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_FRACTION (8)
127 #if defined(CONFIG_DEBUG_INFO_BTF) && \
128 !defined(CONFIG_DEBUG_LOCK_ALLOC) && \
129 !defined(CONFIG_PAHOLE_HAS_ZEROSIZE_PERCPU_SUPPORT)
131 * pahole 1.21 and earlier gets confused by zero-sized per-CPU
132 * variables and produces invalid BTF. Ensure that
133 * sizeof(struct pagesets) != 0 for older versions of pahole.
136 #warning "pahole too old to support zero-sized struct pagesets"
139 static DEFINE_PER_CPU(struct pagesets, pagesets) = {
140 .lock = INIT_LOCAL_LOCK(lock),
143 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
144 DEFINE_PER_CPU(int, numa_node);
145 EXPORT_PER_CPU_SYMBOL(numa_node);
148 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
150 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
152 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
153 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
154 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
155 * defined in <linux/topology.h>.
157 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
158 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
161 /* work_structs for global per-cpu drains */
164 struct work_struct work;
166 static DEFINE_MUTEX(pcpu_drain_mutex);
167 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
169 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
170 volatile unsigned long latent_entropy __latent_entropy;
171 EXPORT_SYMBOL(latent_entropy);
175 * Array of node states.
177 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
178 [N_POSSIBLE] = NODE_MASK_ALL,
179 [N_ONLINE] = { { [0] = 1UL } },
181 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
182 #ifdef CONFIG_HIGHMEM
183 [N_HIGH_MEMORY] = { { [0] = 1UL } },
185 [N_MEMORY] = { { [0] = 1UL } },
186 [N_CPU] = { { [0] = 1UL } },
189 EXPORT_SYMBOL(node_states);
191 atomic_long_t _totalram_pages __read_mostly;
192 EXPORT_SYMBOL(_totalram_pages);
193 unsigned long totalreserve_pages __read_mostly;
194 unsigned long totalcma_pages __read_mostly;
196 int percpu_pagelist_fraction;
197 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
198 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
199 EXPORT_SYMBOL(init_on_alloc);
201 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
202 EXPORT_SYMBOL(init_on_free);
204 static bool _init_on_alloc_enabled_early __read_mostly
205 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
206 static int __init early_init_on_alloc(char *buf)
209 return kstrtobool(buf, &_init_on_alloc_enabled_early);
211 early_param("init_on_alloc", early_init_on_alloc);
213 static bool _init_on_free_enabled_early __read_mostly
214 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
215 static int __init early_init_on_free(char *buf)
217 return kstrtobool(buf, &_init_on_free_enabled_early);
219 early_param("init_on_free", early_init_on_free);
222 * A cached value of the page's pageblock's migratetype, used when the page is
223 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
224 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
225 * Also the migratetype set in the page does not necessarily match the pcplist
226 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
227 * other index - this ensures that it will be put on the correct CMA freelist.
229 static inline int get_pcppage_migratetype(struct page *page)
234 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
236 page->index = migratetype;
239 #ifdef CONFIG_PM_SLEEP
241 * The following functions are used by the suspend/hibernate code to temporarily
242 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
243 * while devices are suspended. To avoid races with the suspend/hibernate code,
244 * they should always be called with system_transition_mutex held
245 * (gfp_allowed_mask also should only be modified with system_transition_mutex
246 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
247 * with that modification).
250 static gfp_t saved_gfp_mask;
252 void pm_restore_gfp_mask(void)
254 WARN_ON(!mutex_is_locked(&system_transition_mutex));
255 if (saved_gfp_mask) {
256 gfp_allowed_mask = saved_gfp_mask;
261 void pm_restrict_gfp_mask(void)
263 WARN_ON(!mutex_is_locked(&system_transition_mutex));
264 WARN_ON(saved_gfp_mask);
265 saved_gfp_mask = gfp_allowed_mask;
266 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
269 bool pm_suspended_storage(void)
271 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
275 #endif /* CONFIG_PM_SLEEP */
277 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
278 unsigned int pageblock_order __read_mostly;
281 static void __free_pages_ok(struct page *page, unsigned int order,
285 * results with 256, 32 in the lowmem_reserve sysctl:
286 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
287 * 1G machine -> (16M dma, 784M normal, 224M high)
288 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
289 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
290 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
292 * TBD: should special case ZONE_DMA32 machines here - in those we normally
293 * don't need any ZONE_NORMAL reservation
295 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
296 #ifdef CONFIG_ZONE_DMA
299 #ifdef CONFIG_ZONE_DMA32
303 #ifdef CONFIG_HIGHMEM
309 static char * const zone_names[MAX_NR_ZONES] = {
310 #ifdef CONFIG_ZONE_DMA
313 #ifdef CONFIG_ZONE_DMA32
317 #ifdef CONFIG_HIGHMEM
321 #ifdef CONFIG_ZONE_DEVICE
326 const char * const migratetype_names[MIGRATE_TYPES] = {
334 #ifdef CONFIG_MEMORY_ISOLATION
339 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
340 [NULL_COMPOUND_DTOR] = NULL,
341 [COMPOUND_PAGE_DTOR] = free_compound_page,
342 #ifdef CONFIG_HUGETLB_PAGE
343 [HUGETLB_PAGE_DTOR] = free_huge_page,
345 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
346 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
350 int min_free_kbytes = 1024;
351 int user_min_free_kbytes = -1;
352 #ifdef CONFIG_DISCONTIGMEM
354 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
355 * are not on separate NUMA nodes. Functionally this works but with
356 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
357 * quite small. By default, do not boost watermarks on discontigmem as in
358 * many cases very high-order allocations like THP are likely to be
359 * unsupported and the premature reclaim offsets the advantage of long-term
360 * fragmentation avoidance.
362 int watermark_boost_factor __read_mostly;
364 int watermark_boost_factor __read_mostly = 15000;
366 int watermark_scale_factor = 10;
368 static unsigned long nr_kernel_pages __initdata;
369 static unsigned long nr_all_pages __initdata;
370 static unsigned long dma_reserve __initdata;
372 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
373 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
374 static unsigned long required_kernelcore __initdata;
375 static unsigned long required_kernelcore_percent __initdata;
376 static unsigned long required_movablecore __initdata;
377 static unsigned long required_movablecore_percent __initdata;
378 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
379 static bool mirrored_kernelcore __meminitdata;
381 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
383 EXPORT_SYMBOL(movable_zone);
386 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
387 unsigned int nr_online_nodes __read_mostly = 1;
388 EXPORT_SYMBOL(nr_node_ids);
389 EXPORT_SYMBOL(nr_online_nodes);
392 int page_group_by_mobility_disabled __read_mostly;
394 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
396 * During boot we initialize deferred pages on-demand, as needed, but once
397 * page_alloc_init_late() has finished, the deferred pages are all initialized,
398 * and we can permanently disable that path.
400 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
403 * Calling kasan_free_pages() only after deferred memory initialization
404 * has completed. Poisoning pages during deferred memory init will greatly
405 * lengthen the process and cause problem in large memory systems as the
406 * deferred pages initialization is done with interrupt disabled.
408 * Assuming that there will be no reference to those newly initialized
409 * pages before they are ever allocated, this should have no effect on
410 * KASAN memory tracking as the poison will be properly inserted at page
411 * allocation time. The only corner case is when pages are allocated by
412 * on-demand allocation and then freed again before the deferred pages
413 * initialization is done, but this is not likely to happen.
415 static inline void kasan_free_nondeferred_pages(struct page *page, int order,
416 bool init, fpi_t fpi_flags)
418 if (static_branch_unlikely(&deferred_pages))
420 if (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
421 (fpi_flags & FPI_SKIP_KASAN_POISON))
423 kasan_free_pages(page, order, init);
426 /* Returns true if the struct page for the pfn is uninitialised */
427 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
429 int nid = early_pfn_to_nid(pfn);
431 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
438 * Returns true when the remaining initialisation should be deferred until
439 * later in the boot cycle when it can be parallelised.
441 static bool __meminit
442 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
444 static unsigned long prev_end_pfn, nr_initialised;
447 * prev_end_pfn static that contains the end of previous zone
448 * No need to protect because called very early in boot before smp_init.
450 if (prev_end_pfn != end_pfn) {
451 prev_end_pfn = end_pfn;
455 /* Always populate low zones for address-constrained allocations */
456 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
459 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
462 * We start only with one section of pages, more pages are added as
463 * needed until the rest of deferred pages are initialized.
466 if ((nr_initialised > PAGES_PER_SECTION) &&
467 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
468 NODE_DATA(nid)->first_deferred_pfn = pfn;
474 static inline void kasan_free_nondeferred_pages(struct page *page, int order,
475 bool init, fpi_t fpi_flags)
477 if (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
478 (fpi_flags & FPI_SKIP_KASAN_POISON))
480 kasan_free_pages(page, order, init);
483 static inline bool early_page_uninitialised(unsigned long pfn)
488 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
494 /* Return a pointer to the bitmap storing bits affecting a block of pages */
495 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
498 #ifdef CONFIG_SPARSEMEM
499 return section_to_usemap(__pfn_to_section(pfn));
501 return page_zone(page)->pageblock_flags;
502 #endif /* CONFIG_SPARSEMEM */
505 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
507 #ifdef CONFIG_SPARSEMEM
508 pfn &= (PAGES_PER_SECTION-1);
510 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
511 #endif /* CONFIG_SPARSEMEM */
512 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
515 static __always_inline
516 unsigned long __get_pfnblock_flags_mask(const struct page *page,
520 unsigned long *bitmap;
521 unsigned long bitidx, word_bitidx;
524 bitmap = get_pageblock_bitmap(page, pfn);
525 bitidx = pfn_to_bitidx(page, pfn);
526 word_bitidx = bitidx / BITS_PER_LONG;
527 bitidx &= (BITS_PER_LONG-1);
529 word = bitmap[word_bitidx];
530 return (word >> bitidx) & mask;
534 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
535 * @page: The page within the block of interest
536 * @pfn: The target page frame number
537 * @mask: mask of bits that the caller is interested in
539 * Return: pageblock_bits flags
541 unsigned long get_pfnblock_flags_mask(const struct page *page,
542 unsigned long pfn, unsigned long mask)
544 return __get_pfnblock_flags_mask(page, pfn, mask);
547 static __always_inline int get_pfnblock_migratetype(const struct page *page,
550 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
554 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
555 * @page: The page within the block of interest
556 * @flags: The flags to set
557 * @pfn: The target page frame number
558 * @mask: mask of bits that the caller is interested in
560 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
564 unsigned long *bitmap;
565 unsigned long bitidx, word_bitidx;
566 unsigned long old_word, word;
568 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
569 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
571 bitmap = get_pageblock_bitmap(page, pfn);
572 bitidx = pfn_to_bitidx(page, pfn);
573 word_bitidx = bitidx / BITS_PER_LONG;
574 bitidx &= (BITS_PER_LONG-1);
576 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
581 word = READ_ONCE(bitmap[word_bitidx]);
583 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
584 if (word == old_word)
590 void set_pageblock_migratetype(struct page *page, int migratetype)
592 if (unlikely(page_group_by_mobility_disabled &&
593 migratetype < MIGRATE_PCPTYPES))
594 migratetype = MIGRATE_UNMOVABLE;
596 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
597 page_to_pfn(page), MIGRATETYPE_MASK);
600 #ifdef CONFIG_DEBUG_VM
601 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
605 unsigned long pfn = page_to_pfn(page);
606 unsigned long sp, start_pfn;
609 seq = zone_span_seqbegin(zone);
610 start_pfn = zone->zone_start_pfn;
611 sp = zone->spanned_pages;
612 if (!zone_spans_pfn(zone, pfn))
614 } while (zone_span_seqretry(zone, seq));
617 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
618 pfn, zone_to_nid(zone), zone->name,
619 start_pfn, start_pfn + sp);
624 static int page_is_consistent(struct zone *zone, struct page *page)
626 if (!pfn_valid_within(page_to_pfn(page)))
628 if (zone != page_zone(page))
634 * Temporary debugging check for pages not lying within a given zone.
636 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
638 if (page_outside_zone_boundaries(zone, page))
640 if (!page_is_consistent(zone, page))
646 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
652 static void bad_page(struct page *page, const char *reason)
654 static unsigned long resume;
655 static unsigned long nr_shown;
656 static unsigned long nr_unshown;
659 * Allow a burst of 60 reports, then keep quiet for that minute;
660 * or allow a steady drip of one report per second.
662 if (nr_shown == 60) {
663 if (time_before(jiffies, resume)) {
669 "BUG: Bad page state: %lu messages suppressed\n",
676 resume = jiffies + 60 * HZ;
678 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
679 current->comm, page_to_pfn(page));
680 dump_page(page, reason);
685 /* Leave bad fields for debug, except PageBuddy could make trouble */
686 page_mapcount_reset(page); /* remove PageBuddy */
687 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
691 * Higher-order pages are called "compound pages". They are structured thusly:
693 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
695 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
696 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
698 * The first tail page's ->compound_dtor holds the offset in array of compound
699 * page destructors. See compound_page_dtors.
701 * The first tail page's ->compound_order holds the order of allocation.
702 * This usage means that zero-order pages may not be compound.
705 void free_compound_page(struct page *page)
707 mem_cgroup_uncharge(page);
708 __free_pages_ok(page, compound_order(page), FPI_NONE);
711 void prep_compound_page(struct page *page, unsigned int order)
714 int nr_pages = 1 << order;
717 for (i = 1; i < nr_pages; i++) {
718 struct page *p = page + i;
719 set_page_count(p, 0);
720 p->mapping = TAIL_MAPPING;
721 set_compound_head(p, page);
724 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
725 set_compound_order(page, order);
726 atomic_set(compound_mapcount_ptr(page), -1);
727 if (hpage_pincount_available(page))
728 atomic_set(compound_pincount_ptr(page), 0);
731 #ifdef CONFIG_DEBUG_PAGEALLOC
732 unsigned int _debug_guardpage_minorder;
734 bool _debug_pagealloc_enabled_early __read_mostly
735 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
736 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
737 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
738 EXPORT_SYMBOL(_debug_pagealloc_enabled);
740 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
742 static int __init early_debug_pagealloc(char *buf)
744 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
746 early_param("debug_pagealloc", early_debug_pagealloc);
748 static int __init debug_guardpage_minorder_setup(char *buf)
752 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
753 pr_err("Bad debug_guardpage_minorder value\n");
756 _debug_guardpage_minorder = res;
757 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
760 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
762 static inline bool set_page_guard(struct zone *zone, struct page *page,
763 unsigned int order, int migratetype)
765 if (!debug_guardpage_enabled())
768 if (order >= debug_guardpage_minorder())
771 __SetPageGuard(page);
772 INIT_LIST_HEAD(&page->lru);
773 set_page_private(page, order);
774 /* Guard pages are not available for any usage */
775 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
780 static inline void clear_page_guard(struct zone *zone, struct page *page,
781 unsigned int order, int migratetype)
783 if (!debug_guardpage_enabled())
786 __ClearPageGuard(page);
788 set_page_private(page, 0);
789 if (!is_migrate_isolate(migratetype))
790 __mod_zone_freepage_state(zone, (1 << order), migratetype);
793 static inline bool set_page_guard(struct zone *zone, struct page *page,
794 unsigned int order, int migratetype) { return false; }
795 static inline void clear_page_guard(struct zone *zone, struct page *page,
796 unsigned int order, int migratetype) {}
800 * Enable static keys related to various memory debugging and hardening options.
801 * Some override others, and depend on early params that are evaluated in the
802 * order of appearance. So we need to first gather the full picture of what was
803 * enabled, and then make decisions.
805 void init_mem_debugging_and_hardening(void)
807 bool page_poisoning_requested = false;
809 #ifdef CONFIG_PAGE_POISONING
811 * Page poisoning is debug page alloc for some arches. If
812 * either of those options are enabled, enable poisoning.
814 if (page_poisoning_enabled() ||
815 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
816 debug_pagealloc_enabled())) {
817 static_branch_enable(&_page_poisoning_enabled);
818 page_poisoning_requested = true;
822 if (_init_on_alloc_enabled_early) {
823 if (page_poisoning_requested)
824 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
825 "will take precedence over init_on_alloc\n");
827 static_branch_enable(&init_on_alloc);
829 if (_init_on_free_enabled_early) {
830 if (page_poisoning_requested)
831 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
832 "will take precedence over init_on_free\n");
834 static_branch_enable(&init_on_free);
837 #ifdef CONFIG_DEBUG_PAGEALLOC
838 if (!debug_pagealloc_enabled())
841 static_branch_enable(&_debug_pagealloc_enabled);
843 if (!debug_guardpage_minorder())
846 static_branch_enable(&_debug_guardpage_enabled);
850 static inline void set_buddy_order(struct page *page, unsigned int order)
852 set_page_private(page, order);
853 __SetPageBuddy(page);
857 * This function checks whether a page is free && is the buddy
858 * we can coalesce a page and its buddy if
859 * (a) the buddy is not in a hole (check before calling!) &&
860 * (b) the buddy is in the buddy system &&
861 * (c) a page and its buddy have the same order &&
862 * (d) a page and its buddy are in the same zone.
864 * For recording whether a page is in the buddy system, we set PageBuddy.
865 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
867 * For recording page's order, we use page_private(page).
869 static inline bool page_is_buddy(struct page *page, struct page *buddy,
872 if (!page_is_guard(buddy) && !PageBuddy(buddy))
875 if (buddy_order(buddy) != order)
879 * zone check is done late to avoid uselessly calculating
880 * zone/node ids for pages that could never merge.
882 if (page_zone_id(page) != page_zone_id(buddy))
885 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
890 #ifdef CONFIG_COMPACTION
891 static inline struct capture_control *task_capc(struct zone *zone)
893 struct capture_control *capc = current->capture_control;
895 return unlikely(capc) &&
896 !(current->flags & PF_KTHREAD) &&
898 capc->cc->zone == zone ? capc : NULL;
902 compaction_capture(struct capture_control *capc, struct page *page,
903 int order, int migratetype)
905 if (!capc || order != capc->cc->order)
908 /* Do not accidentally pollute CMA or isolated regions*/
909 if (is_migrate_cma(migratetype) ||
910 is_migrate_isolate(migratetype))
914 * Do not let lower order allocations pollute a movable pageblock.
915 * This might let an unmovable request use a reclaimable pageblock
916 * and vice-versa but no more than normal fallback logic which can
917 * have trouble finding a high-order free page.
919 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
927 static inline struct capture_control *task_capc(struct zone *zone)
933 compaction_capture(struct capture_control *capc, struct page *page,
934 int order, int migratetype)
938 #endif /* CONFIG_COMPACTION */
940 /* Used for pages not on another list */
941 static inline void add_to_free_list(struct page *page, struct zone *zone,
942 unsigned int order, int migratetype)
944 struct free_area *area = &zone->free_area[order];
946 list_add(&page->lru, &area->free_list[migratetype]);
950 /* Used for pages not on another list */
951 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
952 unsigned int order, int migratetype)
954 struct free_area *area = &zone->free_area[order];
956 list_add_tail(&page->lru, &area->free_list[migratetype]);
961 * Used for pages which are on another list. Move the pages to the tail
962 * of the list - so the moved pages won't immediately be considered for
963 * allocation again (e.g., optimization for memory onlining).
965 static inline void move_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_move_tail(&page->lru, &area->free_list[migratetype]);
973 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
976 /* clear reported state and update reported page count */
977 if (page_reported(page))
978 __ClearPageReported(page);
980 list_del(&page->lru);
981 __ClearPageBuddy(page);
982 set_page_private(page, 0);
983 zone->free_area[order].nr_free--;
987 * If this is not the largest possible page, check if the buddy
988 * of the next-highest order is free. If it is, it's possible
989 * that pages are being freed that will coalesce soon. In case,
990 * that is happening, add the free page to the tail of the list
991 * so it's less likely to be used soon and more likely to be merged
992 * as a higher order page
995 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
996 struct page *page, unsigned int order)
998 struct page *higher_page, *higher_buddy;
999 unsigned long combined_pfn;
1001 if (order >= MAX_ORDER - 2)
1004 if (!pfn_valid_within(buddy_pfn))
1007 combined_pfn = buddy_pfn & pfn;
1008 higher_page = page + (combined_pfn - pfn);
1009 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1010 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1012 return pfn_valid_within(buddy_pfn) &&
1013 page_is_buddy(higher_page, higher_buddy, order + 1);
1017 * Freeing function for a buddy system allocator.
1019 * The concept of a buddy system is to maintain direct-mapped table
1020 * (containing bit values) for memory blocks of various "orders".
1021 * The bottom level table contains the map for the smallest allocatable
1022 * units of memory (here, pages), and each level above it describes
1023 * pairs of units from the levels below, hence, "buddies".
1024 * At a high level, all that happens here is marking the table entry
1025 * at the bottom level available, and propagating the changes upward
1026 * as necessary, plus some accounting needed to play nicely with other
1027 * parts of the VM system.
1028 * At each level, we keep a list of pages, which are heads of continuous
1029 * free pages of length of (1 << order) and marked with PageBuddy.
1030 * Page's order is recorded in page_private(page) field.
1031 * So when we are allocating or freeing one, we can derive the state of the
1032 * other. That is, if we allocate a small block, and both were
1033 * free, the remainder of the region must be split into blocks.
1034 * If a block is freed, and its buddy is also free, then this
1035 * triggers coalescing into a block of larger size.
1040 static inline void __free_one_page(struct page *page,
1042 struct zone *zone, unsigned int order,
1043 int migratetype, fpi_t fpi_flags)
1045 struct capture_control *capc = task_capc(zone);
1046 unsigned long buddy_pfn;
1047 unsigned long combined_pfn;
1048 unsigned int max_order;
1052 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1054 VM_BUG_ON(!zone_is_initialized(zone));
1055 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1057 VM_BUG_ON(migratetype == -1);
1058 if (likely(!is_migrate_isolate(migratetype)))
1059 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1061 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1062 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1065 while (order < max_order) {
1066 if (compaction_capture(capc, page, order, migratetype)) {
1067 __mod_zone_freepage_state(zone, -(1 << order),
1071 buddy_pfn = __find_buddy_pfn(pfn, order);
1072 buddy = page + (buddy_pfn - pfn);
1074 if (!pfn_valid_within(buddy_pfn))
1076 if (!page_is_buddy(page, buddy, order))
1079 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1080 * merge with it and move up one order.
1082 if (page_is_guard(buddy))
1083 clear_page_guard(zone, buddy, order, migratetype);
1085 del_page_from_free_list(buddy, zone, order);
1086 combined_pfn = buddy_pfn & pfn;
1087 page = page + (combined_pfn - pfn);
1091 if (order < MAX_ORDER - 1) {
1092 /* If we are here, it means order is >= pageblock_order.
1093 * We want to prevent merge between freepages on isolate
1094 * pageblock and normal pageblock. Without this, pageblock
1095 * isolation could cause incorrect freepage or CMA accounting.
1097 * We don't want to hit this code for the more frequent
1098 * low-order merging.
1100 if (unlikely(has_isolate_pageblock(zone))) {
1103 buddy_pfn = __find_buddy_pfn(pfn, order);
1104 buddy = page + (buddy_pfn - pfn);
1105 buddy_mt = get_pageblock_migratetype(buddy);
1107 if (migratetype != buddy_mt
1108 && (is_migrate_isolate(migratetype) ||
1109 is_migrate_isolate(buddy_mt)))
1112 max_order = order + 1;
1113 goto continue_merging;
1117 set_buddy_order(page, order);
1119 if (fpi_flags & FPI_TO_TAIL)
1121 else if (is_shuffle_order(order))
1122 to_tail = shuffle_pick_tail();
1124 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1127 add_to_free_list_tail(page, zone, order, migratetype);
1129 add_to_free_list(page, zone, order, migratetype);
1131 /* Notify page reporting subsystem of freed page */
1132 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1133 page_reporting_notify_free(order);
1137 * A bad page could be due to a number of fields. Instead of multiple branches,
1138 * try and check multiple fields with one check. The caller must do a detailed
1139 * check if necessary.
1141 static inline bool page_expected_state(struct page *page,
1142 unsigned long check_flags)
1144 if (unlikely(atomic_read(&page->_mapcount) != -1))
1147 if (unlikely((unsigned long)page->mapping |
1148 page_ref_count(page) |
1152 (page->flags & check_flags)))
1158 static const char *page_bad_reason(struct page *page, unsigned long flags)
1160 const char *bad_reason = NULL;
1162 if (unlikely(atomic_read(&page->_mapcount) != -1))
1163 bad_reason = "nonzero mapcount";
1164 if (unlikely(page->mapping != NULL))
1165 bad_reason = "non-NULL mapping";
1166 if (unlikely(page_ref_count(page) != 0))
1167 bad_reason = "nonzero _refcount";
1168 if (unlikely(page->flags & flags)) {
1169 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1170 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1172 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1175 if (unlikely(page->memcg_data))
1176 bad_reason = "page still charged to cgroup";
1181 static void check_free_page_bad(struct page *page)
1184 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1187 static inline int check_free_page(struct page *page)
1189 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1192 /* Something has gone sideways, find it */
1193 check_free_page_bad(page);
1197 static int free_tail_pages_check(struct page *head_page, struct page *page)
1202 * We rely page->lru.next never has bit 0 set, unless the page
1203 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1205 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1207 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1211 switch (page - head_page) {
1213 /* the first tail page: ->mapping may be compound_mapcount() */
1214 if (unlikely(compound_mapcount(page))) {
1215 bad_page(page, "nonzero compound_mapcount");
1221 * the second tail page: ->mapping is
1222 * deferred_list.next -- ignore value.
1226 if (page->mapping != TAIL_MAPPING) {
1227 bad_page(page, "corrupted mapping in tail page");
1232 if (unlikely(!PageTail(page))) {
1233 bad_page(page, "PageTail not set");
1236 if (unlikely(compound_head(page) != head_page)) {
1237 bad_page(page, "compound_head not consistent");
1242 page->mapping = NULL;
1243 clear_compound_head(page);
1247 static void kernel_init_free_pages(struct page *page, int numpages)
1251 /* s390's use of memset() could override KASAN redzones. */
1252 kasan_disable_current();
1253 for (i = 0; i < numpages; i++) {
1254 u8 tag = page_kasan_tag(page + i);
1255 page_kasan_tag_reset(page + i);
1256 clear_highpage(page + i);
1257 page_kasan_tag_set(page + i, tag);
1259 kasan_enable_current();
1262 static __always_inline bool free_pages_prepare(struct page *page,
1263 unsigned int order, bool check_free, fpi_t fpi_flags)
1268 VM_BUG_ON_PAGE(PageTail(page), page);
1270 trace_mm_page_free(page, order);
1272 if (unlikely(PageHWPoison(page)) && !order) {
1274 * Do not let hwpoison pages hit pcplists/buddy
1275 * Untie memcg state and reset page's owner
1277 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1278 __memcg_kmem_uncharge_page(page, order);
1279 reset_page_owner(page, order);
1284 * Check tail pages before head page information is cleared to
1285 * avoid checking PageCompound for order-0 pages.
1287 if (unlikely(order)) {
1288 bool compound = PageCompound(page);
1291 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1294 ClearPageDoubleMap(page);
1295 for (i = 1; i < (1 << order); i++) {
1297 bad += free_tail_pages_check(page, page + i);
1298 if (unlikely(check_free_page(page + i))) {
1302 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1305 if (PageMappingFlags(page))
1306 page->mapping = NULL;
1307 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1308 __memcg_kmem_uncharge_page(page, order);
1310 bad += check_free_page(page);
1314 page_cpupid_reset_last(page);
1315 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1316 reset_page_owner(page, order);
1318 if (!PageHighMem(page)) {
1319 debug_check_no_locks_freed(page_address(page),
1320 PAGE_SIZE << order);
1321 debug_check_no_obj_freed(page_address(page),
1322 PAGE_SIZE << order);
1325 kernel_poison_pages(page, 1 << order);
1328 * As memory initialization might be integrated into KASAN,
1329 * kasan_free_pages and kernel_init_free_pages must be
1330 * kept together to avoid discrepancies in behavior.
1332 * With hardware tag-based KASAN, memory tags must be set before the
1333 * page becomes unavailable via debug_pagealloc or arch_free_page.
1335 init = want_init_on_free();
1336 if (init && !kasan_has_integrated_init())
1337 kernel_init_free_pages(page, 1 << order);
1338 kasan_free_nondeferred_pages(page, order, init, fpi_flags);
1341 * arch_free_page() can make the page's contents inaccessible. s390
1342 * does this. So nothing which can access the page's contents should
1343 * happen after this.
1345 arch_free_page(page, order);
1347 debug_pagealloc_unmap_pages(page, 1 << order);
1352 #ifdef CONFIG_DEBUG_VM
1354 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1355 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1356 * moved from pcp lists to free lists.
1358 static bool free_pcp_prepare(struct page *page)
1360 return free_pages_prepare(page, 0, true, FPI_NONE);
1363 static bool bulkfree_pcp_prepare(struct page *page)
1365 if (debug_pagealloc_enabled_static())
1366 return check_free_page(page);
1372 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1373 * moving from pcp lists to free list in order to reduce overhead. With
1374 * debug_pagealloc enabled, they are checked also immediately when being freed
1377 static bool free_pcp_prepare(struct page *page)
1379 if (debug_pagealloc_enabled_static())
1380 return free_pages_prepare(page, 0, true, FPI_NONE);
1382 return free_pages_prepare(page, 0, false, FPI_NONE);
1385 static bool bulkfree_pcp_prepare(struct page *page)
1387 return check_free_page(page);
1389 #endif /* CONFIG_DEBUG_VM */
1391 static inline void prefetch_buddy(struct page *page)
1393 unsigned long pfn = page_to_pfn(page);
1394 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1395 struct page *buddy = page + (buddy_pfn - pfn);
1401 * Frees a number of pages from the PCP lists
1402 * Assumes all pages on list are in same zone, and of same order.
1403 * count is the number of pages to free.
1405 * If the zone was previously in an "all pages pinned" state then look to
1406 * see if this freeing clears that state.
1408 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1409 * pinned" detection logic.
1411 static void free_pcppages_bulk(struct zone *zone, int count,
1412 struct per_cpu_pages *pcp)
1414 int migratetype = 0;
1416 int prefetch_nr = READ_ONCE(pcp->batch);
1417 bool isolated_pageblocks;
1418 struct page *page, *tmp;
1422 * Ensure proper count is passed which otherwise would stuck in the
1423 * below while (list_empty(list)) loop.
1425 count = min(pcp->count, count);
1427 struct list_head *list;
1430 * Remove pages from lists in a round-robin fashion. A
1431 * batch_free count is maintained that is incremented when an
1432 * empty list is encountered. This is so more pages are freed
1433 * off fuller lists instead of spinning excessively around empty
1438 if (++migratetype == MIGRATE_PCPTYPES)
1440 list = &pcp->lists[migratetype];
1441 } while (list_empty(list));
1443 /* This is the only non-empty list. Free them all. */
1444 if (batch_free == MIGRATE_PCPTYPES)
1448 page = list_last_entry(list, struct page, lru);
1449 /* must delete to avoid corrupting pcp list */
1450 list_del(&page->lru);
1453 if (bulkfree_pcp_prepare(page))
1456 list_add_tail(&page->lru, &head);
1459 * We are going to put the page back to the global
1460 * pool, prefetch its buddy to speed up later access
1461 * under zone->lock. It is believed the overhead of
1462 * an additional test and calculating buddy_pfn here
1463 * can be offset by reduced memory latency later. To
1464 * avoid excessive prefetching due to large count, only
1465 * prefetch buddy for the first pcp->batch nr of pages.
1468 prefetch_buddy(page);
1471 } while (--count && --batch_free && !list_empty(list));
1475 * local_lock_irq held so equivalent to spin_lock_irqsave for
1476 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1478 spin_lock(&zone->lock);
1479 isolated_pageblocks = has_isolate_pageblock(zone);
1482 * Use safe version since after __free_one_page(),
1483 * page->lru.next will not point to original list.
1485 list_for_each_entry_safe(page, tmp, &head, lru) {
1486 int mt = get_pcppage_migratetype(page);
1487 /* MIGRATE_ISOLATE page should not go to pcplists */
1488 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1489 /* Pageblock could have been isolated meanwhile */
1490 if (unlikely(isolated_pageblocks))
1491 mt = get_pageblock_migratetype(page);
1493 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1494 trace_mm_page_pcpu_drain(page, 0, mt);
1496 spin_unlock(&zone->lock);
1499 static void free_one_page(struct zone *zone,
1500 struct page *page, unsigned long pfn,
1502 int migratetype, fpi_t fpi_flags)
1504 unsigned long flags;
1506 spin_lock_irqsave(&zone->lock, flags);
1507 if (unlikely(has_isolate_pageblock(zone) ||
1508 is_migrate_isolate(migratetype))) {
1509 migratetype = get_pfnblock_migratetype(page, pfn);
1511 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1512 spin_unlock_irqrestore(&zone->lock, flags);
1515 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1516 unsigned long zone, int nid)
1518 mm_zero_struct_page(page);
1519 set_page_links(page, zone, nid, pfn);
1520 init_page_count(page);
1521 page_mapcount_reset(page);
1522 page_cpupid_reset_last(page);
1523 page_kasan_tag_reset(page);
1525 INIT_LIST_HEAD(&page->lru);
1526 #ifdef WANT_PAGE_VIRTUAL
1527 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1528 if (!is_highmem_idx(zone))
1529 set_page_address(page, __va(pfn << PAGE_SHIFT));
1533 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1534 static void __meminit init_reserved_page(unsigned long pfn)
1539 if (!early_page_uninitialised(pfn))
1542 nid = early_pfn_to_nid(pfn);
1543 pgdat = NODE_DATA(nid);
1545 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1546 struct zone *zone = &pgdat->node_zones[zid];
1548 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1551 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1554 static inline void init_reserved_page(unsigned long pfn)
1557 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1560 * Initialised pages do not have PageReserved set. This function is
1561 * called for each range allocated by the bootmem allocator and
1562 * marks the pages PageReserved. The remaining valid pages are later
1563 * sent to the buddy page allocator.
1565 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1567 unsigned long start_pfn = PFN_DOWN(start);
1568 unsigned long end_pfn = PFN_UP(end);
1570 for (; start_pfn < end_pfn; start_pfn++) {
1571 if (pfn_valid(start_pfn)) {
1572 struct page *page = pfn_to_page(start_pfn);
1574 init_reserved_page(start_pfn);
1576 /* Avoid false-positive PageTail() */
1577 INIT_LIST_HEAD(&page->lru);
1580 * no need for atomic set_bit because the struct
1581 * page is not visible yet so nobody should
1584 __SetPageReserved(page);
1589 static void __free_pages_ok(struct page *page, unsigned int order,
1592 unsigned long flags;
1594 unsigned long pfn = page_to_pfn(page);
1595 struct zone *zone = page_zone(page);
1597 if (!free_pages_prepare(page, order, true, fpi_flags))
1600 migratetype = get_pfnblock_migratetype(page, pfn);
1602 spin_lock_irqsave(&zone->lock, flags);
1603 if (unlikely(has_isolate_pageblock(zone) ||
1604 is_migrate_isolate(migratetype))) {
1605 migratetype = get_pfnblock_migratetype(page, pfn);
1607 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1608 spin_unlock_irqrestore(&zone->lock, flags);
1610 __count_vm_events(PGFREE, 1 << order);
1613 void __free_pages_core(struct page *page, unsigned int order)
1615 unsigned int nr_pages = 1 << order;
1616 struct page *p = page;
1620 * When initializing the memmap, __init_single_page() sets the refcount
1621 * of all pages to 1 ("allocated"/"not free"). We have to set the
1622 * refcount of all involved pages to 0.
1625 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1627 __ClearPageReserved(p);
1628 set_page_count(p, 0);
1630 __ClearPageReserved(p);
1631 set_page_count(p, 0);
1633 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1636 * Bypass PCP and place fresh pages right to the tail, primarily
1637 * relevant for memory onlining.
1639 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1642 #ifdef CONFIG_NEED_MULTIPLE_NODES
1645 * During memory init memblocks map pfns to nids. The search is expensive and
1646 * this caches recent lookups. The implementation of __early_pfn_to_nid
1647 * treats start/end as pfns.
1649 struct mminit_pfnnid_cache {
1650 unsigned long last_start;
1651 unsigned long last_end;
1655 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1658 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1660 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1661 struct mminit_pfnnid_cache *state)
1663 unsigned long start_pfn, end_pfn;
1666 if (state->last_start <= pfn && pfn < state->last_end)
1667 return state->last_nid;
1669 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1670 if (nid != NUMA_NO_NODE) {
1671 state->last_start = start_pfn;
1672 state->last_end = end_pfn;
1673 state->last_nid = nid;
1679 int __meminit early_pfn_to_nid(unsigned long pfn)
1681 static DEFINE_SPINLOCK(early_pfn_lock);
1684 spin_lock(&early_pfn_lock);
1685 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1687 nid = first_online_node;
1688 spin_unlock(&early_pfn_lock);
1692 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1694 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1697 if (early_page_uninitialised(pfn))
1699 __free_pages_core(page, order);
1703 * Check that the whole (or subset of) a pageblock given by the interval of
1704 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1705 * with the migration of free compaction scanner. The scanners then need to
1706 * use only pfn_valid_within() check for arches that allow holes within
1709 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1711 * It's possible on some configurations to have a setup like node0 node1 node0
1712 * i.e. it's possible that all pages within a zones range of pages do not
1713 * belong to a single zone. We assume that a border between node0 and node1
1714 * can occur within a single pageblock, but not a node0 node1 node0
1715 * interleaving within a single pageblock. It is therefore sufficient to check
1716 * the first and last page of a pageblock and avoid checking each individual
1717 * page in a pageblock.
1719 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1720 unsigned long end_pfn, struct zone *zone)
1722 struct page *start_page;
1723 struct page *end_page;
1725 /* end_pfn is one past the range we are checking */
1728 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1731 start_page = pfn_to_online_page(start_pfn);
1735 if (page_zone(start_page) != zone)
1738 end_page = pfn_to_page(end_pfn);
1740 /* This gives a shorter code than deriving page_zone(end_page) */
1741 if (page_zone_id(start_page) != page_zone_id(end_page))
1747 void set_zone_contiguous(struct zone *zone)
1749 unsigned long block_start_pfn = zone->zone_start_pfn;
1750 unsigned long block_end_pfn;
1752 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1753 for (; block_start_pfn < zone_end_pfn(zone);
1754 block_start_pfn = block_end_pfn,
1755 block_end_pfn += pageblock_nr_pages) {
1757 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1759 if (!__pageblock_pfn_to_page(block_start_pfn,
1760 block_end_pfn, zone))
1765 /* We confirm that there is no hole */
1766 zone->contiguous = true;
1769 void clear_zone_contiguous(struct zone *zone)
1771 zone->contiguous = false;
1774 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1775 static void __init deferred_free_range(unsigned long pfn,
1776 unsigned long nr_pages)
1784 page = pfn_to_page(pfn);
1786 /* Free a large naturally-aligned chunk if possible */
1787 if (nr_pages == pageblock_nr_pages &&
1788 (pfn & (pageblock_nr_pages - 1)) == 0) {
1789 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1790 __free_pages_core(page, pageblock_order);
1794 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1795 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1796 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1797 __free_pages_core(page, 0);
1801 /* Completion tracking for deferred_init_memmap() threads */
1802 static atomic_t pgdat_init_n_undone __initdata;
1803 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1805 static inline void __init pgdat_init_report_one_done(void)
1807 if (atomic_dec_and_test(&pgdat_init_n_undone))
1808 complete(&pgdat_init_all_done_comp);
1812 * Returns true if page needs to be initialized or freed to buddy allocator.
1814 * First we check if pfn is valid on architectures where it is possible to have
1815 * holes within pageblock_nr_pages. On systems where it is not possible, this
1816 * function is optimized out.
1818 * Then, we check if a current large page is valid by only checking the validity
1821 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1823 if (!pfn_valid_within(pfn))
1825 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1831 * Free pages to buddy allocator. Try to free aligned pages in
1832 * pageblock_nr_pages sizes.
1834 static void __init deferred_free_pages(unsigned long pfn,
1835 unsigned long end_pfn)
1837 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1838 unsigned long nr_free = 0;
1840 for (; pfn < end_pfn; pfn++) {
1841 if (!deferred_pfn_valid(pfn)) {
1842 deferred_free_range(pfn - nr_free, nr_free);
1844 } else if (!(pfn & nr_pgmask)) {
1845 deferred_free_range(pfn - nr_free, nr_free);
1851 /* Free the last block of pages to allocator */
1852 deferred_free_range(pfn - nr_free, nr_free);
1856 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1857 * by performing it only once every pageblock_nr_pages.
1858 * Return number of pages initialized.
1860 static unsigned long __init deferred_init_pages(struct zone *zone,
1862 unsigned long end_pfn)
1864 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1865 int nid = zone_to_nid(zone);
1866 unsigned long nr_pages = 0;
1867 int zid = zone_idx(zone);
1868 struct page *page = NULL;
1870 for (; pfn < end_pfn; pfn++) {
1871 if (!deferred_pfn_valid(pfn)) {
1874 } else if (!page || !(pfn & nr_pgmask)) {
1875 page = pfn_to_page(pfn);
1879 __init_single_page(page, pfn, zid, nid);
1886 * This function is meant to pre-load the iterator for the zone init.
1887 * Specifically it walks through the ranges until we are caught up to the
1888 * first_init_pfn value and exits there. If we never encounter the value we
1889 * return false indicating there are no valid ranges left.
1892 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1893 unsigned long *spfn, unsigned long *epfn,
1894 unsigned long first_init_pfn)
1899 * Start out by walking through the ranges in this zone that have
1900 * already been initialized. We don't need to do anything with them
1901 * so we just need to flush them out of the system.
1903 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1904 if (*epfn <= first_init_pfn)
1906 if (*spfn < first_init_pfn)
1907 *spfn = first_init_pfn;
1916 * Initialize and free pages. We do it in two loops: first we initialize
1917 * struct page, then free to buddy allocator, because while we are
1918 * freeing pages we can access pages that are ahead (computing buddy
1919 * page in __free_one_page()).
1921 * In order to try and keep some memory in the cache we have the loop
1922 * broken along max page order boundaries. This way we will not cause
1923 * any issues with the buddy page computation.
1925 static unsigned long __init
1926 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1927 unsigned long *end_pfn)
1929 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1930 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1931 unsigned long nr_pages = 0;
1934 /* First we loop through and initialize the page values */
1935 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1938 if (mo_pfn <= *start_pfn)
1941 t = min(mo_pfn, *end_pfn);
1942 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1944 if (mo_pfn < *end_pfn) {
1945 *start_pfn = mo_pfn;
1950 /* Reset values and now loop through freeing pages as needed */
1953 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1959 t = min(mo_pfn, epfn);
1960 deferred_free_pages(spfn, t);
1970 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1973 unsigned long spfn, epfn;
1974 struct zone *zone = arg;
1977 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1980 * Initialize and free pages in MAX_ORDER sized increments so that we
1981 * can avoid introducing any issues with the buddy allocator.
1983 while (spfn < end_pfn) {
1984 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1989 /* An arch may override for more concurrency. */
1991 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1996 /* Initialise remaining memory on a node */
1997 static int __init deferred_init_memmap(void *data)
1999 pg_data_t *pgdat = data;
2000 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2001 unsigned long spfn = 0, epfn = 0;
2002 unsigned long first_init_pfn, flags;
2003 unsigned long start = jiffies;
2005 int zid, max_threads;
2008 /* Bind memory initialisation thread to a local node if possible */
2009 if (!cpumask_empty(cpumask))
2010 set_cpus_allowed_ptr(current, cpumask);
2012 pgdat_resize_lock(pgdat, &flags);
2013 first_init_pfn = pgdat->first_deferred_pfn;
2014 if (first_init_pfn == ULONG_MAX) {
2015 pgdat_resize_unlock(pgdat, &flags);
2016 pgdat_init_report_one_done();
2020 /* Sanity check boundaries */
2021 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2022 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2023 pgdat->first_deferred_pfn = ULONG_MAX;
2026 * Once we unlock here, the zone cannot be grown anymore, thus if an
2027 * interrupt thread must allocate this early in boot, zone must be
2028 * pre-grown prior to start of deferred page initialization.
2030 pgdat_resize_unlock(pgdat, &flags);
2032 /* Only the highest zone is deferred so find it */
2033 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2034 zone = pgdat->node_zones + zid;
2035 if (first_init_pfn < zone_end_pfn(zone))
2039 /* If the zone is empty somebody else may have cleared out the zone */
2040 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2044 max_threads = deferred_page_init_max_threads(cpumask);
2046 while (spfn < epfn) {
2047 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2048 struct padata_mt_job job = {
2049 .thread_fn = deferred_init_memmap_chunk,
2052 .size = epfn_align - spfn,
2053 .align = PAGES_PER_SECTION,
2054 .min_chunk = PAGES_PER_SECTION,
2055 .max_threads = max_threads,
2058 padata_do_multithreaded(&job);
2059 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2063 /* Sanity check that the next zone really is unpopulated */
2064 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2066 pr_info("node %d deferred pages initialised in %ums\n",
2067 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2069 pgdat_init_report_one_done();
2074 * If this zone has deferred pages, try to grow it by initializing enough
2075 * deferred pages to satisfy the allocation specified by order, rounded up to
2076 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2077 * of SECTION_SIZE bytes by initializing struct pages in increments of
2078 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2080 * Return true when zone was grown, otherwise return false. We return true even
2081 * when we grow less than requested, to let the caller decide if there are
2082 * enough pages to satisfy the allocation.
2084 * Note: We use noinline because this function is needed only during boot, and
2085 * it is called from a __ref function _deferred_grow_zone. This way we are
2086 * making sure that it is not inlined into permanent text section.
2088 static noinline bool __init
2089 deferred_grow_zone(struct zone *zone, unsigned int order)
2091 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2092 pg_data_t *pgdat = zone->zone_pgdat;
2093 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2094 unsigned long spfn, epfn, flags;
2095 unsigned long nr_pages = 0;
2098 /* Only the last zone may have deferred pages */
2099 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2102 pgdat_resize_lock(pgdat, &flags);
2105 * If someone grew this zone while we were waiting for spinlock, return
2106 * true, as there might be enough pages already.
2108 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2109 pgdat_resize_unlock(pgdat, &flags);
2113 /* If the zone is empty somebody else may have cleared out the zone */
2114 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2115 first_deferred_pfn)) {
2116 pgdat->first_deferred_pfn = ULONG_MAX;
2117 pgdat_resize_unlock(pgdat, &flags);
2118 /* Retry only once. */
2119 return first_deferred_pfn != ULONG_MAX;
2123 * Initialize and free pages in MAX_ORDER sized increments so
2124 * that we can avoid introducing any issues with the buddy
2127 while (spfn < epfn) {
2128 /* update our first deferred PFN for this section */
2129 first_deferred_pfn = spfn;
2131 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2132 touch_nmi_watchdog();
2134 /* We should only stop along section boundaries */
2135 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2138 /* If our quota has been met we can stop here */
2139 if (nr_pages >= nr_pages_needed)
2143 pgdat->first_deferred_pfn = spfn;
2144 pgdat_resize_unlock(pgdat, &flags);
2146 return nr_pages > 0;
2150 * deferred_grow_zone() is __init, but it is called from
2151 * get_page_from_freelist() during early boot until deferred_pages permanently
2152 * disables this call. This is why we have refdata wrapper to avoid warning,
2153 * and to ensure that the function body gets unloaded.
2156 _deferred_grow_zone(struct zone *zone, unsigned int order)
2158 return deferred_grow_zone(zone, order);
2161 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2163 void __init page_alloc_init_late(void)
2168 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2170 /* There will be num_node_state(N_MEMORY) threads */
2171 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2172 for_each_node_state(nid, N_MEMORY) {
2173 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2176 /* Block until all are initialised */
2177 wait_for_completion(&pgdat_init_all_done_comp);
2180 * The number of managed pages has changed due to the initialisation
2181 * so the pcpu batch and high limits needs to be updated or the limits
2182 * will be artificially small.
2184 for_each_populated_zone(zone)
2185 zone_pcp_update(zone);
2188 * We initialized the rest of the deferred pages. Permanently disable
2189 * on-demand struct page initialization.
2191 static_branch_disable(&deferred_pages);
2193 /* Reinit limits that are based on free pages after the kernel is up */
2194 files_maxfiles_init();
2199 /* Discard memblock private memory */
2202 for_each_node_state(nid, N_MEMORY)
2203 shuffle_free_memory(NODE_DATA(nid));
2205 for_each_populated_zone(zone)
2206 set_zone_contiguous(zone);
2210 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2211 void __init init_cma_reserved_pageblock(struct page *page)
2213 unsigned i = pageblock_nr_pages;
2214 struct page *p = page;
2217 __ClearPageReserved(p);
2218 set_page_count(p, 0);
2221 set_pageblock_migratetype(page, MIGRATE_CMA);
2223 if (pageblock_order >= MAX_ORDER) {
2224 i = pageblock_nr_pages;
2227 set_page_refcounted(p);
2228 __free_pages(p, MAX_ORDER - 1);
2229 p += MAX_ORDER_NR_PAGES;
2230 } while (i -= MAX_ORDER_NR_PAGES);
2232 set_page_refcounted(page);
2233 __free_pages(page, pageblock_order);
2236 adjust_managed_page_count(page, pageblock_nr_pages);
2237 page_zone(page)->cma_pages += pageblock_nr_pages;
2242 * The order of subdivision here is critical for the IO subsystem.
2243 * Please do not alter this order without good reasons and regression
2244 * testing. Specifically, as large blocks of memory are subdivided,
2245 * the order in which smaller blocks are delivered depends on the order
2246 * they're subdivided in this function. This is the primary factor
2247 * influencing the order in which pages are delivered to the IO
2248 * subsystem according to empirical testing, and this is also justified
2249 * by considering the behavior of a buddy system containing a single
2250 * large block of memory acted on by a series of small allocations.
2251 * This behavior is a critical factor in sglist merging's success.
2255 static inline void expand(struct zone *zone, struct page *page,
2256 int low, int high, int migratetype)
2258 unsigned long size = 1 << high;
2260 while (high > low) {
2263 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2266 * Mark as guard pages (or page), that will allow to
2267 * merge back to allocator when buddy will be freed.
2268 * Corresponding page table entries will not be touched,
2269 * pages will stay not present in virtual address space
2271 if (set_page_guard(zone, &page[size], high, migratetype))
2274 add_to_free_list(&page[size], zone, high, migratetype);
2275 set_buddy_order(&page[size], high);
2279 static void check_new_page_bad(struct page *page)
2281 if (unlikely(page->flags & __PG_HWPOISON)) {
2282 /* Don't complain about hwpoisoned pages */
2283 page_mapcount_reset(page); /* remove PageBuddy */
2288 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2292 * This page is about to be returned from the page allocator
2294 static inline int check_new_page(struct page *page)
2296 if (likely(page_expected_state(page,
2297 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2300 check_new_page_bad(page);
2304 #ifdef CONFIG_DEBUG_VM
2306 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2307 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2308 * also checked when pcp lists are refilled from the free lists.
2310 static inline bool check_pcp_refill(struct page *page)
2312 if (debug_pagealloc_enabled_static())
2313 return check_new_page(page);
2318 static inline bool check_new_pcp(struct page *page)
2320 return check_new_page(page);
2324 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2325 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2326 * enabled, they are also checked when being allocated from the pcp lists.
2328 static inline bool check_pcp_refill(struct page *page)
2330 return check_new_page(page);
2332 static inline bool check_new_pcp(struct page *page)
2334 if (debug_pagealloc_enabled_static())
2335 return check_new_page(page);
2339 #endif /* CONFIG_DEBUG_VM */
2341 static bool check_new_pages(struct page *page, unsigned int order)
2344 for (i = 0; i < (1 << order); i++) {
2345 struct page *p = page + i;
2347 if (unlikely(check_new_page(p)))
2354 inline void post_alloc_hook(struct page *page, unsigned int order,
2359 set_page_private(page, 0);
2360 set_page_refcounted(page);
2362 arch_alloc_page(page, order);
2363 debug_pagealloc_map_pages(page, 1 << order);
2366 * Page unpoisoning must happen before memory initialization.
2367 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2368 * allocations and the page unpoisoning code will complain.
2370 kernel_unpoison_pages(page, 1 << order);
2373 * As memory initialization might be integrated into KASAN,
2374 * kasan_alloc_pages and kernel_init_free_pages must be
2375 * kept together to avoid discrepancies in behavior.
2377 init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2378 kasan_alloc_pages(page, order, init);
2379 if (init && !kasan_has_integrated_init())
2380 kernel_init_free_pages(page, 1 << order);
2382 set_page_owner(page, order, gfp_flags);
2385 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2386 unsigned int alloc_flags)
2388 post_alloc_hook(page, order, gfp_flags);
2390 if (order && (gfp_flags & __GFP_COMP))
2391 prep_compound_page(page, order);
2394 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2395 * allocate the page. The expectation is that the caller is taking
2396 * steps that will free more memory. The caller should avoid the page
2397 * being used for !PFMEMALLOC purposes.
2399 if (alloc_flags & ALLOC_NO_WATERMARKS)
2400 set_page_pfmemalloc(page);
2402 clear_page_pfmemalloc(page);
2406 * Go through the free lists for the given migratetype and remove
2407 * the smallest available page from the freelists
2409 static __always_inline
2410 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2413 unsigned int current_order;
2414 struct free_area *area;
2417 /* Find a page of the appropriate size in the preferred list */
2418 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2419 area = &(zone->free_area[current_order]);
2420 page = get_page_from_free_area(area, migratetype);
2423 del_page_from_free_list(page, zone, current_order);
2424 expand(zone, page, order, current_order, migratetype);
2425 set_pcppage_migratetype(page, migratetype);
2434 * This array describes the order lists are fallen back to when
2435 * the free lists for the desirable migrate type are depleted
2437 static int fallbacks[MIGRATE_TYPES][3] = {
2438 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2439 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2440 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2442 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2444 #ifdef CONFIG_MEMORY_ISOLATION
2445 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2450 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2453 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2456 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2457 unsigned int order) { return NULL; }
2461 * Move the free pages in a range to the freelist tail of the requested type.
2462 * Note that start_page and end_pages are not aligned on a pageblock
2463 * boundary. If alignment is required, use move_freepages_block()
2465 static int move_freepages(struct zone *zone,
2466 unsigned long start_pfn, unsigned long end_pfn,
2467 int migratetype, int *num_movable)
2472 int pages_moved = 0;
2474 for (pfn = start_pfn; pfn <= end_pfn;) {
2475 if (!pfn_valid_within(pfn)) {
2480 page = pfn_to_page(pfn);
2481 if (!PageBuddy(page)) {
2483 * We assume that pages that could be isolated for
2484 * migration are movable. But we don't actually try
2485 * isolating, as that would be expensive.
2488 (PageLRU(page) || __PageMovable(page)))
2494 /* Make sure we are not inadvertently changing nodes */
2495 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2496 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2498 order = buddy_order(page);
2499 move_to_free_list(page, zone, order, migratetype);
2501 pages_moved += 1 << order;
2507 int move_freepages_block(struct zone *zone, struct page *page,
2508 int migratetype, int *num_movable)
2510 unsigned long start_pfn, end_pfn, pfn;
2515 pfn = page_to_pfn(page);
2516 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2517 end_pfn = start_pfn + pageblock_nr_pages - 1;
2519 /* Do not cross zone boundaries */
2520 if (!zone_spans_pfn(zone, start_pfn))
2522 if (!zone_spans_pfn(zone, end_pfn))
2525 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2529 static void change_pageblock_range(struct page *pageblock_page,
2530 int start_order, int migratetype)
2532 int nr_pageblocks = 1 << (start_order - pageblock_order);
2534 while (nr_pageblocks--) {
2535 set_pageblock_migratetype(pageblock_page, migratetype);
2536 pageblock_page += pageblock_nr_pages;
2541 * When we are falling back to another migratetype during allocation, try to
2542 * steal extra free pages from the same pageblocks to satisfy further
2543 * allocations, instead of polluting multiple pageblocks.
2545 * If we are stealing a relatively large buddy page, it is likely there will
2546 * be more free pages in the pageblock, so try to steal them all. For
2547 * reclaimable and unmovable allocations, we steal regardless of page size,
2548 * as fragmentation caused by those allocations polluting movable pageblocks
2549 * is worse than movable allocations stealing from unmovable and reclaimable
2552 static bool can_steal_fallback(unsigned int order, int start_mt)
2555 * Leaving this order check is intended, although there is
2556 * relaxed order check in next check. The reason is that
2557 * we can actually steal whole pageblock if this condition met,
2558 * but, below check doesn't guarantee it and that is just heuristic
2559 * so could be changed anytime.
2561 if (order >= pageblock_order)
2564 if (order >= pageblock_order / 2 ||
2565 start_mt == MIGRATE_RECLAIMABLE ||
2566 start_mt == MIGRATE_UNMOVABLE ||
2567 page_group_by_mobility_disabled)
2573 static inline bool boost_watermark(struct zone *zone)
2575 unsigned long max_boost;
2577 if (!watermark_boost_factor)
2580 * Don't bother in zones that are unlikely to produce results.
2581 * On small machines, including kdump capture kernels running
2582 * in a small area, boosting the watermark can cause an out of
2583 * memory situation immediately.
2585 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2588 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2589 watermark_boost_factor, 10000);
2592 * high watermark may be uninitialised if fragmentation occurs
2593 * very early in boot so do not boost. We do not fall
2594 * through and boost by pageblock_nr_pages as failing
2595 * allocations that early means that reclaim is not going
2596 * to help and it may even be impossible to reclaim the
2597 * boosted watermark resulting in a hang.
2602 max_boost = max(pageblock_nr_pages, max_boost);
2604 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2611 * This function implements actual steal behaviour. If order is large enough,
2612 * we can steal whole pageblock. If not, we first move freepages in this
2613 * pageblock to our migratetype and determine how many already-allocated pages
2614 * are there in the pageblock with a compatible migratetype. If at least half
2615 * of pages are free or compatible, we can change migratetype of the pageblock
2616 * itself, so pages freed in the future will be put on the correct free list.
2618 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2619 unsigned int alloc_flags, int start_type, bool whole_block)
2621 unsigned int current_order = buddy_order(page);
2622 int free_pages, movable_pages, alike_pages;
2625 old_block_type = get_pageblock_migratetype(page);
2628 * This can happen due to races and we want to prevent broken
2629 * highatomic accounting.
2631 if (is_migrate_highatomic(old_block_type))
2634 /* Take ownership for orders >= pageblock_order */
2635 if (current_order >= pageblock_order) {
2636 change_pageblock_range(page, current_order, start_type);
2641 * Boost watermarks to increase reclaim pressure to reduce the
2642 * likelihood of future fallbacks. Wake kswapd now as the node
2643 * may be balanced overall and kswapd will not wake naturally.
2645 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2646 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2648 /* We are not allowed to try stealing from the whole block */
2652 free_pages = move_freepages_block(zone, page, start_type,
2655 * Determine how many pages are compatible with our allocation.
2656 * For movable allocation, it's the number of movable pages which
2657 * we just obtained. For other types it's a bit more tricky.
2659 if (start_type == MIGRATE_MOVABLE) {
2660 alike_pages = movable_pages;
2663 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2664 * to MOVABLE pageblock, consider all non-movable pages as
2665 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2666 * vice versa, be conservative since we can't distinguish the
2667 * exact migratetype of non-movable pages.
2669 if (old_block_type == MIGRATE_MOVABLE)
2670 alike_pages = pageblock_nr_pages
2671 - (free_pages + movable_pages);
2676 /* moving whole block can fail due to zone boundary conditions */
2681 * If a sufficient number of pages in the block are either free or of
2682 * comparable migratability as our allocation, claim the whole block.
2684 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2685 page_group_by_mobility_disabled)
2686 set_pageblock_migratetype(page, start_type);
2691 move_to_free_list(page, zone, current_order, start_type);
2695 * Check whether there is a suitable fallback freepage with requested order.
2696 * If only_stealable is true, this function returns fallback_mt only if
2697 * we can steal other freepages all together. This would help to reduce
2698 * fragmentation due to mixed migratetype pages in one pageblock.
2700 int find_suitable_fallback(struct free_area *area, unsigned int order,
2701 int migratetype, bool only_stealable, bool *can_steal)
2706 if (area->nr_free == 0)
2711 fallback_mt = fallbacks[migratetype][i];
2712 if (fallback_mt == MIGRATE_TYPES)
2715 if (free_area_empty(area, fallback_mt))
2718 if (can_steal_fallback(order, migratetype))
2721 if (!only_stealable)
2732 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2733 * there are no empty page blocks that contain a page with a suitable order
2735 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2736 unsigned int alloc_order)
2739 unsigned long max_managed, flags;
2742 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2743 * Check is race-prone but harmless.
2745 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2746 if (zone->nr_reserved_highatomic >= max_managed)
2749 spin_lock_irqsave(&zone->lock, flags);
2751 /* Recheck the nr_reserved_highatomic limit under the lock */
2752 if (zone->nr_reserved_highatomic >= max_managed)
2756 mt = get_pageblock_migratetype(page);
2757 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2758 && !is_migrate_cma(mt)) {
2759 zone->nr_reserved_highatomic += pageblock_nr_pages;
2760 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2761 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2765 spin_unlock_irqrestore(&zone->lock, flags);
2769 * Used when an allocation is about to fail under memory pressure. This
2770 * potentially hurts the reliability of high-order allocations when under
2771 * intense memory pressure but failed atomic allocations should be easier
2772 * to recover from than an OOM.
2774 * If @force is true, try to unreserve a pageblock even though highatomic
2775 * pageblock is exhausted.
2777 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2780 struct zonelist *zonelist = ac->zonelist;
2781 unsigned long flags;
2788 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2791 * Preserve at least one pageblock unless memory pressure
2794 if (!force && zone->nr_reserved_highatomic <=
2798 spin_lock_irqsave(&zone->lock, flags);
2799 for (order = 0; order < MAX_ORDER; order++) {
2800 struct free_area *area = &(zone->free_area[order]);
2802 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2807 * In page freeing path, migratetype change is racy so
2808 * we can counter several free pages in a pageblock
2809 * in this loop although we changed the pageblock type
2810 * from highatomic to ac->migratetype. So we should
2811 * adjust the count once.
2813 if (is_migrate_highatomic_page(page)) {
2815 * It should never happen but changes to
2816 * locking could inadvertently allow a per-cpu
2817 * drain to add pages to MIGRATE_HIGHATOMIC
2818 * while unreserving so be safe and watch for
2821 zone->nr_reserved_highatomic -= min(
2823 zone->nr_reserved_highatomic);
2827 * Convert to ac->migratetype and avoid the normal
2828 * pageblock stealing heuristics. Minimally, the caller
2829 * is doing the work and needs the pages. More
2830 * importantly, if the block was always converted to
2831 * MIGRATE_UNMOVABLE or another type then the number
2832 * of pageblocks that cannot be completely freed
2835 set_pageblock_migratetype(page, ac->migratetype);
2836 ret = move_freepages_block(zone, page, ac->migratetype,
2839 spin_unlock_irqrestore(&zone->lock, flags);
2843 spin_unlock_irqrestore(&zone->lock, flags);
2850 * Try finding a free buddy page on the fallback list and put it on the free
2851 * list of requested migratetype, possibly along with other pages from the same
2852 * block, depending on fragmentation avoidance heuristics. Returns true if
2853 * fallback was found so that __rmqueue_smallest() can grab it.
2855 * The use of signed ints for order and current_order is a deliberate
2856 * deviation from the rest of this file, to make the for loop
2857 * condition simpler.
2859 static __always_inline bool
2860 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2861 unsigned int alloc_flags)
2863 struct free_area *area;
2865 int min_order = order;
2871 * Do not steal pages from freelists belonging to other pageblocks
2872 * i.e. orders < pageblock_order. If there are no local zones free,
2873 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2875 if (alloc_flags & ALLOC_NOFRAGMENT)
2876 min_order = pageblock_order;
2879 * Find the largest available free page in the other list. This roughly
2880 * approximates finding the pageblock with the most free pages, which
2881 * would be too costly to do exactly.
2883 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2885 area = &(zone->free_area[current_order]);
2886 fallback_mt = find_suitable_fallback(area, current_order,
2887 start_migratetype, false, &can_steal);
2888 if (fallback_mt == -1)
2892 * We cannot steal all free pages from the pageblock and the
2893 * requested migratetype is movable. In that case it's better to
2894 * steal and split the smallest available page instead of the
2895 * largest available page, because even if the next movable
2896 * allocation falls back into a different pageblock than this
2897 * one, it won't cause permanent fragmentation.
2899 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2900 && current_order > order)
2909 for (current_order = order; current_order < MAX_ORDER;
2911 area = &(zone->free_area[current_order]);
2912 fallback_mt = find_suitable_fallback(area, current_order,
2913 start_migratetype, false, &can_steal);
2914 if (fallback_mt != -1)
2919 * This should not happen - we already found a suitable fallback
2920 * when looking for the largest page.
2922 VM_BUG_ON(current_order == MAX_ORDER);
2925 page = get_page_from_free_area(area, fallback_mt);
2927 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2930 trace_mm_page_alloc_extfrag(page, order, current_order,
2931 start_migratetype, fallback_mt);
2938 * Do the hard work of removing an element from the buddy allocator.
2939 * Call me with the zone->lock already held.
2941 static __always_inline struct page *
2942 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2943 unsigned int alloc_flags)
2947 if (IS_ENABLED(CONFIG_CMA)) {
2949 * Balance movable allocations between regular and CMA areas by
2950 * allocating from CMA when over half of the zone's free memory
2951 * is in the CMA area.
2953 if (alloc_flags & ALLOC_CMA &&
2954 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2955 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2956 page = __rmqueue_cma_fallback(zone, order);
2962 page = __rmqueue_smallest(zone, order, migratetype);
2963 if (unlikely(!page)) {
2964 if (alloc_flags & ALLOC_CMA)
2965 page = __rmqueue_cma_fallback(zone, order);
2967 if (!page && __rmqueue_fallback(zone, order, migratetype,
2973 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2978 * Obtain a specified number of elements from the buddy allocator, all under
2979 * a single hold of the lock, for efficiency. Add them to the supplied list.
2980 * Returns the number of new pages which were placed at *list.
2982 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2983 unsigned long count, struct list_head *list,
2984 int migratetype, unsigned int alloc_flags)
2986 int i, allocated = 0;
2989 * local_lock_irq held so equivalent to spin_lock_irqsave for
2990 * both PREEMPT_RT and non-PREEMPT_RT configurations.
2992 spin_lock(&zone->lock);
2993 for (i = 0; i < count; ++i) {
2994 struct page *page = __rmqueue(zone, order, migratetype,
2996 if (unlikely(page == NULL))
2999 if (unlikely(check_pcp_refill(page)))
3003 * Split buddy pages returned by expand() are received here in
3004 * physical page order. The page is added to the tail of
3005 * caller's list. From the callers perspective, the linked list
3006 * is ordered by page number under some conditions. This is
3007 * useful for IO devices that can forward direction from the
3008 * head, thus also in the physical page order. This is useful
3009 * for IO devices that can merge IO requests if the physical
3010 * pages are ordered properly.
3012 list_add_tail(&page->lru, list);
3014 if (is_migrate_cma(get_pcppage_migratetype(page)))
3015 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3020 * i pages were removed from the buddy list even if some leak due
3021 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3022 * on i. Do not confuse with 'allocated' which is the number of
3023 * pages added to the pcp list.
3025 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3026 spin_unlock(&zone->lock);
3032 * Called from the vmstat counter updater to drain pagesets of this
3033 * currently executing processor on remote nodes after they have
3036 * Note that this function must be called with the thread pinned to
3037 * a single processor.
3039 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3041 unsigned long flags;
3042 int to_drain, batch;
3044 local_lock_irqsave(&pagesets.lock, flags);
3045 batch = READ_ONCE(pcp->batch);
3046 to_drain = min(pcp->count, batch);
3048 free_pcppages_bulk(zone, to_drain, pcp);
3049 local_unlock_irqrestore(&pagesets.lock, flags);
3054 * Drain pcplists of the indicated processor and zone.
3056 * The processor must either be the current processor and the
3057 * thread pinned to the current processor or a processor that
3060 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3062 unsigned long flags;
3063 struct per_cpu_pages *pcp;
3065 local_lock_irqsave(&pagesets.lock, flags);
3067 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3069 free_pcppages_bulk(zone, pcp->count, pcp);
3071 local_unlock_irqrestore(&pagesets.lock, flags);
3075 * Drain pcplists of all zones on the indicated processor.
3077 * The processor must either be the current processor and the
3078 * thread pinned to the current processor or a processor that
3081 static void drain_pages(unsigned int cpu)
3085 for_each_populated_zone(zone) {
3086 drain_pages_zone(cpu, zone);
3091 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3093 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3094 * the single zone's pages.
3096 void drain_local_pages(struct zone *zone)
3098 int cpu = smp_processor_id();
3101 drain_pages_zone(cpu, zone);
3106 static void drain_local_pages_wq(struct work_struct *work)
3108 struct pcpu_drain *drain;
3110 drain = container_of(work, struct pcpu_drain, work);
3113 * drain_all_pages doesn't use proper cpu hotplug protection so
3114 * we can race with cpu offline when the WQ can move this from
3115 * a cpu pinned worker to an unbound one. We can operate on a different
3116 * cpu which is alright but we also have to make sure to not move to
3120 drain_local_pages(drain->zone);
3125 * The implementation of drain_all_pages(), exposing an extra parameter to
3126 * drain on all cpus.
3128 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3129 * not empty. The check for non-emptiness can however race with a free to
3130 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3131 * that need the guarantee that every CPU has drained can disable the
3132 * optimizing racy check.
3134 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3139 * Allocate in the BSS so we wont require allocation in
3140 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3142 static cpumask_t cpus_with_pcps;
3145 * Make sure nobody triggers this path before mm_percpu_wq is fully
3148 if (WARN_ON_ONCE(!mm_percpu_wq))
3152 * Do not drain if one is already in progress unless it's specific to
3153 * a zone. Such callers are primarily CMA and memory hotplug and need
3154 * the drain to be complete when the call returns.
3156 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3159 mutex_lock(&pcpu_drain_mutex);
3163 * We don't care about racing with CPU hotplug event
3164 * as offline notification will cause the notified
3165 * cpu to drain that CPU pcps and on_each_cpu_mask
3166 * disables preemption as part of its processing
3168 for_each_online_cpu(cpu) {
3169 struct per_cpu_pages *pcp;
3171 bool has_pcps = false;
3173 if (force_all_cpus) {
3175 * The pcp.count check is racy, some callers need a
3176 * guarantee that no cpu is missed.
3180 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3184 for_each_populated_zone(z) {
3185 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3194 cpumask_set_cpu(cpu, &cpus_with_pcps);
3196 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3199 for_each_cpu(cpu, &cpus_with_pcps) {
3200 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3203 INIT_WORK(&drain->work, drain_local_pages_wq);
3204 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3206 for_each_cpu(cpu, &cpus_with_pcps)
3207 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3209 mutex_unlock(&pcpu_drain_mutex);
3213 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3215 * When zone parameter is non-NULL, spill just the single zone's pages.
3217 * Note that this can be extremely slow as the draining happens in a workqueue.
3219 void drain_all_pages(struct zone *zone)
3221 __drain_all_pages(zone, false);
3224 #ifdef CONFIG_HIBERNATION
3227 * Touch the watchdog for every WD_PAGE_COUNT pages.
3229 #define WD_PAGE_COUNT (128*1024)
3231 void mark_free_pages(struct zone *zone)
3233 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3234 unsigned long flags;
3235 unsigned int order, t;
3238 if (zone_is_empty(zone))
3241 spin_lock_irqsave(&zone->lock, flags);
3243 max_zone_pfn = zone_end_pfn(zone);
3244 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3245 if (pfn_valid(pfn)) {
3246 page = pfn_to_page(pfn);
3248 if (!--page_count) {
3249 touch_nmi_watchdog();
3250 page_count = WD_PAGE_COUNT;
3253 if (page_zone(page) != zone)
3256 if (!swsusp_page_is_forbidden(page))
3257 swsusp_unset_page_free(page);
3260 for_each_migratetype_order(order, t) {
3261 list_for_each_entry(page,
3262 &zone->free_area[order].free_list[t], lru) {
3265 pfn = page_to_pfn(page);
3266 for (i = 0; i < (1UL << order); i++) {
3267 if (!--page_count) {
3268 touch_nmi_watchdog();
3269 page_count = WD_PAGE_COUNT;
3271 swsusp_set_page_free(pfn_to_page(pfn + i));
3275 spin_unlock_irqrestore(&zone->lock, flags);
3277 #endif /* CONFIG_PM */
3279 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3283 if (!free_pcp_prepare(page))
3286 migratetype = get_pfnblock_migratetype(page, pfn);
3287 set_pcppage_migratetype(page, migratetype);
3291 static void free_unref_page_commit(struct page *page, unsigned long pfn,
3294 struct zone *zone = page_zone(page);
3295 struct per_cpu_pages *pcp;
3297 __count_vm_event(PGFREE);
3298 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3299 list_add(&page->lru, &pcp->lists[migratetype]);
3301 if (pcp->count >= READ_ONCE(pcp->high))
3302 free_pcppages_bulk(zone, READ_ONCE(pcp->batch), pcp);
3306 * Free a 0-order page
3308 void free_unref_page(struct page *page)
3310 unsigned long flags;
3311 unsigned long pfn = page_to_pfn(page);
3314 if (!free_unref_page_prepare(page, pfn))
3318 * We only track unmovable, reclaimable and movable on pcp lists.
3319 * Place ISOLATE pages on the isolated list because they are being
3320 * offlined but treat HIGHATOMIC as movable pages so we can get those
3321 * areas back if necessary. Otherwise, we may have to free
3322 * excessively into the page allocator
3324 migratetype = get_pcppage_migratetype(page);
3325 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3326 if (unlikely(is_migrate_isolate(migratetype))) {
3327 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3330 migratetype = MIGRATE_MOVABLE;
3333 local_lock_irqsave(&pagesets.lock, flags);
3334 free_unref_page_commit(page, pfn, migratetype);
3335 local_unlock_irqrestore(&pagesets.lock, flags);
3339 * Free a list of 0-order pages
3341 void free_unref_page_list(struct list_head *list)
3343 struct page *page, *next;
3344 unsigned long flags, pfn;
3345 int batch_count = 0;
3348 /* Prepare pages for freeing */
3349 list_for_each_entry_safe(page, next, list, lru) {
3350 pfn = page_to_pfn(page);
3351 if (!free_unref_page_prepare(page, pfn))
3352 list_del(&page->lru);
3355 * Free isolated pages directly to the allocator, see
3356 * comment in free_unref_page.
3358 migratetype = get_pcppage_migratetype(page);
3359 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3360 if (unlikely(is_migrate_isolate(migratetype))) {
3361 list_del(&page->lru);
3362 free_one_page(page_zone(page), page, pfn, 0,
3363 migratetype, FPI_NONE);
3368 * Non-isolated types over MIGRATE_PCPTYPES get added
3369 * to the MIGRATE_MOVABLE pcp list.
3371 set_pcppage_migratetype(page, MIGRATE_MOVABLE);
3374 set_page_private(page, pfn);
3377 local_lock_irqsave(&pagesets.lock, flags);
3378 list_for_each_entry_safe(page, next, list, lru) {
3379 pfn = page_private(page);
3380 set_page_private(page, 0);
3381 migratetype = get_pcppage_migratetype(page);
3382 trace_mm_page_free_batched(page);
3383 free_unref_page_commit(page, pfn, migratetype);
3386 * Guard against excessive IRQ disabled times when we get
3387 * a large list of pages to free.
3389 if (++batch_count == SWAP_CLUSTER_MAX) {
3390 local_unlock_irqrestore(&pagesets.lock, flags);
3392 local_lock_irqsave(&pagesets.lock, flags);
3395 local_unlock_irqrestore(&pagesets.lock, flags);
3399 * split_page takes a non-compound higher-order page, and splits it into
3400 * n (1<<order) sub-pages: page[0..n]
3401 * Each sub-page must be freed individually.
3403 * Note: this is probably too low level an operation for use in drivers.
3404 * Please consult with lkml before using this in your driver.
3406 void split_page(struct page *page, unsigned int order)
3410 VM_BUG_ON_PAGE(PageCompound(page), page);
3411 VM_BUG_ON_PAGE(!page_count(page), page);
3413 for (i = 1; i < (1 << order); i++)
3414 set_page_refcounted(page + i);
3415 split_page_owner(page, 1 << order);
3416 split_page_memcg(page, 1 << order);
3418 EXPORT_SYMBOL_GPL(split_page);
3420 int __isolate_free_page(struct page *page, unsigned int order)
3422 unsigned long watermark;
3426 BUG_ON(!PageBuddy(page));
3428 zone = page_zone(page);
3429 mt = get_pageblock_migratetype(page);
3431 if (!is_migrate_isolate(mt)) {
3433 * Obey watermarks as if the page was being allocated. We can
3434 * emulate a high-order watermark check with a raised order-0
3435 * watermark, because we already know our high-order page
3438 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3439 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3442 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3445 /* Remove page from free list */
3447 del_page_from_free_list(page, zone, order);
3450 * Set the pageblock if the isolated page is at least half of a
3453 if (order >= pageblock_order - 1) {
3454 struct page *endpage = page + (1 << order) - 1;
3455 for (; page < endpage; page += pageblock_nr_pages) {
3456 int mt = get_pageblock_migratetype(page);
3457 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3458 && !is_migrate_highatomic(mt))
3459 set_pageblock_migratetype(page,
3465 return 1UL << order;
3469 * __putback_isolated_page - Return a now-isolated page back where we got it
3470 * @page: Page that was isolated
3471 * @order: Order of the isolated page
3472 * @mt: The page's pageblock's migratetype
3474 * This function is meant to return a page pulled from the free lists via
3475 * __isolate_free_page back to the free lists they were pulled from.
3477 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3479 struct zone *zone = page_zone(page);
3481 /* zone lock should be held when this function is called */
3482 lockdep_assert_held(&zone->lock);
3484 /* Return isolated page to tail of freelist. */
3485 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3486 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3490 * Update NUMA hit/miss statistics
3492 * Must be called with interrupts disabled.
3494 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3498 enum numa_stat_item local_stat = NUMA_LOCAL;
3500 /* skip numa counters update if numa stats is disabled */
3501 if (!static_branch_likely(&vm_numa_stat_key))
3504 if (zone_to_nid(z) != numa_node_id())
3505 local_stat = NUMA_OTHER;
3507 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3508 __count_numa_events(z, NUMA_HIT, nr_account);
3510 __count_numa_events(z, NUMA_MISS, nr_account);
3511 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3513 __count_numa_events(z, local_stat, nr_account);
3517 /* Remove page from the per-cpu list, caller must protect the list */
3519 struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3520 unsigned int alloc_flags,
3521 struct per_cpu_pages *pcp,
3522 struct list_head *list)
3527 if (list_empty(list)) {
3528 pcp->count += rmqueue_bulk(zone, 0,
3529 READ_ONCE(pcp->batch), list,
3530 migratetype, alloc_flags);
3531 if (unlikely(list_empty(list)))
3535 page = list_first_entry(list, struct page, lru);
3536 list_del(&page->lru);
3538 } while (check_new_pcp(page));
3543 /* Lock and remove page from the per-cpu list */
3544 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3545 struct zone *zone, gfp_t gfp_flags,
3546 int migratetype, unsigned int alloc_flags)
3548 struct per_cpu_pages *pcp;
3549 struct list_head *list;
3551 unsigned long flags;
3553 local_lock_irqsave(&pagesets.lock, flags);
3554 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3555 list = &pcp->lists[migratetype];
3556 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3557 local_unlock_irqrestore(&pagesets.lock, flags);
3559 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3560 zone_statistics(preferred_zone, zone, 1);
3566 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3569 struct page *rmqueue(struct zone *preferred_zone,
3570 struct zone *zone, unsigned int order,
3571 gfp_t gfp_flags, unsigned int alloc_flags,
3574 unsigned long flags;
3577 if (likely(order == 0)) {
3579 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3580 * we need to skip it when CMA area isn't allowed.
3582 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3583 migratetype != MIGRATE_MOVABLE) {
3584 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3585 migratetype, alloc_flags);
3591 * We most definitely don't want callers attempting to
3592 * allocate greater than order-1 page units with __GFP_NOFAIL.
3594 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3595 spin_lock_irqsave(&zone->lock, flags);
3600 * order-0 request can reach here when the pcplist is skipped
3601 * due to non-CMA allocation context. HIGHATOMIC area is
3602 * reserved for high-order atomic allocation, so order-0
3603 * request should skip it.
3605 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3606 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3608 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3611 page = __rmqueue(zone, order, migratetype, alloc_flags);
3612 } while (page && check_new_pages(page, order));
3616 __mod_zone_freepage_state(zone, -(1 << order),
3617 get_pcppage_migratetype(page));
3618 spin_unlock_irqrestore(&zone->lock, flags);
3620 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3621 zone_statistics(preferred_zone, zone, 1);
3624 /* Separate test+clear to avoid unnecessary atomics */
3625 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3626 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3627 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3630 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3634 spin_unlock_irqrestore(&zone->lock, flags);
3638 #ifdef CONFIG_FAIL_PAGE_ALLOC
3641 struct fault_attr attr;
3643 bool ignore_gfp_highmem;
3644 bool ignore_gfp_reclaim;
3646 } fail_page_alloc = {
3647 .attr = FAULT_ATTR_INITIALIZER,
3648 .ignore_gfp_reclaim = true,
3649 .ignore_gfp_highmem = true,
3653 static int __init setup_fail_page_alloc(char *str)
3655 return setup_fault_attr(&fail_page_alloc.attr, str);
3657 __setup("fail_page_alloc=", setup_fail_page_alloc);
3659 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3661 if (order < fail_page_alloc.min_order)
3663 if (gfp_mask & __GFP_NOFAIL)
3665 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3667 if (fail_page_alloc.ignore_gfp_reclaim &&
3668 (gfp_mask & __GFP_DIRECT_RECLAIM))
3671 return should_fail(&fail_page_alloc.attr, 1 << order);
3674 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3676 static int __init fail_page_alloc_debugfs(void)
3678 umode_t mode = S_IFREG | 0600;
3681 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3682 &fail_page_alloc.attr);
3684 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3685 &fail_page_alloc.ignore_gfp_reclaim);
3686 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3687 &fail_page_alloc.ignore_gfp_highmem);
3688 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3693 late_initcall(fail_page_alloc_debugfs);
3695 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3697 #else /* CONFIG_FAIL_PAGE_ALLOC */
3699 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3704 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3706 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3708 return __should_fail_alloc_page(gfp_mask, order);
3710 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3712 static inline long __zone_watermark_unusable_free(struct zone *z,
3713 unsigned int order, unsigned int alloc_flags)
3715 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3716 long unusable_free = (1 << order) - 1;
3719 * If the caller does not have rights to ALLOC_HARDER then subtract
3720 * the high-atomic reserves. This will over-estimate the size of the
3721 * atomic reserve but it avoids a search.
3723 if (likely(!alloc_harder))
3724 unusable_free += z->nr_reserved_highatomic;
3727 /* If allocation can't use CMA areas don't use free CMA pages */
3728 if (!(alloc_flags & ALLOC_CMA))
3729 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3732 return unusable_free;
3736 * Return true if free base pages are above 'mark'. For high-order checks it
3737 * will return true of the order-0 watermark is reached and there is at least
3738 * one free page of a suitable size. Checking now avoids taking the zone lock
3739 * to check in the allocation paths if no pages are free.
3741 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3742 int highest_zoneidx, unsigned int alloc_flags,
3747 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3749 /* free_pages may go negative - that's OK */
3750 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3752 if (alloc_flags & ALLOC_HIGH)
3755 if (unlikely(alloc_harder)) {
3757 * OOM victims can try even harder than normal ALLOC_HARDER
3758 * users on the grounds that it's definitely going to be in
3759 * the exit path shortly and free memory. Any allocation it
3760 * makes during the free path will be small and short-lived.
3762 if (alloc_flags & ALLOC_OOM)
3769 * Check watermarks for an order-0 allocation request. If these
3770 * are not met, then a high-order request also cannot go ahead
3771 * even if a suitable page happened to be free.
3773 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3776 /* If this is an order-0 request then the watermark is fine */
3780 /* For a high-order request, check at least one suitable page is free */
3781 for (o = order; o < MAX_ORDER; o++) {
3782 struct free_area *area = &z->free_area[o];
3788 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3789 if (!free_area_empty(area, mt))
3794 if ((alloc_flags & ALLOC_CMA) &&
3795 !free_area_empty(area, MIGRATE_CMA)) {
3799 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3805 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3806 int highest_zoneidx, unsigned int alloc_flags)
3808 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3809 zone_page_state(z, NR_FREE_PAGES));
3812 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3813 unsigned long mark, int highest_zoneidx,
3814 unsigned int alloc_flags, gfp_t gfp_mask)
3818 free_pages = zone_page_state(z, NR_FREE_PAGES);
3821 * Fast check for order-0 only. If this fails then the reserves
3822 * need to be calculated.
3827 fast_free = free_pages;
3828 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3829 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3833 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3837 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3838 * when checking the min watermark. The min watermark is the
3839 * point where boosting is ignored so that kswapd is woken up
3840 * when below the low watermark.
3842 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3843 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3844 mark = z->_watermark[WMARK_MIN];
3845 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3846 alloc_flags, free_pages);
3852 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3853 unsigned long mark, int highest_zoneidx)
3855 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3857 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3858 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3860 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3865 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3867 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3868 node_reclaim_distance;
3870 #else /* CONFIG_NUMA */
3871 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3875 #endif /* CONFIG_NUMA */
3878 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3879 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3880 * premature use of a lower zone may cause lowmem pressure problems that
3881 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3882 * probably too small. It only makes sense to spread allocations to avoid
3883 * fragmentation between the Normal and DMA32 zones.
3885 static inline unsigned int
3886 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3888 unsigned int alloc_flags;
3891 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3894 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3896 #ifdef CONFIG_ZONE_DMA32
3900 if (zone_idx(zone) != ZONE_NORMAL)
3904 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3905 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3906 * on UMA that if Normal is populated then so is DMA32.
3908 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3909 if (nr_online_nodes > 1 && !populated_zone(--zone))
3912 alloc_flags |= ALLOC_NOFRAGMENT;
3913 #endif /* CONFIG_ZONE_DMA32 */
3917 /* Must be called after current_gfp_context() which can change gfp_mask */
3918 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3919 unsigned int alloc_flags)
3922 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3923 alloc_flags |= ALLOC_CMA;
3929 * get_page_from_freelist goes through the zonelist trying to allocate
3932 static struct page *
3933 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3934 const struct alloc_context *ac)
3938 struct pglist_data *last_pgdat_dirty_limit = NULL;
3943 * Scan zonelist, looking for a zone with enough free.
3944 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3946 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3947 z = ac->preferred_zoneref;
3948 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3953 if (cpusets_enabled() &&
3954 (alloc_flags & ALLOC_CPUSET) &&
3955 !__cpuset_zone_allowed(zone, gfp_mask))
3958 * When allocating a page cache page for writing, we
3959 * want to get it from a node that is within its dirty
3960 * limit, such that no single node holds more than its
3961 * proportional share of globally allowed dirty pages.
3962 * The dirty limits take into account the node's
3963 * lowmem reserves and high watermark so that kswapd
3964 * should be able to balance it without having to
3965 * write pages from its LRU list.
3967 * XXX: For now, allow allocations to potentially
3968 * exceed the per-node dirty limit in the slowpath
3969 * (spread_dirty_pages unset) before going into reclaim,
3970 * which is important when on a NUMA setup the allowed
3971 * nodes are together not big enough to reach the
3972 * global limit. The proper fix for these situations
3973 * will require awareness of nodes in the
3974 * dirty-throttling and the flusher threads.
3976 if (ac->spread_dirty_pages) {
3977 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3980 if (!node_dirty_ok(zone->zone_pgdat)) {
3981 last_pgdat_dirty_limit = zone->zone_pgdat;
3986 if (no_fallback && nr_online_nodes > 1 &&
3987 zone != ac->preferred_zoneref->zone) {
3991 * If moving to a remote node, retry but allow
3992 * fragmenting fallbacks. Locality is more important
3993 * than fragmentation avoidance.
3995 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3996 if (zone_to_nid(zone) != local_nid) {
3997 alloc_flags &= ~ALLOC_NOFRAGMENT;
4002 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4003 if (!zone_watermark_fast(zone, order, mark,
4004 ac->highest_zoneidx, alloc_flags,
4008 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4010 * Watermark failed for this zone, but see if we can
4011 * grow this zone if it contains deferred pages.
4013 if (static_branch_unlikely(&deferred_pages)) {
4014 if (_deferred_grow_zone(zone, order))
4018 /* Checked here to keep the fast path fast */
4019 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4020 if (alloc_flags & ALLOC_NO_WATERMARKS)
4023 if (!node_reclaim_enabled() ||
4024 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4027 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4029 case NODE_RECLAIM_NOSCAN:
4032 case NODE_RECLAIM_FULL:
4033 /* scanned but unreclaimable */
4036 /* did we reclaim enough */
4037 if (zone_watermark_ok(zone, order, mark,
4038 ac->highest_zoneidx, alloc_flags))
4046 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4047 gfp_mask, alloc_flags, ac->migratetype);
4049 prep_new_page(page, order, gfp_mask, alloc_flags);
4052 * If this is a high-order atomic allocation then check
4053 * if the pageblock should be reserved for the future
4055 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4056 reserve_highatomic_pageblock(page, zone, order);
4060 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4061 /* Try again if zone has deferred pages */
4062 if (static_branch_unlikely(&deferred_pages)) {
4063 if (_deferred_grow_zone(zone, order))
4071 * It's possible on a UMA machine to get through all zones that are
4072 * fragmented. If avoiding fragmentation, reset and try again.
4075 alloc_flags &= ~ALLOC_NOFRAGMENT;
4082 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4084 unsigned int filter = SHOW_MEM_FILTER_NODES;
4087 * This documents exceptions given to allocations in certain
4088 * contexts that are allowed to allocate outside current's set
4091 if (!(gfp_mask & __GFP_NOMEMALLOC))
4092 if (tsk_is_oom_victim(current) ||
4093 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4094 filter &= ~SHOW_MEM_FILTER_NODES;
4095 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4096 filter &= ~SHOW_MEM_FILTER_NODES;
4098 show_mem(filter, nodemask);
4101 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4103 struct va_format vaf;
4105 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4107 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4110 va_start(args, fmt);
4113 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4114 current->comm, &vaf, gfp_mask, &gfp_mask,
4115 nodemask_pr_args(nodemask));
4118 cpuset_print_current_mems_allowed();
4121 warn_alloc_show_mem(gfp_mask, nodemask);
4124 static inline struct page *
4125 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4126 unsigned int alloc_flags,
4127 const struct alloc_context *ac)
4131 page = get_page_from_freelist(gfp_mask, order,
4132 alloc_flags|ALLOC_CPUSET, ac);
4134 * fallback to ignore cpuset restriction if our nodes
4138 page = get_page_from_freelist(gfp_mask, order,
4144 static inline struct page *
4145 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4146 const struct alloc_context *ac, unsigned long *did_some_progress)
4148 struct oom_control oc = {
4149 .zonelist = ac->zonelist,
4150 .nodemask = ac->nodemask,
4152 .gfp_mask = gfp_mask,
4157 *did_some_progress = 0;
4160 * Acquire the oom lock. If that fails, somebody else is
4161 * making progress for us.
4163 if (!mutex_trylock(&oom_lock)) {
4164 *did_some_progress = 1;
4165 schedule_timeout_uninterruptible(1);
4170 * Go through the zonelist yet one more time, keep very high watermark
4171 * here, this is only to catch a parallel oom killing, we must fail if
4172 * we're still under heavy pressure. But make sure that this reclaim
4173 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4174 * allocation which will never fail due to oom_lock already held.
4176 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4177 ~__GFP_DIRECT_RECLAIM, order,
4178 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4182 /* Coredumps can quickly deplete all memory reserves */
4183 if (current->flags & PF_DUMPCORE)
4185 /* The OOM killer will not help higher order allocs */
4186 if (order > PAGE_ALLOC_COSTLY_ORDER)
4189 * We have already exhausted all our reclaim opportunities without any
4190 * success so it is time to admit defeat. We will skip the OOM killer
4191 * because it is very likely that the caller has a more reasonable
4192 * fallback than shooting a random task.
4194 * The OOM killer may not free memory on a specific node.
4196 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4198 /* The OOM killer does not needlessly kill tasks for lowmem */
4199 if (ac->highest_zoneidx < ZONE_NORMAL)
4201 if (pm_suspended_storage())
4204 * XXX: GFP_NOFS allocations should rather fail than rely on
4205 * other request to make a forward progress.
4206 * We are in an unfortunate situation where out_of_memory cannot
4207 * do much for this context but let's try it to at least get
4208 * access to memory reserved if the current task is killed (see
4209 * out_of_memory). Once filesystems are ready to handle allocation
4210 * failures more gracefully we should just bail out here.
4213 /* Exhausted what can be done so it's blame time */
4214 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4215 *did_some_progress = 1;
4218 * Help non-failing allocations by giving them access to memory
4221 if (gfp_mask & __GFP_NOFAIL)
4222 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4223 ALLOC_NO_WATERMARKS, ac);
4226 mutex_unlock(&oom_lock);
4231 * Maximum number of compaction retries with a progress before OOM
4232 * killer is consider as the only way to move forward.
4234 #define MAX_COMPACT_RETRIES 16
4236 #ifdef CONFIG_COMPACTION
4237 /* Try memory compaction for high-order allocations before reclaim */
4238 static struct page *
4239 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4240 unsigned int alloc_flags, const struct alloc_context *ac,
4241 enum compact_priority prio, enum compact_result *compact_result)
4243 struct page *page = NULL;
4244 unsigned long pflags;
4245 unsigned int noreclaim_flag;
4250 psi_memstall_enter(&pflags);
4251 noreclaim_flag = memalloc_noreclaim_save();
4253 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4256 memalloc_noreclaim_restore(noreclaim_flag);
4257 psi_memstall_leave(&pflags);
4259 if (*compact_result == COMPACT_SKIPPED)
4262 * At least in one zone compaction wasn't deferred or skipped, so let's
4263 * count a compaction stall
4265 count_vm_event(COMPACTSTALL);
4267 /* Prep a captured page if available */
4269 prep_new_page(page, order, gfp_mask, alloc_flags);
4271 /* Try get a page from the freelist if available */
4273 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4276 struct zone *zone = page_zone(page);
4278 zone->compact_blockskip_flush = false;
4279 compaction_defer_reset(zone, order, true);
4280 count_vm_event(COMPACTSUCCESS);
4285 * It's bad if compaction run occurs and fails. The most likely reason
4286 * is that pages exist, but not enough to satisfy watermarks.
4288 count_vm_event(COMPACTFAIL);
4296 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4297 enum compact_result compact_result,
4298 enum compact_priority *compact_priority,
4299 int *compaction_retries)
4301 int max_retries = MAX_COMPACT_RETRIES;
4304 int retries = *compaction_retries;
4305 enum compact_priority priority = *compact_priority;
4310 if (fatal_signal_pending(current))
4313 if (compaction_made_progress(compact_result))
4314 (*compaction_retries)++;
4317 * compaction considers all the zone as desperately out of memory
4318 * so it doesn't really make much sense to retry except when the
4319 * failure could be caused by insufficient priority
4321 if (compaction_failed(compact_result))
4322 goto check_priority;
4325 * compaction was skipped because there are not enough order-0 pages
4326 * to work with, so we retry only if it looks like reclaim can help.
4328 if (compaction_needs_reclaim(compact_result)) {
4329 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4334 * make sure the compaction wasn't deferred or didn't bail out early
4335 * due to locks contention before we declare that we should give up.
4336 * But the next retry should use a higher priority if allowed, so
4337 * we don't just keep bailing out endlessly.
4339 if (compaction_withdrawn(compact_result)) {
4340 goto check_priority;
4344 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4345 * costly ones because they are de facto nofail and invoke OOM
4346 * killer to move on while costly can fail and users are ready
4347 * to cope with that. 1/4 retries is rather arbitrary but we
4348 * would need much more detailed feedback from compaction to
4349 * make a better decision.
4351 if (order > PAGE_ALLOC_COSTLY_ORDER)
4353 if (*compaction_retries <= max_retries) {
4359 * Make sure there are attempts at the highest priority if we exhausted
4360 * all retries or failed at the lower priorities.
4363 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4364 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4366 if (*compact_priority > min_priority) {
4367 (*compact_priority)--;
4368 *compaction_retries = 0;
4372 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4376 static inline struct page *
4377 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4378 unsigned int alloc_flags, const struct alloc_context *ac,
4379 enum compact_priority prio, enum compact_result *compact_result)
4381 *compact_result = COMPACT_SKIPPED;
4386 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4387 enum compact_result compact_result,
4388 enum compact_priority *compact_priority,
4389 int *compaction_retries)
4394 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4398 * There are setups with compaction disabled which would prefer to loop
4399 * inside the allocator rather than hit the oom killer prematurely.
4400 * Let's give them a good hope and keep retrying while the order-0
4401 * watermarks are OK.
4403 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4404 ac->highest_zoneidx, ac->nodemask) {
4405 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4406 ac->highest_zoneidx, alloc_flags))
4411 #endif /* CONFIG_COMPACTION */
4413 #ifdef CONFIG_LOCKDEP
4414 static struct lockdep_map __fs_reclaim_map =
4415 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4417 static bool __need_reclaim(gfp_t gfp_mask)
4419 /* no reclaim without waiting on it */
4420 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4423 /* this guy won't enter reclaim */
4424 if (current->flags & PF_MEMALLOC)
4427 if (gfp_mask & __GFP_NOLOCKDEP)
4433 void __fs_reclaim_acquire(void)
4435 lock_map_acquire(&__fs_reclaim_map);
4438 void __fs_reclaim_release(void)
4440 lock_map_release(&__fs_reclaim_map);
4443 void fs_reclaim_acquire(gfp_t gfp_mask)
4445 gfp_mask = current_gfp_context(gfp_mask);
4447 if (__need_reclaim(gfp_mask)) {
4448 if (gfp_mask & __GFP_FS)
4449 __fs_reclaim_acquire();
4451 #ifdef CONFIG_MMU_NOTIFIER
4452 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4453 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4458 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4460 void fs_reclaim_release(gfp_t gfp_mask)
4462 gfp_mask = current_gfp_context(gfp_mask);
4464 if (__need_reclaim(gfp_mask)) {
4465 if (gfp_mask & __GFP_FS)
4466 __fs_reclaim_release();
4469 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4472 /* Perform direct synchronous page reclaim */
4473 static unsigned long
4474 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4475 const struct alloc_context *ac)
4477 unsigned int noreclaim_flag;
4478 unsigned long pflags, progress;
4482 /* We now go into synchronous reclaim */
4483 cpuset_memory_pressure_bump();
4484 psi_memstall_enter(&pflags);
4485 fs_reclaim_acquire(gfp_mask);
4486 noreclaim_flag = memalloc_noreclaim_save();
4488 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4491 memalloc_noreclaim_restore(noreclaim_flag);
4492 fs_reclaim_release(gfp_mask);
4493 psi_memstall_leave(&pflags);
4500 /* The really slow allocator path where we enter direct reclaim */
4501 static inline struct page *
4502 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4503 unsigned int alloc_flags, const struct alloc_context *ac,
4504 unsigned long *did_some_progress)
4506 struct page *page = NULL;
4507 bool drained = false;
4509 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4510 if (unlikely(!(*did_some_progress)))
4514 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4517 * If an allocation failed after direct reclaim, it could be because
4518 * pages are pinned on the per-cpu lists or in high alloc reserves.
4519 * Shrink them and try again
4521 if (!page && !drained) {
4522 unreserve_highatomic_pageblock(ac, false);
4523 drain_all_pages(NULL);
4531 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4532 const struct alloc_context *ac)
4536 pg_data_t *last_pgdat = NULL;
4537 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4539 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4541 if (last_pgdat != zone->zone_pgdat)
4542 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4543 last_pgdat = zone->zone_pgdat;
4547 static inline unsigned int
4548 gfp_to_alloc_flags(gfp_t gfp_mask)
4550 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4553 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4554 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4555 * to save two branches.
4557 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4558 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4561 * The caller may dip into page reserves a bit more if the caller
4562 * cannot run direct reclaim, or if the caller has realtime scheduling
4563 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4564 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4566 alloc_flags |= (__force int)
4567 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4569 if (gfp_mask & __GFP_ATOMIC) {
4571 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4572 * if it can't schedule.
4574 if (!(gfp_mask & __GFP_NOMEMALLOC))
4575 alloc_flags |= ALLOC_HARDER;
4577 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4578 * comment for __cpuset_node_allowed().
4580 alloc_flags &= ~ALLOC_CPUSET;
4581 } else if (unlikely(rt_task(current)) && !in_interrupt())
4582 alloc_flags |= ALLOC_HARDER;
4584 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4589 static bool oom_reserves_allowed(struct task_struct *tsk)
4591 if (!tsk_is_oom_victim(tsk))
4595 * !MMU doesn't have oom reaper so give access to memory reserves
4596 * only to the thread with TIF_MEMDIE set
4598 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4605 * Distinguish requests which really need access to full memory
4606 * reserves from oom victims which can live with a portion of it
4608 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4610 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4612 if (gfp_mask & __GFP_MEMALLOC)
4613 return ALLOC_NO_WATERMARKS;
4614 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4615 return ALLOC_NO_WATERMARKS;
4616 if (!in_interrupt()) {
4617 if (current->flags & PF_MEMALLOC)
4618 return ALLOC_NO_WATERMARKS;
4619 else if (oom_reserves_allowed(current))
4626 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4628 return !!__gfp_pfmemalloc_flags(gfp_mask);
4632 * Checks whether it makes sense to retry the reclaim to make a forward progress
4633 * for the given allocation request.
4635 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4636 * without success, or when we couldn't even meet the watermark if we
4637 * reclaimed all remaining pages on the LRU lists.
4639 * Returns true if a retry is viable or false to enter the oom path.
4642 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4643 struct alloc_context *ac, int alloc_flags,
4644 bool did_some_progress, int *no_progress_loops)
4651 * Costly allocations might have made a progress but this doesn't mean
4652 * their order will become available due to high fragmentation so
4653 * always increment the no progress counter for them
4655 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4656 *no_progress_loops = 0;
4658 (*no_progress_loops)++;
4661 * Make sure we converge to OOM if we cannot make any progress
4662 * several times in the row.
4664 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4665 /* Before OOM, exhaust highatomic_reserve */
4666 return unreserve_highatomic_pageblock(ac, true);
4670 * Keep reclaiming pages while there is a chance this will lead
4671 * somewhere. If none of the target zones can satisfy our allocation
4672 * request even if all reclaimable pages are considered then we are
4673 * screwed and have to go OOM.
4675 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4676 ac->highest_zoneidx, ac->nodemask) {
4677 unsigned long available;
4678 unsigned long reclaimable;
4679 unsigned long min_wmark = min_wmark_pages(zone);
4682 available = reclaimable = zone_reclaimable_pages(zone);
4683 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4686 * Would the allocation succeed if we reclaimed all
4687 * reclaimable pages?
4689 wmark = __zone_watermark_ok(zone, order, min_wmark,
4690 ac->highest_zoneidx, alloc_flags, available);
4691 trace_reclaim_retry_zone(z, order, reclaimable,
4692 available, min_wmark, *no_progress_loops, wmark);
4695 * If we didn't make any progress and have a lot of
4696 * dirty + writeback pages then we should wait for
4697 * an IO to complete to slow down the reclaim and
4698 * prevent from pre mature OOM
4700 if (!did_some_progress) {
4701 unsigned long write_pending;
4703 write_pending = zone_page_state_snapshot(zone,
4704 NR_ZONE_WRITE_PENDING);
4706 if (2 * write_pending > reclaimable) {
4707 congestion_wait(BLK_RW_ASYNC, HZ/10);
4719 * Memory allocation/reclaim might be called from a WQ context and the
4720 * current implementation of the WQ concurrency control doesn't
4721 * recognize that a particular WQ is congested if the worker thread is
4722 * looping without ever sleeping. Therefore we have to do a short sleep
4723 * here rather than calling cond_resched().
4725 if (current->flags & PF_WQ_WORKER)
4726 schedule_timeout_uninterruptible(1);
4733 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4736 * It's possible that cpuset's mems_allowed and the nodemask from
4737 * mempolicy don't intersect. This should be normally dealt with by
4738 * policy_nodemask(), but it's possible to race with cpuset update in
4739 * such a way the check therein was true, and then it became false
4740 * before we got our cpuset_mems_cookie here.
4741 * This assumes that for all allocations, ac->nodemask can come only
4742 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4743 * when it does not intersect with the cpuset restrictions) or the
4744 * caller can deal with a violated nodemask.
4746 if (cpusets_enabled() && ac->nodemask &&
4747 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4748 ac->nodemask = NULL;
4753 * When updating a task's mems_allowed or mempolicy nodemask, it is
4754 * possible to race with parallel threads in such a way that our
4755 * allocation can fail while the mask is being updated. If we are about
4756 * to fail, check if the cpuset changed during allocation and if so,
4759 if (read_mems_allowed_retry(cpuset_mems_cookie))
4765 static inline struct page *
4766 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4767 struct alloc_context *ac)
4769 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4770 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4771 struct page *page = NULL;
4772 unsigned int alloc_flags;
4773 unsigned long did_some_progress;
4774 enum compact_priority compact_priority;
4775 enum compact_result compact_result;
4776 int compaction_retries;
4777 int no_progress_loops;
4778 unsigned int cpuset_mems_cookie;
4782 * We also sanity check to catch abuse of atomic reserves being used by
4783 * callers that are not in atomic context.
4785 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4786 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4787 gfp_mask &= ~__GFP_ATOMIC;
4790 compaction_retries = 0;
4791 no_progress_loops = 0;
4792 compact_priority = DEF_COMPACT_PRIORITY;
4793 cpuset_mems_cookie = read_mems_allowed_begin();
4796 * The fast path uses conservative alloc_flags to succeed only until
4797 * kswapd needs to be woken up, and to avoid the cost of setting up
4798 * alloc_flags precisely. So we do that now.
4800 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4803 * We need to recalculate the starting point for the zonelist iterator
4804 * because we might have used different nodemask in the fast path, or
4805 * there was a cpuset modification and we are retrying - otherwise we
4806 * could end up iterating over non-eligible zones endlessly.
4808 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4809 ac->highest_zoneidx, ac->nodemask);
4810 if (!ac->preferred_zoneref->zone)
4813 if (alloc_flags & ALLOC_KSWAPD)
4814 wake_all_kswapds(order, gfp_mask, ac);
4817 * The adjusted alloc_flags might result in immediate success, so try
4820 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4825 * For costly allocations, try direct compaction first, as it's likely
4826 * that we have enough base pages and don't need to reclaim. For non-
4827 * movable high-order allocations, do that as well, as compaction will
4828 * try prevent permanent fragmentation by migrating from blocks of the
4830 * Don't try this for allocations that are allowed to ignore
4831 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4833 if (can_direct_reclaim &&
4835 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4836 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4837 page = __alloc_pages_direct_compact(gfp_mask, order,
4839 INIT_COMPACT_PRIORITY,
4845 * Checks for costly allocations with __GFP_NORETRY, which
4846 * includes some THP page fault allocations
4848 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4850 * If allocating entire pageblock(s) and compaction
4851 * failed because all zones are below low watermarks
4852 * or is prohibited because it recently failed at this
4853 * order, fail immediately unless the allocator has
4854 * requested compaction and reclaim retry.
4857 * - potentially very expensive because zones are far
4858 * below their low watermarks or this is part of very
4859 * bursty high order allocations,
4860 * - not guaranteed to help because isolate_freepages()
4861 * may not iterate over freed pages as part of its
4863 * - unlikely to make entire pageblocks free on its
4866 if (compact_result == COMPACT_SKIPPED ||
4867 compact_result == COMPACT_DEFERRED)
4871 * Looks like reclaim/compaction is worth trying, but
4872 * sync compaction could be very expensive, so keep
4873 * using async compaction.
4875 compact_priority = INIT_COMPACT_PRIORITY;
4880 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4881 if (alloc_flags & ALLOC_KSWAPD)
4882 wake_all_kswapds(order, gfp_mask, ac);
4884 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4886 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
4889 * Reset the nodemask and zonelist iterators if memory policies can be
4890 * ignored. These allocations are high priority and system rather than
4893 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4894 ac->nodemask = NULL;
4895 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4896 ac->highest_zoneidx, ac->nodemask);
4899 /* Attempt with potentially adjusted zonelist and alloc_flags */
4900 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4904 /* Caller is not willing to reclaim, we can't balance anything */
4905 if (!can_direct_reclaim)
4908 /* Avoid recursion of direct reclaim */
4909 if (current->flags & PF_MEMALLOC)
4912 /* Try direct reclaim and then allocating */
4913 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4914 &did_some_progress);
4918 /* Try direct compaction and then allocating */
4919 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4920 compact_priority, &compact_result);
4924 /* Do not loop if specifically requested */
4925 if (gfp_mask & __GFP_NORETRY)
4929 * Do not retry costly high order allocations unless they are
4930 * __GFP_RETRY_MAYFAIL
4932 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4935 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4936 did_some_progress > 0, &no_progress_loops))
4940 * It doesn't make any sense to retry for the compaction if the order-0
4941 * reclaim is not able to make any progress because the current
4942 * implementation of the compaction depends on the sufficient amount
4943 * of free memory (see __compaction_suitable)
4945 if (did_some_progress > 0 &&
4946 should_compact_retry(ac, order, alloc_flags,
4947 compact_result, &compact_priority,
4948 &compaction_retries))
4952 /* Deal with possible cpuset update races before we start OOM killing */
4953 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4956 /* Reclaim has failed us, start killing things */
4957 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4961 /* Avoid allocations with no watermarks from looping endlessly */
4962 if (tsk_is_oom_victim(current) &&
4963 (alloc_flags & ALLOC_OOM ||
4964 (gfp_mask & __GFP_NOMEMALLOC)))
4967 /* Retry as long as the OOM killer is making progress */
4968 if (did_some_progress) {
4969 no_progress_loops = 0;
4974 /* Deal with possible cpuset update races before we fail */
4975 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4979 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4982 if (gfp_mask & __GFP_NOFAIL) {
4984 * All existing users of the __GFP_NOFAIL are blockable, so warn
4985 * of any new users that actually require GFP_NOWAIT
4987 if (WARN_ON_ONCE(!can_direct_reclaim))
4991 * PF_MEMALLOC request from this context is rather bizarre
4992 * because we cannot reclaim anything and only can loop waiting
4993 * for somebody to do a work for us
4995 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4998 * non failing costly orders are a hard requirement which we
4999 * are not prepared for much so let's warn about these users
5000 * so that we can identify them and convert them to something
5003 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5006 * Help non-failing allocations by giving them access to memory
5007 * reserves but do not use ALLOC_NO_WATERMARKS because this
5008 * could deplete whole memory reserves which would just make
5009 * the situation worse
5011 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5019 warn_alloc(gfp_mask, ac->nodemask,
5020 "page allocation failure: order:%u", order);
5025 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5026 int preferred_nid, nodemask_t *nodemask,
5027 struct alloc_context *ac, gfp_t *alloc_gfp,
5028 unsigned int *alloc_flags)
5030 ac->highest_zoneidx = gfp_zone(gfp_mask);
5031 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5032 ac->nodemask = nodemask;
5033 ac->migratetype = gfp_migratetype(gfp_mask);
5035 if (cpusets_enabled()) {
5036 *alloc_gfp |= __GFP_HARDWALL;
5038 * When we are in the interrupt context, it is irrelevant
5039 * to the current task context. It means that any node ok.
5041 if (!in_interrupt() && !ac->nodemask)
5042 ac->nodemask = &cpuset_current_mems_allowed;
5044 *alloc_flags |= ALLOC_CPUSET;
5047 fs_reclaim_acquire(gfp_mask);
5048 fs_reclaim_release(gfp_mask);
5050 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5052 if (should_fail_alloc_page(gfp_mask, order))
5055 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5057 /* Dirty zone balancing only done in the fast path */
5058 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5061 * The preferred zone is used for statistics but crucially it is
5062 * also used as the starting point for the zonelist iterator. It
5063 * may get reset for allocations that ignore memory policies.
5065 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5066 ac->highest_zoneidx, ac->nodemask);
5072 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5073 * @gfp: GFP flags for the allocation
5074 * @preferred_nid: The preferred NUMA node ID to allocate from
5075 * @nodemask: Set of nodes to allocate from, may be NULL
5076 * @nr_pages: The number of pages desired on the list or array
5077 * @page_list: Optional list to store the allocated pages
5078 * @page_array: Optional array to store the pages
5080 * This is a batched version of the page allocator that attempts to
5081 * allocate nr_pages quickly. Pages are added to page_list if page_list
5082 * is not NULL, otherwise it is assumed that the page_array is valid.
5084 * For lists, nr_pages is the number of pages that should be allocated.
5086 * For arrays, only NULL elements are populated with pages and nr_pages
5087 * is the maximum number of pages that will be stored in the array.
5089 * Returns the number of pages on the list or array.
5091 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5092 nodemask_t *nodemask, int nr_pages,
5093 struct list_head *page_list,
5094 struct page **page_array)
5097 unsigned long flags;
5100 struct per_cpu_pages *pcp;
5101 struct list_head *pcp_list;
5102 struct alloc_context ac;
5104 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5105 int nr_populated = 0, nr_account = 0;
5107 if (unlikely(nr_pages <= 0))
5111 * Skip populated array elements to determine if any pages need
5112 * to be allocated before disabling IRQs.
5114 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5117 /* Already populated array? */
5118 if (unlikely(page_array && nr_pages - nr_populated == 0))
5119 return nr_populated;
5121 /* Use the single page allocator for one page. */
5122 if (nr_pages - nr_populated == 1)
5125 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5126 gfp &= gfp_allowed_mask;
5128 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5132 /* Find an allowed local zone that meets the low watermark. */
5133 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5136 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5137 !__cpuset_zone_allowed(zone, gfp)) {
5141 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5142 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5146 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5147 if (zone_watermark_fast(zone, 0, mark,
5148 zonelist_zone_idx(ac.preferred_zoneref),
5149 alloc_flags, gfp)) {
5155 * If there are no allowed local zones that meets the watermarks then
5156 * try to allocate a single page and reclaim if necessary.
5158 if (unlikely(!zone))
5161 /* Attempt the batch allocation */
5162 local_lock_irqsave(&pagesets.lock, flags);
5163 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5164 pcp_list = &pcp->lists[ac.migratetype];
5166 while (nr_populated < nr_pages) {
5168 /* Skip existing pages */
5169 if (page_array && page_array[nr_populated]) {
5174 page = __rmqueue_pcplist(zone, ac.migratetype, alloc_flags,
5176 if (unlikely(!page)) {
5177 /* Try and get at least one page */
5184 prep_new_page(page, 0, gfp, 0);
5186 list_add(&page->lru, page_list);
5188 page_array[nr_populated] = page;
5192 local_unlock_irqrestore(&pagesets.lock, flags);
5194 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5195 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5197 return nr_populated;
5200 local_unlock_irqrestore(&pagesets.lock, flags);
5203 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5206 list_add(&page->lru, page_list);
5208 page_array[nr_populated] = page;
5212 return nr_populated;
5214 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5217 * This is the 'heart' of the zoned buddy allocator.
5219 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5220 nodemask_t *nodemask)
5223 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5224 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5225 struct alloc_context ac = { };
5228 * There are several places where we assume that the order value is sane
5229 * so bail out early if the request is out of bound.
5231 if (unlikely(order >= MAX_ORDER)) {
5232 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5236 gfp &= gfp_allowed_mask;
5238 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5239 * resp. GFP_NOIO which has to be inherited for all allocation requests
5240 * from a particular context which has been marked by
5241 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5242 * movable zones are not used during allocation.
5244 gfp = current_gfp_context(gfp);
5246 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5247 &alloc_gfp, &alloc_flags))
5251 * Forbid the first pass from falling back to types that fragment
5252 * memory until all local zones are considered.
5254 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5256 /* First allocation attempt */
5257 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5262 ac.spread_dirty_pages = false;
5265 * Restore the original nodemask if it was potentially replaced with
5266 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5268 ac.nodemask = nodemask;
5270 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5273 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5274 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5275 __free_pages(page, order);
5279 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5283 EXPORT_SYMBOL(__alloc_pages);
5286 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5287 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5288 * you need to access high mem.
5290 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5294 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5297 return (unsigned long) page_address(page);
5299 EXPORT_SYMBOL(__get_free_pages);
5301 unsigned long get_zeroed_page(gfp_t gfp_mask)
5303 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5305 EXPORT_SYMBOL(get_zeroed_page);
5307 static inline void free_the_page(struct page *page, unsigned int order)
5309 if (order == 0) /* Via pcp? */
5310 free_unref_page(page);
5312 __free_pages_ok(page, order, FPI_NONE);
5316 * __free_pages - Free pages allocated with alloc_pages().
5317 * @page: The page pointer returned from alloc_pages().
5318 * @order: The order of the allocation.
5320 * This function can free multi-page allocations that are not compound
5321 * pages. It does not check that the @order passed in matches that of
5322 * the allocation, so it is easy to leak memory. Freeing more memory
5323 * than was allocated will probably emit a warning.
5325 * If the last reference to this page is speculative, it will be released
5326 * by put_page() which only frees the first page of a non-compound
5327 * allocation. To prevent the remaining pages from being leaked, we free
5328 * the subsequent pages here. If you want to use the page's reference
5329 * count to decide when to free the allocation, you should allocate a
5330 * compound page, and use put_page() instead of __free_pages().
5332 * Context: May be called in interrupt context or while holding a normal
5333 * spinlock, but not in NMI context or while holding a raw spinlock.
5335 void __free_pages(struct page *page, unsigned int order)
5337 if (put_page_testzero(page))
5338 free_the_page(page, order);
5339 else if (!PageHead(page))
5341 free_the_page(page + (1 << order), order);
5343 EXPORT_SYMBOL(__free_pages);
5345 void free_pages(unsigned long addr, unsigned int order)
5348 VM_BUG_ON(!virt_addr_valid((void *)addr));
5349 __free_pages(virt_to_page((void *)addr), order);
5353 EXPORT_SYMBOL(free_pages);
5357 * An arbitrary-length arbitrary-offset area of memory which resides
5358 * within a 0 or higher order page. Multiple fragments within that page
5359 * are individually refcounted, in the page's reference counter.
5361 * The page_frag functions below provide a simple allocation framework for
5362 * page fragments. This is used by the network stack and network device
5363 * drivers to provide a backing region of memory for use as either an
5364 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5366 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5369 struct page *page = NULL;
5370 gfp_t gfp = gfp_mask;
5372 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5373 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5375 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5376 PAGE_FRAG_CACHE_MAX_ORDER);
5377 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5379 if (unlikely(!page))
5380 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5382 nc->va = page ? page_address(page) : NULL;
5387 void __page_frag_cache_drain(struct page *page, unsigned int count)
5389 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5391 if (page_ref_sub_and_test(page, count))
5392 free_the_page(page, compound_order(page));
5394 EXPORT_SYMBOL(__page_frag_cache_drain);
5396 void *page_frag_alloc_align(struct page_frag_cache *nc,
5397 unsigned int fragsz, gfp_t gfp_mask,
5398 unsigned int align_mask)
5400 unsigned int size = PAGE_SIZE;
5404 if (unlikely(!nc->va)) {
5406 page = __page_frag_cache_refill(nc, gfp_mask);
5410 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5411 /* if size can vary use size else just use PAGE_SIZE */
5414 /* Even if we own the page, we do not use atomic_set().
5415 * This would break get_page_unless_zero() users.
5417 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5419 /* reset page count bias and offset to start of new frag */
5420 nc->pfmemalloc = page_is_pfmemalloc(page);
5421 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5425 offset = nc->offset - fragsz;
5426 if (unlikely(offset < 0)) {
5427 page = virt_to_page(nc->va);
5429 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5432 if (unlikely(nc->pfmemalloc)) {
5433 free_the_page(page, compound_order(page));
5437 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5438 /* if size can vary use size else just use PAGE_SIZE */
5441 /* OK, page count is 0, we can safely set it */
5442 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5444 /* reset page count bias and offset to start of new frag */
5445 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5446 offset = size - fragsz;
5450 offset &= align_mask;
5451 nc->offset = offset;
5453 return nc->va + offset;
5455 EXPORT_SYMBOL(page_frag_alloc_align);
5458 * Frees a page fragment allocated out of either a compound or order 0 page.
5460 void page_frag_free(void *addr)
5462 struct page *page = virt_to_head_page(addr);
5464 if (unlikely(put_page_testzero(page)))
5465 free_the_page(page, compound_order(page));
5467 EXPORT_SYMBOL(page_frag_free);
5469 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5473 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5474 unsigned long used = addr + PAGE_ALIGN(size);
5476 split_page(virt_to_page((void *)addr), order);
5477 while (used < alloc_end) {
5482 return (void *)addr;
5486 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5487 * @size: the number of bytes to allocate
5488 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5490 * This function is similar to alloc_pages(), except that it allocates the
5491 * minimum number of pages to satisfy the request. alloc_pages() can only
5492 * allocate memory in power-of-two pages.
5494 * This function is also limited by MAX_ORDER.
5496 * Memory allocated by this function must be released by free_pages_exact().
5498 * Return: pointer to the allocated area or %NULL in case of error.
5500 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5502 unsigned int order = get_order(size);
5505 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5506 gfp_mask &= ~__GFP_COMP;
5508 addr = __get_free_pages(gfp_mask, order);
5509 return make_alloc_exact(addr, order, size);
5511 EXPORT_SYMBOL(alloc_pages_exact);
5514 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5516 * @nid: the preferred node ID where memory should be allocated
5517 * @size: the number of bytes to allocate
5518 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5520 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5523 * Return: pointer to the allocated area or %NULL in case of error.
5525 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5527 unsigned int order = get_order(size);
5530 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5531 gfp_mask &= ~__GFP_COMP;
5533 p = alloc_pages_node(nid, gfp_mask, order);
5536 return make_alloc_exact((unsigned long)page_address(p), order, size);
5540 * free_pages_exact - release memory allocated via alloc_pages_exact()
5541 * @virt: the value returned by alloc_pages_exact.
5542 * @size: size of allocation, same value as passed to alloc_pages_exact().
5544 * Release the memory allocated by a previous call to alloc_pages_exact.
5546 void free_pages_exact(void *virt, size_t size)
5548 unsigned long addr = (unsigned long)virt;
5549 unsigned long end = addr + PAGE_ALIGN(size);
5551 while (addr < end) {
5556 EXPORT_SYMBOL(free_pages_exact);
5559 * nr_free_zone_pages - count number of pages beyond high watermark
5560 * @offset: The zone index of the highest zone
5562 * nr_free_zone_pages() counts the number of pages which are beyond the
5563 * high watermark within all zones at or below a given zone index. For each
5564 * zone, the number of pages is calculated as:
5566 * nr_free_zone_pages = managed_pages - high_pages
5568 * Return: number of pages beyond high watermark.
5570 static unsigned long nr_free_zone_pages(int offset)
5575 /* Just pick one node, since fallback list is circular */
5576 unsigned long sum = 0;
5578 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5580 for_each_zone_zonelist(zone, z, zonelist, offset) {
5581 unsigned long size = zone_managed_pages(zone);
5582 unsigned long high = high_wmark_pages(zone);
5591 * nr_free_buffer_pages - count number of pages beyond high watermark
5593 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5594 * watermark within ZONE_DMA and ZONE_NORMAL.
5596 * Return: number of pages beyond high watermark within ZONE_DMA and
5599 unsigned long nr_free_buffer_pages(void)
5601 return nr_free_zone_pages(gfp_zone(GFP_USER));
5603 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5605 static inline void show_node(struct zone *zone)
5607 if (IS_ENABLED(CONFIG_NUMA))
5608 printk("Node %d ", zone_to_nid(zone));
5611 long si_mem_available(void)
5614 unsigned long pagecache;
5615 unsigned long wmark_low = 0;
5616 unsigned long pages[NR_LRU_LISTS];
5617 unsigned long reclaimable;
5621 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5622 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5625 wmark_low += low_wmark_pages(zone);
5628 * Estimate the amount of memory available for userspace allocations,
5629 * without causing swapping.
5631 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5634 * Not all the page cache can be freed, otherwise the system will
5635 * start swapping. Assume at least half of the page cache, or the
5636 * low watermark worth of cache, needs to stay.
5638 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5639 pagecache -= min(pagecache / 2, wmark_low);
5640 available += pagecache;
5643 * Part of the reclaimable slab and other kernel memory consists of
5644 * items that are in use, and cannot be freed. Cap this estimate at the
5647 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5648 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5649 available += reclaimable - min(reclaimable / 2, wmark_low);
5655 EXPORT_SYMBOL_GPL(si_mem_available);
5657 void si_meminfo(struct sysinfo *val)
5659 val->totalram = totalram_pages();
5660 val->sharedram = global_node_page_state(NR_SHMEM);
5661 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5662 val->bufferram = nr_blockdev_pages();
5663 val->totalhigh = totalhigh_pages();
5664 val->freehigh = nr_free_highpages();
5665 val->mem_unit = PAGE_SIZE;
5668 EXPORT_SYMBOL(si_meminfo);
5671 void si_meminfo_node(struct sysinfo *val, int nid)
5673 int zone_type; /* needs to be signed */
5674 unsigned long managed_pages = 0;
5675 unsigned long managed_highpages = 0;
5676 unsigned long free_highpages = 0;
5677 pg_data_t *pgdat = NODE_DATA(nid);
5679 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5680 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5681 val->totalram = managed_pages;
5682 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5683 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5684 #ifdef CONFIG_HIGHMEM
5685 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5686 struct zone *zone = &pgdat->node_zones[zone_type];
5688 if (is_highmem(zone)) {
5689 managed_highpages += zone_managed_pages(zone);
5690 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5693 val->totalhigh = managed_highpages;
5694 val->freehigh = free_highpages;
5696 val->totalhigh = managed_highpages;
5697 val->freehigh = free_highpages;
5699 val->mem_unit = PAGE_SIZE;
5704 * Determine whether the node should be displayed or not, depending on whether
5705 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5707 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5709 if (!(flags & SHOW_MEM_FILTER_NODES))
5713 * no node mask - aka implicit memory numa policy. Do not bother with
5714 * the synchronization - read_mems_allowed_begin - because we do not
5715 * have to be precise here.
5718 nodemask = &cpuset_current_mems_allowed;
5720 return !node_isset(nid, *nodemask);
5723 #define K(x) ((x) << (PAGE_SHIFT-10))
5725 static void show_migration_types(unsigned char type)
5727 static const char types[MIGRATE_TYPES] = {
5728 [MIGRATE_UNMOVABLE] = 'U',
5729 [MIGRATE_MOVABLE] = 'M',
5730 [MIGRATE_RECLAIMABLE] = 'E',
5731 [MIGRATE_HIGHATOMIC] = 'H',
5733 [MIGRATE_CMA] = 'C',
5735 #ifdef CONFIG_MEMORY_ISOLATION
5736 [MIGRATE_ISOLATE] = 'I',
5739 char tmp[MIGRATE_TYPES + 1];
5743 for (i = 0; i < MIGRATE_TYPES; i++) {
5744 if (type & (1 << i))
5749 printk(KERN_CONT "(%s) ", tmp);
5753 * Show free area list (used inside shift_scroll-lock stuff)
5754 * We also calculate the percentage fragmentation. We do this by counting the
5755 * memory on each free list with the exception of the first item on the list.
5758 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5761 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5763 unsigned long free_pcp = 0;
5768 for_each_populated_zone(zone) {
5769 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5772 for_each_online_cpu(cpu)
5773 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5776 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5777 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5778 " unevictable:%lu dirty:%lu writeback:%lu\n"
5779 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5780 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5781 " free:%lu free_pcp:%lu free_cma:%lu\n",
5782 global_node_page_state(NR_ACTIVE_ANON),
5783 global_node_page_state(NR_INACTIVE_ANON),
5784 global_node_page_state(NR_ISOLATED_ANON),
5785 global_node_page_state(NR_ACTIVE_FILE),
5786 global_node_page_state(NR_INACTIVE_FILE),
5787 global_node_page_state(NR_ISOLATED_FILE),
5788 global_node_page_state(NR_UNEVICTABLE),
5789 global_node_page_state(NR_FILE_DIRTY),
5790 global_node_page_state(NR_WRITEBACK),
5791 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5792 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5793 global_node_page_state(NR_FILE_MAPPED),
5794 global_node_page_state(NR_SHMEM),
5795 global_node_page_state(NR_PAGETABLE),
5796 global_zone_page_state(NR_BOUNCE),
5797 global_zone_page_state(NR_FREE_PAGES),
5799 global_zone_page_state(NR_FREE_CMA_PAGES));
5801 for_each_online_pgdat(pgdat) {
5802 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5806 " active_anon:%lukB"
5807 " inactive_anon:%lukB"
5808 " active_file:%lukB"
5809 " inactive_file:%lukB"
5810 " unevictable:%lukB"
5811 " isolated(anon):%lukB"
5812 " isolated(file):%lukB"
5817 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5819 " shmem_pmdmapped: %lukB"
5822 " writeback_tmp:%lukB"
5823 " kernel_stack:%lukB"
5824 #ifdef CONFIG_SHADOW_CALL_STACK
5825 " shadow_call_stack:%lukB"
5828 " all_unreclaimable? %s"
5831 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5832 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5833 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5834 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5835 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5836 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5837 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5838 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5839 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5840 K(node_page_state(pgdat, NR_WRITEBACK)),
5841 K(node_page_state(pgdat, NR_SHMEM)),
5842 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5843 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5844 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5845 K(node_page_state(pgdat, NR_ANON_THPS)),
5847 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5848 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5849 #ifdef CONFIG_SHADOW_CALL_STACK
5850 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5852 K(node_page_state(pgdat, NR_PAGETABLE)),
5853 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5857 for_each_populated_zone(zone) {
5860 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5864 for_each_online_cpu(cpu)
5865 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5874 " reserved_highatomic:%luKB"
5875 " active_anon:%lukB"
5876 " inactive_anon:%lukB"
5877 " active_file:%lukB"
5878 " inactive_file:%lukB"
5879 " unevictable:%lukB"
5880 " writepending:%lukB"
5890 K(zone_page_state(zone, NR_FREE_PAGES)),
5891 K(min_wmark_pages(zone)),
5892 K(low_wmark_pages(zone)),
5893 K(high_wmark_pages(zone)),
5894 K(zone->nr_reserved_highatomic),
5895 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5896 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5897 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5898 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5899 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5900 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5901 K(zone->present_pages),
5902 K(zone_managed_pages(zone)),
5903 K(zone_page_state(zone, NR_MLOCK)),
5904 K(zone_page_state(zone, NR_BOUNCE)),
5906 K(this_cpu_read(zone->per_cpu_pageset->count)),
5907 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5908 printk("lowmem_reserve[]:");
5909 for (i = 0; i < MAX_NR_ZONES; i++)
5910 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5911 printk(KERN_CONT "\n");
5914 for_each_populated_zone(zone) {
5916 unsigned long nr[MAX_ORDER], flags, total = 0;
5917 unsigned char types[MAX_ORDER];
5919 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5922 printk(KERN_CONT "%s: ", zone->name);
5924 spin_lock_irqsave(&zone->lock, flags);
5925 for (order = 0; order < MAX_ORDER; order++) {
5926 struct free_area *area = &zone->free_area[order];
5929 nr[order] = area->nr_free;
5930 total += nr[order] << order;
5933 for (type = 0; type < MIGRATE_TYPES; type++) {
5934 if (!free_area_empty(area, type))
5935 types[order] |= 1 << type;
5938 spin_unlock_irqrestore(&zone->lock, flags);
5939 for (order = 0; order < MAX_ORDER; order++) {
5940 printk(KERN_CONT "%lu*%lukB ",
5941 nr[order], K(1UL) << order);
5943 show_migration_types(types[order]);
5945 printk(KERN_CONT "= %lukB\n", K(total));
5948 hugetlb_show_meminfo();
5950 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5952 show_swap_cache_info();
5955 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5957 zoneref->zone = zone;
5958 zoneref->zone_idx = zone_idx(zone);
5962 * Builds allocation fallback zone lists.
5964 * Add all populated zones of a node to the zonelist.
5966 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5969 enum zone_type zone_type = MAX_NR_ZONES;
5974 zone = pgdat->node_zones + zone_type;
5975 if (managed_zone(zone)) {
5976 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5977 check_highest_zone(zone_type);
5979 } while (zone_type);
5986 static int __parse_numa_zonelist_order(char *s)
5989 * We used to support different zonelists modes but they turned
5990 * out to be just not useful. Let's keep the warning in place
5991 * if somebody still use the cmd line parameter so that we do
5992 * not fail it silently
5994 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5995 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6001 char numa_zonelist_order[] = "Node";
6004 * sysctl handler for numa_zonelist_order
6006 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6007 void *buffer, size_t *length, loff_t *ppos)
6010 return __parse_numa_zonelist_order(buffer);
6011 return proc_dostring(table, write, buffer, length, ppos);
6015 #define MAX_NODE_LOAD (nr_online_nodes)
6016 static int node_load[MAX_NUMNODES];
6019 * find_next_best_node - find the next node that should appear in a given node's fallback list
6020 * @node: node whose fallback list we're appending
6021 * @used_node_mask: nodemask_t of already used nodes
6023 * We use a number of factors to determine which is the next node that should
6024 * appear on a given node's fallback list. The node should not have appeared
6025 * already in @node's fallback list, and it should be the next closest node
6026 * according to the distance array (which contains arbitrary distance values
6027 * from each node to each node in the system), and should also prefer nodes
6028 * with no CPUs, since presumably they'll have very little allocation pressure
6029 * on them otherwise.
6031 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6033 static int find_next_best_node(int node, nodemask_t *used_node_mask)
6036 int min_val = INT_MAX;
6037 int best_node = NUMA_NO_NODE;
6039 /* Use the local node if we haven't already */
6040 if (!node_isset(node, *used_node_mask)) {
6041 node_set(node, *used_node_mask);
6045 for_each_node_state(n, N_MEMORY) {
6047 /* Don't want a node to appear more than once */
6048 if (node_isset(n, *used_node_mask))
6051 /* Use the distance array to find the distance */
6052 val = node_distance(node, n);
6054 /* Penalize nodes under us ("prefer the next node") */
6057 /* Give preference to headless and unused nodes */
6058 if (!cpumask_empty(cpumask_of_node(n)))
6059 val += PENALTY_FOR_NODE_WITH_CPUS;
6061 /* Slight preference for less loaded node */
6062 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6063 val += node_load[n];
6065 if (val < min_val) {
6072 node_set(best_node, *used_node_mask);
6079 * Build zonelists ordered by node and zones within node.
6080 * This results in maximum locality--normal zone overflows into local
6081 * DMA zone, if any--but risks exhausting DMA zone.
6083 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6086 struct zoneref *zonerefs;
6089 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6091 for (i = 0; i < nr_nodes; i++) {
6094 pg_data_t *node = NODE_DATA(node_order[i]);
6096 nr_zones = build_zonerefs_node(node, zonerefs);
6097 zonerefs += nr_zones;
6099 zonerefs->zone = NULL;
6100 zonerefs->zone_idx = 0;
6104 * Build gfp_thisnode zonelists
6106 static void build_thisnode_zonelists(pg_data_t *pgdat)
6108 struct zoneref *zonerefs;
6111 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6112 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6113 zonerefs += nr_zones;
6114 zonerefs->zone = NULL;
6115 zonerefs->zone_idx = 0;
6119 * Build zonelists ordered by zone and nodes within zones.
6120 * This results in conserving DMA zone[s] until all Normal memory is
6121 * exhausted, but results in overflowing to remote node while memory
6122 * may still exist in local DMA zone.
6125 static void build_zonelists(pg_data_t *pgdat)
6127 static int node_order[MAX_NUMNODES];
6128 int node, load, nr_nodes = 0;
6129 nodemask_t used_mask = NODE_MASK_NONE;
6130 int local_node, prev_node;
6132 /* NUMA-aware ordering of nodes */
6133 local_node = pgdat->node_id;
6134 load = nr_online_nodes;
6135 prev_node = local_node;
6137 memset(node_order, 0, sizeof(node_order));
6138 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6140 * We don't want to pressure a particular node.
6141 * So adding penalty to the first node in same
6142 * distance group to make it round-robin.
6144 if (node_distance(local_node, node) !=
6145 node_distance(local_node, prev_node))
6146 node_load[node] = load;
6148 node_order[nr_nodes++] = node;
6153 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6154 build_thisnode_zonelists(pgdat);
6157 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6159 * Return node id of node used for "local" allocations.
6160 * I.e., first node id of first zone in arg node's generic zonelist.
6161 * Used for initializing percpu 'numa_mem', which is used primarily
6162 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6164 int local_memory_node(int node)
6168 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6169 gfp_zone(GFP_KERNEL),
6171 return zone_to_nid(z->zone);
6175 static void setup_min_unmapped_ratio(void);
6176 static void setup_min_slab_ratio(void);
6177 #else /* CONFIG_NUMA */
6179 static void build_zonelists(pg_data_t *pgdat)
6181 int node, local_node;
6182 struct zoneref *zonerefs;
6185 local_node = pgdat->node_id;
6187 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6188 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6189 zonerefs += nr_zones;
6192 * Now we build the zonelist so that it contains the zones
6193 * of all the other nodes.
6194 * We don't want to pressure a particular node, so when
6195 * building the zones for node N, we make sure that the
6196 * zones coming right after the local ones are those from
6197 * node N+1 (modulo N)
6199 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6200 if (!node_online(node))
6202 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6203 zonerefs += nr_zones;
6205 for (node = 0; node < local_node; node++) {
6206 if (!node_online(node))
6208 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6209 zonerefs += nr_zones;
6212 zonerefs->zone = NULL;
6213 zonerefs->zone_idx = 0;
6216 #endif /* CONFIG_NUMA */
6219 * Boot pageset table. One per cpu which is going to be used for all
6220 * zones and all nodes. The parameters will be set in such a way
6221 * that an item put on a list will immediately be handed over to
6222 * the buddy list. This is safe since pageset manipulation is done
6223 * with interrupts disabled.
6225 * The boot_pagesets must be kept even after bootup is complete for
6226 * unused processors and/or zones. They do play a role for bootstrapping
6227 * hotplugged processors.
6229 * zoneinfo_show() and maybe other functions do
6230 * not check if the processor is online before following the pageset pointer.
6231 * Other parts of the kernel may not check if the zone is available.
6233 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6234 /* These effectively disable the pcplists in the boot pageset completely */
6235 #define BOOT_PAGESET_HIGH 0
6236 #define BOOT_PAGESET_BATCH 1
6237 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6238 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6239 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6241 static void __build_all_zonelists(void *data)
6244 int __maybe_unused cpu;
6245 pg_data_t *self = data;
6246 static DEFINE_SPINLOCK(lock);
6251 memset(node_load, 0, sizeof(node_load));
6255 * This node is hotadded and no memory is yet present. So just
6256 * building zonelists is fine - no need to touch other nodes.
6258 if (self && !node_online(self->node_id)) {
6259 build_zonelists(self);
6261 for_each_online_node(nid) {
6262 pg_data_t *pgdat = NODE_DATA(nid);
6264 build_zonelists(pgdat);
6267 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6269 * We now know the "local memory node" for each node--
6270 * i.e., the node of the first zone in the generic zonelist.
6271 * Set up numa_mem percpu variable for on-line cpus. During
6272 * boot, only the boot cpu should be on-line; we'll init the
6273 * secondary cpus' numa_mem as they come on-line. During
6274 * node/memory hotplug, we'll fixup all on-line cpus.
6276 for_each_online_cpu(cpu)
6277 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6284 static noinline void __init
6285 build_all_zonelists_init(void)
6289 __build_all_zonelists(NULL);
6292 * Initialize the boot_pagesets that are going to be used
6293 * for bootstrapping processors. The real pagesets for
6294 * each zone will be allocated later when the per cpu
6295 * allocator is available.
6297 * boot_pagesets are used also for bootstrapping offline
6298 * cpus if the system is already booted because the pagesets
6299 * are needed to initialize allocators on a specific cpu too.
6300 * F.e. the percpu allocator needs the page allocator which
6301 * needs the percpu allocator in order to allocate its pagesets
6302 * (a chicken-egg dilemma).
6304 for_each_possible_cpu(cpu)
6305 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6307 mminit_verify_zonelist();
6308 cpuset_init_current_mems_allowed();
6312 * unless system_state == SYSTEM_BOOTING.
6314 * __ref due to call of __init annotated helper build_all_zonelists_init
6315 * [protected by SYSTEM_BOOTING].
6317 void __ref build_all_zonelists(pg_data_t *pgdat)
6319 unsigned long vm_total_pages;
6321 if (system_state == SYSTEM_BOOTING) {
6322 build_all_zonelists_init();
6324 __build_all_zonelists(pgdat);
6325 /* cpuset refresh routine should be here */
6327 /* Get the number of free pages beyond high watermark in all zones. */
6328 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6330 * Disable grouping by mobility if the number of pages in the
6331 * system is too low to allow the mechanism to work. It would be
6332 * more accurate, but expensive to check per-zone. This check is
6333 * made on memory-hotadd so a system can start with mobility
6334 * disabled and enable it later
6336 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6337 page_group_by_mobility_disabled = 1;
6339 page_group_by_mobility_disabled = 0;
6341 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6343 page_group_by_mobility_disabled ? "off" : "on",
6346 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6350 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6351 static bool __meminit
6352 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6354 static struct memblock_region *r;
6356 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6357 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6358 for_each_mem_region(r) {
6359 if (*pfn < memblock_region_memory_end_pfn(r))
6363 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6364 memblock_is_mirror(r)) {
6365 *pfn = memblock_region_memory_end_pfn(r);
6373 * Initially all pages are reserved - free ones are freed
6374 * up by memblock_free_all() once the early boot process is
6375 * done. Non-atomic initialization, single-pass.
6377 * All aligned pageblocks are initialized to the specified migratetype
6378 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6379 * zone stats (e.g., nr_isolate_pageblock) are touched.
6381 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6382 unsigned long start_pfn, unsigned long zone_end_pfn,
6383 enum meminit_context context,
6384 struct vmem_altmap *altmap, int migratetype)
6386 unsigned long pfn, end_pfn = start_pfn + size;
6389 if (highest_memmap_pfn < end_pfn - 1)
6390 highest_memmap_pfn = end_pfn - 1;
6392 #ifdef CONFIG_ZONE_DEVICE
6394 * Honor reservation requested by the driver for this ZONE_DEVICE
6395 * memory. We limit the total number of pages to initialize to just
6396 * those that might contain the memory mapping. We will defer the
6397 * ZONE_DEVICE page initialization until after we have released
6400 if (zone == ZONE_DEVICE) {
6404 if (start_pfn == altmap->base_pfn)
6405 start_pfn += altmap->reserve;
6406 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6410 for (pfn = start_pfn; pfn < end_pfn; ) {
6412 * There can be holes in boot-time mem_map[]s handed to this
6413 * function. They do not exist on hotplugged memory.
6415 if (context == MEMINIT_EARLY) {
6416 if (overlap_memmap_init(zone, &pfn))
6418 if (defer_init(nid, pfn, zone_end_pfn))
6422 page = pfn_to_page(pfn);
6423 __init_single_page(page, pfn, zone, nid);
6424 if (context == MEMINIT_HOTPLUG)
6425 __SetPageReserved(page);
6428 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6429 * such that unmovable allocations won't be scattered all
6430 * over the place during system boot.
6432 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6433 set_pageblock_migratetype(page, migratetype);
6440 #ifdef CONFIG_ZONE_DEVICE
6441 void __ref memmap_init_zone_device(struct zone *zone,
6442 unsigned long start_pfn,
6443 unsigned long nr_pages,
6444 struct dev_pagemap *pgmap)
6446 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6447 struct pglist_data *pgdat = zone->zone_pgdat;
6448 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6449 unsigned long zone_idx = zone_idx(zone);
6450 unsigned long start = jiffies;
6451 int nid = pgdat->node_id;
6453 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6457 * The call to memmap_init should have already taken care
6458 * of the pages reserved for the memmap, so we can just jump to
6459 * the end of that region and start processing the device pages.
6462 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6463 nr_pages = end_pfn - start_pfn;
6466 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6467 struct page *page = pfn_to_page(pfn);
6469 __init_single_page(page, pfn, zone_idx, nid);
6472 * Mark page reserved as it will need to wait for onlining
6473 * phase for it to be fully associated with a zone.
6475 * We can use the non-atomic __set_bit operation for setting
6476 * the flag as we are still initializing the pages.
6478 __SetPageReserved(page);
6481 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6482 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6483 * ever freed or placed on a driver-private list.
6485 page->pgmap = pgmap;
6486 page->zone_device_data = NULL;
6489 * Mark the block movable so that blocks are reserved for
6490 * movable at startup. This will force kernel allocations
6491 * to reserve their blocks rather than leaking throughout
6492 * the address space during boot when many long-lived
6493 * kernel allocations are made.
6495 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6496 * because this is done early in section_activate()
6498 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6499 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6504 pr_info("%s initialised %lu pages in %ums\n", __func__,
6505 nr_pages, jiffies_to_msecs(jiffies - start));
6509 static void __meminit zone_init_free_lists(struct zone *zone)
6511 unsigned int order, t;
6512 for_each_migratetype_order(order, t) {
6513 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6514 zone->free_area[order].nr_free = 0;
6518 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6520 * Only struct pages that correspond to ranges defined by memblock.memory
6521 * are zeroed and initialized by going through __init_single_page() during
6522 * memmap_init_zone_range().
6524 * But, there could be struct pages that correspond to holes in
6525 * memblock.memory. This can happen because of the following reasons:
6526 * - physical memory bank size is not necessarily the exact multiple of the
6527 * arbitrary section size
6528 * - early reserved memory may not be listed in memblock.memory
6529 * - memory layouts defined with memmap= kernel parameter may not align
6530 * nicely with memmap sections
6532 * Explicitly initialize those struct pages so that:
6533 * - PG_Reserved is set
6534 * - zone and node links point to zone and node that span the page if the
6535 * hole is in the middle of a zone
6536 * - zone and node links point to adjacent zone/node if the hole falls on
6537 * the zone boundary; the pages in such holes will be prepended to the
6538 * zone/node above the hole except for the trailing pages in the last
6539 * section that will be appended to the zone/node below.
6541 static void __init init_unavailable_range(unsigned long spfn,
6548 for (pfn = spfn; pfn < epfn; pfn++) {
6549 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6550 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6551 + pageblock_nr_pages - 1;
6554 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6555 __SetPageReserved(pfn_to_page(pfn));
6560 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6561 node, zone_names[zone], pgcnt);
6564 static inline void init_unavailable_range(unsigned long spfn,
6571 static void __init memmap_init_zone_range(struct zone *zone,
6572 unsigned long start_pfn,
6573 unsigned long end_pfn,
6574 unsigned long *hole_pfn)
6576 unsigned long zone_start_pfn = zone->zone_start_pfn;
6577 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6578 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6580 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6581 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6583 if (start_pfn >= end_pfn)
6586 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6587 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6589 if (*hole_pfn < start_pfn)
6590 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6592 *hole_pfn = end_pfn;
6595 static void __init memmap_init(void)
6597 unsigned long start_pfn, end_pfn;
6598 unsigned long hole_pfn = 0;
6599 int i, j, zone_id, nid;
6601 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6602 struct pglist_data *node = NODE_DATA(nid);
6604 for (j = 0; j < MAX_NR_ZONES; j++) {
6605 struct zone *zone = node->node_zones + j;
6607 if (!populated_zone(zone))
6610 memmap_init_zone_range(zone, start_pfn, end_pfn,
6616 #ifdef CONFIG_SPARSEMEM
6618 * Initialize the memory map for hole in the range [memory_end,
6620 * Append the pages in this hole to the highest zone in the last
6622 * The call to init_unavailable_range() is outside the ifdef to
6623 * silence the compiler warining about zone_id set but not used;
6624 * for FLATMEM it is a nop anyway
6626 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6627 if (hole_pfn < end_pfn)
6629 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6632 static int zone_batchsize(struct zone *zone)
6638 * The per-cpu-pages pools are set to around 1000th of the
6641 batch = zone_managed_pages(zone) / 1024;
6642 /* But no more than a meg. */
6643 if (batch * PAGE_SIZE > 1024 * 1024)
6644 batch = (1024 * 1024) / PAGE_SIZE;
6645 batch /= 4; /* We effectively *= 4 below */
6650 * Clamp the batch to a 2^n - 1 value. Having a power
6651 * of 2 value was found to be more likely to have
6652 * suboptimal cache aliasing properties in some cases.
6654 * For example if 2 tasks are alternately allocating
6655 * batches of pages, one task can end up with a lot
6656 * of pages of one half of the possible page colors
6657 * and the other with pages of the other colors.
6659 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6664 /* The deferral and batching of frees should be suppressed under NOMMU
6667 * The problem is that NOMMU needs to be able to allocate large chunks
6668 * of contiguous memory as there's no hardware page translation to
6669 * assemble apparent contiguous memory from discontiguous pages.
6671 * Queueing large contiguous runs of pages for batching, however,
6672 * causes the pages to actually be freed in smaller chunks. As there
6673 * can be a significant delay between the individual batches being
6674 * recycled, this leads to the once large chunks of space being
6675 * fragmented and becoming unavailable for high-order allocations.
6682 * pcp->high and pcp->batch values are related and generally batch is lower
6683 * than high. They are also related to pcp->count such that count is lower
6684 * than high, and as soon as it reaches high, the pcplist is flushed.
6686 * However, guaranteeing these relations at all times would require e.g. write
6687 * barriers here but also careful usage of read barriers at the read side, and
6688 * thus be prone to error and bad for performance. Thus the update only prevents
6689 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6690 * can cope with those fields changing asynchronously, and fully trust only the
6691 * pcp->count field on the local CPU with interrupts disabled.
6693 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6694 * outside of boot time (or some other assurance that no concurrent updaters
6697 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6698 unsigned long batch)
6700 WRITE_ONCE(pcp->batch, batch);
6701 WRITE_ONCE(pcp->high, high);
6704 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6708 memset(pcp, 0, sizeof(*pcp));
6709 memset(pzstats, 0, sizeof(*pzstats));
6711 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6712 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6715 * Set batch and high values safe for a boot pageset. A true percpu
6716 * pageset's initialization will update them subsequently. Here we don't
6717 * need to be as careful as pageset_update() as nobody can access the
6720 pcp->high = BOOT_PAGESET_HIGH;
6721 pcp->batch = BOOT_PAGESET_BATCH;
6724 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6725 unsigned long batch)
6727 struct per_cpu_pages *pcp;
6730 for_each_possible_cpu(cpu) {
6731 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6732 pageset_update(pcp, high, batch);
6737 * Calculate and set new high and batch values for all per-cpu pagesets of a
6738 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6740 static void zone_set_pageset_high_and_batch(struct zone *zone)
6742 unsigned long new_high, new_batch;
6744 if (percpu_pagelist_fraction) {
6745 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6746 new_batch = max(1UL, new_high / 4);
6747 if ((new_high / 4) > (PAGE_SHIFT * 8))
6748 new_batch = PAGE_SHIFT * 8;
6750 new_batch = zone_batchsize(zone);
6751 new_high = 6 * new_batch;
6752 new_batch = max(1UL, 1 * new_batch);
6755 if (zone->pageset_high == new_high &&
6756 zone->pageset_batch == new_batch)
6759 zone->pageset_high = new_high;
6760 zone->pageset_batch = new_batch;
6762 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6765 void __meminit setup_zone_pageset(struct zone *zone)
6769 /* Size may be 0 on !SMP && !NUMA */
6770 if (sizeof(struct per_cpu_zonestat) > 0)
6771 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6773 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6774 for_each_possible_cpu(cpu) {
6775 struct per_cpu_pages *pcp;
6776 struct per_cpu_zonestat *pzstats;
6778 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6779 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6780 per_cpu_pages_init(pcp, pzstats);
6783 zone_set_pageset_high_and_batch(zone);
6787 * Allocate per cpu pagesets and initialize them.
6788 * Before this call only boot pagesets were available.
6790 void __init setup_per_cpu_pageset(void)
6792 struct pglist_data *pgdat;
6794 int __maybe_unused cpu;
6796 for_each_populated_zone(zone)
6797 setup_zone_pageset(zone);
6801 * Unpopulated zones continue using the boot pagesets.
6802 * The numa stats for these pagesets need to be reset.
6803 * Otherwise, they will end up skewing the stats of
6804 * the nodes these zones are associated with.
6806 for_each_possible_cpu(cpu) {
6807 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
6808 memset(pzstats->vm_numa_event, 0,
6809 sizeof(pzstats->vm_numa_event));
6813 for_each_online_pgdat(pgdat)
6814 pgdat->per_cpu_nodestats =
6815 alloc_percpu(struct per_cpu_nodestat);
6818 static __meminit void zone_pcp_init(struct zone *zone)
6821 * per cpu subsystem is not up at this point. The following code
6822 * relies on the ability of the linker to provide the
6823 * offset of a (static) per cpu variable into the per cpu area.
6825 zone->per_cpu_pageset = &boot_pageset;
6826 zone->per_cpu_zonestats = &boot_zonestats;
6827 zone->pageset_high = BOOT_PAGESET_HIGH;
6828 zone->pageset_batch = BOOT_PAGESET_BATCH;
6830 if (populated_zone(zone))
6831 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
6832 zone->present_pages, zone_batchsize(zone));
6835 void __meminit init_currently_empty_zone(struct zone *zone,
6836 unsigned long zone_start_pfn,
6839 struct pglist_data *pgdat = zone->zone_pgdat;
6840 int zone_idx = zone_idx(zone) + 1;
6842 if (zone_idx > pgdat->nr_zones)
6843 pgdat->nr_zones = zone_idx;
6845 zone->zone_start_pfn = zone_start_pfn;
6847 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6848 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6850 (unsigned long)zone_idx(zone),
6851 zone_start_pfn, (zone_start_pfn + size));
6853 zone_init_free_lists(zone);
6854 zone->initialized = 1;
6858 * get_pfn_range_for_nid - Return the start and end page frames for a node
6859 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6860 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6861 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6863 * It returns the start and end page frame of a node based on information
6864 * provided by memblock_set_node(). If called for a node
6865 * with no available memory, a warning is printed and the start and end
6868 void __init get_pfn_range_for_nid(unsigned int nid,
6869 unsigned long *start_pfn, unsigned long *end_pfn)
6871 unsigned long this_start_pfn, this_end_pfn;
6877 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6878 *start_pfn = min(*start_pfn, this_start_pfn);
6879 *end_pfn = max(*end_pfn, this_end_pfn);
6882 if (*start_pfn == -1UL)
6887 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6888 * assumption is made that zones within a node are ordered in monotonic
6889 * increasing memory addresses so that the "highest" populated zone is used
6891 static void __init find_usable_zone_for_movable(void)
6894 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6895 if (zone_index == ZONE_MOVABLE)
6898 if (arch_zone_highest_possible_pfn[zone_index] >
6899 arch_zone_lowest_possible_pfn[zone_index])
6903 VM_BUG_ON(zone_index == -1);
6904 movable_zone = zone_index;
6908 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6909 * because it is sized independent of architecture. Unlike the other zones,
6910 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6911 * in each node depending on the size of each node and how evenly kernelcore
6912 * is distributed. This helper function adjusts the zone ranges
6913 * provided by the architecture for a given node by using the end of the
6914 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6915 * zones within a node are in order of monotonic increases memory addresses
6917 static void __init adjust_zone_range_for_zone_movable(int nid,
6918 unsigned long zone_type,
6919 unsigned long node_start_pfn,
6920 unsigned long node_end_pfn,
6921 unsigned long *zone_start_pfn,
6922 unsigned long *zone_end_pfn)
6924 /* Only adjust if ZONE_MOVABLE is on this node */
6925 if (zone_movable_pfn[nid]) {
6926 /* Size ZONE_MOVABLE */
6927 if (zone_type == ZONE_MOVABLE) {
6928 *zone_start_pfn = zone_movable_pfn[nid];
6929 *zone_end_pfn = min(node_end_pfn,
6930 arch_zone_highest_possible_pfn[movable_zone]);
6932 /* Adjust for ZONE_MOVABLE starting within this range */
6933 } else if (!mirrored_kernelcore &&
6934 *zone_start_pfn < zone_movable_pfn[nid] &&
6935 *zone_end_pfn > zone_movable_pfn[nid]) {
6936 *zone_end_pfn = zone_movable_pfn[nid];
6938 /* Check if this whole range is within ZONE_MOVABLE */
6939 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6940 *zone_start_pfn = *zone_end_pfn;
6945 * Return the number of pages a zone spans in a node, including holes
6946 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6948 static unsigned long __init zone_spanned_pages_in_node(int nid,
6949 unsigned long zone_type,
6950 unsigned long node_start_pfn,
6951 unsigned long node_end_pfn,
6952 unsigned long *zone_start_pfn,
6953 unsigned long *zone_end_pfn)
6955 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6956 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6957 /* When hotadd a new node from cpu_up(), the node should be empty */
6958 if (!node_start_pfn && !node_end_pfn)
6961 /* Get the start and end of the zone */
6962 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6963 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6964 adjust_zone_range_for_zone_movable(nid, zone_type,
6965 node_start_pfn, node_end_pfn,
6966 zone_start_pfn, zone_end_pfn);
6968 /* Check that this node has pages within the zone's required range */
6969 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6972 /* Move the zone boundaries inside the node if necessary */
6973 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6974 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6976 /* Return the spanned pages */
6977 return *zone_end_pfn - *zone_start_pfn;
6981 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6982 * then all holes in the requested range will be accounted for.
6984 unsigned long __init __absent_pages_in_range(int nid,
6985 unsigned long range_start_pfn,
6986 unsigned long range_end_pfn)
6988 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6989 unsigned long start_pfn, end_pfn;
6992 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6993 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6994 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6995 nr_absent -= end_pfn - start_pfn;
7001 * absent_pages_in_range - Return number of page frames in holes within a range
7002 * @start_pfn: The start PFN to start searching for holes
7003 * @end_pfn: The end PFN to stop searching for holes
7005 * Return: the number of pages frames in memory holes within a range.
7007 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7008 unsigned long end_pfn)
7010 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7013 /* Return the number of page frames in holes in a zone on a node */
7014 static unsigned long __init zone_absent_pages_in_node(int nid,
7015 unsigned long zone_type,
7016 unsigned long node_start_pfn,
7017 unsigned long node_end_pfn)
7019 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7020 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7021 unsigned long zone_start_pfn, zone_end_pfn;
7022 unsigned long nr_absent;
7024 /* When hotadd a new node from cpu_up(), the node should be empty */
7025 if (!node_start_pfn && !node_end_pfn)
7028 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7029 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7031 adjust_zone_range_for_zone_movable(nid, zone_type,
7032 node_start_pfn, node_end_pfn,
7033 &zone_start_pfn, &zone_end_pfn);
7034 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7037 * ZONE_MOVABLE handling.
7038 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7041 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7042 unsigned long start_pfn, end_pfn;
7043 struct memblock_region *r;
7045 for_each_mem_region(r) {
7046 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7047 zone_start_pfn, zone_end_pfn);
7048 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7049 zone_start_pfn, zone_end_pfn);
7051 if (zone_type == ZONE_MOVABLE &&
7052 memblock_is_mirror(r))
7053 nr_absent += end_pfn - start_pfn;
7055 if (zone_type == ZONE_NORMAL &&
7056 !memblock_is_mirror(r))
7057 nr_absent += end_pfn - start_pfn;
7064 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7065 unsigned long node_start_pfn,
7066 unsigned long node_end_pfn)
7068 unsigned long realtotalpages = 0, totalpages = 0;
7071 for (i = 0; i < MAX_NR_ZONES; i++) {
7072 struct zone *zone = pgdat->node_zones + i;
7073 unsigned long zone_start_pfn, zone_end_pfn;
7074 unsigned long spanned, absent;
7075 unsigned long size, real_size;
7077 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7082 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7087 real_size = size - absent;
7090 zone->zone_start_pfn = zone_start_pfn;
7092 zone->zone_start_pfn = 0;
7093 zone->spanned_pages = size;
7094 zone->present_pages = real_size;
7097 realtotalpages += real_size;
7100 pgdat->node_spanned_pages = totalpages;
7101 pgdat->node_present_pages = realtotalpages;
7102 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7105 #ifndef CONFIG_SPARSEMEM
7107 * Calculate the size of the zone->blockflags rounded to an unsigned long
7108 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7109 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7110 * round what is now in bits to nearest long in bits, then return it in
7113 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7115 unsigned long usemapsize;
7117 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7118 usemapsize = roundup(zonesize, pageblock_nr_pages);
7119 usemapsize = usemapsize >> pageblock_order;
7120 usemapsize *= NR_PAGEBLOCK_BITS;
7121 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7123 return usemapsize / 8;
7126 static void __ref setup_usemap(struct zone *zone)
7128 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7129 zone->spanned_pages);
7130 zone->pageblock_flags = NULL;
7132 zone->pageblock_flags =
7133 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7135 if (!zone->pageblock_flags)
7136 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7137 usemapsize, zone->name, zone_to_nid(zone));
7141 static inline void setup_usemap(struct zone *zone) {}
7142 #endif /* CONFIG_SPARSEMEM */
7144 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7146 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7147 void __init set_pageblock_order(void)
7151 /* Check that pageblock_nr_pages has not already been setup */
7152 if (pageblock_order)
7155 if (HPAGE_SHIFT > PAGE_SHIFT)
7156 order = HUGETLB_PAGE_ORDER;
7158 order = MAX_ORDER - 1;
7161 * Assume the largest contiguous order of interest is a huge page.
7162 * This value may be variable depending on boot parameters on IA64 and
7165 pageblock_order = order;
7167 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7170 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7171 * is unused as pageblock_order is set at compile-time. See
7172 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7175 void __init set_pageblock_order(void)
7179 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7181 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7182 unsigned long present_pages)
7184 unsigned long pages = spanned_pages;
7187 * Provide a more accurate estimation if there are holes within
7188 * the zone and SPARSEMEM is in use. If there are holes within the
7189 * zone, each populated memory region may cost us one or two extra
7190 * memmap pages due to alignment because memmap pages for each
7191 * populated regions may not be naturally aligned on page boundary.
7192 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7194 if (spanned_pages > present_pages + (present_pages >> 4) &&
7195 IS_ENABLED(CONFIG_SPARSEMEM))
7196 pages = present_pages;
7198 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7201 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7202 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7204 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7206 spin_lock_init(&ds_queue->split_queue_lock);
7207 INIT_LIST_HEAD(&ds_queue->split_queue);
7208 ds_queue->split_queue_len = 0;
7211 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7214 #ifdef CONFIG_COMPACTION
7215 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7217 init_waitqueue_head(&pgdat->kcompactd_wait);
7220 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7223 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7225 pgdat_resize_init(pgdat);
7227 pgdat_init_split_queue(pgdat);
7228 pgdat_init_kcompactd(pgdat);
7230 init_waitqueue_head(&pgdat->kswapd_wait);
7231 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7233 pgdat_page_ext_init(pgdat);
7234 lruvec_init(&pgdat->__lruvec);
7237 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7238 unsigned long remaining_pages)
7240 atomic_long_set(&zone->managed_pages, remaining_pages);
7241 zone_set_nid(zone, nid);
7242 zone->name = zone_names[idx];
7243 zone->zone_pgdat = NODE_DATA(nid);
7244 spin_lock_init(&zone->lock);
7245 zone_seqlock_init(zone);
7246 zone_pcp_init(zone);
7250 * Set up the zone data structures
7251 * - init pgdat internals
7252 * - init all zones belonging to this node
7254 * NOTE: this function is only called during memory hotplug
7256 #ifdef CONFIG_MEMORY_HOTPLUG
7257 void __ref free_area_init_core_hotplug(int nid)
7260 pg_data_t *pgdat = NODE_DATA(nid);
7262 pgdat_init_internals(pgdat);
7263 for (z = 0; z < MAX_NR_ZONES; z++)
7264 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7269 * Set up the zone data structures:
7270 * - mark all pages reserved
7271 * - mark all memory queues empty
7272 * - clear the memory bitmaps
7274 * NOTE: pgdat should get zeroed by caller.
7275 * NOTE: this function is only called during early init.
7277 static void __init free_area_init_core(struct pglist_data *pgdat)
7280 int nid = pgdat->node_id;
7282 pgdat_init_internals(pgdat);
7283 pgdat->per_cpu_nodestats = &boot_nodestats;
7285 for (j = 0; j < MAX_NR_ZONES; j++) {
7286 struct zone *zone = pgdat->node_zones + j;
7287 unsigned long size, freesize, memmap_pages;
7289 size = zone->spanned_pages;
7290 freesize = zone->present_pages;
7293 * Adjust freesize so that it accounts for how much memory
7294 * is used by this zone for memmap. This affects the watermark
7295 * and per-cpu initialisations
7297 memmap_pages = calc_memmap_size(size, freesize);
7298 if (!is_highmem_idx(j)) {
7299 if (freesize >= memmap_pages) {
7300 freesize -= memmap_pages;
7302 pr_debug(" %s zone: %lu pages used for memmap\n",
7303 zone_names[j], memmap_pages);
7305 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
7306 zone_names[j], memmap_pages, freesize);
7309 /* Account for reserved pages */
7310 if (j == 0 && freesize > dma_reserve) {
7311 freesize -= dma_reserve;
7312 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7315 if (!is_highmem_idx(j))
7316 nr_kernel_pages += freesize;
7317 /* Charge for highmem memmap if there are enough kernel pages */
7318 else if (nr_kernel_pages > memmap_pages * 2)
7319 nr_kernel_pages -= memmap_pages;
7320 nr_all_pages += freesize;
7323 * Set an approximate value for lowmem here, it will be adjusted
7324 * when the bootmem allocator frees pages into the buddy system.
7325 * And all highmem pages will be managed by the buddy system.
7327 zone_init_internals(zone, j, nid, freesize);
7332 set_pageblock_order();
7334 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7338 #ifdef CONFIG_FLAT_NODE_MEM_MAP
7339 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7341 unsigned long __maybe_unused start = 0;
7342 unsigned long __maybe_unused offset = 0;
7344 /* Skip empty nodes */
7345 if (!pgdat->node_spanned_pages)
7348 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7349 offset = pgdat->node_start_pfn - start;
7350 /* ia64 gets its own node_mem_map, before this, without bootmem */
7351 if (!pgdat->node_mem_map) {
7352 unsigned long size, end;
7356 * The zone's endpoints aren't required to be MAX_ORDER
7357 * aligned but the node_mem_map endpoints must be in order
7358 * for the buddy allocator to function correctly.
7360 end = pgdat_end_pfn(pgdat);
7361 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7362 size = (end - start) * sizeof(struct page);
7363 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7366 panic("Failed to allocate %ld bytes for node %d memory map\n",
7367 size, pgdat->node_id);
7368 pgdat->node_mem_map = map + offset;
7370 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7371 __func__, pgdat->node_id, (unsigned long)pgdat,
7372 (unsigned long)pgdat->node_mem_map);
7373 #ifndef CONFIG_NEED_MULTIPLE_NODES
7375 * With no DISCONTIG, the global mem_map is just set as node 0's
7377 if (pgdat == NODE_DATA(0)) {
7378 mem_map = NODE_DATA(0)->node_mem_map;
7379 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7385 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7386 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7388 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7389 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7391 pgdat->first_deferred_pfn = ULONG_MAX;
7394 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7397 static void __init free_area_init_node(int nid)
7399 pg_data_t *pgdat = NODE_DATA(nid);
7400 unsigned long start_pfn = 0;
7401 unsigned long end_pfn = 0;
7403 /* pg_data_t should be reset to zero when it's allocated */
7404 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7406 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7408 pgdat->node_id = nid;
7409 pgdat->node_start_pfn = start_pfn;
7410 pgdat->per_cpu_nodestats = NULL;
7412 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7413 (u64)start_pfn << PAGE_SHIFT,
7414 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7415 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7417 alloc_node_mem_map(pgdat);
7418 pgdat_set_deferred_range(pgdat);
7420 free_area_init_core(pgdat);
7423 void __init free_area_init_memoryless_node(int nid)
7425 free_area_init_node(nid);
7428 #if MAX_NUMNODES > 1
7430 * Figure out the number of possible node ids.
7432 void __init setup_nr_node_ids(void)
7434 unsigned int highest;
7436 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7437 nr_node_ids = highest + 1;
7442 * node_map_pfn_alignment - determine the maximum internode alignment
7444 * This function should be called after node map is populated and sorted.
7445 * It calculates the maximum power of two alignment which can distinguish
7448 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7449 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7450 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7451 * shifted, 1GiB is enough and this function will indicate so.
7453 * This is used to test whether pfn -> nid mapping of the chosen memory
7454 * model has fine enough granularity to avoid incorrect mapping for the
7455 * populated node map.
7457 * Return: the determined alignment in pfn's. 0 if there is no alignment
7458 * requirement (single node).
7460 unsigned long __init node_map_pfn_alignment(void)
7462 unsigned long accl_mask = 0, last_end = 0;
7463 unsigned long start, end, mask;
7464 int last_nid = NUMA_NO_NODE;
7467 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7468 if (!start || last_nid < 0 || last_nid == nid) {
7475 * Start with a mask granular enough to pin-point to the
7476 * start pfn and tick off bits one-by-one until it becomes
7477 * too coarse to separate the current node from the last.
7479 mask = ~((1 << __ffs(start)) - 1);
7480 while (mask && last_end <= (start & (mask << 1)))
7483 /* accumulate all internode masks */
7487 /* convert mask to number of pages */
7488 return ~accl_mask + 1;
7492 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7494 * Return: the minimum PFN based on information provided via
7495 * memblock_set_node().
7497 unsigned long __init find_min_pfn_with_active_regions(void)
7499 return PHYS_PFN(memblock_start_of_DRAM());
7503 * early_calculate_totalpages()
7504 * Sum pages in active regions for movable zone.
7505 * Populate N_MEMORY for calculating usable_nodes.
7507 static unsigned long __init early_calculate_totalpages(void)
7509 unsigned long totalpages = 0;
7510 unsigned long start_pfn, end_pfn;
7513 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7514 unsigned long pages = end_pfn - start_pfn;
7516 totalpages += pages;
7518 node_set_state(nid, N_MEMORY);
7524 * Find the PFN the Movable zone begins in each node. Kernel memory
7525 * is spread evenly between nodes as long as the nodes have enough
7526 * memory. When they don't, some nodes will have more kernelcore than
7529 static void __init find_zone_movable_pfns_for_nodes(void)
7532 unsigned long usable_startpfn;
7533 unsigned long kernelcore_node, kernelcore_remaining;
7534 /* save the state before borrow the nodemask */
7535 nodemask_t saved_node_state = node_states[N_MEMORY];
7536 unsigned long totalpages = early_calculate_totalpages();
7537 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7538 struct memblock_region *r;
7540 /* Need to find movable_zone earlier when movable_node is specified. */
7541 find_usable_zone_for_movable();
7544 * If movable_node is specified, ignore kernelcore and movablecore
7547 if (movable_node_is_enabled()) {
7548 for_each_mem_region(r) {
7549 if (!memblock_is_hotpluggable(r))
7552 nid = memblock_get_region_node(r);
7554 usable_startpfn = PFN_DOWN(r->base);
7555 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7556 min(usable_startpfn, zone_movable_pfn[nid]) :
7564 * If kernelcore=mirror is specified, ignore movablecore option
7566 if (mirrored_kernelcore) {
7567 bool mem_below_4gb_not_mirrored = false;
7569 for_each_mem_region(r) {
7570 if (memblock_is_mirror(r))
7573 nid = memblock_get_region_node(r);
7575 usable_startpfn = memblock_region_memory_base_pfn(r);
7577 if (usable_startpfn < 0x100000) {
7578 mem_below_4gb_not_mirrored = true;
7582 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7583 min(usable_startpfn, zone_movable_pfn[nid]) :
7587 if (mem_below_4gb_not_mirrored)
7588 pr_warn("This configuration results in unmirrored kernel memory.\n");
7594 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7595 * amount of necessary memory.
7597 if (required_kernelcore_percent)
7598 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7600 if (required_movablecore_percent)
7601 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7605 * If movablecore= was specified, calculate what size of
7606 * kernelcore that corresponds so that memory usable for
7607 * any allocation type is evenly spread. If both kernelcore
7608 * and movablecore are specified, then the value of kernelcore
7609 * will be used for required_kernelcore if it's greater than
7610 * what movablecore would have allowed.
7612 if (required_movablecore) {
7613 unsigned long corepages;
7616 * Round-up so that ZONE_MOVABLE is at least as large as what
7617 * was requested by the user
7619 required_movablecore =
7620 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7621 required_movablecore = min(totalpages, required_movablecore);
7622 corepages = totalpages - required_movablecore;
7624 required_kernelcore = max(required_kernelcore, corepages);
7628 * If kernelcore was not specified or kernelcore size is larger
7629 * than totalpages, there is no ZONE_MOVABLE.
7631 if (!required_kernelcore || required_kernelcore >= totalpages)
7634 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7635 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7638 /* Spread kernelcore memory as evenly as possible throughout nodes */
7639 kernelcore_node = required_kernelcore / usable_nodes;
7640 for_each_node_state(nid, N_MEMORY) {
7641 unsigned long start_pfn, end_pfn;
7644 * Recalculate kernelcore_node if the division per node
7645 * now exceeds what is necessary to satisfy the requested
7646 * amount of memory for the kernel
7648 if (required_kernelcore < kernelcore_node)
7649 kernelcore_node = required_kernelcore / usable_nodes;
7652 * As the map is walked, we track how much memory is usable
7653 * by the kernel using kernelcore_remaining. When it is
7654 * 0, the rest of the node is usable by ZONE_MOVABLE
7656 kernelcore_remaining = kernelcore_node;
7658 /* Go through each range of PFNs within this node */
7659 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7660 unsigned long size_pages;
7662 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7663 if (start_pfn >= end_pfn)
7666 /* Account for what is only usable for kernelcore */
7667 if (start_pfn < usable_startpfn) {
7668 unsigned long kernel_pages;
7669 kernel_pages = min(end_pfn, usable_startpfn)
7672 kernelcore_remaining -= min(kernel_pages,
7673 kernelcore_remaining);
7674 required_kernelcore -= min(kernel_pages,
7675 required_kernelcore);
7677 /* Continue if range is now fully accounted */
7678 if (end_pfn <= usable_startpfn) {
7681 * Push zone_movable_pfn to the end so
7682 * that if we have to rebalance
7683 * kernelcore across nodes, we will
7684 * not double account here
7686 zone_movable_pfn[nid] = end_pfn;
7689 start_pfn = usable_startpfn;
7693 * The usable PFN range for ZONE_MOVABLE is from
7694 * start_pfn->end_pfn. Calculate size_pages as the
7695 * number of pages used as kernelcore
7697 size_pages = end_pfn - start_pfn;
7698 if (size_pages > kernelcore_remaining)
7699 size_pages = kernelcore_remaining;
7700 zone_movable_pfn[nid] = start_pfn + size_pages;
7703 * Some kernelcore has been met, update counts and
7704 * break if the kernelcore for this node has been
7707 required_kernelcore -= min(required_kernelcore,
7709 kernelcore_remaining -= size_pages;
7710 if (!kernelcore_remaining)
7716 * If there is still required_kernelcore, we do another pass with one
7717 * less node in the count. This will push zone_movable_pfn[nid] further
7718 * along on the nodes that still have memory until kernelcore is
7722 if (usable_nodes && required_kernelcore > usable_nodes)
7726 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7727 for (nid = 0; nid < MAX_NUMNODES; nid++)
7728 zone_movable_pfn[nid] =
7729 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7732 /* restore the node_state */
7733 node_states[N_MEMORY] = saved_node_state;
7736 /* Any regular or high memory on that node ? */
7737 static void check_for_memory(pg_data_t *pgdat, int nid)
7739 enum zone_type zone_type;
7741 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7742 struct zone *zone = &pgdat->node_zones[zone_type];
7743 if (populated_zone(zone)) {
7744 if (IS_ENABLED(CONFIG_HIGHMEM))
7745 node_set_state(nid, N_HIGH_MEMORY);
7746 if (zone_type <= ZONE_NORMAL)
7747 node_set_state(nid, N_NORMAL_MEMORY);
7754 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7755 * such cases we allow max_zone_pfn sorted in the descending order
7757 bool __weak arch_has_descending_max_zone_pfns(void)
7763 * free_area_init - Initialise all pg_data_t and zone data
7764 * @max_zone_pfn: an array of max PFNs for each zone
7766 * This will call free_area_init_node() for each active node in the system.
7767 * Using the page ranges provided by memblock_set_node(), the size of each
7768 * zone in each node and their holes is calculated. If the maximum PFN
7769 * between two adjacent zones match, it is assumed that the zone is empty.
7770 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7771 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7772 * starts where the previous one ended. For example, ZONE_DMA32 starts
7773 * at arch_max_dma_pfn.
7775 void __init free_area_init(unsigned long *max_zone_pfn)
7777 unsigned long start_pfn, end_pfn;
7781 /* Record where the zone boundaries are */
7782 memset(arch_zone_lowest_possible_pfn, 0,
7783 sizeof(arch_zone_lowest_possible_pfn));
7784 memset(arch_zone_highest_possible_pfn, 0,
7785 sizeof(arch_zone_highest_possible_pfn));
7787 start_pfn = find_min_pfn_with_active_regions();
7788 descending = arch_has_descending_max_zone_pfns();
7790 for (i = 0; i < MAX_NR_ZONES; i++) {
7792 zone = MAX_NR_ZONES - i - 1;
7796 if (zone == ZONE_MOVABLE)
7799 end_pfn = max(max_zone_pfn[zone], start_pfn);
7800 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7801 arch_zone_highest_possible_pfn[zone] = end_pfn;
7803 start_pfn = end_pfn;
7806 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7807 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7808 find_zone_movable_pfns_for_nodes();
7810 /* Print out the zone ranges */
7811 pr_info("Zone ranges:\n");
7812 for (i = 0; i < MAX_NR_ZONES; i++) {
7813 if (i == ZONE_MOVABLE)
7815 pr_info(" %-8s ", zone_names[i]);
7816 if (arch_zone_lowest_possible_pfn[i] ==
7817 arch_zone_highest_possible_pfn[i])
7820 pr_cont("[mem %#018Lx-%#018Lx]\n",
7821 (u64)arch_zone_lowest_possible_pfn[i]
7823 ((u64)arch_zone_highest_possible_pfn[i]
7824 << PAGE_SHIFT) - 1);
7827 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7828 pr_info("Movable zone start for each node\n");
7829 for (i = 0; i < MAX_NUMNODES; i++) {
7830 if (zone_movable_pfn[i])
7831 pr_info(" Node %d: %#018Lx\n", i,
7832 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7836 * Print out the early node map, and initialize the
7837 * subsection-map relative to active online memory ranges to
7838 * enable future "sub-section" extensions of the memory map.
7840 pr_info("Early memory node ranges\n");
7841 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7842 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7843 (u64)start_pfn << PAGE_SHIFT,
7844 ((u64)end_pfn << PAGE_SHIFT) - 1);
7845 subsection_map_init(start_pfn, end_pfn - start_pfn);
7848 /* Initialise every node */
7849 mminit_verify_pageflags_layout();
7850 setup_nr_node_ids();
7851 for_each_online_node(nid) {
7852 pg_data_t *pgdat = NODE_DATA(nid);
7853 free_area_init_node(nid);
7855 /* Any memory on that node */
7856 if (pgdat->node_present_pages)
7857 node_set_state(nid, N_MEMORY);
7858 check_for_memory(pgdat, nid);
7864 static int __init cmdline_parse_core(char *p, unsigned long *core,
7865 unsigned long *percent)
7867 unsigned long long coremem;
7873 /* Value may be a percentage of total memory, otherwise bytes */
7874 coremem = simple_strtoull(p, &endptr, 0);
7875 if (*endptr == '%') {
7876 /* Paranoid check for percent values greater than 100 */
7877 WARN_ON(coremem > 100);
7881 coremem = memparse(p, &p);
7882 /* Paranoid check that UL is enough for the coremem value */
7883 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7885 *core = coremem >> PAGE_SHIFT;
7892 * kernelcore=size sets the amount of memory for use for allocations that
7893 * cannot be reclaimed or migrated.
7895 static int __init cmdline_parse_kernelcore(char *p)
7897 /* parse kernelcore=mirror */
7898 if (parse_option_str(p, "mirror")) {
7899 mirrored_kernelcore = true;
7903 return cmdline_parse_core(p, &required_kernelcore,
7904 &required_kernelcore_percent);
7908 * movablecore=size sets the amount of memory for use for allocations that
7909 * can be reclaimed or migrated.
7911 static int __init cmdline_parse_movablecore(char *p)
7913 return cmdline_parse_core(p, &required_movablecore,
7914 &required_movablecore_percent);
7917 early_param("kernelcore", cmdline_parse_kernelcore);
7918 early_param("movablecore", cmdline_parse_movablecore);
7920 void adjust_managed_page_count(struct page *page, long count)
7922 atomic_long_add(count, &page_zone(page)->managed_pages);
7923 totalram_pages_add(count);
7924 #ifdef CONFIG_HIGHMEM
7925 if (PageHighMem(page))
7926 totalhigh_pages_add(count);
7929 EXPORT_SYMBOL(adjust_managed_page_count);
7931 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7934 unsigned long pages = 0;
7936 start = (void *)PAGE_ALIGN((unsigned long)start);
7937 end = (void *)((unsigned long)end & PAGE_MASK);
7938 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7939 struct page *page = virt_to_page(pos);
7940 void *direct_map_addr;
7943 * 'direct_map_addr' might be different from 'pos'
7944 * because some architectures' virt_to_page()
7945 * work with aliases. Getting the direct map
7946 * address ensures that we get a _writeable_
7947 * alias for the memset().
7949 direct_map_addr = page_address(page);
7951 * Perform a kasan-unchecked memset() since this memory
7952 * has not been initialized.
7954 direct_map_addr = kasan_reset_tag(direct_map_addr);
7955 if ((unsigned int)poison <= 0xFF)
7956 memset(direct_map_addr, poison, PAGE_SIZE);
7958 free_reserved_page(page);
7962 pr_info("Freeing %s memory: %ldK\n",
7963 s, pages << (PAGE_SHIFT - 10));
7968 void __init mem_init_print_info(void)
7970 unsigned long physpages, codesize, datasize, rosize, bss_size;
7971 unsigned long init_code_size, init_data_size;
7973 physpages = get_num_physpages();
7974 codesize = _etext - _stext;
7975 datasize = _edata - _sdata;
7976 rosize = __end_rodata - __start_rodata;
7977 bss_size = __bss_stop - __bss_start;
7978 init_data_size = __init_end - __init_begin;
7979 init_code_size = _einittext - _sinittext;
7982 * Detect special cases and adjust section sizes accordingly:
7983 * 1) .init.* may be embedded into .data sections
7984 * 2) .init.text.* may be out of [__init_begin, __init_end],
7985 * please refer to arch/tile/kernel/vmlinux.lds.S.
7986 * 3) .rodata.* may be embedded into .text or .data sections.
7988 #define adj_init_size(start, end, size, pos, adj) \
7990 if (start <= pos && pos < end && size > adj) \
7994 adj_init_size(__init_begin, __init_end, init_data_size,
7995 _sinittext, init_code_size);
7996 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7997 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7998 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7999 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8001 #undef adj_init_size
8003 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8004 #ifdef CONFIG_HIGHMEM
8008 nr_free_pages() << (PAGE_SHIFT - 10),
8009 physpages << (PAGE_SHIFT - 10),
8010 codesize >> 10, datasize >> 10, rosize >> 10,
8011 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8012 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
8013 totalcma_pages << (PAGE_SHIFT - 10)
8014 #ifdef CONFIG_HIGHMEM
8015 , totalhigh_pages() << (PAGE_SHIFT - 10)
8021 * set_dma_reserve - set the specified number of pages reserved in the first zone
8022 * @new_dma_reserve: The number of pages to mark reserved
8024 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8025 * In the DMA zone, a significant percentage may be consumed by kernel image
8026 * and other unfreeable allocations which can skew the watermarks badly. This
8027 * function may optionally be used to account for unfreeable pages in the
8028 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8029 * smaller per-cpu batchsize.
8031 void __init set_dma_reserve(unsigned long new_dma_reserve)
8033 dma_reserve = new_dma_reserve;
8036 static int page_alloc_cpu_dead(unsigned int cpu)
8039 lru_add_drain_cpu(cpu);
8043 * Spill the event counters of the dead processor
8044 * into the current processors event counters.
8045 * This artificially elevates the count of the current
8048 vm_events_fold_cpu(cpu);
8051 * Zero the differential counters of the dead processor
8052 * so that the vm statistics are consistent.
8054 * This is only okay since the processor is dead and cannot
8055 * race with what we are doing.
8057 cpu_vm_stats_fold(cpu);
8062 int hashdist = HASHDIST_DEFAULT;
8064 static int __init set_hashdist(char *str)
8068 hashdist = simple_strtoul(str, &str, 0);
8071 __setup("hashdist=", set_hashdist);
8074 void __init page_alloc_init(void)
8079 if (num_node_state(N_MEMORY) == 1)
8083 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
8084 "mm/page_alloc:dead", NULL,
8085 page_alloc_cpu_dead);
8090 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8091 * or min_free_kbytes changes.
8093 static void calculate_totalreserve_pages(void)
8095 struct pglist_data *pgdat;
8096 unsigned long reserve_pages = 0;
8097 enum zone_type i, j;
8099 for_each_online_pgdat(pgdat) {
8101 pgdat->totalreserve_pages = 0;
8103 for (i = 0; i < MAX_NR_ZONES; i++) {
8104 struct zone *zone = pgdat->node_zones + i;
8106 unsigned long managed_pages = zone_managed_pages(zone);
8108 /* Find valid and maximum lowmem_reserve in the zone */
8109 for (j = i; j < MAX_NR_ZONES; j++) {
8110 if (zone->lowmem_reserve[j] > max)
8111 max = zone->lowmem_reserve[j];
8114 /* we treat the high watermark as reserved pages. */
8115 max += high_wmark_pages(zone);
8117 if (max > managed_pages)
8118 max = managed_pages;
8120 pgdat->totalreserve_pages += max;
8122 reserve_pages += max;
8125 totalreserve_pages = reserve_pages;
8129 * setup_per_zone_lowmem_reserve - called whenever
8130 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8131 * has a correct pages reserved value, so an adequate number of
8132 * pages are left in the zone after a successful __alloc_pages().
8134 static void setup_per_zone_lowmem_reserve(void)
8136 struct pglist_data *pgdat;
8137 enum zone_type i, j;
8139 for_each_online_pgdat(pgdat) {
8140 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8141 struct zone *zone = &pgdat->node_zones[i];
8142 int ratio = sysctl_lowmem_reserve_ratio[i];
8143 bool clear = !ratio || !zone_managed_pages(zone);
8144 unsigned long managed_pages = 0;
8146 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8148 zone->lowmem_reserve[j] = 0;
8150 struct zone *upper_zone = &pgdat->node_zones[j];
8152 managed_pages += zone_managed_pages(upper_zone);
8153 zone->lowmem_reserve[j] = managed_pages / ratio;
8159 /* update totalreserve_pages */
8160 calculate_totalreserve_pages();
8163 static void __setup_per_zone_wmarks(void)
8165 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8166 unsigned long lowmem_pages = 0;
8168 unsigned long flags;
8170 /* Calculate total number of !ZONE_HIGHMEM pages */
8171 for_each_zone(zone) {
8172 if (!is_highmem(zone))
8173 lowmem_pages += zone_managed_pages(zone);
8176 for_each_zone(zone) {
8179 spin_lock_irqsave(&zone->lock, flags);
8180 tmp = (u64)pages_min * zone_managed_pages(zone);
8181 do_div(tmp, lowmem_pages);
8182 if (is_highmem(zone)) {
8184 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8185 * need highmem pages, so cap pages_min to a small
8188 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8189 * deltas control async page reclaim, and so should
8190 * not be capped for highmem.
8192 unsigned long min_pages;
8194 min_pages = zone_managed_pages(zone) / 1024;
8195 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8196 zone->_watermark[WMARK_MIN] = min_pages;
8199 * If it's a lowmem zone, reserve a number of pages
8200 * proportionate to the zone's size.
8202 zone->_watermark[WMARK_MIN] = tmp;
8206 * Set the kswapd watermarks distance according to the
8207 * scale factor in proportion to available memory, but
8208 * ensure a minimum size on small systems.
8210 tmp = max_t(u64, tmp >> 2,
8211 mult_frac(zone_managed_pages(zone),
8212 watermark_scale_factor, 10000));
8214 zone->watermark_boost = 0;
8215 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8216 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8218 spin_unlock_irqrestore(&zone->lock, flags);
8221 /* update totalreserve_pages */
8222 calculate_totalreserve_pages();
8226 * setup_per_zone_wmarks - called when min_free_kbytes changes
8227 * or when memory is hot-{added|removed}
8229 * Ensures that the watermark[min,low,high] values for each zone are set
8230 * correctly with respect to min_free_kbytes.
8232 void setup_per_zone_wmarks(void)
8234 static DEFINE_SPINLOCK(lock);
8237 __setup_per_zone_wmarks();
8242 * Initialise min_free_kbytes.
8244 * For small machines we want it small (128k min). For large machines
8245 * we want it large (256MB max). But it is not linear, because network
8246 * bandwidth does not increase linearly with machine size. We use
8248 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8249 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8265 int __meminit init_per_zone_wmark_min(void)
8267 unsigned long lowmem_kbytes;
8268 int new_min_free_kbytes;
8270 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8271 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8273 if (new_min_free_kbytes > user_min_free_kbytes) {
8274 min_free_kbytes = new_min_free_kbytes;
8275 if (min_free_kbytes < 128)
8276 min_free_kbytes = 128;
8277 if (min_free_kbytes > 262144)
8278 min_free_kbytes = 262144;
8280 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8281 new_min_free_kbytes, user_min_free_kbytes);
8283 setup_per_zone_wmarks();
8284 refresh_zone_stat_thresholds();
8285 setup_per_zone_lowmem_reserve();
8288 setup_min_unmapped_ratio();
8289 setup_min_slab_ratio();
8292 khugepaged_min_free_kbytes_update();
8296 postcore_initcall(init_per_zone_wmark_min)
8299 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8300 * that we can call two helper functions whenever min_free_kbytes
8303 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8304 void *buffer, size_t *length, loff_t *ppos)
8308 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8313 user_min_free_kbytes = min_free_kbytes;
8314 setup_per_zone_wmarks();
8319 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8320 void *buffer, size_t *length, loff_t *ppos)
8324 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8329 setup_per_zone_wmarks();
8335 static void setup_min_unmapped_ratio(void)
8340 for_each_online_pgdat(pgdat)
8341 pgdat->min_unmapped_pages = 0;
8344 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8345 sysctl_min_unmapped_ratio) / 100;
8349 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8350 void *buffer, size_t *length, loff_t *ppos)
8354 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8358 setup_min_unmapped_ratio();
8363 static void setup_min_slab_ratio(void)
8368 for_each_online_pgdat(pgdat)
8369 pgdat->min_slab_pages = 0;
8372 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8373 sysctl_min_slab_ratio) / 100;
8376 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8377 void *buffer, size_t *length, loff_t *ppos)
8381 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8385 setup_min_slab_ratio();
8392 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8393 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8394 * whenever sysctl_lowmem_reserve_ratio changes.
8396 * The reserve ratio obviously has absolutely no relation with the
8397 * minimum watermarks. The lowmem reserve ratio can only make sense
8398 * if in function of the boot time zone sizes.
8400 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8401 void *buffer, size_t *length, loff_t *ppos)
8405 proc_dointvec_minmax(table, write, buffer, length, ppos);
8407 for (i = 0; i < MAX_NR_ZONES; i++) {
8408 if (sysctl_lowmem_reserve_ratio[i] < 1)
8409 sysctl_lowmem_reserve_ratio[i] = 0;
8412 setup_per_zone_lowmem_reserve();
8417 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8418 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8419 * pagelist can have before it gets flushed back to buddy allocator.
8421 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8422 void *buffer, size_t *length, loff_t *ppos)
8425 int old_percpu_pagelist_fraction;
8428 mutex_lock(&pcp_batch_high_lock);
8429 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8431 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8432 if (!write || ret < 0)
8435 /* Sanity checking to avoid pcp imbalance */
8436 if (percpu_pagelist_fraction &&
8437 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8438 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8444 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8447 for_each_populated_zone(zone)
8448 zone_set_pageset_high_and_batch(zone);
8450 mutex_unlock(&pcp_batch_high_lock);
8454 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8456 * Returns the number of pages that arch has reserved but
8457 * is not known to alloc_large_system_hash().
8459 static unsigned long __init arch_reserved_kernel_pages(void)
8466 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8467 * machines. As memory size is increased the scale is also increased but at
8468 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8469 * quadruples the scale is increased by one, which means the size of hash table
8470 * only doubles, instead of quadrupling as well.
8471 * Because 32-bit systems cannot have large physical memory, where this scaling
8472 * makes sense, it is disabled on such platforms.
8474 #if __BITS_PER_LONG > 32
8475 #define ADAPT_SCALE_BASE (64ul << 30)
8476 #define ADAPT_SCALE_SHIFT 2
8477 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8481 * allocate a large system hash table from bootmem
8482 * - it is assumed that the hash table must contain an exact power-of-2
8483 * quantity of entries
8484 * - limit is the number of hash buckets, not the total allocation size
8486 void *__init alloc_large_system_hash(const char *tablename,
8487 unsigned long bucketsize,
8488 unsigned long numentries,
8491 unsigned int *_hash_shift,
8492 unsigned int *_hash_mask,
8493 unsigned long low_limit,
8494 unsigned long high_limit)
8496 unsigned long long max = high_limit;
8497 unsigned long log2qty, size;
8503 /* allow the kernel cmdline to have a say */
8505 /* round applicable memory size up to nearest megabyte */
8506 numentries = nr_kernel_pages;
8507 numentries -= arch_reserved_kernel_pages();
8509 /* It isn't necessary when PAGE_SIZE >= 1MB */
8510 if (PAGE_SHIFT < 20)
8511 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8513 #if __BITS_PER_LONG > 32
8515 unsigned long adapt;
8517 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8518 adapt <<= ADAPT_SCALE_SHIFT)
8523 /* limit to 1 bucket per 2^scale bytes of low memory */
8524 if (scale > PAGE_SHIFT)
8525 numentries >>= (scale - PAGE_SHIFT);
8527 numentries <<= (PAGE_SHIFT - scale);
8529 /* Make sure we've got at least a 0-order allocation.. */
8530 if (unlikely(flags & HASH_SMALL)) {
8531 /* Makes no sense without HASH_EARLY */
8532 WARN_ON(!(flags & HASH_EARLY));
8533 if (!(numentries >> *_hash_shift)) {
8534 numentries = 1UL << *_hash_shift;
8535 BUG_ON(!numentries);
8537 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8538 numentries = PAGE_SIZE / bucketsize;
8540 numentries = roundup_pow_of_two(numentries);
8542 /* limit allocation size to 1/16 total memory by default */
8544 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8545 do_div(max, bucketsize);
8547 max = min(max, 0x80000000ULL);
8549 if (numentries < low_limit)
8550 numentries = low_limit;
8551 if (numentries > max)
8554 log2qty = ilog2(numentries);
8556 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8559 size = bucketsize << log2qty;
8560 if (flags & HASH_EARLY) {
8561 if (flags & HASH_ZERO)
8562 table = memblock_alloc(size, SMP_CACHE_BYTES);
8564 table = memblock_alloc_raw(size,
8566 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8567 table = __vmalloc(size, gfp_flags);
8569 huge = is_vm_area_hugepages(table);
8572 * If bucketsize is not a power-of-two, we may free
8573 * some pages at the end of hash table which
8574 * alloc_pages_exact() automatically does
8576 table = alloc_pages_exact(size, gfp_flags);
8577 kmemleak_alloc(table, size, 1, gfp_flags);
8579 } while (!table && size > PAGE_SIZE && --log2qty);
8582 panic("Failed to allocate %s hash table\n", tablename);
8584 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8585 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8586 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8589 *_hash_shift = log2qty;
8591 *_hash_mask = (1 << log2qty) - 1;
8597 * This function checks whether pageblock includes unmovable pages or not.
8599 * PageLRU check without isolation or lru_lock could race so that
8600 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8601 * check without lock_page also may miss some movable non-lru pages at
8602 * race condition. So you can't expect this function should be exact.
8604 * Returns a page without holding a reference. If the caller wants to
8605 * dereference that page (e.g., dumping), it has to make sure that it
8606 * cannot get removed (e.g., via memory unplug) concurrently.
8609 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8610 int migratetype, int flags)
8612 unsigned long iter = 0;
8613 unsigned long pfn = page_to_pfn(page);
8614 unsigned long offset = pfn % pageblock_nr_pages;
8616 if (is_migrate_cma_page(page)) {
8618 * CMA allocations (alloc_contig_range) really need to mark
8619 * isolate CMA pageblocks even when they are not movable in fact
8620 * so consider them movable here.
8622 if (is_migrate_cma(migratetype))
8628 for (; iter < pageblock_nr_pages - offset; iter++) {
8629 if (!pfn_valid_within(pfn + iter))
8632 page = pfn_to_page(pfn + iter);
8635 * Both, bootmem allocations and memory holes are marked
8636 * PG_reserved and are unmovable. We can even have unmovable
8637 * allocations inside ZONE_MOVABLE, for example when
8638 * specifying "movablecore".
8640 if (PageReserved(page))
8644 * If the zone is movable and we have ruled out all reserved
8645 * pages then it should be reasonably safe to assume the rest
8648 if (zone_idx(zone) == ZONE_MOVABLE)
8652 * Hugepages are not in LRU lists, but they're movable.
8653 * THPs are on the LRU, but need to be counted as #small pages.
8654 * We need not scan over tail pages because we don't
8655 * handle each tail page individually in migration.
8657 if (PageHuge(page) || PageTransCompound(page)) {
8658 struct page *head = compound_head(page);
8659 unsigned int skip_pages;
8661 if (PageHuge(page)) {
8662 if (!hugepage_migration_supported(page_hstate(head)))
8664 } else if (!PageLRU(head) && !__PageMovable(head)) {
8668 skip_pages = compound_nr(head) - (page - head);
8669 iter += skip_pages - 1;
8674 * We can't use page_count without pin a page
8675 * because another CPU can free compound page.
8676 * This check already skips compound tails of THP
8677 * because their page->_refcount is zero at all time.
8679 if (!page_ref_count(page)) {
8680 if (PageBuddy(page))
8681 iter += (1 << buddy_order(page)) - 1;
8686 * The HWPoisoned page may be not in buddy system, and
8687 * page_count() is not 0.
8689 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8693 * We treat all PageOffline() pages as movable when offlining
8694 * to give drivers a chance to decrement their reference count
8695 * in MEM_GOING_OFFLINE in order to indicate that these pages
8696 * can be offlined as there are no direct references anymore.
8697 * For actually unmovable PageOffline() where the driver does
8698 * not support this, we will fail later when trying to actually
8699 * move these pages that still have a reference count > 0.
8700 * (false negatives in this function only)
8702 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8705 if (__PageMovable(page) || PageLRU(page))
8709 * If there are RECLAIMABLE pages, we need to check
8710 * it. But now, memory offline itself doesn't call
8711 * shrink_node_slabs() and it still to be fixed.
8718 #ifdef CONFIG_CONTIG_ALLOC
8719 static unsigned long pfn_max_align_down(unsigned long pfn)
8721 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8722 pageblock_nr_pages) - 1);
8725 static unsigned long pfn_max_align_up(unsigned long pfn)
8727 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8728 pageblock_nr_pages));
8731 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8732 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8733 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8734 static void alloc_contig_dump_pages(struct list_head *page_list)
8736 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8738 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8742 list_for_each_entry(page, page_list, lru)
8743 dump_page(page, "migration failure");
8747 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8752 /* [start, end) must belong to a single zone. */
8753 static int __alloc_contig_migrate_range(struct compact_control *cc,
8754 unsigned long start, unsigned long end)
8756 /* This function is based on compact_zone() from compaction.c. */
8757 unsigned int nr_reclaimed;
8758 unsigned long pfn = start;
8759 unsigned int tries = 0;
8761 struct migration_target_control mtc = {
8762 .nid = zone_to_nid(cc->zone),
8763 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8766 lru_cache_disable();
8768 while (pfn < end || !list_empty(&cc->migratepages)) {
8769 if (fatal_signal_pending(current)) {
8774 if (list_empty(&cc->migratepages)) {
8775 cc->nr_migratepages = 0;
8776 ret = isolate_migratepages_range(cc, pfn, end);
8777 if (ret && ret != -EAGAIN)
8779 pfn = cc->migrate_pfn;
8781 } else if (++tries == 5) {
8786 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8788 cc->nr_migratepages -= nr_reclaimed;
8790 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8791 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8794 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
8795 * to retry again over this error, so do the same here.
8804 alloc_contig_dump_pages(&cc->migratepages);
8805 putback_movable_pages(&cc->migratepages);
8812 * alloc_contig_range() -- tries to allocate given range of pages
8813 * @start: start PFN to allocate
8814 * @end: one-past-the-last PFN to allocate
8815 * @migratetype: migratetype of the underlying pageblocks (either
8816 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8817 * in range must have the same migratetype and it must
8818 * be either of the two.
8819 * @gfp_mask: GFP mask to use during compaction
8821 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8822 * aligned. The PFN range must belong to a single zone.
8824 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8825 * pageblocks in the range. Once isolated, the pageblocks should not
8826 * be modified by others.
8828 * Return: zero on success or negative error code. On success all
8829 * pages which PFN is in [start, end) are allocated for the caller and
8830 * need to be freed with free_contig_range().
8832 int alloc_contig_range(unsigned long start, unsigned long end,
8833 unsigned migratetype, gfp_t gfp_mask)
8835 unsigned long outer_start, outer_end;
8839 struct compact_control cc = {
8840 .nr_migratepages = 0,
8842 .zone = page_zone(pfn_to_page(start)),
8843 .mode = MIGRATE_SYNC,
8844 .ignore_skip_hint = true,
8845 .no_set_skip_hint = true,
8846 .gfp_mask = current_gfp_context(gfp_mask),
8847 .alloc_contig = true,
8849 INIT_LIST_HEAD(&cc.migratepages);
8852 * What we do here is we mark all pageblocks in range as
8853 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8854 * have different sizes, and due to the way page allocator
8855 * work, we align the range to biggest of the two pages so
8856 * that page allocator won't try to merge buddies from
8857 * different pageblocks and change MIGRATE_ISOLATE to some
8858 * other migration type.
8860 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8861 * migrate the pages from an unaligned range (ie. pages that
8862 * we are interested in). This will put all the pages in
8863 * range back to page allocator as MIGRATE_ISOLATE.
8865 * When this is done, we take the pages in range from page
8866 * allocator removing them from the buddy system. This way
8867 * page allocator will never consider using them.
8869 * This lets us mark the pageblocks back as
8870 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8871 * aligned range but not in the unaligned, original range are
8872 * put back to page allocator so that buddy can use them.
8875 ret = start_isolate_page_range(pfn_max_align_down(start),
8876 pfn_max_align_up(end), migratetype, 0);
8880 drain_all_pages(cc.zone);
8883 * In case of -EBUSY, we'd like to know which page causes problem.
8884 * So, just fall through. test_pages_isolated() has a tracepoint
8885 * which will report the busy page.
8887 * It is possible that busy pages could become available before
8888 * the call to test_pages_isolated, and the range will actually be
8889 * allocated. So, if we fall through be sure to clear ret so that
8890 * -EBUSY is not accidentally used or returned to caller.
8892 ret = __alloc_contig_migrate_range(&cc, start, end);
8893 if (ret && ret != -EBUSY)
8898 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8899 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8900 * more, all pages in [start, end) are free in page allocator.
8901 * What we are going to do is to allocate all pages from
8902 * [start, end) (that is remove them from page allocator).
8904 * The only problem is that pages at the beginning and at the
8905 * end of interesting range may be not aligned with pages that
8906 * page allocator holds, ie. they can be part of higher order
8907 * pages. Because of this, we reserve the bigger range and
8908 * once this is done free the pages we are not interested in.
8910 * We don't have to hold zone->lock here because the pages are
8911 * isolated thus they won't get removed from buddy.
8915 outer_start = start;
8916 while (!PageBuddy(pfn_to_page(outer_start))) {
8917 if (++order >= MAX_ORDER) {
8918 outer_start = start;
8921 outer_start &= ~0UL << order;
8924 if (outer_start != start) {
8925 order = buddy_order(pfn_to_page(outer_start));
8928 * outer_start page could be small order buddy page and
8929 * it doesn't include start page. Adjust outer_start
8930 * in this case to report failed page properly
8931 * on tracepoint in test_pages_isolated()
8933 if (outer_start + (1UL << order) <= start)
8934 outer_start = start;
8937 /* Make sure the range is really isolated. */
8938 if (test_pages_isolated(outer_start, end, 0)) {
8943 /* Grab isolated pages from freelists. */
8944 outer_end = isolate_freepages_range(&cc, outer_start, end);
8950 /* Free head and tail (if any) */
8951 if (start != outer_start)
8952 free_contig_range(outer_start, start - outer_start);
8953 if (end != outer_end)
8954 free_contig_range(end, outer_end - end);
8957 undo_isolate_page_range(pfn_max_align_down(start),
8958 pfn_max_align_up(end), migratetype);
8961 EXPORT_SYMBOL(alloc_contig_range);
8963 static int __alloc_contig_pages(unsigned long start_pfn,
8964 unsigned long nr_pages, gfp_t gfp_mask)
8966 unsigned long end_pfn = start_pfn + nr_pages;
8968 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8972 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8973 unsigned long nr_pages)
8975 unsigned long i, end_pfn = start_pfn + nr_pages;
8978 for (i = start_pfn; i < end_pfn; i++) {
8979 page = pfn_to_online_page(i);
8983 if (page_zone(page) != z)
8986 if (PageReserved(page))
8992 static bool zone_spans_last_pfn(const struct zone *zone,
8993 unsigned long start_pfn, unsigned long nr_pages)
8995 unsigned long last_pfn = start_pfn + nr_pages - 1;
8997 return zone_spans_pfn(zone, last_pfn);
9001 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9002 * @nr_pages: Number of contiguous pages to allocate
9003 * @gfp_mask: GFP mask to limit search and used during compaction
9005 * @nodemask: Mask for other possible nodes
9007 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9008 * on an applicable zonelist to find a contiguous pfn range which can then be
9009 * tried for allocation with alloc_contig_range(). This routine is intended
9010 * for allocation requests which can not be fulfilled with the buddy allocator.
9012 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9013 * power of two then the alignment is guaranteed to be to the given nr_pages
9014 * (e.g. 1GB request would be aligned to 1GB).
9016 * Allocated pages can be freed with free_contig_range() or by manually calling
9017 * __free_page() on each allocated page.
9019 * Return: pointer to contiguous pages on success, or NULL if not successful.
9021 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9022 int nid, nodemask_t *nodemask)
9024 unsigned long ret, pfn, flags;
9025 struct zonelist *zonelist;
9029 zonelist = node_zonelist(nid, gfp_mask);
9030 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9031 gfp_zone(gfp_mask), nodemask) {
9032 spin_lock_irqsave(&zone->lock, flags);
9034 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9035 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9036 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9038 * We release the zone lock here because
9039 * alloc_contig_range() will also lock the zone
9040 * at some point. If there's an allocation
9041 * spinning on this lock, it may win the race
9042 * and cause alloc_contig_range() to fail...
9044 spin_unlock_irqrestore(&zone->lock, flags);
9045 ret = __alloc_contig_pages(pfn, nr_pages,
9048 return pfn_to_page(pfn);
9049 spin_lock_irqsave(&zone->lock, flags);
9053 spin_unlock_irqrestore(&zone->lock, flags);
9057 #endif /* CONFIG_CONTIG_ALLOC */
9059 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9061 unsigned long count = 0;
9063 for (; nr_pages--; pfn++) {
9064 struct page *page = pfn_to_page(pfn);
9066 count += page_count(page) != 1;
9069 WARN(count != 0, "%lu pages are still in use!\n", count);
9071 EXPORT_SYMBOL(free_contig_range);
9074 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9075 * page high values need to be recalculated.
9077 void __meminit zone_pcp_update(struct zone *zone)
9079 mutex_lock(&pcp_batch_high_lock);
9080 zone_set_pageset_high_and_batch(zone);
9081 mutex_unlock(&pcp_batch_high_lock);
9085 * Effectively disable pcplists for the zone by setting the high limit to 0
9086 * and draining all cpus. A concurrent page freeing on another CPU that's about
9087 * to put the page on pcplist will either finish before the drain and the page
9088 * will be drained, or observe the new high limit and skip the pcplist.
9090 * Must be paired with a call to zone_pcp_enable().
9092 void zone_pcp_disable(struct zone *zone)
9094 mutex_lock(&pcp_batch_high_lock);
9095 __zone_set_pageset_high_and_batch(zone, 0, 1);
9096 __drain_all_pages(zone, true);
9099 void zone_pcp_enable(struct zone *zone)
9101 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9102 mutex_unlock(&pcp_batch_high_lock);
9105 void zone_pcp_reset(struct zone *zone)
9108 struct per_cpu_zonestat *pzstats;
9110 if (zone->per_cpu_pageset != &boot_pageset) {
9111 for_each_online_cpu(cpu) {
9112 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9113 drain_zonestat(zone, pzstats);
9115 free_percpu(zone->per_cpu_pageset);
9116 free_percpu(zone->per_cpu_zonestats);
9117 zone->per_cpu_pageset = &boot_pageset;
9118 zone->per_cpu_zonestats = &boot_zonestats;
9122 #ifdef CONFIG_MEMORY_HOTREMOVE
9124 * All pages in the range must be in a single zone, must not contain holes,
9125 * must span full sections, and must be isolated before calling this function.
9127 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9129 unsigned long pfn = start_pfn;
9133 unsigned long flags;
9135 offline_mem_sections(pfn, end_pfn);
9136 zone = page_zone(pfn_to_page(pfn));
9137 spin_lock_irqsave(&zone->lock, flags);
9138 while (pfn < end_pfn) {
9139 page = pfn_to_page(pfn);
9141 * The HWPoisoned page may be not in buddy system, and
9142 * page_count() is not 0.
9144 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9149 * At this point all remaining PageOffline() pages have a
9150 * reference count of 0 and can simply be skipped.
9152 if (PageOffline(page)) {
9153 BUG_ON(page_count(page));
9154 BUG_ON(PageBuddy(page));
9159 BUG_ON(page_count(page));
9160 BUG_ON(!PageBuddy(page));
9161 order = buddy_order(page);
9162 del_page_from_free_list(page, zone, order);
9163 pfn += (1 << order);
9165 spin_unlock_irqrestore(&zone->lock, flags);
9169 bool is_free_buddy_page(struct page *page)
9171 struct zone *zone = page_zone(page);
9172 unsigned long pfn = page_to_pfn(page);
9173 unsigned long flags;
9176 spin_lock_irqsave(&zone->lock, flags);
9177 for (order = 0; order < MAX_ORDER; order++) {
9178 struct page *page_head = page - (pfn & ((1 << order) - 1));
9180 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
9183 spin_unlock_irqrestore(&zone->lock, flags);
9185 return order < MAX_ORDER;
9188 #ifdef CONFIG_MEMORY_FAILURE
9190 * Break down a higher-order page in sub-pages, and keep our target out of
9193 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9194 struct page *target, int low, int high,
9197 unsigned long size = 1 << high;
9198 struct page *current_buddy, *next_page;
9200 while (high > low) {
9204 if (target >= &page[size]) {
9205 next_page = page + size;
9206 current_buddy = page;
9209 current_buddy = page + size;
9212 if (set_page_guard(zone, current_buddy, high, migratetype))
9215 if (current_buddy != target) {
9216 add_to_free_list(current_buddy, zone, high, migratetype);
9217 set_buddy_order(current_buddy, high);
9224 * Take a page that will be marked as poisoned off the buddy allocator.
9226 bool take_page_off_buddy(struct page *page)
9228 struct zone *zone = page_zone(page);
9229 unsigned long pfn = page_to_pfn(page);
9230 unsigned long flags;
9234 spin_lock_irqsave(&zone->lock, flags);
9235 for (order = 0; order < MAX_ORDER; order++) {
9236 struct page *page_head = page - (pfn & ((1 << order) - 1));
9237 int page_order = buddy_order(page_head);
9239 if (PageBuddy(page_head) && page_order >= order) {
9240 unsigned long pfn_head = page_to_pfn(page_head);
9241 int migratetype = get_pfnblock_migratetype(page_head,
9244 del_page_from_free_list(page_head, zone, page_order);
9245 break_down_buddy_pages(zone, page_head, page, 0,
9246 page_order, migratetype);
9247 if (!is_migrate_isolate(migratetype))
9248 __mod_zone_freepage_state(zone, -1, migratetype);
9252 if (page_count(page_head) > 0)
9255 spin_unlock_irqrestore(&zone->lock, flags);