default */
8635 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8636 do_div(max, bucketsize);
8638 max = min(max, 0x80000000ULL);
8640 if (numentries < low_limit)
8641 numentries = low_limit;
8642 if (numentries > max)
8645 log2qty = ilog2(numentries);
8647 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8650 size = bucketsize << log2qty;
8651 if (flags & HASH_EARLY) {
8652 if (flags & HASH_ZERO)
8653 table = memblock_alloc(size, SMP_CACHE_BYTES);
8655 table = memblock_alloc_raw(size,
8657 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8658 table = __vmalloc(size, gfp_flags);
8660 huge = is_vm_area_hugepages(table);
8663 * If bucketsize is not a power-of-two, we may free
8664 * some pages at the end of hash table which
8665 * alloc_pages_exact() automatically does
8667 table = alloc_pages_exact(size, gfp_flags);
8668 kmemleak_alloc(table, size, 1, gfp_flags);
8670 } while (!table && size > PAGE_SIZE && --log2qty);
8673 panic("Failed to allocate %s hash table\n", tablename);
8675 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8676 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8677 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8680 *_hash_shift = log2qty;
8682 *_hash_mask = (1 << log2qty) - 1;
8688 * This function checks whether pageblock includes unmovable pages or not.
8690 * PageLRU check without isolation or lru_lock could race so that
8691 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8692 * check without lock_page also may miss some movable non-lru pages at
8693 * race condition. So you can't expect this function should be exact.
8695 * Returns a page without holding a reference. If the caller wants to
8696 * dereference that page (e.g., dumping), it has to make sure that it
8697 * cannot get removed (e.g., via memory unplug) concurrently.
8700 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8701 int migratetype, int flags)
8703 unsigned long iter = 0;
8704 unsigned long pfn = page_to_pfn(page);
8705 unsigned long offset = pfn % pageblock_nr_pages;
8707 if (is_migrate_cma_page(page)) {
8709 * CMA allocations (alloc_contig_range) really need to mark
8710 * isolate CMA pageblocks even when they are not movable in fact
8711 * so consider them movable here.
8713 if (is_migrate_cma(migratetype))
8719 for (; iter < pageblock_nr_pages - offset; iter++) {
8720 if (!pfn_valid_within(pfn + iter))
8723 page = pfn_to_page(pfn + iter);
8726 * Both, bootmem allocations and memory holes are marked
8727 * PG_reserved and are unmovable. We can even have unmovable
8728 * allocations inside ZONE_MOVABLE, for example when
8729 * specifying "movablecore".
8731 if (PageReserved(page))
8735 * If the zone is movable and we have ruled out all reserved
8736 * pages then it should be reasonably safe to assume the rest
8739 if (zone_idx(zone) == ZONE_MOVABLE)
8743 * Hugepages are not in LRU lists, but they're movable.
8744 * THPs are on the LRU, but need to be counted as #small pages.
8745 * We need not scan over tail pages because we don't
8746 * handle each tail page individually in migration.
8748 if (PageHuge(page) || PageTransCompound(page)) {
8749 struct page *head = compound_head(page);
8750 unsigned int skip_pages;
8752 if (PageHuge(page)) {
8753 if (!hugepage_migration_supported(page_hstate(head)))
8755 } else if (!PageLRU(head) && !__PageMovable(head)) {
8759 skip_pages = compound_nr(head) - (page - head);
8760 iter += skip_pages - 1;
8765 * We can't use page_count without pin a page
8766 * because another CPU can free compound page.
8767 * This check already skips compound tails of THP
8768 * because their page->_refcount is zero at all time.
8770 if (!page_ref_count(page)) {
8771 if (PageBuddy(page))
8772 iter += (1 << buddy_order(page)) - 1;
8777 * The HWPoisoned page may be not in buddy system, and
8778 * page_count() is not 0.
8780 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8784 * We treat all PageOffline() pages as movable when offlining
8785 * to give drivers a chance to decrement their reference count
8786 * in MEM_GOING_OFFLINE in order to indicate that these pages
8787 * can be offlined as there are no direct references anymore.
8788 * For actually unmovable PageOffline() where the driver does
8789 * not support this, we will fail later when trying to actually
8790 * move these pages that still have a reference count > 0.
8791 * (false negatives in this function only)
8793 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8796 if (__PageMovable(page) || PageLRU(page))
8800 * If there are RECLAIMABLE pages, we need to check
8801 * it. But now, memory offline itself doesn't call
8802 * shrink_node_slabs() and it still to be fixed.
8809 #ifdef CONFIG_CONTIG_ALLOC
8810 static unsigned long pfn_max_align_down(unsigned long pfn)
8812 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8813 pageblock_nr_pages) - 1);
8816 static unsigned long pfn_max_align_up(unsigned long pfn)
8818 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8819 pageblock_nr_pages));
8822 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8823 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8824 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8825 static void alloc_contig_dump_pages(struct list_head *page_list)
8827 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8829 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8833 list_for_each_entry(page, page_list, lru)
8834 dump_page(page, "migration failure");
8838 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8843 /* [start, end) must belong to a single zone. */
8844 static int __alloc_contig_migrate_range(struct compact_control *cc,
8845 unsigned long start, unsigned long end)
8847 /* This function is based on compact_zone() from compaction.c. */
8848 unsigned int nr_reclaimed;
8849 unsigned long pfn = start;
8850 unsigned int tries = 0;
8852 struct migration_target_control mtc = {
8853 .nid = zone_to_nid(cc->zone),
8854 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8857 lru_cache_disable();
8859 while (pfn < end || !list_empty(&cc->migratepages)) {
8860 if (fatal_signal_pending(current)) {
8865 if (list_empty(&cc->migratepages)) {
8866 cc->nr_migratepages = 0;
8867 ret = isolate_migratepages_range(cc, pfn, end);
8868 if (ret && ret != -EAGAIN)
8870 pfn = cc->migrate_pfn;
8872 } else if (++tries == 5) {
8877 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8879 cc->nr_migratepages -= nr_reclaimed;
8881 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8882 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8885 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
8886 * to retry again over this error, so do the same here.
8895 alloc_contig_dump_pages(&cc->migratepages);
8896 putback_movable_pages(&cc->migratepages);
8903 * alloc_contig_range() -- tries to allocate given range of pages
8904 * @start: start PFN to allocate
8905 * @end: one-past-the-last PFN to allocate
8906 * @migratetype: migratetype of the underlying pageblocks (either
8907 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8908 * in range must have the same migratetype and it must
8909 * be either of the two.
8910 * @gfp_mask: GFP mask to use during compaction
8912 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8913 * aligned. The PFN range must belong to a single zone.
8915 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8916 * pageblocks in the range. Once isolated, the pageblocks should not
8917 * be modified by others.
8919 * Return: zero on success or negative error code. On success all
8920 * pages which PFN is in [start, end) are allocated for the caller and
8921 * need to be freed with free_contig_range().
8923 int alloc_contig_range(unsigned long start, unsigned long end,
8924 unsigned migratetype, gfp_t gfp_mask)
8926 unsigned long outer_start, outer_end;
8930 struct compact_control cc = {
8931 .nr_migratepages = 0,
8933 .zone = page_zone(pfn_to_page(start)),
8934 .mode = MIGRATE_SYNC,
8935 .ignore_skip_hint = true,
8936 .no_set_skip_hint = true,
8937 .gfp_mask = current_gfp_context(gfp_mask),
8938 .alloc_contig = true,
8940 INIT_LIST_HEAD(&cc.migratepages);
8943 * What we do here is we mark all pageblocks in range as
8944 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8945 * have different sizes, and due to the way page allocator
8946 * work, we align the range to biggest of the two pages so
8947 * that page allocator won't try to merge buddies from
8948 * different pageblocks and change MIGRATE_ISOLATE to some
8949 * other migration type.
8951 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8952 * migrate the pages from an unaligned range (ie. pages that
8953 * we are interested in). This will put all the pages in
8954 * range back to page allocator as MIGRATE_ISOLATE.
8956 * When this is done, we take the pages in range from page
8957 * allocator removing them from the buddy system. This way
8958 * page allocator will never consider using them.
8960 * This lets us mark the pageblocks back as
8961 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8962 * aligned range but not in the unaligned, original range are
8963 * put back to page allocator so that buddy can use them.
8966 ret = start_isolate_page_range(pfn_max_align_down(start),
8967 pfn_max_align_up(end), migratetype, 0);
8971 drain_all_pages(cc.zone);
8974 * In case of -EBUSY, we'd like to know which page causes problem.
8975 * So, just fall through. test_pages_isolated() has a tracepoint
8976 * which will report the busy page.
8978 * It is possible that busy pages could become available before
8979 * the call to test_pages_isolated, and the range will actually be
8980 * allocated. So, if we fall through be sure to clear ret so that
8981 * -EBUSY is not accidentally used or returned to caller.
8983 ret = __alloc_contig_migrate_range(&cc, start, end);
8984 if (ret && ret != -EBUSY)
8989 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8990 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8991 * more, all pages in [start, end) are free in page allocator.
8992 * What we are going to do is to allocate all pages from
8993 * [start, end) (that is remove them from page allocator).
8995 * The only problem is that pages at the beginning and at the
8996 * end of interesting range may be not aligned with pages that
8997 * page allocator holds, ie. they can be part of higher order
8998 * pages. Because of this, we reserve the bigger range and
8999 * once this is done free the pages we are not interested in.
9001 * We don't have to hold zone->lock here because the pages are
9002 * isolated thus they won't get removed from buddy.
9006 outer_start = start;
9007 while (!PageBuddy(pfn_to_page(outer_start))) {
9008 if (++order >= MAX_ORDER) {
9009 outer_start = start;
9012 outer_start &= ~0UL << order;
9015 if (outer_start != start) {
9016 order = buddy_order(pfn_to_page(outer_start));
9019 * outer_start page could be small order buddy page and
9020 * it doesn't include start page. Adjust outer_start
9021 * in this case to report failed page properly
9022 * on tracepoint in test_pages_isolated()
9024 if (outer_start + (1UL << order) <= start)
9025 outer_start = start;
9028 /* Make sure the range is really isolated. */
9029 if (test_pages_isolated(outer_start, end, 0)) {
9034 /* Grab isolated pages from freelists. */
9035 outer_end = isolate_freepages_range(&cc, outer_start, end);
9041 /* Free head and tail (if any) */
9042 if (start != outer_start)
9043 free_contig_range(outer_start, start - outer_start);
9044 if (end != outer_end)
9045 free_contig_range(end, outer_end - end);
9048 undo_isolate_page_range(pfn_max_align_down(start),
9049 pfn_max_align_up(end), migratetype);
9052 EXPORT_SYMBOL(alloc_contig_range);
9054 static int __alloc_contig_pages(unsigned long start_pfn,
9055 unsigned long nr_pages, gfp_t gfp_mask)
9057 unsigned long end_pfn = start_pfn + nr_pages;
9059 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9063 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9064 unsigned long nr_pages)
9066 unsigned long i, end_pfn = start_pfn + nr_pages;
9069 for (i = start_pfn; i < end_pfn; i++) {
9070 page = pfn_to_online_page(i);
9074 if (page_zone(page) != z)
9077 if (PageReserved(page))
9083 static bool zone_spans_last_pfn(const struct zone *zone,
9084 unsigned long start_pfn, unsigned long nr_pages)
9086 unsigned long last_pfn = start_pfn + nr_pages - 1;
9088 return zone_spans_pfn(zone, last_pfn);
9092 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9093 * @nr_pages: Number of contiguous pages to allocate
9094 * @gfp_mask: GFP mask to limit search and used during compaction
9096 * @nodemask: Mask for other possible nodes
9098 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9099 * on an applicable zonelist to find a contiguous pfn range which can then be
9100 * tried for allocation with alloc_contig_range(). This routine is intended
9101 * for allocation requests which can not be fulfilled with the buddy allocator.
9103 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9104 * power of two then the alignment is guaranteed to be to the given nr_pages
9105 * (e.g. 1GB request would be aligned to 1GB).
9107 * Allocated pages can be freed with free_contig_range() or by manually calling
9108 * __free_page() on each allocated page.
9110 * Return: pointer to contiguous pages on success, or NULL if not successful.
9112 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9113 int nid, nodemask_t *nodemask)
9115 unsigned long ret, pfn, flags;
9116 struct zonelist *zonelist;
9120 zonelist = node_zonelist(nid, gfp_mask);
9121 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9122 gfp_zone(gfp_mask), nodemask) {
9123 spin_lock_irqsave(&zone->lock, flags);
9125 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9126 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9127 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9129 * We release the zone lock here because
9130 * alloc_contig_range() will also lock the zone
9131 * at some point. If there's an allocation
9132 * spinning on this lock, it may win the race
9133 * and cause alloc_contig_range() to fail...
9135 spin_unlock_irqrestore(&zone->lock, flags);
9136 ret = __alloc_contig_pages(pfn, nr_pages,
9139 return pfn_to_page(pfn);
9140 spin_lock_irqsave(&zone->lock, flags);
9144 spin_unlock_irqrestore(&zone->lock, flags);
9148 #endif /* CONFIG_CONTIG_ALLOC */
9150 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9152 unsigned long count = 0;
9154 for (; nr_pages--; pfn++) {
9155 struct page *page = pfn_to_page(pfn);
9157 count += page_count(page) != 1;
9160 WARN(count != 0, "%lu pages are still in use!\n", count);
9162 EXPORT_SYMBOL(free_contig_range);
9165 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9166 * page high values need to be recalculated.
9168 void zone_pcp_update(struct zone *zone, int cpu_online)
9170 mutex_lock(&pcp_batch_high_lock);
9171 zone_set_pageset_high_and_batch(zone, cpu_online);
9172 mutex_unlock(&pcp_batch_high_lock);
9176 * Effectively disable pcplists for the zone by setting the high limit to 0
9177 * and draining all cpus. A concurrent page freeing on another CPU that's about
9178 * to put the page on pcplist will either finish before the drain and the page
9179 * will be drained, or observe the new high limit and skip the pcplist.
9181 * Must be paired with a call to zone_pcp_enable().
9183 void zone_pcp_disable(struct zone *zone)
9185 mutex_lock(&pcp_batch_high_lock);
9186 __zone_set_pageset_high_and_batch(zone, 0, 1);
9187 __drain_all_pages(zone, true);
9190 void zone_pcp_enable(struct zone *zone)
9192 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9193 mutex_unlock(&pcp_batch_high_lock);
9196 void zone_pcp_reset(struct zone *zone)
9199 struct per_cpu_zonestat *pzstats;
9201 if (zone->per_cpu_pageset != &boot_pageset) {
9202 for_each_online_cpu(cpu) {
9203 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9204 drain_zonestat(zone, pzstats);
9206 free_percpu(zone->per_cpu_pageset);
9207 free_percpu(zone->per_cpu_zonestats);
9208 zone->per_cpu_pageset = &boot_pageset;
9209 zone->per_cpu_zonestats = &boot_zonestats;
9213 #ifdef CONFIG_MEMORY_HOTREMOVE
9215 * All pages in the range must be in a single zone, must not contain holes,
9216 * must span full sections, and must be isolated before calling this function.
9218 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9220 unsigned long pfn = start_pfn;
9224 unsigned long flags;
9226 offline_mem_sections(pfn, end_pfn);
9227 zone = page_zone(pfn_to_page(pfn));
9228 spin_lock_irqsave(&zone->lock, flags);
9229 while (pfn < end_pfn) {
9230 page = pfn_to_page(pfn);
9232 * The HWPoisoned page may be not in buddy system, and
9233 * page_count() is not 0.
9235 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9240 * At this point all remaining PageOffline() pages have a
9241 * reference count of 0 and can simply be skipped.
9243 if (PageOffline(page)) {
9244 BUG_ON(page_count(page));
9245 BUG_ON(PageBuddy(page));
9250 BUG_ON(page_count(page));
9251 BUG_ON(!PageBuddy(page));
9252 order = buddy_order(page);
9253 del_page_from_free_list(page, zone, order);
9254 pfn += (1 << order);
9256 spin_unlock_irqrestore(&zone->lock, flags);
9260 bool is_free_buddy_page(struct page *page)
9262 struct zone *zone = page_zone(page);
9263 unsigned long pfn = page_to_pfn(page);
9264 unsigned long flags;
9267 spin_lock_irqsave(&zone->lock, flags);
9268 for (order = 0; order < MAX_ORDER; order++) {
9269 struct page *page_head = page - (pfn & ((1 << order) - 1));
9271 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
9274 spin_unlock_irqrestore(&zone->lock, flags);
9276 return order < MAX_ORDER;
9279 #ifdef CONFIG_MEMORY_FAILURE
9281 * Break down a higher-order page in sub-pages, and keep our target out of
9284 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9285 struct page *target, int low, int high,
9288 unsigned long size = 1 << high;
9289 struct page *current_buddy, *next_page;
9291 while (high > low) {
9295 if (target >= &page[size]) {
9296 next_page = page + size;
9297 current_buddy = page;
9300 current_buddy = page + size;
9303 if (set_page_guard(zone, current_buddy, high, migratetype))
9306 if (current_buddy != target) {
9307 add_to_free_list(current_buddy, zone, high, migratetype);
9308 set_buddy_order(current_buddy, high);
9315 * Take a page that will be marked as poisoned off the buddy allocator.
9317 bool take_page_off_buddy(struct page *page)
9319 struct zone *zone = page_zone(page);
9320 unsigned long pfn = page_to_pfn(page);
9321 unsigned long flags;
9325 spin_lock_irqsave(&zone->lock, flags);
9326 for (order = 0; order < MAX_ORDER; order++) {
9327 struct page *page_head = page - (pfn & ((1 << order) - 1));
9328 int page_order = buddy_order(page_head);
9330 if (PageBuddy(page_head) && page_order >= order) {
9331 unsigned long pfn_head = page_to_pfn(page_head);
9332 int migratetype = get_pfnblock_migratetype(page_head,
9335 del_page_from_free_list(page_head, zone, page_order);
9336 break_down_buddy_pages(zone, page_head, page, 0,
9337 page_order, migratetype);
9338 if (!is_migrate_isolate(migratetype))
9339 __mod_zone_freepage_state(zone, -1, migratetype);
9343 if (page_count(page_head) > 0)
9346 spin_unlock_irqrestore(&zone->lock, flags);