1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/mmu_notifier.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
71 #include <linux/psi.h>
72 #include <linux/padata.h>
73 #include <linux/khugepaged.h>
74 #include <linux/buffer_head.h>
75 #include <asm/sections.h>
76 #include <asm/tlbflush.h>
77 #include <asm/div64.h>
80 #include "page_reporting.h"
82 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
83 typedef int __bitwise fpi_t;
85 /* No special request */
86 #define FPI_NONE ((__force fpi_t)0)
89 * Skip free page reporting notification for the (possibly merged) page.
90 * This does not hinder free page reporting from grabbing the page,
91 * reporting it and marking it "reported" - it only skips notifying
92 * the free page reporting infrastructure about a newly freed page. For
93 * example, used when temporarily pulling a page from a freelist and
94 * putting it back unmodified.
96 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
99 * Place the (possibly merged) page to the tail of the freelist. Will ignore
100 * page shuffling (relevant code - e.g., memory onlining - is expected to
101 * shuffle the whole zone).
103 * Note: No code should rely on this flag for correctness - it's purely
104 * to allow for optimizations when handing back either fresh pages
105 * (memory onlining) or untouched pages (page isolation, free page
108 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
111 * Don't poison memory with KASAN (only for the tag-based modes).
112 * During boot, all non-reserved memblock memory is exposed to page_alloc.
113 * Poisoning all that memory lengthens boot time, especially on systems with
114 * large amount of RAM. This flag is used to skip that poisoning.
115 * This is only done for the tag-based KASAN modes, as those are able to
116 * detect memory corruptions with the memory tags assigned by default.
117 * All memory allocated normally after boot gets poisoned as usual.
119 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
121 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
122 static DEFINE_MUTEX(pcp_batch_high_lock);
123 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
128 static DEFINE_PER_CPU(struct pagesets, pagesets) = {
129 .lock = INIT_LOCAL_LOCK(lock),
132 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
133 DEFINE_PER_CPU(int, numa_node);
134 EXPORT_PER_CPU_SYMBOL(numa_node);
137 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
139 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
141 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
142 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
143 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
144 * defined in <linux/topology.h>.
146 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
147 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
150 /* work_structs for global per-cpu drains */
153 struct work_struct work;
155 static DEFINE_MUTEX(pcpu_drain_mutex);
156 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
158 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
159 volatile unsigned long latent_entropy __latent_entropy;
160 EXPORT_SYMBOL(latent_entropy);
164 * Array of node states.
166 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
167 [N_POSSIBLE] = NODE_MASK_ALL,
168 [N_ONLINE] = { { [0] = 1UL } },
170 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
171 #ifdef CONFIG_HIGHMEM
172 [N_HIGH_MEMORY] = { { [0] = 1UL } },
174 [N_MEMORY] = { { [0] = 1UL } },
175 [N_CPU] = { { [0] = 1UL } },
178 EXPORT_SYMBOL(node_states);
180 atomic_long_t _totalram_pages __read_mostly;
181 EXPORT_SYMBOL(_totalram_pages);
182 unsigned long totalreserve_pages __read_mostly;
183 unsigned long totalcma_pages __read_mostly;
185 int percpu_pagelist_high_fraction;
186 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
187 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
188 EXPORT_SYMBOL(init_on_alloc);
190 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
191 EXPORT_SYMBOL(init_on_free);
193 static bool _init_on_alloc_enabled_early __read_mostly
194 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
195 static int __init early_init_on_alloc(char *buf)
198 return kstrtobool(buf, &_init_on_alloc_enabled_early);
200 early_param("init_on_alloc", early_init_on_alloc);
202 static bool _init_on_free_enabled_early __read_mostly
203 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
204 static int __init early_init_on_free(char *buf)
206 return kstrtobool(buf, &_init_on_free_enabled_early);
208 early_param("init_on_free", early_init_on_free);
211 * A cached value of the page's pageblock's migratetype, used when the page is
212 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
213 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
214 * Also the migratetype set in the page does not necessarily match the pcplist
215 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
216 * other index - this ensures that it will be put on the correct CMA freelist.
218 static inline int get_pcppage_migratetype(struct page *page)
223 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
225 page->index = migratetype;
228 #ifdef CONFIG_PM_SLEEP
230 * The following functions are used by the suspend/hibernate code to temporarily
231 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
232 * while devices are suspended. To avoid races with the suspend/hibernate code,
233 * they should always be called with system_transition_mutex held
234 * (gfp_allowed_mask also should only be modified with system_transition_mutex
235 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
236 * with that modification).
239 static gfp_t saved_gfp_mask;
241 void pm_restore_gfp_mask(void)
243 WARN_ON(!mutex_is_locked(&system_transition_mutex));
244 if (saved_gfp_mask) {
245 gfp_allowed_mask = saved_gfp_mask;
250 void pm_restrict_gfp_mask(void)
252 WARN_ON(!mutex_is_locked(&system_transition_mutex));
253 WARN_ON(saved_gfp_mask);
254 saved_gfp_mask = gfp_allowed_mask;
255 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
258 bool pm_suspended_storage(void)
260 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
264 #endif /* CONFIG_PM_SLEEP */
266 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
267 unsigned int pageblock_order __read_mostly;
270 static void __free_pages_ok(struct page *page, unsigned int order,
274 * results with 256, 32 in the lowmem_reserve sysctl:
275 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
276 * 1G machine -> (16M dma, 784M normal, 224M high)
277 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
278 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
279 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
281 * TBD: should special case ZONE_DMA32 machines here - in those we normally
282 * don't need any ZONE_NORMAL reservation
284 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
285 #ifdef CONFIG_ZONE_DMA
288 #ifdef CONFIG_ZONE_DMA32
292 #ifdef CONFIG_HIGHMEM
298 static char * const zone_names[MAX_NR_ZONES] = {
299 #ifdef CONFIG_ZONE_DMA
302 #ifdef CONFIG_ZONE_DMA32
306 #ifdef CONFIG_HIGHMEM
310 #ifdef CONFIG_ZONE_DEVICE
315 const char * const migratetype_names[MIGRATE_TYPES] = {
323 #ifdef CONFIG_MEMORY_ISOLATION
328 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
329 [NULL_COMPOUND_DTOR] = NULL,
330 [COMPOUND_PAGE_DTOR] = free_compound_page,
331 #ifdef CONFIG_HUGETLB_PAGE
332 [HUGETLB_PAGE_DTOR] = free_huge_page,
334 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
335 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
339 int min_free_kbytes = 1024;
340 int user_min_free_kbytes = -1;
341 int watermark_boost_factor __read_mostly = 15000;
342 int watermark_scale_factor = 10;
344 static unsigned long nr_kernel_pages __initdata;
345 static unsigned long nr_all_pages __initdata;
346 static unsigned long dma_reserve __initdata;
348 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
349 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
350 static unsigned long required_kernelcore __initdata;
351 static unsigned long required_kernelcore_percent __initdata;
352 static unsigned long required_movablecore __initdata;
353 static unsigned long required_movablecore_percent __initdata;
354 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
355 static bool mirrored_kernelcore __meminitdata;
357 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
359 EXPORT_SYMBOL(movable_zone);
362 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
363 unsigned int nr_online_nodes __read_mostly = 1;
364 EXPORT_SYMBOL(nr_node_ids);
365 EXPORT_SYMBOL(nr_online_nodes);
368 int page_group_by_mobility_disabled __read_mostly;
370 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
372 * During boot we initialize deferred pages on-demand, as needed, but once
373 * page_alloc_init_late() has finished, the deferred pages are all initialized,
374 * and we can permanently disable that path.
376 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
379 * Calling kasan_poison_pages() only after deferred memory initialization
380 * has completed. Poisoning pages during deferred memory init will greatly
381 * lengthen the process and cause problem in large memory systems as the
382 * deferred pages initialization is done with interrupt disabled.
384 * Assuming that there will be no reference to those newly initialized
385 * pages before they are ever allocated, this should have no effect on
386 * KASAN memory tracking as the poison will be properly inserted at page
387 * allocation time. The only corner case is when pages are allocated by
388 * on-demand allocation and then freed again before the deferred pages
389 * initialization is done, but this is not likely to happen.
391 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
393 return static_branch_unlikely(&deferred_pages) ||
394 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
395 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
396 PageSkipKASanPoison(page);
399 /* Returns true if the struct page for the pfn is uninitialised */
400 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
402 int nid = early_pfn_to_nid(pfn);
404 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
411 * Returns true when the remaining initialisation should be deferred until
412 * later in the boot cycle when it can be parallelised.
414 static bool __meminit
415 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
417 static unsigned long prev_end_pfn, nr_initialised;
420 * prev_end_pfn static that contains the end of previous zone
421 * No need to protect because called very early in boot before smp_init.
423 if (prev_end_pfn != end_pfn) {
424 prev_end_pfn = end_pfn;
428 /* Always populate low zones for address-constrained allocations */
429 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
432 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
435 * We start only with one section of pages, more pages are added as
436 * needed until the rest of deferred pages are initialized.
439 if ((nr_initialised > PAGES_PER_SECTION) &&
440 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
441 NODE_DATA(nid)->first_deferred_pfn = pfn;
447 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
449 return (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
450 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
451 PageSkipKASanPoison(page);
454 static inline bool early_page_uninitialised(unsigned long pfn)
459 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
465 /* Return a pointer to the bitmap storing bits affecting a block of pages */
466 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
469 #ifdef CONFIG_SPARSEMEM
470 return section_to_usemap(__pfn_to_section(pfn));
472 return page_zone(page)->pageblock_flags;
473 #endif /* CONFIG_SPARSEMEM */
476 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
478 #ifdef CONFIG_SPARSEMEM
479 pfn &= (PAGES_PER_SECTION-1);
481 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
482 #endif /* CONFIG_SPARSEMEM */
483 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
486 static __always_inline
487 unsigned long __get_pfnblock_flags_mask(const struct page *page,
491 unsigned long *bitmap;
492 unsigned long bitidx, word_bitidx;
495 bitmap = get_pageblock_bitmap(page, pfn);
496 bitidx = pfn_to_bitidx(page, pfn);
497 word_bitidx = bitidx / BITS_PER_LONG;
498 bitidx &= (BITS_PER_LONG-1);
500 word = bitmap[word_bitidx];
501 return (word >> bitidx) & mask;
505 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
506 * @page: The page within the block of interest
507 * @pfn: The target page frame number
508 * @mask: mask of bits that the caller is interested in
510 * Return: pageblock_bits flags
512 unsigned long get_pfnblock_flags_mask(const struct page *page,
513 unsigned long pfn, unsigned long mask)
515 return __get_pfnblock_flags_mask(page, pfn, mask);
518 static __always_inline int get_pfnblock_migratetype(const struct page *page,
521 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
525 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
526 * @page: The page within the block of interest
527 * @flags: The flags to set
528 * @pfn: The target page frame number
529 * @mask: mask of bits that the caller is interested in
531 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
535 unsigned long *bitmap;
536 unsigned long bitidx, word_bitidx;
537 unsigned long old_word, word;
539 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
540 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
542 bitmap = get_pageblock_bitmap(page, pfn);
543 bitidx = pfn_to_bitidx(page, pfn);
544 word_bitidx = bitidx / BITS_PER_LONG;
545 bitidx &= (BITS_PER_LONG-1);
547 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
552 word = READ_ONCE(bitmap[word_bitidx]);
554 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
555 if (word == old_word)
561 void set_pageblock_migratetype(struct page *page, int migratetype)
563 if (unlikely(page_group_by_mobility_disabled &&
564 migratetype < MIGRATE_PCPTYPES))
565 migratetype = MIGRATE_UNMOVABLE;
567 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
568 page_to_pfn(page), MIGRATETYPE_MASK);
571 #ifdef CONFIG_DEBUG_VM
572 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
576 unsigned long pfn = page_to_pfn(page);
577 unsigned long sp, start_pfn;
580 seq = zone_span_seqbegin(zone);
581 start_pfn = zone->zone_start_pfn;
582 sp = zone->spanned_pages;
583 if (!zone_spans_pfn(zone, pfn))
585 } while (zone_span_seqretry(zone, seq));
588 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
589 pfn, zone_to_nid(zone), zone->name,
590 start_pfn, start_pfn + sp);
595 static int page_is_consistent(struct zone *zone, struct page *page)
597 if (zone != page_zone(page))
603 * Temporary debugging check for pages not lying within a given zone.
605 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
607 if (page_outside_zone_boundaries(zone, page))
609 if (!page_is_consistent(zone, page))
615 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
621 static void bad_page(struct page *page, const char *reason)
623 static unsigned long resume;
624 static unsigned long nr_shown;
625 static unsigned long nr_unshown;
628 * Allow a burst of 60 reports, then keep quiet for that minute;
629 * or allow a steady drip of one report per second.
631 if (nr_shown == 60) {
632 if (time_before(jiffies, resume)) {
638 "BUG: Bad page state: %lu messages suppressed\n",
645 resume = jiffies + 60 * HZ;
647 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
648 current->comm, page_to_pfn(page));
649 dump_page(page, reason);
654 /* Leave bad fields for debug, except PageBuddy could make trouble */
655 page_mapcount_reset(page); /* remove PageBuddy */
656 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
659 static inline unsigned int order_to_pindex(int migratetype, int order)
663 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
664 if (order > PAGE_ALLOC_COSTLY_ORDER) {
665 VM_BUG_ON(order != pageblock_order);
666 base = PAGE_ALLOC_COSTLY_ORDER + 1;
669 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
672 return (MIGRATE_PCPTYPES * base) + migratetype;
675 static inline int pindex_to_order(unsigned int pindex)
677 int order = pindex / MIGRATE_PCPTYPES;
679 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
680 if (order > PAGE_ALLOC_COSTLY_ORDER) {
681 order = pageblock_order;
682 VM_BUG_ON(order != pageblock_order);
685 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
691 static inline bool pcp_allowed_order(unsigned int order)
693 if (order <= PAGE_ALLOC_COSTLY_ORDER)
695 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
696 if (order == pageblock_order)
702 static inline void free_the_page(struct page *page, unsigned int order)
704 if (pcp_allowed_order(order)) /* Via pcp? */
705 free_unref_page(page, order);
707 __free_pages_ok(page, order, FPI_NONE);
711 * Higher-order pages are called "compound pages". They are structured thusly:
713 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
715 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
716 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
718 * The first tail page's ->compound_dtor holds the offset in array of compound
719 * page destructors. See compound_page_dtors.
721 * The first tail page's ->compound_order holds the order of allocation.
722 * This usage means that zero-order pages may not be compound.
725 void free_compound_page(struct page *page)
727 mem_cgroup_uncharge(page);
728 free_the_page(page, compound_order(page));
731 void prep_compound_page(struct page *page, unsigned int order)
734 int nr_pages = 1 << order;
737 for (i = 1; i < nr_pages; i++) {
738 struct page *p = page + i;
739 p->mapping = TAIL_MAPPING;
740 set_compound_head(p, page);
743 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
744 set_compound_order(page, order);
745 atomic_set(compound_mapcount_ptr(page), -1);
746 if (hpage_pincount_available(page))
747 atomic_set(compound_pincount_ptr(page), 0);
750 #ifdef CONFIG_DEBUG_PAGEALLOC
751 unsigned int _debug_guardpage_minorder;
753 bool _debug_pagealloc_enabled_early __read_mostly
754 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
755 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
756 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
757 EXPORT_SYMBOL(_debug_pagealloc_enabled);
759 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
761 static int __init early_debug_pagealloc(char *buf)
763 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
765 early_param("debug_pagealloc", early_debug_pagealloc);
767 static int __init debug_guardpage_minorder_setup(char *buf)
771 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
772 pr_err("Bad debug_guardpage_minorder value\n");
775 _debug_guardpage_minorder = res;
776 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
779 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
781 static inline bool set_page_guard(struct zone *zone, struct page *page,
782 unsigned int order, int migratetype)
784 if (!debug_guardpage_enabled())
787 if (order >= debug_guardpage_minorder())
790 __SetPageGuard(page);
791 INIT_LIST_HEAD(&page->lru);
792 set_page_private(page, order);
793 /* Guard pages are not available for any usage */
794 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
799 static inline void clear_page_guard(struct zone *zone, struct page *page,
800 unsigned int order, int migratetype)
802 if (!debug_guardpage_enabled())
805 __ClearPageGuard(page);
807 set_page_private(page, 0);
808 if (!is_migrate_isolate(migratetype))
809 __mod_zone_freepage_state(zone, (1 << order), migratetype);
812 static inline bool set_page_guard(struct zone *zone, struct page *page,
813 unsigned int order, int migratetype) { return false; }
814 static inline void clear_page_guard(struct zone *zone, struct page *page,
815 unsigned int order, int migratetype) {}
819 * Enable static keys related to various memory debugging and hardening options.
820 * Some override others, and depend on early params that are evaluated in the
821 * order of appearance. So we need to first gather the full picture of what was
822 * enabled, and then make decisions.
824 void init_mem_debugging_and_hardening(void)
826 bool page_poisoning_requested = false;
828 #ifdef CONFIG_PAGE_POISONING
830 * Page poisoning is debug page alloc for some arches. If
831 * either of those options are enabled, enable poisoning.
833 if (page_poisoning_enabled() ||
834 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
835 debug_pagealloc_enabled())) {
836 static_branch_enable(&_page_poisoning_enabled);
837 page_poisoning_requested = true;
841 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
842 page_poisoning_requested) {
843 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
844 "will take precedence over init_on_alloc and init_on_free\n");
845 _init_on_alloc_enabled_early = false;
846 _init_on_free_enabled_early = false;
849 if (_init_on_alloc_enabled_early)
850 static_branch_enable(&init_on_alloc);
852 static_branch_disable(&init_on_alloc);
854 if (_init_on_free_enabled_early)
855 static_branch_enable(&init_on_free);
857 static_branch_disable(&init_on_free);
859 #ifdef CONFIG_DEBUG_PAGEALLOC
860 if (!debug_pagealloc_enabled())
863 static_branch_enable(&_debug_pagealloc_enabled);
865 if (!debug_guardpage_minorder())
868 static_branch_enable(&_debug_guardpage_enabled);
872 static inline void set_buddy_order(struct page *page, unsigned int order)
874 set_page_private(page, order);
875 __SetPageBuddy(page);
879 * This function checks whether a page is free && is the buddy
880 * we can coalesce a page and its buddy if
881 * (a) the buddy is not in a hole (check before calling!) &&
882 * (b) the buddy is in the buddy system &&
883 * (c) a page and its buddy have the same order &&
884 * (d) a page and its buddy are in the same zone.
886 * For recording whether a page is in the buddy system, we set PageBuddy.
887 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
889 * For recording page's order, we use page_private(page).
891 static inline bool page_is_buddy(struct page *page, struct page *buddy,
894 if (!page_is_guard(buddy) && !PageBuddy(buddy))
897 if (buddy_order(buddy) != order)
901 * zone check is done late to avoid uselessly calculating
902 * zone/node ids for pages that could never merge.
904 if (page_zone_id(page) != page_zone_id(buddy))
907 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
912 #ifdef CONFIG_COMPACTION
913 static inline struct capture_control *task_capc(struct zone *zone)
915 struct capture_control *capc = current->capture_control;
917 return unlikely(capc) &&
918 !(current->flags & PF_KTHREAD) &&
920 capc->cc->zone == zone ? capc : NULL;
924 compaction_capture(struct capture_control *capc, struct page *page,
925 int order, int migratetype)
927 if (!capc || order != capc->cc->order)
930 /* Do not accidentally pollute CMA or isolated regions*/
931 if (is_migrate_cma(migratetype) ||
932 is_migrate_isolate(migratetype))
936 * Do not let lower order allocations pollute a movable pageblock.
937 * This might let an unmovable request use a reclaimable pageblock
938 * and vice-versa but no more than normal fallback logic which can
939 * have trouble finding a high-order free page.
941 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
949 static inline struct capture_control *task_capc(struct zone *zone)
955 compaction_capture(struct capture_control *capc, struct page *page,
956 int order, int migratetype)
960 #endif /* CONFIG_COMPACTION */
962 /* Used for pages not on another list */
963 static inline void add_to_free_list(struct page *page, struct zone *zone,
964 unsigned int order, int migratetype)
966 struct free_area *area = &zone->free_area[order];
968 list_add(&page->lru, &area->free_list[migratetype]);
972 /* Used for pages not on another list */
973 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
974 unsigned int order, int migratetype)
976 struct free_area *area = &zone->free_area[order];
978 list_add_tail(&page->lru, &area->free_list[migratetype]);
983 * Used for pages which are on another list. Move the pages to the tail
984 * of the list - so the moved pages won't immediately be considered for
985 * allocation again (e.g., optimization for memory onlining).
987 static inline void move_to_free_list(struct page *page, struct zone *zone,
988 unsigned int order, int migratetype)
990 struct free_area *area = &zone->free_area[order];
992 list_move_tail(&page->lru, &area->free_list[migratetype]);
995 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
998 /* clear reported state and update reported page count */
999 if (page_reported(page))
1000 __ClearPageReported(page);
1002 list_del(&page->lru);
1003 __ClearPageBuddy(page);
1004 set_page_private(page, 0);
1005 zone->free_area[order].nr_free--;
1009 * If this is not the largest possible page, check if the buddy
1010 * of the next-highest order is free. If it is, it's possible
1011 * that pages are being freed that will coalesce soon. In case,
1012 * that is happening, add the free page to the tail of the list
1013 * so it's less likely to be used soon and more likely to be merged
1014 * as a higher order page
1017 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1018 struct page *page, unsigned int order)
1020 struct page *higher_page, *higher_buddy;
1021 unsigned long combined_pfn;
1023 if (order >= MAX_ORDER - 2)
1026 combined_pfn = buddy_pfn & pfn;
1027 higher_page = page + (combined_pfn - pfn);
1028 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1029 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1031 return page_is_buddy(higher_page, higher_buddy, order + 1);
1035 * Freeing function for a buddy system allocator.
1037 * The concept of a buddy system is to maintain direct-mapped table
1038 * (containing bit values) for memory blocks of various "orders".
1039 * The bottom level table contains the map for the smallest allocatable
1040 * units of memory (here, pages), and each level above it describes
1041 * pairs of units from the levels below, hence, "buddies".
1042 * At a high level, all that happens here is marking the table entry
1043 * at the bottom level available, and propagating the changes upward
1044 * as necessary, plus some accounting needed to play nicely with other
1045 * parts of the VM system.
1046 * At each level, we keep a list of pages, which are heads of continuous
1047 * free pages of length of (1 << order) and marked with PageBuddy.
1048 * Page's order is recorded in page_private(page) field.
1049 * So when we are allocating or freeing one, we can derive the state of the
1050 * other. That is, if we allocate a small block, and both were
1051 * free, the remainder of the region must be split into blocks.
1052 * If a block is freed, and its buddy is also free, then this
1053 * triggers coalescing into a block of larger size.
1058 static inline void __free_one_page(struct page *page,
1060 struct zone *zone, unsigned int order,
1061 int migratetype, fpi_t fpi_flags)
1063 struct capture_control *capc = task_capc(zone);
1064 unsigned long buddy_pfn;
1065 unsigned long combined_pfn;
1066 unsigned int max_order;
1070 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1072 VM_BUG_ON(!zone_is_initialized(zone));
1073 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1075 VM_BUG_ON(migratetype == -1);
1076 if (likely(!is_migrate_isolate(migratetype)))
1077 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1079 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1080 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1083 while (order < max_order) {
1084 if (compaction_capture(capc, page, order, migratetype)) {
1085 __mod_zone_freepage_state(zone, -(1 << order),
1089 buddy_pfn = __find_buddy_pfn(pfn, order);
1090 buddy = page + (buddy_pfn - pfn);
1092 if (!page_is_buddy(page, buddy, order))
1095 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1096 * merge with it and move up one order.
1098 if (page_is_guard(buddy))
1099 clear_page_guard(zone, buddy, order, migratetype);
1101 del_page_from_free_list(buddy, zone, order);
1102 combined_pfn = buddy_pfn & pfn;
1103 page = page + (combined_pfn - pfn);
1107 if (order < MAX_ORDER - 1) {
1108 /* If we are here, it means order is >= pageblock_order.
1109 * We want to prevent merge between freepages on isolate
1110 * pageblock and normal pageblock. Without this, pageblock
1111 * isolation could cause incorrect freepage or CMA accounting.
1113 * We don't want to hit this code for the more frequent
1114 * low-order merging.
1116 if (unlikely(has_isolate_pageblock(zone))) {
1119 buddy_pfn = __find_buddy_pfn(pfn, order);
1120 buddy = page + (buddy_pfn - pfn);
1121 buddy_mt = get_pageblock_migratetype(buddy);
1123 if (migratetype != buddy_mt
1124 && (is_migrate_isolate(migratetype) ||
1125 is_migrate_isolate(buddy_mt)))
1128 max_order = order + 1;
1129 goto continue_merging;
1133 set_buddy_order(page, order);
1135 if (fpi_flags & FPI_TO_TAIL)
1137 else if (is_shuffle_order(order))
1138 to_tail = shuffle_pick_tail();
1140 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1143 add_to_free_list_tail(page, zone, order, migratetype);
1145 add_to_free_list(page, zone, order, migratetype);
1147 /* Notify page reporting subsystem of freed page */
1148 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1149 page_reporting_notify_free(order);
1153 * A bad page could be due to a number of fields. Instead of multiple branches,
1154 * try and check multiple fields with one check. The caller must do a detailed
1155 * check if necessary.
1157 static inline bool page_expected_state(struct page *page,
1158 unsigned long check_flags)
1160 if (unlikely(atomic_read(&page->_mapcount) != -1))
1163 if (unlikely((unsigned long)page->mapping |
1164 page_ref_count(page) |
1168 (page->flags & check_flags)))
1174 static const char *page_bad_reason(struct page *page, unsigned long flags)
1176 const char *bad_reason = NULL;
1178 if (unlikely(atomic_read(&page->_mapcount) != -1))
1179 bad_reason = "nonzero mapcount";
1180 if (unlikely(page->mapping != NULL))
1181 bad_reason = "non-NULL mapping";
1182 if (unlikely(page_ref_count(page) != 0))
1183 bad_reason = "nonzero _refcount";
1184 if (unlikely(page->flags & flags)) {
1185 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1186 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1188 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1191 if (unlikely(page->memcg_data))
1192 bad_reason = "page still charged to cgroup";
1197 static void check_free_page_bad(struct page *page)
1200 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1203 static inline int check_free_page(struct page *page)
1205 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1208 /* Something has gone sideways, find it */
1209 check_free_page_bad(page);
1213 static int free_tail_pages_check(struct page *head_page, struct page *page)
1218 * We rely page->lru.next never has bit 0 set, unless the page
1219 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1221 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1223 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1227 switch (page - head_page) {
1229 /* the first tail page: ->mapping may be compound_mapcount() */
1230 if (unlikely(compound_mapcount(page))) {
1231 bad_page(page, "nonzero compound_mapcount");
1237 * the second tail page: ->mapping is
1238 * deferred_list.next -- ignore value.
1242 if (page->mapping != TAIL_MAPPING) {
1243 bad_page(page, "corrupted mapping in tail page");
1248 if (unlikely(!PageTail(page))) {
1249 bad_page(page, "PageTail not set");
1252 if (unlikely(compound_head(page) != head_page)) {
1253 bad_page(page, "compound_head not consistent");
1258 page->mapping = NULL;
1259 clear_compound_head(page);
1263 static void kernel_init_free_pages(struct page *page, int numpages, bool zero_tags)
1268 for (i = 0; i < numpages; i++)
1269 tag_clear_highpage(page + i);
1273 /* s390's use of memset() could override KASAN redzones. */
1274 kasan_disable_current();
1275 for (i = 0; i < numpages; i++) {
1276 u8 tag = page_kasan_tag(page + i);
1277 page_kasan_tag_reset(page + i);
1278 clear_highpage(page + i);
1279 page_kasan_tag_set(page + i, tag);
1281 kasan_enable_current();
1284 static __always_inline bool free_pages_prepare(struct page *page,
1285 unsigned int order, bool check_free, fpi_t fpi_flags)
1288 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1290 VM_BUG_ON_PAGE(PageTail(page), page);
1292 trace_mm_page_free(page, order);
1294 if (unlikely(PageHWPoison(page)) && !order) {
1296 * Do not let hwpoison pages hit pcplists/buddy
1297 * Untie memcg state and reset page's owner
1299 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1300 __memcg_kmem_uncharge_page(page, order);
1301 reset_page_owner(page, order);
1306 * Check tail pages before head page information is cleared to
1307 * avoid checking PageCompound for order-0 pages.
1309 if (unlikely(order)) {
1310 bool compound = PageCompound(page);
1313 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1316 ClearPageDoubleMap(page);
1317 ClearPageHasHWPoisoned(page);
1319 for (i = 1; i < (1 << order); i++) {
1321 bad += free_tail_pages_check(page, page + i);
1322 if (unlikely(check_free_page(page + i))) {
1326 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1329 if (PageMappingFlags(page))
1330 page->mapping = NULL;
1331 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1332 __memcg_kmem_uncharge_page(page, order);
1334 bad += check_free_page(page);
1338 page_cpupid_reset_last(page);
1339 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1340 reset_page_owner(page, order);
1342 if (!PageHighMem(page)) {
1343 debug_check_no_locks_freed(page_address(page),
1344 PAGE_SIZE << order);
1345 debug_check_no_obj_freed(page_address(page),
1346 PAGE_SIZE << order);
1349 kernel_poison_pages(page, 1 << order);
1352 * As memory initialization might be integrated into KASAN,
1353 * kasan_free_pages and kernel_init_free_pages must be
1354 * kept together to avoid discrepancies in behavior.
1356 * With hardware tag-based KASAN, memory tags must be set before the
1357 * page becomes unavailable via debug_pagealloc or arch_free_page.
1359 if (kasan_has_integrated_init()) {
1360 if (!skip_kasan_poison)
1361 kasan_free_pages(page, order);
1363 bool init = want_init_on_free();
1366 kernel_init_free_pages(page, 1 << order, false);
1367 if (!skip_kasan_poison)
1368 kasan_poison_pages(page, order, init);
1372 * arch_free_page() can make the page's contents inaccessible. s390
1373 * does this. So nothing which can access the page's contents should
1374 * happen after this.
1376 arch_free_page(page, order);
1378 debug_pagealloc_unmap_pages(page, 1 << order);
1383 #ifdef CONFIG_DEBUG_VM
1385 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1386 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1387 * moved from pcp lists to free lists.
1389 static bool free_pcp_prepare(struct page *page, unsigned int order)
1391 return free_pages_prepare(page, order, true, FPI_NONE);
1394 static bool bulkfree_pcp_prepare(struct page *page)
1396 if (debug_pagealloc_enabled_static())
1397 return check_free_page(page);
1403 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1404 * moving from pcp lists to free list in order to reduce overhead. With
1405 * debug_pagealloc enabled, they are checked also immediately when being freed
1408 static bool free_pcp_prepare(struct page *page, unsigned int order)
1410 if (debug_pagealloc_enabled_static())
1411 return free_pages_prepare(page, order, true, FPI_NONE);
1413 return free_pages_prepare(page, order, false, FPI_NONE);
1416 static bool bulkfree_pcp_prepare(struct page *page)
1418 return check_free_page(page);
1420 #endif /* CONFIG_DEBUG_VM */
1422 static inline void prefetch_buddy(struct page *page)
1424 unsigned long pfn = page_to_pfn(page);
1425 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1426 struct page *buddy = page + (buddy_pfn - pfn);
1432 * Frees a number of pages from the PCP lists
1433 * Assumes all pages on list are in same zone, and of same order.
1434 * count is the number of pages to free.
1436 * If the zone was previously in an "all pages pinned" state then look to
1437 * see if this freeing clears that state.
1439 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1440 * pinned" detection logic.
1442 static void free_pcppages_bulk(struct zone *zone, int count,
1443 struct per_cpu_pages *pcp)
1449 int prefetch_nr = READ_ONCE(pcp->batch);
1450 bool isolated_pageblocks;
1451 struct page *page, *tmp;
1455 * Ensure proper count is passed which otherwise would stuck in the
1456 * below while (list_empty(list)) loop.
1458 count = min(pcp->count, count);
1460 struct list_head *list;
1463 * Remove pages from lists in a round-robin fashion. A
1464 * batch_free count is maintained that is incremented when an
1465 * empty list is encountered. This is so more pages are freed
1466 * off fuller lists instead of spinning excessively around empty
1471 if (++pindex == NR_PCP_LISTS)
1473 list = &pcp->lists[pindex];
1474 } while (list_empty(list));
1476 /* This is the only non-empty list. Free them all. */
1477 if (batch_free == NR_PCP_LISTS)
1480 order = pindex_to_order(pindex);
1481 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1483 page = list_last_entry(list, struct page, lru);
1484 /* must delete to avoid corrupting pcp list */
1485 list_del(&page->lru);
1486 nr_freed += 1 << order;
1487 count -= 1 << order;
1489 if (bulkfree_pcp_prepare(page))
1492 /* Encode order with the migratetype */
1493 page->index <<= NR_PCP_ORDER_WIDTH;
1494 page->index |= order;
1496 list_add_tail(&page->lru, &head);
1499 * We are going to put the page back to the global
1500 * pool, prefetch its buddy to speed up later access
1501 * under zone->lock. It is believed the overhead of
1502 * an additional test and calculating buddy_pfn here
1503 * can be offset by reduced memory latency later. To
1504 * avoid excessive prefetching due to large count, only
1505 * prefetch buddy for the first pcp->batch nr of pages.
1508 prefetch_buddy(page);
1511 } while (count > 0 && --batch_free && !list_empty(list));
1513 pcp->count -= nr_freed;
1516 * local_lock_irq held so equivalent to spin_lock_irqsave for
1517 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1519 spin_lock(&zone->lock);
1520 isolated_pageblocks = has_isolate_pageblock(zone);
1523 * Use safe version since after __free_one_page(),
1524 * page->lru.next will not point to original list.
1526 list_for_each_entry_safe(page, tmp, &head, lru) {
1527 int mt = get_pcppage_migratetype(page);
1529 /* mt has been encoded with the order (see above) */
1530 order = mt & NR_PCP_ORDER_MASK;
1531 mt >>= NR_PCP_ORDER_WIDTH;
1533 /* MIGRATE_ISOLATE page should not go to pcplists */
1534 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1535 /* Pageblock could have been isolated meanwhile */
1536 if (unlikely(isolated_pageblocks))
1537 mt = get_pageblock_migratetype(page);
1539 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1540 trace_mm_page_pcpu_drain(page, order, mt);
1542 spin_unlock(&zone->lock);
1545 static void free_one_page(struct zone *zone,
1546 struct page *page, unsigned long pfn,
1548 int migratetype, fpi_t fpi_flags)
1550 unsigned long flags;
1552 spin_lock_irqsave(&zone->lock, flags);
1553 if (unlikely(has_isolate_pageblock(zone) ||
1554 is_migrate_isolate(migratetype))) {
1555 migratetype = get_pfnblock_migratetype(page, pfn);
1557 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1558 spin_unlock_irqrestore(&zone->lock, flags);
1561 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1562 unsigned long zone, int nid)
1564 mm_zero_struct_page(page);
1565 set_page_links(page, zone, nid, pfn);
1566 init_page_count(page);
1567 page_mapcount_reset(page);
1568 page_cpupid_reset_last(page);
1569 page_kasan_tag_reset(page);
1571 INIT_LIST_HEAD(&page->lru);
1572 #ifdef WANT_PAGE_VIRTUAL
1573 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1574 if (!is_highmem_idx(zone))
1575 set_page_address(page, __va(pfn << PAGE_SHIFT));
1579 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1580 static void __meminit init_reserved_page(unsigned long pfn)
1585 if (!early_page_uninitialised(pfn))
1588 nid = early_pfn_to_nid(pfn);
1589 pgdat = NODE_DATA(nid);
1591 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1592 struct zone *zone = &pgdat->node_zones[zid];
1594 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1597 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1600 static inline void init_reserved_page(unsigned long pfn)
1603 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1606 * Initialised pages do not have PageReserved set. This function is
1607 * called for each range allocated by the bootmem allocator and
1608 * marks the pages PageReserved. The remaining valid pages are later
1609 * sent to the buddy page allocator.
1611 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1613 unsigned long start_pfn = PFN_DOWN(start);
1614 unsigned long end_pfn = PFN_UP(end);
1616 for (; start_pfn < end_pfn; start_pfn++) {
1617 if (pfn_valid(start_pfn)) {
1618 struct page *page = pfn_to_page(start_pfn);
1620 init_reserved_page(start_pfn);
1622 /* Avoid false-positive PageTail() */
1623 INIT_LIST_HEAD(&page->lru);
1626 * no need for atomic set_bit because the struct
1627 * page is not visible yet so nobody should
1630 __SetPageReserved(page);
1635 static void __free_pages_ok(struct page *page, unsigned int order,
1638 unsigned long flags;
1640 unsigned long pfn = page_to_pfn(page);
1641 struct zone *zone = page_zone(page);
1643 if (!free_pages_prepare(page, order, true, fpi_flags))
1646 migratetype = get_pfnblock_migratetype(page, pfn);
1648 spin_lock_irqsave(&zone->lock, flags);
1649 if (unlikely(has_isolate_pageblock(zone) ||
1650 is_migrate_isolate(migratetype))) {
1651 migratetype = get_pfnblock_migratetype(page, pfn);
1653 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1654 spin_unlock_irqrestore(&zone->lock, flags);
1656 __count_vm_events(PGFREE, 1 << order);
1659 void __free_pages_core(struct page *page, unsigned int order)
1661 unsigned int nr_pages = 1 << order;
1662 struct page *p = page;
1666 * When initializing the memmap, __init_single_page() sets the refcount
1667 * of all pages to 1 ("allocated"/"not free"). We have to set the
1668 * refcount of all involved pages to 0.
1671 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1673 __ClearPageReserved(p);
1674 set_page_count(p, 0);
1676 __ClearPageReserved(p);
1677 set_page_count(p, 0);
1679 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1682 * Bypass PCP and place fresh pages right to the tail, primarily
1683 * relevant for memory onlining.
1685 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1691 * During memory init memblocks map pfns to nids. The search is expensive and
1692 * this caches recent lookups. The implementation of __early_pfn_to_nid
1693 * treats start/end as pfns.
1695 struct mminit_pfnnid_cache {
1696 unsigned long last_start;
1697 unsigned long last_end;
1701 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1704 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1706 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1707 struct mminit_pfnnid_cache *state)
1709 unsigned long start_pfn, end_pfn;
1712 if (state->last_start <= pfn && pfn < state->last_end)
1713 return state->last_nid;
1715 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1716 if (nid != NUMA_NO_NODE) {
1717 state->last_start = start_pfn;
1718 state->last_end = end_pfn;
1719 state->last_nid = nid;
1725 int __meminit early_pfn_to_nid(unsigned long pfn)
1727 static DEFINE_SPINLOCK(early_pfn_lock);
1730 spin_lock(&early_pfn_lock);
1731 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1733 nid = first_online_node;
1734 spin_unlock(&early_pfn_lock);
1738 #endif /* CONFIG_NUMA */
1740 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1743 if (early_page_uninitialised(pfn))
1745 __free_pages_core(page, order);
1749 * Check that the whole (or subset of) a pageblock given by the interval of
1750 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1751 * with the migration of free compaction scanner.
1753 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1755 * It's possible on some configurations to have a setup like node0 node1 node0
1756 * i.e. it's possible that all pages within a zones range of pages do not
1757 * belong to a single zone. We assume that a border between node0 and node1
1758 * can occur within a single pageblock, but not a node0 node1 node0
1759 * interleaving within a single pageblock. It is therefore sufficient to check
1760 * the first and last page of a pageblock and avoid checking each individual
1761 * page in a pageblock.
1763 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1764 unsigned long end_pfn, struct zone *zone)
1766 struct page *start_page;
1767 struct page *end_page;
1769 /* end_pfn is one past the range we are checking */
1772 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1775 start_page = pfn_to_online_page(start_pfn);
1779 if (page_zone(start_page) != zone)
1782 end_page = pfn_to_page(end_pfn);
1784 /* This gives a shorter code than deriving page_zone(end_page) */
1785 if (page_zone_id(start_page) != page_zone_id(end_page))
1791 void set_zone_contiguous(struct zone *zone)
1793 unsigned long block_start_pfn = zone->zone_start_pfn;
1794 unsigned long block_end_pfn;
1796 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1797 for (; block_start_pfn < zone_end_pfn(zone);
1798 block_start_pfn = block_end_pfn,
1799 block_end_pfn += pageblock_nr_pages) {
1801 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1803 if (!__pageblock_pfn_to_page(block_start_pfn,
1804 block_end_pfn, zone))
1809 /* We confirm that there is no hole */
1810 zone->contiguous = true;
1813 void clear_zone_contiguous(struct zone *zone)
1815 zone->contiguous = false;
1818 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1819 static void __init deferred_free_range(unsigned long pfn,
1820 unsigned long nr_pages)
1828 page = pfn_to_page(pfn);
1830 /* Free a large naturally-aligned chunk if possible */
1831 if (nr_pages == pageblock_nr_pages &&
1832 (pfn & (pageblock_nr_pages - 1)) == 0) {
1833 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1834 __free_pages_core(page, pageblock_order);
1838 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1839 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1840 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1841 __free_pages_core(page, 0);
1845 /* Completion tracking for deferred_init_memmap() threads */
1846 static atomic_t pgdat_init_n_undone __initdata;
1847 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1849 static inline void __init pgdat_init_report_one_done(void)
1851 if (atomic_dec_and_test(&pgdat_init_n_undone))
1852 complete(&pgdat_init_all_done_comp);
1856 * Returns true if page needs to be initialized or freed to buddy allocator.
1858 * First we check if pfn is valid on architectures where it is possible to have
1859 * holes within pageblock_nr_pages. On systems where it is not possible, this
1860 * function is optimized out.
1862 * Then, we check if a current large page is valid by only checking the validity
1865 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1867 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1873 * Free pages to buddy allocator. Try to free aligned pages in
1874 * pageblock_nr_pages sizes.
1876 static void __init deferred_free_pages(unsigned long pfn,
1877 unsigned long end_pfn)
1879 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1880 unsigned long nr_free = 0;
1882 for (; pfn < end_pfn; pfn++) {
1883 if (!deferred_pfn_valid(pfn)) {
1884 deferred_free_range(pfn - nr_free, nr_free);
1886 } else if (!(pfn & nr_pgmask)) {
1887 deferred_free_range(pfn - nr_free, nr_free);
1893 /* Free the last block of pages to allocator */
1894 deferred_free_range(pfn - nr_free, nr_free);
1898 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1899 * by performing it only once every pageblock_nr_pages.
1900 * Return number of pages initialized.
1902 static unsigned long __init deferred_init_pages(struct zone *zone,
1904 unsigned long end_pfn)
1906 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1907 int nid = zone_to_nid(zone);
1908 unsigned long nr_pages = 0;
1909 int zid = zone_idx(zone);
1910 struct page *page = NULL;
1912 for (; pfn < end_pfn; pfn++) {
1913 if (!deferred_pfn_valid(pfn)) {
1916 } else if (!page || !(pfn & nr_pgmask)) {
1917 page = pfn_to_page(pfn);
1921 __init_single_page(page, pfn, zid, nid);
1928 * This function is meant to pre-load the iterator for the zone init.
1929 * Specifically it walks through the ranges until we are caught up to the
1930 * first_init_pfn value and exits there. If we never encounter the value we
1931 * return false indicating there are no valid ranges left.
1934 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1935 unsigned long *spfn, unsigned long *epfn,
1936 unsigned long first_init_pfn)
1941 * Start out by walking through the ranges in this zone that have
1942 * already been initialized. We don't need to do anything with them
1943 * so we just need to flush them out of the system.
1945 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1946 if (*epfn <= first_init_pfn)
1948 if (*spfn < first_init_pfn)
1949 *spfn = first_init_pfn;
1958 * Initialize and free pages. We do it in two loops: first we initialize
1959 * struct page, then free to buddy allocator, because while we are
1960 * freeing pages we can access pages that are ahead (computing buddy
1961 * page in __free_one_page()).
1963 * In order to try and keep some memory in the cache we have the loop
1964 * broken along max page order boundaries. This way we will not cause
1965 * any issues with the buddy page computation.
1967 static unsigned long __init
1968 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1969 unsigned long *end_pfn)
1971 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1972 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1973 unsigned long nr_pages = 0;
1976 /* First we loop through and initialize the page values */
1977 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1980 if (mo_pfn <= *start_pfn)
1983 t = min(mo_pfn, *end_pfn);
1984 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1986 if (mo_pfn < *end_pfn) {
1987 *start_pfn = mo_pfn;
1992 /* Reset values and now loop through freeing pages as needed */
1995 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2001 t = min(mo_pfn, epfn);
2002 deferred_free_pages(spfn, t);
2012 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2015 unsigned long spfn, epfn;
2016 struct zone *zone = arg;
2019 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2022 * Initialize and free pages in MAX_ORDER sized increments so that we
2023 * can avoid introducing any issues with the buddy allocator.
2025 while (spfn < end_pfn) {
2026 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2031 /* An arch may override for more concurrency. */
2033 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2038 /* Initialise remaining memory on a node */
2039 static int __init deferred_init_memmap(void *data)
2041 pg_data_t *pgdat = data;
2042 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2043 unsigned long spfn = 0, epfn = 0;
2044 unsigned long first_init_pfn, flags;
2045 unsigned long start = jiffies;
2047 int zid, max_threads;
2050 /* Bind memory initialisation thread to a local node if possible */
2051 if (!cpumask_empty(cpumask))
2052 set_cpus_allowed_ptr(current, cpumask);
2054 pgdat_resize_lock(pgdat, &flags);
2055 first_init_pfn = pgdat->first_deferred_pfn;
2056 if (first_init_pfn == ULONG_MAX) {
2057 pgdat_resize_unlock(pgdat, &flags);
2058 pgdat_init_report_one_done();
2062 /* Sanity check boundaries */
2063 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2064 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2065 pgdat->first_deferred_pfn = ULONG_MAX;
2068 * Once we unlock here, the zone cannot be grown anymore, thus if an
2069 * interrupt thread must allocate this early in boot, zone must be
2070 * pre-grown prior to start of deferred page initialization.
2072 pgdat_resize_unlock(pgdat, &flags);
2074 /* Only the highest zone is deferred so find it */
2075 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2076 zone = pgdat->node_zones + zid;
2077 if (first_init_pfn < zone_end_pfn(zone))
2081 /* If the zone is empty somebody else may have cleared out the zone */
2082 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2086 max_threads = deferred_page_init_max_threads(cpumask);
2088 while (spfn < epfn) {
2089 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2090 struct padata_mt_job job = {
2091 .thread_fn = deferred_init_memmap_chunk,
2094 .size = epfn_align - spfn,
2095 .align = PAGES_PER_SECTION,
2096 .min_chunk = PAGES_PER_SECTION,
2097 .max_threads = max_threads,
2100 padata_do_multithreaded(&job);
2101 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2105 /* Sanity check that the next zone really is unpopulated */
2106 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2108 pr_info("node %d deferred pages initialised in %ums\n",
2109 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2111 pgdat_init_report_one_done();
2116 * If this zone has deferred pages, try to grow it by initializing enough
2117 * deferred pages to satisfy the allocation specified by order, rounded up to
2118 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2119 * of SECTION_SIZE bytes by initializing struct pages in increments of
2120 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2122 * Return true when zone was grown, otherwise return false. We return true even
2123 * when we grow less than requested, to let the caller decide if there are
2124 * enough pages to satisfy the allocation.
2126 * Note: We use noinline because this function is needed only during boot, and
2127 * it is called from a __ref function _deferred_grow_zone. This way we are
2128 * making sure that it is not inlined into permanent text section.
2130 static noinline bool __init
2131 deferred_grow_zone(struct zone *zone, unsigned int order)
2133 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2134 pg_data_t *pgdat = zone->zone_pgdat;
2135 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2136 unsigned long spfn, epfn, flags;
2137 unsigned long nr_pages = 0;
2140 /* Only the last zone may have deferred pages */
2141 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2144 pgdat_resize_lock(pgdat, &flags);
2147 * If someone grew this zone while we were waiting for spinlock, return
2148 * true, as there might be enough pages already.
2150 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2151 pgdat_resize_unlock(pgdat, &flags);
2155 /* If the zone is empty somebody else may have cleared out the zone */
2156 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2157 first_deferred_pfn)) {
2158 pgdat->first_deferred_pfn = ULONG_MAX;
2159 pgdat_resize_unlock(pgdat, &flags);
2160 /* Retry only once. */
2161 return first_deferred_pfn != ULONG_MAX;
2165 * Initialize and free pages in MAX_ORDER sized increments so
2166 * that we can avoid introducing any issues with the buddy
2169 while (spfn < epfn) {
2170 /* update our first deferred PFN for this section */
2171 first_deferred_pfn = spfn;
2173 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2174 touch_nmi_watchdog();
2176 /* We should only stop along section boundaries */
2177 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2180 /* If our quota has been met we can stop here */
2181 if (nr_pages >= nr_pages_needed)
2185 pgdat->first_deferred_pfn = spfn;
2186 pgdat_resize_unlock(pgdat, &flags);
2188 return nr_pages > 0;
2192 * deferred_grow_zone() is __init, but it is called from
2193 * get_page_from_freelist() during early boot until deferred_pages permanently
2194 * disables this call. This is why we have refdata wrapper to avoid warning,
2195 * and to ensure that the function body gets unloaded.
2198 _deferred_grow_zone(struct zone *zone, unsigned int order)
2200 return deferred_grow_zone(zone, order);
2203 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2205 void __init page_alloc_init_late(void)
2210 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2212 /* There will be num_node_state(N_MEMORY) threads */
2213 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2214 for_each_node_state(nid, N_MEMORY) {
2215 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2218 /* Block until all are initialised */
2219 wait_for_completion(&pgdat_init_all_done_comp);
2222 * We initialized the rest of the deferred pages. Permanently disable
2223 * on-demand struct page initialization.
2225 static_branch_disable(&deferred_pages);
2227 /* Reinit limits that are based on free pages after the kernel is up */
2228 files_maxfiles_init();
2233 /* Discard memblock private memory */
2236 for_each_node_state(nid, N_MEMORY)
2237 shuffle_free_memory(NODE_DATA(nid));
2239 for_each_populated_zone(zone)
2240 set_zone_contiguous(zone);
2244 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2245 void __init init_cma_reserved_pageblock(struct page *page)
2247 unsigned i = pageblock_nr_pages;
2248 struct page *p = page;
2251 __ClearPageReserved(p);
2252 set_page_count(p, 0);
2255 set_pageblock_migratetype(page, MIGRATE_CMA);
2257 if (pageblock_order >= MAX_ORDER) {
2258 i = pageblock_nr_pages;
2261 set_page_refcounted(p);
2262 __free_pages(p, MAX_ORDER - 1);
2263 p += MAX_ORDER_NR_PAGES;
2264 } while (i -= MAX_ORDER_NR_PAGES);
2266 set_page_refcounted(page);
2267 __free_pages(page, pageblock_order);
2270 adjust_managed_page_count(page, pageblock_nr_pages);
2271 page_zone(page)->cma_pages += pageblock_nr_pages;
2276 * The order of subdivision here is critical for the IO subsystem.
2277 * Please do not alter this order without good reasons and regression
2278 * testing. Specifically, as large blocks of memory are subdivided,
2279 * the order in which smaller blocks are delivered depends on the order
2280 * they're subdivided in this function. This is the primary factor
2281 * influencing the order in which pages are delivered to the IO
2282 * subsystem according to empirical testing, and this is also justified
2283 * by considering the behavior of a buddy system containing a single
2284 * large block of memory acted on by a series of small allocations.
2285 * This behavior is a critical factor in sglist merging's success.
2289 static inline void expand(struct zone *zone, struct page *page,
2290 int low, int high, int migratetype)
2292 unsigned long size = 1 << high;
2294 while (high > low) {
2297 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2300 * Mark as guard pages (or page), that will allow to
2301 * merge back to allocator when buddy will be freed.
2302 * Corresponding page table entries will not be touched,
2303 * pages will stay not present in virtual address space
2305 if (set_page_guard(zone, &page[size], high, migratetype))
2308 add_to_free_list(&page[size], zone, high, migratetype);
2309 set_buddy_order(&page[size], high);
2313 static void check_new_page_bad(struct page *page)
2315 if (unlikely(page->flags & __PG_HWPOISON)) {
2316 /* Don't complain about hwpoisoned pages */
2317 page_mapcount_reset(page); /* remove PageBuddy */
2322 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2326 * This page is about to be returned from the page allocator
2328 static inline int check_new_page(struct page *page)
2330 if (likely(page_expected_state(page,
2331 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2334 check_new_page_bad(page);
2338 #ifdef CONFIG_DEBUG_VM
2340 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2341 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2342 * also checked when pcp lists are refilled from the free lists.
2344 static inline bool check_pcp_refill(struct page *page)
2346 if (debug_pagealloc_enabled_static())
2347 return check_new_page(page);
2352 static inline bool check_new_pcp(struct page *page)
2354 return check_new_page(page);
2358 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2359 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2360 * enabled, they are also checked when being allocated from the pcp lists.
2362 static inline bool check_pcp_refill(struct page *page)
2364 return check_new_page(page);
2366 static inline bool check_new_pcp(struct page *page)
2368 if (debug_pagealloc_enabled_static())
2369 return check_new_page(page);
2373 #endif /* CONFIG_DEBUG_VM */
2375 static bool check_new_pages(struct page *page, unsigned int order)
2378 for (i = 0; i < (1 << order); i++) {
2379 struct page *p = page + i;
2381 if (unlikely(check_new_page(p)))
2388 inline void post_alloc_hook(struct page *page, unsigned int order,
2391 set_page_private(page, 0);
2392 set_page_refcounted(page);
2394 arch_alloc_page(page, order);
2395 debug_pagealloc_map_pages(page, 1 << order);
2398 * Page unpoisoning must happen before memory initialization.
2399 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2400 * allocations and the page unpoisoning code will complain.
2402 kernel_unpoison_pages(page, 1 << order);
2405 * As memory initialization might be integrated into KASAN,
2406 * kasan_alloc_pages and kernel_init_free_pages must be
2407 * kept together to avoid discrepancies in behavior.
2409 if (kasan_has_integrated_init()) {
2410 kasan_alloc_pages(page, order, gfp_flags);
2412 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2414 kasan_unpoison_pages(page, order, init);
2416 kernel_init_free_pages(page, 1 << order,
2417 gfp_flags & __GFP_ZEROTAGS);
2420 set_page_owner(page, order, gfp_flags);
2423 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2424 unsigned int alloc_flags)
2426 post_alloc_hook(page, order, gfp_flags);
2428 if (order && (gfp_flags & __GFP_COMP))
2429 prep_compound_page(page, order);
2432 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2433 * allocate the page. The expectation is that the caller is taking
2434 * steps that will free more memory. The caller should avoid the page
2435 * being used for !PFMEMALLOC purposes.
2437 if (alloc_flags & ALLOC_NO_WATERMARKS)
2438 set_page_pfmemalloc(page);
2440 clear_page_pfmemalloc(page);
2444 * Go through the free lists for the given migratetype and remove
2445 * the smallest available page from the freelists
2447 static __always_inline
2448 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2451 unsigned int current_order;
2452 struct free_area *area;
2455 /* Find a page of the appropriate size in the preferred list */
2456 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2457 area = &(zone->free_area[current_order]);
2458 page = get_page_from_free_area(area, migratetype);
2461 del_page_from_free_list(page, zone, current_order);
2462 expand(zone, page, order, current_order, migratetype);
2463 set_pcppage_migratetype(page, migratetype);
2472 * This array describes the order lists are fallen back to when
2473 * the free lists for the desirable migrate type are depleted
2475 static int fallbacks[MIGRATE_TYPES][3] = {
2476 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2477 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2478 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2480 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2482 #ifdef CONFIG_MEMORY_ISOLATION
2483 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2488 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2491 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2494 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2495 unsigned int order) { return NULL; }
2499 * Move the free pages in a range to the freelist tail of the requested type.
2500 * Note that start_page and end_pages are not aligned on a pageblock
2501 * boundary. If alignment is required, use move_freepages_block()
2503 static int move_freepages(struct zone *zone,
2504 unsigned long start_pfn, unsigned long end_pfn,
2505 int migratetype, int *num_movable)
2510 int pages_moved = 0;
2512 for (pfn = start_pfn; pfn <= end_pfn;) {
2513 page = pfn_to_page(pfn);
2514 if (!PageBuddy(page)) {
2516 * We assume that pages that could be isolated for
2517 * migration are movable. But we don't actually try
2518 * isolating, as that would be expensive.
2521 (PageLRU(page) || __PageMovable(page)))
2527 /* Make sure we are not inadvertently changing nodes */
2528 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2529 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2531 order = buddy_order(page);
2532 move_to_free_list(page, zone, order, migratetype);
2534 pages_moved += 1 << order;
2540 int move_freepages_block(struct zone *zone, struct page *page,
2541 int migratetype, int *num_movable)
2543 unsigned long start_pfn, end_pfn, pfn;
2548 pfn = page_to_pfn(page);
2549 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2550 end_pfn = start_pfn + pageblock_nr_pages - 1;
2552 /* Do not cross zone boundaries */
2553 if (!zone_spans_pfn(zone, start_pfn))
2555 if (!zone_spans_pfn(zone, end_pfn))
2558 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2562 static void change_pageblock_range(struct page *pageblock_page,
2563 int start_order, int migratetype)
2565 int nr_pageblocks = 1 << (start_order - pageblock_order);
2567 while (nr_pageblocks--) {
2568 set_pageblock_migratetype(pageblock_page, migratetype);
2569 pageblock_page += pageblock_nr_pages;
2574 * When we are falling back to another migratetype during allocation, try to
2575 * steal extra free pages from the same pageblocks to satisfy further
2576 * allocations, instead of polluting multiple pageblocks.
2578 * If we are stealing a relatively large buddy page, it is likely there will
2579 * be more free pages in the pageblock, so try to steal them all. For
2580 * reclaimable and unmovable allocations, we steal regardless of page size,
2581 * as fragmentation caused by those allocations polluting movable pageblocks
2582 * is worse than movable allocations stealing from unmovable and reclaimable
2585 static bool can_steal_fallback(unsigned int order, int start_mt)
2588 * Leaving this order check is intended, although there is
2589 * relaxed order check in next check. The reason is that
2590 * we can actually steal whole pageblock if this condition met,
2591 * but, below check doesn't guarantee it and that is just heuristic
2592 * so could be changed anytime.
2594 if (order >= pageblock_order)
2597 if (order >= pageblock_order / 2 ||
2598 start_mt == MIGRATE_RECLAIMABLE ||
2599 start_mt == MIGRATE_UNMOVABLE ||
2600 page_group_by_mobility_disabled)
2606 static inline bool boost_watermark(struct zone *zone)
2608 unsigned long max_boost;
2610 if (!watermark_boost_factor)
2613 * Don't bother in zones that are unlikely to produce results.
2614 * On small machines, including kdump capture kernels running
2615 * in a small area, boosting the watermark can cause an out of
2616 * memory situation immediately.
2618 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2621 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2622 watermark_boost_factor, 10000);
2625 * high watermark may be uninitialised if fragmentation occurs
2626 * very early in boot so do not boost. We do not fall
2627 * through and boost by pageblock_nr_pages as failing
2628 * allocations that early means that reclaim is not going
2629 * to help and it may even be impossible to reclaim the
2630 * boosted watermark resulting in a hang.
2635 max_boost = max(pageblock_nr_pages, max_boost);
2637 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2644 * This function implements actual steal behaviour. If order is large enough,
2645 * we can steal whole pageblock. If not, we first move freepages in this
2646 * pageblock to our migratetype and determine how many already-allocated pages
2647 * are there in the pageblock with a compatible migratetype. If at least half
2648 * of pages are free or compatible, we can change migratetype of the pageblock
2649 * itself, so pages freed in the future will be put on the correct free list.
2651 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2652 unsigned int alloc_flags, int start_type, bool whole_block)
2654 unsigned int current_order = buddy_order(page);
2655 int free_pages, movable_pages, alike_pages;
2658 old_block_type = get_pageblock_migratetype(page);
2661 * This can happen due to races and we want to prevent broken
2662 * highatomic accounting.
2664 if (is_migrate_highatomic(old_block_type))
2667 /* Take ownership for orders >= pageblock_order */
2668 if (current_order >= pageblock_order) {
2669 change_pageblock_range(page, current_order, start_type);
2674 * Boost watermarks to increase reclaim pressure to reduce the
2675 * likelihood of future fallbacks. Wake kswapd now as the node
2676 * may be balanced overall and kswapd will not wake naturally.
2678 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2679 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2681 /* We are not allowed to try stealing from the whole block */
2685 free_pages = move_freepages_block(zone, page, start_type,
2688 * Determine how many pages are compatible with our allocation.
2689 * For movable allocation, it's the number of movable pages which
2690 * we just obtained. For other types it's a bit more tricky.
2692 if (start_type == MIGRATE_MOVABLE) {
2693 alike_pages = movable_pages;
2696 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2697 * to MOVABLE pageblock, consider all non-movable pages as
2698 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2699 * vice versa, be conservative since we can't distinguish the
2700 * exact migratetype of non-movable pages.
2702 if (old_block_type == MIGRATE_MOVABLE)
2703 alike_pages = pageblock_nr_pages
2704 - (free_pages + movable_pages);
2709 /* moving whole block can fail due to zone boundary conditions */
2714 * If a sufficient number of pages in the block are either free or of
2715 * comparable migratability as our allocation, claim the whole block.
2717 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2718 page_group_by_mobility_disabled)
2719 set_pageblock_migratetype(page, start_type);
2724 move_to_free_list(page, zone, current_order, start_type);
2728 * Check whether there is a suitable fallback freepage with requested order.
2729 * If only_stealable is true, this function returns fallback_mt only if
2730 * we can steal other freepages all together. This would help to reduce
2731 * fragmentation due to mixed migratetype pages in one pageblock.
2733 int find_suitable_fallback(struct free_area *area, unsigned int order,
2734 int migratetype, bool only_stealable, bool *can_steal)
2739 if (area->nr_free == 0)
2744 fallback_mt = fallbacks[migratetype][i];
2745 if (fallback_mt == MIGRATE_TYPES)
2748 if (free_area_empty(area, fallback_mt))
2751 if (can_steal_fallback(order, migratetype))
2754 if (!only_stealable)
2765 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2766 * there are no empty page blocks that contain a page with a suitable order
2768 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2769 unsigned int alloc_order)
2772 unsigned long max_managed, flags;
2775 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2776 * Check is race-prone but harmless.
2778 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2779 if (zone->nr_reserved_highatomic >= max_managed)
2782 spin_lock_irqsave(&zone->lock, flags);
2784 /* Recheck the nr_reserved_highatomic limit under the lock */
2785 if (zone->nr_reserved_highatomic >= max_managed)
2789 mt = get_pageblock_migratetype(page);
2790 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2791 && !is_migrate_cma(mt)) {
2792 zone->nr_reserved_highatomic += pageblock_nr_pages;
2793 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2794 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2798 spin_unlock_irqrestore(&zone->lock, flags);
2802 * Used when an allocation is about to fail under memory pressure. This
2803 * potentially hurts the reliability of high-order allocations when under
2804 * intense memory pressure but failed atomic allocations should be easier
2805 * to recover from than an OOM.
2807 * If @force is true, try to unreserve a pageblock even though highatomic
2808 * pageblock is exhausted.
2810 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2813 struct zonelist *zonelist = ac->zonelist;
2814 unsigned long flags;
2821 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2824 * Preserve at least one pageblock unless memory pressure
2827 if (!force && zone->nr_reserved_highatomic <=
2831 spin_lock_irqsave(&zone->lock, flags);
2832 for (order = 0; order < MAX_ORDER; order++) {
2833 struct free_area *area = &(zone->free_area[order]);
2835 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2840 * In page freeing path, migratetype change is racy so
2841 * we can counter several free pages in a pageblock
2842 * in this loop although we changed the pageblock type
2843 * from highatomic to ac->migratetype. So we should
2844 * adjust the count once.
2846 if (is_migrate_highatomic_page(page)) {
2848 * It should never happen but changes to
2849 * locking could inadvertently allow a per-cpu
2850 * drain to add pages to MIGRATE_HIGHATOMIC
2851 * while unreserving so be safe and watch for
2854 zone->nr_reserved_highatomic -= min(
2856 zone->nr_reserved_highatomic);
2860 * Convert to ac->migratetype and avoid the normal
2861 * pageblock stealing heuristics. Minimally, the caller
2862 * is doing the work and needs the pages. More
2863 * importantly, if the block was always converted to
2864 * MIGRATE_UNMOVABLE or another type then the number
2865 * of pageblocks that cannot be completely freed
2868 set_pageblock_migratetype(page, ac->migratetype);
2869 ret = move_freepages_block(zone, page, ac->migratetype,
2872 spin_unlock_irqrestore(&zone->lock, flags);
2876 spin_unlock_irqrestore(&zone->lock, flags);
2883 * Try finding a free buddy page on the fallback list and put it on the free
2884 * list of requested migratetype, possibly along with other pages from the same
2885 * block, depending on fragmentation avoidance heuristics. Returns true if
2886 * fallback was found so that __rmqueue_smallest() can grab it.
2888 * The use of signed ints for order and current_order is a deliberate
2889 * deviation from the rest of this file, to make the for loop
2890 * condition simpler.
2892 static __always_inline bool
2893 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2894 unsigned int alloc_flags)
2896 struct free_area *area;
2898 int min_order = order;
2904 * Do not steal pages from freelists belonging to other pageblocks
2905 * i.e. orders < pageblock_order. If there are no local zones free,
2906 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2908 if (alloc_flags & ALLOC_NOFRAGMENT)
2909 min_order = pageblock_order;
2912 * Find the largest available free page in the other list. This roughly
2913 * approximates finding the pageblock with the most free pages, which
2914 * would be too costly to do exactly.
2916 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2918 area = &(zone->free_area[current_order]);
2919 fallback_mt = find_suitable_fallback(area, current_order,
2920 start_migratetype, false, &can_steal);
2921 if (fallback_mt == -1)
2925 * We cannot steal all free pages from the pageblock and the
2926 * requested migratetype is movable. In that case it's better to
2927 * steal and split the smallest available page instead of the
2928 * largest available page, because even if the next movable
2929 * allocation falls back into a different pageblock than this
2930 * one, it won't cause permanent fragmentation.
2932 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2933 && current_order > order)
2942 for (current_order = order; current_order < MAX_ORDER;
2944 area = &(zone->free_area[current_order]);
2945 fallback_mt = find_suitable_fallback(area, current_order,
2946 start_migratetype, false, &can_steal);
2947 if (fallback_mt != -1)
2952 * This should not happen - we already found a suitable fallback
2953 * when looking for the largest page.
2955 VM_BUG_ON(current_order == MAX_ORDER);
2958 page = get_page_from_free_area(area, fallback_mt);
2960 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2963 trace_mm_page_alloc_extfrag(page, order, current_order,
2964 start_migratetype, fallback_mt);
2971 * Do the hard work of removing an element from the buddy allocator.
2972 * Call me with the zone->lock already held.
2974 static __always_inline struct page *
2975 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2976 unsigned int alloc_flags)
2980 if (IS_ENABLED(CONFIG_CMA)) {
2982 * Balance movable allocations between regular and CMA areas by
2983 * allocating from CMA when over half of the zone's free memory
2984 * is in the CMA area.
2986 if (alloc_flags & ALLOC_CMA &&
2987 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2988 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2989 page = __rmqueue_cma_fallback(zone, order);
2995 page = __rmqueue_smallest(zone, order, migratetype);
2996 if (unlikely(!page)) {
2997 if (alloc_flags & ALLOC_CMA)
2998 page = __rmqueue_cma_fallback(zone, order);
3000 if (!page && __rmqueue_fallback(zone, order, migratetype,
3006 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3011 * Obtain a specified number of elements from the buddy allocator, all under
3012 * a single hold of the lock, for efficiency. Add them to the supplied list.
3013 * Returns the number of new pages which were placed at *list.
3015 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3016 unsigned long count, struct list_head *list,
3017 int migratetype, unsigned int alloc_flags)
3019 int i, allocated = 0;
3022 * local_lock_irq held so equivalent to spin_lock_irqsave for
3023 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3025 spin_lock(&zone->lock);
3026 for (i = 0; i < count; ++i) {
3027 struct page *page = __rmqueue(zone, order, migratetype,
3029 if (unlikely(page == NULL))
3032 if (unlikely(check_pcp_refill(page)))
3036 * Split buddy pages returned by expand() are received here in
3037 * physical page order. The page is added to the tail of
3038 * caller's list. From the callers perspective, the linked list
3039 * is ordered by page number under some conditions. This is
3040 * useful for IO devices that can forward direction from the
3041 * head, thus also in the physical page order. This is useful
3042 * for IO devices that can merge IO requests if the physical
3043 * pages are ordered properly.
3045 list_add_tail(&page->lru, list);
3047 if (is_migrate_cma(get_pcppage_migratetype(page)))
3048 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3053 * i pages were removed from the buddy list even if some leak due
3054 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3055 * on i. Do not confuse with 'allocated' which is the number of
3056 * pages added to the pcp list.
3058 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3059 spin_unlock(&zone->lock);
3065 * Called from the vmstat counter updater to drain pagesets of this
3066 * currently executing processor on remote nodes after they have
3069 * Note that this function must be called with the thread pinned to
3070 * a single processor.
3072 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3074 unsigned long flags;
3075 int to_drain, batch;
3077 local_lock_irqsave(&pagesets.lock, flags);
3078 batch = READ_ONCE(pcp->batch);
3079 to_drain = min(pcp->count, batch);
3081 free_pcppages_bulk(zone, to_drain, pcp);
3082 local_unlock_irqrestore(&pagesets.lock, flags);
3087 * Drain pcplists of the indicated processor and zone.
3089 * The processor must either be the current processor and the
3090 * thread pinned to the current processor or a processor that
3093 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3095 unsigned long flags;
3096 struct per_cpu_pages *pcp;
3098 local_lock_irqsave(&pagesets.lock, flags);
3100 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3102 free_pcppages_bulk(zone, pcp->count, pcp);
3104 local_unlock_irqrestore(&pagesets.lock, flags);
3108 * Drain pcplists of all zones on the indicated processor.
3110 * The processor must either be the current processor and the
3111 * thread pinned to the current processor or a processor that
3114 static void drain_pages(unsigned int cpu)
3118 for_each_populated_zone(zone) {
3119 drain_pages_zone(cpu, zone);
3124 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3126 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3127 * the single zone's pages.
3129 void drain_local_pages(struct zone *zone)
3131 int cpu = smp_processor_id();
3134 drain_pages_zone(cpu, zone);
3139 static void drain_local_pages_wq(struct work_struct *work)
3141 struct pcpu_drain *drain;
3143 drain = container_of(work, struct pcpu_drain, work);
3146 * drain_all_pages doesn't use proper cpu hotplug protection so
3147 * we can race with cpu offline when the WQ can move this from
3148 * a cpu pinned worker to an unbound one. We can operate on a different
3149 * cpu which is alright but we also have to make sure to not move to
3153 drain_local_pages(drain->zone);
3158 * The implementation of drain_all_pages(), exposing an extra parameter to
3159 * drain on all cpus.
3161 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3162 * not empty. The check for non-emptiness can however race with a free to
3163 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3164 * that need the guarantee that every CPU has drained can disable the
3165 * optimizing racy check.
3167 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3172 * Allocate in the BSS so we won't require allocation in
3173 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3175 static cpumask_t cpus_with_pcps;
3178 * Make sure nobody triggers this path before mm_percpu_wq is fully
3181 if (WARN_ON_ONCE(!mm_percpu_wq))
3185 * Do not drain if one is already in progress unless it's specific to
3186 * a zone. Such callers are primarily CMA and memory hotplug and need
3187 * the drain to be complete when the call returns.
3189 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3192 mutex_lock(&pcpu_drain_mutex);
3196 * We don't care about racing with CPU hotplug event
3197 * as offline notification will cause the notified
3198 * cpu to drain that CPU pcps and on_each_cpu_mask
3199 * disables preemption as part of its processing
3201 for_each_online_cpu(cpu) {
3202 struct per_cpu_pages *pcp;
3204 bool has_pcps = false;
3206 if (force_all_cpus) {
3208 * The pcp.count check is racy, some callers need a
3209 * guarantee that no cpu is missed.
3213 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3217 for_each_populated_zone(z) {
3218 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3227 cpumask_set_cpu(cpu, &cpus_with_pcps);
3229 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3232 for_each_cpu(cpu, &cpus_with_pcps) {
3233 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3236 INIT_WORK(&drain->work, drain_local_pages_wq);
3237 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3239 for_each_cpu(cpu, &cpus_with_pcps)
3240 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3242 mutex_unlock(&pcpu_drain_mutex);
3246 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3248 * When zone parameter is non-NULL, spill just the single zone's pages.
3250 * Note that this can be extremely slow as the draining happens in a workqueue.
3252 void drain_all_pages(struct zone *zone)
3254 __drain_all_pages(zone, false);
3257 #ifdef CONFIG_HIBERNATION
3260 * Touch the watchdog for every WD_PAGE_COUNT pages.
3262 #define WD_PAGE_COUNT (128*1024)
3264 void mark_free_pages(struct zone *zone)
3266 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3267 unsigned long flags;
3268 unsigned int order, t;
3271 if (zone_is_empty(zone))
3274 spin_lock_irqsave(&zone->lock, flags);
3276 max_zone_pfn = zone_end_pfn(zone);
3277 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3278 if (pfn_valid(pfn)) {
3279 page = pfn_to_page(pfn);
3281 if (!--page_count) {
3282 touch_nmi_watchdog();
3283 page_count = WD_PAGE_COUNT;
3286 if (page_zone(page) != zone)
3289 if (!swsusp_page_is_forbidden(page))
3290 swsusp_unset_page_free(page);
3293 for_each_migratetype_order(order, t) {
3294 list_for_each_entry(page,
3295 &zone->free_area[order].free_list[t], lru) {
3298 pfn = page_to_pfn(page);
3299 for (i = 0; i < (1UL << order); i++) {
3300 if (!--page_count) {
3301 touch_nmi_watchdog();
3302 page_count = WD_PAGE_COUNT;
3304 swsusp_set_page_free(pfn_to_page(pfn + i));
3308 spin_unlock_irqrestore(&zone->lock, flags);
3310 #endif /* CONFIG_PM */
3312 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3317 if (!free_pcp_prepare(page, order))
3320 migratetype = get_pfnblock_migratetype(page, pfn);
3321 set_pcppage_migratetype(page, migratetype);
3325 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch)
3327 int min_nr_free, max_nr_free;
3329 /* Check for PCP disabled or boot pageset */
3330 if (unlikely(high < batch))
3333 /* Leave at least pcp->batch pages on the list */
3334 min_nr_free = batch;
3335 max_nr_free = high - batch;
3338 * Double the number of pages freed each time there is subsequent
3339 * freeing of pages without any allocation.
3341 batch <<= pcp->free_factor;
3342 if (batch < max_nr_free)
3344 batch = clamp(batch, min_nr_free, max_nr_free);
3349 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone)
3351 int high = READ_ONCE(pcp->high);
3353 if (unlikely(!high))
3356 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3360 * If reclaim is active, limit the number of pages that can be
3361 * stored on pcp lists
3363 return min(READ_ONCE(pcp->batch) << 2, high);
3366 static void free_unref_page_commit(struct page *page, unsigned long pfn,
3367 int migratetype, unsigned int order)
3369 struct zone *zone = page_zone(page);
3370 struct per_cpu_pages *pcp;
3374 __count_vm_event(PGFREE);
3375 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3376 pindex = order_to_pindex(migratetype, order);
3377 list_add(&page->lru, &pcp->lists[pindex]);
3378 pcp->count += 1 << order;
3379 high = nr_pcp_high(pcp, zone);
3380 if (pcp->count >= high) {
3381 int batch = READ_ONCE(pcp->batch);
3383 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch), pcp);
3390 void free_unref_page(struct page *page, unsigned int order)
3392 unsigned long flags;
3393 unsigned long pfn = page_to_pfn(page);
3396 if (!free_unref_page_prepare(page, pfn, order))
3400 * We only track unmovable, reclaimable and movable on pcp lists.
3401 * Place ISOLATE pages on the isolated list because they are being
3402 * offlined but treat HIGHATOMIC as movable pages so we can get those
3403 * areas back if necessary. Otherwise, we may have to free
3404 * excessively into the page allocator
3406 migratetype = get_pcppage_migratetype(page);
3407 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3408 if (unlikely(is_migrate_isolate(migratetype))) {
3409 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3412 migratetype = MIGRATE_MOVABLE;
3415 local_lock_irqsave(&pagesets.lock, flags);
3416 free_unref_page_commit(page, pfn, migratetype, order);
3417 local_unlock_irqrestore(&pagesets.lock, flags);
3421 * Free a list of 0-order pages
3423 void free_unref_page_list(struct list_head *list)
3425 struct page *page, *next;
3426 unsigned long flags, pfn;
3427 int batch_count = 0;
3430 /* Prepare pages for freeing */
3431 list_for_each_entry_safe(page, next, list, lru) {
3432 pfn = page_to_pfn(page);
3433 if (!free_unref_page_prepare(page, pfn, 0)) {
3434 list_del(&page->lru);
3439 * Free isolated pages directly to the allocator, see
3440 * comment in free_unref_page.
3442 migratetype = get_pcppage_migratetype(page);
3443 if (unlikely(is_migrate_isolate(migratetype))) {
3444 list_del(&page->lru);
3445 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3449 set_page_private(page, pfn);
3452 local_lock_irqsave(&pagesets.lock, flags);
3453 list_for_each_entry_safe(page, next, list, lru) {
3454 pfn = page_private(page);
3455 set_page_private(page, 0);
3458 * Non-isolated types over MIGRATE_PCPTYPES get added
3459 * to the MIGRATE_MOVABLE pcp list.
3461 migratetype = get_pcppage_migratetype(page);
3462 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3463 migratetype = MIGRATE_MOVABLE;
3465 trace_mm_page_free_batched(page);
3466 free_unref_page_commit(page, pfn, migratetype, 0);
3469 * Guard against excessive IRQ disabled times when we get
3470 * a large list of pages to free.
3472 if (++batch_count == SWAP_CLUSTER_MAX) {
3473 local_unlock_irqrestore(&pagesets.lock, flags);
3475 local_lock_irqsave(&pagesets.lock, flags);
3478 local_unlock_irqrestore(&pagesets.lock, flags);
3482 * split_page takes a non-compound higher-order page, and splits it into
3483 * n (1<<order) sub-pages: page[0..n]
3484 * Each sub-page must be freed individually.
3486 * Note: this is probably too low level an operation for use in drivers.
3487 * Please consult with lkml before using this in your driver.
3489 void split_page(struct page *page, unsigned int order)
3493 VM_BUG_ON_PAGE(PageCompound(page), page);
3494 VM_BUG_ON_PAGE(!page_count(page), page);
3496 for (i = 1; i < (1 << order); i++)
3497 set_page_refcounted(page + i);
3498 split_page_owner(page, 1 << order);
3499 split_page_memcg(page, 1 << order);
3501 EXPORT_SYMBOL_GPL(split_page);
3503 int __isolate_free_page(struct page *page, unsigned int order)
3505 unsigned long watermark;
3509 BUG_ON(!PageBuddy(page));
3511 zone = page_zone(page);
3512 mt = get_pageblock_migratetype(page);
3514 if (!is_migrate_isolate(mt)) {
3516 * Obey watermarks as if the page was being allocated. We can
3517 * emulate a high-order watermark check with a raised order-0
3518 * watermark, because we already know our high-order page
3521 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3522 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3525 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3528 /* Remove page from free list */
3530 del_page_from_free_list(page, zone, order);
3533 * Set the pageblock if the isolated page is at least half of a
3536 if (order >= pageblock_order - 1) {
3537 struct page *endpage = page + (1 << order) - 1;
3538 for (; page < endpage; page += pageblock_nr_pages) {
3539 int mt = get_pageblock_migratetype(page);
3540 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3541 && !is_migrate_highatomic(mt))
3542 set_pageblock_migratetype(page,
3548 return 1UL << order;
3552 * __putback_isolated_page - Return a now-isolated page back where we got it
3553 * @page: Page that was isolated
3554 * @order: Order of the isolated page
3555 * @mt: The page's pageblock's migratetype
3557 * This function is meant to return a page pulled from the free lists via
3558 * __isolate_free_page back to the free lists they were pulled from.
3560 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3562 struct zone *zone = page_zone(page);
3564 /* zone lock should be held when this function is called */
3565 lockdep_assert_held(&zone->lock);
3567 /* Return isolated page to tail of freelist. */
3568 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3569 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3573 * Update NUMA hit/miss statistics
3575 * Must be called with interrupts disabled.
3577 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3581 enum numa_stat_item local_stat = NUMA_LOCAL;
3583 /* skip numa counters update if numa stats is disabled */
3584 if (!static_branch_likely(&vm_numa_stat_key))
3587 if (zone_to_nid(z) != numa_node_id())
3588 local_stat = NUMA_OTHER;
3590 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3591 __count_numa_events(z, NUMA_HIT, nr_account);
3593 __count_numa_events(z, NUMA_MISS, nr_account);
3594 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3596 __count_numa_events(z, local_stat, nr_account);
3600 /* Remove page from the per-cpu list, caller must protect the list */
3602 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3604 unsigned int alloc_flags,
3605 struct per_cpu_pages *pcp,
3606 struct list_head *list)
3611 if (list_empty(list)) {
3612 int batch = READ_ONCE(pcp->batch);
3616 * Scale batch relative to order if batch implies
3617 * free pages can be stored on the PCP. Batch can
3618 * be 1 for small zones or for boot pagesets which
3619 * should never store free pages as the pages may
3620 * belong to arbitrary zones.
3623 batch = max(batch >> order, 2);
3624 alloced = rmqueue_bulk(zone, order,
3626 migratetype, alloc_flags);
3628 pcp->count += alloced << order;
3629 if (unlikely(list_empty(list)))
3633 page = list_first_entry(list, struct page, lru);
3634 list_del(&page->lru);
3635 pcp->count -= 1 << order;
3636 } while (check_new_pcp(page));
3641 /* Lock and remove page from the per-cpu list */
3642 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3643 struct zone *zone, unsigned int order,
3644 gfp_t gfp_flags, int migratetype,
3645 unsigned int alloc_flags)
3647 struct per_cpu_pages *pcp;
3648 struct list_head *list;
3650 unsigned long flags;
3652 local_lock_irqsave(&pagesets.lock, flags);
3655 * On allocation, reduce the number of pages that are batch freed.
3656 * See nr_pcp_free() where free_factor is increased for subsequent
3659 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3660 pcp->free_factor >>= 1;
3661 list = &pcp->lists[order_to_pindex(migratetype, order)];
3662 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3663 local_unlock_irqrestore(&pagesets.lock, flags);
3665 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3666 zone_statistics(preferred_zone, zone, 1);
3672 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3675 struct page *rmqueue(struct zone *preferred_zone,
3676 struct zone *zone, unsigned int order,
3677 gfp_t gfp_flags, unsigned int alloc_flags,
3680 unsigned long flags;
3683 if (likely(pcp_allowed_order(order))) {
3685 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3686 * we need to skip it when CMA area isn't allowed.
3688 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3689 migratetype != MIGRATE_MOVABLE) {
3690 page = rmqueue_pcplist(preferred_zone, zone, order,
3691 gfp_flags, migratetype, alloc_flags);
3697 * We most definitely don't want callers attempting to
3698 * allocate greater than order-1 page units with __GFP_NOFAIL.
3700 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3701 spin_lock_irqsave(&zone->lock, flags);
3706 * order-0 request can reach here when the pcplist is skipped
3707 * due to non-CMA allocation context. HIGHATOMIC area is
3708 * reserved for high-order atomic allocation, so order-0
3709 * request should skip it.
3711 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3712 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3714 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3717 page = __rmqueue(zone, order, migratetype, alloc_flags);
3718 } while (page && check_new_pages(page, order));
3722 __mod_zone_freepage_state(zone, -(1 << order),
3723 get_pcppage_migratetype(page));
3724 spin_unlock_irqrestore(&zone->lock, flags);
3726 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3727 zone_statistics(preferred_zone, zone, 1);
3730 /* Separate test+clear to avoid unnecessary atomics */
3731 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3732 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3733 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3736 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3740 spin_unlock_irqrestore(&zone->lock, flags);
3744 #ifdef CONFIG_FAIL_PAGE_ALLOC
3747 struct fault_attr attr;
3749 bool ignore_gfp_highmem;
3750 bool ignore_gfp_reclaim;
3752 } fail_page_alloc = {
3753 .attr = FAULT_ATTR_INITIALIZER,
3754 .ignore_gfp_reclaim = true,
3755 .ignore_gfp_highmem = true,
3759 static int __init setup_fail_page_alloc(char *str)
3761 return setup_fault_attr(&fail_page_alloc.attr, str);
3763 __setup("fail_page_alloc=", setup_fail_page_alloc);
3765 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3767 if (order < fail_page_alloc.min_order)
3769 if (gfp_mask & __GFP_NOFAIL)
3771 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3773 if (fail_page_alloc.ignore_gfp_reclaim &&
3774 (gfp_mask & __GFP_DIRECT_RECLAIM))
3777 return should_fail(&fail_page_alloc.attr, 1 << order);
3780 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3782 static int __init fail_page_alloc_debugfs(void)
3784 umode_t mode = S_IFREG | 0600;
3787 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3788 &fail_page_alloc.attr);
3790 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3791 &fail_page_alloc.ignore_gfp_reclaim);
3792 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3793 &fail_page_alloc.ignore_gfp_highmem);
3794 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3799 late_initcall(fail_page_alloc_debugfs);
3801 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3803 #else /* CONFIG_FAIL_PAGE_ALLOC */
3805 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3810 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3812 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3814 return __should_fail_alloc_page(gfp_mask, order);
3816 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3818 static inline long __zone_watermark_unusable_free(struct zone *z,
3819 unsigned int order, unsigned int alloc_flags)
3821 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3822 long unusable_free = (1 << order) - 1;
3825 * If the caller does not have rights to ALLOC_HARDER then subtract
3826 * the high-atomic reserves. This will over-estimate the size of the
3827 * atomic reserve but it avoids a search.
3829 if (likely(!alloc_harder))
3830 unusable_free += z->nr_reserved_highatomic;
3833 /* If allocation can't use CMA areas don't use free CMA pages */
3834 if (!(alloc_flags & ALLOC_CMA))
3835 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3838 return unusable_free;
3842 * Return true if free base pages are above 'mark'. For high-order checks it
3843 * will return true of the order-0 watermark is reached and there is at least
3844 * one free page of a suitable size. Checking now avoids taking the zone lock
3845 * to check in the allocation paths if no pages are free.
3847 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3848 int highest_zoneidx, unsigned int alloc_flags,
3853 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3855 /* free_pages may go negative - that's OK */
3856 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3858 if (alloc_flags & ALLOC_HIGH)
3861 if (unlikely(alloc_harder)) {
3863 * OOM victims can try even harder than normal ALLOC_HARDER
3864 * users on the grounds that it's definitely going to be in
3865 * the exit path shortly and free memory. Any allocation it
3866 * makes during the free path will be small and short-lived.
3868 if (alloc_flags & ALLOC_OOM)
3875 * Check watermarks for an order-0 allocation request. If these
3876 * are not met, then a high-order request also cannot go ahead
3877 * even if a suitable page happened to be free.
3879 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3882 /* If this is an order-0 request then the watermark is fine */
3886 /* For a high-order request, check at least one suitable page is free */
3887 for (o = order; o < MAX_ORDER; o++) {
3888 struct free_area *area = &z->free_area[o];
3894 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3895 if (!free_area_empty(area, mt))
3900 if ((alloc_flags & ALLOC_CMA) &&
3901 !free_area_empty(area, MIGRATE_CMA)) {
3905 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3911 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3912 int highest_zoneidx, unsigned int alloc_flags)
3914 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3915 zone_page_state(z, NR_FREE_PAGES));
3918 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3919 unsigned long mark, int highest_zoneidx,
3920 unsigned int alloc_flags, gfp_t gfp_mask)
3924 free_pages = zone_page_state(z, NR_FREE_PAGES);
3927 * Fast check for order-0 only. If this fails then the reserves
3928 * need to be calculated.
3934 usable_free = free_pages;
3935 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3937 /* reserved may over estimate high-atomic reserves. */
3938 usable_free -= min(usable_free, reserved);
3939 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3943 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3947 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3948 * when checking the min watermark. The min watermark is the
3949 * point where boosting is ignored so that kswapd is woken up
3950 * when below the low watermark.
3952 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3953 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3954 mark = z->_watermark[WMARK_MIN];
3955 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3956 alloc_flags, free_pages);
3962 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3963 unsigned long mark, int highest_zoneidx)
3965 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3967 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3968 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3970 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3975 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3977 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3978 node_reclaim_distance;
3980 #else /* CONFIG_NUMA */
3981 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3985 #endif /* CONFIG_NUMA */
3988 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3989 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3990 * premature use of a lower zone may cause lowmem pressure problems that
3991 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3992 * probably too small. It only makes sense to spread allocations to avoid
3993 * fragmentation between the Normal and DMA32 zones.
3995 static inline unsigned int
3996 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3998 unsigned int alloc_flags;
4001 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4004 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4006 #ifdef CONFIG_ZONE_DMA32
4010 if (zone_idx(zone) != ZONE_NORMAL)
4014 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4015 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4016 * on UMA that if Normal is populated then so is DMA32.
4018 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4019 if (nr_online_nodes > 1 && !populated_zone(--zone))
4022 alloc_flags |= ALLOC_NOFRAGMENT;
4023 #endif /* CONFIG_ZONE_DMA32 */
4027 /* Must be called after current_gfp_context() which can change gfp_mask */
4028 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4029 unsigned int alloc_flags)
4032 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4033 alloc_flags |= ALLOC_CMA;
4039 * get_page_from_freelist goes through the zonelist trying to allocate
4042 static struct page *
4043 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4044 const struct alloc_context *ac)
4048 struct pglist_data *last_pgdat_dirty_limit = NULL;
4053 * Scan zonelist, looking for a zone with enough free.
4054 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4056 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4057 z = ac->preferred_zoneref;
4058 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4063 if (cpusets_enabled() &&
4064 (alloc_flags & ALLOC_CPUSET) &&
4065 !__cpuset_zone_allowed(zone, gfp_mask))
4068 * When allocating a page cache page for writing, we
4069 * want to get it from a node that is within its dirty
4070 * limit, such that no single node holds more than its
4071 * proportional share of globally allowed dirty pages.
4072 * The dirty limits take into account the node's
4073 * lowmem reserves and high watermark so that kswapd
4074 * should be able to balance it without having to
4075 * write pages from its LRU list.
4077 * XXX: For now, allow allocations to potentially
4078 * exceed the per-node dirty limit in the slowpath
4079 * (spread_dirty_pages unset) before going into reclaim,
4080 * which is important when on a NUMA setup the allowed
4081 * nodes are together not big enough to reach the
4082 * global limit. The proper fix for these situations
4083 * will require awareness of nodes in the
4084 * dirty-throttling and the flusher threads.
4086 if (ac->spread_dirty_pages) {
4087 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4090 if (!node_dirty_ok(zone->zone_pgdat)) {
4091 last_pgdat_dirty_limit = zone->zone_pgdat;
4096 if (no_fallback && nr_online_nodes > 1 &&
4097 zone != ac->preferred_zoneref->zone) {
4101 * If moving to a remote node, retry but allow
4102 * fragmenting fallbacks. Locality is more important
4103 * than fragmentation avoidance.
4105 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4106 if (zone_to_nid(zone) != local_nid) {
4107 alloc_flags &= ~ALLOC_NOFRAGMENT;
4112 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4113 if (!zone_watermark_fast(zone, order, mark,
4114 ac->highest_zoneidx, alloc_flags,
4118 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4120 * Watermark failed for this zone, but see if we can
4121 * grow this zone if it contains deferred pages.
4123 if (static_branch_unlikely(&deferred_pages)) {
4124 if (_deferred_grow_zone(zone, order))
4128 /* Checked here to keep the fast path fast */
4129 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4130 if (alloc_flags & ALLOC_NO_WATERMARKS)
4133 if (!node_reclaim_enabled() ||
4134 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4137 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4139 case NODE_RECLAIM_NOSCAN:
4142 case NODE_RECLAIM_FULL:
4143 /* scanned but unreclaimable */
4146 /* did we reclaim enough */
4147 if (zone_watermark_ok(zone, order, mark,
4148 ac->highest_zoneidx, alloc_flags))
4156 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4157 gfp_mask, alloc_flags, ac->migratetype);
4159 prep_new_page(page, order, gfp_mask, alloc_flags);
4162 * If this is a high-order atomic allocation then check
4163 * if the pageblock should be reserved for the future
4165 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4166 reserve_highatomic_pageblock(page, zone, order);
4170 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4171 /* Try again if zone has deferred pages */
4172 if (static_branch_unlikely(&deferred_pages)) {
4173 if (_deferred_grow_zone(zone, order))
4181 * It's possible on a UMA machine to get through all zones that are
4182 * fragmented. If avoiding fragmentation, reset and try again.
4185 alloc_flags &= ~ALLOC_NOFRAGMENT;
4192 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4194 unsigned int filter = SHOW_MEM_FILTER_NODES;
4197 * This documents exceptions given to allocations in certain
4198 * contexts that are allowed to allocate outside current's set
4201 if (!(gfp_mask & __GFP_NOMEMALLOC))
4202 if (tsk_is_oom_victim(current) ||
4203 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4204 filter &= ~SHOW_MEM_FILTER_NODES;
4205 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4206 filter &= ~SHOW_MEM_FILTER_NODES;
4208 show_mem(filter, nodemask);
4211 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4213 struct va_format vaf;
4215 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4217 if ((gfp_mask & __GFP_NOWARN) ||
4218 !__ratelimit(&nopage_rs) ||
4219 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4222 va_start(args, fmt);
4225 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4226 current->comm, &vaf, gfp_mask, &gfp_mask,
4227 nodemask_pr_args(nodemask));
4230 cpuset_print_current_mems_allowed();
4233 warn_alloc_show_mem(gfp_mask, nodemask);
4236 static inline struct page *
4237 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4238 unsigned int alloc_flags,
4239 const struct alloc_context *ac)
4243 page = get_page_from_freelist(gfp_mask, order,
4244 alloc_flags|ALLOC_CPUSET, ac);
4246 * fallback to ignore cpuset restriction if our nodes
4250 page = get_page_from_freelist(gfp_mask, order,
4256 static inline struct page *
4257 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4258 const struct alloc_context *ac, unsigned long *did_some_progress)
4260 struct oom_control oc = {
4261 .zonelist = ac->zonelist,
4262 .nodemask = ac->nodemask,
4264 .gfp_mask = gfp_mask,
4269 *did_some_progress = 0;
4272 * Acquire the oom lock. If that fails, somebody else is
4273 * making progress for us.
4275 if (!mutex_trylock(&oom_lock)) {
4276 *did_some_progress = 1;
4277 schedule_timeout_uninterruptible(1);
4282 * Go through the zonelist yet one more time, keep very high watermark
4283 * here, this is only to catch a parallel oom killing, we must fail if
4284 * we're still under heavy pressure. But make sure that this reclaim
4285 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4286 * allocation which will never fail due to oom_lock already held.
4288 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4289 ~__GFP_DIRECT_RECLAIM, order,
4290 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4294 /* Coredumps can quickly deplete all memory reserves */
4295 if (current->flags & PF_DUMPCORE)
4297 /* The OOM killer will not help higher order allocs */
4298 if (order > PAGE_ALLOC_COSTLY_ORDER)
4301 * We have already exhausted all our reclaim opportunities without any
4302 * success so it is time to admit defeat. We will skip the OOM killer
4303 * because it is very likely that the caller has a more reasonable
4304 * fallback than shooting a random task.
4306 * The OOM killer may not free memory on a specific node.
4308 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4310 /* The OOM killer does not needlessly kill tasks for lowmem */
4311 if (ac->highest_zoneidx < ZONE_NORMAL)
4313 if (pm_suspended_storage())
4316 * XXX: GFP_NOFS allocations should rather fail than rely on
4317 * other request to make a forward progress.
4318 * We are in an unfortunate situation where out_of_memory cannot
4319 * do much for this context but let's try it to at least get
4320 * access to memory reserved if the current task is killed (see
4321 * out_of_memory). Once filesystems are ready to handle allocation
4322 * failures more gracefully we should just bail out here.
4325 /* Exhausted what can be done so it's blame time */
4326 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4327 *did_some_progress = 1;
4330 * Help non-failing allocations by giving them access to memory
4333 if (gfp_mask & __GFP_NOFAIL)
4334 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4335 ALLOC_NO_WATERMARKS, ac);
4338 mutex_unlock(&oom_lock);
4343 * Maximum number of compaction retries with a progress before OOM
4344 * killer is consider as the only way to move forward.
4346 #define MAX_COMPACT_RETRIES 16
4348 #ifdef CONFIG_COMPACTION
4349 /* Try memory compaction for high-order allocations before reclaim */
4350 static struct page *
4351 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4352 unsigned int alloc_flags, const struct alloc_context *ac,
4353 enum compact_priority prio, enum compact_result *compact_result)
4355 struct page *page = NULL;
4356 unsigned long pflags;
4357 unsigned int noreclaim_flag;
4362 psi_memstall_enter(&pflags);
4363 noreclaim_flag = memalloc_noreclaim_save();
4365 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4368 memalloc_noreclaim_restore(noreclaim_flag);
4369 psi_memstall_leave(&pflags);
4371 if (*compact_result == COMPACT_SKIPPED)
4374 * At least in one zone compaction wasn't deferred or skipped, so let's
4375 * count a compaction stall
4377 count_vm_event(COMPACTSTALL);
4379 /* Prep a captured page if available */
4381 prep_new_page(page, order, gfp_mask, alloc_flags);
4383 /* Try get a page from the freelist if available */
4385 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4388 struct zone *zone = page_zone(page);
4390 zone->compact_blockskip_flush = false;
4391 compaction_defer_reset(zone, order, true);
4392 count_vm_event(COMPACTSUCCESS);
4397 * It's bad if compaction run occurs and fails. The most likely reason
4398 * is that pages exist, but not enough to satisfy watermarks.
4400 count_vm_event(COMPACTFAIL);
4408 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4409 enum compact_result compact_result,
4410 enum compact_priority *compact_priority,
4411 int *compaction_retries)
4413 int max_retries = MAX_COMPACT_RETRIES;
4416 int retries = *compaction_retries;
4417 enum compact_priority priority = *compact_priority;
4422 if (fatal_signal_pending(current))
4425 if (compaction_made_progress(compact_result))
4426 (*compaction_retries)++;
4429 * compaction considers all the zone as desperately out of memory
4430 * so it doesn't really make much sense to retry except when the
4431 * failure could be caused by insufficient priority
4433 if (compaction_failed(compact_result))
4434 goto check_priority;
4437 * compaction was skipped because there are not enough order-0 pages
4438 * to work with, so we retry only if it looks like reclaim can help.
4440 if (compaction_needs_reclaim(compact_result)) {
4441 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4446 * make sure the compaction wasn't deferred or didn't bail out early
4447 * due to locks contention before we declare that we should give up.
4448 * But the next retry should use a higher priority if allowed, so
4449 * we don't just keep bailing out endlessly.
4451 if (compaction_withdrawn(compact_result)) {
4452 goto check_priority;
4456 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4457 * costly ones because they are de facto nofail and invoke OOM
4458 * killer to move on while costly can fail and users are ready
4459 * to cope with that. 1/4 retries is rather arbitrary but we
4460 * would need much more detailed feedback from compaction to
4461 * make a better decision.
4463 if (order > PAGE_ALLOC_COSTLY_ORDER)
4465 if (*compaction_retries <= max_retries) {
4471 * Make sure there are attempts at the highest priority if we exhausted
4472 * all retries or failed at the lower priorities.
4475 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4476 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4478 if (*compact_priority > min_priority) {
4479 (*compact_priority)--;
4480 *compaction_retries = 0;
4484 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4488 static inline struct page *
4489 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4490 unsigned int alloc_flags, const struct alloc_context *ac,
4491 enum compact_priority prio, enum compact_result *compact_result)
4493 *compact_result = COMPACT_SKIPPED;
4498 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4499 enum compact_result compact_result,
4500 enum compact_priority *compact_priority,
4501 int *compaction_retries)
4506 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4510 * There are setups with compaction disabled which would prefer to loop
4511 * inside the allocator rather than hit the oom killer prematurely.
4512 * Let's give them a good hope and keep retrying while the order-0
4513 * watermarks are OK.
4515 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4516 ac->highest_zoneidx, ac->nodemask) {
4517 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4518 ac->highest_zoneidx, alloc_flags))
4523 #endif /* CONFIG_COMPACTION */
4525 #ifdef CONFIG_LOCKDEP
4526 static struct lockdep_map __fs_reclaim_map =
4527 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4529 static bool __need_reclaim(gfp_t gfp_mask)
4531 /* no reclaim without waiting on it */
4532 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4535 /* this guy won't enter reclaim */
4536 if (current->flags & PF_MEMALLOC)
4539 if (gfp_mask & __GFP_NOLOCKDEP)
4545 void __fs_reclaim_acquire(unsigned long ip)
4547 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4550 void __fs_reclaim_release(unsigned long ip)
4552 lock_release(&__fs_reclaim_map, ip);
4555 void fs_reclaim_acquire(gfp_t gfp_mask)
4557 gfp_mask = current_gfp_context(gfp_mask);
4559 if (__need_reclaim(gfp_mask)) {
4560 if (gfp_mask & __GFP_FS)
4561 __fs_reclaim_acquire(_RET_IP_);
4563 #ifdef CONFIG_MMU_NOTIFIER
4564 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4565 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4570 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4572 void fs_reclaim_release(gfp_t gfp_mask)
4574 gfp_mask = current_gfp_context(gfp_mask);
4576 if (__need_reclaim(gfp_mask)) {
4577 if (gfp_mask & __GFP_FS)
4578 __fs_reclaim_release(_RET_IP_);
4581 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4585 * Zonelists may change due to hotplug during allocation. Detect when zonelists
4586 * have been rebuilt so allocation retries. Reader side does not lock and
4587 * retries the allocation if zonelist changes. Writer side is protected by the
4588 * embedded spin_lock.
4590 static DEFINE_SEQLOCK(zonelist_update_seq);
4592 static unsigned int zonelist_iter_begin(void)
4594 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4595 return read_seqbegin(&zonelist_update_seq);
4600 static unsigned int check_retry_zonelist(unsigned int seq)
4602 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4603 return read_seqretry(&zonelist_update_seq, seq);
4608 /* Perform direct synchronous page reclaim */
4609 static unsigned long
4610 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4611 const struct alloc_context *ac)
4613 unsigned int noreclaim_flag;
4614 unsigned long pflags, progress;
4618 /* We now go into synchronous reclaim */
4619 cpuset_memory_pressure_bump();
4620 psi_memstall_enter(&pflags);
4621 fs_reclaim_acquire(gfp_mask);
4622 noreclaim_flag = memalloc_noreclaim_save();
4624 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4627 memalloc_noreclaim_restore(noreclaim_flag);
4628 fs_reclaim_release(gfp_mask);
4629 psi_memstall_leave(&pflags);
4636 /* The really slow allocator path where we enter direct reclaim */
4637 static inline struct page *
4638 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4639 unsigned int alloc_flags, const struct alloc_context *ac,
4640 unsigned long *did_some_progress)
4642 struct page *page = NULL;
4643 bool drained = false;
4645 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4646 if (unlikely(!(*did_some_progress)))
4650 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4653 * If an allocation failed after direct reclaim, it could be because
4654 * pages are pinned on the per-cpu lists or in high alloc reserves.
4655 * Shrink them and try again
4657 if (!page && !drained) {
4658 unreserve_highatomic_pageblock(ac, false);
4659 drain_all_pages(NULL);
4667 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4668 const struct alloc_context *ac)
4672 pg_data_t *last_pgdat = NULL;
4673 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4675 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4677 if (last_pgdat != zone->zone_pgdat)
4678 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4679 last_pgdat = zone->zone_pgdat;
4683 static inline unsigned int
4684 gfp_to_alloc_flags(gfp_t gfp_mask)
4686 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4689 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4690 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4691 * to save two branches.
4693 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4694 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4697 * The caller may dip into page reserves a bit more if the caller
4698 * cannot run direct reclaim, or if the caller has realtime scheduling
4699 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4700 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4702 alloc_flags |= (__force int)
4703 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4705 if (gfp_mask & __GFP_ATOMIC) {
4707 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4708 * if it can't schedule.
4710 if (!(gfp_mask & __GFP_NOMEMALLOC))
4711 alloc_flags |= ALLOC_HARDER;
4713 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4714 * comment for __cpuset_node_allowed().
4716 alloc_flags &= ~ALLOC_CPUSET;
4717 } else if (unlikely(rt_task(current)) && in_task())
4718 alloc_flags |= ALLOC_HARDER;
4720 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4725 static bool oom_reserves_allowed(struct task_struct *tsk)
4727 if (!tsk_is_oom_victim(tsk))
4731 * !MMU doesn't have oom reaper so give access to memory reserves
4732 * only to the thread with TIF_MEMDIE set
4734 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4741 * Distinguish requests which really need access to full memory
4742 * reserves from oom victims which can live with a portion of it
4744 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4746 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4748 if (gfp_mask & __GFP_MEMALLOC)
4749 return ALLOC_NO_WATERMARKS;
4750 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4751 return ALLOC_NO_WATERMARKS;
4752 if (!in_interrupt()) {
4753 if (current->flags & PF_MEMALLOC)
4754 return ALLOC_NO_WATERMARKS;
4755 else if (oom_reserves_allowed(current))
4762 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4764 return !!__gfp_pfmemalloc_flags(gfp_mask);
4768 * Checks whether it makes sense to retry the reclaim to make a forward progress
4769 * for the given allocation request.
4771 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4772 * without success, or when we couldn't even meet the watermark if we
4773 * reclaimed all remaining pages on the LRU lists.
4775 * Returns true if a retry is viable or false to enter the oom path.
4778 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4779 struct alloc_context *ac, int alloc_flags,
4780 bool did_some_progress, int *no_progress_loops)
4787 * Costly allocations might have made a progress but this doesn't mean
4788 * their order will become available due to high fragmentation so
4789 * always increment the no progress counter for them
4791 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4792 *no_progress_loops = 0;
4794 (*no_progress_loops)++;
4797 * Make sure we converge to OOM if we cannot make any progress
4798 * several times in the row.
4800 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4801 /* Before OOM, exhaust highatomic_reserve */
4802 return unreserve_highatomic_pageblock(ac, true);
4806 * Keep reclaiming pages while there is a chance this will lead
4807 * somewhere. If none of the target zones can satisfy our allocation
4808 * request even if all reclaimable pages are considered then we are
4809 * screwed and have to go OOM.
4811 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4812 ac->highest_zoneidx, ac->nodemask) {
4813 unsigned long available;
4814 unsigned long reclaimable;
4815 unsigned long min_wmark = min_wmark_pages(zone);
4818 available = reclaimable = zone_reclaimable_pages(zone);
4819 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4822 * Would the allocation succeed if we reclaimed all
4823 * reclaimable pages?
4825 wmark = __zone_watermark_ok(zone, order, min_wmark,
4826 ac->highest_zoneidx, alloc_flags, available);
4827 trace_reclaim_retry_zone(z, order, reclaimable,
4828 available, min_wmark, *no_progress_loops, wmark);
4831 * If we didn't make any progress and have a lot of
4832 * dirty + writeback pages then we should wait for
4833 * an IO to complete to slow down the reclaim and
4834 * prevent from pre mature OOM
4836 if (!did_some_progress) {
4837 unsigned long write_pending;
4839 write_pending = zone_page_state_snapshot(zone,
4840 NR_ZONE_WRITE_PENDING);
4842 if (2 * write_pending > reclaimable) {
4843 congestion_wait(BLK_RW_ASYNC, HZ/10);
4855 * Memory allocation/reclaim might be called from a WQ context and the
4856 * current implementation of the WQ concurrency control doesn't
4857 * recognize that a particular WQ is congested if the worker thread is
4858 * looping without ever sleeping. Therefore we have to do a short sleep
4859 * here rather than calling cond_resched().
4861 if (current->flags & PF_WQ_WORKER)
4862 schedule_timeout_uninterruptible(1);
4869 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4872 * It's possible that cpuset's mems_allowed and the nodemask from
4873 * mempolicy don't intersect. This should be normally dealt with by
4874 * policy_nodemask(), but it's possible to race with cpuset update in
4875 * such a way the check therein was true, and then it became false
4876 * before we got our cpuset_mems_cookie here.
4877 * This assumes that for all allocations, ac->nodemask can come only
4878 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4879 * when it does not intersect with the cpuset restrictions) or the
4880 * caller can deal with a violated nodemask.
4882 if (cpusets_enabled() && ac->nodemask &&
4883 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4884 ac->nodemask = NULL;
4889 * When updating a task's mems_allowed or mempolicy nodemask, it is
4890 * possible to race with parallel threads in such a way that our
4891 * allocation can fail while the mask is being updated. If we are about
4892 * to fail, check if the cpuset changed during allocation and if so,
4895 if (read_mems_allowed_retry(cpuset_mems_cookie))
4901 static inline struct page *
4902 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4903 struct alloc_context *ac)
4905 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4906 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4907 struct page *page = NULL;
4908 unsigned int alloc_flags;
4909 unsigned long did_some_progress;
4910 enum compact_priority compact_priority;
4911 enum compact_result compact_result;
4912 int compaction_retries;
4913 int no_progress_loops;
4914 unsigned int cpuset_mems_cookie;
4915 unsigned int zonelist_iter_cookie;
4919 * We also sanity check to catch abuse of atomic reserves being used by
4920 * callers that are not in atomic context.
4922 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4923 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4924 gfp_mask &= ~__GFP_ATOMIC;
4927 compaction_retries = 0;
4928 no_progress_loops = 0;
4929 compact_priority = DEF_COMPACT_PRIORITY;
4930 cpuset_mems_cookie = read_mems_allowed_begin();
4931 zonelist_iter_cookie = zonelist_iter_begin();
4934 * The fast path uses conservative alloc_flags to succeed only until
4935 * kswapd needs to be woken up, and to avoid the cost of setting up
4936 * alloc_flags precisely. So we do that now.
4938 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4941 * We need to recalculate the starting point for the zonelist iterator
4942 * because we might have used different nodemask in the fast path, or
4943 * there was a cpuset modification and we are retrying - otherwise we
4944 * could end up iterating over non-eligible zones endlessly.
4946 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4947 ac->highest_zoneidx, ac->nodemask);
4948 if (!ac->preferred_zoneref->zone)
4951 if (alloc_flags & ALLOC_KSWAPD)
4952 wake_all_kswapds(order, gfp_mask, ac);
4955 * The adjusted alloc_flags might result in immediate success, so try
4958 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4963 * For costly allocations, try direct compaction first, as it's likely
4964 * that we have enough base pages and don't need to reclaim. For non-
4965 * movable high-order allocations, do that as well, as compaction will
4966 * try prevent permanent fragmentation by migrating from blocks of the
4968 * Don't try this for allocations that are allowed to ignore
4969 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4971 if (can_direct_reclaim &&
4973 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4974 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4975 page = __alloc_pages_direct_compact(gfp_mask, order,
4977 INIT_COMPACT_PRIORITY,
4983 * Checks for costly allocations with __GFP_NORETRY, which
4984 * includes some THP page fault allocations
4986 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4988 * If allocating entire pageblock(s) and compaction
4989 * failed because all zones are below low watermarks
4990 * or is prohibited because it recently failed at this
4991 * order, fail immediately unless the allocator has
4992 * requested compaction and reclaim retry.
4995 * - potentially very expensive because zones are far
4996 * below their low watermarks or this is part of very
4997 * bursty high order allocations,
4998 * - not guaranteed to help because isolate_freepages()
4999 * may not iterate over freed pages as part of its
5001 * - unlikely to make entire pageblocks free on its
5004 if (compact_result == COMPACT_SKIPPED ||
5005 compact_result == COMPACT_DEFERRED)
5009 * Looks like reclaim/compaction is worth trying, but
5010 * sync compaction could be very expensive, so keep
5011 * using async compaction.
5013 compact_priority = INIT_COMPACT_PRIORITY;
5018 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5019 if (alloc_flags & ALLOC_KSWAPD)
5020 wake_all_kswapds(order, gfp_mask, ac);
5022 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5024 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
5027 * Reset the nodemask and zonelist iterators if memory policies can be
5028 * ignored. These allocations are high priority and system rather than
5031 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5032 ac->nodemask = NULL;
5033 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5034 ac->highest_zoneidx, ac->nodemask);
5037 /* Attempt with potentially adjusted zonelist and alloc_flags */
5038 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5042 /* Caller is not willing to reclaim, we can't balance anything */
5043 if (!can_direct_reclaim)
5046 /* Avoid recursion of direct reclaim */
5047 if (current->flags & PF_MEMALLOC)
5050 /* Try direct reclaim and then allocating */
5051 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5052 &did_some_progress);
5056 /* Try direct compaction and then allocating */
5057 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5058 compact_priority, &compact_result);
5062 /* Do not loop if specifically requested */
5063 if (gfp_mask & __GFP_NORETRY)
5067 * Do not retry costly high order allocations unless they are
5068 * __GFP_RETRY_MAYFAIL
5070 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5073 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5074 did_some_progress > 0, &no_progress_loops))
5078 * It doesn't make any sense to retry for the compaction if the order-0
5079 * reclaim is not able to make any progress because the current
5080 * implementation of the compaction depends on the sufficient amount
5081 * of free memory (see __compaction_suitable)
5083 if (did_some_progress > 0 &&
5084 should_compact_retry(ac, order, alloc_flags,
5085 compact_result, &compact_priority,
5086 &compaction_retries))
5091 * Deal with possible cpuset update races or zonelist updates to avoid
5092 * a unnecessary OOM kill.
5094 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5095 check_retry_zonelist(zonelist_iter_cookie))
5098 /* Reclaim has failed us, start killing things */
5099 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5103 /* Avoid allocations with no watermarks from looping endlessly */
5104 if (tsk_is_oom_victim(current) &&
5105 (alloc_flags & ALLOC_OOM ||
5106 (gfp_mask & __GFP_NOMEMALLOC)))
5109 /* Retry as long as the OOM killer is making progress */
5110 if (did_some_progress) {
5111 no_progress_loops = 0;
5117 * Deal with possible cpuset update races or zonelist updates to avoid
5118 * a unnecessary OOM kill.
5120 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5121 check_retry_zonelist(zonelist_iter_cookie))
5125 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5128 if (gfp_mask & __GFP_NOFAIL) {
5130 * All existing users of the __GFP_NOFAIL are blockable, so warn
5131 * of any new users that actually require GFP_NOWAIT
5133 if (WARN_ON_ONCE(!can_direct_reclaim))
5137 * PF_MEMALLOC request from this context is rather bizarre
5138 * because we cannot reclaim anything and only can loop waiting
5139 * for somebody to do a work for us
5141 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5144 * non failing costly orders are a hard requirement which we
5145 * are not prepared for much so let's warn about these users
5146 * so that we can identify them and convert them to something
5149 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5152 * Help non-failing allocations by giving them access to memory
5153 * reserves but do not use ALLOC_NO_WATERMARKS because this
5154 * could deplete whole memory reserves which would just make
5155 * the situation worse
5157 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5165 warn_alloc(gfp_mask, ac->nodemask,
5166 "page allocation failure: order:%u", order);
5171 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5172 int preferred_nid, nodemask_t *nodemask,
5173 struct alloc_context *ac, gfp_t *alloc_gfp,
5174 unsigned int *alloc_flags)
5176 ac->highest_zoneidx = gfp_zone(gfp_mask);
5177 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5178 ac->nodemask = nodemask;
5179 ac->migratetype = gfp_migratetype(gfp_mask);
5181 if (cpusets_enabled()) {
5182 *alloc_gfp |= __GFP_HARDWALL;
5184 * When we are in the interrupt context, it is irrelevant
5185 * to the current task context. It means that any node ok.
5187 if (in_task() && !ac->nodemask)
5188 ac->nodemask = &cpuset_current_mems_allowed;
5190 *alloc_flags |= ALLOC_CPUSET;
5193 fs_reclaim_acquire(gfp_mask);
5194 fs_reclaim_release(gfp_mask);
5196 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5198 if (should_fail_alloc_page(gfp_mask, order))
5201 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5203 /* Dirty zone balancing only done in the fast path */
5204 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5207 * The preferred zone is used for statistics but crucially it is
5208 * also used as the starting point for the zonelist iterator. It
5209 * may get reset for allocations that ignore memory policies.
5211 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5212 ac->highest_zoneidx, ac->nodemask);
5218 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5219 * @gfp: GFP flags for the allocation
5220 * @preferred_nid: The preferred NUMA node ID to allocate from
5221 * @nodemask: Set of nodes to allocate from, may be NULL
5222 * @nr_pages: The number of pages desired on the list or array
5223 * @page_list: Optional list to store the allocated pages
5224 * @page_array: Optional array to store the pages
5226 * This is a batched version of the page allocator that attempts to
5227 * allocate nr_pages quickly. Pages are added to page_list if page_list
5228 * is not NULL, otherwise it is assumed that the page_array is valid.
5230 * For lists, nr_pages is the number of pages that should be allocated.
5232 * For arrays, only NULL elements are populated with pages and nr_pages
5233 * is the maximum number of pages that will be stored in the array.
5235 * Returns the number of pages on the list or array.
5237 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5238 nodemask_t *nodemask, int nr_pages,
5239 struct list_head *page_list,
5240 struct page **page_array)
5243 unsigned long flags;
5246 struct per_cpu_pages *pcp;
5247 struct list_head *pcp_list;
5248 struct alloc_context ac;
5250 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5251 int nr_populated = 0, nr_account = 0;
5254 * Skip populated array elements to determine if any pages need
5255 * to be allocated before disabling IRQs.
5257 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5260 /* No pages requested? */
5261 if (unlikely(nr_pages <= 0))
5264 /* Already populated array? */
5265 if (unlikely(page_array && nr_pages - nr_populated == 0))
5268 /* Bulk allocator does not support memcg accounting. */
5269 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5272 /* Use the single page allocator for one page. */
5273 if (nr_pages - nr_populated == 1)
5276 #ifdef CONFIG_PAGE_OWNER
5278 * PAGE_OWNER may recurse into the allocator to allocate space to
5279 * save the stack with pagesets.lock held. Releasing/reacquiring
5280 * removes much of the performance benefit of bulk allocation so
5281 * force the caller to allocate one page at a time as it'll have
5282 * similar performance to added complexity to the bulk allocator.
5284 if (static_branch_unlikely(&page_owner_inited))
5288 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5289 gfp &= gfp_allowed_mask;
5291 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5295 /* Find an allowed local zone that meets the low watermark. */
5296 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5299 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5300 !__cpuset_zone_allowed(zone, gfp)) {
5304 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5305 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5309 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5310 if (zone_watermark_fast(zone, 0, mark,
5311 zonelist_zone_idx(ac.preferred_zoneref),
5312 alloc_flags, gfp)) {
5318 * If there are no allowed local zones that meets the watermarks then
5319 * try to allocate a single page and reclaim if necessary.
5321 if (unlikely(!zone))
5324 /* Attempt the batch allocation */
5325 local_lock_irqsave(&pagesets.lock, flags);
5326 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5327 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5329 while (nr_populated < nr_pages) {
5331 /* Skip existing pages */
5332 if (page_array && page_array[nr_populated]) {
5337 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5339 if (unlikely(!page)) {
5340 /* Try and allocate at least one page */
5347 prep_new_page(page, 0, gfp, 0);
5349 list_add(&page->lru, page_list);
5351 page_array[nr_populated] = page;
5355 local_unlock_irqrestore(&pagesets.lock, flags);
5357 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5358 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5361 return nr_populated;
5364 local_unlock_irqrestore(&pagesets.lock, flags);
5367 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5370 list_add(&page->lru, page_list);
5372 page_array[nr_populated] = page;
5378 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5381 * This is the 'heart' of the zoned buddy allocator.
5383 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5384 nodemask_t *nodemask)
5387 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5388 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5389 struct alloc_context ac = { };
5392 * There are several places where we assume that the order value is sane
5393 * so bail out early if the request is out of bound.
5395 if (unlikely(order >= MAX_ORDER)) {
5396 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5400 gfp &= gfp_allowed_mask;
5402 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5403 * resp. GFP_NOIO which has to be inherited for all allocation requests
5404 * from a particular context which has been marked by
5405 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5406 * movable zones are not used during allocation.
5408 gfp = current_gfp_context(gfp);
5410 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5411 &alloc_gfp, &alloc_flags))
5415 * Forbid the first pass from falling back to types that fragment
5416 * memory until all local zones are considered.
5418 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5420 /* First allocation attempt */
5421 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5426 ac.spread_dirty_pages = false;
5429 * Restore the original nodemask if it was potentially replaced with
5430 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5432 ac.nodemask = nodemask;
5434 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5437 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5438 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5439 __free_pages(page, order);
5443 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5447 EXPORT_SYMBOL(__alloc_pages);
5450 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5451 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5452 * you need to access high mem.
5454 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5458 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5461 return (unsigned long) page_address(page);
5463 EXPORT_SYMBOL(__get_free_pages);
5465 unsigned long get_zeroed_page(gfp_t gfp_mask)
5467 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5469 EXPORT_SYMBOL(get_zeroed_page);
5472 * __free_pages - Free pages allocated with alloc_pages().
5473 * @page: The page pointer returned from alloc_pages().
5474 * @order: The order of the allocation.
5476 * This function can free multi-page allocations that are not compound
5477 * pages. It does not check that the @order passed in matches that of
5478 * the allocation, so it is easy to leak memory. Freeing more memory
5479 * than was allocated will probably emit a warning.
5481 * If the last reference to this page is speculative, it will be released
5482 * by put_page() which only frees the first page of a non-compound
5483 * allocation. To prevent the remaining pages from being leaked, we free
5484 * the subsequent pages here. If you want to use the page's reference
5485 * count to decide when to free the allocation, you should allocate a
5486 * compound page, and use put_page() instead of __free_pages().
5488 * Context: May be called in interrupt context or while holding a normal
5489 * spinlock, but not in NMI context or while holding a raw spinlock.
5491 void __free_pages(struct page *page, unsigned int order)
5493 if (put_page_testzero(page))
5494 free_the_page(page, order);
5495 else if (!PageHead(page))
5497 free_the_page(page + (1 << order), order);
5499 EXPORT_SYMBOL(__free_pages);
5501 void free_pages(unsigned long addr, unsigned int order)
5504 VM_BUG_ON(!virt_addr_valid((void *)addr));
5505 __free_pages(virt_to_page((void *)addr), order);
5509 EXPORT_SYMBOL(free_pages);
5513 * An arbitrary-length arbitrary-offset area of memory which resides
5514 * within a 0 or higher order page. Multiple fragments within that page
5515 * are individually refcounted, in the page's reference counter.
5517 * The page_frag functions below provide a simple allocation framework for
5518 * page fragments. This is used by the network stack and network device
5519 * drivers to provide a backing region of memory for use as either an
5520 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5522 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5525 struct page *page = NULL;
5526 gfp_t gfp = gfp_mask;
5528 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5529 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5531 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5532 PAGE_FRAG_CACHE_MAX_ORDER);
5533 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5535 if (unlikely(!page))
5536 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5538 nc->va = page ? page_address(page) : NULL;
5543 void __page_frag_cache_drain(struct page *page, unsigned int count)
5545 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5547 if (page_ref_sub_and_test(page, count))
5548 free_the_page(page, compound_order(page));
5550 EXPORT_SYMBOL(__page_frag_cache_drain);
5552 void *page_frag_alloc_align(struct page_frag_cache *nc,
5553 unsigned int fragsz, gfp_t gfp_mask,
5554 unsigned int align_mask)
5556 unsigned int size = PAGE_SIZE;
5560 if (unlikely(!nc->va)) {
5562 page = __page_frag_cache_refill(nc, gfp_mask);
5566 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5567 /* if size can vary use size else just use PAGE_SIZE */
5570 /* Even if we own the page, we do not use atomic_set().
5571 * This would break get_page_unless_zero() users.
5573 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5575 /* reset page count bias and offset to start of new frag */
5576 nc->pfmemalloc = page_is_pfmemalloc(page);
5577 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5581 offset = nc->offset - fragsz;
5582 if (unlikely(offset < 0)) {
5583 page = virt_to_page(nc->va);
5585 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5588 if (unlikely(nc->pfmemalloc)) {
5589 free_the_page(page, compound_order(page));
5593 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5594 /* if size can vary use size else just use PAGE_SIZE */
5597 /* OK, page count is 0, we can safely set it */
5598 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5600 /* reset page count bias and offset to start of new frag */
5601 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5602 offset = size - fragsz;
5603 if (unlikely(offset < 0)) {
5605 * The caller is trying to allocate a fragment
5606 * with fragsz > PAGE_SIZE but the cache isn't big
5607 * enough to satisfy the request, this may
5608 * happen in low memory conditions.
5609 * We don't release the cache page because
5610 * it could make memory pressure worse
5611 * so we simply return NULL here.
5618 offset &= align_mask;
5619 nc->offset = offset;
5621 return nc->va + offset;
5623 EXPORT_SYMBOL(page_frag_alloc_align);
5626 * Frees a page fragment allocated out of either a compound or order 0 page.
5628 void page_frag_free(void *addr)
5630 struct page *page = virt_to_head_page(addr);
5632 if (unlikely(put_page_testzero(page)))
5633 free_the_page(page, compound_order(page));
5635 EXPORT_SYMBOL(page_frag_free);
5637 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5641 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5642 unsigned long used = addr + PAGE_ALIGN(size);
5644 split_page(virt_to_page((void *)addr), order);
5645 while (used < alloc_end) {
5650 return (void *)addr;
5654 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5655 * @size: the number of bytes to allocate
5656 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5658 * This function is similar to alloc_pages(), except that it allocates the
5659 * minimum number of pages to satisfy the request. alloc_pages() can only
5660 * allocate memory in power-of-two pages.
5662 * This function is also limited by MAX_ORDER.
5664 * Memory allocated by this function must be released by free_pages_exact().
5666 * Return: pointer to the allocated area or %NULL in case of error.
5668 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5670 unsigned int order = get_order(size);
5673 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5674 gfp_mask &= ~__GFP_COMP;
5676 addr = __get_free_pages(gfp_mask, order);
5677 return make_alloc_exact(addr, order, size);
5679 EXPORT_SYMBOL(alloc_pages_exact);
5682 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5684 * @nid: the preferred node ID where memory should be allocated
5685 * @size: the number of bytes to allocate
5686 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5688 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5691 * Return: pointer to the allocated area or %NULL in case of error.
5693 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5695 unsigned int order = get_order(size);
5698 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5699 gfp_mask &= ~__GFP_COMP;
5701 p = alloc_pages_node(nid, gfp_mask, order);
5704 return make_alloc_exact((unsigned long)page_address(p), order, size);
5708 * free_pages_exact - release memory allocated via alloc_pages_exact()
5709 * @virt: the value returned by alloc_pages_exact.
5710 * @size: size of allocation, same value as passed to alloc_pages_exact().
5712 * Release the memory allocated by a previous call to alloc_pages_exact.
5714 void free_pages_exact(void *virt, size_t size)
5716 unsigned long addr = (unsigned long)virt;
5717 unsigned long end = addr + PAGE_ALIGN(size);
5719 while (addr < end) {
5724 EXPORT_SYMBOL(free_pages_exact);
5727 * nr_free_zone_pages - count number of pages beyond high watermark
5728 * @offset: The zone index of the highest zone
5730 * nr_free_zone_pages() counts the number of pages which are beyond the
5731 * high watermark within all zones at or below a given zone index. For each
5732 * zone, the number of pages is calculated as:
5734 * nr_free_zone_pages = managed_pages - high_pages
5736 * Return: number of pages beyond high watermark.
5738 static unsigned long nr_free_zone_pages(int offset)
5743 /* Just pick one node, since fallback list is circular */
5744 unsigned long sum = 0;
5746 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5748 for_each_zone_zonelist(zone, z, zonelist, offset) {
5749 unsigned long size = zone_managed_pages(zone);
5750 unsigned long high = high_wmark_pages(zone);
5759 * nr_free_buffer_pages - count number of pages beyond high watermark
5761 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5762 * watermark within ZONE_DMA and ZONE_NORMAL.
5764 * Return: number of pages beyond high watermark within ZONE_DMA and
5767 unsigned long nr_free_buffer_pages(void)
5769 return nr_free_zone_pages(gfp_zone(GFP_USER));
5771 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5773 static inline void show_node(struct zone *zone)
5775 if (IS_ENABLED(CONFIG_NUMA))
5776 printk("Node %d ", zone_to_nid(zone));
5779 long si_mem_available(void)
5782 unsigned long pagecache;
5783 unsigned long wmark_low = 0;
5784 unsigned long pages[NR_LRU_LISTS];
5785 unsigned long reclaimable;
5789 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5790 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5793 wmark_low += low_wmark_pages(zone);
5796 * Estimate the amount of memory available for userspace allocations,
5797 * without causing swapping.
5799 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5802 * Not all the page cache can be freed, otherwise the system will
5803 * start swapping. Assume at least half of the page cache, or the
5804 * low watermark worth of cache, needs to stay.
5806 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5807 pagecache -= min(pagecache / 2, wmark_low);
5808 available += pagecache;
5811 * Part of the reclaimable slab and other kernel memory consists of
5812 * items that are in use, and cannot be freed. Cap this estimate at the
5815 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5816 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5817 available += reclaimable - min(reclaimable / 2, wmark_low);
5823 EXPORT_SYMBOL_GPL(si_mem_available);
5825 void si_meminfo(struct sysinfo *val)
5827 val->totalram = totalram_pages();
5828 val->sharedram = global_node_page_state(NR_SHMEM);
5829 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5830 val->bufferram = nr_blockdev_pages();
5831 val->totalhigh = totalhigh_pages();
5832 val->freehigh = nr_free_highpages();
5833 val->mem_unit = PAGE_SIZE;
5836 EXPORT_SYMBOL(si_meminfo);
5839 void si_meminfo_node(struct sysinfo *val, int nid)
5841 int zone_type; /* needs to be signed */
5842 unsigned long managed_pages = 0;
5843 unsigned long managed_highpages = 0;
5844 unsigned long free_highpages = 0;
5845 pg_data_t *pgdat = NODE_DATA(nid);
5847 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5848 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5849 val->totalram = managed_pages;
5850 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5851 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5852 #ifdef CONFIG_HIGHMEM
5853 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5854 struct zone *zone = &pgdat->node_zones[zone_type];
5856 if (is_highmem(zone)) {
5857 managed_highpages += zone_managed_pages(zone);
5858 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5861 val->totalhigh = managed_highpages;
5862 val->freehigh = free_highpages;
5864 val->totalhigh = managed_highpages;
5865 val->freehigh = free_highpages;
5867 val->mem_unit = PAGE_SIZE;
5872 * Determine whether the node should be displayed or not, depending on whether
5873 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5875 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5877 if (!(flags & SHOW_MEM_FILTER_NODES))
5881 * no node mask - aka implicit memory numa policy. Do not bother with
5882 * the synchronization - read_mems_allowed_begin - because we do not
5883 * have to be precise here.
5886 nodemask = &cpuset_current_mems_allowed;
5888 return !node_isset(nid, *nodemask);
5891 #define K(x) ((x) << (PAGE_SHIFT-10))
5893 static void show_migration_types(unsigned char type)
5895 static const char types[MIGRATE_TYPES] = {
5896 [MIGRATE_UNMOVABLE] = 'U',
5897 [MIGRATE_MOVABLE] = 'M',
5898 [MIGRATE_RECLAIMABLE] = 'E',
5899 [MIGRATE_HIGHATOMIC] = 'H',
5901 [MIGRATE_CMA] = 'C',
5903 #ifdef CONFIG_MEMORY_ISOLATION
5904 [MIGRATE_ISOLATE] = 'I',
5907 char tmp[MIGRATE_TYPES + 1];
5911 for (i = 0; i < MIGRATE_TYPES; i++) {
5912 if (type & (1 << i))
5917 printk(KERN_CONT "(%s) ", tmp);
5921 * Show free area list (used inside shift_scroll-lock stuff)
5922 * We also calculate the percentage fragmentation. We do this by counting the
5923 * memory on each free list with the exception of the first item on the list.
5926 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5929 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5931 unsigned long free_pcp = 0;
5936 for_each_populated_zone(zone) {
5937 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5940 for_each_online_cpu(cpu)
5941 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5944 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5945 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5946 " unevictable:%lu dirty:%lu writeback:%lu\n"
5947 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5948 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5949 " kernel_misc_reclaimable:%lu\n"
5950 " free:%lu free_pcp:%lu free_cma:%lu\n",
5951 global_node_page_state(NR_ACTIVE_ANON),
5952 global_node_page_state(NR_INACTIVE_ANON),
5953 global_node_page_state(NR_ISOLATED_ANON),
5954 global_node_page_state(NR_ACTIVE_FILE),
5955 global_node_page_state(NR_INACTIVE_FILE),
5956 global_node_page_state(NR_ISOLATED_FILE),
5957 global_node_page_state(NR_UNEVICTABLE),
5958 global_node_page_state(NR_FILE_DIRTY),
5959 global_node_page_state(NR_WRITEBACK),
5960 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5961 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5962 global_node_page_state(NR_FILE_MAPPED),
5963 global_node_page_state(NR_SHMEM),
5964 global_node_page_state(NR_PAGETABLE),
5965 global_zone_page_state(NR_BOUNCE),
5966 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5967 global_zone_page_state(NR_FREE_PAGES),
5969 global_zone_page_state(NR_FREE_CMA_PAGES));
5971 for_each_online_pgdat(pgdat) {
5972 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5976 " active_anon:%lukB"
5977 " inactive_anon:%lukB"
5978 " active_file:%lukB"
5979 " inactive_file:%lukB"
5980 " unevictable:%lukB"
5981 " isolated(anon):%lukB"
5982 " isolated(file):%lukB"
5987 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5989 " shmem_pmdmapped: %lukB"
5992 " writeback_tmp:%lukB"
5993 " kernel_stack:%lukB"
5994 #ifdef CONFIG_SHADOW_CALL_STACK
5995 " shadow_call_stack:%lukB"
5998 " all_unreclaimable? %s"
6001 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
6002 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
6003 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
6004 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
6005 K(node_page_state(pgdat, NR_UNEVICTABLE)),
6006 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
6007 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
6008 K(node_page_state(pgdat, NR_FILE_MAPPED)),
6009 K(node_page_state(pgdat, NR_FILE_DIRTY)),
6010 K(node_page_state(pgdat, NR_WRITEBACK)),
6011 K(node_page_state(pgdat, NR_SHMEM)),
6012 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6013 K(node_page_state(pgdat, NR_SHMEM_THPS)),
6014 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6015 K(node_page_state(pgdat, NR_ANON_THPS)),
6017 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6018 node_page_state(pgdat, NR_KERNEL_STACK_KB),
6019 #ifdef CONFIG_SHADOW_CALL_STACK
6020 node_page_state(pgdat, NR_KERNEL_SCS_KB),
6022 K(node_page_state(pgdat, NR_PAGETABLE)),
6023 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6027 for_each_populated_zone(zone) {
6030 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6034 for_each_online_cpu(cpu)
6035 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6044 " reserved_highatomic:%luKB"
6045 " active_anon:%lukB"
6046 " inactive_anon:%lukB"
6047 " active_file:%lukB"
6048 " inactive_file:%lukB"
6049 " unevictable:%lukB"
6050 " writepending:%lukB"
6060 K(zone_page_state(zone, NR_FREE_PAGES)),
6061 K(min_wmark_pages(zone)),
6062 K(low_wmark_pages(zone)),
6063 K(high_wmark_pages(zone)),
6064 K(zone->nr_reserved_highatomic),
6065 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6066 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6067 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6068 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6069 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6070 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6071 K(zone->present_pages),
6072 K(zone_managed_pages(zone)),
6073 K(zone_page_state(zone, NR_MLOCK)),
6074 K(zone_page_state(zone, NR_BOUNCE)),
6076 K(this_cpu_read(zone->per_cpu_pageset->count)),
6077 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6078 printk("lowmem_reserve[]:");
6079 for (i = 0; i < MAX_NR_ZONES; i++)
6080 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6081 printk(KERN_CONT "\n");
6084 for_each_populated_zone(zone) {
6086 unsigned long nr[MAX_ORDER], flags, total = 0;
6087 unsigned char types[MAX_ORDER];
6089 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6092 printk(KERN_CONT "%s: ", zone->name);
6094 spin_lock_irqsave(&zone->lock, flags);
6095 for (order = 0; order < MAX_ORDER; order++) {
6096 struct free_area *area = &zone->free_area[order];
6099 nr[order] = area->nr_free;
6100 total += nr[order] << order;
6103 for (type = 0; type < MIGRATE_TYPES; type++) {
6104 if (!free_area_empty(area, type))
6105 types[order] |= 1 << type;
6108 spin_unlock_irqrestore(&zone->lock, flags);
6109 for (order = 0; order < MAX_ORDER; order++) {
6110 printk(KERN_CONT "%lu*%lukB ",
6111 nr[order], K(1UL) << order);
6113 show_migration_types(types[order]);
6115 printk(KERN_CONT "= %lukB\n", K(total));
6118 hugetlb_show_meminfo();
6120 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6122 show_swap_cache_info();
6125 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6127 zoneref->zone = zone;
6128 zoneref->zone_idx = zone_idx(zone);
6132 * Builds allocation fallback zone lists.
6134 * Add all populated zones of a node to the zonelist.
6136 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6139 enum zone_type zone_type = MAX_NR_ZONES;
6144 zone = pgdat->node_zones + zone_type;
6145 if (populated_zone(zone)) {
6146 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6147 check_highest_zone(zone_type);
6149 } while (zone_type);
6156 static int __parse_numa_zonelist_order(char *s)
6159 * We used to support different zonelists modes but they turned
6160 * out to be just not useful. Let's keep the warning in place
6161 * if somebody still use the cmd line parameter so that we do
6162 * not fail it silently
6164 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6165 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6171 char numa_zonelist_order[] = "Node";
6174 * sysctl handler for numa_zonelist_order
6176 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6177 void *buffer, size_t *length, loff_t *ppos)
6180 return __parse_numa_zonelist_order(buffer);
6181 return proc_dostring(table, write, buffer, length, ppos);
6185 #define MAX_NODE_LOAD (nr_online_nodes)
6186 static int node_load[MAX_NUMNODES];
6189 * find_next_best_node - find the next node that should appear in a given node's fallback list
6190 * @node: node whose fallback list we're appending
6191 * @used_node_mask: nodemask_t of already used nodes
6193 * We use a number of factors to determine which is the next node that should
6194 * appear on a given node's fallback list. The node should not have appeared
6195 * already in @node's fallback list, and it should be the next closest node
6196 * according to the distance array (which contains arbitrary distance values
6197 * from each node to each node in the system), and should also prefer nodes
6198 * with no CPUs, since presumably they'll have very little allocation pressure
6199 * on them otherwise.
6201 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6203 int find_next_best_node(int node, nodemask_t *used_node_mask)
6206 int min_val = INT_MAX;
6207 int best_node = NUMA_NO_NODE;
6209 /* Use the local node if we haven't already */
6210 if (!node_isset(node, *used_node_mask)) {
6211 node_set(node, *used_node_mask);
6215 for_each_node_state(n, N_MEMORY) {
6217 /* Don't want a node to appear more than once */
6218 if (node_isset(n, *used_node_mask))
6221 /* Use the distance array to find the distance */
6222 val = node_distance(node, n);
6224 /* Penalize nodes under us ("prefer the next node") */
6227 /* Give preference to headless and unused nodes */
6228 if (!cpumask_empty(cpumask_of_node(n)))
6229 val += PENALTY_FOR_NODE_WITH_CPUS;
6231 /* Slight preference for less loaded node */
6232 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6233 val += node_load[n];
6235 if (val < min_val) {
6242 node_set(best_node, *used_node_mask);
6249 * Build zonelists ordered by node and zones within node.
6250 * This results in maximum locality--normal zone overflows into local
6251 * DMA zone, if any--but risks exhausting DMA zone.
6253 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6256 struct zoneref *zonerefs;
6259 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6261 for (i = 0; i < nr_nodes; i++) {
6264 pg_data_t *node = NODE_DATA(node_order[i]);
6266 nr_zones = build_zonerefs_node(node, zonerefs);
6267 zonerefs += nr_zones;
6269 zonerefs->zone = NULL;
6270 zonerefs->zone_idx = 0;
6274 * Build gfp_thisnode zonelists
6276 static void build_thisnode_zonelists(pg_data_t *pgdat)
6278 struct zoneref *zonerefs;
6281 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6282 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6283 zonerefs += nr_zones;
6284 zonerefs->zone = NULL;
6285 zonerefs->zone_idx = 0;
6289 * Build zonelists ordered by zone and nodes within zones.
6290 * This results in conserving DMA zone[s] until all Normal memory is
6291 * exhausted, but results in overflowing to remote node while memory
6292 * may still exist in local DMA zone.
6295 static void build_zonelists(pg_data_t *pgdat)
6297 static int node_order[MAX_NUMNODES];
6298 int node, load, nr_nodes = 0;
6299 nodemask_t used_mask = NODE_MASK_NONE;
6300 int local_node, prev_node;
6302 /* NUMA-aware ordering of nodes */
6303 local_node = pgdat->node_id;
6304 load = nr_online_nodes;
6305 prev_node = local_node;
6307 memset(node_order, 0, sizeof(node_order));
6308 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6310 * We don't want to pressure a particular node.
6311 * So adding penalty to the first node in same
6312 * distance group to make it round-robin.
6314 if (node_distance(local_node, node) !=
6315 node_distance(local_node, prev_node))
6316 node_load[node] = load;
6318 node_order[nr_nodes++] = node;
6323 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6324 build_thisnode_zonelists(pgdat);
6327 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6329 * Return node id of node used for "local" allocations.
6330 * I.e., first node id of first zone in arg node's generic zonelist.
6331 * Used for initializing percpu 'numa_mem', which is used primarily
6332 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6334 int local_memory_node(int node)
6338 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6339 gfp_zone(GFP_KERNEL),
6341 return zone_to_nid(z->zone);
6345 static void setup_min_unmapped_ratio(void);
6346 static void setup_min_slab_ratio(void);
6347 #else /* CONFIG_NUMA */
6349 static void build_zonelists(pg_data_t *pgdat)
6351 int node, local_node;
6352 struct zoneref *zonerefs;
6355 local_node = pgdat->node_id;
6357 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6358 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6359 zonerefs += nr_zones;
6362 * Now we build the zonelist so that it contains the zones
6363 * of all the other nodes.
6364 * We don't want to pressure a particular node, so when
6365 * building the zones for node N, we make sure that the
6366 * zones coming right after the local ones are those from
6367 * node N+1 (modulo N)
6369 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6370 if (!node_online(node))
6372 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6373 zonerefs += nr_zones;
6375 for (node = 0; node < local_node; node++) {
6376 if (!node_online(node))
6378 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6379 zonerefs += nr_zones;
6382 zonerefs->zone = NULL;
6383 zonerefs->zone_idx = 0;
6386 #endif /* CONFIG_NUMA */
6389 * Boot pageset table. One per cpu which is going to be used for all
6390 * zones and all nodes. The parameters will be set in such a way
6391 * that an item put on a list will immediately be handed over to
6392 * the buddy list. This is safe since pageset manipulation is done
6393 * with interrupts disabled.
6395 * The boot_pagesets must be kept even after bootup is complete for
6396 * unused processors and/or zones. They do play a role for bootstrapping
6397 * hotplugged processors.
6399 * zoneinfo_show() and maybe other functions do
6400 * not check if the processor is online before following the pageset pointer.
6401 * Other parts of the kernel may not check if the zone is available.
6403 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6404 /* These effectively disable the pcplists in the boot pageset completely */
6405 #define BOOT_PAGESET_HIGH 0
6406 #define BOOT_PAGESET_BATCH 1
6407 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6408 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6409 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6411 static void __build_all_zonelists(void *data)
6414 int __maybe_unused cpu;
6415 pg_data_t *self = data;
6417 write_seqlock(&zonelist_update_seq);
6420 memset(node_load, 0, sizeof(node_load));
6424 * This node is hotadded and no memory is yet present. So just
6425 * building zonelists is fine - no need to touch other nodes.
6427 if (self && !node_online(self->node_id)) {
6428 build_zonelists(self);
6430 for_each_online_node(nid) {
6431 pg_data_t *pgdat = NODE_DATA(nid);
6433 build_zonelists(pgdat);
6436 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6438 * We now know the "local memory node" for each node--
6439 * i.e., the node of the first zone in the generic zonelist.
6440 * Set up numa_mem percpu variable for on-line cpus. During
6441 * boot, only the boot cpu should be on-line; we'll init the
6442 * secondary cpus' numa_mem as they come on-line. During
6443 * node/memory hotplug, we'll fixup all on-line cpus.
6445 for_each_online_cpu(cpu)
6446 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6450 write_sequnlock(&zonelist_update_seq);
6453 static noinline void __init
6454 build_all_zonelists_init(void)
6458 __build_all_zonelists(NULL);
6461 * Initialize the boot_pagesets that are going to be used
6462 * for bootstrapping processors. The real pagesets for
6463 * each zone will be allocated later when the per cpu
6464 * allocator is available.
6466 * boot_pagesets are used also for bootstrapping offline
6467 * cpus if the system is already booted because the pagesets
6468 * are needed to initialize allocators on a specific cpu too.
6469 * F.e. the percpu allocator needs the page allocator which
6470 * needs the percpu allocator in order to allocate its pagesets
6471 * (a chicken-egg dilemma).
6473 for_each_possible_cpu(cpu)
6474 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6476 mminit_verify_zonelist();
6477 cpuset_init_current_mems_allowed();
6481 * unless system_state == SYSTEM_BOOTING.
6483 * __ref due to call of __init annotated helper build_all_zonelists_init
6484 * [protected by SYSTEM_BOOTING].
6486 void __ref build_all_zonelists(pg_data_t *pgdat)
6488 unsigned long vm_total_pages;
6490 if (system_state == SYSTEM_BOOTING) {
6491 build_all_zonelists_init();
6493 __build_all_zonelists(pgdat);
6494 /* cpuset refresh routine should be here */
6496 /* Get the number of free pages beyond high watermark in all zones. */
6497 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6499 * Disable grouping by mobility if the number of pages in the
6500 * system is too low to allow the mechanism to work. It would be
6501 * more accurate, but expensive to check per-zone. This check is
6502 * made on memory-hotadd so a system can start with mobility
6503 * disabled and enable it later
6505 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6506 page_group_by_mobility_disabled = 1;
6508 page_group_by_mobility_disabled = 0;
6510 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6512 page_group_by_mobility_disabled ? "off" : "on",
6515 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6519 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6520 static bool __meminit
6521 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6523 static struct memblock_region *r;
6525 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6526 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6527 for_each_mem_region(r) {
6528 if (*pfn < memblock_region_memory_end_pfn(r))
6532 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6533 memblock_is_mirror(r)) {
6534 *pfn = memblock_region_memory_end_pfn(r);
6542 * Initially all pages are reserved - free ones are freed
6543 * up by memblock_free_all() once the early boot process is
6544 * done. Non-atomic initialization, single-pass.
6546 * All aligned pageblocks are initialized to the specified migratetype
6547 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6548 * zone stats (e.g., nr_isolate_pageblock) are touched.
6550 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6551 unsigned long start_pfn, unsigned long zone_end_pfn,
6552 enum meminit_context context,
6553 struct vmem_altmap *altmap, int migratetype)
6555 unsigned long pfn, end_pfn = start_pfn + size;
6558 if (highest_memmap_pfn < end_pfn - 1)
6559 highest_memmap_pfn = end_pfn - 1;
6561 #ifdef CONFIG_ZONE_DEVICE
6563 * Honor reservation requested by the driver for this ZONE_DEVICE
6564 * memory. We limit the total number of pages to initialize to just
6565 * those that might contain the memory mapping. We will defer the
6566 * ZONE_DEVICE page initialization until after we have released
6569 if (zone == ZONE_DEVICE) {
6573 if (start_pfn == altmap->base_pfn)
6574 start_pfn += altmap->reserve;
6575 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6579 for (pfn = start_pfn; pfn < end_pfn; ) {
6581 * There can be holes in boot-time mem_map[]s handed to this
6582 * function. They do not exist on hotplugged memory.
6584 if (context == MEMINIT_EARLY) {
6585 if (overlap_memmap_init(zone, &pfn))
6587 if (defer_init(nid, pfn, zone_end_pfn))
6591 page = pfn_to_page(pfn);
6592 __init_single_page(page, pfn, zone, nid);
6593 if (context == MEMINIT_HOTPLUG)
6594 __SetPageReserved(page);
6597 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6598 * such that unmovable allocations won't be scattered all
6599 * over the place during system boot.
6601 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6602 set_pageblock_migratetype(page, migratetype);
6609 #ifdef CONFIG_ZONE_DEVICE
6610 void __ref memmap_init_zone_device(struct zone *zone,
6611 unsigned long start_pfn,
6612 unsigned long nr_pages,
6613 struct dev_pagemap *pgmap)
6615 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6616 struct pglist_data *pgdat = zone->zone_pgdat;
6617 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6618 unsigned long zone_idx = zone_idx(zone);
6619 unsigned long start = jiffies;
6620 int nid = pgdat->node_id;
6622 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6626 * The call to memmap_init should have already taken care
6627 * of the pages reserved for the memmap, so we can just jump to
6628 * the end of that region and start processing the device pages.
6631 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6632 nr_pages = end_pfn - start_pfn;
6635 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6636 struct page *page = pfn_to_page(pfn);
6638 __init_single_page(page, pfn, zone_idx, nid);
6641 * Mark page reserved as it will need to wait for onlining
6642 * phase for it to be fully associated with a zone.
6644 * We can use the non-atomic __set_bit operation for setting
6645 * the flag as we are still initializing the pages.
6647 __SetPageReserved(page);
6650 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6651 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6652 * ever freed or placed on a driver-private list.
6654 page->pgmap = pgmap;
6655 page->zone_device_data = NULL;
6658 * Mark the block movable so that blocks are reserved for
6659 * movable at startup. This will force kernel allocations
6660 * to reserve their blocks rather than leaking throughout
6661 * the address space during boot when many long-lived
6662 * kernel allocations are made.
6664 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6665 * because this is done early in section_activate()
6667 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6668 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6673 pr_info("%s initialised %lu pages in %ums\n", __func__,
6674 nr_pages, jiffies_to_msecs(jiffies - start));
6678 static void __meminit zone_init_free_lists(struct zone *zone)
6680 unsigned int order, t;
6681 for_each_migratetype_order(order, t) {
6682 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6683 zone->free_area[order].nr_free = 0;
6688 * Only struct pages that correspond to ranges defined by memblock.memory
6689 * are zeroed and initialized by going through __init_single_page() during
6690 * memmap_init_zone_range().
6692 * But, there could be struct pages that correspond to holes in
6693 * memblock.memory. This can happen because of the following reasons:
6694 * - physical memory bank size is not necessarily the exact multiple of the
6695 * arbitrary section size
6696 * - early reserved memory may not be listed in memblock.memory
6697 * - memory layouts defined with memmap= kernel parameter may not align
6698 * nicely with memmap sections
6700 * Explicitly initialize those struct pages so that:
6701 * - PG_Reserved is set
6702 * - zone and node links point to zone and node that span the page if the
6703 * hole is in the middle of a zone
6704 * - zone and node links point to adjacent zone/node if the hole falls on
6705 * the zone boundary; the pages in such holes will be prepended to the
6706 * zone/node above the hole except for the trailing pages in the last
6707 * section that will be appended to the zone/node below.
6709 static void __init init_unavailable_range(unsigned long spfn,
6716 for (pfn = spfn; pfn < epfn; pfn++) {
6717 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6718 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6719 + pageblock_nr_pages - 1;
6722 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6723 __SetPageReserved(pfn_to_page(pfn));
6728 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6729 node, zone_names[zone], pgcnt);
6732 static void __init memmap_init_zone_range(struct zone *zone,
6733 unsigned long start_pfn,
6734 unsigned long end_pfn,
6735 unsigned long *hole_pfn)
6737 unsigned long zone_start_pfn = zone->zone_start_pfn;
6738 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6739 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6741 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6742 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6744 if (start_pfn >= end_pfn)
6747 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6748 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6750 if (*hole_pfn < start_pfn)
6751 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6753 *hole_pfn = end_pfn;
6756 static void __init memmap_init(void)
6758 unsigned long start_pfn, end_pfn;
6759 unsigned long hole_pfn = 0;
6760 int i, j, zone_id = 0, nid;
6762 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6763 struct pglist_data *node = NODE_DATA(nid);
6765 for (j = 0; j < MAX_NR_ZONES; j++) {
6766 struct zone *zone = node->node_zones + j;
6768 if (!populated_zone(zone))
6771 memmap_init_zone_range(zone, start_pfn, end_pfn,
6777 #ifdef CONFIG_SPARSEMEM
6779 * Initialize the memory map for hole in the range [memory_end,
6781 * Append the pages in this hole to the highest zone in the last
6783 * The call to init_unavailable_range() is outside the ifdef to
6784 * silence the compiler warining about zone_id set but not used;
6785 * for FLATMEM it is a nop anyway
6787 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6788 if (hole_pfn < end_pfn)
6790 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6793 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6794 phys_addr_t min_addr, int nid, bool exact_nid)
6799 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6800 MEMBLOCK_ALLOC_ACCESSIBLE,
6803 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6804 MEMBLOCK_ALLOC_ACCESSIBLE,
6807 if (ptr && size > 0)
6808 page_init_poison(ptr, size);
6813 static int zone_batchsize(struct zone *zone)
6819 * The number of pages to batch allocate is either ~0.1%
6820 * of the zone or 1MB, whichever is smaller. The batch
6821 * size is striking a balance between allocation latency
6822 * and zone lock contention.
6824 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6825 batch /= 4; /* We effectively *= 4 below */
6830 * Clamp the batch to a 2^n - 1 value. Having a power
6831 * of 2 value was found to be more likely to have
6832 * suboptimal cache aliasing properties in some cases.
6834 * For example if 2 tasks are alternately allocating
6835 * batches of pages, one task can end up with a lot
6836 * of pages of one half of the possible page colors
6837 * and the other with pages of the other colors.
6839 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6844 /* The deferral and batching of frees should be suppressed under NOMMU
6847 * The problem is that NOMMU needs to be able to allocate large chunks
6848 * of contiguous memory as there's no hardware page translation to
6849 * assemble apparent contiguous memory from discontiguous pages.
6851 * Queueing large contiguous runs of pages for batching, however,
6852 * causes the pages to actually be freed in smaller chunks. As there
6853 * can be a significant delay between the individual batches being
6854 * recycled, this leads to the once large chunks of space being
6855 * fragmented and becoming unavailable for high-order allocations.
6861 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6866 unsigned long total_pages;
6868 if (!percpu_pagelist_high_fraction) {
6870 * By default, the high value of the pcp is based on the zone
6871 * low watermark so that if they are full then background
6872 * reclaim will not be started prematurely.
6874 total_pages = low_wmark_pages(zone);
6877 * If percpu_pagelist_high_fraction is configured, the high
6878 * value is based on a fraction of the managed pages in the
6881 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6885 * Split the high value across all online CPUs local to the zone. Note
6886 * that early in boot that CPUs may not be online yet and that during
6887 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6888 * onlined. For memory nodes that have no CPUs, split pcp->high across
6889 * all online CPUs to mitigate the risk that reclaim is triggered
6890 * prematurely due to pages stored on pcp lists.
6892 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6894 nr_split_cpus = num_online_cpus();
6895 high = total_pages / nr_split_cpus;
6898 * Ensure high is at least batch*4. The multiple is based on the
6899 * historical relationship between high and batch.
6901 high = max(high, batch << 2);
6910 * pcp->high and pcp->batch values are related and generally batch is lower
6911 * than high. They are also related to pcp->count such that count is lower
6912 * than high, and as soon as it reaches high, the pcplist is flushed.
6914 * However, guaranteeing these relations at all times would require e.g. write
6915 * barriers here but also careful usage of read barriers at the read side, and
6916 * thus be prone to error and bad for performance. Thus the update only prevents
6917 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6918 * can cope with those fields changing asynchronously, and fully trust only the
6919 * pcp->count field on the local CPU with interrupts disabled.
6921 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6922 * outside of boot time (or some other assurance that no concurrent updaters
6925 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6926 unsigned long batch)
6928 WRITE_ONCE(pcp->batch, batch);
6929 WRITE_ONCE(pcp->high, high);
6932 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6936 memset(pcp, 0, sizeof(*pcp));
6937 memset(pzstats, 0, sizeof(*pzstats));
6939 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6940 INIT_LIST_HEAD(&pcp->lists[pindex]);
6943 * Set batch and high values safe for a boot pageset. A true percpu
6944 * pageset's initialization will update them subsequently. Here we don't
6945 * need to be as careful as pageset_update() as nobody can access the
6948 pcp->high = BOOT_PAGESET_HIGH;
6949 pcp->batch = BOOT_PAGESET_BATCH;
6950 pcp->free_factor = 0;
6953 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6954 unsigned long batch)
6956 struct per_cpu_pages *pcp;
6959 for_each_possible_cpu(cpu) {
6960 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6961 pageset_update(pcp, high, batch);
6966 * Calculate and set new high and batch values for all per-cpu pagesets of a
6967 * zone based on the zone's size.
6969 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6971 int new_high, new_batch;
6973 new_batch = max(1, zone_batchsize(zone));
6974 new_high = zone_highsize(zone, new_batch, cpu_online);
6976 if (zone->pageset_high == new_high &&
6977 zone->pageset_batch == new_batch)
6980 zone->pageset_high = new_high;
6981 zone->pageset_batch = new_batch;
6983 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6986 void __meminit setup_zone_pageset(struct zone *zone)
6990 /* Size may be 0 on !SMP && !NUMA */
6991 if (sizeof(struct per_cpu_zonestat) > 0)
6992 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6994 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6995 for_each_possible_cpu(cpu) {
6996 struct per_cpu_pages *pcp;
6997 struct per_cpu_zonestat *pzstats;
6999 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7000 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7001 per_cpu_pages_init(pcp, pzstats);
7004 zone_set_pageset_high_and_batch(zone, 0);
7008 * Allocate per cpu pagesets and initialize them.
7009 * Before this call only boot pagesets were available.
7011 void __init setup_per_cpu_pageset(void)
7013 struct pglist_data *pgdat;
7015 int __maybe_unused cpu;
7017 for_each_populated_zone(zone)
7018 setup_zone_pageset(zone);
7022 * Unpopulated zones continue using the boot pagesets.
7023 * The numa stats for these pagesets need to be reset.
7024 * Otherwise, they will end up skewing the stats of
7025 * the nodes these zones are associated with.
7027 for_each_possible_cpu(cpu) {
7028 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7029 memset(pzstats->vm_numa_event, 0,
7030 sizeof(pzstats->vm_numa_event));
7034 for_each_online_pgdat(pgdat)
7035 pgdat->per_cpu_nodestats =
7036 alloc_percpu(struct per_cpu_nodestat);
7039 static __meminit void zone_pcp_init(struct zone *zone)
7042 * per cpu subsystem is not up at this point. The following code
7043 * relies on the ability of the linker to provide the
7044 * offset of a (static) per cpu variable into the per cpu area.
7046 zone->per_cpu_pageset = &boot_pageset;
7047 zone->per_cpu_zonestats = &boot_zonestats;
7048 zone->pageset_high = BOOT_PAGESET_HIGH;
7049 zone->pageset_batch = BOOT_PAGESET_BATCH;
7051 if (populated_zone(zone))
7052 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7053 zone->present_pages, zone_batchsize(zone));
7056 void __meminit init_currently_empty_zone(struct zone *zone,
7057 unsigned long zone_start_pfn,
7060 struct pglist_data *pgdat = zone->zone_pgdat;
7061 int zone_idx = zone_idx(zone) + 1;
7063 if (zone_idx > pgdat->nr_zones)
7064 pgdat->nr_zones = zone_idx;
7066 zone->zone_start_pfn = zone_start_pfn;
7068 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7069 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7071 (unsigned long)zone_idx(zone),
7072 zone_start_pfn, (zone_start_pfn + size));
7074 zone_init_free_lists(zone);
7075 zone->initialized = 1;
7079 * get_pfn_range_for_nid - Return the start and end page frames for a node
7080 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7081 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7082 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7084 * It returns the start and end page frame of a node based on information
7085 * provided by memblock_set_node(). If called for a node
7086 * with no available memory, a warning is printed and the start and end
7089 void __init get_pfn_range_for_nid(unsigned int nid,
7090 unsigned long *start_pfn, unsigned long *end_pfn)
7092 unsigned long this_start_pfn, this_end_pfn;
7098 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7099 *start_pfn = min(*start_pfn, this_start_pfn);
7100 *end_pfn = max(*end_pfn, this_end_pfn);
7103 if (*start_pfn == -1UL)
7108 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7109 * assumption is made that zones within a node are ordered in monotonic
7110 * increasing memory addresses so that the "highest" populated zone is used
7112 static void __init find_usable_zone_for_movable(void)
7115 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7116 if (zone_index == ZONE_MOVABLE)
7119 if (arch_zone_highest_possible_pfn[zone_index] >
7120 arch_zone_lowest_possible_pfn[zone_index])
7124 VM_BUG_ON(zone_index == -1);
7125 movable_zone = zone_index;
7129 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7130 * because it is sized independent of architecture. Unlike the other zones,
7131 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7132 * in each node depending on the size of each node and how evenly kernelcore
7133 * is distributed. This helper function adjusts the zone ranges
7134 * provided by the architecture for a given node by using the end of the
7135 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7136 * zones within a node are in order of monotonic increases memory addresses
7138 static void __init adjust_zone_range_for_zone_movable(int nid,
7139 unsigned long zone_type,
7140 unsigned long node_start_pfn,
7141 unsigned long node_end_pfn,
7142 unsigned long *zone_start_pfn,
7143 unsigned long *zone_end_pfn)
7145 /* Only adjust if ZONE_MOVABLE is on this node */
7146 if (zone_movable_pfn[nid]) {
7147 /* Size ZONE_MOVABLE */
7148 if (zone_type == ZONE_MOVABLE) {
7149 *zone_start_pfn = zone_movable_pfn[nid];
7150 *zone_end_pfn = min(node_end_pfn,
7151 arch_zone_highest_possible_pfn[movable_zone]);
7153 /* Adjust for ZONE_MOVABLE starting within this range */
7154 } else if (!mirrored_kernelcore &&
7155 *zone_start_pfn < zone_movable_pfn[nid] &&
7156 *zone_end_pfn > zone_movable_pfn[nid]) {
7157 *zone_end_pfn = zone_movable_pfn[nid];
7159 /* Check if this whole range is within ZONE_MOVABLE */
7160 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7161 *zone_start_pfn = *zone_end_pfn;
7166 * Return the number of pages a zone spans in a node, including holes
7167 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7169 static unsigned long __init zone_spanned_pages_in_node(int nid,
7170 unsigned long zone_type,
7171 unsigned long node_start_pfn,
7172 unsigned long node_end_pfn,
7173 unsigned long *zone_start_pfn,
7174 unsigned long *zone_end_pfn)
7176 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7177 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7178 /* When hotadd a new node from cpu_up(), the node should be empty */
7179 if (!node_start_pfn && !node_end_pfn)
7182 /* Get the start and end of the zone */
7183 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7184 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7185 adjust_zone_range_for_zone_movable(nid, zone_type,
7186 node_start_pfn, node_end_pfn,
7187 zone_start_pfn, zone_end_pfn);
7189 /* Check that this node has pages within the zone's required range */
7190 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7193 /* Move the zone boundaries inside the node if necessary */
7194 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7195 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7197 /* Return the spanned pages */
7198 return *zone_end_pfn - *zone_start_pfn;
7202 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7203 * then all holes in the requested range will be accounted for.
7205 unsigned long __init __absent_pages_in_range(int nid,
7206 unsigned long range_start_pfn,
7207 unsigned long range_end_pfn)
7209 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7210 unsigned long start_pfn, end_pfn;
7213 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7214 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7215 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7216 nr_absent -= end_pfn - start_pfn;
7222 * absent_pages_in_range - Return number of page frames in holes within a range
7223 * @start_pfn: The start PFN to start searching for holes
7224 * @end_pfn: The end PFN to stop searching for holes
7226 * Return: the number of pages frames in memory holes within a range.
7228 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7229 unsigned long end_pfn)
7231 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7234 /* Return the number of page frames in holes in a zone on a node */
7235 static unsigned long __init zone_absent_pages_in_node(int nid,
7236 unsigned long zone_type,
7237 unsigned long node_start_pfn,
7238 unsigned long node_end_pfn)
7240 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7241 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7242 unsigned long zone_start_pfn, zone_end_pfn;
7243 unsigned long nr_absent;
7245 /* When hotadd a new node from cpu_up(), the node should be empty */
7246 if (!node_start_pfn && !node_end_pfn)
7249 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7250 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7252 adjust_zone_range_for_zone_movable(nid, zone_type,
7253 node_start_pfn, node_end_pfn,
7254 &zone_start_pfn, &zone_end_pfn);
7255 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7258 * ZONE_MOVABLE handling.
7259 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7262 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7263 unsigned long start_pfn, end_pfn;
7264 struct memblock_region *r;
7266 for_each_mem_region(r) {
7267 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7268 zone_start_pfn, zone_end_pfn);
7269 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7270 zone_start_pfn, zone_end_pfn);
7272 if (zone_type == ZONE_MOVABLE &&
7273 memblock_is_mirror(r))
7274 nr_absent += end_pfn - start_pfn;
7276 if (zone_type == ZONE_NORMAL &&
7277 !memblock_is_mirror(r))
7278 nr_absent += end_pfn - start_pfn;
7285 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7286 unsigned long node_start_pfn,
7287 unsigned long node_end_pfn)
7289 unsigned long realtotalpages = 0, totalpages = 0;
7292 for (i = 0; i < MAX_NR_ZONES; i++) {
7293 struct zone *zone = pgdat->node_zones + i;
7294 unsigned long zone_start_pfn, zone_end_pfn;
7295 unsigned long spanned, absent;
7296 unsigned long size, real_size;
7298 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7303 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7308 real_size = size - absent;
7311 zone->zone_start_pfn = zone_start_pfn;
7313 zone->zone_start_pfn = 0;
7314 zone->spanned_pages = size;
7315 zone->present_pages = real_size;
7316 #if defined(CONFIG_MEMORY_HOTPLUG)
7317 zone->present_early_pages = real_size;
7321 realtotalpages += real_size;
7324 pgdat->node_spanned_pages = totalpages;
7325 pgdat->node_present_pages = realtotalpages;
7326 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7329 #ifndef CONFIG_SPARSEMEM
7331 * Calculate the size of the zone->blockflags rounded to an unsigned long
7332 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7333 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7334 * round what is now in bits to nearest long in bits, then return it in
7337 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7339 unsigned long usemapsize;
7341 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7342 usemapsize = roundup(zonesize, pageblock_nr_pages);
7343 usemapsize = usemapsize >> pageblock_order;
7344 usemapsize *= NR_PAGEBLOCK_BITS;
7345 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7347 return usemapsize / 8;
7350 static void __ref setup_usemap(struct zone *zone)
7352 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7353 zone->spanned_pages);
7354 zone->pageblock_flags = NULL;
7356 zone->pageblock_flags =
7357 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7359 if (!zone->pageblock_flags)
7360 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7361 usemapsize, zone->name, zone_to_nid(zone));
7365 static inline void setup_usemap(struct zone *zone) {}
7366 #endif /* CONFIG_SPARSEMEM */
7368 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7370 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7371 void __init set_pageblock_order(void)
7375 /* Check that pageblock_nr_pages has not already been setup */
7376 if (pageblock_order)
7379 if (HPAGE_SHIFT > PAGE_SHIFT)
7380 order = HUGETLB_PAGE_ORDER;
7382 order = MAX_ORDER - 1;
7385 * Assume the largest contiguous order of interest is a huge page.
7386 * This value may be variable depending on boot parameters on IA64 and
7389 pageblock_order = order;
7391 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7394 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7395 * is unused as pageblock_order is set at compile-time. See
7396 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7399 void __init set_pageblock_order(void)
7403 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7405 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7406 unsigned long present_pages)
7408 unsigned long pages = spanned_pages;
7411 * Provide a more accurate estimation if there are holes within
7412 * the zone and SPARSEMEM is in use. If there are holes within the
7413 * zone, each populated memory region may cost us one or two extra
7414 * memmap pages due to alignment because memmap pages for each
7415 * populated regions may not be naturally aligned on page boundary.
7416 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7418 if (spanned_pages > present_pages + (present_pages >> 4) &&
7419 IS_ENABLED(CONFIG_SPARSEMEM))
7420 pages = present_pages;
7422 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7425 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7426 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7428 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7430 spin_lock_init(&ds_queue->split_queue_lock);
7431 INIT_LIST_HEAD(&ds_queue->split_queue);
7432 ds_queue->split_queue_len = 0;
7435 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7438 #ifdef CONFIG_COMPACTION
7439 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7441 init_waitqueue_head(&pgdat->kcompactd_wait);
7444 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7447 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7449 pgdat_resize_init(pgdat);
7451 pgdat_init_split_queue(pgdat);
7452 pgdat_init_kcompactd(pgdat);
7454 init_waitqueue_head(&pgdat->kswapd_wait);
7455 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7457 pgdat_page_ext_init(pgdat);
7458 lruvec_init(&pgdat->__lruvec);
7461 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7462 unsigned long remaining_pages)
7464 atomic_long_set(&zone->managed_pages, remaining_pages);
7465 zone_set_nid(zone, nid);
7466 zone->name = zone_names[idx];
7467 zone->zone_pgdat = NODE_DATA(nid);
7468 spin_lock_init(&zone->lock);
7469 zone_seqlock_init(zone);
7470 zone_pcp_init(zone);
7474 * Set up the zone data structures
7475 * - init pgdat internals
7476 * - init all zones belonging to this node
7478 * NOTE: this function is only called during memory hotplug
7480 #ifdef CONFIG_MEMORY_HOTPLUG
7481 void __ref free_area_init_core_hotplug(int nid)
7484 pg_data_t *pgdat = NODE_DATA(nid);
7486 pgdat_init_internals(pgdat);
7487 for (z = 0; z < MAX_NR_ZONES; z++)
7488 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7493 * Set up the zone data structures:
7494 * - mark all pages reserved
7495 * - mark all memory queues empty
7496 * - clear the memory bitmaps
7498 * NOTE: pgdat should get zeroed by caller.
7499 * NOTE: this function is only called during early init.
7501 static void __init free_area_init_core(struct pglist_data *pgdat)
7504 int nid = pgdat->node_id;
7506 pgdat_init_internals(pgdat);
7507 pgdat->per_cpu_nodestats = &boot_nodestats;
7509 for (j = 0; j < MAX_NR_ZONES; j++) {
7510 struct zone *zone = pgdat->node_zones + j;
7511 unsigned long size, freesize, memmap_pages;
7513 size = zone->spanned_pages;
7514 freesize = zone->present_pages;
7517 * Adjust freesize so that it accounts for how much memory
7518 * is used by this zone for memmap. This affects the watermark
7519 * and per-cpu initialisations
7521 memmap_pages = calc_memmap_size(size, freesize);
7522 if (!is_highmem_idx(j)) {
7523 if (freesize >= memmap_pages) {
7524 freesize -= memmap_pages;
7526 pr_debug(" %s zone: %lu pages used for memmap\n",
7527 zone_names[j], memmap_pages);
7529 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7530 zone_names[j], memmap_pages, freesize);
7533 /* Account for reserved pages */
7534 if (j == 0 && freesize > dma_reserve) {
7535 freesize -= dma_reserve;
7536 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7539 if (!is_highmem_idx(j))
7540 nr_kernel_pages += freesize;
7541 /* Charge for highmem memmap if there are enough kernel pages */
7542 else if (nr_kernel_pages > memmap_pages * 2)
7543 nr_kernel_pages -= memmap_pages;
7544 nr_all_pages += freesize;
7547 * Set an approximate value for lowmem here, it will be adjusted
7548 * when the bootmem allocator frees pages into the buddy system.
7549 * And all highmem pages will be managed by the buddy system.
7551 zone_init_internals(zone, j, nid, freesize);
7556 set_pageblock_order();
7558 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7562 #ifdef CONFIG_FLATMEM
7563 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7565 unsigned long __maybe_unused start = 0;
7566 unsigned long __maybe_unused offset = 0;
7568 /* Skip empty nodes */
7569 if (!pgdat->node_spanned_pages)
7572 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7573 offset = pgdat->node_start_pfn - start;
7574 /* ia64 gets its own node_mem_map, before this, without bootmem */
7575 if (!pgdat->node_mem_map) {
7576 unsigned long size, end;
7580 * The zone's endpoints aren't required to be MAX_ORDER
7581 * aligned but the node_mem_map endpoints must be in order
7582 * for the buddy allocator to function correctly.
7584 end = pgdat_end_pfn(pgdat);
7585 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7586 size = (end - start) * sizeof(struct page);
7587 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7588 pgdat->node_id, false);
7590 panic("Failed to allocate %ld bytes for node %d memory map\n",
7591 size, pgdat->node_id);
7592 pgdat->node_mem_map = map + offset;
7594 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7595 __func__, pgdat->node_id, (unsigned long)pgdat,
7596 (unsigned long)pgdat->node_mem_map);
7599 * With no DISCONTIG, the global mem_map is just set as node 0's
7601 if (pgdat == NODE_DATA(0)) {
7602 mem_map = NODE_DATA(0)->node_mem_map;
7603 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7609 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7610 #endif /* CONFIG_FLATMEM */
7612 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7613 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7615 pgdat->first_deferred_pfn = ULONG_MAX;
7618 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7621 static void __init free_area_init_node(int nid)
7623 pg_data_t *pgdat = NODE_DATA(nid);
7624 unsigned long start_pfn = 0;
7625 unsigned long end_pfn = 0;
7627 /* pg_data_t should be reset to zero when it's allocated */
7628 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7630 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7632 pgdat->node_id = nid;
7633 pgdat->node_start_pfn = start_pfn;
7634 pgdat->per_cpu_nodestats = NULL;
7636 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7637 (u64)start_pfn << PAGE_SHIFT,
7638 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7639 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7641 alloc_node_mem_map(pgdat);
7642 pgdat_set_deferred_range(pgdat);
7644 free_area_init_core(pgdat);
7647 void __init free_area_init_memoryless_node(int nid)
7649 free_area_init_node(nid);
7652 #if MAX_NUMNODES > 1
7654 * Figure out the number of possible node ids.
7656 void __init setup_nr_node_ids(void)
7658 unsigned int highest;
7660 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7661 nr_node_ids = highest + 1;
7666 * node_map_pfn_alignment - determine the maximum internode alignment
7668 * This function should be called after node map is populated and sorted.
7669 * It calculates the maximum power of two alignment which can distinguish
7672 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7673 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7674 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7675 * shifted, 1GiB is enough and this function will indicate so.
7677 * This is used to test whether pfn -> nid mapping of the chosen memory
7678 * model has fine enough granularity to avoid incorrect mapping for the
7679 * populated node map.
7681 * Return: the determined alignment in pfn's. 0 if there is no alignment
7682 * requirement (single node).
7684 unsigned long __init node_map_pfn_alignment(void)
7686 unsigned long accl_mask = 0, last_end = 0;
7687 unsigned long start, end, mask;
7688 int last_nid = NUMA_NO_NODE;
7691 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7692 if (!start || last_nid < 0 || last_nid == nid) {
7699 * Start with a mask granular enough to pin-point to the
7700 * start pfn and tick off bits one-by-one until it becomes
7701 * too coarse to separate the current node from the last.
7703 mask = ~((1 << __ffs(start)) - 1);
7704 while (mask && last_end <= (start & (mask << 1)))
7707 /* accumulate all internode masks */
7711 /* convert mask to number of pages */
7712 return ~accl_mask + 1;
7716 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7718 * Return: the minimum PFN based on information provided via
7719 * memblock_set_node().
7721 unsigned long __init find_min_pfn_with_active_regions(void)
7723 return PHYS_PFN(memblock_start_of_DRAM());
7727 * early_calculate_totalpages()
7728 * Sum pages in active regions for movable zone.
7729 * Populate N_MEMORY for calculating usable_nodes.
7731 static unsigned long __init early_calculate_totalpages(void)
7733 unsigned long totalpages = 0;
7734 unsigned long start_pfn, end_pfn;
7737 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7738 unsigned long pages = end_pfn - start_pfn;
7740 totalpages += pages;
7742 node_set_state(nid, N_MEMORY);
7748 * Find the PFN the Movable zone begins in each node. Kernel memory
7749 * is spread evenly between nodes as long as the nodes have enough
7750 * memory. When they don't, some nodes will have more kernelcore than
7753 static void __init find_zone_movable_pfns_for_nodes(void)
7756 unsigned long usable_startpfn;
7757 unsigned long kernelcore_node, kernelcore_remaining;
7758 /* save the state before borrow the nodemask */
7759 nodemask_t saved_node_state = node_states[N_MEMORY];
7760 unsigned long totalpages = early_calculate_totalpages();
7761 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7762 struct memblock_region *r;
7764 /* Need to find movable_zone earlier when movable_node is specified. */
7765 find_usable_zone_for_movable();
7768 * If movable_node is specified, ignore kernelcore and movablecore
7771 if (movable_node_is_enabled()) {
7772 for_each_mem_region(r) {
7773 if (!memblock_is_hotpluggable(r))
7776 nid = memblock_get_region_node(r);
7778 usable_startpfn = PFN_DOWN(r->base);
7779 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7780 min(usable_startpfn, zone_movable_pfn[nid]) :
7788 * If kernelcore=mirror is specified, ignore movablecore option
7790 if (mirrored_kernelcore) {
7791 bool mem_below_4gb_not_mirrored = false;
7793 for_each_mem_region(r) {
7794 if (memblock_is_mirror(r))
7797 nid = memblock_get_region_node(r);
7799 usable_startpfn = memblock_region_memory_base_pfn(r);
7801 if (usable_startpfn < 0x100000) {
7802 mem_below_4gb_not_mirrored = true;
7806 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7807 min(usable_startpfn, zone_movable_pfn[nid]) :
7811 if (mem_below_4gb_not_mirrored)
7812 pr_warn("This configuration results in unmirrored kernel memory.\n");
7818 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7819 * amount of necessary memory.
7821 if (required_kernelcore_percent)
7822 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7824 if (required_movablecore_percent)
7825 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7829 * If movablecore= was specified, calculate what size of
7830 * kernelcore that corresponds so that memory usable for
7831 * any allocation type is evenly spread. If both kernelcore
7832 * and movablecore are specified, then the value of kernelcore
7833 * will be used for required_kernelcore if it's greater than
7834 * what movablecore would have allowed.
7836 if (required_movablecore) {
7837 unsigned long corepages;
7840 * Round-up so that ZONE_MOVABLE is at least as large as what
7841 * was requested by the user
7843 required_movablecore =
7844 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7845 required_movablecore = min(totalpages, required_movablecore);
7846 corepages = totalpages - required_movablecore;
7848 required_kernelcore = max(required_kernelcore, corepages);
7852 * If kernelcore was not specified or kernelcore size is larger
7853 * than totalpages, there is no ZONE_MOVABLE.
7855 if (!required_kernelcore || required_kernelcore >= totalpages)
7858 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7859 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7862 /* Spread kernelcore memory as evenly as possible throughout nodes */
7863 kernelcore_node = required_kernelcore / usable_nodes;
7864 for_each_node_state(nid, N_MEMORY) {
7865 unsigned long start_pfn, end_pfn;
7868 * Recalculate kernelcore_node if the division per node
7869 * now exceeds what is necessary to satisfy the requested
7870 * amount of memory for the kernel
7872 if (required_kernelcore < kernelcore_node)
7873 kernelcore_node = required_kernelcore / usable_nodes;
7876 * As the map is walked, we track how much memory is usable
7877 * by the kernel using kernelcore_remaining. When it is
7878 * 0, the rest of the node is usable by ZONE_MOVABLE
7880 kernelcore_remaining = kernelcore_node;
7882 /* Go through each range of PFNs within this node */
7883 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7884 unsigned long size_pages;
7886 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7887 if (start_pfn >= end_pfn)
7890 /* Account for what is only usable for kernelcore */
7891 if (start_pfn < usable_startpfn) {
7892 unsigned long kernel_pages;
7893 kernel_pages = min(end_pfn, usable_startpfn)
7896 kernelcore_remaining -= min(kernel_pages,
7897 kernelcore_remaining);
7898 required_kernelcore -= min(kernel_pages,
7899 required_kernelcore);
7901 /* Continue if range is now fully accounted */
7902 if (end_pfn <= usable_startpfn) {
7905 * Push zone_movable_pfn to the end so
7906 * that if we have to rebalance
7907 * kernelcore across nodes, we will
7908 * not double account here
7910 zone_movable_pfn[nid] = end_pfn;
7913 start_pfn = usable_startpfn;
7917 * The usable PFN range for ZONE_MOVABLE is from
7918 * start_pfn->end_pfn. Calculate size_pages as the
7919 * number of pages used as kernelcore
7921 size_pages = end_pfn - start_pfn;
7922 if (size_pages > kernelcore_remaining)
7923 size_pages = kernelcore_remaining;
7924 zone_movable_pfn[nid] = start_pfn + size_pages;
7927 * Some kernelcore has been met, update counts and
7928 * break if the kernelcore for this node has been
7931 required_kernelcore -= min(required_kernelcore,
7933 kernelcore_remaining -= size_pages;
7934 if (!kernelcore_remaining)
7940 * If there is still required_kernelcore, we do another pass with one
7941 * less node in the count. This will push zone_movable_pfn[nid] further
7942 * along on the nodes that still have memory until kernelcore is
7946 if (usable_nodes && required_kernelcore > usable_nodes)
7950 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7951 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7952 unsigned long start_pfn, end_pfn;
7954 zone_movable_pfn[nid] =
7955 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7957 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7958 if (zone_movable_pfn[nid] >= end_pfn)
7959 zone_movable_pfn[nid] = 0;
7963 /* restore the node_state */
7964 node_states[N_MEMORY] = saved_node_state;
7967 /* Any regular or high memory on that node ? */
7968 static void check_for_memory(pg_data_t *pgdat, int nid)
7970 enum zone_type zone_type;
7972 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7973 struct zone *zone = &pgdat->node_zones[zone_type];
7974 if (populated_zone(zone)) {
7975 if (IS_ENABLED(CONFIG_HIGHMEM))
7976 node_set_state(nid, N_HIGH_MEMORY);
7977 if (zone_type <= ZONE_NORMAL)
7978 node_set_state(nid, N_NORMAL_MEMORY);
7985 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7986 * such cases we allow max_zone_pfn sorted in the descending order
7988 bool __weak arch_has_descending_max_zone_pfns(void)
7994 * free_area_init - Initialise all pg_data_t and zone data
7995 * @max_zone_pfn: an array of max PFNs for each zone
7997 * This will call free_area_init_node() for each active node in the system.
7998 * Using the page ranges provided by memblock_set_node(), the size of each
7999 * zone in each node and their holes is calculated. If the maximum PFN
8000 * between two adjacent zones match, it is assumed that the zone is empty.
8001 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8002 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8003 * starts where the previous one ended. For example, ZONE_DMA32 starts
8004 * at arch_max_dma_pfn.
8006 void __init free_area_init(unsigned long *max_zone_pfn)
8008 unsigned long start_pfn, end_pfn;
8012 /* Record where the zone boundaries are */
8013 memset(arch_zone_lowest_possible_pfn, 0,
8014 sizeof(arch_zone_lowest_possible_pfn));
8015 memset(arch_zone_highest_possible_pfn, 0,
8016 sizeof(arch_zone_highest_possible_pfn));
8018 start_pfn = find_min_pfn_with_active_regions();
8019 descending = arch_has_descending_max_zone_pfns();
8021 for (i = 0; i < MAX_NR_ZONES; i++) {
8023 zone = MAX_NR_ZONES - i - 1;
8027 if (zone == ZONE_MOVABLE)
8030 end_pfn = max(max_zone_pfn[zone], start_pfn);
8031 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8032 arch_zone_highest_possible_pfn[zone] = end_pfn;
8034 start_pfn = end_pfn;
8037 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8038 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8039 find_zone_movable_pfns_for_nodes();
8041 /* Print out the zone ranges */
8042 pr_info("Zone ranges:\n");
8043 for (i = 0; i < MAX_NR_ZONES; i++) {
8044 if (i == ZONE_MOVABLE)
8046 pr_info(" %-8s ", zone_names[i]);
8047 if (arch_zone_lowest_possible_pfn[i] ==
8048 arch_zone_highest_possible_pfn[i])
8051 pr_cont("[mem %#018Lx-%#018Lx]\n",
8052 (u64)arch_zone_lowest_possible_pfn[i]
8054 ((u64)arch_zone_highest_possible_pfn[i]
8055 << PAGE_SHIFT) - 1);
8058 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8059 pr_info("Movable zone start for each node\n");
8060 for (i = 0; i < MAX_NUMNODES; i++) {
8061 if (zone_movable_pfn[i])
8062 pr_info(" Node %d: %#018Lx\n", i,
8063 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8067 * Print out the early node map, and initialize the
8068 * subsection-map relative to active online memory ranges to
8069 * enable future "sub-section" extensions of the memory map.
8071 pr_info("Early memory node ranges\n");
8072 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8073 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8074 (u64)start_pfn << PAGE_SHIFT,
8075 ((u64)end_pfn << PAGE_SHIFT) - 1);
8076 subsection_map_init(start_pfn, end_pfn - start_pfn);
8079 /* Initialise every node */
8080 mminit_verify_pageflags_layout();
8081 setup_nr_node_ids();
8082 for_each_online_node(nid) {
8083 pg_data_t *pgdat = NODE_DATA(nid);
8084 free_area_init_node(nid);
8086 /* Any memory on that node */
8087 if (pgdat->node_present_pages)
8088 node_set_state(nid, N_MEMORY);
8089 check_for_memory(pgdat, nid);
8095 static int __init cmdline_parse_core(char *p, unsigned long *core,
8096 unsigned long *percent)
8098 unsigned long long coremem;
8104 /* Value may be a percentage of total memory, otherwise bytes */
8105 coremem = simple_strtoull(p, &endptr, 0);
8106 if (*endptr == '%') {
8107 /* Paranoid check for percent values greater than 100 */
8108 WARN_ON(coremem > 100);
8112 coremem = memparse(p, &p);
8113 /* Paranoid check that UL is enough for the coremem value */
8114 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8116 *core = coremem >> PAGE_SHIFT;
8123 * kernelcore=size sets the amount of memory for use for allocations that
8124 * cannot be reclaimed or migrated.
8126 static int __init cmdline_parse_kernelcore(char *p)
8128 /* parse kernelcore=mirror */
8129 if (parse_option_str(p, "mirror")) {
8130 mirrored_kernelcore = true;
8134 return cmdline_parse_core(p, &required_kernelcore,
8135 &required_kernelcore_percent);
8139 * movablecore=size sets the amount of memory for use for allocations that
8140 * can be reclaimed or migrated.
8142 static int __init cmdline_parse_movablecore(char *p)
8144 return cmdline_parse_core(p, &required_movablecore,
8145 &required_movablecore_percent);
8148 early_param("kernelcore", cmdline_parse_kernelcore);
8149 early_param("movablecore", cmdline_parse_movablecore);
8151 void adjust_managed_page_count(struct page *page, long count)
8153 atomic_long_add(count, &page_zone(page)->managed_pages);
8154 totalram_pages_add(count);
8155 #ifdef CONFIG_HIGHMEM
8156 if (PageHighMem(page))
8157 totalhigh_pages_add(count);
8160 EXPORT_SYMBOL(adjust_managed_page_count);
8162 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8165 unsigned long pages = 0;
8167 start = (void *)PAGE_ALIGN((unsigned long)start);
8168 end = (void *)((unsigned long)end & PAGE_MASK);
8169 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8170 struct page *page = virt_to_page(pos);
8171 void *direct_map_addr;
8174 * 'direct_map_addr' might be different from 'pos'
8175 * because some architectures' virt_to_page()
8176 * work with aliases. Getting the direct map
8177 * address ensures that we get a _writeable_
8178 * alias for the memset().
8180 direct_map_addr = page_address(page);
8182 * Perform a kasan-unchecked memset() since this memory
8183 * has not been initialized.
8185 direct_map_addr = kasan_reset_tag(direct_map_addr);
8186 if ((unsigned int)poison <= 0xFF)
8187 memset(direct_map_addr, poison, PAGE_SIZE);
8189 free_reserved_page(page);
8193 pr_info("Freeing %s memory: %ldK\n",
8194 s, pages << (PAGE_SHIFT - 10));
8199 void __init mem_init_print_info(void)
8201 unsigned long physpages, codesize, datasize, rosize, bss_size;
8202 unsigned long init_code_size, init_data_size;
8204 physpages = get_num_physpages();
8205 codesize = _etext - _stext;
8206 datasize = _edata - _sdata;
8207 rosize = __end_rodata - __start_rodata;
8208 bss_size = __bss_stop - __bss_start;
8209 init_data_size = __init_end - __init_begin;
8210 init_code_size = _einittext - _sinittext;
8213 * Detect special cases and adjust section sizes accordingly:
8214 * 1) .init.* may be embedded into .data sections
8215 * 2) .init.text.* may be out of [__init_begin, __init_end],
8216 * please refer to arch/tile/kernel/vmlinux.lds.S.
8217 * 3) .rodata.* may be embedded into .text or .data sections.
8219 #define adj_init_size(start, end, size, pos, adj) \
8221 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8225 adj_init_size(__init_begin, __init_end, init_data_size,
8226 _sinittext, init_code_size);
8227 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8228 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8229 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8230 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8232 #undef adj_init_size
8234 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8235 #ifdef CONFIG_HIGHMEM
8239 nr_free_pages() << (PAGE_SHIFT - 10),
8240 physpages << (PAGE_SHIFT - 10),
8241 codesize >> 10, datasize >> 10, rosize >> 10,
8242 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8243 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
8244 totalcma_pages << (PAGE_SHIFT - 10)
8245 #ifdef CONFIG_HIGHMEM
8246 , totalhigh_pages() << (PAGE_SHIFT - 10)
8252 * set_dma_reserve - set the specified number of pages reserved in the first zone
8253 * @new_dma_reserve: The number of pages to mark reserved
8255 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8256 * In the DMA zone, a significant percentage may be consumed by kernel image
8257 * and other unfreeable allocations which can skew the watermarks badly. This
8258 * function may optionally be used to account for unfreeable pages in the
8259 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8260 * smaller per-cpu batchsize.
8262 void __init set_dma_reserve(unsigned long new_dma_reserve)
8264 dma_reserve = new_dma_reserve;
8267 static int page_alloc_cpu_dead(unsigned int cpu)
8271 lru_add_drain_cpu(cpu);
8275 * Spill the event counters of the dead processor
8276 * into the current processors event counters.
8277 * This artificially elevates the count of the current
8280 vm_events_fold_cpu(cpu);
8283 * Zero the differential counters of the dead processor
8284 * so that the vm statistics are consistent.
8286 * This is only okay since the processor is dead and cannot
8287 * race with what we are doing.
8289 cpu_vm_stats_fold(cpu);
8291 for_each_populated_zone(zone)
8292 zone_pcp_update(zone, 0);
8297 static int page_alloc_cpu_online(unsigned int cpu)
8301 for_each_populated_zone(zone)
8302 zone_pcp_update(zone, 1);
8307 int hashdist = HASHDIST_DEFAULT;
8309 static int __init set_hashdist(char *str)
8313 hashdist = simple_strtoul(str, &str, 0);
8316 __setup("hashdist=", set_hashdist);
8319 void __init page_alloc_init(void)
8324 if (num_node_state(N_MEMORY) == 1)
8328 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8329 "mm/page_alloc:pcp",
8330 page_alloc_cpu_online,
8331 page_alloc_cpu_dead);
8336 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8337 * or min_free_kbytes changes.
8339 static void calculate_totalreserve_pages(void)
8341 struct pglist_data *pgdat;
8342 unsigned long reserve_pages = 0;
8343 enum zone_type i, j;
8345 for_each_online_pgdat(pgdat) {
8347 pgdat->totalreserve_pages = 0;
8349 for (i = 0; i < MAX_NR_ZONES; i++) {
8350 struct zone *zone = pgdat->node_zones + i;
8352 unsigned long managed_pages = zone_managed_pages(zone);
8354 /* Find valid and maximum lowmem_reserve in the zone */
8355 for (j = i; j < MAX_NR_ZONES; j++) {
8356 if (zone->lowmem_reserve[j] > max)
8357 max = zone->lowmem_reserve[j];
8360 /* we treat the high watermark as reserved pages. */
8361 max += high_wmark_pages(zone);
8363 if (max > managed_pages)
8364 max = managed_pages;
8366 pgdat->totalreserve_pages += max;
8368 reserve_pages += max;
8371 totalreserve_pages = reserve_pages;
8375 * setup_per_zone_lowmem_reserve - called whenever
8376 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8377 * has a correct pages reserved value, so an adequate number of
8378 * pages are left in the zone after a successful __alloc_pages().
8380 static void setup_per_zone_lowmem_reserve(void)
8382 struct pglist_data *pgdat;
8383 enum zone_type i, j;
8385 for_each_online_pgdat(pgdat) {
8386 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8387 struct zone *zone = &pgdat->node_zones[i];
8388 int ratio = sysctl_lowmem_reserve_ratio[i];
8389 bool clear = !ratio || !zone_managed_pages(zone);
8390 unsigned long managed_pages = 0;
8392 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8393 struct zone *upper_zone = &pgdat->node_zones[j];
8395 managed_pages += zone_managed_pages(upper_zone);
8398 zone->lowmem_reserve[j] = 0;
8400 zone->lowmem_reserve[j] = managed_pages / ratio;
8405 /* update totalreserve_pages */
8406 calculate_totalreserve_pages();
8409 static void __setup_per_zone_wmarks(void)
8411 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8412 unsigned long lowmem_pages = 0;
8414 unsigned long flags;
8416 /* Calculate total number of !ZONE_HIGHMEM pages */
8417 for_each_zone(zone) {
8418 if (!is_highmem(zone))
8419 lowmem_pages += zone_managed_pages(zone);
8422 for_each_zone(zone) {
8425 spin_lock_irqsave(&zone->lock, flags);
8426 tmp = (u64)pages_min * zone_managed_pages(zone);
8427 do_div(tmp, lowmem_pages);
8428 if (is_highmem(zone)) {
8430 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8431 * need highmem pages, so cap pages_min to a small
8434 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8435 * deltas control async page reclaim, and so should
8436 * not be capped for highmem.
8438 unsigned long min_pages;
8440 min_pages = zone_managed_pages(zone) / 1024;
8441 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8442 zone->_watermark[WMARK_MIN] = min_pages;
8445 * If it's a lowmem zone, reserve a number of pages
8446 * proportionate to the zone's size.
8448 zone->_watermark[WMARK_MIN] = tmp;
8452 * Set the kswapd watermarks distance according to the
8453 * scale factor in proportion to available memory, but
8454 * ensure a minimum size on small systems.
8456 tmp = max_t(u64, tmp >> 2,
8457 mult_frac(zone_managed_pages(zone),
8458 watermark_scale_factor, 10000));
8460 zone->watermark_boost = 0;
8461 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8462 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8464 spin_unlock_irqrestore(&zone->lock, flags);
8467 /* update totalreserve_pages */
8468 calculate_totalreserve_pages();
8472 * setup_per_zone_wmarks - called when min_free_kbytes changes
8473 * or when memory is hot-{added|removed}
8475 * Ensures that the watermark[min,low,high] values for each zone are set
8476 * correctly with respect to min_free_kbytes.
8478 void setup_per_zone_wmarks(void)
8481 static DEFINE_SPINLOCK(lock);
8484 __setup_per_zone_wmarks();
8488 * The watermark size have changed so update the pcpu batch
8489 * and high limits or the limits may be inappropriate.
8492 zone_pcp_update(zone, 0);
8496 * Initialise min_free_kbytes.
8498 * For small machines we want it small (128k min). For large machines
8499 * we want it large (256MB max). But it is not linear, because network
8500 * bandwidth does not increase linearly with machine size. We use
8502 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8503 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8519 int __meminit init_per_zone_wmark_min(void)
8521 unsigned long lowmem_kbytes;
8522 int new_min_free_kbytes;
8524 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8525 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8527 if (new_min_free_kbytes > user_min_free_kbytes) {
8528 min_free_kbytes = new_min_free_kbytes;
8529 if (min_free_kbytes < 128)
8530 min_free_kbytes = 128;
8531 if (min_free_kbytes > 262144)
8532 min_free_kbytes = 262144;
8534 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8535 new_min_free_kbytes, user_min_free_kbytes);
8537 setup_per_zone_wmarks();
8538 refresh_zone_stat_thresholds();
8539 setup_per_zone_lowmem_reserve();
8542 setup_min_unmapped_ratio();
8543 setup_min_slab_ratio();
8546 khugepaged_min_free_kbytes_update();
8550 postcore_initcall(init_per_zone_wmark_min)
8553 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8554 * that we can call two helper functions whenever min_free_kbytes
8557 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8558 void *buffer, size_t *length, loff_t *ppos)
8562 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8567 user_min_free_kbytes = min_free_kbytes;
8568 setup_per_zone_wmarks();
8573 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8574 void *buffer, size_t *length, loff_t *ppos)
8578 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8583 setup_per_zone_wmarks();
8589 static void setup_min_unmapped_ratio(void)
8594 for_each_online_pgdat(pgdat)
8595 pgdat->min_unmapped_pages = 0;
8598 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8599 sysctl_min_unmapped_ratio) / 100;
8603 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8604 void *buffer, size_t *length, loff_t *ppos)
8608 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8612 setup_min_unmapped_ratio();
8617 static void setup_min_slab_ratio(void)
8622 for_each_online_pgdat(pgdat)
8623 pgdat->min_slab_pages = 0;
8626 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8627 sysctl_min_slab_ratio) / 100;
8630 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8631 void *buffer, size_t *length, loff_t *ppos)
8635 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8639 setup_min_slab_ratio();
8646 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8647 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8648 * whenever sysctl_lowmem_reserve_ratio changes.
8650 * The reserve ratio obviously has absolutely no relation with the
8651 * minimum watermarks. The lowmem reserve ratio can only make sense
8652 * if in function of the boot time zone sizes.
8654 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8655 void *buffer, size_t *length, loff_t *ppos)
8659 proc_dointvec_minmax(table, write, buffer, length, ppos);
8661 for (i = 0; i < MAX_NR_ZONES; i++) {
8662 if (sysctl_lowmem_reserve_ratio[i] < 1)
8663 sysctl_lowmem_reserve_ratio[i] = 0;
8666 setup_per_zone_lowmem_reserve();
8671 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8672 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8673 * pagelist can have before it gets flushed back to buddy allocator.
8675 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8676 int write, void *buffer, size_t *length, loff_t *ppos)
8679 int old_percpu_pagelist_high_fraction;
8682 mutex_lock(&pcp_batch_high_lock);
8683 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8685 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8686 if (!write || ret < 0)
8689 /* Sanity checking to avoid pcp imbalance */
8690 if (percpu_pagelist_high_fraction &&
8691 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8692 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8698 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8701 for_each_populated_zone(zone)
8702 zone_set_pageset_high_and_batch(zone, 0);
8704 mutex_unlock(&pcp_batch_high_lock);
8708 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8710 * Returns the number of pages that arch has reserved but
8711 * is not known to alloc_large_system_hash().
8713 static unsigned long __init arch_reserved_kernel_pages(void)
8720 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8721 * machines. As memory size is increased the scale is also increased but at
8722 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8723 * quadruples the scale is increased by one, which means the size of hash table
8724 * only doubles, instead of quadrupling as well.
8725 * Because 32-bit systems cannot have large physical memory, where this scaling
8726 * makes sense, it is disabled on such platforms.
8728 #if __BITS_PER_LONG > 32
8729 #define ADAPT_SCALE_BASE (64ul << 30)
8730 #define ADAPT_SCALE_SHIFT 2
8731 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8735 * allocate a large system hash table from bootmem
8736 * - it is assumed that the hash table must contain an exact power-of-2
8737 * quantity of entries
8738 * - limit is the number of hash buckets, not the total allocation size
8740 void *__init alloc_large_system_hash(const char *tablename,
8741 unsigned long bucketsize,
8742 unsigned long numentries,
8745 unsigned int *_hash_shift,
8746 unsigned int *_hash_mask,
8747 unsigned long low_limit,
8748 unsigned long high_limit)
8750 unsigned long long max = high_limit;
8751 unsigned long log2qty, size;
8757 /* allow the kernel cmdline to have a say */
8759 /* round applicable memory size up to nearest megabyte */
8760 numentries = nr_kernel_pages;
8761 numentries -= arch_reserved_kernel_pages();
8763 /* It isn't necessary when PAGE_SIZE >= 1MB */
8764 if (PAGE_SHIFT < 20)
8765 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8767 #if __BITS_PER_LONG > 32
8769 unsigned long adapt;
8771 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8772 adapt <<= ADAPT_SCALE_SHIFT)
8777 /* limit to 1 bucket per 2^scale bytes of low memory */
8778 if (scale > PAGE_SHIFT)
8779 numentries >>= (scale - PAGE_SHIFT);
8781 numentries <<= (PAGE_SHIFT - scale);
8783 /* Make sure we've got at least a 0-order allocation.. */
8784 if (unlikely(flags & HASH_SMALL)) {
8785 /* Makes no sense without HASH_EARLY */
8786 WARN_ON(!(flags & HASH_EARLY));
8787 if (!(numentries >> *_hash_shift)) {
8788 numentries = 1UL << *_hash_shift;
8789 BUG_ON(!numentries);
8791 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8792 numentries = PAGE_SIZE / bucketsize;
8794 numentries = roundup_pow_of_two(numentries);
8796 /* limit allocation size to 1/16 total memory by default */
8798 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8799 do_div(max, bucketsize);
8801 max = min(max, 0x80000000ULL);
8803 if (numentries < low_limit)
8804 numentries = low_limit;
8805 if (numentries > max)
8808 log2qty = ilog2(numentries);
8810 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8813 size = bucketsize << log2qty;
8814 if (flags & HASH_EARLY) {
8815 if (flags & HASH_ZERO)
8816 table = memblock_alloc(size, SMP_CACHE_BYTES);
8818 table = memblock_alloc_raw(size,
8820 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8821 table = __vmalloc(size, gfp_flags);
8823 huge = is_vm_area_hugepages(table);
8826 * If bucketsize is not a power-of-two, we may free
8827 * some pages at the end of hash table which
8828 * alloc_pages_exact() automatically does
8830 table = alloc_pages_exact(size, gfp_flags);
8831 kmemleak_alloc(table, size, 1, gfp_flags);
8833 } while (!table && size > PAGE_SIZE && --log2qty);
8836 panic("Failed to allocate %s hash table\n", tablename);
8838 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8839 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8840 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8843 *_hash_shift = log2qty;
8845 *_hash_mask = (1 << log2qty) - 1;
8851 * This function checks whether pageblock includes unmovable pages or not.
8853 * PageLRU check without isolation or lru_lock could race so that
8854 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8855 * check without lock_page also may miss some movable non-lru pages at
8856 * race condition. So you can't expect this function should be exact.
8858 * Returns a page without holding a reference. If the caller wants to
8859 * dereference that page (e.g., dumping), it has to make sure that it
8860 * cannot get removed (e.g., via memory unplug) concurrently.
8863 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8864 int migratetype, int flags)
8866 unsigned long iter = 0;
8867 unsigned long pfn = page_to_pfn(page);
8868 unsigned long offset = pfn % pageblock_nr_pages;
8870 if (is_migrate_cma_page(page)) {
8872 * CMA allocations (alloc_contig_range) really need to mark
8873 * isolate CMA pageblocks even when they are not movable in fact
8874 * so consider them movable here.
8876 if (is_migrate_cma(migratetype))
8882 for (; iter < pageblock_nr_pages - offset; iter++) {
8883 page = pfn_to_page(pfn + iter);
8886 * Both, bootmem allocations and memory holes are marked
8887 * PG_reserved and are unmovable. We can even have unmovable
8888 * allocations inside ZONE_MOVABLE, for example when
8889 * specifying "movablecore".
8891 if (PageReserved(page))
8895 * If the zone is movable and we have ruled out all reserved
8896 * pages then it should be reasonably safe to assume the rest
8899 if (zone_idx(zone) == ZONE_MOVABLE)
8903 * Hugepages are not in LRU lists, but they're movable.
8904 * THPs are on the LRU, but need to be counted as #small pages.
8905 * We need not scan over tail pages because we don't
8906 * handle each tail page individually in migration.
8908 if (PageHuge(page) || PageTransCompound(page)) {
8909 struct page *head = compound_head(page);
8910 unsigned int skip_pages;
8912 if (PageHuge(page)) {
8913 if (!hugepage_migration_supported(page_hstate(head)))
8915 } else if (!PageLRU(head) && !__PageMovable(head)) {
8919 skip_pages = compound_nr(head) - (page - head);
8920 iter += skip_pages - 1;
8925 * We can't use page_count without pin a page
8926 * because another CPU can free compound page.
8927 * This check already skips compound tails of THP
8928 * because their page->_refcount is zero at all time.
8930 if (!page_ref_count(page)) {
8931 if (PageBuddy(page))
8932 iter += (1 << buddy_order(page)) - 1;
8937 * The HWPoisoned page may be not in buddy system, and
8938 * page_count() is not 0.
8940 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8944 * We treat all PageOffline() pages as movable when offlining
8945 * to give drivers a chance to decrement their reference count
8946 * in MEM_GOING_OFFLINE in order to indicate that these pages
8947 * can be offlined as there are no direct references anymore.
8948 * For actually unmovable PageOffline() where the driver does
8949 * not support this, we will fail later when trying to actually
8950 * move these pages that still have a reference count > 0.
8951 * (false negatives in this function only)
8953 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8956 if (__PageMovable(page) || PageLRU(page))
8960 * If there are RECLAIMABLE pages, we need to check
8961 * it. But now, memory offline itself doesn't call
8962 * shrink_node_slabs() and it still to be fixed.
8969 #ifdef CONFIG_CONTIG_ALLOC
8970 static unsigned long pfn_max_align_down(unsigned long pfn)
8972 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8973 pageblock_nr_pages) - 1);
8976 static unsigned long pfn_max_align_up(unsigned long pfn)
8978 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8979 pageblock_nr_pages));
8982 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8983 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8984 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8985 static void alloc_contig_dump_pages(struct list_head *page_list)
8987 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8989 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8993 list_for_each_entry(page, page_list, lru)
8994 dump_page(page, "migration failure");
8998 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9003 /* [start, end) must belong to a single zone. */
9004 static int __alloc_contig_migrate_range(struct compact_control *cc,
9005 unsigned long start, unsigned long end)
9007 /* This function is based on compact_zone() from compaction.c. */
9008 unsigned int nr_reclaimed;
9009 unsigned long pfn = start;
9010 unsigned int tries = 0;
9012 struct migration_target_control mtc = {
9013 .nid = zone_to_nid(cc->zone),
9014 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9017 lru_cache_disable();
9019 while (pfn < end || !list_empty(&cc->migratepages)) {
9020 if (fatal_signal_pending(current)) {
9025 if (list_empty(&cc->migratepages)) {
9026 cc->nr_migratepages = 0;
9027 ret = isolate_migratepages_range(cc, pfn, end);
9028 if (ret && ret != -EAGAIN)
9030 pfn = cc->migrate_pfn;
9032 } else if (++tries == 5) {
9037 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9039 cc->nr_migratepages -= nr_reclaimed;
9041 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9042 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9045 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9046 * to retry again over this error, so do the same here.
9055 alloc_contig_dump_pages(&cc->migratepages);
9056 putback_movable_pages(&cc->migratepages);
9063 * alloc_contig_range() -- tries to allocate given range of pages
9064 * @start: start PFN to allocate
9065 * @end: one-past-the-last PFN to allocate
9066 * @migratetype: migratetype of the underlying pageblocks (either
9067 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9068 * in range must have the same migratetype and it must
9069 * be either of the two.
9070 * @gfp_mask: GFP mask to use during compaction
9072 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9073 * aligned. The PFN range must belong to a single zone.
9075 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9076 * pageblocks in the range. Once isolated, the pageblocks should not
9077 * be modified by others.
9079 * Return: zero on success or negative error code. On success all
9080 * pages which PFN is in [start, end) are allocated for the caller and
9081 * need to be freed with free_contig_range().
9083 int alloc_contig_range(unsigned long start, unsigned long end,
9084 unsigned migratetype, gfp_t gfp_mask)
9086 unsigned long outer_start, outer_end;
9090 struct compact_control cc = {
9091 .nr_migratepages = 0,
9093 .zone = page_zone(pfn_to_page(start)),
9094 .mode = MIGRATE_SYNC,
9095 .ignore_skip_hint = true,
9096 .no_set_skip_hint = true,
9097 .gfp_mask = current_gfp_context(gfp_mask),
9098 .alloc_contig = true,
9100 INIT_LIST_HEAD(&cc.migratepages);
9103 * What we do here is we mark all pageblocks in range as
9104 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9105 * have different sizes, and due to the way page allocator
9106 * work, we align the range to biggest of the two pages so
9107 * that page allocator won't try to merge buddies from
9108 * different pageblocks and change MIGRATE_ISOLATE to some
9109 * other migration type.
9111 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9112 * migrate the pages from an unaligned range (ie. pages that
9113 * we are interested in). This will put all the pages in
9114 * range back to page allocator as MIGRATE_ISOLATE.
9116 * When this is done, we take the pages in range from page
9117 * allocator removing them from the buddy system. This way
9118 * page allocator will never consider using them.
9120 * This lets us mark the pageblocks back as
9121 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9122 * aligned range but not in the unaligned, original range are
9123 * put back to page allocator so that buddy can use them.
9126 ret = start_isolate_page_range(pfn_max_align_down(start),
9127 pfn_max_align_up(end), migratetype, 0);
9131 drain_all_pages(cc.zone);
9134 * In case of -EBUSY, we'd like to know which page causes problem.
9135 * So, just fall through. test_pages_isolated() has a tracepoint
9136 * which will report the busy page.
9138 * It is possible that busy pages could become available before
9139 * the call to test_pages_isolated, and the range will actually be
9140 * allocated. So, if we fall through be sure to clear ret so that
9141 * -EBUSY is not accidentally used or returned to caller.
9143 ret = __alloc_contig_migrate_range(&cc, start, end);
9144 if (ret && ret != -EBUSY)
9149 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9150 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9151 * more, all pages in [start, end) are free in page allocator.
9152 * What we are going to do is to allocate all pages from
9153 * [start, end) (that is remove them from page allocator).
9155 * The only problem is that pages at the beginning and at the
9156 * end of interesting range may be not aligned with pages that
9157 * page allocator holds, ie. they can be part of higher order
9158 * pages. Because of this, we reserve the bigger range and
9159 * once this is done free the pages we are not interested in.
9161 * We don't have to hold zone->lock here because the pages are
9162 * isolated thus they won't get removed from buddy.
9166 outer_start = start;
9167 while (!PageBuddy(pfn_to_page(outer_start))) {
9168 if (++order >= MAX_ORDER) {
9169 outer_start = start;
9172 outer_start &= ~0UL << order;
9175 if (outer_start != start) {
9176 order = buddy_order(pfn_to_page(outer_start));
9179 * outer_start page could be small order buddy page and
9180 * it doesn't include start page. Adjust outer_start
9181 * in this case to report failed page properly
9182 * on tracepoint in test_pages_isolated()
9184 if (outer_start + (1UL << order) <= start)
9185 outer_start = start;
9188 /* Make sure the range is really isolated. */
9189 if (test_pages_isolated(outer_start, end, 0)) {
9194 /* Grab isolated pages from freelists. */
9195 outer_end = isolate_freepages_range(&cc, outer_start, end);
9201 /* Free head and tail (if any) */
9202 if (start != outer_start)
9203 free_contig_range(outer_start, start - outer_start);
9204 if (end != outer_end)
9205 free_contig_range(end, outer_end - end);
9208 undo_isolate_page_range(pfn_max_align_down(start),
9209 pfn_max_align_up(end), migratetype);
9212 EXPORT_SYMBOL(alloc_contig_range);
9214 static int __alloc_contig_pages(unsigned long start_pfn,
9215 unsigned long nr_pages, gfp_t gfp_mask)
9217 unsigned long end_pfn = start_pfn + nr_pages;
9219 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9223 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9224 unsigned long nr_pages)
9226 unsigned long i, end_pfn = start_pfn + nr_pages;
9229 for (i = start_pfn; i < end_pfn; i++) {
9230 page = pfn_to_online_page(i);
9234 if (page_zone(page) != z)
9237 if (PageReserved(page))
9243 static bool zone_spans_last_pfn(const struct zone *zone,
9244 unsigned long start_pfn, unsigned long nr_pages)
9246 unsigned long last_pfn = start_pfn + nr_pages - 1;
9248 return zone_spans_pfn(zone, last_pfn);
9252 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9253 * @nr_pages: Number of contiguous pages to allocate
9254 * @gfp_mask: GFP mask to limit search and used during compaction
9256 * @nodemask: Mask for other possible nodes
9258 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9259 * on an applicable zonelist to find a contiguous pfn range which can then be
9260 * tried for allocation with alloc_contig_range(). This routine is intended
9261 * for allocation requests which can not be fulfilled with the buddy allocator.
9263 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9264 * power of two then the alignment is guaranteed to be to the given nr_pages
9265 * (e.g. 1GB request would be aligned to 1GB).
9267 * Allocated pages can be freed with free_contig_range() or by manually calling
9268 * __free_page() on each allocated page.
9270 * Return: pointer to contiguous pages on success, or NULL if not successful.
9272 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9273 int nid, nodemask_t *nodemask)
9275 unsigned long ret, pfn, flags;
9276 struct zonelist *zonelist;
9280 zonelist = node_zonelist(nid, gfp_mask);
9281 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9282 gfp_zone(gfp_mask), nodemask) {
9283 spin_lock_irqsave(&zone->lock, flags);
9285 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9286 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9287 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9289 * We release the zone lock here because
9290 * alloc_contig_range() will also lock the zone
9291 * at some point. If there's an allocation
9292 * spinning on this lock, it may win the race
9293 * and cause alloc_contig_range() to fail...
9295 spin_unlock_irqrestore(&zone->lock, flags);
9296 ret = __alloc_contig_pages(pfn, nr_pages,
9299 return pfn_to_page(pfn);
9300 spin_lock_irqsave(&zone->lock, flags);
9304 spin_unlock_irqrestore(&zone->lock, flags);
9308 #endif /* CONFIG_CONTIG_ALLOC */
9310 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9312 unsigned long count = 0;
9314 for (; nr_pages--; pfn++) {
9315 struct page *page = pfn_to_page(pfn);
9317 count += page_count(page) != 1;
9320 WARN(count != 0, "%lu pages are still in use!\n", count);
9322 EXPORT_SYMBOL(free_contig_range);
9325 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9326 * page high values need to be recalculated.
9328 void zone_pcp_update(struct zone *zone, int cpu_online)
9330 mutex_lock(&pcp_batch_high_lock);
9331 zone_set_pageset_high_and_batch(zone, cpu_online);
9332 mutex_unlock(&pcp_batch_high_lock);
9336 * Effectively disable pcplists for the zone by setting the high limit to 0
9337 * and draining all cpus. A concurrent page freeing on another CPU that's about
9338 * to put the page on pcplist will either finish before the drain and the page
9339 * will be drained, or observe the new high limit and skip the pcplist.
9341 * Must be paired with a call to zone_pcp_enable().
9343 void zone_pcp_disable(struct zone *zone)
9345 mutex_lock(&pcp_batch_high_lock);
9346 __zone_set_pageset_high_and_batch(zone, 0, 1);
9347 __drain_all_pages(zone, true);
9350 void zone_pcp_enable(struct zone *zone)
9352 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9353 mutex_unlock(&pcp_batch_high_lock);
9356 void zone_pcp_reset(struct zone *zone)
9359 struct per_cpu_zonestat *pzstats;
9361 if (zone->per_cpu_pageset != &boot_pageset) {
9362 for_each_online_cpu(cpu) {
9363 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9364 drain_zonestat(zone, pzstats);
9366 free_percpu(zone->per_cpu_pageset);
9367 free_percpu(zone->per_cpu_zonestats);
9368 zone->per_cpu_pageset = &boot_pageset;
9369 zone->per_cpu_zonestats = &boot_zonestats;
9373 #ifdef CONFIG_MEMORY_HOTREMOVE
9375 * All pages in the range must be in a single zone, must not contain holes,
9376 * must span full sections, and must be isolated before calling this function.
9378 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9380 unsigned long pfn = start_pfn;
9384 unsigned long flags;
9386 offline_mem_sections(pfn, end_pfn);
9387 zone = page_zone(pfn_to_page(pfn));
9388 spin_lock_irqsave(&zone->lock, flags);
9389 while (pfn < end_pfn) {
9390 page = pfn_to_page(pfn);
9392 * The HWPoisoned page may be not in buddy system, and
9393 * page_count() is not 0.
9395 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9400 * At this point all remaining PageOffline() pages have a
9401 * reference count of 0 and can simply be skipped.
9403 if (PageOffline(page)) {
9404 BUG_ON(page_count(page));
9405 BUG_ON(PageBuddy(page));
9410 BUG_ON(page_count(page));
9411 BUG_ON(!PageBuddy(page));
9412 order = buddy_order(page);
9413 del_page_from_free_list(page, zone, order);
9414 pfn += (1 << order);
9416 spin_unlock_irqrestore(&zone->lock, flags);
9420 bool is_free_buddy_page(struct page *page)
9422 struct zone *zone = page_zone(page);
9423 unsigned long pfn = page_to_pfn(page);
9424 unsigned long flags;
9427 spin_lock_irqsave(&zone->lock, flags);
9428 for (order = 0; order < MAX_ORDER; order++) {
9429 struct page *page_head = page - (pfn & ((1 << order) - 1));
9431 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
9434 spin_unlock_irqrestore(&zone->lock, flags);
9436 return order < MAX_ORDER;
9439 #ifdef CONFIG_MEMORY_FAILURE
9441 * Break down a higher-order page in sub-pages, and keep our target out of
9444 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9445 struct page *target, int low, int high,
9448 unsigned long size = 1 << high;
9449 struct page *current_buddy, *next_page;
9451 while (high > low) {
9455 if (target >= &page[size]) {
9456 next_page = page + size;
9457 current_buddy = page;
9460 current_buddy = page + size;
9463 if (set_page_guard(zone, current_buddy, high, migratetype))
9466 if (current_buddy != target) {
9467 add_to_free_list(current_buddy, zone, high, migratetype);
9468 set_buddy_order(current_buddy, high);
9475 * Take a page that will be marked as poisoned off the buddy allocator.
9477 bool take_page_off_buddy(struct page *page)
9479 struct zone *zone = page_zone(page);
9480 unsigned long pfn = page_to_pfn(page);
9481 unsigned long flags;
9485 spin_lock_irqsave(&zone->lock, flags);
9486 for (order = 0; order < MAX_ORDER; order++) {
9487 struct page *page_head = page - (pfn & ((1 << order) - 1));
9488 int page_order = buddy_order(page_head);
9490 if (PageBuddy(page_head) && page_order >= order) {
9491 unsigned long pfn_head = page_to_pfn(page_head);
9492 int migratetype = get_pfnblock_migratetype(page_head,
9495 del_page_from_free_list(page_head, zone, page_order);
9496 break_down_buddy_pages(zone, page_head, page, 0,
9497 page_order, migratetype);
9498 if (!is_migrate_isolate(migratetype))
9499 __mod_zone_freepage_state(zone, -1, migratetype);
9503 if (page_count(page_head) > 0)
9506 spin_unlock_irqrestore(&zone->lock, flags);
9511 #ifdef CONFIG_ZONE_DMA
9512 bool has_managed_dma(void)
9514 struct pglist_data *pgdat;
9516 for_each_online_pgdat(pgdat) {
9517 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9519 if (managed_zone(zone))
9524 #endif /* CONFIG_ZONE_DMA */