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 spin_lock(&zone->lock);
1505 if (unlikely(has_isolate_pageblock(zone) ||
1506 is_migrate_isolate(migratetype))) {
1507 migratetype = get_pfnblock_migratetype(page, pfn);
1509 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1510 spin_unlock(&zone->lock);
1513 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1514 unsigned long zone, int nid)
1516 mm_zero_struct_page(page);
1517 set_page_links(page, zone, nid, pfn);
1518 init_page_count(page);
1519 page_mapcount_reset(page);
1520 page_cpupid_reset_last(page);
1521 page_kasan_tag_reset(page);
1523 INIT_LIST_HEAD(&page->lru);
1524 #ifdef WANT_PAGE_VIRTUAL
1525 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1526 if (!is_highmem_idx(zone))
1527 set_page_address(page, __va(pfn << PAGE_SHIFT));
1531 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1532 static void __meminit init_reserved_page(unsigned long pfn)
1537 if (!early_page_uninitialised(pfn))
1540 nid = early_pfn_to_nid(pfn);
1541 pgdat = NODE_DATA(nid);
1543 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1544 struct zone *zone = &pgdat->node_zones[zid];
1546 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1549 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1552 static inline void init_reserved_page(unsigned long pfn)
1555 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1558 * Initialised pages do not have PageReserved set. This function is
1559 * called for each range allocated by the bootmem allocator and
1560 * marks the pages PageReserved. The remaining valid pages are later
1561 * sent to the buddy page allocator.
1563 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1565 unsigned long start_pfn = PFN_DOWN(start);
1566 unsigned long end_pfn = PFN_UP(end);
1568 for (; start_pfn < end_pfn; start_pfn++) {
1569 if (pfn_valid(start_pfn)) {
1570 struct page *page = pfn_to_page(start_pfn);
1572 init_reserved_page(start_pfn);
1574 /* Avoid false-positive PageTail() */
1575 INIT_LIST_HEAD(&page->lru);
1578 * no need for atomic set_bit because the struct
1579 * page is not visible yet so nobody should
1582 __SetPageReserved(page);
1587 static void __free_pages_ok(struct page *page, unsigned int order,
1590 unsigned long flags;
1592 unsigned long pfn = page_to_pfn(page);
1593 struct zone *zone = page_zone(page);
1595 if (!free_pages_prepare(page, order, true, fpi_flags))
1598 migratetype = get_pfnblock_migratetype(page, pfn);
1600 spin_lock_irqsave(&zone->lock, flags);
1601 __count_vm_events(PGFREE, 1 << order);
1602 if (unlikely(has_isolate_pageblock(zone) ||
1603 is_migrate_isolate(migratetype))) {
1604 migratetype = get_pfnblock_migratetype(page, pfn);
1606 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1607 spin_unlock_irqrestore(&zone->lock, flags);
1610 void __free_pages_core(struct page *page, unsigned int order)
1612 unsigned int nr_pages = 1 << order;
1613 struct page *p = page;
1617 * When initializing the memmap, __init_single_page() sets the refcount
1618 * of all pages to 1 ("allocated"/"not free"). We have to set the
1619 * refcount of all involved pages to 0.
1622 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1624 __ClearPageReserved(p);
1625 set_page_count(p, 0);
1627 __ClearPageReserved(p);
1628 set_page_count(p, 0);
1630 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1633 * Bypass PCP and place fresh pages right to the tail, primarily
1634 * relevant for memory onlining.
1636 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1639 #ifdef CONFIG_NEED_MULTIPLE_NODES
1642 * During memory init memblocks map pfns to nids. The search is expensive and
1643 * this caches recent lookups. The implementation of __early_pfn_to_nid
1644 * treats start/end as pfns.
1646 struct mminit_pfnnid_cache {
1647 unsigned long last_start;
1648 unsigned long last_end;
1652 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1655 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1657 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1658 struct mminit_pfnnid_cache *state)
1660 unsigned long start_pfn, end_pfn;
1663 if (state->last_start <= pfn && pfn < state->last_end)
1664 return state->last_nid;
1666 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1667 if (nid != NUMA_NO_NODE) {
1668 state->last_start = start_pfn;
1669 state->last_end = end_pfn;
1670 state->last_nid = nid;
1676 int __meminit early_pfn_to_nid(unsigned long pfn)
1678 static DEFINE_SPINLOCK(early_pfn_lock);
1681 spin_lock(&early_pfn_lock);
1682 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1684 nid = first_online_node;
1685 spin_unlock(&early_pfn_lock);
1689 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1691 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1694 if (early_page_uninitialised(pfn))
1696 __free_pages_core(page, order);
1700 * Check that the whole (or subset of) a pageblock given by the interval of
1701 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1702 * with the migration of free compaction scanner. The scanners then need to
1703 * use only pfn_valid_within() check for arches that allow holes within
1706 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1708 * It's possible on some configurations to have a setup like node0 node1 node0
1709 * i.e. it's possible that all pages within a zones range of pages do not
1710 * belong to a single zone. We assume that a border between node0 and node1
1711 * can occur within a single pageblock, but not a node0 node1 node0
1712 * interleaving within a single pageblock. It is therefore sufficient to check
1713 * the first and last page of a pageblock and avoid checking each individual
1714 * page in a pageblock.
1716 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1717 unsigned long end_pfn, struct zone *zone)
1719 struct page *start_page;
1720 struct page *end_page;
1722 /* end_pfn is one past the range we are checking */
1725 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1728 start_page = pfn_to_online_page(start_pfn);
1732 if (page_zone(start_page) != zone)
1735 end_page = pfn_to_page(end_pfn);
1737 /* This gives a shorter code than deriving page_zone(end_page) */
1738 if (page_zone_id(start_page) != page_zone_id(end_page))
1744 void set_zone_contiguous(struct zone *zone)
1746 unsigned long block_start_pfn = zone->zone_start_pfn;
1747 unsigned long block_end_pfn;
1749 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1750 for (; block_start_pfn < zone_end_pfn(zone);
1751 block_start_pfn = block_end_pfn,
1752 block_end_pfn += pageblock_nr_pages) {
1754 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1756 if (!__pageblock_pfn_to_page(block_start_pfn,
1757 block_end_pfn, zone))
1762 /* We confirm that there is no hole */
1763 zone->contiguous = true;
1766 void clear_zone_contiguous(struct zone *zone)
1768 zone->contiguous = false;
1771 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1772 static void __init deferred_free_range(unsigned long pfn,
1773 unsigned long nr_pages)
1781 page = pfn_to_page(pfn);
1783 /* Free a large naturally-aligned chunk if possible */
1784 if (nr_pages == pageblock_nr_pages &&
1785 (pfn & (pageblock_nr_pages - 1)) == 0) {
1786 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1787 __free_pages_core(page, pageblock_order);
1791 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1792 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1793 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1794 __free_pages_core(page, 0);
1798 /* Completion tracking for deferred_init_memmap() threads */
1799 static atomic_t pgdat_init_n_undone __initdata;
1800 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1802 static inline void __init pgdat_init_report_one_done(void)
1804 if (atomic_dec_and_test(&pgdat_init_n_undone))
1805 complete(&pgdat_init_all_done_comp);
1809 * Returns true if page needs to be initialized or freed to buddy allocator.
1811 * First we check if pfn is valid on architectures where it is possible to have
1812 * holes within pageblock_nr_pages. On systems where it is not possible, this
1813 * function is optimized out.
1815 * Then, we check if a current large page is valid by only checking the validity
1818 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1820 if (!pfn_valid_within(pfn))
1822 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1828 * Free pages to buddy allocator. Try to free aligned pages in
1829 * pageblock_nr_pages sizes.
1831 static void __init deferred_free_pages(unsigned long pfn,
1832 unsigned long end_pfn)
1834 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1835 unsigned long nr_free = 0;
1837 for (; pfn < end_pfn; pfn++) {
1838 if (!deferred_pfn_valid(pfn)) {
1839 deferred_free_range(pfn - nr_free, nr_free);
1841 } else if (!(pfn & nr_pgmask)) {
1842 deferred_free_range(pfn - nr_free, nr_free);
1848 /* Free the last block of pages to allocator */
1849 deferred_free_range(pfn - nr_free, nr_free);
1853 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1854 * by performing it only once every pageblock_nr_pages.
1855 * Return number of pages initialized.
1857 static unsigned long __init deferred_init_pages(struct zone *zone,
1859 unsigned long end_pfn)
1861 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1862 int nid = zone_to_nid(zone);
1863 unsigned long nr_pages = 0;
1864 int zid = zone_idx(zone);
1865 struct page *page = NULL;
1867 for (; pfn < end_pfn; pfn++) {
1868 if (!deferred_pfn_valid(pfn)) {
1871 } else if (!page || !(pfn & nr_pgmask)) {
1872 page = pfn_to_page(pfn);
1876 __init_single_page(page, pfn, zid, nid);
1883 * This function is meant to pre-load the iterator for the zone init.
1884 * Specifically it walks through the ranges until we are caught up to the
1885 * first_init_pfn value and exits there. If we never encounter the value we
1886 * return false indicating there are no valid ranges left.
1889 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1890 unsigned long *spfn, unsigned long *epfn,
1891 unsigned long first_init_pfn)
1896 * Start out by walking through the ranges in this zone that have
1897 * already been initialized. We don't need to do anything with them
1898 * so we just need to flush them out of the system.
1900 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1901 if (*epfn <= first_init_pfn)
1903 if (*spfn < first_init_pfn)
1904 *spfn = first_init_pfn;
1913 * Initialize and free pages. We do it in two loops: first we initialize
1914 * struct page, then free to buddy allocator, because while we are
1915 * freeing pages we can access pages that are ahead (computing buddy
1916 * page in __free_one_page()).
1918 * In order to try and keep some memory in the cache we have the loop
1919 * broken along max page order boundaries. This way we will not cause
1920 * any issues with the buddy page computation.
1922 static unsigned long __init
1923 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1924 unsigned long *end_pfn)
1926 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1927 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1928 unsigned long nr_pages = 0;
1931 /* First we loop through and initialize the page values */
1932 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1935 if (mo_pfn <= *start_pfn)
1938 t = min(mo_pfn, *end_pfn);
1939 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1941 if (mo_pfn < *end_pfn) {
1942 *start_pfn = mo_pfn;
1947 /* Reset values and now loop through freeing pages as needed */
1950 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1956 t = min(mo_pfn, epfn);
1957 deferred_free_pages(spfn, t);
1967 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1970 unsigned long spfn, epfn;
1971 struct zone *zone = arg;
1974 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1977 * Initialize and free pages in MAX_ORDER sized increments so that we
1978 * can avoid introducing any issues with the buddy allocator.
1980 while (spfn < end_pfn) {
1981 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1986 /* An arch may override for more concurrency. */
1988 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1993 /* Initialise remaining memory on a node */
1994 static int __init deferred_init_memmap(void *data)
1996 pg_data_t *pgdat = data;
1997 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1998 unsigned long spfn = 0, epfn = 0;
1999 unsigned long first_init_pfn, flags;
2000 unsigned long start = jiffies;
2002 int zid, max_threads;
2005 /* Bind memory initialisation thread to a local node if possible */
2006 if (!cpumask_empty(cpumask))
2007 set_cpus_allowed_ptr(current, cpumask);
2009 pgdat_resize_lock(pgdat, &flags);
2010 first_init_pfn = pgdat->first_deferred_pfn;
2011 if (first_init_pfn == ULONG_MAX) {
2012 pgdat_resize_unlock(pgdat, &flags);
2013 pgdat_init_report_one_done();
2017 /* Sanity check boundaries */
2018 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2019 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2020 pgdat->first_deferred_pfn = ULONG_MAX;
2023 * Once we unlock here, the zone cannot be grown anymore, thus if an
2024 * interrupt thread must allocate this early in boot, zone must be
2025 * pre-grown prior to start of deferred page initialization.
2027 pgdat_resize_unlock(pgdat, &flags);
2029 /* Only the highest zone is deferred so find it */
2030 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2031 zone = pgdat->node_zones + zid;
2032 if (first_init_pfn < zone_end_pfn(zone))
2036 /* If the zone is empty somebody else may have cleared out the zone */
2037 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2041 max_threads = deferred_page_init_max_threads(cpumask);
2043 while (spfn < epfn) {
2044 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2045 struct padata_mt_job job = {
2046 .thread_fn = deferred_init_memmap_chunk,
2049 .size = epfn_align - spfn,
2050 .align = PAGES_PER_SECTION,
2051 .min_chunk = PAGES_PER_SECTION,
2052 .max_threads = max_threads,
2055 padata_do_multithreaded(&job);
2056 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2060 /* Sanity check that the next zone really is unpopulated */
2061 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2063 pr_info("node %d deferred pages initialised in %ums\n",
2064 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2066 pgdat_init_report_one_done();
2071 * If this zone has deferred pages, try to grow it by initializing enough
2072 * deferred pages to satisfy the allocation specified by order, rounded up to
2073 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2074 * of SECTION_SIZE bytes by initializing struct pages in increments of
2075 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2077 * Return true when zone was grown, otherwise return false. We return true even
2078 * when we grow less than requested, to let the caller decide if there are
2079 * enough pages to satisfy the allocation.
2081 * Note: We use noinline because this function is needed only during boot, and
2082 * it is called from a __ref function _deferred_grow_zone. This way we are
2083 * making sure that it is not inlined into permanent text section.
2085 static noinline bool __init
2086 deferred_grow_zone(struct zone *zone, unsigned int order)
2088 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2089 pg_data_t *pgdat = zone->zone_pgdat;
2090 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2091 unsigned long spfn, epfn, flags;
2092 unsigned long nr_pages = 0;
2095 /* Only the last zone may have deferred pages */
2096 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2099 pgdat_resize_lock(pgdat, &flags);
2102 * If someone grew this zone while we were waiting for spinlock, return
2103 * true, as there might be enough pages already.
2105 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2106 pgdat_resize_unlock(pgdat, &flags);
2110 /* If the zone is empty somebody else may have cleared out the zone */
2111 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2112 first_deferred_pfn)) {
2113 pgdat->first_deferred_pfn = ULONG_MAX;
2114 pgdat_resize_unlock(pgdat, &flags);
2115 /* Retry only once. */
2116 return first_deferred_pfn != ULONG_MAX;
2120 * Initialize and free pages in MAX_ORDER sized increments so
2121 * that we can avoid introducing any issues with the buddy
2124 while (spfn < epfn) {
2125 /* update our first deferred PFN for this section */
2126 first_deferred_pfn = spfn;
2128 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2129 touch_nmi_watchdog();
2131 /* We should only stop along section boundaries */
2132 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2135 /* If our quota has been met we can stop here */
2136 if (nr_pages >= nr_pages_needed)
2140 pgdat->first_deferred_pfn = spfn;
2141 pgdat_resize_unlock(pgdat, &flags);
2143 return nr_pages > 0;
2147 * deferred_grow_zone() is __init, but it is called from
2148 * get_page_from_freelist() during early boot until deferred_pages permanently
2149 * disables this call. This is why we have refdata wrapper to avoid warning,
2150 * and to ensure that the function body gets unloaded.
2153 _deferred_grow_zone(struct zone *zone, unsigned int order)
2155 return deferred_grow_zone(zone, order);
2158 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2160 void __init page_alloc_init_late(void)
2165 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2167 /* There will be num_node_state(N_MEMORY) threads */
2168 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2169 for_each_node_state(nid, N_MEMORY) {
2170 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2173 /* Block until all are initialised */
2174 wait_for_completion(&pgdat_init_all_done_comp);
2177 * The number of managed pages has changed due to the initialisation
2178 * so the pcpu batch and high limits needs to be updated or the limits
2179 * will be artificially small.
2181 for_each_populated_zone(zone)
2182 zone_pcp_update(zone);
2185 * We initialized the rest of the deferred pages. Permanently disable
2186 * on-demand struct page initialization.
2188 static_branch_disable(&deferred_pages);
2190 /* Reinit limits that are based on free pages after the kernel is up */
2191 files_maxfiles_init();
2196 /* Discard memblock private memory */
2199 for_each_node_state(nid, N_MEMORY)
2200 shuffle_free_memory(NODE_DATA(nid));
2202 for_each_populated_zone(zone)
2203 set_zone_contiguous(zone);
2207 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2208 void __init init_cma_reserved_pageblock(struct page *page)
2210 unsigned i = pageblock_nr_pages;
2211 struct page *p = page;
2214 __ClearPageReserved(p);
2215 set_page_count(p, 0);
2218 set_pageblock_migratetype(page, MIGRATE_CMA);
2220 if (pageblock_order >= MAX_ORDER) {
2221 i = pageblock_nr_pages;
2224 set_page_refcounted(p);
2225 __free_pages(p, MAX_ORDER - 1);
2226 p += MAX_ORDER_NR_PAGES;
2227 } while (i -= MAX_ORDER_NR_PAGES);
2229 set_page_refcounted(page);
2230 __free_pages(page, pageblock_order);
2233 adjust_managed_page_count(page, pageblock_nr_pages);
2234 page_zone(page)->cma_pages += pageblock_nr_pages;
2239 * The order of subdivision here is critical for the IO subsystem.
2240 * Please do not alter this order without good reasons and regression
2241 * testing. Specifically, as large blocks of memory are subdivided,
2242 * the order in which smaller blocks are delivered depends on the order
2243 * they're subdivided in this function. This is the primary factor
2244 * influencing the order in which pages are delivered to the IO
2245 * subsystem according to empirical testing, and this is also justified
2246 * by considering the behavior of a buddy system containing a single
2247 * large block of memory acted on by a series of small allocations.
2248 * This behavior is a critical factor in sglist merging's success.
2252 static inline void expand(struct zone *zone, struct page *page,
2253 int low, int high, int migratetype)
2255 unsigned long size = 1 << high;
2257 while (high > low) {
2260 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2263 * Mark as guard pages (or page), that will allow to
2264 * merge back to allocator when buddy will be freed.
2265 * Corresponding page table entries will not be touched,
2266 * pages will stay not present in virtual address space
2268 if (set_page_guard(zone, &page[size], high, migratetype))
2271 add_to_free_list(&page[size], zone, high, migratetype);
2272 set_buddy_order(&page[size], high);
2276 static void check_new_page_bad(struct page *page)
2278 if (unlikely(page->flags & __PG_HWPOISON)) {
2279 /* Don't complain about hwpoisoned pages */
2280 page_mapcount_reset(page); /* remove PageBuddy */
2285 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2289 * This page is about to be returned from the page allocator
2291 static inline int check_new_page(struct page *page)
2293 if (likely(page_expected_state(page,
2294 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2297 check_new_page_bad(page);
2301 #ifdef CONFIG_DEBUG_VM
2303 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2304 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2305 * also checked when pcp lists are refilled from the free lists.
2307 static inline bool check_pcp_refill(struct page *page)
2309 if (debug_pagealloc_enabled_static())
2310 return check_new_page(page);
2315 static inline bool check_new_pcp(struct page *page)
2317 return check_new_page(page);
2321 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2322 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2323 * enabled, they are also checked when being allocated from the pcp lists.
2325 static inline bool check_pcp_refill(struct page *page)
2327 return check_new_page(page);
2329 static inline bool check_new_pcp(struct page *page)
2331 if (debug_pagealloc_enabled_static())
2332 return check_new_page(page);
2336 #endif /* CONFIG_DEBUG_VM */
2338 static bool check_new_pages(struct page *page, unsigned int order)
2341 for (i = 0; i < (1 << order); i++) {
2342 struct page *p = page + i;
2344 if (unlikely(check_new_page(p)))
2351 inline void post_alloc_hook(struct page *page, unsigned int order,
2356 set_page_private(page, 0);
2357 set_page_refcounted(page);
2359 arch_alloc_page(page, order);
2360 debug_pagealloc_map_pages(page, 1 << order);
2363 * Page unpoisoning must happen before memory initialization.
2364 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2365 * allocations and the page unpoisoning code will complain.
2367 kernel_unpoison_pages(page, 1 << order);
2370 * As memory initialization might be integrated into KASAN,
2371 * kasan_alloc_pages and kernel_init_free_pages must be
2372 * kept together to avoid discrepancies in behavior.
2374 init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2375 kasan_alloc_pages(page, order, init);
2376 if (init && !kasan_has_integrated_init())
2377 kernel_init_free_pages(page, 1 << order);
2379 set_page_owner(page, order, gfp_flags);
2382 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2383 unsigned int alloc_flags)
2385 post_alloc_hook(page, order, gfp_flags);
2387 if (order && (gfp_flags & __GFP_COMP))
2388 prep_compound_page(page, order);
2391 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2392 * allocate the page. The expectation is that the caller is taking
2393 * steps that will free more memory. The caller should avoid the page
2394 * being used for !PFMEMALLOC purposes.
2396 if (alloc_flags & ALLOC_NO_WATERMARKS)
2397 set_page_pfmemalloc(page);
2399 clear_page_pfmemalloc(page);
2403 * Go through the free lists for the given migratetype and remove
2404 * the smallest available page from the freelists
2406 static __always_inline
2407 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2410 unsigned int current_order;
2411 struct free_area *area;
2414 /* Find a page of the appropriate size in the preferred list */
2415 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2416 area = &(zone->free_area[current_order]);
2417 page = get_page_from_free_area(area, migratetype);
2420 del_page_from_free_list(page, zone, current_order);
2421 expand(zone, page, order, current_order, migratetype);
2422 set_pcppage_migratetype(page, migratetype);
2431 * This array describes the order lists are fallen back to when
2432 * the free lists for the desirable migrate type are depleted
2434 static int fallbacks[MIGRATE_TYPES][3] = {
2435 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2436 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2437 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2439 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2441 #ifdef CONFIG_MEMORY_ISOLATION
2442 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2447 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2450 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2453 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2454 unsigned int order) { return NULL; }
2458 * Move the free pages in a range to the freelist tail of the requested type.
2459 * Note that start_page and end_pages are not aligned on a pageblock
2460 * boundary. If alignment is required, use move_freepages_block()
2462 static int move_freepages(struct zone *zone,
2463 unsigned long start_pfn, unsigned long end_pfn,
2464 int migratetype, int *num_movable)
2469 int pages_moved = 0;
2471 for (pfn = start_pfn; pfn <= end_pfn;) {
2472 if (!pfn_valid_within(pfn)) {
2477 page = pfn_to_page(pfn);
2478 if (!PageBuddy(page)) {
2480 * We assume that pages that could be isolated for
2481 * migration are movable. But we don't actually try
2482 * isolating, as that would be expensive.
2485 (PageLRU(page) || __PageMovable(page)))
2491 /* Make sure we are not inadvertently changing nodes */
2492 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2493 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2495 order = buddy_order(page);
2496 move_to_free_list(page, zone, order, migratetype);
2498 pages_moved += 1 << order;
2504 int move_freepages_block(struct zone *zone, struct page *page,
2505 int migratetype, int *num_movable)
2507 unsigned long start_pfn, end_pfn, pfn;
2512 pfn = page_to_pfn(page);
2513 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2514 end_pfn = start_pfn + pageblock_nr_pages - 1;
2516 /* Do not cross zone boundaries */
2517 if (!zone_spans_pfn(zone, start_pfn))
2519 if (!zone_spans_pfn(zone, end_pfn))
2522 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2526 static void change_pageblock_range(struct page *pageblock_page,
2527 int start_order, int migratetype)
2529 int nr_pageblocks = 1 << (start_order - pageblock_order);
2531 while (nr_pageblocks--) {
2532 set_pageblock_migratetype(pageblock_page, migratetype);
2533 pageblock_page += pageblock_nr_pages;
2538 * When we are falling back to another migratetype during allocation, try to
2539 * steal extra free pages from the same pageblocks to satisfy further
2540 * allocations, instead of polluting multiple pageblocks.
2542 * If we are stealing a relatively large buddy page, it is likely there will
2543 * be more free pages in the pageblock, so try to steal them all. For
2544 * reclaimable and unmovable allocations, we steal regardless of page size,
2545 * as fragmentation caused by those allocations polluting movable pageblocks
2546 * is worse than movable allocations stealing from unmovable and reclaimable
2549 static bool can_steal_fallback(unsigned int order, int start_mt)
2552 * Leaving this order check is intended, although there is
2553 * relaxed order check in next check. The reason is that
2554 * we can actually steal whole pageblock if this condition met,
2555 * but, below check doesn't guarantee it and that is just heuristic
2556 * so could be changed anytime.
2558 if (order >= pageblock_order)
2561 if (order >= pageblock_order / 2 ||
2562 start_mt == MIGRATE_RECLAIMABLE ||
2563 start_mt == MIGRATE_UNMOVABLE ||
2564 page_group_by_mobility_disabled)
2570 static inline bool boost_watermark(struct zone *zone)
2572 unsigned long max_boost;
2574 if (!watermark_boost_factor)
2577 * Don't bother in zones that are unlikely to produce results.
2578 * On small machines, including kdump capture kernels running
2579 * in a small area, boosting the watermark can cause an out of
2580 * memory situation immediately.
2582 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2585 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2586 watermark_boost_factor, 10000);
2589 * high watermark may be uninitialised if fragmentation occurs
2590 * very early in boot so do not boost. We do not fall
2591 * through and boost by pageblock_nr_pages as failing
2592 * allocations that early means that reclaim is not going
2593 * to help and it may even be impossible to reclaim the
2594 * boosted watermark resulting in a hang.
2599 max_boost = max(pageblock_nr_pages, max_boost);
2601 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2608 * This function implements actual steal behaviour. If order is large enough,
2609 * we can steal whole pageblock. If not, we first move freepages in this
2610 * pageblock to our migratetype and determine how many already-allocated pages
2611 * are there in the pageblock with a compatible migratetype. If at least half
2612 * of pages are free or compatible, we can change migratetype of the pageblock
2613 * itself, so pages freed in the future will be put on the correct free list.
2615 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2616 unsigned int alloc_flags, int start_type, bool whole_block)
2618 unsigned int current_order = buddy_order(page);
2619 int free_pages, movable_pages, alike_pages;
2622 old_block_type = get_pageblock_migratetype(page);
2625 * This can happen due to races and we want to prevent broken
2626 * highatomic accounting.
2628 if (is_migrate_highatomic(old_block_type))
2631 /* Take ownership for orders >= pageblock_order */
2632 if (current_order >= pageblock_order) {
2633 change_pageblock_range(page, current_order, start_type);
2638 * Boost watermarks to increase reclaim pressure to reduce the
2639 * likelihood of future fallbacks. Wake kswapd now as the node
2640 * may be balanced overall and kswapd will not wake naturally.
2642 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2643 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2645 /* We are not allowed to try stealing from the whole block */
2649 free_pages = move_freepages_block(zone, page, start_type,
2652 * Determine how many pages are compatible with our allocation.
2653 * For movable allocation, it's the number of movable pages which
2654 * we just obtained. For other types it's a bit more tricky.
2656 if (start_type == MIGRATE_MOVABLE) {
2657 alike_pages = movable_pages;
2660 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2661 * to MOVABLE pageblock, consider all non-movable pages as
2662 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2663 * vice versa, be conservative since we can't distinguish the
2664 * exact migratetype of non-movable pages.
2666 if (old_block_type == MIGRATE_MOVABLE)
2667 alike_pages = pageblock_nr_pages
2668 - (free_pages + movable_pages);
2673 /* moving whole block can fail due to zone boundary conditions */
2678 * If a sufficient number of pages in the block are either free or of
2679 * comparable migratability as our allocation, claim the whole block.
2681 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2682 page_group_by_mobility_disabled)
2683 set_pageblock_migratetype(page, start_type);
2688 move_to_free_list(page, zone, current_order, start_type);
2692 * Check whether there is a suitable fallback freepage with requested order.
2693 * If only_stealable is true, this function returns fallback_mt only if
2694 * we can steal other freepages all together. This would help to reduce
2695 * fragmentation due to mixed migratetype pages in one pageblock.
2697 int find_suitable_fallback(struct free_area *area, unsigned int order,
2698 int migratetype, bool only_stealable, bool *can_steal)
2703 if (area->nr_free == 0)
2708 fallback_mt = fallbacks[migratetype][i];
2709 if (fallback_mt == MIGRATE_TYPES)
2712 if (free_area_empty(area, fallback_mt))
2715 if (can_steal_fallback(order, migratetype))
2718 if (!only_stealable)
2729 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2730 * there are no empty page blocks that contain a page with a suitable order
2732 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2733 unsigned int alloc_order)
2736 unsigned long max_managed, flags;
2739 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2740 * Check is race-prone but harmless.
2742 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2743 if (zone->nr_reserved_highatomic >= max_managed)
2746 spin_lock_irqsave(&zone->lock, flags);
2748 /* Recheck the nr_reserved_highatomic limit under the lock */
2749 if (zone->nr_reserved_highatomic >= max_managed)
2753 mt = get_pageblock_migratetype(page);
2754 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2755 && !is_migrate_cma(mt)) {
2756 zone->nr_reserved_highatomic += pageblock_nr_pages;
2757 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2758 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2762 spin_unlock_irqrestore(&zone->lock, flags);
2766 * Used when an allocation is about to fail under memory pressure. This
2767 * potentially hurts the reliability of high-order allocations when under
2768 * intense memory pressure but failed atomic allocations should be easier
2769 * to recover from than an OOM.
2771 * If @force is true, try to unreserve a pageblock even though highatomic
2772 * pageblock is exhausted.
2774 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2777 struct zonelist *zonelist = ac->zonelist;
2778 unsigned long flags;
2785 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2788 * Preserve at least one pageblock unless memory pressure
2791 if (!force && zone->nr_reserved_highatomic <=
2795 spin_lock_irqsave(&zone->lock, flags);
2796 for (order = 0; order < MAX_ORDER; order++) {
2797 struct free_area *area = &(zone->free_area[order]);
2799 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2804 * In page freeing path, migratetype change is racy so
2805 * we can counter several free pages in a pageblock
2806 * in this loop although we changed the pageblock type
2807 * from highatomic to ac->migratetype. So we should
2808 * adjust the count once.
2810 if (is_migrate_highatomic_page(page)) {
2812 * It should never happen but changes to
2813 * locking could inadvertently allow a per-cpu
2814 * drain to add pages to MIGRATE_HIGHATOMIC
2815 * while unreserving so be safe and watch for
2818 zone->nr_reserved_highatomic -= min(
2820 zone->nr_reserved_highatomic);
2824 * Convert to ac->migratetype and avoid the normal
2825 * pageblock stealing heuristics. Minimally, the caller
2826 * is doing the work and needs the pages. More
2827 * importantly, if the block was always converted to
2828 * MIGRATE_UNMOVABLE or another type then the number
2829 * of pageblocks that cannot be completely freed
2832 set_pageblock_migratetype(page, ac->migratetype);
2833 ret = move_freepages_block(zone, page, ac->migratetype,
2836 spin_unlock_irqrestore(&zone->lock, flags);
2840 spin_unlock_irqrestore(&zone->lock, flags);
2847 * Try finding a free buddy page on the fallback list and put it on the free
2848 * list of requested migratetype, possibly along with other pages from the same
2849 * block, depending on fragmentation avoidance heuristics. Returns true if
2850 * fallback was found so that __rmqueue_smallest() can grab it.
2852 * The use of signed ints for order and current_order is a deliberate
2853 * deviation from the rest of this file, to make the for loop
2854 * condition simpler.
2856 static __always_inline bool
2857 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2858 unsigned int alloc_flags)
2860 struct free_area *area;
2862 int min_order = order;
2868 * Do not steal pages from freelists belonging to other pageblocks
2869 * i.e. orders < pageblock_order. If there are no local zones free,
2870 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2872 if (alloc_flags & ALLOC_NOFRAGMENT)
2873 min_order = pageblock_order;
2876 * Find the largest available free page in the other list. This roughly
2877 * approximates finding the pageblock with the most free pages, which
2878 * would be too costly to do exactly.
2880 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2882 area = &(zone->free_area[current_order]);
2883 fallback_mt = find_suitable_fallback(area, current_order,
2884 start_migratetype, false, &can_steal);
2885 if (fallback_mt == -1)
2889 * We cannot steal all free pages from the pageblock and the
2890 * requested migratetype is movable. In that case it's better to
2891 * steal and split the smallest available page instead of the
2892 * largest available page, because even if the next movable
2893 * allocation falls back into a different pageblock than this
2894 * one, it won't cause permanent fragmentation.
2896 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2897 && current_order > order)
2906 for (current_order = order; current_order < MAX_ORDER;
2908 area = &(zone->free_area[current_order]);
2909 fallback_mt = find_suitable_fallback(area, current_order,
2910 start_migratetype, false, &can_steal);
2911 if (fallback_mt != -1)
2916 * This should not happen - we already found a suitable fallback
2917 * when looking for the largest page.
2919 VM_BUG_ON(current_order == MAX_ORDER);
2922 page = get_page_from_free_area(area, fallback_mt);
2924 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2927 trace_mm_page_alloc_extfrag(page, order, current_order,
2928 start_migratetype, fallback_mt);
2935 * Do the hard work of removing an element from the buddy allocator.
2936 * Call me with the zone->lock already held.
2938 static __always_inline struct page *
2939 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2940 unsigned int alloc_flags)
2944 if (IS_ENABLED(CONFIG_CMA)) {
2946 * Balance movable allocations between regular and CMA areas by
2947 * allocating from CMA when over half of the zone's free memory
2948 * is in the CMA area.
2950 if (alloc_flags & ALLOC_CMA &&
2951 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2952 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2953 page = __rmqueue_cma_fallback(zone, order);
2959 page = __rmqueue_smallest(zone, order, migratetype);
2960 if (unlikely(!page)) {
2961 if (alloc_flags & ALLOC_CMA)
2962 page = __rmqueue_cma_fallback(zone, order);
2964 if (!page && __rmqueue_fallback(zone, order, migratetype,
2970 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2975 * Obtain a specified number of elements from the buddy allocator, all under
2976 * a single hold of the lock, for efficiency. Add them to the supplied list.
2977 * Returns the number of new pages which were placed at *list.
2979 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2980 unsigned long count, struct list_head *list,
2981 int migratetype, unsigned int alloc_flags)
2983 int i, allocated = 0;
2986 * local_lock_irq held so equivalent to spin_lock_irqsave for
2987 * both PREEMPT_RT and non-PREEMPT_RT configurations.
2989 spin_lock(&zone->lock);
2990 for (i = 0; i < count; ++i) {
2991 struct page *page = __rmqueue(zone, order, migratetype,
2993 if (unlikely(page == NULL))
2996 if (unlikely(check_pcp_refill(page)))
3000 * Split buddy pages returned by expand() are received here in
3001 * physical page order. The page is added to the tail of
3002 * caller's list. From the callers perspective, the linked list
3003 * is ordered by page number under some conditions. This is
3004 * useful for IO devices that can forward direction from the
3005 * head, thus also in the physical page order. This is useful
3006 * for IO devices that can merge IO requests if the physical
3007 * pages are ordered properly.
3009 list_add_tail(&page->lru, list);
3011 if (is_migrate_cma(get_pcppage_migratetype(page)))
3012 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3017 * i pages were removed from the buddy list even if some leak due
3018 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3019 * on i. Do not confuse with 'allocated' which is the number of
3020 * pages added to the pcp list.
3022 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3023 spin_unlock(&zone->lock);
3029 * Called from the vmstat counter updater to drain pagesets of this
3030 * currently executing processor on remote nodes after they have
3033 * Note that this function must be called with the thread pinned to
3034 * a single processor.
3036 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3038 unsigned long flags;
3039 int to_drain, batch;
3041 local_lock_irqsave(&pagesets.lock, flags);
3042 batch = READ_ONCE(pcp->batch);
3043 to_drain = min(pcp->count, batch);
3045 free_pcppages_bulk(zone, to_drain, pcp);
3046 local_unlock_irqrestore(&pagesets.lock, flags);
3051 * Drain pcplists of the indicated processor and zone.
3053 * The processor must either be the current processor and the
3054 * thread pinned to the current processor or a processor that
3057 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3059 unsigned long flags;
3060 struct per_cpu_pages *pcp;
3062 local_lock_irqsave(&pagesets.lock, flags);
3064 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3066 free_pcppages_bulk(zone, pcp->count, pcp);
3068 local_unlock_irqrestore(&pagesets.lock, flags);
3072 * Drain pcplists of all zones on the indicated processor.
3074 * The processor must either be the current processor and the
3075 * thread pinned to the current processor or a processor that
3078 static void drain_pages(unsigned int cpu)
3082 for_each_populated_zone(zone) {
3083 drain_pages_zone(cpu, zone);
3088 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3090 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3091 * the single zone's pages.
3093 void drain_local_pages(struct zone *zone)
3095 int cpu = smp_processor_id();
3098 drain_pages_zone(cpu, zone);
3103 static void drain_local_pages_wq(struct work_struct *work)
3105 struct pcpu_drain *drain;
3107 drain = container_of(work, struct pcpu_drain, work);
3110 * drain_all_pages doesn't use proper cpu hotplug protection so
3111 * we can race with cpu offline when the WQ can move this from
3112 * a cpu pinned worker to an unbound one. We can operate on a different
3113 * cpu which is alright but we also have to make sure to not move to
3117 drain_local_pages(drain->zone);
3122 * The implementation of drain_all_pages(), exposing an extra parameter to
3123 * drain on all cpus.
3125 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3126 * not empty. The check for non-emptiness can however race with a free to
3127 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3128 * that need the guarantee that every CPU has drained can disable the
3129 * optimizing racy check.
3131 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3136 * Allocate in the BSS so we wont require allocation in
3137 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3139 static cpumask_t cpus_with_pcps;
3142 * Make sure nobody triggers this path before mm_percpu_wq is fully
3145 if (WARN_ON_ONCE(!mm_percpu_wq))
3149 * Do not drain if one is already in progress unless it's specific to
3150 * a zone. Such callers are primarily CMA and memory hotplug and need
3151 * the drain to be complete when the call returns.
3153 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3156 mutex_lock(&pcpu_drain_mutex);
3160 * We don't care about racing with CPU hotplug event
3161 * as offline notification will cause the notified
3162 * cpu to drain that CPU pcps and on_each_cpu_mask
3163 * disables preemption as part of its processing
3165 for_each_online_cpu(cpu) {
3166 struct per_cpu_pages *pcp;
3168 bool has_pcps = false;
3170 if (force_all_cpus) {
3172 * The pcp.count check is racy, some callers need a
3173 * guarantee that no cpu is missed.
3177 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3181 for_each_populated_zone(z) {
3182 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3191 cpumask_set_cpu(cpu, &cpus_with_pcps);
3193 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3196 for_each_cpu(cpu, &cpus_with_pcps) {
3197 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3200 INIT_WORK(&drain->work, drain_local_pages_wq);
3201 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3203 for_each_cpu(cpu, &cpus_with_pcps)
3204 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3206 mutex_unlock(&pcpu_drain_mutex);
3210 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3212 * When zone parameter is non-NULL, spill just the single zone's pages.
3214 * Note that this can be extremely slow as the draining happens in a workqueue.
3216 void drain_all_pages(struct zone *zone)
3218 __drain_all_pages(zone, false);
3221 #ifdef CONFIG_HIBERNATION
3224 * Touch the watchdog for every WD_PAGE_COUNT pages.
3226 #define WD_PAGE_COUNT (128*1024)
3228 void mark_free_pages(struct zone *zone)
3230 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3231 unsigned long flags;
3232 unsigned int order, t;
3235 if (zone_is_empty(zone))
3238 spin_lock_irqsave(&zone->lock, flags);
3240 max_zone_pfn = zone_end_pfn(zone);
3241 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3242 if (pfn_valid(pfn)) {
3243 page = pfn_to_page(pfn);
3245 if (!--page_count) {
3246 touch_nmi_watchdog();
3247 page_count = WD_PAGE_COUNT;
3250 if (page_zone(page) != zone)
3253 if (!swsusp_page_is_forbidden(page))
3254 swsusp_unset_page_free(page);
3257 for_each_migratetype_order(order, t) {
3258 list_for_each_entry(page,
3259 &zone->free_area[order].free_list[t], lru) {
3262 pfn = page_to_pfn(page);
3263 for (i = 0; i < (1UL << order); i++) {
3264 if (!--page_count) {
3265 touch_nmi_watchdog();
3266 page_count = WD_PAGE_COUNT;
3268 swsusp_set_page_free(pfn_to_page(pfn + i));
3272 spin_unlock_irqrestore(&zone->lock, flags);
3274 #endif /* CONFIG_PM */
3276 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3280 if (!free_pcp_prepare(page))
3283 migratetype = get_pfnblock_migratetype(page, pfn);
3284 set_pcppage_migratetype(page, migratetype);
3288 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3290 struct zone *zone = page_zone(page);
3291 struct per_cpu_pages *pcp;
3294 migratetype = get_pcppage_migratetype(page);
3295 __count_vm_event(PGFREE);
3298 * We only track unmovable, reclaimable and movable on pcp lists.
3299 * Free ISOLATE pages back to the allocator because they are being
3300 * offlined but treat HIGHATOMIC as movable pages so we can get those
3301 * areas back if necessary. Otherwise, we may have to free
3302 * excessively into the page allocator
3304 if (migratetype >= MIGRATE_PCPTYPES) {
3305 if (unlikely(is_migrate_isolate(migratetype))) {
3306 free_one_page(zone, page, pfn, 0, migratetype,
3310 migratetype = MIGRATE_MOVABLE;
3313 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3314 list_add(&page->lru, &pcp->lists[migratetype]);
3316 if (pcp->count >= READ_ONCE(pcp->high))
3317 free_pcppages_bulk(zone, READ_ONCE(pcp->batch), pcp);
3321 * Free a 0-order page
3323 void free_unref_page(struct page *page)
3325 unsigned long flags;
3326 unsigned long pfn = page_to_pfn(page);
3328 if (!free_unref_page_prepare(page, pfn))
3331 local_lock_irqsave(&pagesets.lock, flags);
3332 free_unref_page_commit(page, pfn);
3333 local_unlock_irqrestore(&pagesets.lock, flags);
3337 * Free a list of 0-order pages
3339 void free_unref_page_list(struct list_head *list)
3341 struct page *page, *next;
3342 unsigned long flags, pfn;
3343 int batch_count = 0;
3345 /* Prepare pages for freeing */
3346 list_for_each_entry_safe(page, next, list, lru) {
3347 pfn = page_to_pfn(page);
3348 if (!free_unref_page_prepare(page, pfn))
3349 list_del(&page->lru);
3350 set_page_private(page, pfn);
3353 local_lock_irqsave(&pagesets.lock, flags);
3354 list_for_each_entry_safe(page, next, list, lru) {
3355 unsigned long pfn = page_private(page);
3357 set_page_private(page, 0);
3358 trace_mm_page_free_batched(page);
3359 free_unref_page_commit(page, pfn);
3362 * Guard against excessive IRQ disabled times when we get
3363 * a large list of pages to free.
3365 if (++batch_count == SWAP_CLUSTER_MAX) {
3366 local_unlock_irqrestore(&pagesets.lock, flags);
3368 local_lock_irqsave(&pagesets.lock, flags);
3371 local_unlock_irqrestore(&pagesets.lock, flags);
3375 * split_page takes a non-compound higher-order page, and splits it into
3376 * n (1<<order) sub-pages: page[0..n]
3377 * Each sub-page must be freed individually.
3379 * Note: this is probably too low level an operation for use in drivers.
3380 * Please consult with lkml before using this in your driver.
3382 void split_page(struct page *page, unsigned int order)
3386 VM_BUG_ON_PAGE(PageCompound(page), page);
3387 VM_BUG_ON_PAGE(!page_count(page), page);
3389 for (i = 1; i < (1 << order); i++)
3390 set_page_refcounted(page + i);
3391 split_page_owner(page, 1 << order);
3392 split_page_memcg(page, 1 << order);
3394 EXPORT_SYMBOL_GPL(split_page);
3396 int __isolate_free_page(struct page *page, unsigned int order)
3398 unsigned long watermark;
3402 BUG_ON(!PageBuddy(page));
3404 zone = page_zone(page);
3405 mt = get_pageblock_migratetype(page);
3407 if (!is_migrate_isolate(mt)) {
3409 * Obey watermarks as if the page was being allocated. We can
3410 * emulate a high-order watermark check with a raised order-0
3411 * watermark, because we already know our high-order page
3414 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3415 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3418 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3421 /* Remove page from free list */
3423 del_page_from_free_list(page, zone, order);
3426 * Set the pageblock if the isolated page is at least half of a
3429 if (order >= pageblock_order - 1) {
3430 struct page *endpage = page + (1 << order) - 1;
3431 for (; page < endpage; page += pageblock_nr_pages) {
3432 int mt = get_pageblock_migratetype(page);
3433 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3434 && !is_migrate_highatomic(mt))
3435 set_pageblock_migratetype(page,
3441 return 1UL << order;
3445 * __putback_isolated_page - Return a now-isolated page back where we got it
3446 * @page: Page that was isolated
3447 * @order: Order of the isolated page
3448 * @mt: The page's pageblock's migratetype
3450 * This function is meant to return a page pulled from the free lists via
3451 * __isolate_free_page back to the free lists they were pulled from.
3453 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3455 struct zone *zone = page_zone(page);
3457 /* zone lock should be held when this function is called */
3458 lockdep_assert_held(&zone->lock);
3460 /* Return isolated page to tail of freelist. */
3461 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3462 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3466 * Update NUMA hit/miss statistics
3468 * Must be called with interrupts disabled.
3470 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3474 enum numa_stat_item local_stat = NUMA_LOCAL;
3476 /* skip numa counters update if numa stats is disabled */
3477 if (!static_branch_likely(&vm_numa_stat_key))
3480 if (zone_to_nid(z) != numa_node_id())
3481 local_stat = NUMA_OTHER;
3483 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3484 __count_numa_events(z, NUMA_HIT, nr_account);
3486 __count_numa_events(z, NUMA_MISS, nr_account);
3487 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3489 __count_numa_events(z, local_stat, nr_account);
3493 /* Remove page from the per-cpu list, caller must protect the list */
3495 struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3496 unsigned int alloc_flags,
3497 struct per_cpu_pages *pcp,
3498 struct list_head *list)
3503 if (list_empty(list)) {
3504 pcp->count += rmqueue_bulk(zone, 0,
3505 READ_ONCE(pcp->batch), list,
3506 migratetype, alloc_flags);
3507 if (unlikely(list_empty(list)))
3511 page = list_first_entry(list, struct page, lru);
3512 list_del(&page->lru);
3514 } while (check_new_pcp(page));
3519 /* Lock and remove page from the per-cpu list */
3520 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3521 struct zone *zone, gfp_t gfp_flags,
3522 int migratetype, unsigned int alloc_flags)
3524 struct per_cpu_pages *pcp;
3525 struct list_head *list;
3527 unsigned long flags;
3529 local_lock_irqsave(&pagesets.lock, flags);
3530 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3531 list = &pcp->lists[migratetype];
3532 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3533 local_unlock_irqrestore(&pagesets.lock, flags);
3535 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3536 zone_statistics(preferred_zone, zone, 1);
3542 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3545 struct page *rmqueue(struct zone *preferred_zone,
3546 struct zone *zone, unsigned int order,
3547 gfp_t gfp_flags, unsigned int alloc_flags,
3550 unsigned long flags;
3553 if (likely(order == 0)) {
3555 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3556 * we need to skip it when CMA area isn't allowed.
3558 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3559 migratetype != MIGRATE_MOVABLE) {
3560 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3561 migratetype, alloc_flags);
3567 * We most definitely don't want callers attempting to
3568 * allocate greater than order-1 page units with __GFP_NOFAIL.
3570 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3571 spin_lock_irqsave(&zone->lock, flags);
3576 * order-0 request can reach here when the pcplist is skipped
3577 * due to non-CMA allocation context. HIGHATOMIC area is
3578 * reserved for high-order atomic allocation, so order-0
3579 * request should skip it.
3581 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3582 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3584 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3587 page = __rmqueue(zone, order, migratetype, alloc_flags);
3588 } while (page && check_new_pages(page, order));
3592 __mod_zone_freepage_state(zone, -(1 << order),
3593 get_pcppage_migratetype(page));
3594 spin_unlock_irqrestore(&zone->lock, flags);
3596 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3597 zone_statistics(preferred_zone, zone, 1);
3600 /* Separate test+clear to avoid unnecessary atomics */
3601 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3602 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3603 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3606 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3610 spin_unlock_irqrestore(&zone->lock, flags);
3614 #ifdef CONFIG_FAIL_PAGE_ALLOC
3617 struct fault_attr attr;
3619 bool ignore_gfp_highmem;
3620 bool ignore_gfp_reclaim;
3622 } fail_page_alloc = {
3623 .attr = FAULT_ATTR_INITIALIZER,
3624 .ignore_gfp_reclaim = true,
3625 .ignore_gfp_highmem = true,
3629 static int __init setup_fail_page_alloc(char *str)
3631 return setup_fault_attr(&fail_page_alloc.attr, str);
3633 __setup("fail_page_alloc=", setup_fail_page_alloc);
3635 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3637 if (order < fail_page_alloc.min_order)
3639 if (gfp_mask & __GFP_NOFAIL)
3641 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3643 if (fail_page_alloc.ignore_gfp_reclaim &&
3644 (gfp_mask & __GFP_DIRECT_RECLAIM))
3647 return should_fail(&fail_page_alloc.attr, 1 << order);
3650 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3652 static int __init fail_page_alloc_debugfs(void)
3654 umode_t mode = S_IFREG | 0600;
3657 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3658 &fail_page_alloc.attr);
3660 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3661 &fail_page_alloc.ignore_gfp_reclaim);
3662 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3663 &fail_page_alloc.ignore_gfp_highmem);
3664 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3669 late_initcall(fail_page_alloc_debugfs);
3671 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3673 #else /* CONFIG_FAIL_PAGE_ALLOC */
3675 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3680 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3682 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3684 return __should_fail_alloc_page(gfp_mask, order);
3686 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3688 static inline long __zone_watermark_unusable_free(struct zone *z,
3689 unsigned int order, unsigned int alloc_flags)
3691 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3692 long unusable_free = (1 << order) - 1;
3695 * If the caller does not have rights to ALLOC_HARDER then subtract
3696 * the high-atomic reserves. This will over-estimate the size of the
3697 * atomic reserve but it avoids a search.
3699 if (likely(!alloc_harder))
3700 unusable_free += z->nr_reserved_highatomic;
3703 /* If allocation can't use CMA areas don't use free CMA pages */
3704 if (!(alloc_flags & ALLOC_CMA))
3705 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3708 return unusable_free;
3712 * Return true if free base pages are above 'mark'. For high-order checks it
3713 * will return true of the order-0 watermark is reached and there is at least
3714 * one free page of a suitable size. Checking now avoids taking the zone lock
3715 * to check in the allocation paths if no pages are free.
3717 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3718 int highest_zoneidx, unsigned int alloc_flags,
3723 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3725 /* free_pages may go negative - that's OK */
3726 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3728 if (alloc_flags & ALLOC_HIGH)
3731 if (unlikely(alloc_harder)) {
3733 * OOM victims can try even harder than normal ALLOC_HARDER
3734 * users on the grounds that it's definitely going to be in
3735 * the exit path shortly and free memory. Any allocation it
3736 * makes during the free path will be small and short-lived.
3738 if (alloc_flags & ALLOC_OOM)
3745 * Check watermarks for an order-0 allocation request. If these
3746 * are not met, then a high-order request also cannot go ahead
3747 * even if a suitable page happened to be free.
3749 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3752 /* If this is an order-0 request then the watermark is fine */
3756 /* For a high-order request, check at least one suitable page is free */
3757 for (o = order; o < MAX_ORDER; o++) {
3758 struct free_area *area = &z->free_area[o];
3764 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3765 if (!free_area_empty(area, mt))
3770 if ((alloc_flags & ALLOC_CMA) &&
3771 !free_area_empty(area, MIGRATE_CMA)) {
3775 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3781 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3782 int highest_zoneidx, unsigned int alloc_flags)
3784 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3785 zone_page_state(z, NR_FREE_PAGES));
3788 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3789 unsigned long mark, int highest_zoneidx,
3790 unsigned int alloc_flags, gfp_t gfp_mask)
3794 free_pages = zone_page_state(z, NR_FREE_PAGES);
3797 * Fast check for order-0 only. If this fails then the reserves
3798 * need to be calculated.
3803 fast_free = free_pages;
3804 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3805 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3809 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3813 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3814 * when checking the min watermark. The min watermark is the
3815 * point where boosting is ignored so that kswapd is woken up
3816 * when below the low watermark.
3818 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3819 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3820 mark = z->_watermark[WMARK_MIN];
3821 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3822 alloc_flags, free_pages);
3828 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3829 unsigned long mark, int highest_zoneidx)
3831 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3833 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3834 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3836 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3841 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3843 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3844 node_reclaim_distance;
3846 #else /* CONFIG_NUMA */
3847 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3851 #endif /* CONFIG_NUMA */
3854 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3855 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3856 * premature use of a lower zone may cause lowmem pressure problems that
3857 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3858 * probably too small. It only makes sense to spread allocations to avoid
3859 * fragmentation between the Normal and DMA32 zones.
3861 static inline unsigned int
3862 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3864 unsigned int alloc_flags;
3867 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3870 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3872 #ifdef CONFIG_ZONE_DMA32
3876 if (zone_idx(zone) != ZONE_NORMAL)
3880 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3881 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3882 * on UMA that if Normal is populated then so is DMA32.
3884 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3885 if (nr_online_nodes > 1 && !populated_zone(--zone))
3888 alloc_flags |= ALLOC_NOFRAGMENT;
3889 #endif /* CONFIG_ZONE_DMA32 */
3893 /* Must be called after current_gfp_context() which can change gfp_mask */
3894 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3895 unsigned int alloc_flags)
3898 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3899 alloc_flags |= ALLOC_CMA;
3905 * get_page_from_freelist goes through the zonelist trying to allocate
3908 static struct page *
3909 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3910 const struct alloc_context *ac)
3914 struct pglist_data *last_pgdat_dirty_limit = NULL;
3919 * Scan zonelist, looking for a zone with enough free.
3920 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3922 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3923 z = ac->preferred_zoneref;
3924 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3929 if (cpusets_enabled() &&
3930 (alloc_flags & ALLOC_CPUSET) &&
3931 !__cpuset_zone_allowed(zone, gfp_mask))
3934 * When allocating a page cache page for writing, we
3935 * want to get it from a node that is within its dirty
3936 * limit, such that no single node holds more than its
3937 * proportional share of globally allowed dirty pages.
3938 * The dirty limits take into account the node's
3939 * lowmem reserves and high watermark so that kswapd
3940 * should be able to balance it without having to
3941 * write pages from its LRU list.
3943 * XXX: For now, allow allocations to potentially
3944 * exceed the per-node dirty limit in the slowpath
3945 * (spread_dirty_pages unset) before going into reclaim,
3946 * which is important when on a NUMA setup the allowed
3947 * nodes are together not big enough to reach the
3948 * global limit. The proper fix for these situations
3949 * will require awareness of nodes in the
3950 * dirty-throttling and the flusher threads.
3952 if (ac->spread_dirty_pages) {
3953 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3956 if (!node_dirty_ok(zone->zone_pgdat)) {
3957 last_pgdat_dirty_limit = zone->zone_pgdat;
3962 if (no_fallback && nr_online_nodes > 1 &&
3963 zone != ac->preferred_zoneref->zone) {
3967 * If moving to a remote node, retry but allow
3968 * fragmenting fallbacks. Locality is more important
3969 * than fragmentation avoidance.
3971 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3972 if (zone_to_nid(zone) != local_nid) {
3973 alloc_flags &= ~ALLOC_NOFRAGMENT;
3978 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3979 if (!zone_watermark_fast(zone, order, mark,
3980 ac->highest_zoneidx, alloc_flags,
3984 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3986 * Watermark failed for this zone, but see if we can
3987 * grow this zone if it contains deferred pages.
3989 if (static_branch_unlikely(&deferred_pages)) {
3990 if (_deferred_grow_zone(zone, order))
3994 /* Checked here to keep the fast path fast */
3995 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3996 if (alloc_flags & ALLOC_NO_WATERMARKS)
3999 if (!node_reclaim_enabled() ||
4000 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4003 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4005 case NODE_RECLAIM_NOSCAN:
4008 case NODE_RECLAIM_FULL:
4009 /* scanned but unreclaimable */
4012 /* did we reclaim enough */
4013 if (zone_watermark_ok(zone, order, mark,
4014 ac->highest_zoneidx, alloc_flags))
4022 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4023 gfp_mask, alloc_flags, ac->migratetype);
4025 prep_new_page(page, order, gfp_mask, alloc_flags);
4028 * If this is a high-order atomic allocation then check
4029 * if the pageblock should be reserved for the future
4031 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4032 reserve_highatomic_pageblock(page, zone, order);
4036 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4037 /* Try again if zone has deferred pages */
4038 if (static_branch_unlikely(&deferred_pages)) {
4039 if (_deferred_grow_zone(zone, order))
4047 * It's possible on a UMA machine to get through all zones that are
4048 * fragmented. If avoiding fragmentation, reset and try again.
4051 alloc_flags &= ~ALLOC_NOFRAGMENT;
4058 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4060 unsigned int filter = SHOW_MEM_FILTER_NODES;
4063 * This documents exceptions given to allocations in certain
4064 * contexts that are allowed to allocate outside current's set
4067 if (!(gfp_mask & __GFP_NOMEMALLOC))
4068 if (tsk_is_oom_victim(current) ||
4069 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4070 filter &= ~SHOW_MEM_FILTER_NODES;
4071 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4072 filter &= ~SHOW_MEM_FILTER_NODES;
4074 show_mem(filter, nodemask);
4077 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4079 struct va_format vaf;
4081 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4083 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4086 va_start(args, fmt);
4089 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4090 current->comm, &vaf, gfp_mask, &gfp_mask,
4091 nodemask_pr_args(nodemask));
4094 cpuset_print_current_mems_allowed();
4097 warn_alloc_show_mem(gfp_mask, nodemask);
4100 static inline struct page *
4101 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4102 unsigned int alloc_flags,
4103 const struct alloc_context *ac)
4107 page = get_page_from_freelist(gfp_mask, order,
4108 alloc_flags|ALLOC_CPUSET, ac);
4110 * fallback to ignore cpuset restriction if our nodes
4114 page = get_page_from_freelist(gfp_mask, order,
4120 static inline struct page *
4121 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4122 const struct alloc_context *ac, unsigned long *did_some_progress)
4124 struct oom_control oc = {
4125 .zonelist = ac->zonelist,
4126 .nodemask = ac->nodemask,
4128 .gfp_mask = gfp_mask,
4133 *did_some_progress = 0;
4136 * Acquire the oom lock. If that fails, somebody else is
4137 * making progress for us.
4139 if (!mutex_trylock(&oom_lock)) {
4140 *did_some_progress = 1;
4141 schedule_timeout_uninterruptible(1);
4146 * Go through the zonelist yet one more time, keep very high watermark
4147 * here, this is only to catch a parallel oom killing, we must fail if
4148 * we're still under heavy pressure. But make sure that this reclaim
4149 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4150 * allocation which will never fail due to oom_lock already held.
4152 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4153 ~__GFP_DIRECT_RECLAIM, order,
4154 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4158 /* Coredumps can quickly deplete all memory reserves */
4159 if (current->flags & PF_DUMPCORE)
4161 /* The OOM killer will not help higher order allocs */
4162 if (order > PAGE_ALLOC_COSTLY_ORDER)
4165 * We have already exhausted all our reclaim opportunities without any
4166 * success so it is time to admit defeat. We will skip the OOM killer
4167 * because it is very likely that the caller has a more reasonable
4168 * fallback than shooting a random task.
4170 * The OOM killer may not free memory on a specific node.
4172 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4174 /* The OOM killer does not needlessly kill tasks for lowmem */
4175 if (ac->highest_zoneidx < ZONE_NORMAL)
4177 if (pm_suspended_storage())
4180 * XXX: GFP_NOFS allocations should rather fail than rely on
4181 * other request to make a forward progress.
4182 * We are in an unfortunate situation where out_of_memory cannot
4183 * do much for this context but let's try it to at least get
4184 * access to memory reserved if the current task is killed (see
4185 * out_of_memory). Once filesystems are ready to handle allocation
4186 * failures more gracefully we should just bail out here.
4189 /* Exhausted what can be done so it's blame time */
4190 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4191 *did_some_progress = 1;
4194 * Help non-failing allocations by giving them access to memory
4197 if (gfp_mask & __GFP_NOFAIL)
4198 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4199 ALLOC_NO_WATERMARKS, ac);
4202 mutex_unlock(&oom_lock);
4207 * Maximum number of compaction retries with a progress before OOM
4208 * killer is consider as the only way to move forward.
4210 #define MAX_COMPACT_RETRIES 16
4212 #ifdef CONFIG_COMPACTION
4213 /* Try memory compaction for high-order allocations before reclaim */
4214 static struct page *
4215 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4216 unsigned int alloc_flags, const struct alloc_context *ac,
4217 enum compact_priority prio, enum compact_result *compact_result)
4219 struct page *page = NULL;
4220 unsigned long pflags;
4221 unsigned int noreclaim_flag;
4226 psi_memstall_enter(&pflags);
4227 noreclaim_flag = memalloc_noreclaim_save();
4229 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4232 memalloc_noreclaim_restore(noreclaim_flag);
4233 psi_memstall_leave(&pflags);
4235 if (*compact_result == COMPACT_SKIPPED)
4238 * At least in one zone compaction wasn't deferred or skipped, so let's
4239 * count a compaction stall
4241 count_vm_event(COMPACTSTALL);
4243 /* Prep a captured page if available */
4245 prep_new_page(page, order, gfp_mask, alloc_flags);
4247 /* Try get a page from the freelist if available */
4249 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4252 struct zone *zone = page_zone(page);
4254 zone->compact_blockskip_flush = false;
4255 compaction_defer_reset(zone, order, true);
4256 count_vm_event(COMPACTSUCCESS);
4261 * It's bad if compaction run occurs and fails. The most likely reason
4262 * is that pages exist, but not enough to satisfy watermarks.
4264 count_vm_event(COMPACTFAIL);
4272 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4273 enum compact_result compact_result,
4274 enum compact_priority *compact_priority,
4275 int *compaction_retries)
4277 int max_retries = MAX_COMPACT_RETRIES;
4280 int retries = *compaction_retries;
4281 enum compact_priority priority = *compact_priority;
4286 if (fatal_signal_pending(current))
4289 if (compaction_made_progress(compact_result))
4290 (*compaction_retries)++;
4293 * compaction considers all the zone as desperately out of memory
4294 * so it doesn't really make much sense to retry except when the
4295 * failure could be caused by insufficient priority
4297 if (compaction_failed(compact_result))
4298 goto check_priority;
4301 * compaction was skipped because there are not enough order-0 pages
4302 * to work with, so we retry only if it looks like reclaim can help.
4304 if (compaction_needs_reclaim(compact_result)) {
4305 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4310 * make sure the compaction wasn't deferred or didn't bail out early
4311 * due to locks contention before we declare that we should give up.
4312 * But the next retry should use a higher priority if allowed, so
4313 * we don't just keep bailing out endlessly.
4315 if (compaction_withdrawn(compact_result)) {
4316 goto check_priority;
4320 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4321 * costly ones because they are de facto nofail and invoke OOM
4322 * killer to move on while costly can fail and users are ready
4323 * to cope with that. 1/4 retries is rather arbitrary but we
4324 * would need much more detailed feedback from compaction to
4325 * make a better decision.
4327 if (order > PAGE_ALLOC_COSTLY_ORDER)
4329 if (*compaction_retries <= max_retries) {
4335 * Make sure there are attempts at the highest priority if we exhausted
4336 * all retries or failed at the lower priorities.
4339 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4340 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4342 if (*compact_priority > min_priority) {
4343 (*compact_priority)--;
4344 *compaction_retries = 0;
4348 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4352 static inline struct page *
4353 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4354 unsigned int alloc_flags, const struct alloc_context *ac,
4355 enum compact_priority prio, enum compact_result *compact_result)
4357 *compact_result = COMPACT_SKIPPED;
4362 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4363 enum compact_result compact_result,
4364 enum compact_priority *compact_priority,
4365 int *compaction_retries)
4370 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4374 * There are setups with compaction disabled which would prefer to loop
4375 * inside the allocator rather than hit the oom killer prematurely.
4376 * Let's give them a good hope and keep retrying while the order-0
4377 * watermarks are OK.
4379 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4380 ac->highest_zoneidx, ac->nodemask) {
4381 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4382 ac->highest_zoneidx, alloc_flags))
4387 #endif /* CONFIG_COMPACTION */
4389 #ifdef CONFIG_LOCKDEP
4390 static struct lockdep_map __fs_reclaim_map =
4391 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4393 static bool __need_reclaim(gfp_t gfp_mask)
4395 /* no reclaim without waiting on it */
4396 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4399 /* this guy won't enter reclaim */
4400 if (current->flags & PF_MEMALLOC)
4403 if (gfp_mask & __GFP_NOLOCKDEP)
4409 void __fs_reclaim_acquire(void)
4411 lock_map_acquire(&__fs_reclaim_map);
4414 void __fs_reclaim_release(void)
4416 lock_map_release(&__fs_reclaim_map);
4419 void fs_reclaim_acquire(gfp_t gfp_mask)
4421 gfp_mask = current_gfp_context(gfp_mask);
4423 if (__need_reclaim(gfp_mask)) {
4424 if (gfp_mask & __GFP_FS)
4425 __fs_reclaim_acquire();
4427 #ifdef CONFIG_MMU_NOTIFIER
4428 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4429 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4434 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4436 void fs_reclaim_release(gfp_t gfp_mask)
4438 gfp_mask = current_gfp_context(gfp_mask);
4440 if (__need_reclaim(gfp_mask)) {
4441 if (gfp_mask & __GFP_FS)
4442 __fs_reclaim_release();
4445 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4448 /* Perform direct synchronous page reclaim */
4449 static unsigned long
4450 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4451 const struct alloc_context *ac)
4453 unsigned int noreclaim_flag;
4454 unsigned long pflags, progress;
4458 /* We now go into synchronous reclaim */
4459 cpuset_memory_pressure_bump();
4460 psi_memstall_enter(&pflags);
4461 fs_reclaim_acquire(gfp_mask);
4462 noreclaim_flag = memalloc_noreclaim_save();
4464 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4467 memalloc_noreclaim_restore(noreclaim_flag);
4468 fs_reclaim_release(gfp_mask);
4469 psi_memstall_leave(&pflags);
4476 /* The really slow allocator path where we enter direct reclaim */
4477 static inline struct page *
4478 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4479 unsigned int alloc_flags, const struct alloc_context *ac,
4480 unsigned long *did_some_progress)
4482 struct page *page = NULL;
4483 bool drained = false;
4485 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4486 if (unlikely(!(*did_some_progress)))
4490 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4493 * If an allocation failed after direct reclaim, it could be because
4494 * pages are pinned on the per-cpu lists or in high alloc reserves.
4495 * Shrink them and try again
4497 if (!page && !drained) {
4498 unreserve_highatomic_pageblock(ac, false);
4499 drain_all_pages(NULL);
4507 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4508 const struct alloc_context *ac)
4512 pg_data_t *last_pgdat = NULL;
4513 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4515 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4517 if (last_pgdat != zone->zone_pgdat)
4518 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4519 last_pgdat = zone->zone_pgdat;
4523 static inline unsigned int
4524 gfp_to_alloc_flags(gfp_t gfp_mask)
4526 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4529 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4530 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4531 * to save two branches.
4533 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4534 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4537 * The caller may dip into page reserves a bit more if the caller
4538 * cannot run direct reclaim, or if the caller has realtime scheduling
4539 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4540 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4542 alloc_flags |= (__force int)
4543 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4545 if (gfp_mask & __GFP_ATOMIC) {
4547 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4548 * if it can't schedule.
4550 if (!(gfp_mask & __GFP_NOMEMALLOC))
4551 alloc_flags |= ALLOC_HARDER;
4553 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4554 * comment for __cpuset_node_allowed().
4556 alloc_flags &= ~ALLOC_CPUSET;
4557 } else if (unlikely(rt_task(current)) && !in_interrupt())
4558 alloc_flags |= ALLOC_HARDER;
4560 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4565 static bool oom_reserves_allowed(struct task_struct *tsk)
4567 if (!tsk_is_oom_victim(tsk))
4571 * !MMU doesn't have oom reaper so give access to memory reserves
4572 * only to the thread with TIF_MEMDIE set
4574 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4581 * Distinguish requests which really need access to full memory
4582 * reserves from oom victims which can live with a portion of it
4584 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4586 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4588 if (gfp_mask & __GFP_MEMALLOC)
4589 return ALLOC_NO_WATERMARKS;
4590 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4591 return ALLOC_NO_WATERMARKS;
4592 if (!in_interrupt()) {
4593 if (current->flags & PF_MEMALLOC)
4594 return ALLOC_NO_WATERMARKS;
4595 else if (oom_reserves_allowed(current))
4602 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4604 return !!__gfp_pfmemalloc_flags(gfp_mask);
4608 * Checks whether it makes sense to retry the reclaim to make a forward progress
4609 * for the given allocation request.
4611 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4612 * without success, or when we couldn't even meet the watermark if we
4613 * reclaimed all remaining pages on the LRU lists.
4615 * Returns true if a retry is viable or false to enter the oom path.
4618 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4619 struct alloc_context *ac, int alloc_flags,
4620 bool did_some_progress, int *no_progress_loops)
4627 * Costly allocations might have made a progress but this doesn't mean
4628 * their order will become available due to high fragmentation so
4629 * always increment the no progress counter for them
4631 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4632 *no_progress_loops = 0;
4634 (*no_progress_loops)++;
4637 * Make sure we converge to OOM if we cannot make any progress
4638 * several times in the row.
4640 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4641 /* Before OOM, exhaust highatomic_reserve */
4642 return unreserve_highatomic_pageblock(ac, true);
4646 * Keep reclaiming pages while there is a chance this will lead
4647 * somewhere. If none of the target zones can satisfy our allocation
4648 * request even if all reclaimable pages are considered then we are
4649 * screwed and have to go OOM.
4651 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4652 ac->highest_zoneidx, ac->nodemask) {
4653 unsigned long available;
4654 unsigned long reclaimable;
4655 unsigned long min_wmark = min_wmark_pages(zone);
4658 available = reclaimable = zone_reclaimable_pages(zone);
4659 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4662 * Would the allocation succeed if we reclaimed all
4663 * reclaimable pages?
4665 wmark = __zone_watermark_ok(zone, order, min_wmark,
4666 ac->highest_zoneidx, alloc_flags, available);
4667 trace_reclaim_retry_zone(z, order, reclaimable,
4668 available, min_wmark, *no_progress_loops, wmark);
4671 * If we didn't make any progress and have a lot of
4672 * dirty + writeback pages then we should wait for
4673 * an IO to complete to slow down the reclaim and
4674 * prevent from pre mature OOM
4676 if (!did_some_progress) {
4677 unsigned long write_pending;
4679 write_pending = zone_page_state_snapshot(zone,
4680 NR_ZONE_WRITE_PENDING);
4682 if (2 * write_pending > reclaimable) {
4683 congestion_wait(BLK_RW_ASYNC, HZ/10);
4695 * Memory allocation/reclaim might be called from a WQ context and the
4696 * current implementation of the WQ concurrency control doesn't
4697 * recognize that a particular WQ is congested if the worker thread is
4698 * looping without ever sleeping. Therefore we have to do a short sleep
4699 * here rather than calling cond_resched().
4701 if (current->flags & PF_WQ_WORKER)
4702 schedule_timeout_uninterruptible(1);
4709 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4712 * It's possible that cpuset's mems_allowed and the nodemask from
4713 * mempolicy don't intersect. This should be normally dealt with by
4714 * policy_nodemask(), but it's possible to race with cpuset update in
4715 * such a way the check therein was true, and then it became false
4716 * before we got our cpuset_mems_cookie here.
4717 * This assumes that for all allocations, ac->nodemask can come only
4718 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4719 * when it does not intersect with the cpuset restrictions) or the
4720 * caller can deal with a violated nodemask.
4722 if (cpusets_enabled() && ac->nodemask &&
4723 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4724 ac->nodemask = NULL;
4729 * When updating a task's mems_allowed or mempolicy nodemask, it is
4730 * possible to race with parallel threads in such a way that our
4731 * allocation can fail while the mask is being updated. If we are about
4732 * to fail, check if the cpuset changed during allocation and if so,
4735 if (read_mems_allowed_retry(cpuset_mems_cookie))
4741 static inline struct page *
4742 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4743 struct alloc_context *ac)
4745 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4746 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4747 struct page *page = NULL;
4748 unsigned int alloc_flags;
4749 unsigned long did_some_progress;
4750 enum compact_priority compact_priority;
4751 enum compact_result compact_result;
4752 int compaction_retries;
4753 int no_progress_loops;
4754 unsigned int cpuset_mems_cookie;
4758 * We also sanity check to catch abuse of atomic reserves being used by
4759 * callers that are not in atomic context.
4761 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4762 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4763 gfp_mask &= ~__GFP_ATOMIC;
4766 compaction_retries = 0;
4767 no_progress_loops = 0;
4768 compact_priority = DEF_COMPACT_PRIORITY;
4769 cpuset_mems_cookie = read_mems_allowed_begin();
4772 * The fast path uses conservative alloc_flags to succeed only until
4773 * kswapd needs to be woken up, and to avoid the cost of setting up
4774 * alloc_flags precisely. So we do that now.
4776 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4779 * We need to recalculate the starting point for the zonelist iterator
4780 * because we might have used different nodemask in the fast path, or
4781 * there was a cpuset modification and we are retrying - otherwise we
4782 * could end up iterating over non-eligible zones endlessly.
4784 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4785 ac->highest_zoneidx, ac->nodemask);
4786 if (!ac->preferred_zoneref->zone)
4789 if (alloc_flags & ALLOC_KSWAPD)
4790 wake_all_kswapds(order, gfp_mask, ac);
4793 * The adjusted alloc_flags might result in immediate success, so try
4796 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4801 * For costly allocations, try direct compaction first, as it's likely
4802 * that we have enough base pages and don't need to reclaim. For non-
4803 * movable high-order allocations, do that as well, as compaction will
4804 * try prevent permanent fragmentation by migrating from blocks of the
4806 * Don't try this for allocations that are allowed to ignore
4807 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4809 if (can_direct_reclaim &&
4811 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4812 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4813 page = __alloc_pages_direct_compact(gfp_mask, order,
4815 INIT_COMPACT_PRIORITY,
4821 * Checks for costly allocations with __GFP_NORETRY, which
4822 * includes some THP page fault allocations
4824 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4826 * If allocating entire pageblock(s) and compaction
4827 * failed because all zones are below low watermarks
4828 * or is prohibited because it recently failed at this
4829 * order, fail immediately unless the allocator has
4830 * requested compaction and reclaim retry.
4833 * - potentially very expensive because zones are far
4834 * below their low watermarks or this is part of very
4835 * bursty high order allocations,
4836 * - not guaranteed to help because isolate_freepages()
4837 * may not iterate over freed pages as part of its
4839 * - unlikely to make entire pageblocks free on its
4842 if (compact_result == COMPACT_SKIPPED ||
4843 compact_result == COMPACT_DEFERRED)
4847 * Looks like reclaim/compaction is worth trying, but
4848 * sync compaction could be very expensive, so keep
4849 * using async compaction.
4851 compact_priority = INIT_COMPACT_PRIORITY;
4856 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4857 if (alloc_flags & ALLOC_KSWAPD)
4858 wake_all_kswapds(order, gfp_mask, ac);
4860 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4862 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
4865 * Reset the nodemask and zonelist iterators if memory policies can be
4866 * ignored. These allocations are high priority and system rather than
4869 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4870 ac->nodemask = NULL;
4871 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4872 ac->highest_zoneidx, ac->nodemask);
4875 /* Attempt with potentially adjusted zonelist and alloc_flags */
4876 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4880 /* Caller is not willing to reclaim, we can't balance anything */
4881 if (!can_direct_reclaim)
4884 /* Avoid recursion of direct reclaim */
4885 if (current->flags & PF_MEMALLOC)
4888 /* Try direct reclaim and then allocating */
4889 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4890 &did_some_progress);
4894 /* Try direct compaction and then allocating */
4895 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4896 compact_priority, &compact_result);
4900 /* Do not loop if specifically requested */
4901 if (gfp_mask & __GFP_NORETRY)
4905 * Do not retry costly high order allocations unless they are
4906 * __GFP_RETRY_MAYFAIL
4908 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4911 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4912 did_some_progress > 0, &no_progress_loops))
4916 * It doesn't make any sense to retry for the compaction if the order-0
4917 * reclaim is not able to make any progress because the current
4918 * implementation of the compaction depends on the sufficient amount
4919 * of free memory (see __compaction_suitable)
4921 if (did_some_progress > 0 &&
4922 should_compact_retry(ac, order, alloc_flags,
4923 compact_result, &compact_priority,
4924 &compaction_retries))
4928 /* Deal with possible cpuset update races before we start OOM killing */
4929 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4932 /* Reclaim has failed us, start killing things */
4933 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4937 /* Avoid allocations with no watermarks from looping endlessly */
4938 if (tsk_is_oom_victim(current) &&
4939 (alloc_flags & ALLOC_OOM ||
4940 (gfp_mask & __GFP_NOMEMALLOC)))
4943 /* Retry as long as the OOM killer is making progress */
4944 if (did_some_progress) {
4945 no_progress_loops = 0;
4950 /* Deal with possible cpuset update races before we fail */
4951 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4955 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4958 if (gfp_mask & __GFP_NOFAIL) {
4960 * All existing users of the __GFP_NOFAIL are blockable, so warn
4961 * of any new users that actually require GFP_NOWAIT
4963 if (WARN_ON_ONCE(!can_direct_reclaim))
4967 * PF_MEMALLOC request from this context is rather bizarre
4968 * because we cannot reclaim anything and only can loop waiting
4969 * for somebody to do a work for us
4971 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4974 * non failing costly orders are a hard requirement which we
4975 * are not prepared for much so let's warn about these users
4976 * so that we can identify them and convert them to something
4979 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4982 * Help non-failing allocations by giving them access to memory
4983 * reserves but do not use ALLOC_NO_WATERMARKS because this
4984 * could deplete whole memory reserves which would just make
4985 * the situation worse
4987 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4995 warn_alloc(gfp_mask, ac->nodemask,
4996 "page allocation failure: order:%u", order);
5001 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5002 int preferred_nid, nodemask_t *nodemask,
5003 struct alloc_context *ac, gfp_t *alloc_gfp,
5004 unsigned int *alloc_flags)
5006 ac->highest_zoneidx = gfp_zone(gfp_mask);
5007 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5008 ac->nodemask = nodemask;
5009 ac->migratetype = gfp_migratetype(gfp_mask);
5011 if (cpusets_enabled()) {
5012 *alloc_gfp |= __GFP_HARDWALL;
5014 * When we are in the interrupt context, it is irrelevant
5015 * to the current task context. It means that any node ok.
5017 if (!in_interrupt() && !ac->nodemask)
5018 ac->nodemask = &cpuset_current_mems_allowed;
5020 *alloc_flags |= ALLOC_CPUSET;
5023 fs_reclaim_acquire(gfp_mask);
5024 fs_reclaim_release(gfp_mask);
5026 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5028 if (should_fail_alloc_page(gfp_mask, order))
5031 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5033 /* Dirty zone balancing only done in the fast path */
5034 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5037 * The preferred zone is used for statistics but crucially it is
5038 * also used as the starting point for the zonelist iterator. It
5039 * may get reset for allocations that ignore memory policies.
5041 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5042 ac->highest_zoneidx, ac->nodemask);
5048 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5049 * @gfp: GFP flags for the allocation
5050 * @preferred_nid: The preferred NUMA node ID to allocate from
5051 * @nodemask: Set of nodes to allocate from, may be NULL
5052 * @nr_pages: The number of pages desired on the list or array
5053 * @page_list: Optional list to store the allocated pages
5054 * @page_array: Optional array to store the pages
5056 * This is a batched version of the page allocator that attempts to
5057 * allocate nr_pages quickly. Pages are added to page_list if page_list
5058 * is not NULL, otherwise it is assumed that the page_array is valid.
5060 * For lists, nr_pages is the number of pages that should be allocated.
5062 * For arrays, only NULL elements are populated with pages and nr_pages
5063 * is the maximum number of pages that will be stored in the array.
5065 * Returns the number of pages on the list or array.
5067 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5068 nodemask_t *nodemask, int nr_pages,
5069 struct list_head *page_list,
5070 struct page **page_array)
5073 unsigned long flags;
5076 struct per_cpu_pages *pcp;
5077 struct list_head *pcp_list;
5078 struct alloc_context ac;
5080 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5081 int nr_populated = 0, nr_account = 0;
5083 if (unlikely(nr_pages <= 0))
5087 * Skip populated array elements to determine if any pages need
5088 * to be allocated before disabling IRQs.
5090 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5093 /* Already populated array? */
5094 if (unlikely(page_array && nr_pages - nr_populated == 0))
5095 return nr_populated;
5097 /* Use the single page allocator for one page. */
5098 if (nr_pages - nr_populated == 1)
5101 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5102 gfp &= gfp_allowed_mask;
5104 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5108 /* Find an allowed local zone that meets the low watermark. */
5109 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5112 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5113 !__cpuset_zone_allowed(zone, gfp)) {
5117 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5118 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5122 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5123 if (zone_watermark_fast(zone, 0, mark,
5124 zonelist_zone_idx(ac.preferred_zoneref),
5125 alloc_flags, gfp)) {
5131 * If there are no allowed local zones that meets the watermarks then
5132 * try to allocate a single page and reclaim if necessary.
5134 if (unlikely(!zone))
5137 /* Attempt the batch allocation */
5138 local_lock_irqsave(&pagesets.lock, flags);
5139 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5140 pcp_list = &pcp->lists[ac.migratetype];
5142 while (nr_populated < nr_pages) {
5144 /* Skip existing pages */
5145 if (page_array && page_array[nr_populated]) {
5150 page = __rmqueue_pcplist(zone, ac.migratetype, alloc_flags,
5152 if (unlikely(!page)) {
5153 /* Try and get at least one page */
5160 prep_new_page(page, 0, gfp, 0);
5162 list_add(&page->lru, page_list);
5164 page_array[nr_populated] = page;
5168 local_unlock_irqrestore(&pagesets.lock, flags);
5170 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5171 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5173 return nr_populated;
5176 local_unlock_irqrestore(&pagesets.lock, flags);
5179 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5182 list_add(&page->lru, page_list);
5184 page_array[nr_populated] = page;
5188 return nr_populated;
5190 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5193 * This is the 'heart' of the zoned buddy allocator.
5195 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5196 nodemask_t *nodemask)
5199 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5200 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5201 struct alloc_context ac = { };
5204 * There are several places where we assume that the order value is sane
5205 * so bail out early if the request is out of bound.
5207 if (unlikely(order >= MAX_ORDER)) {
5208 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5212 gfp &= gfp_allowed_mask;
5214 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5215 * resp. GFP_NOIO which has to be inherited for all allocation requests
5216 * from a particular context which has been marked by
5217 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5218 * movable zones are not used during allocation.
5220 gfp = current_gfp_context(gfp);
5222 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5223 &alloc_gfp, &alloc_flags))
5227 * Forbid the first pass from falling back to types that fragment
5228 * memory until all local zones are considered.
5230 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5232 /* First allocation attempt */
5233 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5238 ac.spread_dirty_pages = false;
5241 * Restore the original nodemask if it was potentially replaced with
5242 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5244 ac.nodemask = nodemask;
5246 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5249 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5250 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5251 __free_pages(page, order);
5255 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5259 EXPORT_SYMBOL(__alloc_pages);
5262 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5263 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5264 * you need to access high mem.
5266 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5270 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5273 return (unsigned long) page_address(page);
5275 EXPORT_SYMBOL(__get_free_pages);
5277 unsigned long get_zeroed_page(gfp_t gfp_mask)
5279 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5281 EXPORT_SYMBOL(get_zeroed_page);
5283 static inline void free_the_page(struct page *page, unsigned int order)
5285 if (order == 0) /* Via pcp? */
5286 free_unref_page(page);
5288 __free_pages_ok(page, order, FPI_NONE);
5292 * __free_pages - Free pages allocated with alloc_pages().
5293 * @page: The page pointer returned from alloc_pages().
5294 * @order: The order of the allocation.
5296 * This function can free multi-page allocations that are not compound
5297 * pages. It does not check that the @order passed in matches that of
5298 * the allocation, so it is easy to leak memory. Freeing more memory
5299 * than was allocated will probably emit a warning.
5301 * If the last reference to this page is speculative, it will be released
5302 * by put_page() which only frees the first page of a non-compound
5303 * allocation. To prevent the remaining pages from being leaked, we free
5304 * the subsequent pages here. If you want to use the page's reference
5305 * count to decide when to free the allocation, you should allocate a
5306 * compound page, and use put_page() instead of __free_pages().
5308 * Context: May be called in interrupt context or while holding a normal
5309 * spinlock, but not in NMI context or while holding a raw spinlock.
5311 void __free_pages(struct page *page, unsigned int order)
5313 if (put_page_testzero(page))
5314 free_the_page(page, order);
5315 else if (!PageHead(page))
5317 free_the_page(page + (1 << order), order);
5319 EXPORT_SYMBOL(__free_pages);
5321 void free_pages(unsigned long addr, unsigned int order)
5324 VM_BUG_ON(!virt_addr_valid((void *)addr));
5325 __free_pages(virt_to_page((void *)addr), order);
5329 EXPORT_SYMBOL(free_pages);
5333 * An arbitrary-length arbitrary-offset area of memory which resides
5334 * within a 0 or higher order page. Multiple fragments within that page
5335 * are individually refcounted, in the page's reference counter.
5337 * The page_frag functions below provide a simple allocation framework for
5338 * page fragments. This is used by the network stack and network device
5339 * drivers to provide a backing region of memory for use as either an
5340 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5342 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5345 struct page *page = NULL;
5346 gfp_t gfp = gfp_mask;
5348 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5349 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5351 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5352 PAGE_FRAG_CACHE_MAX_ORDER);
5353 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5355 if (unlikely(!page))
5356 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5358 nc->va = page ? page_address(page) : NULL;
5363 void __page_frag_cache_drain(struct page *page, unsigned int count)
5365 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5367 if (page_ref_sub_and_test(page, count))
5368 free_the_page(page, compound_order(page));
5370 EXPORT_SYMBOL(__page_frag_cache_drain);
5372 void *page_frag_alloc_align(struct page_frag_cache *nc,
5373 unsigned int fragsz, gfp_t gfp_mask,
5374 unsigned int align_mask)
5376 unsigned int size = PAGE_SIZE;
5380 if (unlikely(!nc->va)) {
5382 page = __page_frag_cache_refill(nc, gfp_mask);
5386 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5387 /* if size can vary use size else just use PAGE_SIZE */
5390 /* Even if we own the page, we do not use atomic_set().
5391 * This would break get_page_unless_zero() users.
5393 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5395 /* reset page count bias and offset to start of new frag */
5396 nc->pfmemalloc = page_is_pfmemalloc(page);
5397 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5401 offset = nc->offset - fragsz;
5402 if (unlikely(offset < 0)) {
5403 page = virt_to_page(nc->va);
5405 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5408 if (unlikely(nc->pfmemalloc)) {
5409 free_the_page(page, compound_order(page));
5413 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5414 /* if size can vary use size else just use PAGE_SIZE */
5417 /* OK, page count is 0, we can safely set it */
5418 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5420 /* reset page count bias and offset to start of new frag */
5421 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5422 offset = size - fragsz;
5426 offset &= align_mask;
5427 nc->offset = offset;
5429 return nc->va + offset;
5431 EXPORT_SYMBOL(page_frag_alloc_align);
5434 * Frees a page fragment allocated out of either a compound or order 0 page.
5436 void page_frag_free(void *addr)
5438 struct page *page = virt_to_head_page(addr);
5440 if (unlikely(put_page_testzero(page)))
5441 free_the_page(page, compound_order(page));
5443 EXPORT_SYMBOL(page_frag_free);
5445 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5449 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5450 unsigned long used = addr + PAGE_ALIGN(size);
5452 split_page(virt_to_page((void *)addr), order);
5453 while (used < alloc_end) {
5458 return (void *)addr;
5462 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5463 * @size: the number of bytes to allocate
5464 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5466 * This function is similar to alloc_pages(), except that it allocates the
5467 * minimum number of pages to satisfy the request. alloc_pages() can only
5468 * allocate memory in power-of-two pages.
5470 * This function is also limited by MAX_ORDER.
5472 * Memory allocated by this function must be released by free_pages_exact().
5474 * Return: pointer to the allocated area or %NULL in case of error.
5476 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5478 unsigned int order = get_order(size);
5481 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5482 gfp_mask &= ~__GFP_COMP;
5484 addr = __get_free_pages(gfp_mask, order);
5485 return make_alloc_exact(addr, order, size);
5487 EXPORT_SYMBOL(alloc_pages_exact);
5490 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5492 * @nid: the preferred node ID where memory should be allocated
5493 * @size: the number of bytes to allocate
5494 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5496 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5499 * Return: pointer to the allocated area or %NULL in case of error.
5501 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5503 unsigned int order = get_order(size);
5506 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5507 gfp_mask &= ~__GFP_COMP;
5509 p = alloc_pages_node(nid, gfp_mask, order);
5512 return make_alloc_exact((unsigned long)page_address(p), order, size);
5516 * free_pages_exact - release memory allocated via alloc_pages_exact()
5517 * @virt: the value returned by alloc_pages_exact.
5518 * @size: size of allocation, same value as passed to alloc_pages_exact().
5520 * Release the memory allocated by a previous call to alloc_pages_exact.
5522 void free_pages_exact(void *virt, size_t size)
5524 unsigned long addr = (unsigned long)virt;
5525 unsigned long end = addr + PAGE_ALIGN(size);
5527 while (addr < end) {
5532 EXPORT_SYMBOL(free_pages_exact);
5535 * nr_free_zone_pages - count number of pages beyond high watermark
5536 * @offset: The zone index of the highest zone
5538 * nr_free_zone_pages() counts the number of pages which are beyond the
5539 * high watermark within all zones at or below a given zone index. For each
5540 * zone, the number of pages is calculated as:
5542 * nr_free_zone_pages = managed_pages - high_pages
5544 * Return: number of pages beyond high watermark.
5546 static unsigned long nr_free_zone_pages(int offset)
5551 /* Just pick one node, since fallback list is circular */
5552 unsigned long sum = 0;
5554 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5556 for_each_zone_zonelist(zone, z, zonelist, offset) {
5557 unsigned long size = zone_managed_pages(zone);
5558 unsigned long high = high_wmark_pages(zone);
5567 * nr_free_buffer_pages - count number of pages beyond high watermark
5569 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5570 * watermark within ZONE_DMA and ZONE_NORMAL.
5572 * Return: number of pages beyond high watermark within ZONE_DMA and
5575 unsigned long nr_free_buffer_pages(void)
5577 return nr_free_zone_pages(gfp_zone(GFP_USER));
5579 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5581 static inline void show_node(struct zone *zone)
5583 if (IS_ENABLED(CONFIG_NUMA))
5584 printk("Node %d ", zone_to_nid(zone));
5587 long si_mem_available(void)
5590 unsigned long pagecache;
5591 unsigned long wmark_low = 0;
5592 unsigned long pages[NR_LRU_LISTS];
5593 unsigned long reclaimable;
5597 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5598 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5601 wmark_low += low_wmark_pages(zone);
5604 * Estimate the amount of memory available for userspace allocations,
5605 * without causing swapping.
5607 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5610 * Not all the page cache can be freed, otherwise the system will
5611 * start swapping. Assume at least half of the page cache, or the
5612 * low watermark worth of cache, needs to stay.
5614 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5615 pagecache -= min(pagecache / 2, wmark_low);
5616 available += pagecache;
5619 * Part of the reclaimable slab and other kernel memory consists of
5620 * items that are in use, and cannot be freed. Cap this estimate at the
5623 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5624 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5625 available += reclaimable - min(reclaimable / 2, wmark_low);
5631 EXPORT_SYMBOL_GPL(si_mem_available);
5633 void si_meminfo(struct sysinfo *val)
5635 val->totalram = totalram_pages();
5636 val->sharedram = global_node_page_state(NR_SHMEM);
5637 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5638 val->bufferram = nr_blockdev_pages();
5639 val->totalhigh = totalhigh_pages();
5640 val->freehigh = nr_free_highpages();
5641 val->mem_unit = PAGE_SIZE;
5644 EXPORT_SYMBOL(si_meminfo);
5647 void si_meminfo_node(struct sysinfo *val, int nid)
5649 int zone_type; /* needs to be signed */
5650 unsigned long managed_pages = 0;
5651 unsigned long managed_highpages = 0;
5652 unsigned long free_highpages = 0;
5653 pg_data_t *pgdat = NODE_DATA(nid);
5655 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5656 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5657 val->totalram = managed_pages;
5658 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5659 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5660 #ifdef CONFIG_HIGHMEM
5661 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5662 struct zone *zone = &pgdat->node_zones[zone_type];
5664 if (is_highmem(zone)) {
5665 managed_highpages += zone_managed_pages(zone);
5666 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5669 val->totalhigh = managed_highpages;
5670 val->freehigh = free_highpages;
5672 val->totalhigh = managed_highpages;
5673 val->freehigh = free_highpages;
5675 val->mem_unit = PAGE_SIZE;
5680 * Determine whether the node should be displayed or not, depending on whether
5681 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5683 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5685 if (!(flags & SHOW_MEM_FILTER_NODES))
5689 * no node mask - aka implicit memory numa policy. Do not bother with
5690 * the synchronization - read_mems_allowed_begin - because we do not
5691 * have to be precise here.
5694 nodemask = &cpuset_current_mems_allowed;
5696 return !node_isset(nid, *nodemask);
5699 #define K(x) ((x) << (PAGE_SHIFT-10))
5701 static void show_migration_types(unsigned char type)
5703 static const char types[MIGRATE_TYPES] = {
5704 [MIGRATE_UNMOVABLE] = 'U',
5705 [MIGRATE_MOVABLE] = 'M',
5706 [MIGRATE_RECLAIMABLE] = 'E',
5707 [MIGRATE_HIGHATOMIC] = 'H',
5709 [MIGRATE_CMA] = 'C',
5711 #ifdef CONFIG_MEMORY_ISOLATION
5712 [MIGRATE_ISOLATE] = 'I',
5715 char tmp[MIGRATE_TYPES + 1];
5719 for (i = 0; i < MIGRATE_TYPES; i++) {
5720 if (type & (1 << i))
5725 printk(KERN_CONT "(%s) ", tmp);
5729 * Show free area list (used inside shift_scroll-lock stuff)
5730 * We also calculate the percentage fragmentation. We do this by counting the
5731 * memory on each free list with the exception of the first item on the list.
5734 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5737 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5739 unsigned long free_pcp = 0;
5744 for_each_populated_zone(zone) {
5745 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5748 for_each_online_cpu(cpu)
5749 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5752 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5753 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5754 " unevictable:%lu dirty:%lu writeback:%lu\n"
5755 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5756 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5757 " free:%lu free_pcp:%lu free_cma:%lu\n",
5758 global_node_page_state(NR_ACTIVE_ANON),
5759 global_node_page_state(NR_INACTIVE_ANON),
5760 global_node_page_state(NR_ISOLATED_ANON),
5761 global_node_page_state(NR_ACTIVE_FILE),
5762 global_node_page_state(NR_INACTIVE_FILE),
5763 global_node_page_state(NR_ISOLATED_FILE),
5764 global_node_page_state(NR_UNEVICTABLE),
5765 global_node_page_state(NR_FILE_DIRTY),
5766 global_node_page_state(NR_WRITEBACK),
5767 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5768 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5769 global_node_page_state(NR_FILE_MAPPED),
5770 global_node_page_state(NR_SHMEM),
5771 global_node_page_state(NR_PAGETABLE),
5772 global_zone_page_state(NR_BOUNCE),
5773 global_zone_page_state(NR_FREE_PAGES),
5775 global_zone_page_state(NR_FREE_CMA_PAGES));
5777 for_each_online_pgdat(pgdat) {
5778 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5782 " active_anon:%lukB"
5783 " inactive_anon:%lukB"
5784 " active_file:%lukB"
5785 " inactive_file:%lukB"
5786 " unevictable:%lukB"
5787 " isolated(anon):%lukB"
5788 " isolated(file):%lukB"
5793 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5795 " shmem_pmdmapped: %lukB"
5798 " writeback_tmp:%lukB"
5799 " kernel_stack:%lukB"
5800 #ifdef CONFIG_SHADOW_CALL_STACK
5801 " shadow_call_stack:%lukB"
5804 " all_unreclaimable? %s"
5807 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5808 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5809 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5810 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5811 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5812 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5813 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5814 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5815 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5816 K(node_page_state(pgdat, NR_WRITEBACK)),
5817 K(node_page_state(pgdat, NR_SHMEM)),
5818 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5819 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5820 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5821 K(node_page_state(pgdat, NR_ANON_THPS)),
5823 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5824 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5825 #ifdef CONFIG_SHADOW_CALL_STACK
5826 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5828 K(node_page_state(pgdat, NR_PAGETABLE)),
5829 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5833 for_each_populated_zone(zone) {
5836 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5840 for_each_online_cpu(cpu)
5841 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5850 " reserved_highatomic:%luKB"
5851 " active_anon:%lukB"
5852 " inactive_anon:%lukB"
5853 " active_file:%lukB"
5854 " inactive_file:%lukB"
5855 " unevictable:%lukB"
5856 " writepending:%lukB"
5866 K(zone_page_state(zone, NR_FREE_PAGES)),
5867 K(min_wmark_pages(zone)),
5868 K(low_wmark_pages(zone)),
5869 K(high_wmark_pages(zone)),
5870 K(zone->nr_reserved_highatomic),
5871 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5872 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5873 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5874 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5875 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5876 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5877 K(zone->present_pages),
5878 K(zone_managed_pages(zone)),
5879 K(zone_page_state(zone, NR_MLOCK)),
5880 K(zone_page_state(zone, NR_BOUNCE)),
5882 K(this_cpu_read(zone->per_cpu_pageset->count)),
5883 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5884 printk("lowmem_reserve[]:");
5885 for (i = 0; i < MAX_NR_ZONES; i++)
5886 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5887 printk(KERN_CONT "\n");
5890 for_each_populated_zone(zone) {
5892 unsigned long nr[MAX_ORDER], flags, total = 0;
5893 unsigned char types[MAX_ORDER];
5895 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5898 printk(KERN_CONT "%s: ", zone->name);
5900 spin_lock_irqsave(&zone->lock, flags);
5901 for (order = 0; order < MAX_ORDER; order++) {
5902 struct free_area *area = &zone->free_area[order];
5905 nr[order] = area->nr_free;
5906 total += nr[order] << order;
5909 for (type = 0; type < MIGRATE_TYPES; type++) {
5910 if (!free_area_empty(area, type))
5911 types[order] |= 1 << type;
5914 spin_unlock_irqrestore(&zone->lock, flags);
5915 for (order = 0; order < MAX_ORDER; order++) {
5916 printk(KERN_CONT "%lu*%lukB ",
5917 nr[order], K(1UL) << order);
5919 show_migration_types(types[order]);
5921 printk(KERN_CONT "= %lukB\n", K(total));
5924 hugetlb_show_meminfo();
5926 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5928 show_swap_cache_info();
5931 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5933 zoneref->zone = zone;
5934 zoneref->zone_idx = zone_idx(zone);
5938 * Builds allocation fallback zone lists.
5940 * Add all populated zones of a node to the zonelist.
5942 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5945 enum zone_type zone_type = MAX_NR_ZONES;
5950 zone = pgdat->node_zones + zone_type;
5951 if (managed_zone(zone)) {
5952 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5953 check_highest_zone(zone_type);
5955 } while (zone_type);
5962 static int __parse_numa_zonelist_order(char *s)
5965 * We used to support different zonelists modes but they turned
5966 * out to be just not useful. Let's keep the warning in place
5967 * if somebody still use the cmd line parameter so that we do
5968 * not fail it silently
5970 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5971 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5977 char numa_zonelist_order[] = "Node";
5980 * sysctl handler for numa_zonelist_order
5982 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5983 void *buffer, size_t *length, loff_t *ppos)
5986 return __parse_numa_zonelist_order(buffer);
5987 return proc_dostring(table, write, buffer, length, ppos);
5991 #define MAX_NODE_LOAD (nr_online_nodes)
5992 static int node_load[MAX_NUMNODES];
5995 * find_next_best_node - find the next node that should appear in a given node's fallback list
5996 * @node: node whose fallback list we're appending
5997 * @used_node_mask: nodemask_t of already used nodes
5999 * We use a number of factors to determine which is the next node that should
6000 * appear on a given node's fallback list. The node should not have appeared
6001 * already in @node's fallback list, and it should be the next closest node
6002 * according to the distance array (which contains arbitrary distance values
6003 * from each node to each node in the system), and should also prefer nodes
6004 * with no CPUs, since presumably they'll have very little allocation pressure
6005 * on them otherwise.
6007 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6009 static int find_next_best_node(int node, nodemask_t *used_node_mask)
6012 int min_val = INT_MAX;
6013 int best_node = NUMA_NO_NODE;
6015 /* Use the local node if we haven't already */
6016 if (!node_isset(node, *used_node_mask)) {
6017 node_set(node, *used_node_mask);
6021 for_each_node_state(n, N_MEMORY) {
6023 /* Don't want a node to appear more than once */
6024 if (node_isset(n, *used_node_mask))
6027 /* Use the distance array to find the distance */
6028 val = node_distance(node, n);
6030 /* Penalize nodes under us ("prefer the next node") */
6033 /* Give preference to headless and unused nodes */
6034 if (!cpumask_empty(cpumask_of_node(n)))
6035 val += PENALTY_FOR_NODE_WITH_CPUS;
6037 /* Slight preference for less loaded node */
6038 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6039 val += node_load[n];
6041 if (val < min_val) {
6048 node_set(best_node, *used_node_mask);
6055 * Build zonelists ordered by node and zones within node.
6056 * This results in maximum locality--normal zone overflows into local
6057 * DMA zone, if any--but risks exhausting DMA zone.
6059 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6062 struct zoneref *zonerefs;
6065 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6067 for (i = 0; i < nr_nodes; i++) {
6070 pg_data_t *node = NODE_DATA(node_order[i]);
6072 nr_zones = build_zonerefs_node(node, zonerefs);
6073 zonerefs += nr_zones;
6075 zonerefs->zone = NULL;
6076 zonerefs->zone_idx = 0;
6080 * Build gfp_thisnode zonelists
6082 static void build_thisnode_zonelists(pg_data_t *pgdat)
6084 struct zoneref *zonerefs;
6087 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6088 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6089 zonerefs += nr_zones;
6090 zonerefs->zone = NULL;
6091 zonerefs->zone_idx = 0;
6095 * Build zonelists ordered by zone and nodes within zones.
6096 * This results in conserving DMA zone[s] until all Normal memory is
6097 * exhausted, but results in overflowing to remote node while memory
6098 * may still exist in local DMA zone.
6101 static void build_zonelists(pg_data_t *pgdat)
6103 static int node_order[MAX_NUMNODES];
6104 int node, load, nr_nodes = 0;
6105 nodemask_t used_mask = NODE_MASK_NONE;
6106 int local_node, prev_node;
6108 /* NUMA-aware ordering of nodes */
6109 local_node = pgdat->node_id;
6110 load = nr_online_nodes;
6111 prev_node = local_node;
6113 memset(node_order, 0, sizeof(node_order));
6114 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6116 * We don't want to pressure a particular node.
6117 * So adding penalty to the first node in same
6118 * distance group to make it round-robin.
6120 if (node_distance(local_node, node) !=
6121 node_distance(local_node, prev_node))
6122 node_load[node] = load;
6124 node_order[nr_nodes++] = node;
6129 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6130 build_thisnode_zonelists(pgdat);
6133 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6135 * Return node id of node used for "local" allocations.
6136 * I.e., first node id of first zone in arg node's generic zonelist.
6137 * Used for initializing percpu 'numa_mem', which is used primarily
6138 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6140 int local_memory_node(int node)
6144 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6145 gfp_zone(GFP_KERNEL),
6147 return zone_to_nid(z->zone);
6151 static void setup_min_unmapped_ratio(void);
6152 static void setup_min_slab_ratio(void);
6153 #else /* CONFIG_NUMA */
6155 static void build_zonelists(pg_data_t *pgdat)
6157 int node, local_node;
6158 struct zoneref *zonerefs;
6161 local_node = pgdat->node_id;
6163 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6164 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6165 zonerefs += nr_zones;
6168 * Now we build the zonelist so that it contains the zones
6169 * of all the other nodes.
6170 * We don't want to pressure a particular node, so when
6171 * building the zones for node N, we make sure that the
6172 * zones coming right after the local ones are those from
6173 * node N+1 (modulo N)
6175 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6176 if (!node_online(node))
6178 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6179 zonerefs += nr_zones;
6181 for (node = 0; node < local_node; node++) {
6182 if (!node_online(node))
6184 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6185 zonerefs += nr_zones;
6188 zonerefs->zone = NULL;
6189 zonerefs->zone_idx = 0;
6192 #endif /* CONFIG_NUMA */
6195 * Boot pageset table. One per cpu which is going to be used for all
6196 * zones and all nodes. The parameters will be set in such a way
6197 * that an item put on a list will immediately be handed over to
6198 * the buddy list. This is safe since pageset manipulation is done
6199 * with interrupts disabled.
6201 * The boot_pagesets must be kept even after bootup is complete for
6202 * unused processors and/or zones. They do play a role for bootstrapping
6203 * hotplugged processors.
6205 * zoneinfo_show() and maybe other functions do
6206 * not check if the processor is online before following the pageset pointer.
6207 * Other parts of the kernel may not check if the zone is available.
6209 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6210 /* These effectively disable the pcplists in the boot pageset completely */
6211 #define BOOT_PAGESET_HIGH 0
6212 #define BOOT_PAGESET_BATCH 1
6213 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6214 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6215 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6217 static void __build_all_zonelists(void *data)
6220 int __maybe_unused cpu;
6221 pg_data_t *self = data;
6222 static DEFINE_SPINLOCK(lock);
6227 memset(node_load, 0, sizeof(node_load));
6231 * This node is hotadded and no memory is yet present. So just
6232 * building zonelists is fine - no need to touch other nodes.
6234 if (self && !node_online(self->node_id)) {
6235 build_zonelists(self);
6237 for_each_online_node(nid) {
6238 pg_data_t *pgdat = NODE_DATA(nid);
6240 build_zonelists(pgdat);
6243 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6245 * We now know the "local memory node" for each node--
6246 * i.e., the node of the first zone in the generic zonelist.
6247 * Set up numa_mem percpu variable for on-line cpus. During
6248 * boot, only the boot cpu should be on-line; we'll init the
6249 * secondary cpus' numa_mem as they come on-line. During
6250 * node/memory hotplug, we'll fixup all on-line cpus.
6252 for_each_online_cpu(cpu)
6253 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6260 static noinline void __init
6261 build_all_zonelists_init(void)
6265 __build_all_zonelists(NULL);
6268 * Initialize the boot_pagesets that are going to be used
6269 * for bootstrapping processors. The real pagesets for
6270 * each zone will be allocated later when the per cpu
6271 * allocator is available.
6273 * boot_pagesets are used also for bootstrapping offline
6274 * cpus if the system is already booted because the pagesets
6275 * are needed to initialize allocators on a specific cpu too.
6276 * F.e. the percpu allocator needs the page allocator which
6277 * needs the percpu allocator in order to allocate its pagesets
6278 * (a chicken-egg dilemma).
6280 for_each_possible_cpu(cpu)
6281 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6283 mminit_verify_zonelist();
6284 cpuset_init_current_mems_allowed();
6288 * unless system_state == SYSTEM_BOOTING.
6290 * __ref due to call of __init annotated helper build_all_zonelists_init
6291 * [protected by SYSTEM_BOOTING].
6293 void __ref build_all_zonelists(pg_data_t *pgdat)
6295 unsigned long vm_total_pages;
6297 if (system_state == SYSTEM_BOOTING) {
6298 build_all_zonelists_init();
6300 __build_all_zonelists(pgdat);
6301 /* cpuset refresh routine should be here */
6303 /* Get the number of free pages beyond high watermark in all zones. */
6304 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6306 * Disable grouping by mobility if the number of pages in the
6307 * system is too low to allow the mechanism to work. It would be
6308 * more accurate, but expensive to check per-zone. This check is
6309 * made on memory-hotadd so a system can start with mobility
6310 * disabled and enable it later
6312 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6313 page_group_by_mobility_disabled = 1;
6315 page_group_by_mobility_disabled = 0;
6317 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6319 page_group_by_mobility_disabled ? "off" : "on",
6322 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6326 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6327 static bool __meminit
6328 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6330 static struct memblock_region *r;
6332 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6333 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6334 for_each_mem_region(r) {
6335 if (*pfn < memblock_region_memory_end_pfn(r))
6339 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6340 memblock_is_mirror(r)) {
6341 *pfn = memblock_region_memory_end_pfn(r);
6349 * Initially all pages are reserved - free ones are freed
6350 * up by memblock_free_all() once the early boot process is
6351 * done. Non-atomic initialization, single-pass.
6353 * All aligned pageblocks are initialized to the specified migratetype
6354 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6355 * zone stats (e.g., nr_isolate_pageblock) are touched.
6357 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6358 unsigned long start_pfn, unsigned long zone_end_pfn,
6359 enum meminit_context context,
6360 struct vmem_altmap *altmap, int migratetype)
6362 unsigned long pfn, end_pfn = start_pfn + size;
6365 if (highest_memmap_pfn < end_pfn - 1)
6366 highest_memmap_pfn = end_pfn - 1;
6368 #ifdef CONFIG_ZONE_DEVICE
6370 * Honor reservation requested by the driver for this ZONE_DEVICE
6371 * memory. We limit the total number of pages to initialize to just
6372 * those that might contain the memory mapping. We will defer the
6373 * ZONE_DEVICE page initialization until after we have released
6376 if (zone == ZONE_DEVICE) {
6380 if (start_pfn == altmap->base_pfn)
6381 start_pfn += altmap->reserve;
6382 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6386 for (pfn = start_pfn; pfn < end_pfn; ) {
6388 * There can be holes in boot-time mem_map[]s handed to this
6389 * function. They do not exist on hotplugged memory.
6391 if (context == MEMINIT_EARLY) {
6392 if (overlap_memmap_init(zone, &pfn))
6394 if (defer_init(nid, pfn, zone_end_pfn))
6398 page = pfn_to_page(pfn);
6399 __init_single_page(page, pfn, zone, nid);
6400 if (context == MEMINIT_HOTPLUG)
6401 __SetPageReserved(page);
6404 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6405 * such that unmovable allocations won't be scattered all
6406 * over the place during system boot.
6408 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6409 set_pageblock_migratetype(page, migratetype);
6416 #ifdef CONFIG_ZONE_DEVICE
6417 void __ref memmap_init_zone_device(struct zone *zone,
6418 unsigned long start_pfn,
6419 unsigned long nr_pages,
6420 struct dev_pagemap *pgmap)
6422 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6423 struct pglist_data *pgdat = zone->zone_pgdat;
6424 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6425 unsigned long zone_idx = zone_idx(zone);
6426 unsigned long start = jiffies;
6427 int nid = pgdat->node_id;
6429 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6433 * The call to memmap_init should have already taken care
6434 * of the pages reserved for the memmap, so we can just jump to
6435 * the end of that region and start processing the device pages.
6438 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6439 nr_pages = end_pfn - start_pfn;
6442 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6443 struct page *page = pfn_to_page(pfn);
6445 __init_single_page(page, pfn, zone_idx, nid);
6448 * Mark page reserved as it will need to wait for onlining
6449 * phase for it to be fully associated with a zone.
6451 * We can use the non-atomic __set_bit operation for setting
6452 * the flag as we are still initializing the pages.
6454 __SetPageReserved(page);
6457 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6458 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6459 * ever freed or placed on a driver-private list.
6461 page->pgmap = pgmap;
6462 page->zone_device_data = NULL;
6465 * Mark the block movable so that blocks are reserved for
6466 * movable at startup. This will force kernel allocations
6467 * to reserve their blocks rather than leaking throughout
6468 * the address space during boot when many long-lived
6469 * kernel allocations are made.
6471 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6472 * because this is done early in section_activate()
6474 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6475 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6480 pr_info("%s initialised %lu pages in %ums\n", __func__,
6481 nr_pages, jiffies_to_msecs(jiffies - start));
6485 static void __meminit zone_init_free_lists(struct zone *zone)
6487 unsigned int order, t;
6488 for_each_migratetype_order(order, t) {
6489 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6490 zone->free_area[order].nr_free = 0;
6494 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6496 * Only struct pages that correspond to ranges defined by memblock.memory
6497 * are zeroed and initialized by going through __init_single_page() during
6498 * memmap_init_zone_range().
6500 * But, there could be struct pages that correspond to holes in
6501 * memblock.memory. This can happen because of the following reasons:
6502 * - physical memory bank size is not necessarily the exact multiple of the
6503 * arbitrary section size
6504 * - early reserved memory may not be listed in memblock.memory
6505 * - memory layouts defined with memmap= kernel parameter may not align
6506 * nicely with memmap sections
6508 * Explicitly initialize those struct pages so that:
6509 * - PG_Reserved is set
6510 * - zone and node links point to zone and node that span the page if the
6511 * hole is in the middle of a zone
6512 * - zone and node links point to adjacent zone/node if the hole falls on
6513 * the zone boundary; the pages in such holes will be prepended to the
6514 * zone/node above the hole except for the trailing pages in the last
6515 * section that will be appended to the zone/node below.
6517 static void __init init_unavailable_range(unsigned long spfn,
6524 for (pfn = spfn; pfn < epfn; pfn++) {
6525 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6526 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6527 + pageblock_nr_pages - 1;
6530 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6531 __SetPageReserved(pfn_to_page(pfn));
6536 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6537 node, zone_names[zone], pgcnt);
6540 static inline void init_unavailable_range(unsigned long spfn,
6547 static void __init memmap_init_zone_range(struct zone *zone,
6548 unsigned long start_pfn,
6549 unsigned long end_pfn,
6550 unsigned long *hole_pfn)
6552 unsigned long zone_start_pfn = zone->zone_start_pfn;
6553 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6554 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6556 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6557 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6559 if (start_pfn >= end_pfn)
6562 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6563 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6565 if (*hole_pfn < start_pfn)
6566 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6568 *hole_pfn = end_pfn;
6571 static void __init memmap_init(void)
6573 unsigned long start_pfn, end_pfn;
6574 unsigned long hole_pfn = 0;
6575 int i, j, zone_id, nid;
6577 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6578 struct pglist_data *node = NODE_DATA(nid);
6580 for (j = 0; j < MAX_NR_ZONES; j++) {
6581 struct zone *zone = node->node_zones + j;
6583 if (!populated_zone(zone))
6586 memmap_init_zone_range(zone, start_pfn, end_pfn,
6592 #ifdef CONFIG_SPARSEMEM
6594 * Initialize the memory map for hole in the range [memory_end,
6596 * Append the pages in this hole to the highest zone in the last
6598 * The call to init_unavailable_range() is outside the ifdef to
6599 * silence the compiler warining about zone_id set but not used;
6600 * for FLATMEM it is a nop anyway
6602 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6603 if (hole_pfn < end_pfn)
6605 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6608 static int zone_batchsize(struct zone *zone)
6614 * The per-cpu-pages pools are set to around 1000th of the
6617 batch = zone_managed_pages(zone) / 1024;
6618 /* But no more than a meg. */
6619 if (batch * PAGE_SIZE > 1024 * 1024)
6620 batch = (1024 * 1024) / PAGE_SIZE;
6621 batch /= 4; /* We effectively *= 4 below */
6626 * Clamp the batch to a 2^n - 1 value. Having a power
6627 * of 2 value was found to be more likely to have
6628 * suboptimal cache aliasing properties in some cases.
6630 * For example if 2 tasks are alternately allocating
6631 * batches of pages, one task can end up with a lot
6632 * of pages of one half of the possible page colors
6633 * and the other with pages of the other colors.
6635 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6640 /* The deferral and batching of frees should be suppressed under NOMMU
6643 * The problem is that NOMMU needs to be able to allocate large chunks
6644 * of contiguous memory as there's no hardware page translation to
6645 * assemble apparent contiguous memory from discontiguous pages.
6647 * Queueing large contiguous runs of pages for batching, however,
6648 * causes the pages to actually be freed in smaller chunks. As there
6649 * can be a significant delay between the individual batches being
6650 * recycled, this leads to the once large chunks of space being
6651 * fragmented and becoming unavailable for high-order allocations.
6658 * pcp->high and pcp->batch values are related and generally batch is lower
6659 * than high. They are also related to pcp->count such that count is lower
6660 * than high, and as soon as it reaches high, the pcplist is flushed.
6662 * However, guaranteeing these relations at all times would require e.g. write
6663 * barriers here but also careful usage of read barriers at the read side, and
6664 * thus be prone to error and bad for performance. Thus the update only prevents
6665 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6666 * can cope with those fields changing asynchronously, and fully trust only the
6667 * pcp->count field on the local CPU with interrupts disabled.
6669 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6670 * outside of boot time (or some other assurance that no concurrent updaters
6673 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6674 unsigned long batch)
6676 WRITE_ONCE(pcp->batch, batch);
6677 WRITE_ONCE(pcp->high, high);
6680 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6684 memset(pcp, 0, sizeof(*pcp));
6685 memset(pzstats, 0, sizeof(*pzstats));
6687 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6688 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6691 * Set batch and high values safe for a boot pageset. A true percpu
6692 * pageset's initialization will update them subsequently. Here we don't
6693 * need to be as careful as pageset_update() as nobody can access the
6696 pcp->high = BOOT_PAGESET_HIGH;
6697 pcp->batch = BOOT_PAGESET_BATCH;
6700 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6701 unsigned long batch)
6703 struct per_cpu_pages *pcp;
6706 for_each_possible_cpu(cpu) {
6707 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6708 pageset_update(pcp, high, batch);
6713 * Calculate and set new high and batch values for all per-cpu pagesets of a
6714 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6716 static void zone_set_pageset_high_and_batch(struct zone *zone)
6718 unsigned long new_high, new_batch;
6720 if (percpu_pagelist_fraction) {
6721 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6722 new_batch = max(1UL, new_high / 4);
6723 if ((new_high / 4) > (PAGE_SHIFT * 8))
6724 new_batch = PAGE_SHIFT * 8;
6726 new_batch = zone_batchsize(zone);
6727 new_high = 6 * new_batch;
6728 new_batch = max(1UL, 1 * new_batch);
6731 if (zone->pageset_high == new_high &&
6732 zone->pageset_batch == new_batch)
6735 zone->pageset_high = new_high;
6736 zone->pageset_batch = new_batch;
6738 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6741 void __meminit setup_zone_pageset(struct zone *zone)
6745 /* Size may be 0 on !SMP && !NUMA */
6746 if (sizeof(struct per_cpu_zonestat) > 0)
6747 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6749 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6750 for_each_possible_cpu(cpu) {
6751 struct per_cpu_pages *pcp;
6752 struct per_cpu_zonestat *pzstats;
6754 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6755 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6756 per_cpu_pages_init(pcp, pzstats);
6759 zone_set_pageset_high_and_batch(zone);
6763 * Allocate per cpu pagesets and initialize them.
6764 * Before this call only boot pagesets were available.
6766 void __init setup_per_cpu_pageset(void)
6768 struct pglist_data *pgdat;
6770 int __maybe_unused cpu;
6772 for_each_populated_zone(zone)
6773 setup_zone_pageset(zone);
6777 * Unpopulated zones continue using the boot pagesets.
6778 * The numa stats for these pagesets need to be reset.
6779 * Otherwise, they will end up skewing the stats of
6780 * the nodes these zones are associated with.
6782 for_each_possible_cpu(cpu) {
6783 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
6784 memset(pzstats->vm_numa_event, 0,
6785 sizeof(pzstats->vm_numa_event));
6789 for_each_online_pgdat(pgdat)
6790 pgdat->per_cpu_nodestats =
6791 alloc_percpu(struct per_cpu_nodestat);
6794 static __meminit void zone_pcp_init(struct zone *zone)
6797 * per cpu subsystem is not up at this point. The following code
6798 * relies on the ability of the linker to provide the
6799 * offset of a (static) per cpu variable into the per cpu area.
6801 zone->per_cpu_pageset = &boot_pageset;
6802 zone->per_cpu_zonestats = &boot_zonestats;
6803 zone->pageset_high = BOOT_PAGESET_HIGH;
6804 zone->pageset_batch = BOOT_PAGESET_BATCH;
6806 if (populated_zone(zone))
6807 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
6808 zone->present_pages, zone_batchsize(zone));
6811 void __meminit init_currently_empty_zone(struct zone *zone,
6812 unsigned long zone_start_pfn,
6815 struct pglist_data *pgdat = zone->zone_pgdat;
6816 int zone_idx = zone_idx(zone) + 1;
6818 if (zone_idx > pgdat->nr_zones)
6819 pgdat->nr_zones = zone_idx;
6821 zone->zone_start_pfn = zone_start_pfn;
6823 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6824 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6826 (unsigned long)zone_idx(zone),
6827 zone_start_pfn, (zone_start_pfn + size));
6829 zone_init_free_lists(zone);
6830 zone->initialized = 1;
6834 * get_pfn_range_for_nid - Return the start and end page frames for a node
6835 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6836 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6837 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6839 * It returns the start and end page frame of a node based on information
6840 * provided by memblock_set_node(). If called for a node
6841 * with no available memory, a warning is printed and the start and end
6844 void __init get_pfn_range_for_nid(unsigned int nid,
6845 unsigned long *start_pfn, unsigned long *end_pfn)
6847 unsigned long this_start_pfn, this_end_pfn;
6853 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6854 *start_pfn = min(*start_pfn, this_start_pfn);
6855 *end_pfn = max(*end_pfn, this_end_pfn);
6858 if (*start_pfn == -1UL)
6863 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6864 * assumption is made that zones within a node are ordered in monotonic
6865 * increasing memory addresses so that the "highest" populated zone is used
6867 static void __init find_usable_zone_for_movable(void)
6870 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6871 if (zone_index == ZONE_MOVABLE)
6874 if (arch_zone_highest_possible_pfn[zone_index] >
6875 arch_zone_lowest_possible_pfn[zone_index])
6879 VM_BUG_ON(zone_index == -1);
6880 movable_zone = zone_index;
6884 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6885 * because it is sized independent of architecture. Unlike the other zones,
6886 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6887 * in each node depending on the size of each node and how evenly kernelcore
6888 * is distributed. This helper function adjusts the zone ranges
6889 * provided by the architecture for a given node by using the end of the
6890 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6891 * zones within a node are in order of monotonic increases memory addresses
6893 static void __init adjust_zone_range_for_zone_movable(int nid,
6894 unsigned long zone_type,
6895 unsigned long node_start_pfn,
6896 unsigned long node_end_pfn,
6897 unsigned long *zone_start_pfn,
6898 unsigned long *zone_end_pfn)
6900 /* Only adjust if ZONE_MOVABLE is on this node */
6901 if (zone_movable_pfn[nid]) {
6902 /* Size ZONE_MOVABLE */
6903 if (zone_type == ZONE_MOVABLE) {
6904 *zone_start_pfn = zone_movable_pfn[nid];
6905 *zone_end_pfn = min(node_end_pfn,
6906 arch_zone_highest_possible_pfn[movable_zone]);
6908 /* Adjust for ZONE_MOVABLE starting within this range */
6909 } else if (!mirrored_kernelcore &&
6910 *zone_start_pfn < zone_movable_pfn[nid] &&
6911 *zone_end_pfn > zone_movable_pfn[nid]) {
6912 *zone_end_pfn = zone_movable_pfn[nid];
6914 /* Check if this whole range is within ZONE_MOVABLE */
6915 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6916 *zone_start_pfn = *zone_end_pfn;
6921 * Return the number of pages a zone spans in a node, including holes
6922 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6924 static unsigned long __init zone_spanned_pages_in_node(int nid,
6925 unsigned long zone_type,
6926 unsigned long node_start_pfn,
6927 unsigned long node_end_pfn,
6928 unsigned long *zone_start_pfn,
6929 unsigned long *zone_end_pfn)
6931 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6932 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6933 /* When hotadd a new node from cpu_up(), the node should be empty */
6934 if (!node_start_pfn && !node_end_pfn)
6937 /* Get the start and end of the zone */
6938 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6939 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6940 adjust_zone_range_for_zone_movable(nid, zone_type,
6941 node_start_pfn, node_end_pfn,
6942 zone_start_pfn, zone_end_pfn);
6944 /* Check that this node has pages within the zone's required range */
6945 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6948 /* Move the zone boundaries inside the node if necessary */
6949 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6950 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6952 /* Return the spanned pages */
6953 return *zone_end_pfn - *zone_start_pfn;
6957 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6958 * then all holes in the requested range will be accounted for.
6960 unsigned long __init __absent_pages_in_range(int nid,
6961 unsigned long range_start_pfn,
6962 unsigned long range_end_pfn)
6964 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6965 unsigned long start_pfn, end_pfn;
6968 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6969 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6970 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6971 nr_absent -= end_pfn - start_pfn;
6977 * absent_pages_in_range - Return number of page frames in holes within a range
6978 * @start_pfn: The start PFN to start searching for holes
6979 * @end_pfn: The end PFN to stop searching for holes
6981 * Return: the number of pages frames in memory holes within a range.
6983 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6984 unsigned long end_pfn)
6986 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6989 /* Return the number of page frames in holes in a zone on a node */
6990 static unsigned long __init zone_absent_pages_in_node(int nid,
6991 unsigned long zone_type,
6992 unsigned long node_start_pfn,
6993 unsigned long node_end_pfn)
6995 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6996 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6997 unsigned long zone_start_pfn, zone_end_pfn;
6998 unsigned long nr_absent;
7000 /* When hotadd a new node from cpu_up(), the node should be empty */
7001 if (!node_start_pfn && !node_end_pfn)
7004 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7005 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7007 adjust_zone_range_for_zone_movable(nid, zone_type,
7008 node_start_pfn, node_end_pfn,
7009 &zone_start_pfn, &zone_end_pfn);
7010 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7013 * ZONE_MOVABLE handling.
7014 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7017 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7018 unsigned long start_pfn, end_pfn;
7019 struct memblock_region *r;
7021 for_each_mem_region(r) {
7022 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7023 zone_start_pfn, zone_end_pfn);
7024 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7025 zone_start_pfn, zone_end_pfn);
7027 if (zone_type == ZONE_MOVABLE &&
7028 memblock_is_mirror(r))
7029 nr_absent += end_pfn - start_pfn;
7031 if (zone_type == ZONE_NORMAL &&
7032 !memblock_is_mirror(r))
7033 nr_absent += end_pfn - start_pfn;
7040 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7041 unsigned long node_start_pfn,
7042 unsigned long node_end_pfn)
7044 unsigned long realtotalpages = 0, totalpages = 0;
7047 for (i = 0; i < MAX_NR_ZONES; i++) {
7048 struct zone *zone = pgdat->node_zones + i;
7049 unsigned long zone_start_pfn, zone_end_pfn;
7050 unsigned long spanned, absent;
7051 unsigned long size, real_size;
7053 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7058 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7063 real_size = size - absent;
7066 zone->zone_start_pfn = zone_start_pfn;
7068 zone->zone_start_pfn = 0;
7069 zone->spanned_pages = size;
7070 zone->present_pages = real_size;
7073 realtotalpages += real_size;
7076 pgdat->node_spanned_pages = totalpages;
7077 pgdat->node_present_pages = realtotalpages;
7078 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7081 #ifndef CONFIG_SPARSEMEM
7083 * Calculate the size of the zone->blockflags rounded to an unsigned long
7084 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7085 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7086 * round what is now in bits to nearest long in bits, then return it in
7089 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7091 unsigned long usemapsize;
7093 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7094 usemapsize = roundup(zonesize, pageblock_nr_pages);
7095 usemapsize = usemapsize >> pageblock_order;
7096 usemapsize *= NR_PAGEBLOCK_BITS;
7097 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7099 return usemapsize / 8;
7102 static void __ref setup_usemap(struct zone *zone)
7104 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7105 zone->spanned_pages);
7106 zone->pageblock_flags = NULL;
7108 zone->pageblock_flags =
7109 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7111 if (!zone->pageblock_flags)
7112 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7113 usemapsize, zone->name, zone_to_nid(zone));
7117 static inline void setup_usemap(struct zone *zone) {}
7118 #endif /* CONFIG_SPARSEMEM */
7120 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7122 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7123 void __init set_pageblock_order(void)
7127 /* Check that pageblock_nr_pages has not already been setup */
7128 if (pageblock_order)
7131 if (HPAGE_SHIFT > PAGE_SHIFT)
7132 order = HUGETLB_PAGE_ORDER;
7134 order = MAX_ORDER - 1;
7137 * Assume the largest contiguous order of interest is a huge page.
7138 * This value may be variable depending on boot parameters on IA64 and
7141 pageblock_order = order;
7143 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7146 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7147 * is unused as pageblock_order is set at compile-time. See
7148 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7151 void __init set_pageblock_order(void)
7155 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7157 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7158 unsigned long present_pages)
7160 unsigned long pages = spanned_pages;
7163 * Provide a more accurate estimation if there are holes within
7164 * the zone and SPARSEMEM is in use. If there are holes within the
7165 * zone, each populated memory region may cost us one or two extra
7166 * memmap pages due to alignment because memmap pages for each
7167 * populated regions may not be naturally aligned on page boundary.
7168 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7170 if (spanned_pages > present_pages + (present_pages >> 4) &&
7171 IS_ENABLED(CONFIG_SPARSEMEM))
7172 pages = present_pages;
7174 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7177 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7178 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7180 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7182 spin_lock_init(&ds_queue->split_queue_lock);
7183 INIT_LIST_HEAD(&ds_queue->split_queue);
7184 ds_queue->split_queue_len = 0;
7187 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7190 #ifdef CONFIG_COMPACTION
7191 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7193 init_waitqueue_head(&pgdat->kcompactd_wait);
7196 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7199 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7201 pgdat_resize_init(pgdat);
7203 pgdat_init_split_queue(pgdat);
7204 pgdat_init_kcompactd(pgdat);
7206 init_waitqueue_head(&pgdat->kswapd_wait);
7207 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7209 pgdat_page_ext_init(pgdat);
7210 lruvec_init(&pgdat->__lruvec);
7213 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7214 unsigned long remaining_pages)
7216 atomic_long_set(&zone->managed_pages, remaining_pages);
7217 zone_set_nid(zone, nid);
7218 zone->name = zone_names[idx];
7219 zone->zone_pgdat = NODE_DATA(nid);
7220 spin_lock_init(&zone->lock);
7221 zone_seqlock_init(zone);
7222 zone_pcp_init(zone);
7226 * Set up the zone data structures
7227 * - init pgdat internals
7228 * - init all zones belonging to this node
7230 * NOTE: this function is only called during memory hotplug
7232 #ifdef CONFIG_MEMORY_HOTPLUG
7233 void __ref free_area_init_core_hotplug(int nid)
7236 pg_data_t *pgdat = NODE_DATA(nid);
7238 pgdat_init_internals(pgdat);
7239 for (z = 0; z < MAX_NR_ZONES; z++)
7240 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7245 * Set up the zone data structures:
7246 * - mark all pages reserved
7247 * - mark all memory queues empty
7248 * - clear the memory bitmaps
7250 * NOTE: pgdat should get zeroed by caller.
7251 * NOTE: this function is only called during early init.
7253 static void __init free_area_init_core(struct pglist_data *pgdat)
7256 int nid = pgdat->node_id;
7258 pgdat_init_internals(pgdat);
7259 pgdat->per_cpu_nodestats = &boot_nodestats;
7261 for (j = 0; j < MAX_NR_ZONES; j++) {
7262 struct zone *zone = pgdat->node_zones + j;
7263 unsigned long size, freesize, memmap_pages;
7265 size = zone->spanned_pages;
7266 freesize = zone->present_pages;
7269 * Adjust freesize so that it accounts for how much memory
7270 * is used by this zone for memmap. This affects the watermark
7271 * and per-cpu initialisations
7273 memmap_pages = calc_memmap_size(size, freesize);
7274 if (!is_highmem_idx(j)) {
7275 if (freesize >= memmap_pages) {
7276 freesize -= memmap_pages;
7278 pr_debug(" %s zone: %lu pages used for memmap\n",
7279 zone_names[j], memmap_pages);
7281 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
7282 zone_names[j], memmap_pages, freesize);
7285 /* Account for reserved pages */
7286 if (j == 0 && freesize > dma_reserve) {
7287 freesize -= dma_reserve;
7288 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7291 if (!is_highmem_idx(j))
7292 nr_kernel_pages += freesize;
7293 /* Charge for highmem memmap if there are enough kernel pages */
7294 else if (nr_kernel_pages > memmap_pages * 2)
7295 nr_kernel_pages -= memmap_pages;
7296 nr_all_pages += freesize;
7299 * Set an approximate value for lowmem here, it will be adjusted
7300 * when the bootmem allocator frees pages into the buddy system.
7301 * And all highmem pages will be managed by the buddy system.
7303 zone_init_internals(zone, j, nid, freesize);
7308 set_pageblock_order();
7310 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7314 #ifdef CONFIG_FLAT_NODE_MEM_MAP
7315 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
7317 unsigned long __maybe_unused start = 0;
7318 unsigned long __maybe_unused offset = 0;
7320 /* Skip empty nodes */
7321 if (!pgdat->node_spanned_pages)
7324 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7325 offset = pgdat->node_start_pfn - start;
7326 /* ia64 gets its own node_mem_map, before this, without bootmem */
7327 if (!pgdat->node_mem_map) {
7328 unsigned long size, end;
7332 * The zone's endpoints aren't required to be MAX_ORDER
7333 * aligned but the node_mem_map endpoints must be in order
7334 * for the buddy allocator to function correctly.
7336 end = pgdat_end_pfn(pgdat);
7337 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7338 size = (end - start) * sizeof(struct page);
7339 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7342 panic("Failed to allocate %ld bytes for node %d memory map\n",
7343 size, pgdat->node_id);
7344 pgdat->node_mem_map = map + offset;
7346 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7347 __func__, pgdat->node_id, (unsigned long)pgdat,
7348 (unsigned long)pgdat->node_mem_map);
7349 #ifndef CONFIG_NEED_MULTIPLE_NODES
7351 * With no DISCONTIG, the global mem_map is just set as node 0's
7353 if (pgdat == NODE_DATA(0)) {
7354 mem_map = NODE_DATA(0)->node_mem_map;
7355 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7361 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7362 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7364 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7365 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7367 pgdat->first_deferred_pfn = ULONG_MAX;
7370 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7373 static void __init free_area_init_node(int nid)
7375 pg_data_t *pgdat = NODE_DATA(nid);
7376 unsigned long start_pfn = 0;
7377 unsigned long end_pfn = 0;
7379 /* pg_data_t should be reset to zero when it's allocated */
7380 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7382 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7384 pgdat->node_id = nid;
7385 pgdat->node_start_pfn = start_pfn;
7386 pgdat->per_cpu_nodestats = NULL;
7388 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7389 (u64)start_pfn << PAGE_SHIFT,
7390 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7391 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7393 alloc_node_mem_map(pgdat);
7394 pgdat_set_deferred_range(pgdat);
7396 free_area_init_core(pgdat);
7399 void __init free_area_init_memoryless_node(int nid)
7401 free_area_init_node(nid);
7404 #if MAX_NUMNODES > 1
7406 * Figure out the number of possible node ids.
7408 void __init setup_nr_node_ids(void)
7410 unsigned int highest;
7412 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7413 nr_node_ids = highest + 1;
7418 * node_map_pfn_alignment - determine the maximum internode alignment
7420 * This function should be called after node map is populated and sorted.
7421 * It calculates the maximum power of two alignment which can distinguish
7424 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7425 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7426 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7427 * shifted, 1GiB is enough and this function will indicate so.
7429 * This is used to test whether pfn -> nid mapping of the chosen memory
7430 * model has fine enough granularity to avoid incorrect mapping for the
7431 * populated node map.
7433 * Return: the determined alignment in pfn's. 0 if there is no alignment
7434 * requirement (single node).
7436 unsigned long __init node_map_pfn_alignment(void)
7438 unsigned long accl_mask = 0, last_end = 0;
7439 unsigned long start, end, mask;
7440 int last_nid = NUMA_NO_NODE;
7443 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7444 if (!start || last_nid < 0 || last_nid == nid) {
7451 * Start with a mask granular enough to pin-point to the
7452 * start pfn and tick off bits one-by-one until it becomes
7453 * too coarse to separate the current node from the last.
7455 mask = ~((1 << __ffs(start)) - 1);
7456 while (mask && last_end <= (start & (mask << 1)))
7459 /* accumulate all internode masks */
7463 /* convert mask to number of pages */
7464 return ~accl_mask + 1;
7468 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7470 * Return: the minimum PFN based on information provided via
7471 * memblock_set_node().
7473 unsigned long __init find_min_pfn_with_active_regions(void)
7475 return PHYS_PFN(memblock_start_of_DRAM());
7479 * early_calculate_totalpages()
7480 * Sum pages in active regions for movable zone.
7481 * Populate N_MEMORY for calculating usable_nodes.
7483 static unsigned long __init early_calculate_totalpages(void)
7485 unsigned long totalpages = 0;
7486 unsigned long start_pfn, end_pfn;
7489 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7490 unsigned long pages = end_pfn - start_pfn;
7492 totalpages += pages;
7494 node_set_state(nid, N_MEMORY);
7500 * Find the PFN the Movable zone begins in each node. Kernel memory
7501 * is spread evenly between nodes as long as the nodes have enough
7502 * memory. When they don't, some nodes will have more kernelcore than
7505 static void __init find_zone_movable_pfns_for_nodes(void)
7508 unsigned long usable_startpfn;
7509 unsigned long kernelcore_node, kernelcore_remaining;
7510 /* save the state before borrow the nodemask */
7511 nodemask_t saved_node_state = node_states[N_MEMORY];
7512 unsigned long totalpages = early_calculate_totalpages();
7513 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7514 struct memblock_region *r;
7516 /* Need to find movable_zone earlier when movable_node is specified. */
7517 find_usable_zone_for_movable();
7520 * If movable_node is specified, ignore kernelcore and movablecore
7523 if (movable_node_is_enabled()) {
7524 for_each_mem_region(r) {
7525 if (!memblock_is_hotpluggable(r))
7528 nid = memblock_get_region_node(r);
7530 usable_startpfn = PFN_DOWN(r->base);
7531 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7532 min(usable_startpfn, zone_movable_pfn[nid]) :
7540 * If kernelcore=mirror is specified, ignore movablecore option
7542 if (mirrored_kernelcore) {
7543 bool mem_below_4gb_not_mirrored = false;
7545 for_each_mem_region(r) {
7546 if (memblock_is_mirror(r))
7549 nid = memblock_get_region_node(r);
7551 usable_startpfn = memblock_region_memory_base_pfn(r);
7553 if (usable_startpfn < 0x100000) {
7554 mem_below_4gb_not_mirrored = true;
7558 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7559 min(usable_startpfn, zone_movable_pfn[nid]) :
7563 if (mem_below_4gb_not_mirrored)
7564 pr_warn("This configuration results in unmirrored kernel memory.\n");
7570 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7571 * amount of necessary memory.
7573 if (required_kernelcore_percent)
7574 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7576 if (required_movablecore_percent)
7577 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7581 * If movablecore= was specified, calculate what size of
7582 * kernelcore that corresponds so that memory usable for
7583 * any allocation type is evenly spread. If both kernelcore
7584 * and movablecore are specified, then the value of kernelcore
7585 * will be used for required_kernelcore if it's greater than
7586 * what movablecore would have allowed.
7588 if (required_movablecore) {
7589 unsigned long corepages;
7592 * Round-up so that ZONE_MOVABLE is at least as large as what
7593 * was requested by the user
7595 required_movablecore =
7596 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7597 required_movablecore = min(totalpages, required_movablecore);
7598 corepages = totalpages - required_movablecore;
7600 required_kernelcore = max(required_kernelcore, corepages);
7604 * If kernelcore was not specified or kernelcore size is larger
7605 * than totalpages, there is no ZONE_MOVABLE.
7607 if (!required_kernelcore || required_kernelcore >= totalpages)
7610 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7611 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7614 /* Spread kernelcore memory as evenly as possible throughout nodes */
7615 kernelcore_node = required_kernelcore / usable_nodes;
7616 for_each_node_state(nid, N_MEMORY) {
7617 unsigned long start_pfn, end_pfn;
7620 * Recalculate kernelcore_node if the division per node
7621 * now exceeds what is necessary to satisfy the requested
7622 * amount of memory for the kernel
7624 if (required_kernelcore < kernelcore_node)
7625 kernelcore_node = required_kernelcore / usable_nodes;
7628 * As the map is walked, we track how much memory is usable
7629 * by the kernel using kernelcore_remaining. When it is
7630 * 0, the rest of the node is usable by ZONE_MOVABLE
7632 kernelcore_remaining = kernelcore_node;
7634 /* Go through each range of PFNs within this node */
7635 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7636 unsigned long size_pages;
7638 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7639 if (start_pfn >= end_pfn)
7642 /* Account for what is only usable for kernelcore */
7643 if (start_pfn < usable_startpfn) {
7644 unsigned long kernel_pages;
7645 kernel_pages = min(end_pfn, usable_startpfn)
7648 kernelcore_remaining -= min(kernel_pages,
7649 kernelcore_remaining);
7650 required_kernelcore -= min(kernel_pages,
7651 required_kernelcore);
7653 /* Continue if range is now fully accounted */
7654 if (end_pfn <= usable_startpfn) {
7657 * Push zone_movable_pfn to the end so
7658 * that if we have to rebalance
7659 * kernelcore across nodes, we will
7660 * not double account here
7662 zone_movable_pfn[nid] = end_pfn;
7665 start_pfn = usable_startpfn;
7669 * The usable PFN range for ZONE_MOVABLE is from
7670 * start_pfn->end_pfn. Calculate size_pages as the
7671 * number of pages used as kernelcore
7673 size_pages = end_pfn - start_pfn;
7674 if (size_pages > kernelcore_remaining)
7675 size_pages = kernelcore_remaining;
7676 zone_movable_pfn[nid] = start_pfn + size_pages;
7679 * Some kernelcore has been met, update counts and
7680 * break if the kernelcore for this node has been
7683 required_kernelcore -= min(required_kernelcore,
7685 kernelcore_remaining -= size_pages;
7686 if (!kernelcore_remaining)
7692 * If there is still required_kernelcore, we do another pass with one
7693 * less node in the count. This will push zone_movable_pfn[nid] further
7694 * along on the nodes that still have memory until kernelcore is
7698 if (usable_nodes && required_kernelcore > usable_nodes)
7702 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7703 for (nid = 0; nid < MAX_NUMNODES; nid++)
7704 zone_movable_pfn[nid] =
7705 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7708 /* restore the node_state */
7709 node_states[N_MEMORY] = saved_node_state;
7712 /* Any regular or high memory on that node ? */
7713 static void check_for_memory(pg_data_t *pgdat, int nid)
7715 enum zone_type zone_type;
7717 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7718 struct zone *zone = &pgdat->node_zones[zone_type];
7719 if (populated_zone(zone)) {
7720 if (IS_ENABLED(CONFIG_HIGHMEM))
7721 node_set_state(nid, N_HIGH_MEMORY);
7722 if (zone_type <= ZONE_NORMAL)
7723 node_set_state(nid, N_NORMAL_MEMORY);
7730 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7731 * such cases we allow max_zone_pfn sorted in the descending order
7733 bool __weak arch_has_descending_max_zone_pfns(void)
7739 * free_area_init - Initialise all pg_data_t and zone data
7740 * @max_zone_pfn: an array of max PFNs for each zone
7742 * This will call free_area_init_node() for each active node in the system.
7743 * Using the page ranges provided by memblock_set_node(), the size of each
7744 * zone in each node and their holes is calculated. If the maximum PFN
7745 * between two adjacent zones match, it is assumed that the zone is empty.
7746 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7747 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7748 * starts where the previous one ended. For example, ZONE_DMA32 starts
7749 * at arch_max_dma_pfn.
7751 void __init free_area_init(unsigned long *max_zone_pfn)
7753 unsigned long start_pfn, end_pfn;
7757 /* Record where the zone boundaries are */
7758 memset(arch_zone_lowest_possible_pfn, 0,
7759 sizeof(arch_zone_lowest_possible_pfn));
7760 memset(arch_zone_highest_possible_pfn, 0,
7761 sizeof(arch_zone_highest_possible_pfn));
7763 start_pfn = find_min_pfn_with_active_regions();
7764 descending = arch_has_descending_max_zone_pfns();
7766 for (i = 0; i < MAX_NR_ZONES; i++) {
7768 zone = MAX_NR_ZONES - i - 1;
7772 if (zone == ZONE_MOVABLE)
7775 end_pfn = max(max_zone_pfn[zone], start_pfn);
7776 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7777 arch_zone_highest_possible_pfn[zone] = end_pfn;
7779 start_pfn = end_pfn;
7782 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7783 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7784 find_zone_movable_pfns_for_nodes();
7786 /* Print out the zone ranges */
7787 pr_info("Zone ranges:\n");
7788 for (i = 0; i < MAX_NR_ZONES; i++) {
7789 if (i == ZONE_MOVABLE)
7791 pr_info(" %-8s ", zone_names[i]);
7792 if (arch_zone_lowest_possible_pfn[i] ==
7793 arch_zone_highest_possible_pfn[i])
7796 pr_cont("[mem %#018Lx-%#018Lx]\n",
7797 (u64)arch_zone_lowest_possible_pfn[i]
7799 ((u64)arch_zone_highest_possible_pfn[i]
7800 << PAGE_SHIFT) - 1);
7803 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7804 pr_info("Movable zone start for each node\n");
7805 for (i = 0; i < MAX_NUMNODES; i++) {
7806 if (zone_movable_pfn[i])
7807 pr_info(" Node %d: %#018Lx\n", i,
7808 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7812 * Print out the early node map, and initialize the
7813 * subsection-map relative to active online memory ranges to
7814 * enable future "sub-section" extensions of the memory map.
7816 pr_info("Early memory node ranges\n");
7817 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7818 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7819 (u64)start_pfn << PAGE_SHIFT,
7820 ((u64)end_pfn << PAGE_SHIFT) - 1);
7821 subsection_map_init(start_pfn, end_pfn - start_pfn);
7824 /* Initialise every node */
7825 mminit_verify_pageflags_layout();
7826 setup_nr_node_ids();
7827 for_each_online_node(nid) {
7828 pg_data_t *pgdat = NODE_DATA(nid);
7829 free_area_init_node(nid);
7831 /* Any memory on that node */
7832 if (pgdat->node_present_pages)
7833 node_set_state(nid, N_MEMORY);
7834 check_for_memory(pgdat, nid);
7840 static int __init cmdline_parse_core(char *p, unsigned long *core,
7841 unsigned long *percent)
7843 unsigned long long coremem;
7849 /* Value may be a percentage of total memory, otherwise bytes */
7850 coremem = simple_strtoull(p, &endptr, 0);
7851 if (*endptr == '%') {
7852 /* Paranoid check for percent values greater than 100 */
7853 WARN_ON(coremem > 100);
7857 coremem = memparse(p, &p);
7858 /* Paranoid check that UL is enough for the coremem value */
7859 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7861 *core = coremem >> PAGE_SHIFT;
7868 * kernelcore=size sets the amount of memory for use for allocations that
7869 * cannot be reclaimed or migrated.
7871 static int __init cmdline_parse_kernelcore(char *p)
7873 /* parse kernelcore=mirror */
7874 if (parse_option_str(p, "mirror")) {
7875 mirrored_kernelcore = true;
7879 return cmdline_parse_core(p, &required_kernelcore,
7880 &required_kernelcore_percent);
7884 * movablecore=size sets the amount of memory for use for allocations that
7885 * can be reclaimed or migrated.
7887 static int __init cmdline_parse_movablecore(char *p)
7889 return cmdline_parse_core(p, &required_movablecore,
7890 &required_movablecore_percent);
7893 early_param("kernelcore", cmdline_parse_kernelcore);
7894 early_param("movablecore", cmdline_parse_movablecore);
7896 void adjust_managed_page_count(struct page *page, long count)
7898 atomic_long_add(count, &page_zone(page)->managed_pages);
7899 totalram_pages_add(count);
7900 #ifdef CONFIG_HIGHMEM
7901 if (PageHighMem(page))
7902 totalhigh_pages_add(count);
7905 EXPORT_SYMBOL(adjust_managed_page_count);
7907 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7910 unsigned long pages = 0;
7912 start = (void *)PAGE_ALIGN((unsigned long)start);
7913 end = (void *)((unsigned long)end & PAGE_MASK);
7914 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7915 struct page *page = virt_to_page(pos);
7916 void *direct_map_addr;
7919 * 'direct_map_addr' might be different from 'pos'
7920 * because some architectures' virt_to_page()
7921 * work with aliases. Getting the direct map
7922 * address ensures that we get a _writeable_
7923 * alias for the memset().
7925 direct_map_addr = page_address(page);
7927 * Perform a kasan-unchecked memset() since this memory
7928 * has not been initialized.
7930 direct_map_addr = kasan_reset_tag(direct_map_addr);
7931 if ((unsigned int)poison <= 0xFF)
7932 memset(direct_map_addr, poison, PAGE_SIZE);
7934 free_reserved_page(page);
7938 pr_info("Freeing %s memory: %ldK\n",
7939 s, pages << (PAGE_SHIFT - 10));
7944 void __init mem_init_print_info(void)
7946 unsigned long physpages, codesize, datasize, rosize, bss_size;
7947 unsigned long init_code_size, init_data_size;
7949 physpages = get_num_physpages();
7950 codesize = _etext - _stext;
7951 datasize = _edata - _sdata;
7952 rosize = __end_rodata - __start_rodata;
7953 bss_size = __bss_stop - __bss_start;
7954 init_data_size = __init_end - __init_begin;
7955 init_code_size = _einittext - _sinittext;
7958 * Detect special cases and adjust section sizes accordingly:
7959 * 1) .init.* may be embedded into .data sections
7960 * 2) .init.text.* may be out of [__init_begin, __init_end],
7961 * please refer to arch/tile/kernel/vmlinux.lds.S.
7962 * 3) .rodata.* may be embedded into .text or .data sections.
7964 #define adj_init_size(start, end, size, pos, adj) \
7966 if (start <= pos && pos < end && size > adj) \
7970 adj_init_size(__init_begin, __init_end, init_data_size,
7971 _sinittext, init_code_size);
7972 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7973 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7974 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7975 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7977 #undef adj_init_size
7979 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7980 #ifdef CONFIG_HIGHMEM
7984 nr_free_pages() << (PAGE_SHIFT - 10),
7985 physpages << (PAGE_SHIFT - 10),
7986 codesize >> 10, datasize >> 10, rosize >> 10,
7987 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7988 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7989 totalcma_pages << (PAGE_SHIFT - 10)
7990 #ifdef CONFIG_HIGHMEM
7991 , totalhigh_pages() << (PAGE_SHIFT - 10)
7997 * set_dma_reserve - set the specified number of pages reserved in the first zone
7998 * @new_dma_reserve: The number of pages to mark reserved
8000 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8001 * In the DMA zone, a significant percentage may be consumed by kernel image
8002 * and other unfreeable allocations which can skew the watermarks badly. This
8003 * function may optionally be used to account for unfreeable pages in the
8004 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8005 * smaller per-cpu batchsize.
8007 void __init set_dma_reserve(unsigned long new_dma_reserve)
8009 dma_reserve = new_dma_reserve;
8012 static int page_alloc_cpu_dead(unsigned int cpu)
8015 lru_add_drain_cpu(cpu);
8019 * Spill the event counters of the dead processor
8020 * into the current processors event counters.
8021 * This artificially elevates the count of the current
8024 vm_events_fold_cpu(cpu);
8027 * Zero the differential counters of the dead processor
8028 * so that the vm statistics are consistent.
8030 * This is only okay since the processor is dead and cannot
8031 * race with what we are doing.
8033 cpu_vm_stats_fold(cpu);
8038 int hashdist = HASHDIST_DEFAULT;
8040 static int __init set_hashdist(char *str)
8044 hashdist = simple_strtoul(str, &str, 0);
8047 __setup("hashdist=", set_hashdist);
8050 void __init page_alloc_init(void)
8055 if (num_node_state(N_MEMORY) == 1)
8059 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
8060 "mm/page_alloc:dead", NULL,
8061 page_alloc_cpu_dead);
8066 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8067 * or min_free_kbytes changes.
8069 static void calculate_totalreserve_pages(void)
8071 struct pglist_data *pgdat;
8072 unsigned long reserve_pages = 0;
8073 enum zone_type i, j;
8075 for_each_online_pgdat(pgdat) {
8077 pgdat->totalreserve_pages = 0;
8079 for (i = 0; i < MAX_NR_ZONES; i++) {
8080 struct zone *zone = pgdat->node_zones + i;
8082 unsigned long managed_pages = zone_managed_pages(zone);
8084 /* Find valid and maximum lowmem_reserve in the zone */
8085 for (j = i; j < MAX_NR_ZONES; j++) {
8086 if (zone->lowmem_reserve[j] > max)
8087 max = zone->lowmem_reserve[j];
8090 /* we treat the high watermark as reserved pages. */
8091 max += high_wmark_pages(zone);
8093 if (max > managed_pages)
8094 max = managed_pages;
8096 pgdat->totalreserve_pages += max;
8098 reserve_pages += max;
8101 totalreserve_pages = reserve_pages;
8105 * setup_per_zone_lowmem_reserve - called whenever
8106 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8107 * has a correct pages reserved value, so an adequate number of
8108 * pages are left in the zone after a successful __alloc_pages().
8110 static void setup_per_zone_lowmem_reserve(void)
8112 struct pglist_data *pgdat;
8113 enum zone_type i, j;
8115 for_each_online_pgdat(pgdat) {
8116 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8117 struct zone *zone = &pgdat->node_zones[i];
8118 int ratio = sysctl_lowmem_reserve_ratio[i];
8119 bool clear = !ratio || !zone_managed_pages(zone);
8120 unsigned long managed_pages = 0;
8122 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8124 zone->lowmem_reserve[j] = 0;
8126 struct zone *upper_zone = &pgdat->node_zones[j];
8128 managed_pages += zone_managed_pages(upper_zone);
8129 zone->lowmem_reserve[j] = managed_pages / ratio;
8135 /* update totalreserve_pages */
8136 calculate_totalreserve_pages();
8139 static void __setup_per_zone_wmarks(void)
8141 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8142 unsigned long lowmem_pages = 0;
8144 unsigned long flags;
8146 /* Calculate total number of !ZONE_HIGHMEM pages */
8147 for_each_zone(zone) {
8148 if (!is_highmem(zone))
8149 lowmem_pages += zone_managed_pages(zone);
8152 for_each_zone(zone) {
8155 spin_lock_irqsave(&zone->lock, flags);
8156 tmp = (u64)pages_min * zone_managed_pages(zone);
8157 do_div(tmp, lowmem_pages);
8158 if (is_highmem(zone)) {
8160 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8161 * need highmem pages, so cap pages_min to a small
8164 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8165 * deltas control async page reclaim, and so should
8166 * not be capped for highmem.
8168 unsigned long min_pages;
8170 min_pages = zone_managed_pages(zone) / 1024;
8171 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8172 zone->_watermark[WMARK_MIN] = min_pages;
8175 * If it's a lowmem zone, reserve a number of pages
8176 * proportionate to the zone's size.
8178 zone->_watermark[WMARK_MIN] = tmp;
8182 * Set the kswapd watermarks distance according to the
8183 * scale factor in proportion to available memory, but
8184 * ensure a minimum size on small systems.
8186 tmp = max_t(u64, tmp >> 2,
8187 mult_frac(zone_managed_pages(zone),
8188 watermark_scale_factor, 10000));
8190 zone->watermark_boost = 0;
8191 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8192 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8194 spin_unlock_irqrestore(&zone->lock, flags);
8197 /* update totalreserve_pages */
8198 calculate_totalreserve_pages();
8202 * setup_per_zone_wmarks - called when min_free_kbytes changes
8203 * or when memory is hot-{added|removed}
8205 * Ensures that the watermark[min,low,high] values for each zone are set
8206 * correctly with respect to min_free_kbytes.
8208 void setup_per_zone_wmarks(void)
8210 static DEFINE_SPINLOCK(lock);
8213 __setup_per_zone_wmarks();
8218 * Initialise min_free_kbytes.
8220 * For small machines we want it small (128k min). For large machines
8221 * we want it large (256MB max). But it is not linear, because network
8222 * bandwidth does not increase linearly with machine size. We use
8224 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8225 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8241 int __meminit init_per_zone_wmark_min(void)
8243 unsigned long lowmem_kbytes;
8244 int new_min_free_kbytes;
8246 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8247 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8249 if (new_min_free_kbytes > user_min_free_kbytes) {
8250 min_free_kbytes = new_min_free_kbytes;
8251 if (min_free_kbytes < 128)
8252 min_free_kbytes = 128;
8253 if (min_free_kbytes > 262144)
8254 min_free_kbytes = 262144;
8256 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8257 new_min_free_kbytes, user_min_free_kbytes);
8259 setup_per_zone_wmarks();
8260 refresh_zone_stat_thresholds();
8261 setup_per_zone_lowmem_reserve();
8264 setup_min_unmapped_ratio();
8265 setup_min_slab_ratio();
8268 khugepaged_min_free_kbytes_update();
8272 postcore_initcall(init_per_zone_wmark_min)
8275 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8276 * that we can call two helper functions whenever min_free_kbytes
8279 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8280 void *buffer, size_t *length, loff_t *ppos)
8284 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8289 user_min_free_kbytes = min_free_kbytes;
8290 setup_per_zone_wmarks();
8295 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8296 void *buffer, size_t *length, loff_t *ppos)
8300 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8305 setup_per_zone_wmarks();
8311 static void setup_min_unmapped_ratio(void)
8316 for_each_online_pgdat(pgdat)
8317 pgdat->min_unmapped_pages = 0;
8320 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8321 sysctl_min_unmapped_ratio) / 100;
8325 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8326 void *buffer, size_t *length, loff_t *ppos)
8330 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8334 setup_min_unmapped_ratio();
8339 static void setup_min_slab_ratio(void)
8344 for_each_online_pgdat(pgdat)
8345 pgdat->min_slab_pages = 0;
8348 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8349 sysctl_min_slab_ratio) / 100;
8352 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8353 void *buffer, size_t *length, loff_t *ppos)
8357 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8361 setup_min_slab_ratio();
8368 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8369 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8370 * whenever sysctl_lowmem_reserve_ratio changes.
8372 * The reserve ratio obviously has absolutely no relation with the
8373 * minimum watermarks. The lowmem reserve ratio can only make sense
8374 * if in function of the boot time zone sizes.
8376 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8377 void *buffer, size_t *length, loff_t *ppos)
8381 proc_dointvec_minmax(table, write, buffer, length, ppos);
8383 for (i = 0; i < MAX_NR_ZONES; i++) {
8384 if (sysctl_lowmem_reserve_ratio[i] < 1)
8385 sysctl_lowmem_reserve_ratio[i] = 0;
8388 setup_per_zone_lowmem_reserve();
8393 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8394 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8395 * pagelist can have before it gets flushed back to buddy allocator.
8397 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8398 void *buffer, size_t *length, loff_t *ppos)
8401 int old_percpu_pagelist_fraction;
8404 mutex_lock(&pcp_batch_high_lock);
8405 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8407 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8408 if (!write || ret < 0)
8411 /* Sanity checking to avoid pcp imbalance */
8412 if (percpu_pagelist_fraction &&
8413 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8414 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8420 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8423 for_each_populated_zone(zone)
8424 zone_set_pageset_high_and_batch(zone);
8426 mutex_unlock(&pcp_batch_high_lock);
8430 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8432 * Returns the number of pages that arch has reserved but
8433 * is not known to alloc_large_system_hash().
8435 static unsigned long __init arch_reserved_kernel_pages(void)
8442 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8443 * machines. As memory size is increased the scale is also increased but at
8444 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8445 * quadruples the scale is increased by one, which means the size of hash table
8446 * only doubles, instead of quadrupling as well.
8447 * Because 32-bit systems cannot have large physical memory, where this scaling
8448 * makes sense, it is disabled on such platforms.
8450 #if __BITS_PER_LONG > 32
8451 #define ADAPT_SCALE_BASE (64ul << 30)
8452 #define ADAPT_SCALE_SHIFT 2
8453 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8457 * allocate a large system hash table from bootmem
8458 * - it is assumed that the hash table must contain an exact power-of-2
8459 * quantity of entries
8460 * - limit is the number of hash buckets, not the total allocation size
8462 void *__init alloc_large_system_hash(const char *tablename,
8463 unsigned long bucketsize,
8464 unsigned long numentries,
8467 unsigned int *_hash_shift,
8468 unsigned int *_hash_mask,
8469 unsigned long low_limit,
8470 unsigned long high_limit)
8472 unsigned long long max = high_limit;
8473 unsigned long log2qty, size;
8479 /* allow the kernel cmdline to have a say */
8481 /* round applicable memory size up to nearest megabyte */
8482 numentries = nr_kernel_pages;
8483 numentries -= arch_reserved_kernel_pages();
8485 /* It isn't necessary when PAGE_SIZE >= 1MB */
8486 if (PAGE_SHIFT < 20)
8487 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8489 #if __BITS_PER_LONG > 32
8491 unsigned long adapt;
8493 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8494 adapt <<= ADAPT_SCALE_SHIFT)
8499 /* limit to 1 bucket per 2^scale bytes of low memory */
8500 if (scale > PAGE_SHIFT)
8501 numentries >>= (scale - PAGE_SHIFT);
8503 numentries <<= (PAGE_SHIFT - scale);
8505 /* Make sure we've got at least a 0-order allocation.. */
8506 if (unlikely(flags & HASH_SMALL)) {
8507 /* Makes no sense without HASH_EARLY */
8508 WARN_ON(!(flags & HASH_EARLY));
8509 if (!(numentries >> *_hash_shift)) {
8510 numentries = 1UL << *_hash_shift;
8511 BUG_ON(!numentries);
8513 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8514 numentries = PAGE_SIZE / bucketsize;
8516 numentries = roundup_pow_of_two(numentries);
8518 /* limit allocation size to 1/16 total memory by default */
8520 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8521 do_div(max, bucketsize);
8523 max = min(max, 0x80000000ULL);
8525 if (numentries < low_limit)
8526 numentries = low_limit;
8527 if (numentries > max)
8530 log2qty = ilog2(numentries);
8532 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8535 size = bucketsize << log2qty;
8536 if (flags & HASH_EARLY) {
8537 if (flags & HASH_ZERO)
8538 table = memblock_alloc(size, SMP_CACHE_BYTES);
8540 table = memblock_alloc_raw(size,
8542 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8543 table = __vmalloc(size, gfp_flags);
8545 huge = is_vm_area_hugepages(table);
8548 * If bucketsize is not a power-of-two, we may free
8549 * some pages at the end of hash table which
8550 * alloc_pages_exact() automatically does
8552 table = alloc_pages_exact(size, gfp_flags);
8553 kmemleak_alloc(table, size, 1, gfp_flags);
8555 } while (!table && size > PAGE_SIZE && --log2qty);
8558 panic("Failed to allocate %s hash table\n", tablename);
8560 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8561 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8562 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8565 *_hash_shift = log2qty;
8567 *_hash_mask = (1 << log2qty) - 1;
8573 * This function checks whether pageblock includes unmovable pages or not.
8575 * PageLRU check without isolation or lru_lock could race so that
8576 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8577 * check without lock_page also may miss some movable non-lru pages at
8578 * race condition. So you can't expect this function should be exact.
8580 * Returns a page without holding a reference. If the caller wants to
8581 * dereference that page (e.g., dumping), it has to make sure that it
8582 * cannot get removed (e.g., via memory unplug) concurrently.
8585 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8586 int migratetype, int flags)
8588 unsigned long iter = 0;
8589 unsigned long pfn = page_to_pfn(page);
8590 unsigned long offset = pfn % pageblock_nr_pages;
8592 if (is_migrate_cma_page(page)) {
8594 * CMA allocations (alloc_contig_range) really need to mark
8595 * isolate CMA pageblocks even when they are not movable in fact
8596 * so consider them movable here.
8598 if (is_migrate_cma(migratetype))
8604 for (; iter < pageblock_nr_pages - offset; iter++) {
8605 if (!pfn_valid_within(pfn + iter))
8608 page = pfn_to_page(pfn + iter);
8611 * Both, bootmem allocations and memory holes are marked
8612 * PG_reserved and are unmovable. We can even have unmovable
8613 * allocations inside ZONE_MOVABLE, for example when
8614 * specifying "movablecore".
8616 if (PageReserved(page))
8620 * If the zone is movable and we have ruled out all reserved
8621 * pages then it should be reasonably safe to assume the rest
8624 if (zone_idx(zone) == ZONE_MOVABLE)
8628 * Hugepages are not in LRU lists, but they're movable.
8629 * THPs are on the LRU, but need to be counted as #small pages.
8630 * We need not scan over tail pages because we don't
8631 * handle each tail page individually in migration.
8633 if (PageHuge(page) || PageTransCompound(page)) {
8634 struct page *head = compound_head(page);
8635 unsigned int skip_pages;
8637 if (PageHuge(page)) {
8638 if (!hugepage_migration_supported(page_hstate(head)))
8640 } else if (!PageLRU(head) && !__PageMovable(head)) {
8644 skip_pages = compound_nr(head) - (page - head);
8645 iter += skip_pages - 1;
8650 * We can't use page_count without pin a page
8651 * because another CPU can free compound page.
8652 * This check already skips compound tails of THP
8653 * because their page->_refcount is zero at all time.
8655 if (!page_ref_count(page)) {
8656 if (PageBuddy(page))
8657 iter += (1 << buddy_order(page)) - 1;
8662 * The HWPoisoned page may be not in buddy system, and
8663 * page_count() is not 0.
8665 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8669 * We treat all PageOffline() pages as movable when offlining
8670 * to give drivers a chance to decrement their reference count
8671 * in MEM_GOING_OFFLINE in order to indicate that these pages
8672 * can be offlined as there are no direct references anymore.
8673 * For actually unmovable PageOffline() where the driver does
8674 * not support this, we will fail later when trying to actually
8675 * move these pages that still have a reference count > 0.
8676 * (false negatives in this function only)
8678 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8681 if (__PageMovable(page) || PageLRU(page))
8685 * If there are RECLAIMABLE pages, we need to check
8686 * it. But now, memory offline itself doesn't call
8687 * shrink_node_slabs() and it still to be fixed.
8694 #ifdef CONFIG_CONTIG_ALLOC
8695 static unsigned long pfn_max_align_down(unsigned long pfn)
8697 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8698 pageblock_nr_pages) - 1);
8701 static unsigned long pfn_max_align_up(unsigned long pfn)
8703 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8704 pageblock_nr_pages));
8707 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8708 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8709 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8710 static void alloc_contig_dump_pages(struct list_head *page_list)
8712 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8714 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8718 list_for_each_entry(page, page_list, lru)
8719 dump_page(page, "migration failure");
8723 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8728 /* [start, end) must belong to a single zone. */
8729 static int __alloc_contig_migrate_range(struct compact_control *cc,
8730 unsigned long start, unsigned long end)
8732 /* This function is based on compact_zone() from compaction.c. */
8733 unsigned int nr_reclaimed;
8734 unsigned long pfn = start;
8735 unsigned int tries = 0;
8737 struct migration_target_control mtc = {
8738 .nid = zone_to_nid(cc->zone),
8739 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8742 lru_cache_disable();
8744 while (pfn < end || !list_empty(&cc->migratepages)) {
8745 if (fatal_signal_pending(current)) {
8750 if (list_empty(&cc->migratepages)) {
8751 cc->nr_migratepages = 0;
8752 ret = isolate_migratepages_range(cc, pfn, end);
8753 if (ret && ret != -EAGAIN)
8755 pfn = cc->migrate_pfn;
8757 } else if (++tries == 5) {
8762 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8764 cc->nr_migratepages -= nr_reclaimed;
8766 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8767 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8770 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
8771 * to retry again over this error, so do the same here.
8779 alloc_contig_dump_pages(&cc->migratepages);
8780 putback_movable_pages(&cc->migratepages);
8787 * alloc_contig_range() -- tries to allocate given range of pages
8788 * @start: start PFN to allocate
8789 * @end: one-past-the-last PFN to allocate
8790 * @migratetype: migratetype of the underlying pageblocks (either
8791 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8792 * in range must have the same migratetype and it must
8793 * be either of the two.
8794 * @gfp_mask: GFP mask to use during compaction
8796 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8797 * aligned. The PFN range must belong to a single zone.
8799 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8800 * pageblocks in the range. Once isolated, the pageblocks should not
8801 * be modified by others.
8803 * Return: zero on success or negative error code. On success all
8804 * pages which PFN is in [start, end) are allocated for the caller and
8805 * need to be freed with free_contig_range().
8807 int alloc_contig_range(unsigned long start, unsigned long end,
8808 unsigned migratetype, gfp_t gfp_mask)
8810 unsigned long outer_start, outer_end;
8814 struct compact_control cc = {
8815 .nr_migratepages = 0,
8817 .zone = page_zone(pfn_to_page(start)),
8818 .mode = MIGRATE_SYNC,
8819 .ignore_skip_hint = true,
8820 .no_set_skip_hint = true,
8821 .gfp_mask = current_gfp_context(gfp_mask),
8822 .alloc_contig = true,
8824 INIT_LIST_HEAD(&cc.migratepages);
8827 * What we do here is we mark all pageblocks in range as
8828 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8829 * have different sizes, and due to the way page allocator
8830 * work, we align the range to biggest of the two pages so
8831 * that page allocator won't try to merge buddies from
8832 * different pageblocks and change MIGRATE_ISOLATE to some
8833 * other migration type.
8835 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8836 * migrate the pages from an unaligned range (ie. pages that
8837 * we are interested in). This will put all the pages in
8838 * range back to page allocator as MIGRATE_ISOLATE.
8840 * When this is done, we take the pages in range from page
8841 * allocator removing them from the buddy system. This way
8842 * page allocator will never consider using them.
8844 * This lets us mark the pageblocks back as
8845 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8846 * aligned range but not in the unaligned, original range are
8847 * put back to page allocator so that buddy can use them.
8850 ret = start_isolate_page_range(pfn_max_align_down(start),
8851 pfn_max_align_up(end), migratetype, 0);
8855 drain_all_pages(cc.zone);
8858 * In case of -EBUSY, we'd like to know which page causes problem.
8859 * So, just fall through. test_pages_isolated() has a tracepoint
8860 * which will report the busy page.
8862 * It is possible that busy pages could become available before
8863 * the call to test_pages_isolated, and the range will actually be
8864 * allocated. So, if we fall through be sure to clear ret so that
8865 * -EBUSY is not accidentally used or returned to caller.
8867 ret = __alloc_contig_migrate_range(&cc, start, end);
8868 if (ret && ret != -EBUSY)
8873 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8874 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8875 * more, all pages in [start, end) are free in page allocator.
8876 * What we are going to do is to allocate all pages from
8877 * [start, end) (that is remove them from page allocator).
8879 * The only problem is that pages at the beginning and at the
8880 * end of interesting range may be not aligned with pages that
8881 * page allocator holds, ie. they can be part of higher order
8882 * pages. Because of this, we reserve the bigger range and
8883 * once this is done free the pages we are not interested in.
8885 * We don't have to hold zone->lock here because the pages are
8886 * isolated thus they won't get removed from buddy.
8890 outer_start = start;
8891 while (!PageBuddy(pfn_to_page(outer_start))) {
8892 if (++order >= MAX_ORDER) {
8893 outer_start = start;
8896 outer_start &= ~0UL << order;
8899 if (outer_start != start) {
8900 order = buddy_order(pfn_to_page(outer_start));
8903 * outer_start page could be small order buddy page and
8904 * it doesn't include start page. Adjust outer_start
8905 * in this case to report failed page properly
8906 * on tracepoint in test_pages_isolated()
8908 if (outer_start + (1UL << order) <= start)
8909 outer_start = start;
8912 /* Make sure the range is really isolated. */
8913 if (test_pages_isolated(outer_start, end, 0)) {
8918 /* Grab isolated pages from freelists. */
8919 outer_end = isolate_freepages_range(&cc, outer_start, end);
8925 /* Free head and tail (if any) */
8926 if (start != outer_start)
8927 free_contig_range(outer_start, start - outer_start);
8928 if (end != outer_end)
8929 free_contig_range(end, outer_end - end);
8932 undo_isolate_page_range(pfn_max_align_down(start),
8933 pfn_max_align_up(end), migratetype);
8936 EXPORT_SYMBOL(alloc_contig_range);
8938 static int __alloc_contig_pages(unsigned long start_pfn,
8939 unsigned long nr_pages, gfp_t gfp_mask)
8941 unsigned long end_pfn = start_pfn + nr_pages;
8943 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8947 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8948 unsigned long nr_pages)
8950 unsigned long i, end_pfn = start_pfn + nr_pages;
8953 for (i = start_pfn; i < end_pfn; i++) {
8954 page = pfn_to_online_page(i);
8958 if (page_zone(page) != z)
8961 if (PageReserved(page))
8967 static bool zone_spans_last_pfn(const struct zone *zone,
8968 unsigned long start_pfn, unsigned long nr_pages)
8970 unsigned long last_pfn = start_pfn + nr_pages - 1;
8972 return zone_spans_pfn(zone, last_pfn);
8976 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8977 * @nr_pages: Number of contiguous pages to allocate
8978 * @gfp_mask: GFP mask to limit search and used during compaction
8980 * @nodemask: Mask for other possible nodes
8982 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8983 * on an applicable zonelist to find a contiguous pfn range which can then be
8984 * tried for allocation with alloc_contig_range(). This routine is intended
8985 * for allocation requests which can not be fulfilled with the buddy allocator.
8987 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8988 * power of two then the alignment is guaranteed to be to the given nr_pages
8989 * (e.g. 1GB request would be aligned to 1GB).
8991 * Allocated pages can be freed with free_contig_range() or by manually calling
8992 * __free_page() on each allocated page.
8994 * Return: pointer to contiguous pages on success, or NULL if not successful.
8996 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8997 int nid, nodemask_t *nodemask)
8999 unsigned long ret, pfn, flags;
9000 struct zonelist *zonelist;
9004 zonelist = node_zonelist(nid, gfp_mask);
9005 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9006 gfp_zone(gfp_mask), nodemask) {
9007 spin_lock_irqsave(&zone->lock, flags);
9009 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9010 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9011 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9013 * We release the zone lock here because
9014 * alloc_contig_range() will also lock the zone
9015 * at some point. If there's an allocation
9016 * spinning on this lock, it may win the race
9017 * and cause alloc_contig_range() to fail...
9019 spin_unlock_irqrestore(&zone->lock, flags);
9020 ret = __alloc_contig_pages(pfn, nr_pages,
9023 return pfn_to_page(pfn);
9024 spin_lock_irqsave(&zone->lock, flags);
9028 spin_unlock_irqrestore(&zone->lock, flags);
9032 #endif /* CONFIG_CONTIG_ALLOC */
9034 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9036 unsigned long count = 0;
9038 for (; nr_pages--; pfn++) {
9039 struct page *page = pfn_to_page(pfn);
9041 count += page_count(page) != 1;
9044 WARN(count != 0, "%lu pages are still in use!\n", count);
9046 EXPORT_SYMBOL(free_contig_range);
9049 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9050 * page high values need to be recalculated.
9052 void __meminit zone_pcp_update(struct zone *zone)
9054 mutex_lock(&pcp_batch_high_lock);
9055 zone_set_pageset_high_and_batch(zone);
9056 mutex_unlock(&pcp_batch_high_lock);
9060 * Effectively disable pcplists for the zone by setting the high limit to 0
9061 * and draining all cpus. A concurrent page freeing on another CPU that's about
9062 * to put the page on pcplist will either finish before the drain and the page
9063 * will be drained, or observe the new high limit and skip the pcplist.
9065 * Must be paired with a call to zone_pcp_enable().
9067 void zone_pcp_disable(struct zone *zone)
9069 mutex_lock(&pcp_batch_high_lock);
9070 __zone_set_pageset_high_and_batch(zone, 0, 1);
9071 __drain_all_pages(zone, true);
9074 void zone_pcp_enable(struct zone *zone)
9076 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9077 mutex_unlock(&pcp_batch_high_lock);
9080 void zone_pcp_reset(struct zone *zone)
9083 struct per_cpu_zonestat *pzstats;
9085 if (zone->per_cpu_pageset != &boot_pageset) {
9086 for_each_online_cpu(cpu) {
9087 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9088 drain_zonestat(zone, pzstats);
9090 free_percpu(zone->per_cpu_pageset);
9091 free_percpu(zone->per_cpu_zonestats);
9092 zone->per_cpu_pageset = &boot_pageset;
9093 zone->per_cpu_zonestats = &boot_zonestats;
9097 #ifdef CONFIG_MEMORY_HOTREMOVE
9099 * All pages in the range must be in a single zone, must not contain holes,
9100 * must span full sections, and must be isolated before calling this function.
9102 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9104 unsigned long pfn = start_pfn;
9108 unsigned long flags;
9110 offline_mem_sections(pfn, end_pfn);
9111 zone = page_zone(pfn_to_page(pfn));
9112 spin_lock_irqsave(&zone->lock, flags);
9113 while (pfn < end_pfn) {
9114 page = pfn_to_page(pfn);
9116 * The HWPoisoned page may be not in buddy system, and
9117 * page_count() is not 0.
9119 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9124 * At this point all remaining PageOffline() pages have a
9125 * reference count of 0 and can simply be skipped.
9127 if (PageOffline(page)) {
9128 BUG_ON(page_count(page));
9129 BUG_ON(PageBuddy(page));
9134 BUG_ON(page_count(page));
9135 BUG_ON(!PageBuddy(page));
9136 order = buddy_order(page);
9137 del_page_from_free_list(page, zone, order);
9138 pfn += (1 << order);
9140 spin_unlock_irqrestore(&zone->lock, flags);
9144 bool is_free_buddy_page(struct page *page)
9146 struct zone *zone = page_zone(page);
9147 unsigned long pfn = page_to_pfn(page);
9148 unsigned long flags;
9151 spin_lock_irqsave(&zone->lock, flags);
9152 for (order = 0; order < MAX_ORDER; order++) {
9153 struct page *page_head = page - (pfn & ((1 << order) - 1));
9155 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
9158 spin_unlock_irqrestore(&zone->lock, flags);
9160 return order < MAX_ORDER;
9163 #ifdef CONFIG_MEMORY_FAILURE
9165 * Break down a higher-order page in sub-pages, and keep our target out of
9168 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9169 struct page *target, int low, int high,
9172 unsigned long size = 1 << high;
9173 struct page *current_buddy, *next_page;
9175 while (high > low) {
9179 if (target >= &page[size]) {
9180 next_page = page + size;
9181 current_buddy = page;
9184 current_buddy = page + size;
9187 if (set_page_guard(zone, current_buddy, high, migratetype))
9190 if (current_buddy != target) {
9191 add_to_free_list(current_buddy, zone, high, migratetype);
9192 set_buddy_order(current_buddy, high);
9199 * Take a page that will be marked as poisoned off the buddy allocator.
9201 bool take_page_off_buddy(struct page *page)
9203 struct zone *zone = page_zone(page);
9204 unsigned long pfn = page_to_pfn(page);
9205 unsigned long flags;
9209 spin_lock_irqsave(&zone->lock, flags);
9210 for (order = 0; order < MAX_ORDER; order++) {
9211 struct page *page_head = page - (pfn & ((1 << order) - 1));
9212 int page_order = buddy_order(page_head);
9214 if (PageBuddy(page_head) && page_order >= order) {
9215 unsigned long pfn_head = page_to_pfn(page_head);
9216 int migratetype = get_pfnblock_migratetype(page_head,
9219 del_page_from_free_list(page_head, zone, page_order);
9220 break_down_buddy_pages(zone, page_head, page, 0,
9221 page_order, migratetype);
9222 if (!is_migrate_isolate(migratetype))
9223 __mod_zone_freepage_state(zone, -1, migratetype);
9227 if (page_count(page_head) > 0)
9230 spin_unlock_irqrestore(&zone->lock, flags);