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)
128 static DEFINE_PER_CPU(struct pagesets, pagesets) = {
129 .lock = INIT_LOCAL_LOCK(lock),
132 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
133 DEFINE_PER_CPU(int, numa_node);
134 EXPORT_PER_CPU_SYMBOL(numa_node);
137 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
139 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
141 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
142 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
143 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
144 * defined in <linux/topology.h>.
146 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
147 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
150 /* work_structs for global per-cpu drains */
153 struct work_struct work;
155 static DEFINE_MUTEX(pcpu_drain_mutex);
156 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
158 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
159 volatile unsigned long latent_entropy __latent_entropy;
160 EXPORT_SYMBOL(latent_entropy);
164 * Array of node states.
166 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
167 [N_POSSIBLE] = NODE_MASK_ALL,
168 [N_ONLINE] = { { [0] = 1UL } },
170 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
171 #ifdef CONFIG_HIGHMEM
172 [N_HIGH_MEMORY] = { { [0] = 1UL } },
174 [N_MEMORY] = { { [0] = 1UL } },
175 [N_CPU] = { { [0] = 1UL } },
178 EXPORT_SYMBOL(node_states);
180 atomic_long_t _totalram_pages __read_mostly;
181 EXPORT_SYMBOL(_totalram_pages);
182 unsigned long totalreserve_pages __read_mostly;
183 unsigned long totalcma_pages __read_mostly;
185 int percpu_pagelist_high_fraction;
186 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
187 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
188 EXPORT_SYMBOL(init_on_alloc);
190 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
191 EXPORT_SYMBOL(init_on_free);
193 static bool _init_on_alloc_enabled_early __read_mostly
194 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
195 static int __init early_init_on_alloc(char *buf)
198 return kstrtobool(buf, &_init_on_alloc_enabled_early);
200 early_param("init_on_alloc", early_init_on_alloc);
202 static bool _init_on_free_enabled_early __read_mostly
203 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
204 static int __init early_init_on_free(char *buf)
206 return kstrtobool(buf, &_init_on_free_enabled_early);
208 early_param("init_on_free", early_init_on_free);
211 * A cached value of the page's pageblock's migratetype, used when the page is
212 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
213 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
214 * Also the migratetype set in the page does not necessarily match the pcplist
215 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
216 * other index - this ensures that it will be put on the correct CMA freelist.
218 static inline int get_pcppage_migratetype(struct page *page)
223 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
225 page->index = migratetype;
228 #ifdef CONFIG_PM_SLEEP
230 * The following functions are used by the suspend/hibernate code to temporarily
231 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
232 * while devices are suspended. To avoid races with the suspend/hibernate code,
233 * they should always be called with system_transition_mutex held
234 * (gfp_allowed_mask also should only be modified with system_transition_mutex
235 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
236 * with that modification).
239 static gfp_t saved_gfp_mask;
241 void pm_restore_gfp_mask(void)
243 WARN_ON(!mutex_is_locked(&system_transition_mutex));
244 if (saved_gfp_mask) {
245 gfp_allowed_mask = saved_gfp_mask;
250 void pm_restrict_gfp_mask(void)
252 WARN_ON(!mutex_is_locked(&system_transition_mutex));
253 WARN_ON(saved_gfp_mask);
254 saved_gfp_mask = gfp_allowed_mask;
255 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
258 bool pm_suspended_storage(void)
260 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
264 #endif /* CONFIG_PM_SLEEP */
266 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
267 unsigned int pageblock_order __read_mostly;
270 static void __free_pages_ok(struct page *page, unsigned int order,
274 * results with 256, 32 in the lowmem_reserve sysctl:
275 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
276 * 1G machine -> (16M dma, 784M normal, 224M high)
277 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
278 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
279 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
281 * TBD: should special case ZONE_DMA32 machines here - in those we normally
282 * don't need any ZONE_NORMAL reservation
284 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
285 #ifdef CONFIG_ZONE_DMA
288 #ifdef CONFIG_ZONE_DMA32
292 #ifdef CONFIG_HIGHMEM
298 static char * const zone_names[MAX_NR_ZONES] = {
299 #ifdef CONFIG_ZONE_DMA
302 #ifdef CONFIG_ZONE_DMA32
306 #ifdef CONFIG_HIGHMEM
310 #ifdef CONFIG_ZONE_DEVICE
315 const char * const migratetype_names[MIGRATE_TYPES] = {
323 #ifdef CONFIG_MEMORY_ISOLATION
328 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
329 [NULL_COMPOUND_DTOR] = NULL,
330 [COMPOUND_PAGE_DTOR] = free_compound_page,
331 #ifdef CONFIG_HUGETLB_PAGE
332 [HUGETLB_PAGE_DTOR] = free_huge_page,
334 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
335 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
339 int min_free_kbytes = 1024;
340 int user_min_free_kbytes = -1;
341 int watermark_boost_factor __read_mostly = 15000;
342 int watermark_scale_factor = 10;
344 static unsigned long nr_kernel_pages __initdata;
345 static unsigned long nr_all_pages __initdata;
346 static unsigned long dma_reserve __initdata;
348 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
349 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
350 static unsigned long required_kernelcore __initdata;
351 static unsigned long required_kernelcore_percent __initdata;
352 static unsigned long required_movablecore __initdata;
353 static unsigned long required_movablecore_percent __initdata;
354 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
355 static bool mirrored_kernelcore __meminitdata;
357 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
359 EXPORT_SYMBOL(movable_zone);
362 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
363 unsigned int nr_online_nodes __read_mostly = 1;
364 EXPORT_SYMBOL(nr_node_ids);
365 EXPORT_SYMBOL(nr_online_nodes);
368 int page_group_by_mobility_disabled __read_mostly;
370 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
372 * During boot we initialize deferred pages on-demand, as needed, but once
373 * page_alloc_init_late() has finished, the deferred pages are all initialized,
374 * and we can permanently disable that path.
376 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
379 * Calling kasan_poison_pages() only after deferred memory initialization
380 * has completed. Poisoning pages during deferred memory init will greatly
381 * lengthen the process and cause problem in large memory systems as the
382 * deferred pages initialization is done with interrupt disabled.
384 * Assuming that there will be no reference to those newly initialized
385 * pages before they are ever allocated, this should have no effect on
386 * KASAN memory tracking as the poison will be properly inserted at page
387 * allocation time. The only corner case is when pages are allocated by
388 * on-demand allocation and then freed again before the deferred pages
389 * initialization is done, but this is not likely to happen.
391 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
393 return static_branch_unlikely(&deferred_pages) ||
394 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
395 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
396 PageSkipKASanPoison(page);
399 /* Returns true if the struct page for the pfn is uninitialised */
400 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
402 int nid = early_pfn_to_nid(pfn);
404 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
411 * Returns true when the remaining initialisation should be deferred until
412 * later in the boot cycle when it can be parallelised.
414 static bool __meminit
415 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
417 static unsigned long prev_end_pfn, nr_initialised;
420 * prev_end_pfn static that contains the end of previous zone
421 * No need to protect because called very early in boot before smp_init.
423 if (prev_end_pfn != end_pfn) {
424 prev_end_pfn = end_pfn;
428 /* Always populate low zones for address-constrained allocations */
429 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
432 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
435 * We start only with one section of pages, more pages are added as
436 * needed until the rest of deferred pages are initialized.
439 if ((nr_initialised > PAGES_PER_SECTION) &&
440 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
441 NODE_DATA(nid)->first_deferred_pfn = pfn;
447 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
449 return (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
450 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
451 PageSkipKASanPoison(page);
454 static inline bool early_page_uninitialised(unsigned long pfn)
459 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
465 /* Return a pointer to the bitmap storing bits affecting a block of pages */
466 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
469 #ifdef CONFIG_SPARSEMEM
470 return section_to_usemap(__pfn_to_section(pfn));
472 return page_zone(page)->pageblock_flags;
473 #endif /* CONFIG_SPARSEMEM */
476 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
478 #ifdef CONFIG_SPARSEMEM
479 pfn &= (PAGES_PER_SECTION-1);
481 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
482 #endif /* CONFIG_SPARSEMEM */
483 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
486 static __always_inline
487 unsigned long __get_pfnblock_flags_mask(const struct page *page,
491 unsigned long *bitmap;
492 unsigned long bitidx, word_bitidx;
495 bitmap = get_pageblock_bitmap(page, pfn);
496 bitidx = pfn_to_bitidx(page, pfn);
497 word_bitidx = bitidx / BITS_PER_LONG;
498 bitidx &= (BITS_PER_LONG-1);
500 word = bitmap[word_bitidx];
501 return (word >> bitidx) & mask;
505 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
506 * @page: The page within the block of interest
507 * @pfn: The target page frame number
508 * @mask: mask of bits that the caller is interested in
510 * Return: pageblock_bits flags
512 unsigned long get_pfnblock_flags_mask(const struct page *page,
513 unsigned long pfn, unsigned long mask)
515 return __get_pfnblock_flags_mask(page, pfn, mask);
518 static __always_inline int get_pfnblock_migratetype(const struct page *page,
521 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
525 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
526 * @page: The page within the block of interest
527 * @flags: The flags to set
528 * @pfn: The target page frame number
529 * @mask: mask of bits that the caller is interested in
531 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
535 unsigned long *bitmap;
536 unsigned long bitidx, word_bitidx;
537 unsigned long old_word, word;
539 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
540 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
542 bitmap = get_pageblock_bitmap(page, pfn);
543 bitidx = pfn_to_bitidx(page, pfn);
544 word_bitidx = bitidx / BITS_PER_LONG;
545 bitidx &= (BITS_PER_LONG-1);
547 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
552 word = READ_ONCE(bitmap[word_bitidx]);
554 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
555 if (word == old_word)
561 void set_pageblock_migratetype(struct page *page, int migratetype)
563 if (unlikely(page_group_by_mobility_disabled &&
564 migratetype < MIGRATE_PCPTYPES))
565 migratetype = MIGRATE_UNMOVABLE;
567 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
568 page_to_pfn(page), MIGRATETYPE_MASK);
571 #ifdef CONFIG_DEBUG_VM
572 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
576 unsigned long pfn = page_to_pfn(page);
577 unsigned long sp, start_pfn;
580 seq = zone_span_seqbegin(zone);
581 start_pfn = zone->zone_start_pfn;
582 sp = zone->spanned_pages;
583 if (!zone_spans_pfn(zone, pfn))
585 } while (zone_span_seqretry(zone, seq));
588 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
589 pfn, zone_to_nid(zone), zone->name,
590 start_pfn, start_pfn + sp);
595 static int page_is_consistent(struct zone *zone, struct page *page)
597 if (!pfn_valid_within(page_to_pfn(page)))
599 if (zone != page_zone(page))
605 * Temporary debugging check for pages not lying within a given zone.
607 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
609 if (page_outside_zone_boundaries(zone, page))
611 if (!page_is_consistent(zone, page))
617 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
623 static void bad_page(struct page *page, const char *reason)
625 static unsigned long resume;
626 static unsigned long nr_shown;
627 static unsigned long nr_unshown;
630 * Allow a burst of 60 reports, then keep quiet for that minute;
631 * or allow a steady drip of one report per second.
633 if (nr_shown == 60) {
634 if (time_before(jiffies, resume)) {
640 "BUG: Bad page state: %lu messages suppressed\n",
647 resume = jiffies + 60 * HZ;
649 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
650 current->comm, page_to_pfn(page));
651 dump_page(page, reason);
656 /* Leave bad fields for debug, except PageBuddy could make trouble */
657 page_mapcount_reset(page); /* remove PageBuddy */
658 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
661 static inline unsigned int order_to_pindex(int migratetype, int order)
665 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
666 if (order > PAGE_ALLOC_COSTLY_ORDER) {
667 VM_BUG_ON(order != pageblock_order);
668 base = PAGE_ALLOC_COSTLY_ORDER + 1;
671 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
674 return (MIGRATE_PCPTYPES * base) + migratetype;
677 static inline int pindex_to_order(unsigned int pindex)
679 int order = pindex / MIGRATE_PCPTYPES;
681 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
682 if (order > PAGE_ALLOC_COSTLY_ORDER) {
683 order = pageblock_order;
684 VM_BUG_ON(order != pageblock_order);
687 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
693 static inline bool pcp_allowed_order(unsigned int order)
695 if (order <= PAGE_ALLOC_COSTLY_ORDER)
697 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
698 if (order == pageblock_order)
704 static inline void free_the_page(struct page *page, unsigned int order)
706 if (pcp_allowed_order(order)) /* Via pcp? */
707 free_unref_page(page, order);
709 __free_pages_ok(page, order, FPI_NONE);
713 * Higher-order pages are called "compound pages". They are structured thusly:
715 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
717 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
718 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
720 * The first tail page's ->compound_dtor holds the offset in array of compound
721 * page destructors. See compound_page_dtors.
723 * The first tail page's ->compound_order holds the order of allocation.
724 * This usage means that zero-order pages may not be compound.
727 void free_compound_page(struct page *page)
729 mem_cgroup_uncharge(page);
730 free_the_page(page, compound_order(page));
733 void prep_compound_page(struct page *page, unsigned int order)
736 int nr_pages = 1 << order;
739 for (i = 1; i < nr_pages; i++) {
740 struct page *p = page + i;
741 p->mapping = TAIL_MAPPING;
742 set_compound_head(p, page);
745 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
746 set_compound_order(page, order);
747 atomic_set(compound_mapcount_ptr(page), -1);
748 if (hpage_pincount_available(page))
749 atomic_set(compound_pincount_ptr(page), 0);
752 #ifdef CONFIG_DEBUG_PAGEALLOC
753 unsigned int _debug_guardpage_minorder;
755 bool _debug_pagealloc_enabled_early __read_mostly
756 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
757 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
758 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
759 EXPORT_SYMBOL(_debug_pagealloc_enabled);
761 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
763 static int __init early_debug_pagealloc(char *buf)
765 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
767 early_param("debug_pagealloc", early_debug_pagealloc);
769 static int __init debug_guardpage_minorder_setup(char *buf)
773 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
774 pr_err("Bad debug_guardpage_minorder value\n");
777 _debug_guardpage_minorder = res;
778 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
781 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
783 static inline bool set_page_guard(struct zone *zone, struct page *page,
784 unsigned int order, int migratetype)
786 if (!debug_guardpage_enabled())
789 if (order >= debug_guardpage_minorder())
792 __SetPageGuard(page);
793 INIT_LIST_HEAD(&page->lru);
794 set_page_private(page, order);
795 /* Guard pages are not available for any usage */
796 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
801 static inline void clear_page_guard(struct zone *zone, struct page *page,
802 unsigned int order, int migratetype)
804 if (!debug_guardpage_enabled())
807 __ClearPageGuard(page);
809 set_page_private(page, 0);
810 if (!is_migrate_isolate(migratetype))
811 __mod_zone_freepage_state(zone, (1 << order), migratetype);
814 static inline bool set_page_guard(struct zone *zone, struct page *page,
815 unsigned int order, int migratetype) { return false; }
816 static inline void clear_page_guard(struct zone *zone, struct page *page,
817 unsigned int order, int migratetype) {}
821 * Enable static keys related to various memory debugging and hardening options.
822 * Some override others, and depend on early params that are evaluated in the
823 * order of appearance. So we need to first gather the full picture of what was
824 * enabled, and then make decisions.
826 void init_mem_debugging_and_hardening(void)
828 bool page_poisoning_requested = false;
830 #ifdef CONFIG_PAGE_POISONING
832 * Page poisoning is debug page alloc for some arches. If
833 * either of those options are enabled, enable poisoning.
835 if (page_poisoning_enabled() ||
836 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
837 debug_pagealloc_enabled())) {
838 static_branch_enable(&_page_poisoning_enabled);
839 page_poisoning_requested = true;
843 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
844 page_poisoning_requested) {
845 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
846 "will take precedence over init_on_alloc and init_on_free\n");
847 _init_on_alloc_enabled_early = false;
848 _init_on_free_enabled_early = false;
851 if (_init_on_alloc_enabled_early)
852 static_branch_enable(&init_on_alloc);
854 static_branch_disable(&init_on_alloc);
856 if (_init_on_free_enabled_early)
857 static_branch_enable(&init_on_free);
859 static_branch_disable(&init_on_free);
861 #ifdef CONFIG_DEBUG_PAGEALLOC
862 if (!debug_pagealloc_enabled())
865 static_branch_enable(&_debug_pagealloc_enabled);
867 if (!debug_guardpage_minorder())
870 static_branch_enable(&_debug_guardpage_enabled);
874 static inline void set_buddy_order(struct page *page, unsigned int order)
876 set_page_private(page, order);
877 __SetPageBuddy(page);
881 * This function checks whether a page is free && is the buddy
882 * we can coalesce a page and its buddy if
883 * (a) the buddy is not in a hole (check before calling!) &&
884 * (b) the buddy is in the buddy system &&
885 * (c) a page and its buddy have the same order &&
886 * (d) a page and its buddy are in the same zone.
888 * For recording whether a page is in the buddy system, we set PageBuddy.
889 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
891 * For recording page's order, we use page_private(page).
893 static inline bool page_is_buddy(struct page *page, struct page *buddy,
896 if (!page_is_guard(buddy) && !PageBuddy(buddy))
899 if (buddy_order(buddy) != order)
903 * zone check is done late to avoid uselessly calculating
904 * zone/node ids for pages that could never merge.
906 if (page_zone_id(page) != page_zone_id(buddy))
909 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
914 #ifdef CONFIG_COMPACTION
915 static inline struct capture_control *task_capc(struct zone *zone)
917 struct capture_control *capc = current->capture_control;
919 return unlikely(capc) &&
920 !(current->flags & PF_KTHREAD) &&
922 capc->cc->zone == zone ? capc : NULL;
926 compaction_capture(struct capture_control *capc, struct page *page,
927 int order, int migratetype)
929 if (!capc || order != capc->cc->order)
932 /* Do not accidentally pollute CMA or isolated regions*/
933 if (is_migrate_cma(migratetype) ||
934 is_migrate_isolate(migratetype))
938 * Do not let lower order allocations pollute a movable pageblock.
939 * This might let an unmovable request use a reclaimable pageblock
940 * and vice-versa but no more than normal fallback logic which can
941 * have trouble finding a high-order free page.
943 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
951 static inline struct capture_control *task_capc(struct zone *zone)
957 compaction_capture(struct capture_control *capc, struct page *page,
958 int order, int migratetype)
962 #endif /* CONFIG_COMPACTION */
964 /* Used for pages not on another list */
965 static inline void add_to_free_list(struct page *page, struct zone *zone,
966 unsigned int order, int migratetype)
968 struct free_area *area = &zone->free_area[order];
970 list_add(&page->lru, &area->free_list[migratetype]);
974 /* Used for pages not on another list */
975 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
976 unsigned int order, int migratetype)
978 struct free_area *area = &zone->free_area[order];
980 list_add_tail(&page->lru, &area->free_list[migratetype]);
985 * Used for pages which are on another list. Move the pages to the tail
986 * of the list - so the moved pages won't immediately be considered for
987 * allocation again (e.g., optimization for memory onlining).
989 static inline void move_to_free_list(struct page *page, struct zone *zone,
990 unsigned int order, int migratetype)
992 struct free_area *area = &zone->free_area[order];
994 list_move_tail(&page->lru, &area->free_list[migratetype]);
997 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
1000 /* clear reported state and update reported page count */
1001 if (page_reported(page))
1002 __ClearPageReported(page);
1004 list_del(&page->lru);
1005 __ClearPageBuddy(page);
1006 set_page_private(page, 0);
1007 zone->free_area[order].nr_free--;
1011 * If this is not the largest possible page, check if the buddy
1012 * of the next-highest order is free. If it is, it's possible
1013 * that pages are being freed that will coalesce soon. In case,
1014 * that is happening, add the free page to the tail of the list
1015 * so it's less likely to be used soon and more likely to be merged
1016 * as a higher order page
1019 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1020 struct page *page, unsigned int order)
1022 struct page *higher_page, *higher_buddy;
1023 unsigned long combined_pfn;
1025 if (order >= MAX_ORDER - 2)
1028 if (!pfn_valid_within(buddy_pfn))
1031 combined_pfn = buddy_pfn & pfn;
1032 higher_page = page + (combined_pfn - pfn);
1033 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1034 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1036 return pfn_valid_within(buddy_pfn) &&
1037 page_is_buddy(higher_page, higher_buddy, order + 1);
1041 * Freeing function for a buddy system allocator.
1043 * The concept of a buddy system is to maintain direct-mapped table
1044 * (containing bit values) for memory blocks of various "orders".
1045 * The bottom level table contains the map for the smallest allocatable
1046 * units of memory (here, pages), and each level above it describes
1047 * pairs of units from the levels below, hence, "buddies".
1048 * At a high level, all that happens here is marking the table entry
1049 * at the bottom level available, and propagating the changes upward
1050 * as necessary, plus some accounting needed to play nicely with other
1051 * parts of the VM system.
1052 * At each level, we keep a list of pages, which are heads of continuous
1053 * free pages of length of (1 << order) and marked with PageBuddy.
1054 * Page's order is recorded in page_private(page) field.
1055 * So when we are allocating or freeing one, we can derive the state of the
1056 * other. That is, if we allocate a small block, and both were
1057 * free, the remainder of the region must be split into blocks.
1058 * If a block is freed, and its buddy is also free, then this
1059 * triggers coalescing into a block of larger size.
1064 static inline void __free_one_page(struct page *page,
1066 struct zone *zone, unsigned int order,
1067 int migratetype, fpi_t fpi_flags)
1069 struct capture_control *capc = task_capc(zone);
1070 unsigned long buddy_pfn;
1071 unsigned long combined_pfn;
1072 unsigned int max_order;
1076 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1078 VM_BUG_ON(!zone_is_initialized(zone));
1079 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1081 VM_BUG_ON(migratetype == -1);
1082 if (likely(!is_migrate_isolate(migratetype)))
1083 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1085 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1086 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1089 while (order < max_order) {
1090 if (compaction_capture(capc, page, order, migratetype)) {
1091 __mod_zone_freepage_state(zone, -(1 << order),
1095 buddy_pfn = __find_buddy_pfn(pfn, order);
1096 buddy = page + (buddy_pfn - pfn);
1098 if (!pfn_valid_within(buddy_pfn))
1100 if (!page_is_buddy(page, buddy, order))
1103 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1104 * merge with it and move up one order.
1106 if (page_is_guard(buddy))
1107 clear_page_guard(zone, buddy, order, migratetype);
1109 del_page_from_free_list(buddy, zone, order);
1110 combined_pfn = buddy_pfn & pfn;
1111 page = page + (combined_pfn - pfn);
1115 if (order < MAX_ORDER - 1) {
1116 /* If we are here, it means order is >= pageblock_order.
1117 * We want to prevent merge between freepages on isolate
1118 * pageblock and normal pageblock. Without this, pageblock
1119 * isolation could cause incorrect freepage or CMA accounting.
1121 * We don't want to hit this code for the more frequent
1122 * low-order merging.
1124 if (unlikely(has_isolate_pageblock(zone))) {
1127 buddy_pfn = __find_buddy_pfn(pfn, order);
1128 buddy = page + (buddy_pfn - pfn);
1129 buddy_mt = get_pageblock_migratetype(buddy);
1131 if (migratetype != buddy_mt
1132 && (is_migrate_isolate(migratetype) ||
1133 is_migrate_isolate(buddy_mt)))
1136 max_order = order + 1;
1137 goto continue_merging;
1141 set_buddy_order(page, order);
1143 if (fpi_flags & FPI_TO_TAIL)
1145 else if (is_shuffle_order(order))
1146 to_tail = shuffle_pick_tail();
1148 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1151 add_to_free_list_tail(page, zone, order, migratetype);
1153 add_to_free_list(page, zone, order, migratetype);
1155 /* Notify page reporting subsystem of freed page */
1156 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1157 page_reporting_notify_free(order);
1161 * A bad page could be due to a number of fields. Instead of multiple branches,
1162 * try and check multiple fields with one check. The caller must do a detailed
1163 * check if necessary.
1165 static inline bool page_expected_state(struct page *page,
1166 unsigned long check_flags)
1168 if (unlikely(atomic_read(&page->_mapcount) != -1))
1171 if (unlikely((unsigned long)page->mapping |
1172 page_ref_count(page) |
1176 (page->flags & check_flags)))
1182 static const char *page_bad_reason(struct page *page, unsigned long flags)
1184 const char *bad_reason = NULL;
1186 if (unlikely(atomic_read(&page->_mapcount) != -1))
1187 bad_reason = "nonzero mapcount";
1188 if (unlikely(page->mapping != NULL))
1189 bad_reason = "non-NULL mapping";
1190 if (unlikely(page_ref_count(page) != 0))
1191 bad_reason = "nonzero _refcount";
1192 if (unlikely(page->flags & flags)) {
1193 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1194 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1196 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1199 if (unlikely(page->memcg_data))
1200 bad_reason = "page still charged to cgroup";
1205 static void check_free_page_bad(struct page *page)
1208 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1211 static inline int check_free_page(struct page *page)
1213 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1216 /* Something has gone sideways, find it */
1217 check_free_page_bad(page);
1221 static int free_tail_pages_check(struct page *head_page, struct page *page)
1226 * We rely page->lru.next never has bit 0 set, unless the page
1227 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1229 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1231 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1235 switch (page - head_page) {
1237 /* the first tail page: ->mapping may be compound_mapcount() */
1238 if (unlikely(compound_mapcount(page))) {
1239 bad_page(page, "nonzero compound_mapcount");
1245 * the second tail page: ->mapping is
1246 * deferred_list.next -- ignore value.
1250 if (page->mapping != TAIL_MAPPING) {
1251 bad_page(page, "corrupted mapping in tail page");
1256 if (unlikely(!PageTail(page))) {
1257 bad_page(page, "PageTail not set");
1260 if (unlikely(compound_head(page) != head_page)) {
1261 bad_page(page, "compound_head not consistent");
1266 page->mapping = NULL;
1267 clear_compound_head(page);
1271 static void kernel_init_free_pages(struct page *page, int numpages, bool zero_tags)
1276 for (i = 0; i < numpages; i++)
1277 tag_clear_highpage(page + i);
1281 /* s390's use of memset() could override KASAN redzones. */
1282 kasan_disable_current();
1283 for (i = 0; i < numpages; i++) {
1284 u8 tag = page_kasan_tag(page + i);
1285 page_kasan_tag_reset(page + i);
1286 clear_highpage(page + i);
1287 page_kasan_tag_set(page + i, tag);
1289 kasan_enable_current();
1292 static __always_inline bool free_pages_prepare(struct page *page,
1293 unsigned int order, bool check_free, fpi_t fpi_flags)
1296 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1298 VM_BUG_ON_PAGE(PageTail(page), page);
1300 trace_mm_page_free(page, order);
1302 if (unlikely(PageHWPoison(page)) && !order) {
1304 * Do not let hwpoison pages hit pcplists/buddy
1305 * Untie memcg state and reset page's owner
1307 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1308 __memcg_kmem_uncharge_page(page, order);
1309 reset_page_owner(page, order);
1314 * Check tail pages before head page information is cleared to
1315 * avoid checking PageCompound for order-0 pages.
1317 if (unlikely(order)) {
1318 bool compound = PageCompound(page);
1321 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1324 ClearPageDoubleMap(page);
1325 for (i = 1; i < (1 << order); i++) {
1327 bad += free_tail_pages_check(page, page + i);
1328 if (unlikely(check_free_page(page + i))) {
1332 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1335 if (PageMappingFlags(page))
1336 page->mapping = NULL;
1337 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1338 __memcg_kmem_uncharge_page(page, order);
1340 bad += check_free_page(page);
1344 page_cpupid_reset_last(page);
1345 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1346 reset_page_owner(page, order);
1348 if (!PageHighMem(page)) {
1349 debug_check_no_locks_freed(page_address(page),
1350 PAGE_SIZE << order);
1351 debug_check_no_obj_freed(page_address(page),
1352 PAGE_SIZE << order);
1355 kernel_poison_pages(page, 1 << order);
1358 * As memory initialization might be integrated into KASAN,
1359 * kasan_free_pages and kernel_init_free_pages must be
1360 * kept together to avoid discrepancies in behavior.
1362 * With hardware tag-based KASAN, memory tags must be set before the
1363 * page becomes unavailable via debug_pagealloc or arch_free_page.
1365 if (kasan_has_integrated_init()) {
1366 if (!skip_kasan_poison)
1367 kasan_free_pages(page, order);
1369 bool init = want_init_on_free();
1372 kernel_init_free_pages(page, 1 << order, false);
1373 if (!skip_kasan_poison)
1374 kasan_poison_pages(page, order, init);
1378 * arch_free_page() can make the page's contents inaccessible. s390
1379 * does this. So nothing which can access the page's contents should
1380 * happen after this.
1382 arch_free_page(page, order);
1384 debug_pagealloc_unmap_pages(page, 1 << order);
1389 #ifdef CONFIG_DEBUG_VM
1391 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1392 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1393 * moved from pcp lists to free lists.
1395 static bool free_pcp_prepare(struct page *page, unsigned int order)
1397 return free_pages_prepare(page, order, true, FPI_NONE);
1400 static bool bulkfree_pcp_prepare(struct page *page)
1402 if (debug_pagealloc_enabled_static())
1403 return check_free_page(page);
1409 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1410 * moving from pcp lists to free list in order to reduce overhead. With
1411 * debug_pagealloc enabled, they are checked also immediately when being freed
1414 static bool free_pcp_prepare(struct page *page, unsigned int order)
1416 if (debug_pagealloc_enabled_static())
1417 return free_pages_prepare(page, order, true, FPI_NONE);
1419 return free_pages_prepare(page, order, false, FPI_NONE);
1422 static bool bulkfree_pcp_prepare(struct page *page)
1424 return check_free_page(page);
1426 #endif /* CONFIG_DEBUG_VM */
1428 static inline void prefetch_buddy(struct page *page)
1430 unsigned long pfn = page_to_pfn(page);
1431 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1432 struct page *buddy = page + (buddy_pfn - pfn);
1438 * Frees a number of pages from the PCP lists
1439 * Assumes all pages on list are in same zone, and of same order.
1440 * count is the number of pages to free.
1442 * If the zone was previously in an "all pages pinned" state then look to
1443 * see if this freeing clears that state.
1445 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1446 * pinned" detection logic.
1448 static void free_pcppages_bulk(struct zone *zone, int count,
1449 struct per_cpu_pages *pcp)
1455 int prefetch_nr = READ_ONCE(pcp->batch);
1456 bool isolated_pageblocks;
1457 struct page *page, *tmp;
1461 * Ensure proper count is passed which otherwise would stuck in the
1462 * below while (list_empty(list)) loop.
1464 count = min(pcp->count, count);
1466 struct list_head *list;
1469 * Remove pages from lists in a round-robin fashion. A
1470 * batch_free count is maintained that is incremented when an
1471 * empty list is encountered. This is so more pages are freed
1472 * off fuller lists instead of spinning excessively around empty
1477 if (++pindex == NR_PCP_LISTS)
1479 list = &pcp->lists[pindex];
1480 } while (list_empty(list));
1482 /* This is the only non-empty list. Free them all. */
1483 if (batch_free == NR_PCP_LISTS)
1486 order = pindex_to_order(pindex);
1487 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1489 page = list_last_entry(list, struct page, lru);
1490 /* must delete to avoid corrupting pcp list */
1491 list_del(&page->lru);
1492 nr_freed += 1 << order;
1493 count -= 1 << order;
1495 if (bulkfree_pcp_prepare(page))
1498 /* Encode order with the migratetype */
1499 page->index <<= NR_PCP_ORDER_WIDTH;
1500 page->index |= order;
1502 list_add_tail(&page->lru, &head);
1505 * We are going to put the page back to the global
1506 * pool, prefetch its buddy to speed up later access
1507 * under zone->lock. It is believed the overhead of
1508 * an additional test and calculating buddy_pfn here
1509 * can be offset by reduced memory latency later. To
1510 * avoid excessive prefetching due to large count, only
1511 * prefetch buddy for the first pcp->batch nr of pages.
1514 prefetch_buddy(page);
1517 } while (count > 0 && --batch_free && !list_empty(list));
1519 pcp->count -= nr_freed;
1522 * local_lock_irq held so equivalent to spin_lock_irqsave for
1523 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1525 spin_lock(&zone->lock);
1526 isolated_pageblocks = has_isolate_pageblock(zone);
1529 * Use safe version since after __free_one_page(),
1530 * page->lru.next will not point to original list.
1532 list_for_each_entry_safe(page, tmp, &head, lru) {
1533 int mt = get_pcppage_migratetype(page);
1535 /* mt has been encoded with the order (see above) */
1536 order = mt & NR_PCP_ORDER_MASK;
1537 mt >>= NR_PCP_ORDER_WIDTH;
1539 /* MIGRATE_ISOLATE page should not go to pcplists */
1540 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1541 /* Pageblock could have been isolated meanwhile */
1542 if (unlikely(isolated_pageblocks))
1543 mt = get_pageblock_migratetype(page);
1545 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1546 trace_mm_page_pcpu_drain(page, order, mt);
1548 spin_unlock(&zone->lock);
1551 static void free_one_page(struct zone *zone,
1552 struct page *page, unsigned long pfn,
1554 int migratetype, fpi_t fpi_flags)
1556 unsigned long flags;
1558 spin_lock_irqsave(&zone->lock, flags);
1559 if (unlikely(has_isolate_pageblock(zone) ||
1560 is_migrate_isolate(migratetype))) {
1561 migratetype = get_pfnblock_migratetype(page, pfn);
1563 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1564 spin_unlock_irqrestore(&zone->lock, flags);
1567 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1568 unsigned long zone, int nid)
1570 mm_zero_struct_page(page);
1571 set_page_links(page, zone, nid, pfn);
1572 init_page_count(page);
1573 page_mapcount_reset(page);
1574 page_cpupid_reset_last(page);
1575 page_kasan_tag_reset(page);
1577 INIT_LIST_HEAD(&page->lru);
1578 #ifdef WANT_PAGE_VIRTUAL
1579 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1580 if (!is_highmem_idx(zone))
1581 set_page_address(page, __va(pfn << PAGE_SHIFT));
1585 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1586 static void __meminit init_reserved_page(unsigned long pfn)
1591 if (!early_page_uninitialised(pfn))
1594 nid = early_pfn_to_nid(pfn);
1595 pgdat = NODE_DATA(nid);
1597 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1598 struct zone *zone = &pgdat->node_zones[zid];
1600 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1603 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1606 static inline void init_reserved_page(unsigned long pfn)
1609 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1612 * Initialised pages do not have PageReserved set. This function is
1613 * called for each range allocated by the bootmem allocator and
1614 * marks the pages PageReserved. The remaining valid pages are later
1615 * sent to the buddy page allocator.
1617 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1619 unsigned long start_pfn = PFN_DOWN(start);
1620 unsigned long end_pfn = PFN_UP(end);
1622 for (; start_pfn < end_pfn; start_pfn++) {
1623 if (pfn_valid(start_pfn)) {
1624 struct page *page = pfn_to_page(start_pfn);
1626 init_reserved_page(start_pfn);
1628 /* Avoid false-positive PageTail() */
1629 INIT_LIST_HEAD(&page->lru);
1632 * no need for atomic set_bit because the struct
1633 * page is not visible yet so nobody should
1636 __SetPageReserved(page);
1641 static void __free_pages_ok(struct page *page, unsigned int order,
1644 unsigned long flags;
1646 unsigned long pfn = page_to_pfn(page);
1647 struct zone *zone = page_zone(page);
1649 if (!free_pages_prepare(page, order, true, fpi_flags))
1652 migratetype = get_pfnblock_migratetype(page, pfn);
1654 spin_lock_irqsave(&zone->lock, flags);
1655 if (unlikely(has_isolate_pageblock(zone) ||
1656 is_migrate_isolate(migratetype))) {
1657 migratetype = get_pfnblock_migratetype(page, pfn);
1659 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1660 spin_unlock_irqrestore(&zone->lock, flags);
1662 __count_vm_events(PGFREE, 1 << order);
1665 void __free_pages_core(struct page *page, unsigned int order)
1667 unsigned int nr_pages = 1 << order;
1668 struct page *p = page;
1672 * When initializing the memmap, __init_single_page() sets the refcount
1673 * of all pages to 1 ("allocated"/"not free"). We have to set the
1674 * refcount of all involved pages to 0.
1677 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1679 __ClearPageReserved(p);
1680 set_page_count(p, 0);
1682 __ClearPageReserved(p);
1683 set_page_count(p, 0);
1685 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1688 * Bypass PCP and place fresh pages right to the tail, primarily
1689 * relevant for memory onlining.
1691 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1697 * During memory init memblocks map pfns to nids. The search is expensive and
1698 * this caches recent lookups. The implementation of __early_pfn_to_nid
1699 * treats start/end as pfns.
1701 struct mminit_pfnnid_cache {
1702 unsigned long last_start;
1703 unsigned long last_end;
1707 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1710 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1712 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1713 struct mminit_pfnnid_cache *state)
1715 unsigned long start_pfn, end_pfn;
1718 if (state->last_start <= pfn && pfn < state->last_end)
1719 return state->last_nid;
1721 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1722 if (nid != NUMA_NO_NODE) {
1723 state->last_start = start_pfn;
1724 state->last_end = end_pfn;
1725 state->last_nid = nid;
1731 int __meminit early_pfn_to_nid(unsigned long pfn)
1733 static DEFINE_SPINLOCK(early_pfn_lock);
1736 spin_lock(&early_pfn_lock);
1737 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1739 nid = first_online_node;
1740 spin_unlock(&early_pfn_lock);
1744 #endif /* CONFIG_NUMA */
1746 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1749 if (early_page_uninitialised(pfn))
1751 __free_pages_core(page, order);
1755 * Check that the whole (or subset of) a pageblock given by the interval of
1756 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1757 * with the migration of free compaction scanner. The scanners then need to
1758 * use only pfn_valid_within() check for arches that allow holes within
1761 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1763 * It's possible on some configurations to have a setup like node0 node1 node0
1764 * i.e. it's possible that all pages within a zones range of pages do not
1765 * belong to a single zone. We assume that a border between node0 and node1
1766 * can occur within a single pageblock, but not a node0 node1 node0
1767 * interleaving within a single pageblock. It is therefore sufficient to check
1768 * the first and last page of a pageblock and avoid checking each individual
1769 * page in a pageblock.
1771 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1772 unsigned long end_pfn, struct zone *zone)
1774 struct page *start_page;
1775 struct page *end_page;
1777 /* end_pfn is one past the range we are checking */
1780 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1783 start_page = pfn_to_online_page(start_pfn);
1787 if (page_zone(start_page) != zone)
1790 end_page = pfn_to_page(end_pfn);
1792 /* This gives a shorter code than deriving page_zone(end_page) */
1793 if (page_zone_id(start_page) != page_zone_id(end_page))
1799 void set_zone_contiguous(struct zone *zone)
1801 unsigned long block_start_pfn = zone->zone_start_pfn;
1802 unsigned long block_end_pfn;
1804 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1805 for (; block_start_pfn < zone_end_pfn(zone);
1806 block_start_pfn = block_end_pfn,
1807 block_end_pfn += pageblock_nr_pages) {
1809 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1811 if (!__pageblock_pfn_to_page(block_start_pfn,
1812 block_end_pfn, zone))
1817 /* We confirm that there is no hole */
1818 zone->contiguous = true;
1821 void clear_zone_contiguous(struct zone *zone)
1823 zone->contiguous = false;
1826 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1827 static void __init deferred_free_range(unsigned long pfn,
1828 unsigned long nr_pages)
1836 page = pfn_to_page(pfn);
1838 /* Free a large naturally-aligned chunk if possible */
1839 if (nr_pages == pageblock_nr_pages &&
1840 (pfn & (pageblock_nr_pages - 1)) == 0) {
1841 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1842 __free_pages_core(page, pageblock_order);
1846 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1847 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1848 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1849 __free_pages_core(page, 0);
1853 /* Completion tracking for deferred_init_memmap() threads */
1854 static atomic_t pgdat_init_n_undone __initdata;
1855 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1857 static inline void __init pgdat_init_report_one_done(void)
1859 if (atomic_dec_and_test(&pgdat_init_n_undone))
1860 complete(&pgdat_init_all_done_comp);
1864 * Returns true if page needs to be initialized or freed to buddy allocator.
1866 * First we check if pfn is valid on architectures where it is possible to have
1867 * holes within pageblock_nr_pages. On systems where it is not possible, this
1868 * function is optimized out.
1870 * Then, we check if a current large page is valid by only checking the validity
1873 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1875 if (!pfn_valid_within(pfn))
1877 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1883 * Free pages to buddy allocator. Try to free aligned pages in
1884 * pageblock_nr_pages sizes.
1886 static void __init deferred_free_pages(unsigned long pfn,
1887 unsigned long end_pfn)
1889 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1890 unsigned long nr_free = 0;
1892 for (; pfn < end_pfn; pfn++) {
1893 if (!deferred_pfn_valid(pfn)) {
1894 deferred_free_range(pfn - nr_free, nr_free);
1896 } else if (!(pfn & nr_pgmask)) {
1897 deferred_free_range(pfn - nr_free, nr_free);
1903 /* Free the last block of pages to allocator */
1904 deferred_free_range(pfn - nr_free, nr_free);
1908 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1909 * by performing it only once every pageblock_nr_pages.
1910 * Return number of pages initialized.
1912 static unsigned long __init deferred_init_pages(struct zone *zone,
1914 unsigned long end_pfn)
1916 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1917 int nid = zone_to_nid(zone);
1918 unsigned long nr_pages = 0;
1919 int zid = zone_idx(zone);
1920 struct page *page = NULL;
1922 for (; pfn < end_pfn; pfn++) {
1923 if (!deferred_pfn_valid(pfn)) {
1926 } else if (!page || !(pfn & nr_pgmask)) {
1927 page = pfn_to_page(pfn);
1931 __init_single_page(page, pfn, zid, nid);
1938 * This function is meant to pre-load the iterator for the zone init.
1939 * Specifically it walks through the ranges until we are caught up to the
1940 * first_init_pfn value and exits there. If we never encounter the value we
1941 * return false indicating there are no valid ranges left.
1944 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1945 unsigned long *spfn, unsigned long *epfn,
1946 unsigned long first_init_pfn)
1951 * Start out by walking through the ranges in this zone that have
1952 * already been initialized. We don't need to do anything with them
1953 * so we just need to flush them out of the system.
1955 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1956 if (*epfn <= first_init_pfn)
1958 if (*spfn < first_init_pfn)
1959 *spfn = first_init_pfn;
1968 * Initialize and free pages. We do it in two loops: first we initialize
1969 * struct page, then free to buddy allocator, because while we are
1970 * freeing pages we can access pages that are ahead (computing buddy
1971 * page in __free_one_page()).
1973 * In order to try and keep some memory in the cache we have the loop
1974 * broken along max page order boundaries. This way we will not cause
1975 * any issues with the buddy page computation.
1977 static unsigned long __init
1978 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1979 unsigned long *end_pfn)
1981 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1982 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1983 unsigned long nr_pages = 0;
1986 /* First we loop through and initialize the page values */
1987 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1990 if (mo_pfn <= *start_pfn)
1993 t = min(mo_pfn, *end_pfn);
1994 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1996 if (mo_pfn < *end_pfn) {
1997 *start_pfn = mo_pfn;
2002 /* Reset values and now loop through freeing pages as needed */
2005 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2011 t = min(mo_pfn, epfn);
2012 deferred_free_pages(spfn, t);
2022 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2025 unsigned long spfn, epfn;
2026 struct zone *zone = arg;
2029 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2032 * Initialize and free pages in MAX_ORDER sized increments so that we
2033 * can avoid introducing any issues with the buddy allocator.
2035 while (spfn < end_pfn) {
2036 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2041 /* An arch may override for more concurrency. */
2043 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2048 /* Initialise remaining memory on a node */
2049 static int __init deferred_init_memmap(void *data)
2051 pg_data_t *pgdat = data;
2052 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2053 unsigned long spfn = 0, epfn = 0;
2054 unsigned long first_init_pfn, flags;
2055 unsigned long start = jiffies;
2057 int zid, max_threads;
2060 /* Bind memory initialisation thread to a local node if possible */
2061 if (!cpumask_empty(cpumask))
2062 set_cpus_allowed_ptr(current, cpumask);
2064 pgdat_resize_lock(pgdat, &flags);
2065 first_init_pfn = pgdat->first_deferred_pfn;
2066 if (first_init_pfn == ULONG_MAX) {
2067 pgdat_resize_unlock(pgdat, &flags);
2068 pgdat_init_report_one_done();
2072 /* Sanity check boundaries */
2073 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2074 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2075 pgdat->first_deferred_pfn = ULONG_MAX;
2078 * Once we unlock here, the zone cannot be grown anymore, thus if an
2079 * interrupt thread must allocate this early in boot, zone must be
2080 * pre-grown prior to start of deferred page initialization.
2082 pgdat_resize_unlock(pgdat, &flags);
2084 /* Only the highest zone is deferred so find it */
2085 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2086 zone = pgdat->node_zones + zid;
2087 if (first_init_pfn < zone_end_pfn(zone))
2091 /* If the zone is empty somebody else may have cleared out the zone */
2092 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2096 max_threads = deferred_page_init_max_threads(cpumask);
2098 while (spfn < epfn) {
2099 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2100 struct padata_mt_job job = {
2101 .thread_fn = deferred_init_memmap_chunk,
2104 .size = epfn_align - spfn,
2105 .align = PAGES_PER_SECTION,
2106 .min_chunk = PAGES_PER_SECTION,
2107 .max_threads = max_threads,
2110 padata_do_multithreaded(&job);
2111 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2115 /* Sanity check that the next zone really is unpopulated */
2116 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2118 pr_info("node %d deferred pages initialised in %ums\n",
2119 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2121 pgdat_init_report_one_done();
2126 * If this zone has deferred pages, try to grow it by initializing enough
2127 * deferred pages to satisfy the allocation specified by order, rounded up to
2128 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2129 * of SECTION_SIZE bytes by initializing struct pages in increments of
2130 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2132 * Return true when zone was grown, otherwise return false. We return true even
2133 * when we grow less than requested, to let the caller decide if there are
2134 * enough pages to satisfy the allocation.
2136 * Note: We use noinline because this function is needed only during boot, and
2137 * it is called from a __ref function _deferred_grow_zone. This way we are
2138 * making sure that it is not inlined into permanent text section.
2140 static noinline bool __init
2141 deferred_grow_zone(struct zone *zone, unsigned int order)
2143 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2144 pg_data_t *pgdat = zone->zone_pgdat;
2145 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2146 unsigned long spfn, epfn, flags;
2147 unsigned long nr_pages = 0;
2150 /* Only the last zone may have deferred pages */
2151 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2154 pgdat_resize_lock(pgdat, &flags);
2157 * If someone grew this zone while we were waiting for spinlock, return
2158 * true, as there might be enough pages already.
2160 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2161 pgdat_resize_unlock(pgdat, &flags);
2165 /* If the zone is empty somebody else may have cleared out the zone */
2166 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2167 first_deferred_pfn)) {
2168 pgdat->first_deferred_pfn = ULONG_MAX;
2169 pgdat_resize_unlock(pgdat, &flags);
2170 /* Retry only once. */
2171 return first_deferred_pfn != ULONG_MAX;
2175 * Initialize and free pages in MAX_ORDER sized increments so
2176 * that we can avoid introducing any issues with the buddy
2179 while (spfn < epfn) {
2180 /* update our first deferred PFN for this section */
2181 first_deferred_pfn = spfn;
2183 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2184 touch_nmi_watchdog();
2186 /* We should only stop along section boundaries */
2187 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2190 /* If our quota has been met we can stop here */
2191 if (nr_pages >= nr_pages_needed)
2195 pgdat->first_deferred_pfn = spfn;
2196 pgdat_resize_unlock(pgdat, &flags);
2198 return nr_pages > 0;
2202 * deferred_grow_zone() is __init, but it is called from
2203 * get_page_from_freelist() during early boot until deferred_pages permanently
2204 * disables this call. This is why we have refdata wrapper to avoid warning,
2205 * and to ensure that the function body gets unloaded.
2208 _deferred_grow_zone(struct zone *zone, unsigned int order)
2210 return deferred_grow_zone(zone, order);
2213 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2215 void __init page_alloc_init_late(void)
2220 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2222 /* There will be num_node_state(N_MEMORY) threads */
2223 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2224 for_each_node_state(nid, N_MEMORY) {
2225 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2228 /* Block until all are initialised */
2229 wait_for_completion(&pgdat_init_all_done_comp);
2232 * We initialized the rest of the deferred pages. Permanently disable
2233 * on-demand struct page initialization.
2235 static_branch_disable(&deferred_pages);
2237 /* Reinit limits that are based on free pages after the kernel is up */
2238 files_maxfiles_init();
2243 /* Discard memblock private memory */
2246 for_each_node_state(nid, N_MEMORY)
2247 shuffle_free_memory(NODE_DATA(nid));
2249 for_each_populated_zone(zone)
2250 set_zone_contiguous(zone);
2254 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2255 void __init init_cma_reserved_pageblock(struct page *page)
2257 unsigned i = pageblock_nr_pages;
2258 struct page *p = page;
2261 __ClearPageReserved(p);
2262 set_page_count(p, 0);
2265 set_pageblock_migratetype(page, MIGRATE_CMA);
2267 if (pageblock_order >= MAX_ORDER) {
2268 i = pageblock_nr_pages;
2271 set_page_refcounted(p);
2272 __free_pages(p, MAX_ORDER - 1);
2273 p += MAX_ORDER_NR_PAGES;
2274 } while (i -= MAX_ORDER_NR_PAGES);
2276 set_page_refcounted(page);
2277 __free_pages(page, pageblock_order);
2280 adjust_managed_page_count(page, pageblock_nr_pages);
2281 page_zone(page)->cma_pages += pageblock_nr_pages;
2286 * The order of subdivision here is critical for the IO subsystem.
2287 * Please do not alter this order without good reasons and regression
2288 * testing. Specifically, as large blocks of memory are subdivided,
2289 * the order in which smaller blocks are delivered depends on the order
2290 * they're subdivided in this function. This is the primary factor
2291 * influencing the order in which pages are delivered to the IO
2292 * subsystem according to empirical testing, and this is also justified
2293 * by considering the behavior of a buddy system containing a single
2294 * large block of memory acted on by a series of small allocations.
2295 * This behavior is a critical factor in sglist merging's success.
2299 static inline void expand(struct zone *zone, struct page *page,
2300 int low, int high, int migratetype)
2302 unsigned long size = 1 << high;
2304 while (high > low) {
2307 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2310 * Mark as guard pages (or page), that will allow to
2311 * merge back to allocator when buddy will be freed.
2312 * Corresponding page table entries will not be touched,
2313 * pages will stay not present in virtual address space
2315 if (set_page_guard(zone, &page[size], high, migratetype))
2318 add_to_free_list(&page[size], zone, high, migratetype);
2319 set_buddy_order(&page[size], high);
2323 static void check_new_page_bad(struct page *page)
2325 if (unlikely(page->flags & __PG_HWPOISON)) {
2326 /* Don't complain about hwpoisoned pages */
2327 page_mapcount_reset(page); /* remove PageBuddy */
2332 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2336 * This page is about to be returned from the page allocator
2338 static inline int check_new_page(struct page *page)
2340 if (likely(page_expected_state(page,
2341 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2344 check_new_page_bad(page);
2348 #ifdef CONFIG_DEBUG_VM
2350 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2351 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2352 * also checked when pcp lists are refilled from the free lists.
2354 static inline bool check_pcp_refill(struct page *page)
2356 if (debug_pagealloc_enabled_static())
2357 return check_new_page(page);
2362 static inline bool check_new_pcp(struct page *page)
2364 return check_new_page(page);
2368 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2369 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2370 * enabled, they are also checked when being allocated from the pcp lists.
2372 static inline bool check_pcp_refill(struct page *page)
2374 return check_new_page(page);
2376 static inline bool check_new_pcp(struct page *page)
2378 if (debug_pagealloc_enabled_static())
2379 return check_new_page(page);
2383 #endif /* CONFIG_DEBUG_VM */
2385 static bool check_new_pages(struct page *page, unsigned int order)
2388 for (i = 0; i < (1 << order); i++) {
2389 struct page *p = page + i;
2391 if (unlikely(check_new_page(p)))
2398 inline void post_alloc_hook(struct page *page, unsigned int order,
2401 set_page_private(page, 0);
2402 set_page_refcounted(page);
2404 arch_alloc_page(page, order);
2405 debug_pagealloc_map_pages(page, 1 << order);
2408 * Page unpoisoning must happen before memory initialization.
2409 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2410 * allocations and the page unpoisoning code will complain.
2412 kernel_unpoison_pages(page, 1 << order);
2415 * As memory initialization might be integrated into KASAN,
2416 * kasan_alloc_pages and kernel_init_free_pages must be
2417 * kept together to avoid discrepancies in behavior.
2419 if (kasan_has_integrated_init()) {
2420 kasan_alloc_pages(page, order, gfp_flags);
2422 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2424 kasan_unpoison_pages(page, order, init);
2426 kernel_init_free_pages(page, 1 << order,
2427 gfp_flags & __GFP_ZEROTAGS);
2430 set_page_owner(page, order, gfp_flags);
2433 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2434 unsigned int alloc_flags)
2436 post_alloc_hook(page, order, gfp_flags);
2438 if (order && (gfp_flags & __GFP_COMP))
2439 prep_compound_page(page, order);
2442 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2443 * allocate the page. The expectation is that the caller is taking
2444 * steps that will free more memory. The caller should avoid the page
2445 * being used for !PFMEMALLOC purposes.
2447 if (alloc_flags & ALLOC_NO_WATERMARKS)
2448 set_page_pfmemalloc(page);
2450 clear_page_pfmemalloc(page);
2454 * Go through the free lists for the given migratetype and remove
2455 * the smallest available page from the freelists
2457 static __always_inline
2458 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2461 unsigned int current_order;
2462 struct free_area *area;
2465 /* Find a page of the appropriate size in the preferred list */
2466 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2467 area = &(zone->free_area[current_order]);
2468 page = get_page_from_free_area(area, migratetype);
2471 del_page_from_free_list(page, zone, current_order);
2472 expand(zone, page, order, current_order, migratetype);
2473 set_pcppage_migratetype(page, migratetype);
2482 * This array describes the order lists are fallen back to when
2483 * the free lists for the desirable migrate type are depleted
2485 static int fallbacks[MIGRATE_TYPES][3] = {
2486 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2487 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2488 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2490 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2492 #ifdef CONFIG_MEMORY_ISOLATION
2493 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2498 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2501 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2504 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2505 unsigned int order) { return NULL; }
2509 * Move the free pages in a range to the freelist tail of the requested type.
2510 * Note that start_page and end_pages are not aligned on a pageblock
2511 * boundary. If alignment is required, use move_freepages_block()
2513 static int move_freepages(struct zone *zone,
2514 unsigned long start_pfn, unsigned long end_pfn,
2515 int migratetype, int *num_movable)
2520 int pages_moved = 0;
2522 for (pfn = start_pfn; pfn <= end_pfn;) {
2523 if (!pfn_valid_within(pfn)) {
2528 page = pfn_to_page(pfn);
2529 if (!PageBuddy(page)) {
2531 * We assume that pages that could be isolated for
2532 * migration are movable. But we don't actually try
2533 * isolating, as that would be expensive.
2536 (PageLRU(page) || __PageMovable(page)))
2542 /* Make sure we are not inadvertently changing nodes */
2543 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2544 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2546 order = buddy_order(page);
2547 move_to_free_list(page, zone, order, migratetype);
2549 pages_moved += 1 << order;
2555 int move_freepages_block(struct zone *zone, struct page *page,
2556 int migratetype, int *num_movable)
2558 unsigned long start_pfn, end_pfn, pfn;
2563 pfn = page_to_pfn(page);
2564 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2565 end_pfn = start_pfn + pageblock_nr_pages - 1;
2567 /* Do not cross zone boundaries */
2568 if (!zone_spans_pfn(zone, start_pfn))
2570 if (!zone_spans_pfn(zone, end_pfn))
2573 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2577 static void change_pageblock_range(struct page *pageblock_page,
2578 int start_order, int migratetype)
2580 int nr_pageblocks = 1 << (start_order - pageblock_order);
2582 while (nr_pageblocks--) {
2583 set_pageblock_migratetype(pageblock_page, migratetype);
2584 pageblock_page += pageblock_nr_pages;
2589 * When we are falling back to another migratetype during allocation, try to
2590 * steal extra free pages from the same pageblocks to satisfy further
2591 * allocations, instead of polluting multiple pageblocks.
2593 * If we are stealing a relatively large buddy page, it is likely there will
2594 * be more free pages in the pageblock, so try to steal them all. For
2595 * reclaimable and unmovable allocations, we steal regardless of page size,
2596 * as fragmentation caused by those allocations polluting movable pageblocks
2597 * is worse than movable allocations stealing from unmovable and reclaimable
2600 static bool can_steal_fallback(unsigned int order, int start_mt)
2603 * Leaving this order check is intended, although there is
2604 * relaxed order check in next check. The reason is that
2605 * we can actually steal whole pageblock if this condition met,
2606 * but, below check doesn't guarantee it and that is just heuristic
2607 * so could be changed anytime.
2609 if (order >= pageblock_order)
2612 if (order >= pageblock_order / 2 ||
2613 start_mt == MIGRATE_RECLAIMABLE ||
2614 start_mt == MIGRATE_UNMOVABLE ||
2615 page_group_by_mobility_disabled)
2621 static inline bool boost_watermark(struct zone *zone)
2623 unsigned long max_boost;
2625 if (!watermark_boost_factor)
2628 * Don't bother in zones that are unlikely to produce results.
2629 * On small machines, including kdump capture kernels running
2630 * in a small area, boosting the watermark can cause an out of
2631 * memory situation immediately.
2633 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2636 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2637 watermark_boost_factor, 10000);
2640 * high watermark may be uninitialised if fragmentation occurs
2641 * very early in boot so do not boost. We do not fall
2642 * through and boost by pageblock_nr_pages as failing
2643 * allocations that early means that reclaim is not going
2644 * to help and it may even be impossible to reclaim the
2645 * boosted watermark resulting in a hang.
2650 max_boost = max(pageblock_nr_pages, max_boost);
2652 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2659 * This function implements actual steal behaviour. If order is large enough,
2660 * we can steal whole pageblock. If not, we first move freepages in this
2661 * pageblock to our migratetype and determine how many already-allocated pages
2662 * are there in the pageblock with a compatible migratetype. If at least half
2663 * of pages are free or compatible, we can change migratetype of the pageblock
2664 * itself, so pages freed in the future will be put on the correct free list.
2666 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2667 unsigned int alloc_flags, int start_type, bool whole_block)
2669 unsigned int current_order = buddy_order(page);
2670 int free_pages, movable_pages, alike_pages;
2673 old_block_type = get_pageblock_migratetype(page);
2676 * This can happen due to races and we want to prevent broken
2677 * highatomic accounting.
2679 if (is_migrate_highatomic(old_block_type))
2682 /* Take ownership for orders >= pageblock_order */
2683 if (current_order >= pageblock_order) {
2684 change_pageblock_range(page, current_order, start_type);
2689 * Boost watermarks to increase reclaim pressure to reduce the
2690 * likelihood of future fallbacks. Wake kswapd now as the node
2691 * may be balanced overall and kswapd will not wake naturally.
2693 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2694 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2696 /* We are not allowed to try stealing from the whole block */
2700 free_pages = move_freepages_block(zone, page, start_type,
2703 * Determine how many pages are compatible with our allocation.
2704 * For movable allocation, it's the number of movable pages which
2705 * we just obtained. For other types it's a bit more tricky.
2707 if (start_type == MIGRATE_MOVABLE) {
2708 alike_pages = movable_pages;
2711 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2712 * to MOVABLE pageblock, consider all non-movable pages as
2713 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2714 * vice versa, be conservative since we can't distinguish the
2715 * exact migratetype of non-movable pages.
2717 if (old_block_type == MIGRATE_MOVABLE)
2718 alike_pages = pageblock_nr_pages
2719 - (free_pages + movable_pages);
2724 /* moving whole block can fail due to zone boundary conditions */
2729 * If a sufficient number of pages in the block are either free or of
2730 * comparable migratability as our allocation, claim the whole block.
2732 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2733 page_group_by_mobility_disabled)
2734 set_pageblock_migratetype(page, start_type);
2739 move_to_free_list(page, zone, current_order, start_type);
2743 * Check whether there is a suitable fallback freepage with requested order.
2744 * If only_stealable is true, this function returns fallback_mt only if
2745 * we can steal other freepages all together. This would help to reduce
2746 * fragmentation due to mixed migratetype pages in one pageblock.
2748 int find_suitable_fallback(struct free_area *area, unsigned int order,
2749 int migratetype, bool only_stealable, bool *can_steal)
2754 if (area->nr_free == 0)
2759 fallback_mt = fallbacks[migratetype][i];
2760 if (fallback_mt == MIGRATE_TYPES)
2763 if (free_area_empty(area, fallback_mt))
2766 if (can_steal_fallback(order, migratetype))
2769 if (!only_stealable)
2780 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2781 * there are no empty page blocks that contain a page with a suitable order
2783 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2784 unsigned int alloc_order)
2787 unsigned long max_managed, flags;
2790 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2791 * Check is race-prone but harmless.
2793 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2794 if (zone->nr_reserved_highatomic >= max_managed)
2797 spin_lock_irqsave(&zone->lock, flags);
2799 /* Recheck the nr_reserved_highatomic limit under the lock */
2800 if (zone->nr_reserved_highatomic >= max_managed)
2804 mt = get_pageblock_migratetype(page);
2805 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2806 && !is_migrate_cma(mt)) {
2807 zone->nr_reserved_highatomic += pageblock_nr_pages;
2808 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2809 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2813 spin_unlock_irqrestore(&zone->lock, flags);
2817 * Used when an allocation is about to fail under memory pressure. This
2818 * potentially hurts the reliability of high-order allocations when under
2819 * intense memory pressure but failed atomic allocations should be easier
2820 * to recover from than an OOM.
2822 * If @force is true, try to unreserve a pageblock even though highatomic
2823 * pageblock is exhausted.
2825 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2828 struct zonelist *zonelist = ac->zonelist;
2829 unsigned long flags;
2836 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2839 * Preserve at least one pageblock unless memory pressure
2842 if (!force && zone->nr_reserved_highatomic <=
2846 spin_lock_irqsave(&zone->lock, flags);
2847 for (order = 0; order < MAX_ORDER; order++) {
2848 struct free_area *area = &(zone->free_area[order]);
2850 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2855 * In page freeing path, migratetype change is racy so
2856 * we can counter several free pages in a pageblock
2857 * in this loop although we changed the pageblock type
2858 * from highatomic to ac->migratetype. So we should
2859 * adjust the count once.
2861 if (is_migrate_highatomic_page(page)) {
2863 * It should never happen but changes to
2864 * locking could inadvertently allow a per-cpu
2865 * drain to add pages to MIGRATE_HIGHATOMIC
2866 * while unreserving so be safe and watch for
2869 zone->nr_reserved_highatomic -= min(
2871 zone->nr_reserved_highatomic);
2875 * Convert to ac->migratetype and avoid the normal
2876 * pageblock stealing heuristics. Minimally, the caller
2877 * is doing the work and needs the pages. More
2878 * importantly, if the block was always converted to
2879 * MIGRATE_UNMOVABLE or another type then the number
2880 * of pageblocks that cannot be completely freed
2883 set_pageblock_migratetype(page, ac->migratetype);
2884 ret = move_freepages_block(zone, page, ac->migratetype,
2887 spin_unlock_irqrestore(&zone->lock, flags);
2891 spin_unlock_irqrestore(&zone->lock, flags);
2898 * Try finding a free buddy page on the fallback list and put it on the free
2899 * list of requested migratetype, possibly along with other pages from the same
2900 * block, depending on fragmentation avoidance heuristics. Returns true if
2901 * fallback was found so that __rmqueue_smallest() can grab it.
2903 * The use of signed ints for order and current_order is a deliberate
2904 * deviation from the rest of this file, to make the for loop
2905 * condition simpler.
2907 static __always_inline bool
2908 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2909 unsigned int alloc_flags)
2911 struct free_area *area;
2913 int min_order = order;
2919 * Do not steal pages from freelists belonging to other pageblocks
2920 * i.e. orders < pageblock_order. If there are no local zones free,
2921 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2923 if (alloc_flags & ALLOC_NOFRAGMENT)
2924 min_order = pageblock_order;
2927 * Find the largest available free page in the other list. This roughly
2928 * approximates finding the pageblock with the most free pages, which
2929 * would be too costly to do exactly.
2931 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2933 area = &(zone->free_area[current_order]);
2934 fallback_mt = find_suitable_fallback(area, current_order,
2935 start_migratetype, false, &can_steal);
2936 if (fallback_mt == -1)
2940 * We cannot steal all free pages from the pageblock and the
2941 * requested migratetype is movable. In that case it's better to
2942 * steal and split the smallest available page instead of the
2943 * largest available page, because even if the next movable
2944 * allocation falls back into a different pageblock than this
2945 * one, it won't cause permanent fragmentation.
2947 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2948 && current_order > order)
2957 for (current_order = order; current_order < MAX_ORDER;
2959 area = &(zone->free_area[current_order]);
2960 fallback_mt = find_suitable_fallback(area, current_order,
2961 start_migratetype, false, &can_steal);
2962 if (fallback_mt != -1)
2967 * This should not happen - we already found a suitable fallback
2968 * when looking for the largest page.
2970 VM_BUG_ON(current_order == MAX_ORDER);
2973 page = get_page_from_free_area(area, fallback_mt);
2975 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2978 trace_mm_page_alloc_extfrag(page, order, current_order,
2979 start_migratetype, fallback_mt);
2986 * Do the hard work of removing an element from the buddy allocator.
2987 * Call me with the zone->lock already held.
2989 static __always_inline struct page *
2990 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2991 unsigned int alloc_flags)
2995 if (IS_ENABLED(CONFIG_CMA)) {
2997 * Balance movable allocations between regular and CMA areas by
2998 * allocating from CMA when over half of the zone's free memory
2999 * is in the CMA area.
3001 if (alloc_flags & ALLOC_CMA &&
3002 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3003 zone_page_state(zone, NR_FREE_PAGES) / 2) {
3004 page = __rmqueue_cma_fallback(zone, order);
3010 page = __rmqueue_smallest(zone, order, migratetype);
3011 if (unlikely(!page)) {
3012 if (alloc_flags & ALLOC_CMA)
3013 page = __rmqueue_cma_fallback(zone, order);
3015 if (!page && __rmqueue_fallback(zone, order, migratetype,
3021 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3026 * Obtain a specified number of elements from the buddy allocator, all under
3027 * a single hold of the lock, for efficiency. Add them to the supplied list.
3028 * Returns the number of new pages which were placed at *list.
3030 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3031 unsigned long count, struct list_head *list,
3032 int migratetype, unsigned int alloc_flags)
3034 int i, allocated = 0;
3037 * local_lock_irq held so equivalent to spin_lock_irqsave for
3038 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3040 spin_lock(&zone->lock);
3041 for (i = 0; i < count; ++i) {
3042 struct page *page = __rmqueue(zone, order, migratetype,
3044 if (unlikely(page == NULL))
3047 if (unlikely(check_pcp_refill(page)))
3051 * Split buddy pages returned by expand() are received here in
3052 * physical page order. The page is added to the tail of
3053 * caller's list. From the callers perspective, the linked list
3054 * is ordered by page number under some conditions. This is
3055 * useful for IO devices that can forward direction from the
3056 * head, thus also in the physical page order. This is useful
3057 * for IO devices that can merge IO requests if the physical
3058 * pages are ordered properly.
3060 list_add_tail(&page->lru, list);
3062 if (is_migrate_cma(get_pcppage_migratetype(page)))
3063 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3068 * i pages were removed from the buddy list even if some leak due
3069 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3070 * on i. Do not confuse with 'allocated' which is the number of
3071 * pages added to the pcp list.
3073 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3074 spin_unlock(&zone->lock);
3080 * Called from the vmstat counter updater to drain pagesets of this
3081 * currently executing processor on remote nodes after they have
3084 * Note that this function must be called with the thread pinned to
3085 * a single processor.
3087 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3089 unsigned long flags;
3090 int to_drain, batch;
3092 local_lock_irqsave(&pagesets.lock, flags);
3093 batch = READ_ONCE(pcp->batch);
3094 to_drain = min(pcp->count, batch);
3096 free_pcppages_bulk(zone, to_drain, pcp);
3097 local_unlock_irqrestore(&pagesets.lock, flags);
3102 * Drain pcplists of the indicated processor and zone.
3104 * The processor must either be the current processor and the
3105 * thread pinned to the current processor or a processor that
3108 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3110 unsigned long flags;
3111 struct per_cpu_pages *pcp;
3113 local_lock_irqsave(&pagesets.lock, flags);
3115 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3117 free_pcppages_bulk(zone, pcp->count, pcp);
3119 local_unlock_irqrestore(&pagesets.lock, flags);
3123 * Drain pcplists of all zones on the indicated processor.
3125 * The processor must either be the current processor and the
3126 * thread pinned to the current processor or a processor that
3129 static void drain_pages(unsigned int cpu)
3133 for_each_populated_zone(zone) {
3134 drain_pages_zone(cpu, zone);
3139 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3141 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3142 * the single zone's pages.
3144 void drain_local_pages(struct zone *zone)
3146 int cpu = smp_processor_id();
3149 drain_pages_zone(cpu, zone);
3154 static void drain_local_pages_wq(struct work_struct *work)
3156 struct pcpu_drain *drain;
3158 drain = container_of(work, struct pcpu_drain, work);
3161 * drain_all_pages doesn't use proper cpu hotplug protection so
3162 * we can race with cpu offline when the WQ can move this from
3163 * a cpu pinned worker to an unbound one. We can operate on a different
3164 * cpu which is alright but we also have to make sure to not move to
3168 drain_local_pages(drain->zone);
3173 * The implementation of drain_all_pages(), exposing an extra parameter to
3174 * drain on all cpus.
3176 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3177 * not empty. The check for non-emptiness can however race with a free to
3178 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3179 * that need the guarantee that every CPU has drained can disable the
3180 * optimizing racy check.
3182 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3187 * Allocate in the BSS so we won't require allocation in
3188 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3190 static cpumask_t cpus_with_pcps;
3193 * Make sure nobody triggers this path before mm_percpu_wq is fully
3196 if (WARN_ON_ONCE(!mm_percpu_wq))
3200 * Do not drain if one is already in progress unless it's specific to
3201 * a zone. Such callers are primarily CMA and memory hotplug and need
3202 * the drain to be complete when the call returns.
3204 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3207 mutex_lock(&pcpu_drain_mutex);
3211 * We don't care about racing with CPU hotplug event
3212 * as offline notification will cause the notified
3213 * cpu to drain that CPU pcps and on_each_cpu_mask
3214 * disables preemption as part of its processing
3216 for_each_online_cpu(cpu) {
3217 struct per_cpu_pages *pcp;
3219 bool has_pcps = false;
3221 if (force_all_cpus) {
3223 * The pcp.count check is racy, some callers need a
3224 * guarantee that no cpu is missed.
3228 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3232 for_each_populated_zone(z) {
3233 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3242 cpumask_set_cpu(cpu, &cpus_with_pcps);
3244 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3247 for_each_cpu(cpu, &cpus_with_pcps) {
3248 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3251 INIT_WORK(&drain->work, drain_local_pages_wq);
3252 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3254 for_each_cpu(cpu, &cpus_with_pcps)
3255 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3257 mutex_unlock(&pcpu_drain_mutex);
3261 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3263 * When zone parameter is non-NULL, spill just the single zone's pages.
3265 * Note that this can be extremely slow as the draining happens in a workqueue.
3267 void drain_all_pages(struct zone *zone)
3269 __drain_all_pages(zone, false);
3272 #ifdef CONFIG_HIBERNATION
3275 * Touch the watchdog for every WD_PAGE_COUNT pages.
3277 #define WD_PAGE_COUNT (128*1024)
3279 void mark_free_pages(struct zone *zone)
3281 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3282 unsigned long flags;
3283 unsigned int order, t;
3286 if (zone_is_empty(zone))
3289 spin_lock_irqsave(&zone->lock, flags);
3291 max_zone_pfn = zone_end_pfn(zone);
3292 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3293 if (pfn_valid(pfn)) {
3294 page = pfn_to_page(pfn);
3296 if (!--page_count) {
3297 touch_nmi_watchdog();
3298 page_count = WD_PAGE_COUNT;
3301 if (page_zone(page) != zone)
3304 if (!swsusp_page_is_forbidden(page))
3305 swsusp_unset_page_free(page);
3308 for_each_migratetype_order(order, t) {
3309 list_for_each_entry(page,
3310 &zone->free_area[order].free_list[t], lru) {
3313 pfn = page_to_pfn(page);
3314 for (i = 0; i < (1UL << order); i++) {
3315 if (!--page_count) {
3316 touch_nmi_watchdog();
3317 page_count = WD_PAGE_COUNT;
3319 swsusp_set_page_free(pfn_to_page(pfn + i));
3323 spin_unlock_irqrestore(&zone->lock, flags);
3325 #endif /* CONFIG_PM */
3327 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3332 if (!free_pcp_prepare(page, order))
3335 migratetype = get_pfnblock_migratetype(page, pfn);
3336 set_pcppage_migratetype(page, migratetype);
3340 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch)
3342 int min_nr_free, max_nr_free;
3344 /* Check for PCP disabled or boot pageset */
3345 if (unlikely(high < batch))
3348 /* Leave at least pcp->batch pages on the list */
3349 min_nr_free = batch;
3350 max_nr_free = high - batch;
3353 * Double the number of pages freed each time there is subsequent
3354 * freeing of pages without any allocation.
3356 batch <<= pcp->free_factor;
3357 if (batch < max_nr_free)
3359 batch = clamp(batch, min_nr_free, max_nr_free);
3364 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone)
3366 int high = READ_ONCE(pcp->high);
3368 if (unlikely(!high))
3371 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3375 * If reclaim is active, limit the number of pages that can be
3376 * stored on pcp lists
3378 return min(READ_ONCE(pcp->batch) << 2, high);
3381 static void free_unref_page_commit(struct page *page, unsigned long pfn,
3382 int migratetype, unsigned int order)
3384 struct zone *zone = page_zone(page);
3385 struct per_cpu_pages *pcp;
3389 __count_vm_event(PGFREE);
3390 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3391 pindex = order_to_pindex(migratetype, order);
3392 list_add(&page->lru, &pcp->lists[pindex]);
3393 pcp->count += 1 << order;
3394 high = nr_pcp_high(pcp, zone);
3395 if (pcp->count >= high) {
3396 int batch = READ_ONCE(pcp->batch);
3398 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch), pcp);
3405 void free_unref_page(struct page *page, unsigned int order)
3407 unsigned long flags;
3408 unsigned long pfn = page_to_pfn(page);
3411 if (!free_unref_page_prepare(page, pfn, order))
3415 * We only track unmovable, reclaimable and movable on pcp lists.
3416 * Place ISOLATE pages on the isolated list because they are being
3417 * offlined but treat HIGHATOMIC as movable pages so we can get those
3418 * areas back if necessary. Otherwise, we may have to free
3419 * excessively into the page allocator
3421 migratetype = get_pcppage_migratetype(page);
3422 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3423 if (unlikely(is_migrate_isolate(migratetype))) {
3424 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3427 migratetype = MIGRATE_MOVABLE;
3430 local_lock_irqsave(&pagesets.lock, flags);
3431 free_unref_page_commit(page, pfn, migratetype, order);
3432 local_unlock_irqrestore(&pagesets.lock, flags);
3436 * Free a list of 0-order pages
3438 void free_unref_page_list(struct list_head *list)
3440 struct page *page, *next;
3441 unsigned long flags, pfn;
3442 int batch_count = 0;
3445 /* Prepare pages for freeing */
3446 list_for_each_entry_safe(page, next, list, lru) {
3447 pfn = page_to_pfn(page);
3448 if (!free_unref_page_prepare(page, pfn, 0))
3449 list_del(&page->lru);
3452 * Free isolated pages directly to the allocator, see
3453 * comment in free_unref_page.
3455 migratetype = get_pcppage_migratetype(page);
3456 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3457 if (unlikely(is_migrate_isolate(migratetype))) {
3458 list_del(&page->lru);
3459 free_one_page(page_zone(page), page, pfn, 0,
3460 migratetype, FPI_NONE);
3465 * Non-isolated types over MIGRATE_PCPTYPES get added
3466 * to the MIGRATE_MOVABLE pcp list.
3468 set_pcppage_migratetype(page, MIGRATE_MOVABLE);
3471 set_page_private(page, pfn);
3474 local_lock_irqsave(&pagesets.lock, flags);
3475 list_for_each_entry_safe(page, next, list, lru) {
3476 pfn = page_private(page);
3477 set_page_private(page, 0);
3478 migratetype = get_pcppage_migratetype(page);
3479 trace_mm_page_free_batched(page);
3480 free_unref_page_commit(page, pfn, migratetype, 0);
3483 * Guard against excessive IRQ disabled times when we get
3484 * a large list of pages to free.
3486 if (++batch_count == SWAP_CLUSTER_MAX) {
3487 local_unlock_irqrestore(&pagesets.lock, flags);
3489 local_lock_irqsave(&pagesets.lock, flags);
3492 local_unlock_irqrestore(&pagesets.lock, flags);
3496 * split_page takes a non-compound higher-order page, and splits it into
3497 * n (1<<order) sub-pages: page[0..n]
3498 * Each sub-page must be freed individually.
3500 * Note: this is probably too low level an operation for use in drivers.
3501 * Please consult with lkml before using this in your driver.
3503 void split_page(struct page *page, unsigned int order)
3507 VM_BUG_ON_PAGE(PageCompound(page), page);
3508 VM_BUG_ON_PAGE(!page_count(page), page);
3510 for (i = 1; i < (1 << order); i++)
3511 set_page_refcounted(page + i);
3512 split_page_owner(page, 1 << order);
3513 split_page_memcg(page, 1 << order);
3515 EXPORT_SYMBOL_GPL(split_page);
3517 int __isolate_free_page(struct page *page, unsigned int order)
3519 unsigned long watermark;
3523 BUG_ON(!PageBuddy(page));
3525 zone = page_zone(page);
3526 mt = get_pageblock_migratetype(page);
3528 if (!is_migrate_isolate(mt)) {
3530 * Obey watermarks as if the page was being allocated. We can
3531 * emulate a high-order watermark check with a raised order-0
3532 * watermark, because we already know our high-order page
3535 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3536 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3539 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3542 /* Remove page from free list */
3544 del_page_from_free_list(page, zone, order);
3547 * Set the pageblock if the isolated page is at least half of a
3550 if (order >= pageblock_order - 1) {
3551 struct page *endpage = page + (1 << order) - 1;
3552 for (; page < endpage; page += pageblock_nr_pages) {
3553 int mt = get_pageblock_migratetype(page);
3554 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3555 && !is_migrate_highatomic(mt))
3556 set_pageblock_migratetype(page,
3562 return 1UL << order;
3566 * __putback_isolated_page - Return a now-isolated page back where we got it
3567 * @page: Page that was isolated
3568 * @order: Order of the isolated page
3569 * @mt: The page's pageblock's migratetype
3571 * This function is meant to return a page pulled from the free lists via
3572 * __isolate_free_page back to the free lists they were pulled from.
3574 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3576 struct zone *zone = page_zone(page);
3578 /* zone lock should be held when this function is called */
3579 lockdep_assert_held(&zone->lock);
3581 /* Return isolated page to tail of freelist. */
3582 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3583 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3587 * Update NUMA hit/miss statistics
3589 * Must be called with interrupts disabled.
3591 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3595 enum numa_stat_item local_stat = NUMA_LOCAL;
3597 /* skip numa counters update if numa stats is disabled */
3598 if (!static_branch_likely(&vm_numa_stat_key))
3601 if (zone_to_nid(z) != numa_node_id())
3602 local_stat = NUMA_OTHER;
3604 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3605 __count_numa_events(z, NUMA_HIT, nr_account);
3607 __count_numa_events(z, NUMA_MISS, nr_account);
3608 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3610 __count_numa_events(z, local_stat, nr_account);
3614 /* Remove page from the per-cpu list, caller must protect the list */
3616 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3618 unsigned int alloc_flags,
3619 struct per_cpu_pages *pcp,
3620 struct list_head *list)
3625 if (list_empty(list)) {
3626 int batch = READ_ONCE(pcp->batch);
3630 * Scale batch relative to order if batch implies
3631 * free pages can be stored on the PCP. Batch can
3632 * be 1 for small zones or for boot pagesets which
3633 * should never store free pages as the pages may
3634 * belong to arbitrary zones.
3637 batch = max(batch >> order, 2);
3638 alloced = rmqueue_bulk(zone, order,
3640 migratetype, alloc_flags);
3642 pcp->count += alloced << order;
3643 if (unlikely(list_empty(list)))
3647 page = list_first_entry(list, struct page, lru);
3648 list_del(&page->lru);
3649 pcp->count -= 1 << order;
3650 } while (check_new_pcp(page));
3655 /* Lock and remove page from the per-cpu list */
3656 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3657 struct zone *zone, unsigned int order,
3658 gfp_t gfp_flags, int migratetype,
3659 unsigned int alloc_flags)
3661 struct per_cpu_pages *pcp;
3662 struct list_head *list;
3664 unsigned long flags;
3666 local_lock_irqsave(&pagesets.lock, flags);
3669 * On allocation, reduce the number of pages that are batch freed.
3670 * See nr_pcp_free() where free_factor is increased for subsequent
3673 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3674 pcp->free_factor >>= 1;
3675 list = &pcp->lists[order_to_pindex(migratetype, order)];
3676 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3677 local_unlock_irqrestore(&pagesets.lock, flags);
3679 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3680 zone_statistics(preferred_zone, zone, 1);
3686 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3689 struct page *rmqueue(struct zone *preferred_zone,
3690 struct zone *zone, unsigned int order,
3691 gfp_t gfp_flags, unsigned int alloc_flags,
3694 unsigned long flags;
3697 if (likely(pcp_allowed_order(order))) {
3699 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3700 * we need to skip it when CMA area isn't allowed.
3702 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3703 migratetype != MIGRATE_MOVABLE) {
3704 page = rmqueue_pcplist(preferred_zone, zone, order,
3705 gfp_flags, migratetype, alloc_flags);
3711 * We most definitely don't want callers attempting to
3712 * allocate greater than order-1 page units with __GFP_NOFAIL.
3714 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3715 spin_lock_irqsave(&zone->lock, flags);
3720 * order-0 request can reach here when the pcplist is skipped
3721 * due to non-CMA allocation context. HIGHATOMIC area is
3722 * reserved for high-order atomic allocation, so order-0
3723 * request should skip it.
3725 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3726 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3728 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3731 page = __rmqueue(zone, order, migratetype, alloc_flags);
3732 } while (page && check_new_pages(page, order));
3736 __mod_zone_freepage_state(zone, -(1 << order),
3737 get_pcppage_migratetype(page));
3738 spin_unlock_irqrestore(&zone->lock, flags);
3740 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3741 zone_statistics(preferred_zone, zone, 1);
3744 /* Separate test+clear to avoid unnecessary atomics */
3745 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3746 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3747 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3750 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3754 spin_unlock_irqrestore(&zone->lock, flags);
3758 #ifdef CONFIG_FAIL_PAGE_ALLOC
3761 struct fault_attr attr;
3763 bool ignore_gfp_highmem;
3764 bool ignore_gfp_reclaim;
3766 } fail_page_alloc = {
3767 .attr = FAULT_ATTR_INITIALIZER,
3768 .ignore_gfp_reclaim = true,
3769 .ignore_gfp_highmem = true,
3773 static int __init setup_fail_page_alloc(char *str)
3775 return setup_fault_attr(&fail_page_alloc.attr, str);
3777 __setup("fail_page_alloc=", setup_fail_page_alloc);
3779 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3781 if (order < fail_page_alloc.min_order)
3783 if (gfp_mask & __GFP_NOFAIL)
3785 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3787 if (fail_page_alloc.ignore_gfp_reclaim &&
3788 (gfp_mask & __GFP_DIRECT_RECLAIM))
3791 return should_fail(&fail_page_alloc.attr, 1 << order);
3794 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3796 static int __init fail_page_alloc_debugfs(void)
3798 umode_t mode = S_IFREG | 0600;
3801 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3802 &fail_page_alloc.attr);
3804 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3805 &fail_page_alloc.ignore_gfp_reclaim);
3806 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3807 &fail_page_alloc.ignore_gfp_highmem);
3808 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3813 late_initcall(fail_page_alloc_debugfs);
3815 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3817 #else /* CONFIG_FAIL_PAGE_ALLOC */
3819 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3824 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3826 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3828 return __should_fail_alloc_page(gfp_mask, order);
3830 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3832 static inline long __zone_watermark_unusable_free(struct zone *z,
3833 unsigned int order, unsigned int alloc_flags)
3835 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3836 long unusable_free = (1 << order) - 1;
3839 * If the caller does not have rights to ALLOC_HARDER then subtract
3840 * the high-atomic reserves. This will over-estimate the size of the
3841 * atomic reserve but it avoids a search.
3843 if (likely(!alloc_harder))
3844 unusable_free += z->nr_reserved_highatomic;
3847 /* If allocation can't use CMA areas don't use free CMA pages */
3848 if (!(alloc_flags & ALLOC_CMA))
3849 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3852 return unusable_free;
3856 * Return true if free base pages are above 'mark'. For high-order checks it
3857 * will return true of the order-0 watermark is reached and there is at least
3858 * one free page of a suitable size. Checking now avoids taking the zone lock
3859 * to check in the allocation paths if no pages are free.
3861 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3862 int highest_zoneidx, unsigned int alloc_flags,
3867 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3869 /* free_pages may go negative - that's OK */
3870 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3872 if (alloc_flags & ALLOC_HIGH)
3875 if (unlikely(alloc_harder)) {
3877 * OOM victims can try even harder than normal ALLOC_HARDER
3878 * users on the grounds that it's definitely going to be in
3879 * the exit path shortly and free memory. Any allocation it
3880 * makes during the free path will be small and short-lived.
3882 if (alloc_flags & ALLOC_OOM)
3889 * Check watermarks for an order-0 allocation request. If these
3890 * are not met, then a high-order request also cannot go ahead
3891 * even if a suitable page happened to be free.
3893 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3896 /* If this is an order-0 request then the watermark is fine */
3900 /* For a high-order request, check at least one suitable page is free */
3901 for (o = order; o < MAX_ORDER; o++) {
3902 struct free_area *area = &z->free_area[o];
3908 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3909 if (!free_area_empty(area, mt))
3914 if ((alloc_flags & ALLOC_CMA) &&
3915 !free_area_empty(area, MIGRATE_CMA)) {
3919 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3925 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3926 int highest_zoneidx, unsigned int alloc_flags)
3928 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3929 zone_page_state(z, NR_FREE_PAGES));
3932 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3933 unsigned long mark, int highest_zoneidx,
3934 unsigned int alloc_flags, gfp_t gfp_mask)
3938 free_pages = zone_page_state(z, NR_FREE_PAGES);
3941 * Fast check for order-0 only. If this fails then the reserves
3942 * need to be calculated.
3947 fast_free = free_pages;
3948 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3949 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3953 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3957 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3958 * when checking the min watermark. The min watermark is the
3959 * point where boosting is ignored so that kswapd is woken up
3960 * when below the low watermark.
3962 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3963 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3964 mark = z->_watermark[WMARK_MIN];
3965 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3966 alloc_flags, free_pages);
3972 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3973 unsigned long mark, int highest_zoneidx)
3975 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3977 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3978 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3980 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3985 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3987 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3988 node_reclaim_distance;
3990 #else /* CONFIG_NUMA */
3991 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3995 #endif /* CONFIG_NUMA */
3998 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3999 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4000 * premature use of a lower zone may cause lowmem pressure problems that
4001 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4002 * probably too small. It only makes sense to spread allocations to avoid
4003 * fragmentation between the Normal and DMA32 zones.
4005 static inline unsigned int
4006 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4008 unsigned int alloc_flags;
4011 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4014 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4016 #ifdef CONFIG_ZONE_DMA32
4020 if (zone_idx(zone) != ZONE_NORMAL)
4024 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4025 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4026 * on UMA that if Normal is populated then so is DMA32.
4028 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4029 if (nr_online_nodes > 1 && !populated_zone(--zone))
4032 alloc_flags |= ALLOC_NOFRAGMENT;
4033 #endif /* CONFIG_ZONE_DMA32 */
4037 /* Must be called after current_gfp_context() which can change gfp_mask */
4038 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4039 unsigned int alloc_flags)
4042 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4043 alloc_flags |= ALLOC_CMA;
4049 * get_page_from_freelist goes through the zonelist trying to allocate
4052 static struct page *
4053 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4054 const struct alloc_context *ac)
4058 struct pglist_data *last_pgdat_dirty_limit = NULL;
4063 * Scan zonelist, looking for a zone with enough free.
4064 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4066 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4067 z = ac->preferred_zoneref;
4068 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4073 if (cpusets_enabled() &&
4074 (alloc_flags & ALLOC_CPUSET) &&
4075 !__cpuset_zone_allowed(zone, gfp_mask))
4078 * When allocating a page cache page for writing, we
4079 * want to get it from a node that is within its dirty
4080 * limit, such that no single node holds more than its
4081 * proportional share of globally allowed dirty pages.
4082 * The dirty limits take into account the node's
4083 * lowmem reserves and high watermark so that kswapd
4084 * should be able to balance it without having to
4085 * write pages from its LRU list.
4087 * XXX: For now, allow allocations to potentially
4088 * exceed the per-node dirty limit in the slowpath
4089 * (spread_dirty_pages unset) before going into reclaim,
4090 * which is important when on a NUMA setup the allowed
4091 * nodes are together not big enough to reach the
4092 * global limit. The proper fix for these situations
4093 * will require awareness of nodes in the
4094 * dirty-throttling and the flusher threads.
4096 if (ac->spread_dirty_pages) {
4097 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4100 if (!node_dirty_ok(zone->zone_pgdat)) {
4101 last_pgdat_dirty_limit = zone->zone_pgdat;
4106 if (no_fallback && nr_online_nodes > 1 &&
4107 zone != ac->preferred_zoneref->zone) {
4111 * If moving to a remote node, retry but allow
4112 * fragmenting fallbacks. Locality is more important
4113 * than fragmentation avoidance.
4115 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4116 if (zone_to_nid(zone) != local_nid) {
4117 alloc_flags &= ~ALLOC_NOFRAGMENT;
4122 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4123 if (!zone_watermark_fast(zone, order, mark,
4124 ac->highest_zoneidx, alloc_flags,
4128 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4130 * Watermark failed for this zone, but see if we can
4131 * grow this zone if it contains deferred pages.
4133 if (static_branch_unlikely(&deferred_pages)) {
4134 if (_deferred_grow_zone(zone, order))
4138 /* Checked here to keep the fast path fast */
4139 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4140 if (alloc_flags & ALLOC_NO_WATERMARKS)
4143 if (!node_reclaim_enabled() ||
4144 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4147 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4149 case NODE_RECLAIM_NOSCAN:
4152 case NODE_RECLAIM_FULL:
4153 /* scanned but unreclaimable */
4156 /* did we reclaim enough */
4157 if (zone_watermark_ok(zone, order, mark,
4158 ac->highest_zoneidx, alloc_flags))
4166 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4167 gfp_mask, alloc_flags, ac->migratetype);
4169 prep_new_page(page, order, gfp_mask, alloc_flags);
4172 * If this is a high-order atomic allocation then check
4173 * if the pageblock should be reserved for the future
4175 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4176 reserve_highatomic_pageblock(page, zone, order);
4180 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4181 /* Try again if zone has deferred pages */
4182 if (static_branch_unlikely(&deferred_pages)) {
4183 if (_deferred_grow_zone(zone, order))
4191 * It's possible on a UMA machine to get through all zones that are
4192 * fragmented. If avoiding fragmentation, reset and try again.
4195 alloc_flags &= ~ALLOC_NOFRAGMENT;
4202 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4204 unsigned int filter = SHOW_MEM_FILTER_NODES;
4207 * This documents exceptions given to allocations in certain
4208 * contexts that are allowed to allocate outside current's set
4211 if (!(gfp_mask & __GFP_NOMEMALLOC))
4212 if (tsk_is_oom_victim(current) ||
4213 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4214 filter &= ~SHOW_MEM_FILTER_NODES;
4215 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4216 filter &= ~SHOW_MEM_FILTER_NODES;
4218 show_mem(filter, nodemask);
4221 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4223 struct va_format vaf;
4225 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4227 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4230 va_start(args, fmt);
4233 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4234 current->comm, &vaf, gfp_mask, &gfp_mask,
4235 nodemask_pr_args(nodemask));
4238 cpuset_print_current_mems_allowed();
4241 warn_alloc_show_mem(gfp_mask, nodemask);
4244 static inline struct page *
4245 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4246 unsigned int alloc_flags,
4247 const struct alloc_context *ac)
4251 page = get_page_from_freelist(gfp_mask, order,
4252 alloc_flags|ALLOC_CPUSET, ac);
4254 * fallback to ignore cpuset restriction if our nodes
4258 page = get_page_from_freelist(gfp_mask, order,
4264 static inline struct page *
4265 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4266 const struct alloc_context *ac, unsigned long *did_some_progress)
4268 struct oom_control oc = {
4269 .zonelist = ac->zonelist,
4270 .nodemask = ac->nodemask,
4272 .gfp_mask = gfp_mask,
4277 *did_some_progress = 0;
4280 * Acquire the oom lock. If that fails, somebody else is
4281 * making progress for us.
4283 if (!mutex_trylock(&oom_lock)) {
4284 *did_some_progress = 1;
4285 schedule_timeout_uninterruptible(1);
4290 * Go through the zonelist yet one more time, keep very high watermark
4291 * here, this is only to catch a parallel oom killing, we must fail if
4292 * we're still under heavy pressure. But make sure that this reclaim
4293 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4294 * allocation which will never fail due to oom_lock already held.
4296 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4297 ~__GFP_DIRECT_RECLAIM, order,
4298 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4302 /* Coredumps can quickly deplete all memory reserves */
4303 if (current->flags & PF_DUMPCORE)
4305 /* The OOM killer will not help higher order allocs */
4306 if (order > PAGE_ALLOC_COSTLY_ORDER)
4309 * We have already exhausted all our reclaim opportunities without any
4310 * success so it is time to admit defeat. We will skip the OOM killer
4311 * because it is very likely that the caller has a more reasonable
4312 * fallback than shooting a random task.
4314 * The OOM killer may not free memory on a specific node.
4316 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4318 /* The OOM killer does not needlessly kill tasks for lowmem */
4319 if (ac->highest_zoneidx < ZONE_NORMAL)
4321 if (pm_suspended_storage())
4324 * XXX: GFP_NOFS allocations should rather fail than rely on
4325 * other request to make a forward progress.
4326 * We are in an unfortunate situation where out_of_memory cannot
4327 * do much for this context but let's try it to at least get
4328 * access to memory reserved if the current task is killed (see
4329 * out_of_memory). Once filesystems are ready to handle allocation
4330 * failures more gracefully we should just bail out here.
4333 /* Exhausted what can be done so it's blame time */
4334 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4335 *did_some_progress = 1;
4338 * Help non-failing allocations by giving them access to memory
4341 if (gfp_mask & __GFP_NOFAIL)
4342 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4343 ALLOC_NO_WATERMARKS, ac);
4346 mutex_unlock(&oom_lock);
4351 * Maximum number of compaction retries with a progress before OOM
4352 * killer is consider as the only way to move forward.
4354 #define MAX_COMPACT_RETRIES 16
4356 #ifdef CONFIG_COMPACTION
4357 /* Try memory compaction for high-order allocations before reclaim */
4358 static struct page *
4359 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4360 unsigned int alloc_flags, const struct alloc_context *ac,
4361 enum compact_priority prio, enum compact_result *compact_result)
4363 struct page *page = NULL;
4364 unsigned long pflags;
4365 unsigned int noreclaim_flag;
4370 psi_memstall_enter(&pflags);
4371 noreclaim_flag = memalloc_noreclaim_save();
4373 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4376 memalloc_noreclaim_restore(noreclaim_flag);
4377 psi_memstall_leave(&pflags);
4379 if (*compact_result == COMPACT_SKIPPED)
4382 * At least in one zone compaction wasn't deferred or skipped, so let's
4383 * count a compaction stall
4385 count_vm_event(COMPACTSTALL);
4387 /* Prep a captured page if available */
4389 prep_new_page(page, order, gfp_mask, alloc_flags);
4391 /* Try get a page from the freelist if available */
4393 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4396 struct zone *zone = page_zone(page);
4398 zone->compact_blockskip_flush = false;
4399 compaction_defer_reset(zone, order, true);
4400 count_vm_event(COMPACTSUCCESS);
4405 * It's bad if compaction run occurs and fails. The most likely reason
4406 * is that pages exist, but not enough to satisfy watermarks.
4408 count_vm_event(COMPACTFAIL);
4416 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4417 enum compact_result compact_result,
4418 enum compact_priority *compact_priority,
4419 int *compaction_retries)
4421 int max_retries = MAX_COMPACT_RETRIES;
4424 int retries = *compaction_retries;
4425 enum compact_priority priority = *compact_priority;
4430 if (fatal_signal_pending(current))
4433 if (compaction_made_progress(compact_result))
4434 (*compaction_retries)++;
4437 * compaction considers all the zone as desperately out of memory
4438 * so it doesn't really make much sense to retry except when the
4439 * failure could be caused by insufficient priority
4441 if (compaction_failed(compact_result))
4442 goto check_priority;
4445 * compaction was skipped because there are not enough order-0 pages
4446 * to work with, so we retry only if it looks like reclaim can help.
4448 if (compaction_needs_reclaim(compact_result)) {
4449 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4454 * make sure the compaction wasn't deferred or didn't bail out early
4455 * due to locks contention before we declare that we should give up.
4456 * But the next retry should use a higher priority if allowed, so
4457 * we don't just keep bailing out endlessly.
4459 if (compaction_withdrawn(compact_result)) {
4460 goto check_priority;
4464 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4465 * costly ones because they are de facto nofail and invoke OOM
4466 * killer to move on while costly can fail and users are ready
4467 * to cope with that. 1/4 retries is rather arbitrary but we
4468 * would need much more detailed feedback from compaction to
4469 * make a better decision.
4471 if (order > PAGE_ALLOC_COSTLY_ORDER)
4473 if (*compaction_retries <= max_retries) {
4479 * Make sure there are attempts at the highest priority if we exhausted
4480 * all retries or failed at the lower priorities.
4483 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4484 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4486 if (*compact_priority > min_priority) {
4487 (*compact_priority)--;
4488 *compaction_retries = 0;
4492 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4496 static inline struct page *
4497 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4498 unsigned int alloc_flags, const struct alloc_context *ac,
4499 enum compact_priority prio, enum compact_result *compact_result)
4501 *compact_result = COMPACT_SKIPPED;
4506 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4507 enum compact_result compact_result,
4508 enum compact_priority *compact_priority,
4509 int *compaction_retries)
4514 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4518 * There are setups with compaction disabled which would prefer to loop
4519 * inside the allocator rather than hit the oom killer prematurely.
4520 * Let's give them a good hope and keep retrying while the order-0
4521 * watermarks are OK.
4523 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4524 ac->highest_zoneidx, ac->nodemask) {
4525 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4526 ac->highest_zoneidx, alloc_flags))
4531 #endif /* CONFIG_COMPACTION */
4533 #ifdef CONFIG_LOCKDEP
4534 static struct lockdep_map __fs_reclaim_map =
4535 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4537 static bool __need_reclaim(gfp_t gfp_mask)
4539 /* no reclaim without waiting on it */
4540 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4543 /* this guy won't enter reclaim */
4544 if (current->flags & PF_MEMALLOC)
4547 if (gfp_mask & __GFP_NOLOCKDEP)
4553 void __fs_reclaim_acquire(void)
4555 lock_map_acquire(&__fs_reclaim_map);
4558 void __fs_reclaim_release(void)
4560 lock_map_release(&__fs_reclaim_map);
4563 void fs_reclaim_acquire(gfp_t gfp_mask)
4565 gfp_mask = current_gfp_context(gfp_mask);
4567 if (__need_reclaim(gfp_mask)) {
4568 if (gfp_mask & __GFP_FS)
4569 __fs_reclaim_acquire();
4571 #ifdef CONFIG_MMU_NOTIFIER
4572 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4573 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4578 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4580 void fs_reclaim_release(gfp_t gfp_mask)
4582 gfp_mask = current_gfp_context(gfp_mask);
4584 if (__need_reclaim(gfp_mask)) {
4585 if (gfp_mask & __GFP_FS)
4586 __fs_reclaim_release();
4589 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4592 /* Perform direct synchronous page reclaim */
4593 static unsigned long
4594 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4595 const struct alloc_context *ac)
4597 unsigned int noreclaim_flag;
4598 unsigned long pflags, progress;
4602 /* We now go into synchronous reclaim */
4603 cpuset_memory_pressure_bump();
4604 psi_memstall_enter(&pflags);
4605 fs_reclaim_acquire(gfp_mask);
4606 noreclaim_flag = memalloc_noreclaim_save();
4608 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4611 memalloc_noreclaim_restore(noreclaim_flag);
4612 fs_reclaim_release(gfp_mask);
4613 psi_memstall_leave(&pflags);
4620 /* The really slow allocator path where we enter direct reclaim */
4621 static inline struct page *
4622 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4623 unsigned int alloc_flags, const struct alloc_context *ac,
4624 unsigned long *did_some_progress)
4626 struct page *page = NULL;
4627 bool drained = false;
4629 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4630 if (unlikely(!(*did_some_progress)))
4634 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4637 * If an allocation failed after direct reclaim, it could be because
4638 * pages are pinned on the per-cpu lists or in high alloc reserves.
4639 * Shrink them and try again
4641 if (!page && !drained) {
4642 unreserve_highatomic_pageblock(ac, false);
4643 drain_all_pages(NULL);
4651 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4652 const struct alloc_context *ac)
4656 pg_data_t *last_pgdat = NULL;
4657 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4659 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4661 if (last_pgdat != zone->zone_pgdat)
4662 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4663 last_pgdat = zone->zone_pgdat;
4667 static inline unsigned int
4668 gfp_to_alloc_flags(gfp_t gfp_mask)
4670 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4673 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4674 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4675 * to save two branches.
4677 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4678 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4681 * The caller may dip into page reserves a bit more if the caller
4682 * cannot run direct reclaim, or if the caller has realtime scheduling
4683 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4684 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4686 alloc_flags |= (__force int)
4687 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4689 if (gfp_mask & __GFP_ATOMIC) {
4691 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4692 * if it can't schedule.
4694 if (!(gfp_mask & __GFP_NOMEMALLOC))
4695 alloc_flags |= ALLOC_HARDER;
4697 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4698 * comment for __cpuset_node_allowed().
4700 alloc_flags &= ~ALLOC_CPUSET;
4701 } else if (unlikely(rt_task(current)) && !in_interrupt())
4702 alloc_flags |= ALLOC_HARDER;
4704 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4709 static bool oom_reserves_allowed(struct task_struct *tsk)
4711 if (!tsk_is_oom_victim(tsk))
4715 * !MMU doesn't have oom reaper so give access to memory reserves
4716 * only to the thread with TIF_MEMDIE set
4718 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4725 * Distinguish requests which really need access to full memory
4726 * reserves from oom victims which can live with a portion of it
4728 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4730 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4732 if (gfp_mask & __GFP_MEMALLOC)
4733 return ALLOC_NO_WATERMARKS;
4734 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4735 return ALLOC_NO_WATERMARKS;
4736 if (!in_interrupt()) {
4737 if (current->flags & PF_MEMALLOC)
4738 return ALLOC_NO_WATERMARKS;
4739 else if (oom_reserves_allowed(current))
4746 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4748 return !!__gfp_pfmemalloc_flags(gfp_mask);
4752 * Checks whether it makes sense to retry the reclaim to make a forward progress
4753 * for the given allocation request.
4755 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4756 * without success, or when we couldn't even meet the watermark if we
4757 * reclaimed all remaining pages on the LRU lists.
4759 * Returns true if a retry is viable or false to enter the oom path.
4762 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4763 struct alloc_context *ac, int alloc_flags,
4764 bool did_some_progress, int *no_progress_loops)
4771 * Costly allocations might have made a progress but this doesn't mean
4772 * their order will become available due to high fragmentation so
4773 * always increment the no progress counter for them
4775 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4776 *no_progress_loops = 0;
4778 (*no_progress_loops)++;
4781 * Make sure we converge to OOM if we cannot make any progress
4782 * several times in the row.
4784 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4785 /* Before OOM, exhaust highatomic_reserve */
4786 return unreserve_highatomic_pageblock(ac, true);
4790 * Keep reclaiming pages while there is a chance this will lead
4791 * somewhere. If none of the target zones can satisfy our allocation
4792 * request even if all reclaimable pages are considered then we are
4793 * screwed and have to go OOM.
4795 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4796 ac->highest_zoneidx, ac->nodemask) {
4797 unsigned long available;
4798 unsigned long reclaimable;
4799 unsigned long min_wmark = min_wmark_pages(zone);
4802 available = reclaimable = zone_reclaimable_pages(zone);
4803 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4806 * Would the allocation succeed if we reclaimed all
4807 * reclaimable pages?
4809 wmark = __zone_watermark_ok(zone, order, min_wmark,
4810 ac->highest_zoneidx, alloc_flags, available);
4811 trace_reclaim_retry_zone(z, order, reclaimable,
4812 available, min_wmark, *no_progress_loops, wmark);
4815 * If we didn't make any progress and have a lot of
4816 * dirty + writeback pages then we should wait for
4817 * an IO to complete to slow down the reclaim and
4818 * prevent from pre mature OOM
4820 if (!did_some_progress) {
4821 unsigned long write_pending;
4823 write_pending = zone_page_state_snapshot(zone,
4824 NR_ZONE_WRITE_PENDING);
4826 if (2 * write_pending > reclaimable) {
4827 congestion_wait(BLK_RW_ASYNC, HZ/10);
4839 * Memory allocation/reclaim might be called from a WQ context and the
4840 * current implementation of the WQ concurrency control doesn't
4841 * recognize that a particular WQ is congested if the worker thread is
4842 * looping without ever sleeping. Therefore we have to do a short sleep
4843 * here rather than calling cond_resched().
4845 if (current->flags & PF_WQ_WORKER)
4846 schedule_timeout_uninterruptible(1);
4853 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4856 * It's possible that cpuset's mems_allowed and the nodemask from
4857 * mempolicy don't intersect. This should be normally dealt with by
4858 * policy_nodemask(), but it's possible to race with cpuset update in
4859 * such a way the check therein was true, and then it became false
4860 * before we got our cpuset_mems_cookie here.
4861 * This assumes that for all allocations, ac->nodemask can come only
4862 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4863 * when it does not intersect with the cpuset restrictions) or the
4864 * caller can deal with a violated nodemask.
4866 if (cpusets_enabled() && ac->nodemask &&
4867 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4868 ac->nodemask = NULL;
4873 * When updating a task's mems_allowed or mempolicy nodemask, it is
4874 * possible to race with parallel threads in such a way that our
4875 * allocation can fail while the mask is being updated. If we are about
4876 * to fail, check if the cpuset changed during allocation and if so,
4879 if (read_mems_allowed_retry(cpuset_mems_cookie))
4885 static inline struct page *
4886 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4887 struct alloc_context *ac)
4889 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4890 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4891 struct page *page = NULL;
4892 unsigned int alloc_flags;
4893 unsigned long did_some_progress;
4894 enum compact_priority compact_priority;
4895 enum compact_result compact_result;
4896 int compaction_retries;
4897 int no_progress_loops;
4898 unsigned int cpuset_mems_cookie;
4902 * We also sanity check to catch abuse of atomic reserves being used by
4903 * callers that are not in atomic context.
4905 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4906 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4907 gfp_mask &= ~__GFP_ATOMIC;
4910 compaction_retries = 0;
4911 no_progress_loops = 0;
4912 compact_priority = DEF_COMPACT_PRIORITY;
4913 cpuset_mems_cookie = read_mems_allowed_begin();
4916 * The fast path uses conservative alloc_flags to succeed only until
4917 * kswapd needs to be woken up, and to avoid the cost of setting up
4918 * alloc_flags precisely. So we do that now.
4920 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4923 * We need to recalculate the starting point for the zonelist iterator
4924 * because we might have used different nodemask in the fast path, or
4925 * there was a cpuset modification and we are retrying - otherwise we
4926 * could end up iterating over non-eligible zones endlessly.
4928 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4929 ac->highest_zoneidx, ac->nodemask);
4930 if (!ac->preferred_zoneref->zone)
4933 if (alloc_flags & ALLOC_KSWAPD)
4934 wake_all_kswapds(order, gfp_mask, ac);
4937 * The adjusted alloc_flags might result in immediate success, so try
4940 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4945 * For costly allocations, try direct compaction first, as it's likely
4946 * that we have enough base pages and don't need to reclaim. For non-
4947 * movable high-order allocations, do that as well, as compaction will
4948 * try prevent permanent fragmentation by migrating from blocks of the
4950 * Don't try this for allocations that are allowed to ignore
4951 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4953 if (can_direct_reclaim &&
4955 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4956 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4957 page = __alloc_pages_direct_compact(gfp_mask, order,
4959 INIT_COMPACT_PRIORITY,
4965 * Checks for costly allocations with __GFP_NORETRY, which
4966 * includes some THP page fault allocations
4968 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4970 * If allocating entire pageblock(s) and compaction
4971 * failed because all zones are below low watermarks
4972 * or is prohibited because it recently failed at this
4973 * order, fail immediately unless the allocator has
4974 * requested compaction and reclaim retry.
4977 * - potentially very expensive because zones are far
4978 * below their low watermarks or this is part of very
4979 * bursty high order allocations,
4980 * - not guaranteed to help because isolate_freepages()
4981 * may not iterate over freed pages as part of its
4983 * - unlikely to make entire pageblocks free on its
4986 if (compact_result == COMPACT_SKIPPED ||
4987 compact_result == COMPACT_DEFERRED)
4991 * Looks like reclaim/compaction is worth trying, but
4992 * sync compaction could be very expensive, so keep
4993 * using async compaction.
4995 compact_priority = INIT_COMPACT_PRIORITY;
5000 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5001 if (alloc_flags & ALLOC_KSWAPD)
5002 wake_all_kswapds(order, gfp_mask, ac);
5004 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5006 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
5009 * Reset the nodemask and zonelist iterators if memory policies can be
5010 * ignored. These allocations are high priority and system rather than
5013 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5014 ac->nodemask = NULL;
5015 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5016 ac->highest_zoneidx, ac->nodemask);
5019 /* Attempt with potentially adjusted zonelist and alloc_flags */
5020 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5024 /* Caller is not willing to reclaim, we can't balance anything */
5025 if (!can_direct_reclaim)
5028 /* Avoid recursion of direct reclaim */
5029 if (current->flags & PF_MEMALLOC)
5032 /* Try direct reclaim and then allocating */
5033 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5034 &did_some_progress);
5038 /* Try direct compaction and then allocating */
5039 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5040 compact_priority, &compact_result);
5044 /* Do not loop if specifically requested */
5045 if (gfp_mask & __GFP_NORETRY)
5049 * Do not retry costly high order allocations unless they are
5050 * __GFP_RETRY_MAYFAIL
5052 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5055 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5056 did_some_progress > 0, &no_progress_loops))
5060 * It doesn't make any sense to retry for the compaction if the order-0
5061 * reclaim is not able to make any progress because the current
5062 * implementation of the compaction depends on the sufficient amount
5063 * of free memory (see __compaction_suitable)
5065 if (did_some_progress > 0 &&
5066 should_compact_retry(ac, order, alloc_flags,
5067 compact_result, &compact_priority,
5068 &compaction_retries))
5072 /* Deal with possible cpuset update races before we start OOM killing */
5073 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5076 /* Reclaim has failed us, start killing things */
5077 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5081 /* Avoid allocations with no watermarks from looping endlessly */
5082 if (tsk_is_oom_victim(current) &&
5083 (alloc_flags & ALLOC_OOM ||
5084 (gfp_mask & __GFP_NOMEMALLOC)))
5087 /* Retry as long as the OOM killer is making progress */
5088 if (did_some_progress) {
5089 no_progress_loops = 0;
5094 /* Deal with possible cpuset update races before we fail */
5095 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5099 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5102 if (gfp_mask & __GFP_NOFAIL) {
5104 * All existing users of the __GFP_NOFAIL are blockable, so warn
5105 * of any new users that actually require GFP_NOWAIT
5107 if (WARN_ON_ONCE(!can_direct_reclaim))
5111 * PF_MEMALLOC request from this context is rather bizarre
5112 * because we cannot reclaim anything and only can loop waiting
5113 * for somebody to do a work for us
5115 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5118 * non failing costly orders are a hard requirement which we
5119 * are not prepared for much so let's warn about these users
5120 * so that we can identify them and convert them to something
5123 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5126 * Help non-failing allocations by giving them access to memory
5127 * reserves but do not use ALLOC_NO_WATERMARKS because this
5128 * could deplete whole memory reserves which would just make
5129 * the situation worse
5131 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5139 warn_alloc(gfp_mask, ac->nodemask,
5140 "page allocation failure: order:%u", order);
5145 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5146 int preferred_nid, nodemask_t *nodemask,
5147 struct alloc_context *ac, gfp_t *alloc_gfp,
5148 unsigned int *alloc_flags)
5150 ac->highest_zoneidx = gfp_zone(gfp_mask);
5151 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5152 ac->nodemask = nodemask;
5153 ac->migratetype = gfp_migratetype(gfp_mask);
5155 if (cpusets_enabled()) {
5156 *alloc_gfp |= __GFP_HARDWALL;
5158 * When we are in the interrupt context, it is irrelevant
5159 * to the current task context. It means that any node ok.
5161 if (!in_interrupt() && !ac->nodemask)
5162 ac->nodemask = &cpuset_current_mems_allowed;
5164 *alloc_flags |= ALLOC_CPUSET;
5167 fs_reclaim_acquire(gfp_mask);
5168 fs_reclaim_release(gfp_mask);
5170 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5172 if (should_fail_alloc_page(gfp_mask, order))
5175 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5177 /* Dirty zone balancing only done in the fast path */
5178 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5181 * The preferred zone is used for statistics but crucially it is
5182 * also used as the starting point for the zonelist iterator. It
5183 * may get reset for allocations that ignore memory policies.
5185 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5186 ac->highest_zoneidx, ac->nodemask);
5192 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5193 * @gfp: GFP flags for the allocation
5194 * @preferred_nid: The preferred NUMA node ID to allocate from
5195 * @nodemask: Set of nodes to allocate from, may be NULL
5196 * @nr_pages: The number of pages desired on the list or array
5197 * @page_list: Optional list to store the allocated pages
5198 * @page_array: Optional array to store the pages
5200 * This is a batched version of the page allocator that attempts to
5201 * allocate nr_pages quickly. Pages are added to page_list if page_list
5202 * is not NULL, otherwise it is assumed that the page_array is valid.
5204 * For lists, nr_pages is the number of pages that should be allocated.
5206 * For arrays, only NULL elements are populated with pages and nr_pages
5207 * is the maximum number of pages that will be stored in the array.
5209 * Returns the number of pages on the list or array.
5211 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5212 nodemask_t *nodemask, int nr_pages,
5213 struct list_head *page_list,
5214 struct page **page_array)
5217 unsigned long flags;
5220 struct per_cpu_pages *pcp;
5221 struct list_head *pcp_list;
5222 struct alloc_context ac;
5224 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5225 int nr_populated = 0, nr_account = 0;
5228 * Skip populated array elements to determine if any pages need
5229 * to be allocated before disabling IRQs.
5231 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5234 /* No pages requested? */
5235 if (unlikely(nr_pages <= 0))
5238 /* Already populated array? */
5239 if (unlikely(page_array && nr_pages - nr_populated == 0))
5242 /* Use the single page allocator for one page. */
5243 if (nr_pages - nr_populated == 1)
5246 #ifdef CONFIG_PAGE_OWNER
5248 * PAGE_OWNER may recurse into the allocator to allocate space to
5249 * save the stack with pagesets.lock held. Releasing/reacquiring
5250 * removes much of the performance benefit of bulk allocation so
5251 * force the caller to allocate one page at a time as it'll have
5252 * similar performance to added complexity to the bulk allocator.
5254 if (static_branch_unlikely(&page_owner_inited))
5258 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5259 gfp &= gfp_allowed_mask;
5261 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5265 /* Find an allowed local zone that meets the low watermark. */
5266 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5269 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5270 !__cpuset_zone_allowed(zone, gfp)) {
5274 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5275 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5279 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5280 if (zone_watermark_fast(zone, 0, mark,
5281 zonelist_zone_idx(ac.preferred_zoneref),
5282 alloc_flags, gfp)) {
5288 * If there are no allowed local zones that meets the watermarks then
5289 * try to allocate a single page and reclaim if necessary.
5291 if (unlikely(!zone))
5294 /* Attempt the batch allocation */
5295 local_lock_irqsave(&pagesets.lock, flags);
5296 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5297 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5299 while (nr_populated < nr_pages) {
5301 /* Skip existing pages */
5302 if (page_array && page_array[nr_populated]) {
5307 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5309 if (unlikely(!page)) {
5310 /* Try and get at least one page */
5317 prep_new_page(page, 0, gfp, 0);
5319 list_add(&page->lru, page_list);
5321 page_array[nr_populated] = page;
5325 local_unlock_irqrestore(&pagesets.lock, flags);
5327 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5328 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5331 return nr_populated;
5334 local_unlock_irqrestore(&pagesets.lock, flags);
5337 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5340 list_add(&page->lru, page_list);
5342 page_array[nr_populated] = page;
5348 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5351 * This is the 'heart' of the zoned buddy allocator.
5353 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5354 nodemask_t *nodemask)
5357 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5358 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5359 struct alloc_context ac = { };
5362 * There are several places where we assume that the order value is sane
5363 * so bail out early if the request is out of bound.
5365 if (unlikely(order >= MAX_ORDER)) {
5366 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5370 gfp &= gfp_allowed_mask;
5372 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5373 * resp. GFP_NOIO which has to be inherited for all allocation requests
5374 * from a particular context which has been marked by
5375 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5376 * movable zones are not used during allocation.
5378 gfp = current_gfp_context(gfp);
5380 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5381 &alloc_gfp, &alloc_flags))
5385 * Forbid the first pass from falling back to types that fragment
5386 * memory until all local zones are considered.
5388 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5390 /* First allocation attempt */
5391 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5396 ac.spread_dirty_pages = false;
5399 * Restore the original nodemask if it was potentially replaced with
5400 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5402 ac.nodemask = nodemask;
5404 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5407 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5408 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5409 __free_pages(page, order);
5413 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5417 EXPORT_SYMBOL(__alloc_pages);
5420 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5421 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5422 * you need to access high mem.
5424 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5428 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5431 return (unsigned long) page_address(page);
5433 EXPORT_SYMBOL(__get_free_pages);
5435 unsigned long get_zeroed_page(gfp_t gfp_mask)
5437 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5439 EXPORT_SYMBOL(get_zeroed_page);
5442 * __free_pages - Free pages allocated with alloc_pages().
5443 * @page: The page pointer returned from alloc_pages().
5444 * @order: The order of the allocation.
5446 * This function can free multi-page allocations that are not compound
5447 * pages. It does not check that the @order passed in matches that of
5448 * the allocation, so it is easy to leak memory. Freeing more memory
5449 * than was allocated will probably emit a warning.
5451 * If the last reference to this page is speculative, it will be released
5452 * by put_page() which only frees the first page of a non-compound
5453 * allocation. To prevent the remaining pages from being leaked, we free
5454 * the subsequent pages here. If you want to use the page's reference
5455 * count to decide when to free the allocation, you should allocate a
5456 * compound page, and use put_page() instead of __free_pages().
5458 * Context: May be called in interrupt context or while holding a normal
5459 * spinlock, but not in NMI context or while holding a raw spinlock.
5461 void __free_pages(struct page *page, unsigned int order)
5463 if (put_page_testzero(page))
5464 free_the_page(page, order);
5465 else if (!PageHead(page))
5467 free_the_page(page + (1 << order), order);
5469 EXPORT_SYMBOL(__free_pages);
5471 void free_pages(unsigned long addr, unsigned int order)
5474 VM_BUG_ON(!virt_addr_valid((void *)addr));
5475 __free_pages(virt_to_page((void *)addr), order);
5479 EXPORT_SYMBOL(free_pages);
5483 * An arbitrary-length arbitrary-offset area of memory which resides
5484 * within a 0 or higher order page. Multiple fragments within that page
5485 * are individually refcounted, in the page's reference counter.
5487 * The page_frag functions below provide a simple allocation framework for
5488 * page fragments. This is used by the network stack and network device
5489 * drivers to provide a backing region of memory for use as either an
5490 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5492 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5495 struct page *page = NULL;
5496 gfp_t gfp = gfp_mask;
5498 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5499 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5501 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5502 PAGE_FRAG_CACHE_MAX_ORDER);
5503 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5505 if (unlikely(!page))
5506 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5508 nc->va = page ? page_address(page) : NULL;
5513 void __page_frag_cache_drain(struct page *page, unsigned int count)
5515 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5517 if (page_ref_sub_and_test(page, count))
5518 free_the_page(page, compound_order(page));
5520 EXPORT_SYMBOL(__page_frag_cache_drain);
5522 void *page_frag_alloc_align(struct page_frag_cache *nc,
5523 unsigned int fragsz, gfp_t gfp_mask,
5524 unsigned int align_mask)
5526 unsigned int size = PAGE_SIZE;
5530 if (unlikely(!nc->va)) {
5532 page = __page_frag_cache_refill(nc, gfp_mask);
5536 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5537 /* if size can vary use size else just use PAGE_SIZE */
5540 /* Even if we own the page, we do not use atomic_set().
5541 * This would break get_page_unless_zero() users.
5543 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5545 /* reset page count bias and offset to start of new frag */
5546 nc->pfmemalloc = page_is_pfmemalloc(page);
5547 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5551 offset = nc->offset - fragsz;
5552 if (unlikely(offset < 0)) {
5553 page = virt_to_page(nc->va);
5555 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5558 if (unlikely(nc->pfmemalloc)) {
5559 free_the_page(page, compound_order(page));
5563 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5564 /* if size can vary use size else just use PAGE_SIZE */
5567 /* OK, page count is 0, we can safely set it */
5568 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5570 /* reset page count bias and offset to start of new frag */
5571 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5572 offset = size - fragsz;
5576 offset &= align_mask;
5577 nc->offset = offset;
5579 return nc->va + offset;
5581 EXPORT_SYMBOL(page_frag_alloc_align);
5584 * Frees a page fragment allocated out of either a compound or order 0 page.
5586 void page_frag_free(void *addr)
5588 struct page *page = virt_to_head_page(addr);
5590 if (unlikely(put_page_testzero(page)))
5591 free_the_page(page, compound_order(page));
5593 EXPORT_SYMBOL(page_frag_free);
5595 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5599 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5600 unsigned long used = addr + PAGE_ALIGN(size);
5602 split_page(virt_to_page((void *)addr), order);
5603 while (used < alloc_end) {
5608 return (void *)addr;
5612 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5613 * @size: the number of bytes to allocate
5614 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5616 * This function is similar to alloc_pages(), except that it allocates the
5617 * minimum number of pages to satisfy the request. alloc_pages() can only
5618 * allocate memory in power-of-two pages.
5620 * This function is also limited by MAX_ORDER.
5622 * Memory allocated by this function must be released by free_pages_exact().
5624 * Return: pointer to the allocated area or %NULL in case of error.
5626 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5628 unsigned int order = get_order(size);
5631 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5632 gfp_mask &= ~__GFP_COMP;
5634 addr = __get_free_pages(gfp_mask, order);
5635 return make_alloc_exact(addr, order, size);
5637 EXPORT_SYMBOL(alloc_pages_exact);
5640 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5642 * @nid: the preferred node ID where memory should be allocated
5643 * @size: the number of bytes to allocate
5644 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5646 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5649 * Return: pointer to the allocated area or %NULL in case of error.
5651 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5653 unsigned int order = get_order(size);
5656 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5657 gfp_mask &= ~__GFP_COMP;
5659 p = alloc_pages_node(nid, gfp_mask, order);
5662 return make_alloc_exact((unsigned long)page_address(p), order, size);
5666 * free_pages_exact - release memory allocated via alloc_pages_exact()
5667 * @virt: the value returned by alloc_pages_exact.
5668 * @size: size of allocation, same value as passed to alloc_pages_exact().
5670 * Release the memory allocated by a previous call to alloc_pages_exact.
5672 void free_pages_exact(void *virt, size_t size)
5674 unsigned long addr = (unsigned long)virt;
5675 unsigned long end = addr + PAGE_ALIGN(size);
5677 while (addr < end) {
5682 EXPORT_SYMBOL(free_pages_exact);
5685 * nr_free_zone_pages - count number of pages beyond high watermark
5686 * @offset: The zone index of the highest zone
5688 * nr_free_zone_pages() counts the number of pages which are beyond the
5689 * high watermark within all zones at or below a given zone index. For each
5690 * zone, the number of pages is calculated as:
5692 * nr_free_zone_pages = managed_pages - high_pages
5694 * Return: number of pages beyond high watermark.
5696 static unsigned long nr_free_zone_pages(int offset)
5701 /* Just pick one node, since fallback list is circular */
5702 unsigned long sum = 0;
5704 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5706 for_each_zone_zonelist(zone, z, zonelist, offset) {
5707 unsigned long size = zone_managed_pages(zone);
5708 unsigned long high = high_wmark_pages(zone);
5717 * nr_free_buffer_pages - count number of pages beyond high watermark
5719 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5720 * watermark within ZONE_DMA and ZONE_NORMAL.
5722 * Return: number of pages beyond high watermark within ZONE_DMA and
5725 unsigned long nr_free_buffer_pages(void)
5727 return nr_free_zone_pages(gfp_zone(GFP_USER));
5729 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5731 static inline void show_node(struct zone *zone)
5733 if (IS_ENABLED(CONFIG_NUMA))
5734 printk("Node %d ", zone_to_nid(zone));
5737 long si_mem_available(void)
5740 unsigned long pagecache;
5741 unsigned long wmark_low = 0;
5742 unsigned long pages[NR_LRU_LISTS];
5743 unsigned long reclaimable;
5747 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5748 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5751 wmark_low += low_wmark_pages(zone);
5754 * Estimate the amount of memory available for userspace allocations,
5755 * without causing swapping.
5757 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5760 * Not all the page cache can be freed, otherwise the system will
5761 * start swapping. Assume at least half of the page cache, or the
5762 * low watermark worth of cache, needs to stay.
5764 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5765 pagecache -= min(pagecache / 2, wmark_low);
5766 available += pagecache;
5769 * Part of the reclaimable slab and other kernel memory consists of
5770 * items that are in use, and cannot be freed. Cap this estimate at the
5773 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5774 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5775 available += reclaimable - min(reclaimable / 2, wmark_low);
5781 EXPORT_SYMBOL_GPL(si_mem_available);
5783 void si_meminfo(struct sysinfo *val)
5785 val->totalram = totalram_pages();
5786 val->sharedram = global_node_page_state(NR_SHMEM);
5787 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5788 val->bufferram = nr_blockdev_pages();
5789 val->totalhigh = totalhigh_pages();
5790 val->freehigh = nr_free_highpages();
5791 val->mem_unit = PAGE_SIZE;
5794 EXPORT_SYMBOL(si_meminfo);
5797 void si_meminfo_node(struct sysinfo *val, int nid)
5799 int zone_type; /* needs to be signed */
5800 unsigned long managed_pages = 0;
5801 unsigned long managed_highpages = 0;
5802 unsigned long free_highpages = 0;
5803 pg_data_t *pgdat = NODE_DATA(nid);
5805 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5806 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5807 val->totalram = managed_pages;
5808 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5809 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5810 #ifdef CONFIG_HIGHMEM
5811 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5812 struct zone *zone = &pgdat->node_zones[zone_type];
5814 if (is_highmem(zone)) {
5815 managed_highpages += zone_managed_pages(zone);
5816 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5819 val->totalhigh = managed_highpages;
5820 val->freehigh = free_highpages;
5822 val->totalhigh = managed_highpages;
5823 val->freehigh = free_highpages;
5825 val->mem_unit = PAGE_SIZE;
5830 * Determine whether the node should be displayed or not, depending on whether
5831 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5833 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5835 if (!(flags & SHOW_MEM_FILTER_NODES))
5839 * no node mask - aka implicit memory numa policy. Do not bother with
5840 * the synchronization - read_mems_allowed_begin - because we do not
5841 * have to be precise here.
5844 nodemask = &cpuset_current_mems_allowed;
5846 return !node_isset(nid, *nodemask);
5849 #define K(x) ((x) << (PAGE_SHIFT-10))
5851 static void show_migration_types(unsigned char type)
5853 static const char types[MIGRATE_TYPES] = {
5854 [MIGRATE_UNMOVABLE] = 'U',
5855 [MIGRATE_MOVABLE] = 'M',
5856 [MIGRATE_RECLAIMABLE] = 'E',
5857 [MIGRATE_HIGHATOMIC] = 'H',
5859 [MIGRATE_CMA] = 'C',
5861 #ifdef CONFIG_MEMORY_ISOLATION
5862 [MIGRATE_ISOLATE] = 'I',
5865 char tmp[MIGRATE_TYPES + 1];
5869 for (i = 0; i < MIGRATE_TYPES; i++) {
5870 if (type & (1 << i))
5875 printk(KERN_CONT "(%s) ", tmp);
5879 * Show free area list (used inside shift_scroll-lock stuff)
5880 * We also calculate the percentage fragmentation. We do this by counting the
5881 * memory on each free list with the exception of the first item on the list.
5884 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5887 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5889 unsigned long free_pcp = 0;
5894 for_each_populated_zone(zone) {
5895 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5898 for_each_online_cpu(cpu)
5899 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5902 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5903 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5904 " unevictable:%lu dirty:%lu writeback:%lu\n"
5905 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5906 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5907 " free:%lu free_pcp:%lu free_cma:%lu\n",
5908 global_node_page_state(NR_ACTIVE_ANON),
5909 global_node_page_state(NR_INACTIVE_ANON),
5910 global_node_page_state(NR_ISOLATED_ANON),
5911 global_node_page_state(NR_ACTIVE_FILE),
5912 global_node_page_state(NR_INACTIVE_FILE),
5913 global_node_page_state(NR_ISOLATED_FILE),
5914 global_node_page_state(NR_UNEVICTABLE),
5915 global_node_page_state(NR_FILE_DIRTY),
5916 global_node_page_state(NR_WRITEBACK),
5917 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5918 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5919 global_node_page_state(NR_FILE_MAPPED),
5920 global_node_page_state(NR_SHMEM),
5921 global_node_page_state(NR_PAGETABLE),
5922 global_zone_page_state(NR_BOUNCE),
5923 global_zone_page_state(NR_FREE_PAGES),
5925 global_zone_page_state(NR_FREE_CMA_PAGES));
5927 for_each_online_pgdat(pgdat) {
5928 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5932 " active_anon:%lukB"
5933 " inactive_anon:%lukB"
5934 " active_file:%lukB"
5935 " inactive_file:%lukB"
5936 " unevictable:%lukB"
5937 " isolated(anon):%lukB"
5938 " isolated(file):%lukB"
5943 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5945 " shmem_pmdmapped: %lukB"
5948 " writeback_tmp:%lukB"
5949 " kernel_stack:%lukB"
5950 #ifdef CONFIG_SHADOW_CALL_STACK
5951 " shadow_call_stack:%lukB"
5954 " all_unreclaimable? %s"
5957 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5958 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5959 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5960 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5961 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5962 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5963 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5964 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5965 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5966 K(node_page_state(pgdat, NR_WRITEBACK)),
5967 K(node_page_state(pgdat, NR_SHMEM)),
5968 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5969 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5970 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5971 K(node_page_state(pgdat, NR_ANON_THPS)),
5973 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5974 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5975 #ifdef CONFIG_SHADOW_CALL_STACK
5976 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5978 K(node_page_state(pgdat, NR_PAGETABLE)),
5979 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5983 for_each_populated_zone(zone) {
5986 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5990 for_each_online_cpu(cpu)
5991 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6000 " reserved_highatomic:%luKB"
6001 " active_anon:%lukB"
6002 " inactive_anon:%lukB"
6003 " active_file:%lukB"
6004 " inactive_file:%lukB"
6005 " unevictable:%lukB"
6006 " writepending:%lukB"
6016 K(zone_page_state(zone, NR_FREE_PAGES)),
6017 K(min_wmark_pages(zone)),
6018 K(low_wmark_pages(zone)),
6019 K(high_wmark_pages(zone)),
6020 K(zone->nr_reserved_highatomic),
6021 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6022 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6023 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6024 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6025 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6026 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6027 K(zone->present_pages),
6028 K(zone_managed_pages(zone)),
6029 K(zone_page_state(zone, NR_MLOCK)),
6030 K(zone_page_state(zone, NR_BOUNCE)),
6032 K(this_cpu_read(zone->per_cpu_pageset->count)),
6033 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6034 printk("lowmem_reserve[]:");
6035 for (i = 0; i < MAX_NR_ZONES; i++)
6036 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6037 printk(KERN_CONT "\n");
6040 for_each_populated_zone(zone) {
6042 unsigned long nr[MAX_ORDER], flags, total = 0;
6043 unsigned char types[MAX_ORDER];
6045 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6048 printk(KERN_CONT "%s: ", zone->name);
6050 spin_lock_irqsave(&zone->lock, flags);
6051 for (order = 0; order < MAX_ORDER; order++) {
6052 struct free_area *area = &zone->free_area[order];
6055 nr[order] = area->nr_free;
6056 total += nr[order] << order;
6059 for (type = 0; type < MIGRATE_TYPES; type++) {
6060 if (!free_area_empty(area, type))
6061 types[order] |= 1 << type;
6064 spin_unlock_irqrestore(&zone->lock, flags);
6065 for (order = 0; order < MAX_ORDER; order++) {
6066 printk(KERN_CONT "%lu*%lukB ",
6067 nr[order], K(1UL) << order);
6069 show_migration_types(types[order]);
6071 printk(KERN_CONT "= %lukB\n", K(total));
6074 hugetlb_show_meminfo();
6076 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6078 show_swap_cache_info();
6081 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6083 zoneref->zone = zone;
6084 zoneref->zone_idx = zone_idx(zone);
6088 * Builds allocation fallback zone lists.
6090 * Add all populated zones of a node to the zonelist.
6092 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6095 enum zone_type zone_type = MAX_NR_ZONES;
6100 zone = pgdat->node_zones + zone_type;
6101 if (managed_zone(zone)) {
6102 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6103 check_highest_zone(zone_type);
6105 } while (zone_type);
6112 static int __parse_numa_zonelist_order(char *s)
6115 * We used to support different zonelists modes but they turned
6116 * out to be just not useful. Let's keep the warning in place
6117 * if somebody still use the cmd line parameter so that we do
6118 * not fail it silently
6120 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6121 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6127 char numa_zonelist_order[] = "Node";
6130 * sysctl handler for numa_zonelist_order
6132 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6133 void *buffer, size_t *length, loff_t *ppos)
6136 return __parse_numa_zonelist_order(buffer);
6137 return proc_dostring(table, write, buffer, length, ppos);
6141 #define MAX_NODE_LOAD (nr_online_nodes)
6142 static int node_load[MAX_NUMNODES];
6145 * find_next_best_node - find the next node that should appear in a given node's fallback list
6146 * @node: node whose fallback list we're appending
6147 * @used_node_mask: nodemask_t of already used nodes
6149 * We use a number of factors to determine which is the next node that should
6150 * appear on a given node's fallback list. The node should not have appeared
6151 * already in @node's fallback list, and it should be the next closest node
6152 * according to the distance array (which contains arbitrary distance values
6153 * from each node to each node in the system), and should also prefer nodes
6154 * with no CPUs, since presumably they'll have very little allocation pressure
6155 * on them otherwise.
6157 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6159 static int find_next_best_node(int node, nodemask_t *used_node_mask)
6162 int min_val = INT_MAX;
6163 int best_node = NUMA_NO_NODE;
6165 /* Use the local node if we haven't already */
6166 if (!node_isset(node, *used_node_mask)) {
6167 node_set(node, *used_node_mask);
6171 for_each_node_state(n, N_MEMORY) {
6173 /* Don't want a node to appear more than once */
6174 if (node_isset(n, *used_node_mask))
6177 /* Use the distance array to find the distance */
6178 val = node_distance(node, n);
6180 /* Penalize nodes under us ("prefer the next node") */
6183 /* Give preference to headless and unused nodes */
6184 if (!cpumask_empty(cpumask_of_node(n)))
6185 val += PENALTY_FOR_NODE_WITH_CPUS;
6187 /* Slight preference for less loaded node */
6188 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6189 val += node_load[n];
6191 if (val < min_val) {
6198 node_set(best_node, *used_node_mask);
6205 * Build zonelists ordered by node and zones within node.
6206 * This results in maximum locality--normal zone overflows into local
6207 * DMA zone, if any--but risks exhausting DMA zone.
6209 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6212 struct zoneref *zonerefs;
6215 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6217 for (i = 0; i < nr_nodes; i++) {
6220 pg_data_t *node = NODE_DATA(node_order[i]);
6222 nr_zones = build_zonerefs_node(node, zonerefs);
6223 zonerefs += nr_zones;
6225 zonerefs->zone = NULL;
6226 zonerefs->zone_idx = 0;
6230 * Build gfp_thisnode zonelists
6232 static void build_thisnode_zonelists(pg_data_t *pgdat)
6234 struct zoneref *zonerefs;
6237 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6238 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6239 zonerefs += nr_zones;
6240 zonerefs->zone = NULL;
6241 zonerefs->zone_idx = 0;
6245 * Build zonelists ordered by zone and nodes within zones.
6246 * This results in conserving DMA zone[s] until all Normal memory is
6247 * exhausted, but results in overflowing to remote node while memory
6248 * may still exist in local DMA zone.
6251 static void build_zonelists(pg_data_t *pgdat)
6253 static int node_order[MAX_NUMNODES];
6254 int node, load, nr_nodes = 0;
6255 nodemask_t used_mask = NODE_MASK_NONE;
6256 int local_node, prev_node;
6258 /* NUMA-aware ordering of nodes */
6259 local_node = pgdat->node_id;
6260 load = nr_online_nodes;
6261 prev_node = local_node;
6263 memset(node_order, 0, sizeof(node_order));
6264 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6266 * We don't want to pressure a particular node.
6267 * So adding penalty to the first node in same
6268 * distance group to make it round-robin.
6270 if (node_distance(local_node, node) !=
6271 node_distance(local_node, prev_node))
6272 node_load[node] = load;
6274 node_order[nr_nodes++] = node;
6279 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6280 build_thisnode_zonelists(pgdat);
6283 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6285 * Return node id of node used for "local" allocations.
6286 * I.e., first node id of first zone in arg node's generic zonelist.
6287 * Used for initializing percpu 'numa_mem', which is used primarily
6288 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6290 int local_memory_node(int node)
6294 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6295 gfp_zone(GFP_KERNEL),
6297 return zone_to_nid(z->zone);
6301 static void setup_min_unmapped_ratio(void);
6302 static void setup_min_slab_ratio(void);
6303 #else /* CONFIG_NUMA */
6305 static void build_zonelists(pg_data_t *pgdat)
6307 int node, local_node;
6308 struct zoneref *zonerefs;
6311 local_node = pgdat->node_id;
6313 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6314 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6315 zonerefs += nr_zones;
6318 * Now we build the zonelist so that it contains the zones
6319 * of all the other nodes.
6320 * We don't want to pressure a particular node, so when
6321 * building the zones for node N, we make sure that the
6322 * zones coming right after the local ones are those from
6323 * node N+1 (modulo N)
6325 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6326 if (!node_online(node))
6328 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6329 zonerefs += nr_zones;
6331 for (node = 0; node < local_node; node++) {
6332 if (!node_online(node))
6334 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6335 zonerefs += nr_zones;
6338 zonerefs->zone = NULL;
6339 zonerefs->zone_idx = 0;
6342 #endif /* CONFIG_NUMA */
6345 * Boot pageset table. One per cpu which is going to be used for all
6346 * zones and all nodes. The parameters will be set in such a way
6347 * that an item put on a list will immediately be handed over to
6348 * the buddy list. This is safe since pageset manipulation is done
6349 * with interrupts disabled.
6351 * The boot_pagesets must be kept even after bootup is complete for
6352 * unused processors and/or zones. They do play a role for bootstrapping
6353 * hotplugged processors.
6355 * zoneinfo_show() and maybe other functions do
6356 * not check if the processor is online before following the pageset pointer.
6357 * Other parts of the kernel may not check if the zone is available.
6359 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6360 /* These effectively disable the pcplists in the boot pageset completely */
6361 #define BOOT_PAGESET_HIGH 0
6362 #define BOOT_PAGESET_BATCH 1
6363 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6364 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6365 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6367 static void __build_all_zonelists(void *data)
6370 int __maybe_unused cpu;
6371 pg_data_t *self = data;
6372 static DEFINE_SPINLOCK(lock);
6377 memset(node_load, 0, sizeof(node_load));
6381 * This node is hotadded and no memory is yet present. So just
6382 * building zonelists is fine - no need to touch other nodes.
6384 if (self && !node_online(self->node_id)) {
6385 build_zonelists(self);
6387 for_each_online_node(nid) {
6388 pg_data_t *pgdat = NODE_DATA(nid);
6390 build_zonelists(pgdat);
6393 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6395 * We now know the "local memory node" for each node--
6396 * i.e., the node of the first zone in the generic zonelist.
6397 * Set up numa_mem percpu variable for on-line cpus. During
6398 * boot, only the boot cpu should be on-line; we'll init the
6399 * secondary cpus' numa_mem as they come on-line. During
6400 * node/memory hotplug, we'll fixup all on-line cpus.
6402 for_each_online_cpu(cpu)
6403 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6410 static noinline void __init
6411 build_all_zonelists_init(void)
6415 __build_all_zonelists(NULL);
6418 * Initialize the boot_pagesets that are going to be used
6419 * for bootstrapping processors. The real pagesets for
6420 * each zone will be allocated later when the per cpu
6421 * allocator is available.
6423 * boot_pagesets are used also for bootstrapping offline
6424 * cpus if the system is already booted because the pagesets
6425 * are needed to initialize allocators on a specific cpu too.
6426 * F.e. the percpu allocator needs the page allocator which
6427 * needs the percpu allocator in order to allocate its pagesets
6428 * (a chicken-egg dilemma).
6430 for_each_possible_cpu(cpu)
6431 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6433 mminit_verify_zonelist();
6434 cpuset_init_current_mems_allowed();
6438 * unless system_state == SYSTEM_BOOTING.
6440 * __ref due to call of __init annotated helper build_all_zonelists_init
6441 * [protected by SYSTEM_BOOTING].
6443 void __ref build_all_zonelists(pg_data_t *pgdat)
6445 unsigned long vm_total_pages;
6447 if (system_state == SYSTEM_BOOTING) {
6448 build_all_zonelists_init();
6450 __build_all_zonelists(pgdat);
6451 /* cpuset refresh routine should be here */
6453 /* Get the number of free pages beyond high watermark in all zones. */
6454 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6456 * Disable grouping by mobility if the number of pages in the
6457 * system is too low to allow the mechanism to work. It would be
6458 * more accurate, but expensive to check per-zone. This check is
6459 * made on memory-hotadd so a system can start with mobility
6460 * disabled and enable it later
6462 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6463 page_group_by_mobility_disabled = 1;
6465 page_group_by_mobility_disabled = 0;
6467 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6469 page_group_by_mobility_disabled ? "off" : "on",
6472 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6476 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6477 static bool __meminit
6478 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6480 static struct memblock_region *r;
6482 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6483 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6484 for_each_mem_region(r) {
6485 if (*pfn < memblock_region_memory_end_pfn(r))
6489 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6490 memblock_is_mirror(r)) {
6491 *pfn = memblock_region_memory_end_pfn(r);
6499 * Initially all pages are reserved - free ones are freed
6500 * up by memblock_free_all() once the early boot process is
6501 * done. Non-atomic initialization, single-pass.
6503 * All aligned pageblocks are initialized to the specified migratetype
6504 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6505 * zone stats (e.g., nr_isolate_pageblock) are touched.
6507 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6508 unsigned long start_pfn, unsigned long zone_end_pfn,
6509 enum meminit_context context,
6510 struct vmem_altmap *altmap, int migratetype)
6512 unsigned long pfn, end_pfn = start_pfn + size;
6515 if (highest_memmap_pfn < end_pfn - 1)
6516 highest_memmap_pfn = end_pfn - 1;
6518 #ifdef CONFIG_ZONE_DEVICE
6520 * Honor reservation requested by the driver for this ZONE_DEVICE
6521 * memory. We limit the total number of pages to initialize to just
6522 * those that might contain the memory mapping. We will defer the
6523 * ZONE_DEVICE page initialization until after we have released
6526 if (zone == ZONE_DEVICE) {
6530 if (start_pfn == altmap->base_pfn)
6531 start_pfn += altmap->reserve;
6532 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6536 for (pfn = start_pfn; pfn < end_pfn; ) {
6538 * There can be holes in boot-time mem_map[]s handed to this
6539 * function. They do not exist on hotplugged memory.
6541 if (context == MEMINIT_EARLY) {
6542 if (overlap_memmap_init(zone, &pfn))
6544 if (defer_init(nid, pfn, zone_end_pfn))
6548 page = pfn_to_page(pfn);
6549 __init_single_page(page, pfn, zone, nid);
6550 if (context == MEMINIT_HOTPLUG)
6551 __SetPageReserved(page);
6554 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6555 * such that unmovable allocations won't be scattered all
6556 * over the place during system boot.
6558 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6559 set_pageblock_migratetype(page, migratetype);
6566 #ifdef CONFIG_ZONE_DEVICE
6567 void __ref memmap_init_zone_device(struct zone *zone,
6568 unsigned long start_pfn,
6569 unsigned long nr_pages,
6570 struct dev_pagemap *pgmap)
6572 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6573 struct pglist_data *pgdat = zone->zone_pgdat;
6574 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6575 unsigned long zone_idx = zone_idx(zone);
6576 unsigned long start = jiffies;
6577 int nid = pgdat->node_id;
6579 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6583 * The call to memmap_init should have already taken care
6584 * of the pages reserved for the memmap, so we can just jump to
6585 * the end of that region and start processing the device pages.
6588 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6589 nr_pages = end_pfn - start_pfn;
6592 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6593 struct page *page = pfn_to_page(pfn);
6595 __init_single_page(page, pfn, zone_idx, nid);
6598 * Mark page reserved as it will need to wait for onlining
6599 * phase for it to be fully associated with a zone.
6601 * We can use the non-atomic __set_bit operation for setting
6602 * the flag as we are still initializing the pages.
6604 __SetPageReserved(page);
6607 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6608 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6609 * ever freed or placed on a driver-private list.
6611 page->pgmap = pgmap;
6612 page->zone_device_data = NULL;
6615 * Mark the block movable so that blocks are reserved for
6616 * movable at startup. This will force kernel allocations
6617 * to reserve their blocks rather than leaking throughout
6618 * the address space during boot when many long-lived
6619 * kernel allocations are made.
6621 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6622 * because this is done early in section_activate()
6624 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6625 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6630 pr_info("%s initialised %lu pages in %ums\n", __func__,
6631 nr_pages, jiffies_to_msecs(jiffies - start));
6635 static void __meminit zone_init_free_lists(struct zone *zone)
6637 unsigned int order, t;
6638 for_each_migratetype_order(order, t) {
6639 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6640 zone->free_area[order].nr_free = 0;
6644 #if !defined(CONFIG_FLATMEM)
6646 * Only struct pages that correspond to ranges defined by memblock.memory
6647 * are zeroed and initialized by going through __init_single_page() during
6648 * memmap_init_zone_range().
6650 * But, there could be struct pages that correspond to holes in
6651 * memblock.memory. This can happen because of the following reasons:
6652 * - physical memory bank size is not necessarily the exact multiple of the
6653 * arbitrary section size
6654 * - early reserved memory may not be listed in memblock.memory
6655 * - memory layouts defined with memmap= kernel parameter may not align
6656 * nicely with memmap sections
6658 * Explicitly initialize those struct pages so that:
6659 * - PG_Reserved is set
6660 * - zone and node links point to zone and node that span the page if the
6661 * hole is in the middle of a zone
6662 * - zone and node links point to adjacent zone/node if the hole falls on
6663 * the zone boundary; the pages in such holes will be prepended to the
6664 * zone/node above the hole except for the trailing pages in the last
6665 * section that will be appended to the zone/node below.
6667 static void __init init_unavailable_range(unsigned long spfn,
6674 for (pfn = spfn; pfn < epfn; pfn++) {
6675 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6676 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6677 + pageblock_nr_pages - 1;
6680 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6681 __SetPageReserved(pfn_to_page(pfn));
6686 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6687 node, zone_names[zone], pgcnt);
6690 static inline void init_unavailable_range(unsigned long spfn,
6697 static void __init memmap_init_zone_range(struct zone *zone,
6698 unsigned long start_pfn,
6699 unsigned long end_pfn,
6700 unsigned long *hole_pfn)
6702 unsigned long zone_start_pfn = zone->zone_start_pfn;
6703 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6704 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6706 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6707 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6709 if (start_pfn >= end_pfn)
6712 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6713 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6715 if (*hole_pfn < start_pfn)
6716 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6718 *hole_pfn = end_pfn;
6721 static void __init memmap_init(void)
6723 unsigned long start_pfn, end_pfn;
6724 unsigned long hole_pfn = 0;
6725 int i, j, zone_id, nid;
6727 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6728 struct pglist_data *node = NODE_DATA(nid);
6730 for (j = 0; j < MAX_NR_ZONES; j++) {
6731 struct zone *zone = node->node_zones + j;
6733 if (!populated_zone(zone))
6736 memmap_init_zone_range(zone, start_pfn, end_pfn,
6742 #ifdef CONFIG_SPARSEMEM
6744 * Initialize the memory map for hole in the range [memory_end,
6746 * Append the pages in this hole to the highest zone in the last
6748 * The call to init_unavailable_range() is outside the ifdef to
6749 * silence the compiler warining about zone_id set but not used;
6750 * for FLATMEM it is a nop anyway
6752 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6753 if (hole_pfn < end_pfn)
6755 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6758 static int zone_batchsize(struct zone *zone)
6764 * The number of pages to batch allocate is either ~0.1%
6765 * of the zone or 1MB, whichever is smaller. The batch
6766 * size is striking a balance between allocation latency
6767 * and zone lock contention.
6769 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6770 batch /= 4; /* We effectively *= 4 below */
6775 * Clamp the batch to a 2^n - 1 value. Having a power
6776 * of 2 value was found to be more likely to have
6777 * suboptimal cache aliasing properties in some cases.
6779 * For example if 2 tasks are alternately allocating
6780 * batches of pages, one task can end up with a lot
6781 * of pages of one half of the possible page colors
6782 * and the other with pages of the other colors.
6784 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6789 /* The deferral and batching of frees should be suppressed under NOMMU
6792 * The problem is that NOMMU needs to be able to allocate large chunks
6793 * of contiguous memory as there's no hardware page translation to
6794 * assemble apparent contiguous memory from discontiguous pages.
6796 * Queueing large contiguous runs of pages for batching, however,
6797 * causes the pages to actually be freed in smaller chunks. As there
6798 * can be a significant delay between the individual batches being
6799 * recycled, this leads to the once large chunks of space being
6800 * fragmented and becoming unavailable for high-order allocations.
6806 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6811 unsigned long total_pages;
6813 if (!percpu_pagelist_high_fraction) {
6815 * By default, the high value of the pcp is based on the zone
6816 * low watermark so that if they are full then background
6817 * reclaim will not be started prematurely.
6819 total_pages = low_wmark_pages(zone);
6822 * If percpu_pagelist_high_fraction is configured, the high
6823 * value is based on a fraction of the managed pages in the
6826 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6830 * Split the high value across all online CPUs local to the zone. Note
6831 * that early in boot that CPUs may not be online yet and that during
6832 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6833 * onlined. For memory nodes that have no CPUs, split pcp->high across
6834 * all online CPUs to mitigate the risk that reclaim is triggered
6835 * prematurely due to pages stored on pcp lists.
6837 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6839 nr_split_cpus = num_online_cpus();
6840 high = total_pages / nr_split_cpus;
6843 * Ensure high is at least batch*4. The multiple is based on the
6844 * historical relationship between high and batch.
6846 high = max(high, batch << 2);
6855 * pcp->high and pcp->batch values are related and generally batch is lower
6856 * than high. They are also related to pcp->count such that count is lower
6857 * than high, and as soon as it reaches high, the pcplist is flushed.
6859 * However, guaranteeing these relations at all times would require e.g. write
6860 * barriers here but also careful usage of read barriers at the read side, and
6861 * thus be prone to error and bad for performance. Thus the update only prevents
6862 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6863 * can cope with those fields changing asynchronously, and fully trust only the
6864 * pcp->count field on the local CPU with interrupts disabled.
6866 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6867 * outside of boot time (or some other assurance that no concurrent updaters
6870 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6871 unsigned long batch)
6873 WRITE_ONCE(pcp->batch, batch);
6874 WRITE_ONCE(pcp->high, high);
6877 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6881 memset(pcp, 0, sizeof(*pcp));
6882 memset(pzstats, 0, sizeof(*pzstats));
6884 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6885 INIT_LIST_HEAD(&pcp->lists[pindex]);
6888 * Set batch and high values safe for a boot pageset. A true percpu
6889 * pageset's initialization will update them subsequently. Here we don't
6890 * need to be as careful as pageset_update() as nobody can access the
6893 pcp->high = BOOT_PAGESET_HIGH;
6894 pcp->batch = BOOT_PAGESET_BATCH;
6895 pcp->free_factor = 0;
6898 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6899 unsigned long batch)
6901 struct per_cpu_pages *pcp;
6904 for_each_possible_cpu(cpu) {
6905 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6906 pageset_update(pcp, high, batch);
6911 * Calculate and set new high and batch values for all per-cpu pagesets of a
6912 * zone based on the zone's size.
6914 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6916 int new_high, new_batch;
6918 new_batch = max(1, zone_batchsize(zone));
6919 new_high = zone_highsize(zone, new_batch, cpu_online);
6921 if (zone->pageset_high == new_high &&
6922 zone->pageset_batch == new_batch)
6925 zone->pageset_high = new_high;
6926 zone->pageset_batch = new_batch;
6928 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6931 void __meminit setup_zone_pageset(struct zone *zone)
6935 /* Size may be 0 on !SMP && !NUMA */
6936 if (sizeof(struct per_cpu_zonestat) > 0)
6937 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6939 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6940 for_each_possible_cpu(cpu) {
6941 struct per_cpu_pages *pcp;
6942 struct per_cpu_zonestat *pzstats;
6944 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6945 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6946 per_cpu_pages_init(pcp, pzstats);
6949 zone_set_pageset_high_and_batch(zone, 0);
6953 * Allocate per cpu pagesets and initialize them.
6954 * Before this call only boot pagesets were available.
6956 void __init setup_per_cpu_pageset(void)
6958 struct pglist_data *pgdat;
6960 int __maybe_unused cpu;
6962 for_each_populated_zone(zone)
6963 setup_zone_pageset(zone);
6967 * Unpopulated zones continue using the boot pagesets.
6968 * The numa stats for these pagesets need to be reset.
6969 * Otherwise, they will end up skewing the stats of
6970 * the nodes these zones are associated with.
6972 for_each_possible_cpu(cpu) {
6973 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
6974 memset(pzstats->vm_numa_event, 0,
6975 sizeof(pzstats->vm_numa_event));
6979 for_each_online_pgdat(pgdat)
6980 pgdat->per_cpu_nodestats =
6981 alloc_percpu(struct per_cpu_nodestat);
6984 static __meminit void zone_pcp_init(struct zone *zone)
6987 * per cpu subsystem is not up at this point. The following code
6988 * relies on the ability of the linker to provide the
6989 * offset of a (static) per cpu variable into the per cpu area.
6991 zone->per_cpu_pageset = &boot_pageset;
6992 zone->per_cpu_zonestats = &boot_zonestats;
6993 zone->pageset_high = BOOT_PAGESET_HIGH;
6994 zone->pageset_batch = BOOT_PAGESET_BATCH;
6996 if (populated_zone(zone))
6997 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
6998 zone->present_pages, zone_batchsize(zone));
7001 void __meminit init_currently_empty_zone(struct zone *zone,
7002 unsigned long zone_start_pfn,
7005 struct pglist_data *pgdat = zone->zone_pgdat;
7006 int zone_idx = zone_idx(zone) + 1;
7008 if (zone_idx > pgdat->nr_zones)
7009 pgdat->nr_zones = zone_idx;
7011 zone->zone_start_pfn = zone_start_pfn;
7013 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7014 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7016 (unsigned long)zone_idx(zone),
7017 zone_start_pfn, (zone_start_pfn + size));
7019 zone_init_free_lists(zone);
7020 zone->initialized = 1;
7024 * get_pfn_range_for_nid - Return the start and end page frames for a node
7025 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7026 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7027 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7029 * It returns the start and end page frame of a node based on information
7030 * provided by memblock_set_node(). If called for a node
7031 * with no available memory, a warning is printed and the start and end
7034 void __init get_pfn_range_for_nid(unsigned int nid,
7035 unsigned long *start_pfn, unsigned long *end_pfn)
7037 unsigned long this_start_pfn, this_end_pfn;
7043 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7044 *start_pfn = min(*start_pfn, this_start_pfn);
7045 *end_pfn = max(*end_pfn, this_end_pfn);
7048 if (*start_pfn == -1UL)
7053 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7054 * assumption is made that zones within a node are ordered in monotonic
7055 * increasing memory addresses so that the "highest" populated zone is used
7057 static void __init find_usable_zone_for_movable(void)
7060 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7061 if (zone_index == ZONE_MOVABLE)
7064 if (arch_zone_highest_possible_pfn[zone_index] >
7065 arch_zone_lowest_possible_pfn[zone_index])
7069 VM_BUG_ON(zone_index == -1);
7070 movable_zone = zone_index;
7074 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7075 * because it is sized independent of architecture. Unlike the other zones,
7076 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7077 * in each node depending on the size of each node and how evenly kernelcore
7078 * is distributed. This helper function adjusts the zone ranges
7079 * provided by the architecture for a given node by using the end of the
7080 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7081 * zones within a node are in order of monotonic increases memory addresses
7083 static void __init adjust_zone_range_for_zone_movable(int nid,
7084 unsigned long zone_type,
7085 unsigned long node_start_pfn,
7086 unsigned long node_end_pfn,
7087 unsigned long *zone_start_pfn,
7088 unsigned long *zone_end_pfn)
7090 /* Only adjust if ZONE_MOVABLE is on this node */
7091 if (zone_movable_pfn[nid]) {
7092 /* Size ZONE_MOVABLE */
7093 if (zone_type == ZONE_MOVABLE) {
7094 *zone_start_pfn = zone_movable_pfn[nid];
7095 *zone_end_pfn = min(node_end_pfn,
7096 arch_zone_highest_possible_pfn[movable_zone]);
7098 /* Adjust for ZONE_MOVABLE starting within this range */
7099 } else if (!mirrored_kernelcore &&
7100 *zone_start_pfn < zone_movable_pfn[nid] &&
7101 *zone_end_pfn > zone_movable_pfn[nid]) {
7102 *zone_end_pfn = zone_movable_pfn[nid];
7104 /* Check if this whole range is within ZONE_MOVABLE */
7105 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7106 *zone_start_pfn = *zone_end_pfn;
7111 * Return the number of pages a zone spans in a node, including holes
7112 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7114 static unsigned long __init zone_spanned_pages_in_node(int nid,
7115 unsigned long zone_type,
7116 unsigned long node_start_pfn,
7117 unsigned long node_end_pfn,
7118 unsigned long *zone_start_pfn,
7119 unsigned long *zone_end_pfn)
7121 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7122 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7123 /* When hotadd a new node from cpu_up(), the node should be empty */
7124 if (!node_start_pfn && !node_end_pfn)
7127 /* Get the start and end of the zone */
7128 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7129 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7130 adjust_zone_range_for_zone_movable(nid, zone_type,
7131 node_start_pfn, node_end_pfn,
7132 zone_start_pfn, zone_end_pfn);
7134 /* Check that this node has pages within the zone's required range */
7135 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7138 /* Move the zone boundaries inside the node if necessary */
7139 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7140 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7142 /* Return the spanned pages */
7143 return *zone_end_pfn - *zone_start_pfn;
7147 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7148 * then all holes in the requested range will be accounted for.
7150 unsigned long __init __absent_pages_in_range(int nid,
7151 unsigned long range_start_pfn,
7152 unsigned long range_end_pfn)
7154 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7155 unsigned long start_pfn, end_pfn;
7158 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7159 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7160 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7161 nr_absent -= end_pfn - start_pfn;
7167 * absent_pages_in_range - Return number of page frames in holes within a range
7168 * @start_pfn: The start PFN to start searching for holes
7169 * @end_pfn: The end PFN to stop searching for holes
7171 * Return: the number of pages frames in memory holes within a range.
7173 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7174 unsigned long end_pfn)
7176 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7179 /* Return the number of page frames in holes in a zone on a node */
7180 static unsigned long __init zone_absent_pages_in_node(int nid,
7181 unsigned long zone_type,
7182 unsigned long node_start_pfn,
7183 unsigned long node_end_pfn)
7185 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7186 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7187 unsigned long zone_start_pfn, zone_end_pfn;
7188 unsigned long nr_absent;
7190 /* When hotadd a new node from cpu_up(), the node should be empty */
7191 if (!node_start_pfn && !node_end_pfn)
7194 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7195 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7197 adjust_zone_range_for_zone_movable(nid, zone_type,
7198 node_start_pfn, node_end_pfn,
7199 &zone_start_pfn, &zone_end_pfn);
7200 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7203 * ZONE_MOVABLE handling.
7204 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7207 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7208 unsigned long start_pfn, end_pfn;
7209 struct memblock_region *r;
7211 for_each_mem_region(r) {
7212 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7213 zone_start_pfn, zone_end_pfn);
7214 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7215 zone_start_pfn, zone_end_pfn);
7217 if (zone_type == ZONE_MOVABLE &&
7218 memblock_is_mirror(r))
7219 nr_absent += end_pfn - start_pfn;
7221 if (zone_type == ZONE_NORMAL &&
7222 !memblock_is_mirror(r))
7223 nr_absent += end_pfn - start_pfn;
7230 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7231 unsigned long node_start_pfn,
7232 unsigned long node_end_pfn)
7234 unsigned long realtotalpages = 0, totalpages = 0;
7237 for (i = 0; i < MAX_NR_ZONES; i++) {
7238 struct zone *zone = pgdat->node_zones + i;
7239 unsigned long zone_start_pfn, zone_end_pfn;
7240 unsigned long spanned, absent;
7241 unsigned long size, real_size;
7243 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7248 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7253 real_size = size - absent;
7256 zone->zone_start_pfn = zone_start_pfn;
7258 zone->zone_start_pfn = 0;
7259 zone->spanned_pages = size;
7260 zone->present_pages = real_size;
7263 realtotalpages += real_size;
7266 pgdat->node_spanned_pages = totalpages;
7267 pgdat->node_present_pages = realtotalpages;
7268 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7271 #ifndef CONFIG_SPARSEMEM
7273 * Calculate the size of the zone->blockflags rounded to an unsigned long
7274 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7275 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7276 * round what is now in bits to nearest long in bits, then return it in
7279 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7281 unsigned long usemapsize;
7283 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7284 usemapsize = roundup(zonesize, pageblock_nr_pages);
7285 usemapsize = usemapsize >> pageblock_order;
7286 usemapsize *= NR_PAGEBLOCK_BITS;
7287 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7289 return usemapsize / 8;
7292 static void __ref setup_usemap(struct zone *zone)
7294 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7295 zone->spanned_pages);
7296 zone->pageblock_flags = NULL;
7298 zone->pageblock_flags =
7299 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7301 if (!zone->pageblock_flags)
7302 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7303 usemapsize, zone->name, zone_to_nid(zone));
7307 static inline void setup_usemap(struct zone *zone) {}
7308 #endif /* CONFIG_SPARSEMEM */
7310 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7312 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7313 void __init set_pageblock_order(void)
7317 /* Check that pageblock_nr_pages has not already been setup */
7318 if (pageblock_order)
7321 if (HPAGE_SHIFT > PAGE_SHIFT)
7322 order = HUGETLB_PAGE_ORDER;
7324 order = MAX_ORDER - 1;
7327 * Assume the largest contiguous order of interest is a huge page.
7328 * This value may be variable depending on boot parameters on IA64 and
7331 pageblock_order = order;
7333 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7336 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7337 * is unused as pageblock_order is set at compile-time. See
7338 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7341 void __init set_pageblock_order(void)
7345 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7347 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7348 unsigned long present_pages)
7350 unsigned long pages = spanned_pages;
7353 * Provide a more accurate estimation if there are holes within
7354 * the zone and SPARSEMEM is in use. If there are holes within the
7355 * zone, each populated memory region may cost us one or two extra
7356 * memmap pages due to alignment because memmap pages for each
7357 * populated regions may not be naturally aligned on page boundary.
7358 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7360 if (spanned_pages > present_pages + (present_pages >> 4) &&
7361 IS_ENABLED(CONFIG_SPARSEMEM))
7362 pages = present_pages;
7364 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7367 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7368 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7370 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7372 spin_lock_init(&ds_queue->split_queue_lock);
7373 INIT_LIST_HEAD(&ds_queue->split_queue);
7374 ds_queue->split_queue_len = 0;
7377 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7380 #ifdef CONFIG_COMPACTION
7381 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7383 init_waitqueue_head(&pgdat->kcompactd_wait);
7386 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7389 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7391 pgdat_resize_init(pgdat);
7393 pgdat_init_split_queue(pgdat);
7394 pgdat_init_kcompactd(pgdat);
7396 init_waitqueue_head(&pgdat->kswapd_wait);
7397 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7399 pgdat_page_ext_init(pgdat);
7400 lruvec_init(&pgdat->__lruvec);
7403 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7404 unsigned long remaining_pages)
7406 atomic_long_set(&zone->managed_pages, remaining_pages);
7407 zone_set_nid(zone, nid);
7408 zone->name = zone_names[idx];
7409 zone->zone_pgdat = NODE_DATA(nid);
7410 spin_lock_init(&zone->lock);
7411 zone_seqlock_init(zone);
7412 zone_pcp_init(zone);
7416 * Set up the zone data structures
7417 * - init pgdat internals
7418 * - init all zones belonging to this node
7420 * NOTE: this function is only called during memory hotplug
7422 #ifdef CONFIG_MEMORY_HOTPLUG
7423 void __ref free_area_init_core_hotplug(int nid)
7426 pg_data_t *pgdat = NODE_DATA(nid);
7428 pgdat_init_internals(pgdat);
7429 for (z = 0; z < MAX_NR_ZONES; z++)
7430 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7435 * Set up the zone data structures:
7436 * - mark all pages reserved
7437 * - mark all memory queues empty
7438 * - clear the memory bitmaps
7440 * NOTE: pgdat should get zeroed by caller.
7441 * NOTE: this function is only called during early init.
7443 static void __init free_area_init_core(struct pglist_data *pgdat)
7446 int nid = pgdat->node_id;
7448 pgdat_init_internals(pgdat);
7449 pgdat->per_cpu_nodestats = &boot_nodestats;
7451 for (j = 0; j < MAX_NR_ZONES; j++) {
7452 struct zone *zone = pgdat->node_zones + j;
7453 unsigned long size, freesize, memmap_pages;
7455 size = zone->spanned_pages;
7456 freesize = zone->present_pages;
7459 * Adjust freesize so that it accounts for how much memory
7460 * is used by this zone for memmap. This affects the watermark
7461 * and per-cpu initialisations
7463 memmap_pages = calc_memmap_size(size, freesize);
7464 if (!is_highmem_idx(j)) {
7465 if (freesize >= memmap_pages) {
7466 freesize -= memmap_pages;
7468 pr_debug(" %s zone: %lu pages used for memmap\n",
7469 zone_names[j], memmap_pages);
7471 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7472 zone_names[j], memmap_pages, freesize);
7475 /* Account for reserved pages */
7476 if (j == 0 && freesize > dma_reserve) {
7477 freesize -= dma_reserve;
7478 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7481 if (!is_highmem_idx(j))
7482 nr_kernel_pages += freesize;
7483 /* Charge for highmem memmap if there are enough kernel pages */
7484 else if (nr_kernel_pages > memmap_pages * 2)
7485 nr_kernel_pages -= memmap_pages;
7486 nr_all_pages += freesize;
7489 * Set an approximate value for lowmem here, it will be adjusted
7490 * when the bootmem allocator frees pages into the buddy system.
7491 * And all highmem pages will be managed by the buddy system.
7493 zone_init_internals(zone, j, nid, freesize);
7498 set_pageblock_order();
7500 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7504 #ifdef CONFIG_FLATMEM
7505 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7507 unsigned long __maybe_unused start = 0;
7508 unsigned long __maybe_unused offset = 0;
7510 /* Skip empty nodes */
7511 if (!pgdat->node_spanned_pages)
7514 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7515 offset = pgdat->node_start_pfn - start;
7516 /* ia64 gets its own node_mem_map, before this, without bootmem */
7517 if (!pgdat->node_mem_map) {
7518 unsigned long size, end;
7522 * The zone's endpoints aren't required to be MAX_ORDER
7523 * aligned but the node_mem_map endpoints must be in order
7524 * for the buddy allocator to function correctly.
7526 end = pgdat_end_pfn(pgdat);
7527 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7528 size = (end - start) * sizeof(struct page);
7529 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7532 panic("Failed to allocate %ld bytes for node %d memory map\n",
7533 size, pgdat->node_id);
7534 pgdat->node_mem_map = map + offset;
7536 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7537 __func__, pgdat->node_id, (unsigned long)pgdat,
7538 (unsigned long)pgdat->node_mem_map);
7541 * With no DISCONTIG, the global mem_map is just set as node 0's
7543 if (pgdat == NODE_DATA(0)) {
7544 mem_map = NODE_DATA(0)->node_mem_map;
7545 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7551 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7552 #endif /* CONFIG_FLATMEM */
7554 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7555 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7557 pgdat->first_deferred_pfn = ULONG_MAX;
7560 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7563 static void __init free_area_init_node(int nid)
7565 pg_data_t *pgdat = NODE_DATA(nid);
7566 unsigned long start_pfn = 0;
7567 unsigned long end_pfn = 0;
7569 /* pg_data_t should be reset to zero when it's allocated */
7570 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7572 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7574 pgdat->node_id = nid;
7575 pgdat->node_start_pfn = start_pfn;
7576 pgdat->per_cpu_nodestats = NULL;
7578 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7579 (u64)start_pfn << PAGE_SHIFT,
7580 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7581 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7583 alloc_node_mem_map(pgdat);
7584 pgdat_set_deferred_range(pgdat);
7586 free_area_init_core(pgdat);
7589 void __init free_area_init_memoryless_node(int nid)
7591 free_area_init_node(nid);
7594 #if MAX_NUMNODES > 1
7596 * Figure out the number of possible node ids.
7598 void __init setup_nr_node_ids(void)
7600 unsigned int highest;
7602 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7603 nr_node_ids = highest + 1;
7608 * node_map_pfn_alignment - determine the maximum internode alignment
7610 * This function should be called after node map is populated and sorted.
7611 * It calculates the maximum power of two alignment which can distinguish
7614 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7615 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7616 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7617 * shifted, 1GiB is enough and this function will indicate so.
7619 * This is used to test whether pfn -> nid mapping of the chosen memory
7620 * model has fine enough granularity to avoid incorrect mapping for the
7621 * populated node map.
7623 * Return: the determined alignment in pfn's. 0 if there is no alignment
7624 * requirement (single node).
7626 unsigned long __init node_map_pfn_alignment(void)
7628 unsigned long accl_mask = 0, last_end = 0;
7629 unsigned long start, end, mask;
7630 int last_nid = NUMA_NO_NODE;
7633 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7634 if (!start || last_nid < 0 || last_nid == nid) {
7641 * Start with a mask granular enough to pin-point to the
7642 * start pfn and tick off bits one-by-one until it becomes
7643 * too coarse to separate the current node from the last.
7645 mask = ~((1 << __ffs(start)) - 1);
7646 while (mask && last_end <= (start & (mask << 1)))
7649 /* accumulate all internode masks */
7653 /* convert mask to number of pages */
7654 return ~accl_mask + 1;
7658 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7660 * Return: the minimum PFN based on information provided via
7661 * memblock_set_node().
7663 unsigned long __init find_min_pfn_with_active_regions(void)
7665 return PHYS_PFN(memblock_start_of_DRAM());
7669 * early_calculate_totalpages()
7670 * Sum pages in active regions for movable zone.
7671 * Populate N_MEMORY for calculating usable_nodes.
7673 static unsigned long __init early_calculate_totalpages(void)
7675 unsigned long totalpages = 0;
7676 unsigned long start_pfn, end_pfn;
7679 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7680 unsigned long pages = end_pfn - start_pfn;
7682 totalpages += pages;
7684 node_set_state(nid, N_MEMORY);
7690 * Find the PFN the Movable zone begins in each node. Kernel memory
7691 * is spread evenly between nodes as long as the nodes have enough
7692 * memory. When they don't, some nodes will have more kernelcore than
7695 static void __init find_zone_movable_pfns_for_nodes(void)
7698 unsigned long usable_startpfn;
7699 unsigned long kernelcore_node, kernelcore_remaining;
7700 /* save the state before borrow the nodemask */
7701 nodemask_t saved_node_state = node_states[N_MEMORY];
7702 unsigned long totalpages = early_calculate_totalpages();
7703 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7704 struct memblock_region *r;
7706 /* Need to find movable_zone earlier when movable_node is specified. */
7707 find_usable_zone_for_movable();
7710 * If movable_node is specified, ignore kernelcore and movablecore
7713 if (movable_node_is_enabled()) {
7714 for_each_mem_region(r) {
7715 if (!memblock_is_hotpluggable(r))
7718 nid = memblock_get_region_node(r);
7720 usable_startpfn = PFN_DOWN(r->base);
7721 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7722 min(usable_startpfn, zone_movable_pfn[nid]) :
7730 * If kernelcore=mirror is specified, ignore movablecore option
7732 if (mirrored_kernelcore) {
7733 bool mem_below_4gb_not_mirrored = false;
7735 for_each_mem_region(r) {
7736 if (memblock_is_mirror(r))
7739 nid = memblock_get_region_node(r);
7741 usable_startpfn = memblock_region_memory_base_pfn(r);
7743 if (usable_startpfn < 0x100000) {
7744 mem_below_4gb_not_mirrored = true;
7748 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7749 min(usable_startpfn, zone_movable_pfn[nid]) :
7753 if (mem_below_4gb_not_mirrored)
7754 pr_warn("This configuration results in unmirrored kernel memory.\n");
7760 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7761 * amount of necessary memory.
7763 if (required_kernelcore_percent)
7764 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7766 if (required_movablecore_percent)
7767 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7771 * If movablecore= was specified, calculate what size of
7772 * kernelcore that corresponds so that memory usable for
7773 * any allocation type is evenly spread. If both kernelcore
7774 * and movablecore are specified, then the value of kernelcore
7775 * will be used for required_kernelcore if it's greater than
7776 * what movablecore would have allowed.
7778 if (required_movablecore) {
7779 unsigned long corepages;
7782 * Round-up so that ZONE_MOVABLE is at least as large as what
7783 * was requested by the user
7785 required_movablecore =
7786 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7787 required_movablecore = min(totalpages, required_movablecore);
7788 corepages = totalpages - required_movablecore;
7790 required_kernelcore = max(required_kernelcore, corepages);
7794 * If kernelcore was not specified or kernelcore size is larger
7795 * than totalpages, there is no ZONE_MOVABLE.
7797 if (!required_kernelcore || required_kernelcore >= totalpages)
7800 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7801 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7804 /* Spread kernelcore memory as evenly as possible throughout nodes */
7805 kernelcore_node = required_kernelcore / usable_nodes;
7806 for_each_node_state(nid, N_MEMORY) {
7807 unsigned long start_pfn, end_pfn;
7810 * Recalculate kernelcore_node if the division per node
7811 * now exceeds what is necessary to satisfy the requested
7812 * amount of memory for the kernel
7814 if (required_kernelcore < kernelcore_node)
7815 kernelcore_node = required_kernelcore / usable_nodes;
7818 * As the map is walked, we track how much memory is usable
7819 * by the kernel using kernelcore_remaining. When it is
7820 * 0, the rest of the node is usable by ZONE_MOVABLE
7822 kernelcore_remaining = kernelcore_node;
7824 /* Go through each range of PFNs within this node */
7825 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7826 unsigned long size_pages;
7828 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7829 if (start_pfn >= end_pfn)
7832 /* Account for what is only usable for kernelcore */
7833 if (start_pfn < usable_startpfn) {
7834 unsigned long kernel_pages;
7835 kernel_pages = min(end_pfn, usable_startpfn)
7838 kernelcore_remaining -= min(kernel_pages,
7839 kernelcore_remaining);
7840 required_kernelcore -= min(kernel_pages,
7841 required_kernelcore);
7843 /* Continue if range is now fully accounted */
7844 if (end_pfn <= usable_startpfn) {
7847 * Push zone_movable_pfn to the end so
7848 * that if we have to rebalance
7849 * kernelcore across nodes, we will
7850 * not double account here
7852 zone_movable_pfn[nid] = end_pfn;
7855 start_pfn = usable_startpfn;
7859 * The usable PFN range for ZONE_MOVABLE is from
7860 * start_pfn->end_pfn. Calculate size_pages as the
7861 * number of pages used as kernelcore
7863 size_pages = end_pfn - start_pfn;
7864 if (size_pages > kernelcore_remaining)
7865 size_pages = kernelcore_remaining;
7866 zone_movable_pfn[nid] = start_pfn + size_pages;
7869 * Some kernelcore has been met, update counts and
7870 * break if the kernelcore for this node has been
7873 required_kernelcore -= min(required_kernelcore,
7875 kernelcore_remaining -= size_pages;
7876 if (!kernelcore_remaining)
7882 * If there is still required_kernelcore, we do another pass with one
7883 * less node in the count. This will push zone_movable_pfn[nid] further
7884 * along on the nodes that still have memory until kernelcore is
7888 if (usable_nodes && required_kernelcore > usable_nodes)
7892 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7893 for (nid = 0; nid < MAX_NUMNODES; nid++)
7894 zone_movable_pfn[nid] =
7895 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7898 /* restore the node_state */
7899 node_states[N_MEMORY] = saved_node_state;
7902 /* Any regular or high memory on that node ? */
7903 static void check_for_memory(pg_data_t *pgdat, int nid)
7905 enum zone_type zone_type;
7907 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7908 struct zone *zone = &pgdat->node_zones[zone_type];
7909 if (populated_zone(zone)) {
7910 if (IS_ENABLED(CONFIG_HIGHMEM))
7911 node_set_state(nid, N_HIGH_MEMORY);
7912 if (zone_type <= ZONE_NORMAL)
7913 node_set_state(nid, N_NORMAL_MEMORY);
7920 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7921 * such cases we allow max_zone_pfn sorted in the descending order
7923 bool __weak arch_has_descending_max_zone_pfns(void)
7929 * free_area_init - Initialise all pg_data_t and zone data
7930 * @max_zone_pfn: an array of max PFNs for each zone
7932 * This will call free_area_init_node() for each active node in the system.
7933 * Using the page ranges provided by memblock_set_node(), the size of each
7934 * zone in each node and their holes is calculated. If the maximum PFN
7935 * between two adjacent zones match, it is assumed that the zone is empty.
7936 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7937 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7938 * starts where the previous one ended. For example, ZONE_DMA32 starts
7939 * at arch_max_dma_pfn.
7941 void __init free_area_init(unsigned long *max_zone_pfn)
7943 unsigned long start_pfn, end_pfn;
7947 /* Record where the zone boundaries are */
7948 memset(arch_zone_lowest_possible_pfn, 0,
7949 sizeof(arch_zone_lowest_possible_pfn));
7950 memset(arch_zone_highest_possible_pfn, 0,
7951 sizeof(arch_zone_highest_possible_pfn));
7953 start_pfn = find_min_pfn_with_active_regions();
7954 descending = arch_has_descending_max_zone_pfns();
7956 for (i = 0; i < MAX_NR_ZONES; i++) {
7958 zone = MAX_NR_ZONES - i - 1;
7962 if (zone == ZONE_MOVABLE)
7965 end_pfn = max(max_zone_pfn[zone], start_pfn);
7966 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7967 arch_zone_highest_possible_pfn[zone] = end_pfn;
7969 start_pfn = end_pfn;
7972 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7973 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7974 find_zone_movable_pfns_for_nodes();
7976 /* Print out the zone ranges */
7977 pr_info("Zone ranges:\n");
7978 for (i = 0; i < MAX_NR_ZONES; i++) {
7979 if (i == ZONE_MOVABLE)
7981 pr_info(" %-8s ", zone_names[i]);
7982 if (arch_zone_lowest_possible_pfn[i] ==
7983 arch_zone_highest_possible_pfn[i])
7986 pr_cont("[mem %#018Lx-%#018Lx]\n",
7987 (u64)arch_zone_lowest_possible_pfn[i]
7989 ((u64)arch_zone_highest_possible_pfn[i]
7990 << PAGE_SHIFT) - 1);
7993 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7994 pr_info("Movable zone start for each node\n");
7995 for (i = 0; i < MAX_NUMNODES; i++) {
7996 if (zone_movable_pfn[i])
7997 pr_info(" Node %d: %#018Lx\n", i,
7998 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8002 * Print out the early node map, and initialize the
8003 * subsection-map relative to active online memory ranges to
8004 * enable future "sub-section" extensions of the memory map.
8006 pr_info("Early memory node ranges\n");
8007 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8008 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8009 (u64)start_pfn << PAGE_SHIFT,
8010 ((u64)end_pfn << PAGE_SHIFT) - 1);
8011 subsection_map_init(start_pfn, end_pfn - start_pfn);
8014 /* Initialise every node */
8015 mminit_verify_pageflags_layout();
8016 setup_nr_node_ids();
8017 for_each_online_node(nid) {
8018 pg_data_t *pgdat = NODE_DATA(nid);
8019 free_area_init_node(nid);
8021 /* Any memory on that node */
8022 if (pgdat->node_present_pages)
8023 node_set_state(nid, N_MEMORY);
8024 check_for_memory(pgdat, nid);
8030 static int __init cmdline_parse_core(char *p, unsigned long *core,
8031 unsigned long *percent)
8033 unsigned long long coremem;
8039 /* Value may be a percentage of total memory, otherwise bytes */
8040 coremem = simple_strtoull(p, &endptr, 0);
8041 if (*endptr == '%') {
8042 /* Paranoid check for percent values greater than 100 */
8043 WARN_ON(coremem > 100);
8047 coremem = memparse(p, &p);
8048 /* Paranoid check that UL is enough for the coremem value */
8049 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8051 *core = coremem >> PAGE_SHIFT;
8058 * kernelcore=size sets the amount of memory for use for allocations that
8059 * cannot be reclaimed or migrated.
8061 static int __init cmdline_parse_kernelcore(char *p)
8063 /* parse kernelcore=mirror */
8064 if (parse_option_str(p, "mirror")) {
8065 mirrored_kernelcore = true;
8069 return cmdline_parse_core(p, &required_kernelcore,
8070 &required_kernelcore_percent);
8074 * movablecore=size sets the amount of memory for use for allocations that
8075 * can be reclaimed or migrated.
8077 static int __init cmdline_parse_movablecore(char *p)
8079 return cmdline_parse_core(p, &required_movablecore,
8080 &required_movablecore_percent);
8083 early_param("kernelcore", cmdline_parse_kernelcore);
8084 early_param("movablecore", cmdline_parse_movablecore);
8086 void adjust_managed_page_count(struct page *page, long count)
8088 atomic_long_add(count, &page_zone(page)->managed_pages);
8089 totalram_pages_add(count);
8090 #ifdef CONFIG_HIGHMEM
8091 if (PageHighMem(page))
8092 totalhigh_pages_add(count);
8095 EXPORT_SYMBOL(adjust_managed_page_count);
8097 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8100 unsigned long pages = 0;
8102 start = (void *)PAGE_ALIGN((unsigned long)start);
8103 end = (void *)((unsigned long)end & PAGE_MASK);
8104 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8105 struct page *page = virt_to_page(pos);
8106 void *direct_map_addr;
8109 * 'direct_map_addr' might be different from 'pos'
8110 * because some architectures' virt_to_page()
8111 * work with aliases. Getting the direct map
8112 * address ensures that we get a _writeable_
8113 * alias for the memset().
8115 direct_map_addr = page_address(page);
8117 * Perform a kasan-unchecked memset() since this memory
8118 * has not been initialized.
8120 direct_map_addr = kasan_reset_tag(direct_map_addr);
8121 if ((unsigned int)poison <= 0xFF)
8122 memset(direct_map_addr, poison, PAGE_SIZE);
8124 free_reserved_page(page);
8128 pr_info("Freeing %s memory: %ldK\n",
8129 s, pages << (PAGE_SHIFT - 10));
8134 void __init mem_init_print_info(void)
8136 unsigned long physpages, codesize, datasize, rosize, bss_size;
8137 unsigned long init_code_size, init_data_size;
8139 physpages = get_num_physpages();
8140 codesize = _etext - _stext;
8141 datasize = _edata - _sdata;
8142 rosize = __end_rodata - __start_rodata;
8143 bss_size = __bss_stop - __bss_start;
8144 init_data_size = __init_end - __init_begin;
8145 init_code_size = _einittext - _sinittext;
8148 * Detect special cases and adjust section sizes accordingly:
8149 * 1) .init.* may be embedded into .data sections
8150 * 2) .init.text.* may be out of [__init_begin, __init_end],
8151 * please refer to arch/tile/kernel/vmlinux.lds.S.
8152 * 3) .rodata.* may be embedded into .text or .data sections.
8154 #define adj_init_size(start, end, size, pos, adj) \
8156 if (start <= pos && pos < end && size > adj) \
8160 adj_init_size(__init_begin, __init_end, init_data_size,
8161 _sinittext, init_code_size);
8162 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8163 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8164 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8165 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8167 #undef adj_init_size
8169 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8170 #ifdef CONFIG_HIGHMEM
8174 nr_free_pages() << (PAGE_SHIFT - 10),
8175 physpages << (PAGE_SHIFT - 10),
8176 codesize >> 10, datasize >> 10, rosize >> 10,
8177 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8178 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
8179 totalcma_pages << (PAGE_SHIFT - 10)
8180 #ifdef CONFIG_HIGHMEM
8181 , totalhigh_pages() << (PAGE_SHIFT - 10)
8187 * set_dma_reserve - set the specified number of pages reserved in the first zone
8188 * @new_dma_reserve: The number of pages to mark reserved
8190 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8191 * In the DMA zone, a significant percentage may be consumed by kernel image
8192 * and other unfreeable allocations which can skew the watermarks badly. This
8193 * function may optionally be used to account for unfreeable pages in the
8194 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8195 * smaller per-cpu batchsize.
8197 void __init set_dma_reserve(unsigned long new_dma_reserve)
8199 dma_reserve = new_dma_reserve;
8202 static int page_alloc_cpu_dead(unsigned int cpu)
8206 lru_add_drain_cpu(cpu);
8210 * Spill the event counters of the dead processor
8211 * into the current processors event counters.
8212 * This artificially elevates the count of the current
8215 vm_events_fold_cpu(cpu);
8218 * Zero the differential counters of the dead processor
8219 * so that the vm statistics are consistent.
8221 * This is only okay since the processor is dead and cannot
8222 * race with what we are doing.
8224 cpu_vm_stats_fold(cpu);
8226 for_each_populated_zone(zone)
8227 zone_pcp_update(zone, 0);
8232 static int page_alloc_cpu_online(unsigned int cpu)
8236 for_each_populated_zone(zone)
8237 zone_pcp_update(zone, 1);
8242 int hashdist = HASHDIST_DEFAULT;
8244 static int __init set_hashdist(char *str)
8248 hashdist = simple_strtoul(str, &str, 0);
8251 __setup("hashdist=", set_hashdist);
8254 void __init page_alloc_init(void)
8259 if (num_node_state(N_MEMORY) == 1)
8263 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8264 "mm/page_alloc:pcp",
8265 page_alloc_cpu_online,
8266 page_alloc_cpu_dead);
8271 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8272 * or min_free_kbytes changes.
8274 static void calculate_totalreserve_pages(void)
8276 struct pglist_data *pgdat;
8277 unsigned long reserve_pages = 0;
8278 enum zone_type i, j;
8280 for_each_online_pgdat(pgdat) {
8282 pgdat->totalreserve_pages = 0;
8284 for (i = 0; i < MAX_NR_ZONES; i++) {
8285 struct zone *zone = pgdat->node_zones + i;
8287 unsigned long managed_pages = zone_managed_pages(zone);
8289 /* Find valid and maximum lowmem_reserve in the zone */
8290 for (j = i; j < MAX_NR_ZONES; j++) {
8291 if (zone->lowmem_reserve[j] > max)
8292 max = zone->lowmem_reserve[j];
8295 /* we treat the high watermark as reserved pages. */
8296 max += high_wmark_pages(zone);
8298 if (max > managed_pages)
8299 max = managed_pages;
8301 pgdat->totalreserve_pages += max;
8303 reserve_pages += max;
8306 totalreserve_pages = reserve_pages;
8310 * setup_per_zone_lowmem_reserve - called whenever
8311 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8312 * has a correct pages reserved value, so an adequate number of
8313 * pages are left in the zone after a successful __alloc_pages().
8315 static void setup_per_zone_lowmem_reserve(void)
8317 struct pglist_data *pgdat;
8318 enum zone_type i, j;
8320 for_each_online_pgdat(pgdat) {
8321 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8322 struct zone *zone = &pgdat->node_zones[i];
8323 int ratio = sysctl_lowmem_reserve_ratio[i];
8324 bool clear = !ratio || !zone_managed_pages(zone);
8325 unsigned long managed_pages = 0;
8327 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8328 struct zone *upper_zone = &pgdat->node_zones[j];
8330 managed_pages += zone_managed_pages(upper_zone);
8333 zone->lowmem_reserve[j] = 0;
8335 zone->lowmem_reserve[j] = managed_pages / ratio;
8340 /* update totalreserve_pages */
8341 calculate_totalreserve_pages();
8344 static void __setup_per_zone_wmarks(void)
8346 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8347 unsigned long lowmem_pages = 0;
8349 unsigned long flags;
8351 /* Calculate total number of !ZONE_HIGHMEM pages */
8352 for_each_zone(zone) {
8353 if (!is_highmem(zone))
8354 lowmem_pages += zone_managed_pages(zone);
8357 for_each_zone(zone) {
8360 spin_lock_irqsave(&zone->lock, flags);
8361 tmp = (u64)pages_min * zone_managed_pages(zone);
8362 do_div(tmp, lowmem_pages);
8363 if (is_highmem(zone)) {
8365 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8366 * need highmem pages, so cap pages_min to a small
8369 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8370 * deltas control async page reclaim, and so should
8371 * not be capped for highmem.
8373 unsigned long min_pages;
8375 min_pages = zone_managed_pages(zone) / 1024;
8376 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8377 zone->_watermark[WMARK_MIN] = min_pages;
8380 * If it's a lowmem zone, reserve a number of pages
8381 * proportionate to the zone's size.
8383 zone->_watermark[WMARK_MIN] = tmp;
8387 * Set the kswapd watermarks distance according to the
8388 * scale factor in proportion to available memory, but
8389 * ensure a minimum size on small systems.
8391 tmp = max_t(u64, tmp >> 2,
8392 mult_frac(zone_managed_pages(zone),
8393 watermark_scale_factor, 10000));
8395 zone->watermark_boost = 0;
8396 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8397 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8399 spin_unlock_irqrestore(&zone->lock, flags);
8402 /* update totalreserve_pages */
8403 calculate_totalreserve_pages();
8407 * setup_per_zone_wmarks - called when min_free_kbytes changes
8408 * or when memory is hot-{added|removed}
8410 * Ensures that the watermark[min,low,high] values for each zone are set
8411 * correctly with respect to min_free_kbytes.
8413 void setup_per_zone_wmarks(void)
8416 static DEFINE_SPINLOCK(lock);
8419 __setup_per_zone_wmarks();
8423 * The watermark size have changed so update the pcpu batch
8424 * and high limits or the limits may be inappropriate.
8427 zone_pcp_update(zone, 0);
8431 * Initialise min_free_kbytes.
8433 * For small machines we want it small (128k min). For large machines
8434 * we want it large (256MB max). But it is not linear, because network
8435 * bandwidth does not increase linearly with machine size. We use
8437 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8438 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8454 int __meminit init_per_zone_wmark_min(void)
8456 unsigned long lowmem_kbytes;
8457 int new_min_free_kbytes;
8459 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8460 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8462 if (new_min_free_kbytes > user_min_free_kbytes) {
8463 min_free_kbytes = new_min_free_kbytes;
8464 if (min_free_kbytes < 128)
8465 min_free_kbytes = 128;
8466 if (min_free_kbytes > 262144)
8467 min_free_kbytes = 262144;
8469 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8470 new_min_free_kbytes, user_min_free_kbytes);
8472 setup_per_zone_wmarks();
8473 refresh_zone_stat_thresholds();
8474 setup_per_zone_lowmem_reserve();
8477 setup_min_unmapped_ratio();
8478 setup_min_slab_ratio();
8481 khugepaged_min_free_kbytes_update();
8485 postcore_initcall(init_per_zone_wmark_min)
8488 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8489 * that we can call two helper functions whenever min_free_kbytes
8492 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8493 void *buffer, size_t *length, loff_t *ppos)
8497 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8502 user_min_free_kbytes = min_free_kbytes;
8503 setup_per_zone_wmarks();
8508 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8509 void *buffer, size_t *length, loff_t *ppos)
8513 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8518 setup_per_zone_wmarks();
8524 static void setup_min_unmapped_ratio(void)
8529 for_each_online_pgdat(pgdat)
8530 pgdat->min_unmapped_pages = 0;
8533 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8534 sysctl_min_unmapped_ratio) / 100;
8538 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8539 void *buffer, size_t *length, loff_t *ppos)
8543 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8547 setup_min_unmapped_ratio();
8552 static void setup_min_slab_ratio(void)
8557 for_each_online_pgdat(pgdat)
8558 pgdat->min_slab_pages = 0;
8561 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8562 sysctl_min_slab_ratio) / 100;
8565 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8566 void *buffer, size_t *length, loff_t *ppos)
8570 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8574 setup_min_slab_ratio();
8581 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8582 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8583 * whenever sysctl_lowmem_reserve_ratio changes.
8585 * The reserve ratio obviously has absolutely no relation with the
8586 * minimum watermarks. The lowmem reserve ratio can only make sense
8587 * if in function of the boot time zone sizes.
8589 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8590 void *buffer, size_t *length, loff_t *ppos)
8594 proc_dointvec_minmax(table, write, buffer, length, ppos);
8596 for (i = 0; i < MAX_NR_ZONES; i++) {
8597 if (sysctl_lowmem_reserve_ratio[i] < 1)
8598 sysctl_lowmem_reserve_ratio[i] = 0;
8601 setup_per_zone_lowmem_reserve();
8606 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8607 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8608 * pagelist can have before it gets flushed back to buddy allocator.
8610 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8611 int write, void *buffer, size_t *length, loff_t *ppos)
8614 int old_percpu_pagelist_high_fraction;
8617 mutex_lock(&pcp_batch_high_lock);
8618 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8620 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8621 if (!write || ret < 0)
8624 /* Sanity checking to avoid pcp imbalance */
8625 if (percpu_pagelist_high_fraction &&
8626 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8627 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8633 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8636 for_each_populated_zone(zone)
8637 zone_set_pageset_high_and_batch(zone, 0);
8639 mutex_unlock(&pcp_batch_high_lock);
8643 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8645 * Returns the number of pages that arch has reserved but
8646 * is not known to alloc_large_system_hash().
8648 static unsigned long __init arch_reserved_kernel_pages(void)
8655 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8656 * machines. As memory size is increased the scale is also increased but at
8657 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8658 * quadruples the scale is increased by one, which means the size of hash table
8659 * only doubles, instead of quadrupling as well.
8660 * Because 32-bit systems cannot have large physical memory, where this scaling
8661 * makes sense, it is disabled on such platforms.
8663 #if __BITS_PER_LONG > 32
8664 #define ADAPT_SCALE_BASE (64ul << 30)
8665 #define ADAPT_SCALE_SHIFT 2
8666 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8670 * allocate a large system hash table from bootmem
8671 * - it is assumed that the hash table must contain an exact power-of-2
8672 * quantity of entries
8673 * - limit is the number of hash buckets, not the total allocation size
8675 void *__init alloc_large_system_hash(const char *tablename,
8676 unsigned long bucketsize,
8677 unsigned long numentries,
8680 unsigned int *_hash_shift,
8681 unsigned int *_hash_mask,
8682 unsigned long low_limit,
8683 unsigned long high_limit)
8685 unsigned long long max = high_limit;
8686 unsigned long log2qty, size;
8692 /* allow the kernel cmdline to have a say */
8694 /* round applicable memory size up to nearest megabyte */
8695 numentries = nr_kernel_pages;
8696 numentries -= arch_reserved_kernel_pages();
8698 /* It isn't necessary when PAGE_SIZE >= 1MB */
8699 if (PAGE_SHIFT < 20)
8700 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8702 #if __BITS_PER_LONG > 32
8704 unsigned long adapt;
8706 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8707 adapt <<= ADAPT_SCALE_SHIFT)
8712 /* limit to 1 bucket per 2^scale bytes of low memory */
8713 if (scale > PAGE_SHIFT)
8714 numentries >>= (scale - PAGE_SHIFT);
8716 numentries <<= (PAGE_SHIFT - scale);
8718 /* Make sure we've got at least a 0-order allocation.. */
8719 if (unlikely(flags & HASH_SMALL)) {
8720 /* Makes no sense without HASH_EARLY */
8721 WARN_ON(!(flags & HASH_EARLY));
8722 if (!(numentries >> *_hash_shift)) {
8723 numentries = 1UL << *_hash_shift;
8724 BUG_ON(!numentries);
8726 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8727 numentries = PAGE_SIZE / bucketsize;
8729 numentries = roundup_pow_of_two(numentries);
8731 /* limit allocation size to 1/16 total memory by default */
8733 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8734 do_div(max, bucketsize);
8736 max = min(max, 0x80000000ULL);
8738 if (numentries < low_limit)
8739 numentries = low_limit;
8740 if (numentries > max)
8743 log2qty = ilog2(numentries);
8745 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8748 size = bucketsize << log2qty;
8749 if (flags & HASH_EARLY) {
8750 if (flags & HASH_ZERO)
8751 table = memblock_alloc(size, SMP_CACHE_BYTES);
8753 table = memblock_alloc_raw(size,
8755 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8756 table = __vmalloc(size, gfp_flags);
8758 huge = is_vm_area_hugepages(table);
8761 * If bucketsize is not a power-of-two, we may free
8762 * some pages at the end of hash table which
8763 * alloc_pages_exact() automatically does
8765 table = alloc_pages_exact(size, gfp_flags);
8766 kmemleak_alloc(table, size, 1, gfp_flags);
8768 } while (!table && size > PAGE_SIZE && --log2qty);
8771 panic("Failed to allocate %s hash table\n", tablename);
8773 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8774 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8775 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8778 *_hash_shift = log2qty;
8780 *_hash_mask = (1 << log2qty) - 1;
8786 * This function checks whether pageblock includes unmovable pages or not.
8788 * PageLRU check without isolation or lru_lock could race so that
8789 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8790 * check without lock_page also may miss some movable non-lru pages at
8791 * race condition. So you can't expect this function should be exact.
8793 * Returns a page without holding a reference. If the caller wants to
8794 * dereference that page (e.g., dumping), it has to make sure that it
8795 * cannot get removed (e.g., via memory unplug) concurrently.
8798 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8799 int migratetype, int flags)
8801 unsigned long iter = 0;
8802 unsigned long pfn = page_to_pfn(page);
8803 unsigned long offset = pfn % pageblock_nr_pages;
8805 if (is_migrate_cma_page(page)) {
8807 * CMA allocations (alloc_contig_range) really need to mark
8808 * isolate CMA pageblocks even when they are not movable in fact
8809 * so consider them movable here.
8811 if (is_migrate_cma(migratetype))
8817 for (; iter < pageblock_nr_pages - offset; iter++) {
8818 if (!pfn_valid_within(pfn + iter))
8821 page = pfn_to_page(pfn + iter);
8824 * Both, bootmem allocations and memory holes are marked
8825 * PG_reserved and are unmovable. We can even have unmovable
8826 * allocations inside ZONE_MOVABLE, for example when
8827 * specifying "movablecore".
8829 if (PageReserved(page))
8833 * If the zone is movable and we have ruled out all reserved
8834 * pages then it should be reasonably safe to assume the rest
8837 if (zone_idx(zone) == ZONE_MOVABLE)
8841 * Hugepages are not in LRU lists, but they're movable.
8842 * THPs are on the LRU, but need to be counted as #small pages.
8843 * We need not scan over tail pages because we don't
8844 * handle each tail page individually in migration.
8846 if (PageHuge(page) || PageTransCompound(page)) {
8847 struct page *head = compound_head(page);
8848 unsigned int skip_pages;
8850 if (PageHuge(page)) {
8851 if (!hugepage_migration_supported(page_hstate(head)))
8853 } else if (!PageLRU(head) && !__PageMovable(head)) {
8857 skip_pages = compound_nr(head) - (page - head);
8858 iter += skip_pages - 1;
8863 * We can't use page_count without pin a page
8864 * because another CPU can free compound page.
8865 * This check already skips compound tails of THP
8866 * because their page->_refcount is zero at all time.
8868 if (!page_ref_count(page)) {
8869 if (PageBuddy(page))
8870 iter += (1 << buddy_order(page)) - 1;
8875 * The HWPoisoned page may be not in buddy system, and
8876 * page_count() is not 0.
8878 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8882 * We treat all PageOffline() pages as movable when offlining
8883 * to give drivers a chance to decrement their reference count
8884 * in MEM_GOING_OFFLINE in order to indicate that these pages
8885 * can be offlined as there are no direct references anymore.
8886 * For actually unmovable PageOffline() where the driver does
8887 * not support this, we will fail later when trying to actually
8888 * move these pages that still have a reference count > 0.
8889 * (false negatives in this function only)
8891 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8894 if (__PageMovable(page) || PageLRU(page))
8898 * If there are RECLAIMABLE pages, we need to check
8899 * it. But now, memory offline itself doesn't call
8900 * shrink_node_slabs() and it still to be fixed.
8907 #ifdef CONFIG_CONTIG_ALLOC
8908 static unsigned long pfn_max_align_down(unsigned long pfn)
8910 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8911 pageblock_nr_pages) - 1);
8914 static unsigned long pfn_max_align_up(unsigned long pfn)
8916 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8917 pageblock_nr_pages));
8920 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8921 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8922 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8923 static void alloc_contig_dump_pages(struct list_head *page_list)
8925 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8927 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8931 list_for_each_entry(page, page_list, lru)
8932 dump_page(page, "migration failure");
8936 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8941 /* [start, end) must belong to a single zone. */
8942 static int __alloc_contig_migrate_range(struct compact_control *cc,
8943 unsigned long start, unsigned long end)
8945 /* This function is based on compact_zone() from compaction.c. */
8946 unsigned int nr_reclaimed;
8947 unsigned long pfn = start;
8948 unsigned int tries = 0;
8950 struct migration_target_control mtc = {
8951 .nid = zone_to_nid(cc->zone),
8952 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8955 lru_cache_disable();
8957 while (pfn < end || !list_empty(&cc->migratepages)) {
8958 if (fatal_signal_pending(current)) {
8963 if (list_empty(&cc->migratepages)) {
8964 cc->nr_migratepages = 0;
8965 ret = isolate_migratepages_range(cc, pfn, end);
8966 if (ret && ret != -EAGAIN)
8968 pfn = cc->migrate_pfn;
8970 } else if (++tries == 5) {
8975 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8977 cc->nr_migratepages -= nr_reclaimed;
8979 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8980 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8983 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
8984 * to retry again over this error, so do the same here.
8993 alloc_contig_dump_pages(&cc->migratepages);
8994 putback_movable_pages(&cc->migratepages);
9001 * alloc_contig_range() -- tries to allocate given range of pages
9002 * @start: start PFN to allocate
9003 * @end: one-past-the-last PFN to allocate
9004 * @migratetype: migratetype of the underlying pageblocks (either
9005 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9006 * in range must have the same migratetype and it must
9007 * be either of the two.
9008 * @gfp_mask: GFP mask to use during compaction
9010 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9011 * aligned. The PFN range must belong to a single zone.
9013 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9014 * pageblocks in the range. Once isolated, the pageblocks should not
9015 * be modified by others.
9017 * Return: zero on success or negative error code. On success all
9018 * pages which PFN is in [start, end) are allocated for the caller and
9019 * need to be freed with free_contig_range().
9021 int alloc_contig_range(unsigned long start, unsigned long end,
9022 unsigned migratetype, gfp_t gfp_mask)
9024 unsigned long outer_start, outer_end;
9028 struct compact_control cc = {
9029 .nr_migratepages = 0,
9031 .zone = page_zone(pfn_to_page(start)),
9032 .mode = MIGRATE_SYNC,
9033 .ignore_skip_hint = true,
9034 .no_set_skip_hint = true,
9035 .gfp_mask = current_gfp_context(gfp_mask),
9036 .alloc_contig = true,
9038 INIT_LIST_HEAD(&cc.migratepages);
9041 * What we do here is we mark all pageblocks in range as
9042 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9043 * have different sizes, and due to the way page allocator
9044 * work, we align the range to biggest of the two pages so
9045 * that page allocator won't try to merge buddies from
9046 * different pageblocks and change MIGRATE_ISOLATE to some
9047 * other migration type.
9049 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9050 * migrate the pages from an unaligned range (ie. pages that
9051 * we are interested in). This will put all the pages in
9052 * range back to page allocator as MIGRATE_ISOLATE.
9054 * When this is done, we take the pages in range from page
9055 * allocator removing them from the buddy system. This way
9056 * page allocator will never consider using them.
9058 * This lets us mark the pageblocks back as
9059 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9060 * aligned range but not in the unaligned, original range are
9061 * put back to page allocator so that buddy can use them.
9064 ret = start_isolate_page_range(pfn_max_align_down(start),
9065 pfn_max_align_up(end), migratetype, 0);
9069 drain_all_pages(cc.zone);
9072 * In case of -EBUSY, we'd like to know which page causes problem.
9073 * So, just fall through. test_pages_isolated() has a tracepoint
9074 * which will report the busy page.
9076 * It is possible that busy pages could become available before
9077 * the call to test_pages_isolated, and the range will actually be
9078 * allocated. So, if we fall through be sure to clear ret so that
9079 * -EBUSY is not accidentally used or returned to caller.
9081 ret = __alloc_contig_migrate_range(&cc, start, end);
9082 if (ret && ret != -EBUSY)
9087 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9088 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9089 * more, all pages in [start, end) are free in page allocator.
9090 * What we are going to do is to allocate all pages from
9091 * [start, end) (that is remove them from page allocator).
9093 * The only problem is that pages at the beginning and at the
9094 * end of interesting range may be not aligned with pages that
9095 * page allocator holds, ie. they can be part of higher order
9096 * pages. Because of this, we reserve the bigger range and
9097 * once this is done free the pages we are not interested in.
9099 * We don't have to hold zone->lock here because the pages are
9100 * isolated thus they won't get removed from buddy.
9104 outer_start = start;
9105 while (!PageBuddy(pfn_to_page(outer_start))) {
9106 if (++order >= MAX_ORDER) {
9107 outer_start = start;
9110 outer_start &= ~0UL << order;
9113 if (outer_start != start) {
9114 order = buddy_order(pfn_to_page(outer_start));
9117 * outer_start page could be small order buddy page and
9118 * it doesn't include start page. Adjust outer_start
9119 * in this case to report failed page properly
9120 * on tracepoint in test_pages_isolated()
9122 if (outer_start + (1UL << order) <= start)
9123 outer_start = start;
9126 /* Make sure the range is really isolated. */
9127 if (test_pages_isolated(outer_start, end, 0)) {
9132 /* Grab isolated pages from freelists. */
9133 outer_end = isolate_freepages_range(&cc, outer_start, end);
9139 /* Free head and tail (if any) */
9140 if (start != outer_start)
9141 free_contig_range(outer_start, start - outer_start);
9142 if (end != outer_end)
9143 free_contig_range(end, outer_end - end);
9146 undo_isolate_page_range(pfn_max_align_down(start),
9147 pfn_max_align_up(end), migratetype);
9150 EXPORT_SYMBOL(alloc_contig_range);
9152 static int __alloc_contig_pages(unsigned long start_pfn,
9153 unsigned long nr_pages, gfp_t gfp_mask)
9155 unsigned long end_pfn = start_pfn + nr_pages;
9157 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9161 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9162 unsigned long nr_pages)
9164 unsigned long i, end_pfn = start_pfn + nr_pages;
9167 for (i = start_pfn; i < end_pfn; i++) {
9168 page = pfn_to_online_page(i);
9172 if (page_zone(page) != z)
9175 if (PageReserved(page))
9181 static bool zone_spans_last_pfn(const struct zone *zone,
9182 unsigned long start_pfn, unsigned long nr_pages)
9184 unsigned long last_pfn = start_pfn + nr_pages - 1;
9186 return zone_spans_pfn(zone, last_pfn);
9190 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9191 * @nr_pages: Number of contiguous pages to allocate
9192 * @gfp_mask: GFP mask to limit search and used during compaction
9194 * @nodemask: Mask for other possible nodes
9196 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9197 * on an applicable zonelist to find a contiguous pfn range which can then be
9198 * tried for allocation with alloc_contig_range(). This routine is intended
9199 * for allocation requests which can not be fulfilled with the buddy allocator.
9201 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9202 * power of two then the alignment is guaranteed to be to the given nr_pages
9203 * (e.g. 1GB request would be aligned to 1GB).
9205 * Allocated pages can be freed with free_contig_range() or by manually calling
9206 * __free_page() on each allocated page.
9208 * Return: pointer to contiguous pages on success, or NULL if not successful.
9210 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9211 int nid, nodemask_t *nodemask)
9213 unsigned long ret, pfn, flags;
9214 struct zonelist *zonelist;
9218 zonelist = node_zonelist(nid, gfp_mask);
9219 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9220 gfp_zone(gfp_mask), nodemask) {
9221 spin_lock_irqsave(&zone->lock, flags);
9223 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9224 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9225 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9227 * We release the zone lock here because
9228 * alloc_contig_range() will also lock the zone
9229 * at some point. If there's an allocation
9230 * spinning on this lock, it may win the race
9231 * and cause alloc_contig_range() to fail...
9233 spin_unlock_irqrestore(&zone->lock, flags);
9234 ret = __alloc_contig_pages(pfn, nr_pages,
9237 return pfn_to_page(pfn);
9238 spin_lock_irqsave(&zone->lock, flags);
9242 spin_unlock_irqrestore(&zone->lock, flags);
9246 #endif /* CONFIG_CONTIG_ALLOC */
9248 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9250 unsigned long count = 0;
9252 for (; nr_pages--; pfn++) {
9253 struct page *page = pfn_to_page(pfn);
9255 count += page_count(page) != 1;
9258 WARN(count != 0, "%lu pages are still in use!\n", count);
9260 EXPORT_SYMBOL(free_contig_range);
9263 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9264 * page high values need to be recalculated.
9266 void zone_pcp_update(struct zone *zone, int cpu_online)
9268 mutex_lock(&pcp_batch_high_lock);
9269 zone_set_pageset_high_and_batch(zone, cpu_online);
9270 mutex_unlock(&pcp_batch_high_lock);
9274 * Effectively disable pcplists for the zone by setting the high limit to 0
9275 * and draining all cpus. A concurrent page freeing on another CPU that's about
9276 * to put the page on pcplist will either finish before the drain and the page
9277 * will be drained, or observe the new high limit and skip the pcplist.
9279 * Must be paired with a call to zone_pcp_enable().
9281 void zone_pcp_disable(struct zone *zone)
9283 mutex_lock(&pcp_batch_high_lock);
9284 __zone_set_pageset_high_and_batch(zone, 0, 1);
9285 __drain_all_pages(zone, true);
9288 void zone_pcp_enable(struct zone *zone)
9290 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9291 mutex_unlock(&pcp_batch_high_lock);
9294 void zone_pcp_reset(struct zone *zone)
9297 struct per_cpu_zonestat *pzstats;
9299 if (zone->per_cpu_pageset != &boot_pageset) {
9300 for_each_online_cpu(cpu) {
9301 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9302 drain_zonestat(zone, pzstats);
9304 free_percpu(zone->per_cpu_pageset);
9305 free_percpu(zone->per_cpu_zonestats);
9306 zone->per_cpu_pageset = &boot_pageset;
9307 zone->per_cpu_zonestats = &boot_zonestats;
9311 #ifdef CONFIG_MEMORY_HOTREMOVE
9313 * All pages in the range must be in a single zone, must not contain holes,
9314 * must span full sections, and must be isolated before calling this function.
9316 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9318 unsigned long pfn = start_pfn;
9322 unsigned long flags;
9324 offline_mem_sections(pfn, end_pfn);
9325 zone = page_zone(pfn_to_page(pfn));
9326 spin_lock_irqsave(&zone->lock, flags);
9327 while (pfn < end_pfn) {
9328 page = pfn_to_page(pfn);
9330 * The HWPoisoned page may be not in buddy system, and
9331 * page_count() is not 0.
9333 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9338 * At this point all remaining PageOffline() pages have a
9339 * reference count of 0 and can simply be skipped.
9341 if (PageOffline(page)) {
9342 BUG_ON(page_count(page));
9343 BUG_ON(PageBuddy(page));
9348 BUG_ON(page_count(page));
9349 BUG_ON(!PageBuddy(page));
9350 order = buddy_order(page);
9351 del_page_from_free_list(page, zone, order);
9352 pfn += (1 << order);
9354 spin_unlock_irqrestore(&zone->lock, flags);
9358 bool is_free_buddy_page(struct page *page)
9360 struct zone *zone = page_zone(page);
9361 unsigned long pfn = page_to_pfn(page);
9362 unsigned long flags;
9365 spin_lock_irqsave(&zone->lock, flags);
9366 for (order = 0; order < MAX_ORDER; order++) {
9367 struct page *page_head = page - (pfn & ((1 << order) - 1));
9369 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
9372 spin_unlock_irqrestore(&zone->lock, flags);
9374 return order < MAX_ORDER;
9377 #ifdef CONFIG_MEMORY_FAILURE
9379 * Break down a higher-order page in sub-pages, and keep our target out of
9382 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9383 struct page *target, int low, int high,
9386 unsigned long size = 1 << high;
9387 struct page *current_buddy, *next_page;
9389 while (high > low) {
9393 if (target >= &page[size]) {
9394 next_page = page + size;
9395 current_buddy = page;
9398 current_buddy = page + size;
9401 if (set_page_guard(zone, current_buddy, high, migratetype))
9404 if (current_buddy != target) {
9405 add_to_free_list(current_buddy, zone, high, migratetype);
9406 set_buddy_order(current_buddy, high);
9413 * Take a page that will be marked as poisoned off the buddy allocator.
9415 bool take_page_off_buddy(struct page *page)
9417 struct zone *zone = page_zone(page);
9418 unsigned long pfn = page_to_pfn(page);
9419 unsigned long flags;
9423 spin_lock_irqsave(&zone->lock, flags);
9424 for (order = 0; order < MAX_ORDER; order++) {
9425 struct page *page_head = page - (pfn & ((1 << order) - 1));
9426 int page_order = buddy_order(page_head);
9428 if (PageBuddy(page_head) && page_order >= order) {
9429 unsigned long pfn_head = page_to_pfn(page_head);
9430 int migratetype = get_pfnblock_migratetype(page_head,
9433 del_page_from_free_list(page_head, zone, page_order);
9434 break_down_buddy_pages(zone, page_head, page, 0,
9435 page_order, migratetype);
9436 if (!is_migrate_isolate(migratetype))
9437 __mod_zone_freepage_state(zone, -1, migratetype);
9441 if (page_count(page_head) > 0)
9444 spin_unlock_irqrestore(&zone->lock, flags);