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/swapops.h>
23 #include <linux/interrupt.h>
24 #include <linux/pagemap.h>
25 #include <linux/jiffies.h>
26 #include <linux/memblock.h>
27 #include <linux/compiler.h>
28 #include <linux/kernel.h>
29 #include <linux/kasan.h>
30 #include <linux/module.h>
31 #include <linux/suspend.h>
32 #include <linux/pagevec.h>
33 #include <linux/blkdev.h>
34 #include <linux/slab.h>
35 #include <linux/ratelimit.h>
36 #include <linux/oom.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/random.h>
49 #include <linux/sort.h>
50 #include <linux/pfn.h>
51 #include <linux/backing-dev.h>
52 #include <linux/fault-inject.h>
53 #include <linux/page-isolation.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/migrate.h>
63 #include <linux/hugetlb.h>
64 #include <linux/sched/rt.h>
65 #include <linux/sched/mm.h>
66 #include <linux/page_owner.h>
67 #include <linux/page_table_check.h>
68 #include <linux/kthread.h>
69 #include <linux/memcontrol.h>
70 #include <linux/ftrace.h>
71 #include <linux/lockdep.h>
72 #include <linux/nmi.h>
73 #include <linux/psi.h>
74 #include <linux/padata.h>
75 #include <linux/khugepaged.h>
76 #include <linux/buffer_head.h>
77 #include <linux/delayacct.h>
78 #include <asm/sections.h>
79 #include <asm/tlbflush.h>
80 #include <asm/div64.h>
83 #include "page_reporting.h"
85 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
86 typedef int __bitwise fpi_t;
88 /* No special request */
89 #define FPI_NONE ((__force fpi_t)0)
92 * Skip free page reporting notification for the (possibly merged) page.
93 * This does not hinder free page reporting from grabbing the page,
94 * reporting it and marking it "reported" - it only skips notifying
95 * the free page reporting infrastructure about a newly freed page. For
96 * example, used when temporarily pulling a page from a freelist and
97 * putting it back unmodified.
99 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
102 * Place the (possibly merged) page to the tail of the freelist. Will ignore
103 * page shuffling (relevant code - e.g., memory onlining - is expected to
104 * shuffle the whole zone).
106 * Note: No code should rely on this flag for correctness - it's purely
107 * to allow for optimizations when handing back either fresh pages
108 * (memory onlining) or untouched pages (page isolation, free page
111 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
114 * Don't poison memory with KASAN (only for the tag-based modes).
115 * During boot, all non-reserved memblock memory is exposed to page_alloc.
116 * Poisoning all that memory lengthens boot time, especially on systems with
117 * large amount of RAM. This flag is used to skip that poisoning.
118 * This is only done for the tag-based KASAN modes, as those are able to
119 * detect memory corruptions with the memory tags assigned by default.
120 * All memory allocated normally after boot gets poisoned as usual.
122 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
124 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
125 static DEFINE_MUTEX(pcp_batch_high_lock);
126 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
131 static DEFINE_PER_CPU(struct pagesets, pagesets) __maybe_unused = {
132 .lock = INIT_LOCAL_LOCK(lock),
135 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
136 DEFINE_PER_CPU(int, numa_node);
137 EXPORT_PER_CPU_SYMBOL(numa_node);
140 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
142 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
144 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
145 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
146 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
147 * defined in <linux/topology.h>.
149 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
150 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
153 /* work_structs for global per-cpu drains */
156 struct work_struct work;
158 static DEFINE_MUTEX(pcpu_drain_mutex);
159 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
161 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
162 volatile unsigned long latent_entropy __latent_entropy;
163 EXPORT_SYMBOL(latent_entropy);
167 * Array of node states.
169 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
170 [N_POSSIBLE] = NODE_MASK_ALL,
171 [N_ONLINE] = { { [0] = 1UL } },
173 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
174 #ifdef CONFIG_HIGHMEM
175 [N_HIGH_MEMORY] = { { [0] = 1UL } },
177 [N_MEMORY] = { { [0] = 1UL } },
178 [N_CPU] = { { [0] = 1UL } },
181 EXPORT_SYMBOL(node_states);
183 atomic_long_t _totalram_pages __read_mostly;
184 EXPORT_SYMBOL(_totalram_pages);
185 unsigned long totalreserve_pages __read_mostly;
186 unsigned long totalcma_pages __read_mostly;
188 int percpu_pagelist_high_fraction;
189 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
190 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
191 EXPORT_SYMBOL(init_on_alloc);
193 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
194 EXPORT_SYMBOL(init_on_free);
196 static bool _init_on_alloc_enabled_early __read_mostly
197 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
198 static int __init early_init_on_alloc(char *buf)
201 return kstrtobool(buf, &_init_on_alloc_enabled_early);
203 early_param("init_on_alloc", early_init_on_alloc);
205 static bool _init_on_free_enabled_early __read_mostly
206 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
207 static int __init early_init_on_free(char *buf)
209 return kstrtobool(buf, &_init_on_free_enabled_early);
211 early_param("init_on_free", early_init_on_free);
214 * A cached value of the page's pageblock's migratetype, used when the page is
215 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
216 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
217 * Also the migratetype set in the page does not necessarily match the pcplist
218 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
219 * other index - this ensures that it will be put on the correct CMA freelist.
221 static inline int get_pcppage_migratetype(struct page *page)
226 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
228 page->index = migratetype;
231 #ifdef CONFIG_PM_SLEEP
233 * The following functions are used by the suspend/hibernate code to temporarily
234 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
235 * while devices are suspended. To avoid races with the suspend/hibernate code,
236 * they should always be called with system_transition_mutex held
237 * (gfp_allowed_mask also should only be modified with system_transition_mutex
238 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
239 * with that modification).
242 static gfp_t saved_gfp_mask;
244 void pm_restore_gfp_mask(void)
246 WARN_ON(!mutex_is_locked(&system_transition_mutex));
247 if (saved_gfp_mask) {
248 gfp_allowed_mask = saved_gfp_mask;
253 void pm_restrict_gfp_mask(void)
255 WARN_ON(!mutex_is_locked(&system_transition_mutex));
256 WARN_ON(saved_gfp_mask);
257 saved_gfp_mask = gfp_allowed_mask;
258 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
261 bool pm_suspended_storage(void)
263 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
267 #endif /* CONFIG_PM_SLEEP */
269 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
270 unsigned int pageblock_order __read_mostly;
273 static void __free_pages_ok(struct page *page, unsigned int order,
277 * results with 256, 32 in the lowmem_reserve sysctl:
278 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
279 * 1G machine -> (16M dma, 784M normal, 224M high)
280 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
281 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
282 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
284 * TBD: should special case ZONE_DMA32 machines here - in those we normally
285 * don't need any ZONE_NORMAL reservation
287 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
288 #ifdef CONFIG_ZONE_DMA
291 #ifdef CONFIG_ZONE_DMA32
295 #ifdef CONFIG_HIGHMEM
301 static char * const zone_names[MAX_NR_ZONES] = {
302 #ifdef CONFIG_ZONE_DMA
305 #ifdef CONFIG_ZONE_DMA32
309 #ifdef CONFIG_HIGHMEM
313 #ifdef CONFIG_ZONE_DEVICE
318 const char * const migratetype_names[MIGRATE_TYPES] = {
326 #ifdef CONFIG_MEMORY_ISOLATION
331 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
332 [NULL_COMPOUND_DTOR] = NULL,
333 [COMPOUND_PAGE_DTOR] = free_compound_page,
334 #ifdef CONFIG_HUGETLB_PAGE
335 [HUGETLB_PAGE_DTOR] = free_huge_page,
337 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
338 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
342 int min_free_kbytes = 1024;
343 int user_min_free_kbytes = -1;
344 int watermark_boost_factor __read_mostly = 15000;
345 int watermark_scale_factor = 10;
347 static unsigned long nr_kernel_pages __initdata;
348 static unsigned long nr_all_pages __initdata;
349 static unsigned long dma_reserve __initdata;
351 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
352 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
353 static unsigned long required_kernelcore __initdata;
354 static unsigned long required_kernelcore_percent __initdata;
355 static unsigned long required_movablecore __initdata;
356 static unsigned long required_movablecore_percent __initdata;
357 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
358 static bool mirrored_kernelcore __meminitdata;
360 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
362 EXPORT_SYMBOL(movable_zone);
365 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
366 unsigned int nr_online_nodes __read_mostly = 1;
367 EXPORT_SYMBOL(nr_node_ids);
368 EXPORT_SYMBOL(nr_online_nodes);
371 int page_group_by_mobility_disabled __read_mostly;
373 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
375 * During boot we initialize deferred pages on-demand, as needed, but once
376 * page_alloc_init_late() has finished, the deferred pages are all initialized,
377 * and we can permanently disable that path.
379 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
382 * Calling kasan_poison_pages() only after deferred memory initialization
383 * has completed. Poisoning pages during deferred memory init will greatly
384 * lengthen the process and cause problem in large memory systems as the
385 * deferred pages initialization is done with interrupt disabled.
387 * Assuming that there will be no reference to those newly initialized
388 * pages before they are ever allocated, this should have no effect on
389 * KASAN memory tracking as the poison will be properly inserted at page
390 * allocation time. The only corner case is when pages are allocated by
391 * on-demand allocation and then freed again before the deferred pages
392 * initialization is done, but this is not likely to happen.
394 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
396 return static_branch_unlikely(&deferred_pages) ||
397 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
398 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
399 PageSkipKASanPoison(page);
402 /* Returns true if the struct page for the pfn is uninitialised */
403 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
405 int nid = early_pfn_to_nid(pfn);
407 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
414 * Returns true when the remaining initialisation should be deferred until
415 * later in the boot cycle when it can be parallelised.
417 static bool __meminit
418 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
420 static unsigned long prev_end_pfn, nr_initialised;
423 * prev_end_pfn static that contains the end of previous zone
424 * No need to protect because called very early in boot before smp_init.
426 if (prev_end_pfn != end_pfn) {
427 prev_end_pfn = end_pfn;
431 /* Always populate low zones for address-constrained allocations */
432 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
435 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
438 * We start only with one section of pages, more pages are added as
439 * needed until the rest of deferred pages are initialized.
442 if ((nr_initialised > PAGES_PER_SECTION) &&
443 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
444 NODE_DATA(nid)->first_deferred_pfn = pfn;
450 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
452 return (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
453 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
454 PageSkipKASanPoison(page);
457 static inline bool early_page_uninitialised(unsigned long pfn)
462 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
468 /* Return a pointer to the bitmap storing bits affecting a block of pages */
469 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
472 #ifdef CONFIG_SPARSEMEM
473 return section_to_usemap(__pfn_to_section(pfn));
475 return page_zone(page)->pageblock_flags;
476 #endif /* CONFIG_SPARSEMEM */
479 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
481 #ifdef CONFIG_SPARSEMEM
482 pfn &= (PAGES_PER_SECTION-1);
484 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
485 #endif /* CONFIG_SPARSEMEM */
486 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
489 static __always_inline
490 unsigned long __get_pfnblock_flags_mask(const struct page *page,
494 unsigned long *bitmap;
495 unsigned long bitidx, word_bitidx;
498 bitmap = get_pageblock_bitmap(page, pfn);
499 bitidx = pfn_to_bitidx(page, pfn);
500 word_bitidx = bitidx / BITS_PER_LONG;
501 bitidx &= (BITS_PER_LONG-1);
503 word = bitmap[word_bitidx];
504 return (word >> bitidx) & mask;
508 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
509 * @page: The page within the block of interest
510 * @pfn: The target page frame number
511 * @mask: mask of bits that the caller is interested in
513 * Return: pageblock_bits flags
515 unsigned long get_pfnblock_flags_mask(const struct page *page,
516 unsigned long pfn, unsigned long mask)
518 return __get_pfnblock_flags_mask(page, pfn, mask);
521 static __always_inline int get_pfnblock_migratetype(const struct page *page,
524 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
528 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
529 * @page: The page within the block of interest
530 * @flags: The flags to set
531 * @pfn: The target page frame number
532 * @mask: mask of bits that the caller is interested in
534 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
538 unsigned long *bitmap;
539 unsigned long bitidx, word_bitidx;
540 unsigned long old_word, word;
542 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
543 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
545 bitmap = get_pageblock_bitmap(page, pfn);
546 bitidx = pfn_to_bitidx(page, pfn);
547 word_bitidx = bitidx / BITS_PER_LONG;
548 bitidx &= (BITS_PER_LONG-1);
550 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
555 word = READ_ONCE(bitmap[word_bitidx]);
557 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
558 if (word == old_word)
564 void set_pageblock_migratetype(struct page *page, int migratetype)
566 if (unlikely(page_group_by_mobility_disabled &&
567 migratetype < MIGRATE_PCPTYPES))
568 migratetype = MIGRATE_UNMOVABLE;
570 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
571 page_to_pfn(page), MIGRATETYPE_MASK);
574 #ifdef CONFIG_DEBUG_VM
575 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
579 unsigned long pfn = page_to_pfn(page);
580 unsigned long sp, start_pfn;
583 seq = zone_span_seqbegin(zone);
584 start_pfn = zone->zone_start_pfn;
585 sp = zone->spanned_pages;
586 if (!zone_spans_pfn(zone, pfn))
588 } while (zone_span_seqretry(zone, seq));
591 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
592 pfn, zone_to_nid(zone), zone->name,
593 start_pfn, start_pfn + sp);
598 static int page_is_consistent(struct zone *zone, struct page *page)
600 if (zone != page_zone(page))
606 * Temporary debugging check for pages not lying within a given zone.
608 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
610 if (page_outside_zone_boundaries(zone, page))
612 if (!page_is_consistent(zone, page))
618 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
624 static void bad_page(struct page *page, const char *reason)
626 static unsigned long resume;
627 static unsigned long nr_shown;
628 static unsigned long nr_unshown;
631 * Allow a burst of 60 reports, then keep quiet for that minute;
632 * or allow a steady drip of one report per second.
634 if (nr_shown == 60) {
635 if (time_before(jiffies, resume)) {
641 "BUG: Bad page state: %lu messages suppressed\n",
648 resume = jiffies + 60 * HZ;
650 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
651 current->comm, page_to_pfn(page));
652 dump_page(page, reason);
657 /* Leave bad fields for debug, except PageBuddy could make trouble */
658 page_mapcount_reset(page); /* remove PageBuddy */
659 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
662 static inline unsigned int order_to_pindex(int migratetype, int order)
666 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
667 if (order > PAGE_ALLOC_COSTLY_ORDER) {
668 VM_BUG_ON(order != pageblock_order);
669 base = PAGE_ALLOC_COSTLY_ORDER + 1;
672 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
675 return (MIGRATE_PCPTYPES * base) + migratetype;
678 static inline int pindex_to_order(unsigned int pindex)
680 int order = pindex / MIGRATE_PCPTYPES;
682 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
683 if (order > PAGE_ALLOC_COSTLY_ORDER)
684 order = pageblock_order;
686 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
692 static inline bool pcp_allowed_order(unsigned int order)
694 if (order <= PAGE_ALLOC_COSTLY_ORDER)
696 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
697 if (order == pageblock_order)
703 static inline void free_the_page(struct page *page, unsigned int order)
705 if (pcp_allowed_order(order)) /* Via pcp? */
706 free_unref_page(page, order);
708 __free_pages_ok(page, order, FPI_NONE);
712 * Higher-order pages are called "compound pages". They are structured thusly:
714 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
716 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
717 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
719 * The first tail page's ->compound_dtor holds the offset in array of compound
720 * page destructors. See compound_page_dtors.
722 * The first tail page's ->compound_order holds the order of allocation.
723 * This usage means that zero-order pages may not be compound.
726 void free_compound_page(struct page *page)
728 mem_cgroup_uncharge(page_folio(page));
729 free_the_page(page, compound_order(page));
732 static void prep_compound_head(struct page *page, unsigned int order)
734 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
735 set_compound_order(page, order);
736 atomic_set(compound_mapcount_ptr(page), -1);
737 if (hpage_pincount_available(page))
738 atomic_set(compound_pincount_ptr(page), 0);
741 static void prep_compound_tail(struct page *head, int tail_idx)
743 struct page *p = head + tail_idx;
745 p->mapping = TAIL_MAPPING;
746 set_compound_head(p, head);
749 void prep_compound_page(struct page *page, unsigned int order)
752 int nr_pages = 1 << order;
755 for (i = 1; i < nr_pages; i++)
756 prep_compound_tail(page, i);
758 prep_compound_head(page, order);
761 #ifdef CONFIG_DEBUG_PAGEALLOC
762 unsigned int _debug_guardpage_minorder;
764 bool _debug_pagealloc_enabled_early __read_mostly
765 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
766 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
767 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
768 EXPORT_SYMBOL(_debug_pagealloc_enabled);
770 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
772 static int __init early_debug_pagealloc(char *buf)
774 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
776 early_param("debug_pagealloc", early_debug_pagealloc);
778 static int __init debug_guardpage_minorder_setup(char *buf)
782 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
783 pr_err("Bad debug_guardpage_minorder value\n");
786 _debug_guardpage_minorder = res;
787 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
790 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
792 static inline bool set_page_guard(struct zone *zone, struct page *page,
793 unsigned int order, int migratetype)
795 if (!debug_guardpage_enabled())
798 if (order >= debug_guardpage_minorder())
801 __SetPageGuard(page);
802 INIT_LIST_HEAD(&page->lru);
803 set_page_private(page, order);
804 /* Guard pages are not available for any usage */
805 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
810 static inline void clear_page_guard(struct zone *zone, struct page *page,
811 unsigned int order, int migratetype)
813 if (!debug_guardpage_enabled())
816 __ClearPageGuard(page);
818 set_page_private(page, 0);
819 if (!is_migrate_isolate(migratetype))
820 __mod_zone_freepage_state(zone, (1 << order), migratetype);
823 static inline bool set_page_guard(struct zone *zone, struct page *page,
824 unsigned int order, int migratetype) { return false; }
825 static inline void clear_page_guard(struct zone *zone, struct page *page,
826 unsigned int order, int migratetype) {}
830 * Enable static keys related to various memory debugging and hardening options.
831 * Some override others, and depend on early params that are evaluated in the
832 * order of appearance. So we need to first gather the full picture of what was
833 * enabled, and then make decisions.
835 void init_mem_debugging_and_hardening(void)
837 bool page_poisoning_requested = false;
839 #ifdef CONFIG_PAGE_POISONING
841 * Page poisoning is debug page alloc for some arches. If
842 * either of those options are enabled, enable poisoning.
844 if (page_poisoning_enabled() ||
845 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
846 debug_pagealloc_enabled())) {
847 static_branch_enable(&_page_poisoning_enabled);
848 page_poisoning_requested = true;
852 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
853 page_poisoning_requested) {
854 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
855 "will take precedence over init_on_alloc and init_on_free\n");
856 _init_on_alloc_enabled_early = false;
857 _init_on_free_enabled_early = false;
860 if (_init_on_alloc_enabled_early)
861 static_branch_enable(&init_on_alloc);
863 static_branch_disable(&init_on_alloc);
865 if (_init_on_free_enabled_early)
866 static_branch_enable(&init_on_free);
868 static_branch_disable(&init_on_free);
870 #ifdef CONFIG_DEBUG_PAGEALLOC
871 if (!debug_pagealloc_enabled())
874 static_branch_enable(&_debug_pagealloc_enabled);
876 if (!debug_guardpage_minorder())
879 static_branch_enable(&_debug_guardpage_enabled);
883 static inline void set_buddy_order(struct page *page, unsigned int order)
885 set_page_private(page, order);
886 __SetPageBuddy(page);
890 * This function checks whether a page is free && is the buddy
891 * we can coalesce a page and its buddy if
892 * (a) the buddy is not in a hole (check before calling!) &&
893 * (b) the buddy is in the buddy system &&
894 * (c) a page and its buddy have the same order &&
895 * (d) a page and its buddy are in the same zone.
897 * For recording whether a page is in the buddy system, we set PageBuddy.
898 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
900 * For recording page's order, we use page_private(page).
902 static inline bool page_is_buddy(struct page *page, struct page *buddy,
905 if (!page_is_guard(buddy) && !PageBuddy(buddy))
908 if (buddy_order(buddy) != order)
912 * zone check is done late to avoid uselessly calculating
913 * zone/node ids for pages that could never merge.
915 if (page_zone_id(page) != page_zone_id(buddy))
918 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
923 #ifdef CONFIG_COMPACTION
924 static inline struct capture_control *task_capc(struct zone *zone)
926 struct capture_control *capc = current->capture_control;
928 return unlikely(capc) &&
929 !(current->flags & PF_KTHREAD) &&
931 capc->cc->zone == zone ? capc : NULL;
935 compaction_capture(struct capture_control *capc, struct page *page,
936 int order, int migratetype)
938 if (!capc || order != capc->cc->order)
941 /* Do not accidentally pollute CMA or isolated regions*/
942 if (is_migrate_cma(migratetype) ||
943 is_migrate_isolate(migratetype))
947 * Do not let lower order allocations pollute a movable pageblock.
948 * This might let an unmovable request use a reclaimable pageblock
949 * and vice-versa but no more than normal fallback logic which can
950 * have trouble finding a high-order free page.
952 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
960 static inline struct capture_control *task_capc(struct zone *zone)
966 compaction_capture(struct capture_control *capc, struct page *page,
967 int order, int migratetype)
971 #endif /* CONFIG_COMPACTION */
973 /* Used for pages not on another list */
974 static inline void add_to_free_list(struct page *page, struct zone *zone,
975 unsigned int order, int migratetype)
977 struct free_area *area = &zone->free_area[order];
979 list_add(&page->lru, &area->free_list[migratetype]);
983 /* Used for pages not on another list */
984 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
985 unsigned int order, int migratetype)
987 struct free_area *area = &zone->free_area[order];
989 list_add_tail(&page->lru, &area->free_list[migratetype]);
994 * Used for pages which are on another list. Move the pages to the tail
995 * of the list - so the moved pages won't immediately be considered for
996 * allocation again (e.g., optimization for memory onlining).
998 static inline void move_to_free_list(struct page *page, struct zone *zone,
999 unsigned int order, int migratetype)
1001 struct free_area *area = &zone->free_area[order];
1003 list_move_tail(&page->lru, &area->free_list[migratetype]);
1006 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
1009 /* clear reported state and update reported page count */
1010 if (page_reported(page))
1011 __ClearPageReported(page);
1013 list_del(&page->lru);
1014 __ClearPageBuddy(page);
1015 set_page_private(page, 0);
1016 zone->free_area[order].nr_free--;
1020 * If this is not the largest possible page, check if the buddy
1021 * of the next-highest order is free. If it is, it's possible
1022 * that pages are being freed that will coalesce soon. In case,
1023 * that is happening, add the free page to the tail of the list
1024 * so it's less likely to be used soon and more likely to be merged
1025 * as a higher order page
1028 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1029 struct page *page, unsigned int order)
1031 struct page *higher_page, *higher_buddy;
1032 unsigned long combined_pfn;
1034 if (order >= MAX_ORDER - 2)
1037 combined_pfn = buddy_pfn & pfn;
1038 higher_page = page + (combined_pfn - pfn);
1039 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1040 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1042 return page_is_buddy(higher_page, higher_buddy, order + 1);
1046 * Freeing function for a buddy system allocator.
1048 * The concept of a buddy system is to maintain direct-mapped table
1049 * (containing bit values) for memory blocks of various "orders".
1050 * The bottom level table contains the map for the smallest allocatable
1051 * units of memory (here, pages), and each level above it describes
1052 * pairs of units from the levels below, hence, "buddies".
1053 * At a high level, all that happens here is marking the table entry
1054 * at the bottom level available, and propagating the changes upward
1055 * as necessary, plus some accounting needed to play nicely with other
1056 * parts of the VM system.
1057 * At each level, we keep a list of pages, which are heads of continuous
1058 * free pages of length of (1 << order) and marked with PageBuddy.
1059 * Page's order is recorded in page_private(page) field.
1060 * So when we are allocating or freeing one, we can derive the state of the
1061 * other. That is, if we allocate a small block, and both were
1062 * free, the remainder of the region must be split into blocks.
1063 * If a block is freed, and its buddy is also free, then this
1064 * triggers coalescing into a block of larger size.
1069 static inline void __free_one_page(struct page *page,
1071 struct zone *zone, unsigned int order,
1072 int migratetype, fpi_t fpi_flags)
1074 struct capture_control *capc = task_capc(zone);
1075 unsigned int max_order = pageblock_order;
1076 unsigned long buddy_pfn;
1077 unsigned long combined_pfn;
1081 VM_BUG_ON(!zone_is_initialized(zone));
1082 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1084 VM_BUG_ON(migratetype == -1);
1085 if (likely(!is_migrate_isolate(migratetype)))
1086 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1088 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1089 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1092 while (order < max_order) {
1093 if (compaction_capture(capc, page, order, migratetype)) {
1094 __mod_zone_freepage_state(zone, -(1 << order),
1098 buddy_pfn = __find_buddy_pfn(pfn, order);
1099 buddy = page + (buddy_pfn - pfn);
1101 if (!page_is_buddy(page, buddy, order))
1104 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1105 * merge with it and move up one order.
1107 if (page_is_guard(buddy))
1108 clear_page_guard(zone, buddy, order, migratetype);
1110 del_page_from_free_list(buddy, zone, order);
1111 combined_pfn = buddy_pfn & pfn;
1112 page = page + (combined_pfn - pfn);
1116 if (order < MAX_ORDER - 1) {
1117 /* If we are here, it means order is >= pageblock_order.
1118 * We want to prevent merge between freepages on pageblock
1119 * without fallbacks and normal pageblock. Without this,
1120 * pageblock isolation could cause incorrect freepage or CMA
1121 * accounting or HIGHATOMIC accounting.
1123 * We don't want to hit this code for the more frequent
1124 * low-order merging.
1128 buddy_pfn = __find_buddy_pfn(pfn, order);
1129 buddy = page + (buddy_pfn - pfn);
1130 buddy_mt = get_pageblock_migratetype(buddy);
1132 if (migratetype != buddy_mt
1133 && (!migratetype_is_mergeable(migratetype) ||
1134 !migratetype_is_mergeable(buddy_mt)))
1136 max_order = order + 1;
1137 goto continue_merging;
1141 set_buddy_order(page, order);
1143 if (fpi_flags & FPI_TO_TAIL)
1145 else if (is_shuffle_order(order))
1146 to_tail = shuffle_pick_tail();
1148 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1151 add_to_free_list_tail(page, zone, order, migratetype);
1153 add_to_free_list(page, zone, order, migratetype);
1155 /* Notify page reporting subsystem of freed page */
1156 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1157 page_reporting_notify_free(order);
1161 * A bad page could be due to a number of fields. Instead of multiple branches,
1162 * try and check multiple fields with one check. The caller must do a detailed
1163 * check if necessary.
1165 static inline bool page_expected_state(struct page *page,
1166 unsigned long check_flags)
1168 if (unlikely(atomic_read(&page->_mapcount) != -1))
1171 if (unlikely((unsigned long)page->mapping |
1172 page_ref_count(page) |
1176 (page->flags & check_flags)))
1182 static const char *page_bad_reason(struct page *page, unsigned long flags)
1184 const char *bad_reason = NULL;
1186 if (unlikely(atomic_read(&page->_mapcount) != -1))
1187 bad_reason = "nonzero mapcount";
1188 if (unlikely(page->mapping != NULL))
1189 bad_reason = "non-NULL mapping";
1190 if (unlikely(page_ref_count(page) != 0))
1191 bad_reason = "nonzero _refcount";
1192 if (unlikely(page->flags & flags)) {
1193 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1194 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1196 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1199 if (unlikely(page->memcg_data))
1200 bad_reason = "page still charged to cgroup";
1205 static void check_free_page_bad(struct page *page)
1208 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1211 static inline int check_free_page(struct page *page)
1213 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1216 /* Something has gone sideways, find it */
1217 check_free_page_bad(page);
1221 static int free_tail_pages_check(struct page *head_page, struct page *page)
1226 * We rely page->lru.next never has bit 0 set, unless the page
1227 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1229 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1231 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1235 switch (page - head_page) {
1237 /* the first tail page: ->mapping may be compound_mapcount() */
1238 if (unlikely(compound_mapcount(page))) {
1239 bad_page(page, "nonzero compound_mapcount");
1245 * the second tail page: ->mapping is
1246 * deferred_list.next -- ignore value.
1250 if (page->mapping != TAIL_MAPPING) {
1251 bad_page(page, "corrupted mapping in tail page");
1256 if (unlikely(!PageTail(page))) {
1257 bad_page(page, "PageTail not set");
1260 if (unlikely(compound_head(page) != head_page)) {
1261 bad_page(page, "compound_head not consistent");
1266 page->mapping = NULL;
1267 clear_compound_head(page);
1271 static void kernel_init_free_pages(struct page *page, int numpages, bool zero_tags)
1276 for (i = 0; i < numpages; i++)
1277 tag_clear_highpage(page + i);
1281 /* s390's use of memset() could override KASAN redzones. */
1282 kasan_disable_current();
1283 for (i = 0; i < numpages; i++) {
1284 u8 tag = page_kasan_tag(page + i);
1285 page_kasan_tag_reset(page + i);
1286 clear_highpage(page + i);
1287 page_kasan_tag_set(page + i, tag);
1289 kasan_enable_current();
1292 static __always_inline bool free_pages_prepare(struct page *page,
1293 unsigned int order, bool check_free, fpi_t fpi_flags)
1296 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1298 VM_BUG_ON_PAGE(PageTail(page), page);
1300 trace_mm_page_free(page, order);
1302 if (unlikely(PageHWPoison(page)) && !order) {
1304 * Do not let hwpoison pages hit pcplists/buddy
1305 * Untie memcg state and reset page's owner
1307 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1308 __memcg_kmem_uncharge_page(page, order);
1309 reset_page_owner(page, order);
1310 page_table_check_free(page, order);
1315 * Check tail pages before head page information is cleared to
1316 * avoid checking PageCompound for order-0 pages.
1318 if (unlikely(order)) {
1319 bool compound = PageCompound(page);
1322 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1325 ClearPageDoubleMap(page);
1326 ClearPageHasHWPoisoned(page);
1328 for (i = 1; i < (1 << order); i++) {
1330 bad += free_tail_pages_check(page, page + i);
1331 if (unlikely(check_free_page(page + i))) {
1335 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1338 if (PageMappingFlags(page))
1339 page->mapping = NULL;
1340 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1341 __memcg_kmem_uncharge_page(page, order);
1343 bad += check_free_page(page);
1347 page_cpupid_reset_last(page);
1348 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1349 reset_page_owner(page, order);
1350 page_table_check_free(page, order);
1352 if (!PageHighMem(page)) {
1353 debug_check_no_locks_freed(page_address(page),
1354 PAGE_SIZE << order);
1355 debug_check_no_obj_freed(page_address(page),
1356 PAGE_SIZE << order);
1359 kernel_poison_pages(page, 1 << order);
1362 * As memory initialization might be integrated into KASAN,
1363 * kasan_free_pages and kernel_init_free_pages must be
1364 * kept together to avoid discrepancies in behavior.
1366 * With hardware tag-based KASAN, memory tags must be set before the
1367 * page becomes unavailable via debug_pagealloc or arch_free_page.
1369 if (kasan_has_integrated_init()) {
1370 if (!skip_kasan_poison)
1371 kasan_free_pages(page, order);
1373 bool init = want_init_on_free();
1376 kernel_init_free_pages(page, 1 << order, false);
1377 if (!skip_kasan_poison)
1378 kasan_poison_pages(page, order, init);
1382 * arch_free_page() can make the page's contents inaccessible. s390
1383 * does this. So nothing which can access the page's contents should
1384 * happen after this.
1386 arch_free_page(page, order);
1388 debug_pagealloc_unmap_pages(page, 1 << order);
1393 #ifdef CONFIG_DEBUG_VM
1395 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1396 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1397 * moved from pcp lists to free lists.
1399 static bool free_pcp_prepare(struct page *page, unsigned int order)
1401 return free_pages_prepare(page, order, true, FPI_NONE);
1404 static bool bulkfree_pcp_prepare(struct page *page)
1406 if (debug_pagealloc_enabled_static())
1407 return check_free_page(page);
1413 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1414 * moving from pcp lists to free list in order to reduce overhead. With
1415 * debug_pagealloc enabled, they are checked also immediately when being freed
1418 static bool free_pcp_prepare(struct page *page, unsigned int order)
1420 if (debug_pagealloc_enabled_static())
1421 return free_pages_prepare(page, order, true, FPI_NONE);
1423 return free_pages_prepare(page, order, false, FPI_NONE);
1426 static bool bulkfree_pcp_prepare(struct page *page)
1428 return check_free_page(page);
1430 #endif /* CONFIG_DEBUG_VM */
1433 * Frees a number of pages from the PCP lists
1434 * Assumes all pages on list are in same zone.
1435 * count is the number of pages to free.
1437 static void free_pcppages_bulk(struct zone *zone, int count,
1438 struct per_cpu_pages *pcp,
1442 int max_pindex = NR_PCP_LISTS - 1;
1444 bool isolated_pageblocks;
1448 * Ensure proper count is passed which otherwise would stuck in the
1449 * below while (list_empty(list)) loop.
1451 count = min(pcp->count, count);
1453 /* Ensure requested pindex is drained first. */
1454 pindex = pindex - 1;
1457 * local_lock_irq held so equivalent to spin_lock_irqsave for
1458 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1460 spin_lock(&zone->lock);
1461 isolated_pageblocks = has_isolate_pageblock(zone);
1464 struct list_head *list;
1467 /* Remove pages from lists in a round-robin fashion. */
1469 if (++pindex > max_pindex)
1470 pindex = min_pindex;
1471 list = &pcp->lists[pindex];
1472 if (!list_empty(list))
1475 if (pindex == max_pindex)
1477 if (pindex == min_pindex)
1481 order = pindex_to_order(pindex);
1482 nr_pages = 1 << order;
1483 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1487 page = list_last_entry(list, struct page, lru);
1488 mt = get_pcppage_migratetype(page);
1490 /* must delete to avoid corrupting pcp list */
1491 list_del(&page->lru);
1493 pcp->count -= nr_pages;
1495 if (bulkfree_pcp_prepare(page))
1498 /* MIGRATE_ISOLATE page should not go to pcplists */
1499 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1500 /* Pageblock could have been isolated meanwhile */
1501 if (unlikely(isolated_pageblocks))
1502 mt = get_pageblock_migratetype(page);
1504 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1505 trace_mm_page_pcpu_drain(page, order, mt);
1506 } while (count > 0 && !list_empty(list));
1509 spin_unlock(&zone->lock);
1512 static void free_one_page(struct zone *zone,
1513 struct page *page, unsigned long pfn,
1515 int migratetype, fpi_t fpi_flags)
1517 unsigned long flags;
1519 spin_lock_irqsave(&zone->lock, flags);
1520 if (unlikely(has_isolate_pageblock(zone) ||
1521 is_migrate_isolate(migratetype))) {
1522 migratetype = get_pfnblock_migratetype(page, pfn);
1524 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1525 spin_unlock_irqrestore(&zone->lock, flags);
1528 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1529 unsigned long zone, int nid)
1531 mm_zero_struct_page(page);
1532 set_page_links(page, zone, nid, pfn);
1533 init_page_count(page);
1534 page_mapcount_reset(page);
1535 page_cpupid_reset_last(page);
1536 page_kasan_tag_reset(page);
1538 INIT_LIST_HEAD(&page->lru);
1539 #ifdef WANT_PAGE_VIRTUAL
1540 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1541 if (!is_highmem_idx(zone))
1542 set_page_address(page, __va(pfn << PAGE_SHIFT));
1546 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1547 static void __meminit init_reserved_page(unsigned long pfn)
1552 if (!early_page_uninitialised(pfn))
1555 nid = early_pfn_to_nid(pfn);
1556 pgdat = NODE_DATA(nid);
1558 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1559 struct zone *zone = &pgdat->node_zones[zid];
1561 if (zone_spans_pfn(zone, pfn))
1564 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1567 static inline void init_reserved_page(unsigned long pfn)
1570 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1573 * Initialised pages do not have PageReserved set. This function is
1574 * called for each range allocated by the bootmem allocator and
1575 * marks the pages PageReserved. The remaining valid pages are later
1576 * sent to the buddy page allocator.
1578 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1580 unsigned long start_pfn = PFN_DOWN(start);
1581 unsigned long end_pfn = PFN_UP(end);
1583 for (; start_pfn < end_pfn; start_pfn++) {
1584 if (pfn_valid(start_pfn)) {
1585 struct page *page = pfn_to_page(start_pfn);
1587 init_reserved_page(start_pfn);
1589 /* Avoid false-positive PageTail() */
1590 INIT_LIST_HEAD(&page->lru);
1593 * no need for atomic set_bit because the struct
1594 * page is not visible yet so nobody should
1597 __SetPageReserved(page);
1602 static void __free_pages_ok(struct page *page, unsigned int order,
1605 unsigned long flags;
1607 unsigned long pfn = page_to_pfn(page);
1608 struct zone *zone = page_zone(page);
1610 if (!free_pages_prepare(page, order, true, fpi_flags))
1613 migratetype = get_pfnblock_migratetype(page, pfn);
1615 spin_lock_irqsave(&zone->lock, flags);
1616 if (unlikely(has_isolate_pageblock(zone) ||
1617 is_migrate_isolate(migratetype))) {
1618 migratetype = get_pfnblock_migratetype(page, pfn);
1620 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1621 spin_unlock_irqrestore(&zone->lock, flags);
1623 __count_vm_events(PGFREE, 1 << order);
1626 void __free_pages_core(struct page *page, unsigned int order)
1628 unsigned int nr_pages = 1 << order;
1629 struct page *p = page;
1633 * When initializing the memmap, __init_single_page() sets the refcount
1634 * of all pages to 1 ("allocated"/"not free"). We have to set the
1635 * refcount of all involved pages to 0.
1638 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1640 __ClearPageReserved(p);
1641 set_page_count(p, 0);
1643 __ClearPageReserved(p);
1644 set_page_count(p, 0);
1646 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1649 * Bypass PCP and place fresh pages right to the tail, primarily
1650 * relevant for memory onlining.
1652 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1658 * During memory init memblocks map pfns to nids. The search is expensive and
1659 * this caches recent lookups. The implementation of __early_pfn_to_nid
1660 * treats start/end as pfns.
1662 struct mminit_pfnnid_cache {
1663 unsigned long last_start;
1664 unsigned long last_end;
1668 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1671 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1673 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1674 struct mminit_pfnnid_cache *state)
1676 unsigned long start_pfn, end_pfn;
1679 if (state->last_start <= pfn && pfn < state->last_end)
1680 return state->last_nid;
1682 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1683 if (nid != NUMA_NO_NODE) {
1684 state->last_start = start_pfn;
1685 state->last_end = end_pfn;
1686 state->last_nid = nid;
1692 int __meminit early_pfn_to_nid(unsigned long pfn)
1694 static DEFINE_SPINLOCK(early_pfn_lock);
1697 spin_lock(&early_pfn_lock);
1698 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1700 nid = first_online_node;
1701 spin_unlock(&early_pfn_lock);
1705 #endif /* CONFIG_NUMA */
1707 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1710 if (early_page_uninitialised(pfn))
1712 __free_pages_core(page, order);
1716 * Check that the whole (or subset of) a pageblock given by the interval of
1717 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1718 * with the migration of free compaction scanner.
1720 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1722 * It's possible on some configurations to have a setup like node0 node1 node0
1723 * i.e. it's possible that all pages within a zones range of pages do not
1724 * belong to a single zone. We assume that a border between node0 and node1
1725 * can occur within a single pageblock, but not a node0 node1 node0
1726 * interleaving within a single pageblock. It is therefore sufficient to check
1727 * the first and last page of a pageblock and avoid checking each individual
1728 * page in a pageblock.
1730 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1731 unsigned long end_pfn, struct zone *zone)
1733 struct page *start_page;
1734 struct page *end_page;
1736 /* end_pfn is one past the range we are checking */
1739 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1742 start_page = pfn_to_online_page(start_pfn);
1746 if (page_zone(start_page) != zone)
1749 end_page = pfn_to_page(end_pfn);
1751 /* This gives a shorter code than deriving page_zone(end_page) */
1752 if (page_zone_id(start_page) != page_zone_id(end_page))
1758 void set_zone_contiguous(struct zone *zone)
1760 unsigned long block_start_pfn = zone->zone_start_pfn;
1761 unsigned long block_end_pfn;
1763 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1764 for (; block_start_pfn < zone_end_pfn(zone);
1765 block_start_pfn = block_end_pfn,
1766 block_end_pfn += pageblock_nr_pages) {
1768 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1770 if (!__pageblock_pfn_to_page(block_start_pfn,
1771 block_end_pfn, zone))
1776 /* We confirm that there is no hole */
1777 zone->contiguous = true;
1780 void clear_zone_contiguous(struct zone *zone)
1782 zone->contiguous = false;
1785 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1786 static void __init deferred_free_range(unsigned long pfn,
1787 unsigned long nr_pages)
1795 page = pfn_to_page(pfn);
1797 /* Free a large naturally-aligned chunk if possible */
1798 if (nr_pages == pageblock_nr_pages &&
1799 (pfn & (pageblock_nr_pages - 1)) == 0) {
1800 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1801 __free_pages_core(page, pageblock_order);
1805 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1806 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1807 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1808 __free_pages_core(page, 0);
1812 /* Completion tracking for deferred_init_memmap() threads */
1813 static atomic_t pgdat_init_n_undone __initdata;
1814 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1816 static inline void __init pgdat_init_report_one_done(void)
1818 if (atomic_dec_and_test(&pgdat_init_n_undone))
1819 complete(&pgdat_init_all_done_comp);
1823 * Returns true if page needs to be initialized or freed to buddy allocator.
1825 * First we check if pfn is valid on architectures where it is possible to have
1826 * holes within pageblock_nr_pages. On systems where it is not possible, this
1827 * function is optimized out.
1829 * Then, we check if a current large page is valid by only checking the validity
1832 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1834 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1840 * Free pages to buddy allocator. Try to free aligned pages in
1841 * pageblock_nr_pages sizes.
1843 static void __init deferred_free_pages(unsigned long pfn,
1844 unsigned long end_pfn)
1846 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1847 unsigned long nr_free = 0;
1849 for (; pfn < end_pfn; pfn++) {
1850 if (!deferred_pfn_valid(pfn)) {
1851 deferred_free_range(pfn - nr_free, nr_free);
1853 } else if (!(pfn & nr_pgmask)) {
1854 deferred_free_range(pfn - nr_free, nr_free);
1860 /* Free the last block of pages to allocator */
1861 deferred_free_range(pfn - nr_free, nr_free);
1865 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1866 * by performing it only once every pageblock_nr_pages.
1867 * Return number of pages initialized.
1869 static unsigned long __init deferred_init_pages(struct zone *zone,
1871 unsigned long end_pfn)
1873 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1874 int nid = zone_to_nid(zone);
1875 unsigned long nr_pages = 0;
1876 int zid = zone_idx(zone);
1877 struct page *page = NULL;
1879 for (; pfn < end_pfn; pfn++) {
1880 if (!deferred_pfn_valid(pfn)) {
1883 } else if (!page || !(pfn & nr_pgmask)) {
1884 page = pfn_to_page(pfn);
1888 __init_single_page(page, pfn, zid, nid);
1895 * This function is meant to pre-load the iterator for the zone init.
1896 * Specifically it walks through the ranges until we are caught up to the
1897 * first_init_pfn value and exits there. If we never encounter the value we
1898 * return false indicating there are no valid ranges left.
1901 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1902 unsigned long *spfn, unsigned long *epfn,
1903 unsigned long first_init_pfn)
1908 * Start out by walking through the ranges in this zone that have
1909 * already been initialized. We don't need to do anything with them
1910 * so we just need to flush them out of the system.
1912 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1913 if (*epfn <= first_init_pfn)
1915 if (*spfn < first_init_pfn)
1916 *spfn = first_init_pfn;
1925 * Initialize and free pages. We do it in two loops: first we initialize
1926 * struct page, then free to buddy allocator, because while we are
1927 * freeing pages we can access pages that are ahead (computing buddy
1928 * page in __free_one_page()).
1930 * In order to try and keep some memory in the cache we have the loop
1931 * broken along max page order boundaries. This way we will not cause
1932 * any issues with the buddy page computation.
1934 static unsigned long __init
1935 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1936 unsigned long *end_pfn)
1938 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1939 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1940 unsigned long nr_pages = 0;
1943 /* First we loop through and initialize the page values */
1944 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1947 if (mo_pfn <= *start_pfn)
1950 t = min(mo_pfn, *end_pfn);
1951 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1953 if (mo_pfn < *end_pfn) {
1954 *start_pfn = mo_pfn;
1959 /* Reset values and now loop through freeing pages as needed */
1962 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1968 t = min(mo_pfn, epfn);
1969 deferred_free_pages(spfn, t);
1979 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1982 unsigned long spfn, epfn;
1983 struct zone *zone = arg;
1986 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1989 * Initialize and free pages in MAX_ORDER sized increments so that we
1990 * can avoid introducing any issues with the buddy allocator.
1992 while (spfn < end_pfn) {
1993 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1998 /* An arch may override for more concurrency. */
2000 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2005 /* Initialise remaining memory on a node */
2006 static int __init deferred_init_memmap(void *data)
2008 pg_data_t *pgdat = data;
2009 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2010 unsigned long spfn = 0, epfn = 0;
2011 unsigned long first_init_pfn, flags;
2012 unsigned long start = jiffies;
2014 int zid, max_threads;
2017 /* Bind memory initialisation thread to a local node if possible */
2018 if (!cpumask_empty(cpumask))
2019 set_cpus_allowed_ptr(current, cpumask);
2021 pgdat_resize_lock(pgdat, &flags);
2022 first_init_pfn = pgdat->first_deferred_pfn;
2023 if (first_init_pfn == ULONG_MAX) {
2024 pgdat_resize_unlock(pgdat, &flags);
2025 pgdat_init_report_one_done();
2029 /* Sanity check boundaries */
2030 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2031 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2032 pgdat->first_deferred_pfn = ULONG_MAX;
2035 * Once we unlock here, the zone cannot be grown anymore, thus if an
2036 * interrupt thread must allocate this early in boot, zone must be
2037 * pre-grown prior to start of deferred page initialization.
2039 pgdat_resize_unlock(pgdat, &flags);
2041 /* Only the highest zone is deferred so find it */
2042 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2043 zone = pgdat->node_zones + zid;
2044 if (first_init_pfn < zone_end_pfn(zone))
2048 /* If the zone is empty somebody else may have cleared out the zone */
2049 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2053 max_threads = deferred_page_init_max_threads(cpumask);
2055 while (spfn < epfn) {
2056 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2057 struct padata_mt_job job = {
2058 .thread_fn = deferred_init_memmap_chunk,
2061 .size = epfn_align - spfn,
2062 .align = PAGES_PER_SECTION,
2063 .min_chunk = PAGES_PER_SECTION,
2064 .max_threads = max_threads,
2067 padata_do_multithreaded(&job);
2068 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2072 /* Sanity check that the next zone really is unpopulated */
2073 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2075 pr_info("node %d deferred pages initialised in %ums\n",
2076 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2078 pgdat_init_report_one_done();
2083 * If this zone has deferred pages, try to grow it by initializing enough
2084 * deferred pages to satisfy the allocation specified by order, rounded up to
2085 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2086 * of SECTION_SIZE bytes by initializing struct pages in increments of
2087 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2089 * Return true when zone was grown, otherwise return false. We return true even
2090 * when we grow less than requested, to let the caller decide if there are
2091 * enough pages to satisfy the allocation.
2093 * Note: We use noinline because this function is needed only during boot, and
2094 * it is called from a __ref function _deferred_grow_zone. This way we are
2095 * making sure that it is not inlined into permanent text section.
2097 static noinline bool __init
2098 deferred_grow_zone(struct zone *zone, unsigned int order)
2100 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2101 pg_data_t *pgdat = zone->zone_pgdat;
2102 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2103 unsigned long spfn, epfn, flags;
2104 unsigned long nr_pages = 0;
2107 /* Only the last zone may have deferred pages */
2108 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2111 pgdat_resize_lock(pgdat, &flags);
2114 * If someone grew this zone while we were waiting for spinlock, return
2115 * true, as there might be enough pages already.
2117 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2118 pgdat_resize_unlock(pgdat, &flags);
2122 /* If the zone is empty somebody else may have cleared out the zone */
2123 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2124 first_deferred_pfn)) {
2125 pgdat->first_deferred_pfn = ULONG_MAX;
2126 pgdat_resize_unlock(pgdat, &flags);
2127 /* Retry only once. */
2128 return first_deferred_pfn != ULONG_MAX;
2132 * Initialize and free pages in MAX_ORDER sized increments so
2133 * that we can avoid introducing any issues with the buddy
2136 while (spfn < epfn) {
2137 /* update our first deferred PFN for this section */
2138 first_deferred_pfn = spfn;
2140 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2141 touch_nmi_watchdog();
2143 /* We should only stop along section boundaries */
2144 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2147 /* If our quota has been met we can stop here */
2148 if (nr_pages >= nr_pages_needed)
2152 pgdat->first_deferred_pfn = spfn;
2153 pgdat_resize_unlock(pgdat, &flags);
2155 return nr_pages > 0;
2159 * deferred_grow_zone() is __init, but it is called from
2160 * get_page_from_freelist() during early boot until deferred_pages permanently
2161 * disables this call. This is why we have refdata wrapper to avoid warning,
2162 * and to ensure that the function body gets unloaded.
2165 _deferred_grow_zone(struct zone *zone, unsigned int order)
2167 return deferred_grow_zone(zone, order);
2170 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2172 void __init page_alloc_init_late(void)
2177 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2179 /* There will be num_node_state(N_MEMORY) threads */
2180 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2181 for_each_node_state(nid, N_MEMORY) {
2182 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2185 /* Block until all are initialised */
2186 wait_for_completion(&pgdat_init_all_done_comp);
2189 * We initialized the rest of the deferred pages. Permanently disable
2190 * on-demand struct page initialization.
2192 static_branch_disable(&deferred_pages);
2194 /* Reinit limits that are based on free pages after the kernel is up */
2195 files_maxfiles_init();
2200 /* Discard memblock private memory */
2203 for_each_node_state(nid, N_MEMORY)
2204 shuffle_free_memory(NODE_DATA(nid));
2206 for_each_populated_zone(zone)
2207 set_zone_contiguous(zone);
2211 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2212 void __init init_cma_reserved_pageblock(struct page *page)
2214 unsigned i = pageblock_nr_pages;
2215 struct page *p = page;
2218 __ClearPageReserved(p);
2219 set_page_count(p, 0);
2222 set_pageblock_migratetype(page, MIGRATE_CMA);
2223 set_page_refcounted(page);
2224 __free_pages(page, pageblock_order);
2226 adjust_managed_page_count(page, pageblock_nr_pages);
2227 page_zone(page)->cma_pages += pageblock_nr_pages;
2232 * The order of subdivision here is critical for the IO subsystem.
2233 * Please do not alter this order without good reasons and regression
2234 * testing. Specifically, as large blocks of memory are subdivided,
2235 * the order in which smaller blocks are delivered depends on the order
2236 * they're subdivided in this function. This is the primary factor
2237 * influencing the order in which pages are delivered to the IO
2238 * subsystem according to empirical testing, and this is also justified
2239 * by considering the behavior of a buddy system containing a single
2240 * large block of memory acted on by a series of small allocations.
2241 * This behavior is a critical factor in sglist merging's success.
2245 static inline void expand(struct zone *zone, struct page *page,
2246 int low, int high, int migratetype)
2248 unsigned long size = 1 << high;
2250 while (high > low) {
2253 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2256 * Mark as guard pages (or page), that will allow to
2257 * merge back to allocator when buddy will be freed.
2258 * Corresponding page table entries will not be touched,
2259 * pages will stay not present in virtual address space
2261 if (set_page_guard(zone, &page[size], high, migratetype))
2264 add_to_free_list(&page[size], zone, high, migratetype);
2265 set_buddy_order(&page[size], high);
2269 static void check_new_page_bad(struct page *page)
2271 if (unlikely(page->flags & __PG_HWPOISON)) {
2272 /* Don't complain about hwpoisoned pages */
2273 page_mapcount_reset(page); /* remove PageBuddy */
2278 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2282 * This page is about to be returned from the page allocator
2284 static inline int check_new_page(struct page *page)
2286 if (likely(page_expected_state(page,
2287 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2290 check_new_page_bad(page);
2294 static bool check_new_pages(struct page *page, unsigned int order)
2297 for (i = 0; i < (1 << order); i++) {
2298 struct page *p = page + i;
2300 if (unlikely(check_new_page(p)))
2307 #ifdef CONFIG_DEBUG_VM
2309 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2310 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2311 * also checked when pcp lists are refilled from the free lists.
2313 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2315 if (debug_pagealloc_enabled_static())
2316 return check_new_pages(page, order);
2321 static inline bool check_new_pcp(struct page *page, unsigned int order)
2323 return check_new_pages(page, order);
2327 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2328 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2329 * enabled, they are also checked when being allocated from the pcp lists.
2331 static inline bool check_pcp_refill(struct page *page, unsigned int order)
2333 return check_new_pages(page, order);
2335 static inline bool check_new_pcp(struct page *page, unsigned int order)
2337 if (debug_pagealloc_enabled_static())
2338 return check_new_pages(page, order);
2342 #endif /* CONFIG_DEBUG_VM */
2344 inline void post_alloc_hook(struct page *page, unsigned int order,
2347 set_page_private(page, 0);
2348 set_page_refcounted(page);
2350 arch_alloc_page(page, order);
2351 debug_pagealloc_map_pages(page, 1 << order);
2354 * Page unpoisoning must happen before memory initialization.
2355 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2356 * allocations and the page unpoisoning code will complain.
2358 kernel_unpoison_pages(page, 1 << order);
2361 * As memory initialization might be integrated into KASAN,
2362 * kasan_alloc_pages and kernel_init_free_pages must be
2363 * kept together to avoid discrepancies in behavior.
2365 if (kasan_has_integrated_init()) {
2366 kasan_alloc_pages(page, order, gfp_flags);
2368 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2370 kasan_unpoison_pages(page, order, init);
2372 kernel_init_free_pages(page, 1 << order,
2373 gfp_flags & __GFP_ZEROTAGS);
2376 set_page_owner(page, order, gfp_flags);
2377 page_table_check_alloc(page, order);
2380 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2381 unsigned int alloc_flags)
2383 post_alloc_hook(page, order, gfp_flags);
2385 if (order && (gfp_flags & __GFP_COMP))
2386 prep_compound_page(page, order);
2389 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2390 * allocate the page. The expectation is that the caller is taking
2391 * steps that will free more memory. The caller should avoid the page
2392 * being used for !PFMEMALLOC purposes.
2394 if (alloc_flags & ALLOC_NO_WATERMARKS)
2395 set_page_pfmemalloc(page);
2397 clear_page_pfmemalloc(page);
2401 * Go through the free lists for the given migratetype and remove
2402 * the smallest available page from the freelists
2404 static __always_inline
2405 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2408 unsigned int current_order;
2409 struct free_area *area;
2412 /* Find a page of the appropriate size in the preferred list */
2413 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2414 area = &(zone->free_area[current_order]);
2415 page = get_page_from_free_area(area, migratetype);
2418 del_page_from_free_list(page, zone, current_order);
2419 expand(zone, page, order, current_order, migratetype);
2420 set_pcppage_migratetype(page, migratetype);
2429 * This array describes the order lists are fallen back to when
2430 * the free lists for the desirable migrate type are depleted
2432 * The other migratetypes do not have fallbacks.
2434 static int fallbacks[MIGRATE_TYPES][3] = {
2435 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2436 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2437 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2441 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2444 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2447 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2448 unsigned int order) { return NULL; }
2452 * Move the free pages in a range to the freelist tail of the requested type.
2453 * Note that start_page and end_pages are not aligned on a pageblock
2454 * boundary. If alignment is required, use move_freepages_block()
2456 static int move_freepages(struct zone *zone,
2457 unsigned long start_pfn, unsigned long end_pfn,
2458 int migratetype, int *num_movable)
2463 int pages_moved = 0;
2465 for (pfn = start_pfn; pfn <= end_pfn;) {
2466 page = pfn_to_page(pfn);
2467 if (!PageBuddy(page)) {
2469 * We assume that pages that could be isolated for
2470 * migration are movable. But we don't actually try
2471 * isolating, as that would be expensive.
2474 (PageLRU(page) || __PageMovable(page)))
2480 /* Make sure we are not inadvertently changing nodes */
2481 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2482 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2484 order = buddy_order(page);
2485 move_to_free_list(page, zone, order, migratetype);
2487 pages_moved += 1 << order;
2493 int move_freepages_block(struct zone *zone, struct page *page,
2494 int migratetype, int *num_movable)
2496 unsigned long start_pfn, end_pfn, pfn;
2501 pfn = page_to_pfn(page);
2502 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2503 end_pfn = start_pfn + pageblock_nr_pages - 1;
2505 /* Do not cross zone boundaries */
2506 if (!zone_spans_pfn(zone, start_pfn))
2508 if (!zone_spans_pfn(zone, end_pfn))
2511 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2515 static void change_pageblock_range(struct page *pageblock_page,
2516 int start_order, int migratetype)
2518 int nr_pageblocks = 1 << (start_order - pageblock_order);
2520 while (nr_pageblocks--) {
2521 set_pageblock_migratetype(pageblock_page, migratetype);
2522 pageblock_page += pageblock_nr_pages;
2527 * When we are falling back to another migratetype during allocation, try to
2528 * steal extra free pages from the same pageblocks to satisfy further
2529 * allocations, instead of polluting multiple pageblocks.
2531 * If we are stealing a relatively large buddy page, it is likely there will
2532 * be more free pages in the pageblock, so try to steal them all. For
2533 * reclaimable and unmovable allocations, we steal regardless of page size,
2534 * as fragmentation caused by those allocations polluting movable pageblocks
2535 * is worse than movable allocations stealing from unmovable and reclaimable
2538 static bool can_steal_fallback(unsigned int order, int start_mt)
2541 * Leaving this order check is intended, although there is
2542 * relaxed order check in next check. The reason is that
2543 * we can actually steal whole pageblock if this condition met,
2544 * but, below check doesn't guarantee it and that is just heuristic
2545 * so could be changed anytime.
2547 if (order >= pageblock_order)
2550 if (order >= pageblock_order / 2 ||
2551 start_mt == MIGRATE_RECLAIMABLE ||
2552 start_mt == MIGRATE_UNMOVABLE ||
2553 page_group_by_mobility_disabled)
2559 static inline bool boost_watermark(struct zone *zone)
2561 unsigned long max_boost;
2563 if (!watermark_boost_factor)
2566 * Don't bother in zones that are unlikely to produce results.
2567 * On small machines, including kdump capture kernels running
2568 * in a small area, boosting the watermark can cause an out of
2569 * memory situation immediately.
2571 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2574 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2575 watermark_boost_factor, 10000);
2578 * high watermark may be uninitialised if fragmentation occurs
2579 * very early in boot so do not boost. We do not fall
2580 * through and boost by pageblock_nr_pages as failing
2581 * allocations that early means that reclaim is not going
2582 * to help and it may even be impossible to reclaim the
2583 * boosted watermark resulting in a hang.
2588 max_boost = max(pageblock_nr_pages, max_boost);
2590 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2597 * This function implements actual steal behaviour. If order is large enough,
2598 * we can steal whole pageblock. If not, we first move freepages in this
2599 * pageblock to our migratetype and determine how many already-allocated pages
2600 * are there in the pageblock with a compatible migratetype. If at least half
2601 * of pages are free or compatible, we can change migratetype of the pageblock
2602 * itself, so pages freed in the future will be put on the correct free list.
2604 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2605 unsigned int alloc_flags, int start_type, bool whole_block)
2607 unsigned int current_order = buddy_order(page);
2608 int free_pages, movable_pages, alike_pages;
2611 old_block_type = get_pageblock_migratetype(page);
2614 * This can happen due to races and we want to prevent broken
2615 * highatomic accounting.
2617 if (is_migrate_highatomic(old_block_type))
2620 /* Take ownership for orders >= pageblock_order */
2621 if (current_order >= pageblock_order) {
2622 change_pageblock_range(page, current_order, start_type);
2627 * Boost watermarks to increase reclaim pressure to reduce the
2628 * likelihood of future fallbacks. Wake kswapd now as the node
2629 * may be balanced overall and kswapd will not wake naturally.
2631 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2632 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2634 /* We are not allowed to try stealing from the whole block */
2638 free_pages = move_freepages_block(zone, page, start_type,
2641 * Determine how many pages are compatible with our allocation.
2642 * For movable allocation, it's the number of movable pages which
2643 * we just obtained. For other types it's a bit more tricky.
2645 if (start_type == MIGRATE_MOVABLE) {
2646 alike_pages = movable_pages;
2649 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2650 * to MOVABLE pageblock, consider all non-movable pages as
2651 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2652 * vice versa, be conservative since we can't distinguish the
2653 * exact migratetype of non-movable pages.
2655 if (old_block_type == MIGRATE_MOVABLE)
2656 alike_pages = pageblock_nr_pages
2657 - (free_pages + movable_pages);
2662 /* moving whole block can fail due to zone boundary conditions */
2667 * If a sufficient number of pages in the block are either free or of
2668 * comparable migratability as our allocation, claim the whole block.
2670 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2671 page_group_by_mobility_disabled)
2672 set_pageblock_migratetype(page, start_type);
2677 move_to_free_list(page, zone, current_order, start_type);
2681 * Check whether there is a suitable fallback freepage with requested order.
2682 * If only_stealable is true, this function returns fallback_mt only if
2683 * we can steal other freepages all together. This would help to reduce
2684 * fragmentation due to mixed migratetype pages in one pageblock.
2686 int find_suitable_fallback(struct free_area *area, unsigned int order,
2687 int migratetype, bool only_stealable, bool *can_steal)
2692 if (area->nr_free == 0)
2697 fallback_mt = fallbacks[migratetype][i];
2698 if (fallback_mt == MIGRATE_TYPES)
2701 if (free_area_empty(area, fallback_mt))
2704 if (can_steal_fallback(order, migratetype))
2707 if (!only_stealable)
2718 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2719 * there are no empty page blocks that contain a page with a suitable order
2721 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2722 unsigned int alloc_order)
2725 unsigned long max_managed, flags;
2728 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2729 * Check is race-prone but harmless.
2731 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2732 if (zone->nr_reserved_highatomic >= max_managed)
2735 spin_lock_irqsave(&zone->lock, flags);
2737 /* Recheck the nr_reserved_highatomic limit under the lock */
2738 if (zone->nr_reserved_highatomic >= max_managed)
2742 mt = get_pageblock_migratetype(page);
2743 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2744 if (migratetype_is_mergeable(mt)) {
2745 zone->nr_reserved_highatomic += pageblock_nr_pages;
2746 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2747 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2751 spin_unlock_irqrestore(&zone->lock, flags);
2755 * Used when an allocation is about to fail under memory pressure. This
2756 * potentially hurts the reliability of high-order allocations when under
2757 * intense memory pressure but failed atomic allocations should be easier
2758 * to recover from than an OOM.
2760 * If @force is true, try to unreserve a pageblock even though highatomic
2761 * pageblock is exhausted.
2763 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2766 struct zonelist *zonelist = ac->zonelist;
2767 unsigned long flags;
2774 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2777 * Preserve at least one pageblock unless memory pressure
2780 if (!force && zone->nr_reserved_highatomic <=
2784 spin_lock_irqsave(&zone->lock, flags);
2785 for (order = 0; order < MAX_ORDER; order++) {
2786 struct free_area *area = &(zone->free_area[order]);
2788 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2793 * In page freeing path, migratetype change is racy so
2794 * we can counter several free pages in a pageblock
2795 * in this loop although we changed the pageblock type
2796 * from highatomic to ac->migratetype. So we should
2797 * adjust the count once.
2799 if (is_migrate_highatomic_page(page)) {
2801 * It should never happen but changes to
2802 * locking could inadvertently allow a per-cpu
2803 * drain to add pages to MIGRATE_HIGHATOMIC
2804 * while unreserving so be safe and watch for
2807 zone->nr_reserved_highatomic -= min(
2809 zone->nr_reserved_highatomic);
2813 * Convert to ac->migratetype and avoid the normal
2814 * pageblock stealing heuristics. Minimally, the caller
2815 * is doing the work and needs the pages. More
2816 * importantly, if the block was always converted to
2817 * MIGRATE_UNMOVABLE or another type then the number
2818 * of pageblocks that cannot be completely freed
2821 set_pageblock_migratetype(page, ac->migratetype);
2822 ret = move_freepages_block(zone, page, ac->migratetype,
2825 spin_unlock_irqrestore(&zone->lock, flags);
2829 spin_unlock_irqrestore(&zone->lock, flags);
2836 * Try finding a free buddy page on the fallback list and put it on the free
2837 * list of requested migratetype, possibly along with other pages from the same
2838 * block, depending on fragmentation avoidance heuristics. Returns true if
2839 * fallback was found so that __rmqueue_smallest() can grab it.
2841 * The use of signed ints for order and current_order is a deliberate
2842 * deviation from the rest of this file, to make the for loop
2843 * condition simpler.
2845 static __always_inline bool
2846 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2847 unsigned int alloc_flags)
2849 struct free_area *area;
2851 int min_order = order;
2857 * Do not steal pages from freelists belonging to other pageblocks
2858 * i.e. orders < pageblock_order. If there are no local zones free,
2859 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2861 if (alloc_flags & ALLOC_NOFRAGMENT)
2862 min_order = pageblock_order;
2865 * Find the largest available free page in the other list. This roughly
2866 * approximates finding the pageblock with the most free pages, which
2867 * would be too costly to do exactly.
2869 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2871 area = &(zone->free_area[current_order]);
2872 fallback_mt = find_suitable_fallback(area, current_order,
2873 start_migratetype, false, &can_steal);
2874 if (fallback_mt == -1)
2878 * We cannot steal all free pages from the pageblock and the
2879 * requested migratetype is movable. In that case it's better to
2880 * steal and split the smallest available page instead of the
2881 * largest available page, because even if the next movable
2882 * allocation falls back into a different pageblock than this
2883 * one, it won't cause permanent fragmentation.
2885 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2886 && current_order > order)
2895 for (current_order = order; current_order < MAX_ORDER;
2897 area = &(zone->free_area[current_order]);
2898 fallback_mt = find_suitable_fallback(area, current_order,
2899 start_migratetype, false, &can_steal);
2900 if (fallback_mt != -1)
2905 * This should not happen - we already found a suitable fallback
2906 * when looking for the largest page.
2908 VM_BUG_ON(current_order == MAX_ORDER);
2911 page = get_page_from_free_area(area, fallback_mt);
2913 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2916 trace_mm_page_alloc_extfrag(page, order, current_order,
2917 start_migratetype, fallback_mt);
2924 * Do the hard work of removing an element from the buddy allocator.
2925 * Call me with the zone->lock already held.
2927 static __always_inline struct page *
2928 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2929 unsigned int alloc_flags)
2933 if (IS_ENABLED(CONFIG_CMA)) {
2935 * Balance movable allocations between regular and CMA areas by
2936 * allocating from CMA when over half of the zone's free memory
2937 * is in the CMA area.
2939 if (alloc_flags & ALLOC_CMA &&
2940 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2941 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2942 page = __rmqueue_cma_fallback(zone, order);
2948 page = __rmqueue_smallest(zone, order, migratetype);
2949 if (unlikely(!page)) {
2950 if (alloc_flags & ALLOC_CMA)
2951 page = __rmqueue_cma_fallback(zone, order);
2953 if (!page && __rmqueue_fallback(zone, order, migratetype,
2959 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2964 * Obtain a specified number of elements from the buddy allocator, all under
2965 * a single hold of the lock, for efficiency. Add them to the supplied list.
2966 * Returns the number of new pages which were placed at *list.
2968 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2969 unsigned long count, struct list_head *list,
2970 int migratetype, unsigned int alloc_flags)
2972 int i, allocated = 0;
2975 * local_lock_irq held so equivalent to spin_lock_irqsave for
2976 * both PREEMPT_RT and non-PREEMPT_RT configurations.
2978 spin_lock(&zone->lock);
2979 for (i = 0; i < count; ++i) {
2980 struct page *page = __rmqueue(zone, order, migratetype,
2982 if (unlikely(page == NULL))
2985 if (unlikely(check_pcp_refill(page, order)))
2989 * Split buddy pages returned by expand() are received here in
2990 * physical page order. The page is added to the tail of
2991 * caller's list. From the callers perspective, the linked list
2992 * is ordered by page number under some conditions. This is
2993 * useful for IO devices that can forward direction from the
2994 * head, thus also in the physical page order. This is useful
2995 * for IO devices that can merge IO requests if the physical
2996 * pages are ordered properly.
2998 list_add_tail(&page->lru, list);
3000 if (is_migrate_cma(get_pcppage_migratetype(page)))
3001 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3006 * i pages were removed from the buddy list even if some leak due
3007 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3008 * on i. Do not confuse with 'allocated' which is the number of
3009 * pages added to the pcp list.
3011 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3012 spin_unlock(&zone->lock);
3018 * Called from the vmstat counter updater to drain pagesets of this
3019 * currently executing processor on remote nodes after they have
3022 * Note that this function must be called with the thread pinned to
3023 * a single processor.
3025 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3027 unsigned long flags;
3028 int to_drain, batch;
3030 local_lock_irqsave(&pagesets.lock, flags);
3031 batch = READ_ONCE(pcp->batch);
3032 to_drain = min(pcp->count, batch);
3034 free_pcppages_bulk(zone, to_drain, pcp, 0);
3035 local_unlock_irqrestore(&pagesets.lock, flags);
3040 * Drain pcplists of the indicated processor and zone.
3042 * The processor must either be the current processor and the
3043 * thread pinned to the current processor or a processor that
3046 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3048 unsigned long flags;
3049 struct per_cpu_pages *pcp;
3051 local_lock_irqsave(&pagesets.lock, flags);
3053 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3055 free_pcppages_bulk(zone, pcp->count, pcp, 0);
3057 local_unlock_irqrestore(&pagesets.lock, flags);
3061 * Drain pcplists of all zones on the indicated processor.
3063 * The processor must either be the current processor and the
3064 * thread pinned to the current processor or a processor that
3067 static void drain_pages(unsigned int cpu)
3071 for_each_populated_zone(zone) {
3072 drain_pages_zone(cpu, zone);
3077 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3079 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3080 * the single zone's pages.
3082 void drain_local_pages(struct zone *zone)
3084 int cpu = smp_processor_id();
3087 drain_pages_zone(cpu, zone);
3092 static void drain_local_pages_wq(struct work_struct *work)
3094 struct pcpu_drain *drain;
3096 drain = container_of(work, struct pcpu_drain, work);
3099 * drain_all_pages doesn't use proper cpu hotplug protection so
3100 * we can race with cpu offline when the WQ can move this from
3101 * a cpu pinned worker to an unbound one. We can operate on a different
3102 * cpu which is alright but we also have to make sure to not move to
3106 drain_local_pages(drain->zone);
3111 * The implementation of drain_all_pages(), exposing an extra parameter to
3112 * drain on all cpus.
3114 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3115 * not empty. The check for non-emptiness can however race with a free to
3116 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3117 * that need the guarantee that every CPU has drained can disable the
3118 * optimizing racy check.
3120 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3125 * Allocate in the BSS so we won't require allocation in
3126 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3128 static cpumask_t cpus_with_pcps;
3131 * Make sure nobody triggers this path before mm_percpu_wq is fully
3134 if (WARN_ON_ONCE(!mm_percpu_wq))
3138 * Do not drain if one is already in progress unless it's specific to
3139 * a zone. Such callers are primarily CMA and memory hotplug and need
3140 * the drain to be complete when the call returns.
3142 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3145 mutex_lock(&pcpu_drain_mutex);
3149 * We don't care about racing with CPU hotplug event
3150 * as offline notification will cause the notified
3151 * cpu to drain that CPU pcps and on_each_cpu_mask
3152 * disables preemption as part of its processing
3154 for_each_online_cpu(cpu) {
3155 struct per_cpu_pages *pcp;
3157 bool has_pcps = false;
3159 if (force_all_cpus) {
3161 * The pcp.count check is racy, some callers need a
3162 * guarantee that no cpu is missed.
3166 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3170 for_each_populated_zone(z) {
3171 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3180 cpumask_set_cpu(cpu, &cpus_with_pcps);
3182 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3185 for_each_cpu(cpu, &cpus_with_pcps) {
3186 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3189 INIT_WORK(&drain->work, drain_local_pages_wq);
3190 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3192 for_each_cpu(cpu, &cpus_with_pcps)
3193 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3195 mutex_unlock(&pcpu_drain_mutex);
3199 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3201 * When zone parameter is non-NULL, spill just the single zone's pages.
3203 * Note that this can be extremely slow as the draining happens in a workqueue.
3205 void drain_all_pages(struct zone *zone)
3207 __drain_all_pages(zone, false);
3210 #ifdef CONFIG_HIBERNATION
3213 * Touch the watchdog for every WD_PAGE_COUNT pages.
3215 #define WD_PAGE_COUNT (128*1024)
3217 void mark_free_pages(struct zone *zone)
3219 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3220 unsigned long flags;
3221 unsigned int order, t;
3224 if (zone_is_empty(zone))
3227 spin_lock_irqsave(&zone->lock, flags);
3229 max_zone_pfn = zone_end_pfn(zone);
3230 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3231 if (pfn_valid(pfn)) {
3232 page = pfn_to_page(pfn);
3234 if (!--page_count) {
3235 touch_nmi_watchdog();
3236 page_count = WD_PAGE_COUNT;
3239 if (page_zone(page) != zone)
3242 if (!swsusp_page_is_forbidden(page))
3243 swsusp_unset_page_free(page);
3246 for_each_migratetype_order(order, t) {
3247 list_for_each_entry(page,
3248 &zone->free_area[order].free_list[t], lru) {
3251 pfn = page_to_pfn(page);
3252 for (i = 0; i < (1UL << order); i++) {
3253 if (!--page_count) {
3254 touch_nmi_watchdog();
3255 page_count = WD_PAGE_COUNT;
3257 swsusp_set_page_free(pfn_to_page(pfn + i));
3261 spin_unlock_irqrestore(&zone->lock, flags);
3263 #endif /* CONFIG_PM */
3265 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3270 if (!free_pcp_prepare(page, order))
3273 migratetype = get_pfnblock_migratetype(page, pfn);
3274 set_pcppage_migratetype(page, migratetype);
3278 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch,
3281 int min_nr_free, max_nr_free;
3283 /* Free everything if batch freeing high-order pages. */
3284 if (unlikely(free_high))
3287 /* Check for PCP disabled or boot pageset */
3288 if (unlikely(high < batch))
3291 /* Leave at least pcp->batch pages on the list */
3292 min_nr_free = batch;
3293 max_nr_free = high - batch;
3296 * Double the number of pages freed each time there is subsequent
3297 * freeing of pages without any allocation.
3299 batch <<= pcp->free_factor;
3300 if (batch < max_nr_free)
3302 batch = clamp(batch, min_nr_free, max_nr_free);
3307 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
3310 int high = READ_ONCE(pcp->high);
3312 if (unlikely(!high || free_high))
3315 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3319 * If reclaim is active, limit the number of pages that can be
3320 * stored on pcp lists
3322 return min(READ_ONCE(pcp->batch) << 2, high);
3325 static void free_unref_page_commit(struct page *page, int migratetype,
3328 struct zone *zone = page_zone(page);
3329 struct per_cpu_pages *pcp;
3334 __count_vm_event(PGFREE);
3335 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3336 pindex = order_to_pindex(migratetype, order);
3337 list_add(&page->lru, &pcp->lists[pindex]);
3338 pcp->count += 1 << order;
3341 * As high-order pages other than THP's stored on PCP can contribute
3342 * to fragmentation, limit the number stored when PCP is heavily
3343 * freeing without allocation. The remainder after bulk freeing
3344 * stops will be drained from vmstat refresh context.
3346 free_high = (pcp->free_factor && order && order <= PAGE_ALLOC_COSTLY_ORDER);
3348 high = nr_pcp_high(pcp, zone, free_high);
3349 if (pcp->count >= high) {
3350 int batch = READ_ONCE(pcp->batch);
3352 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch, free_high), pcp, pindex);
3359 void free_unref_page(struct page *page, unsigned int order)
3361 unsigned long flags;
3362 unsigned long pfn = page_to_pfn(page);
3365 if (!free_unref_page_prepare(page, pfn, order))
3369 * We only track unmovable, reclaimable and movable on pcp lists.
3370 * Place ISOLATE pages on the isolated list because they are being
3371 * offlined but treat HIGHATOMIC as movable pages so we can get those
3372 * areas back if necessary. Otherwise, we may have to free
3373 * excessively into the page allocator
3375 migratetype = get_pcppage_migratetype(page);
3376 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3377 if (unlikely(is_migrate_isolate(migratetype))) {
3378 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3381 migratetype = MIGRATE_MOVABLE;
3384 local_lock_irqsave(&pagesets.lock, flags);
3385 free_unref_page_commit(page, migratetype, order);
3386 local_unlock_irqrestore(&pagesets.lock, flags);
3390 * Free a list of 0-order pages
3392 void free_unref_page_list(struct list_head *list)
3394 struct page *page, *next;
3395 unsigned long flags;
3396 int batch_count = 0;
3399 /* Prepare pages for freeing */
3400 list_for_each_entry_safe(page, next, list, lru) {
3401 unsigned long pfn = page_to_pfn(page);
3402 if (!free_unref_page_prepare(page, pfn, 0)) {
3403 list_del(&page->lru);
3408 * Free isolated pages directly to the allocator, see
3409 * comment in free_unref_page.
3411 migratetype = get_pcppage_migratetype(page);
3412 if (unlikely(is_migrate_isolate(migratetype))) {
3413 list_del(&page->lru);
3414 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3419 local_lock_irqsave(&pagesets.lock, flags);
3420 list_for_each_entry_safe(page, next, list, lru) {
3422 * Non-isolated types over MIGRATE_PCPTYPES get added
3423 * to the MIGRATE_MOVABLE pcp list.
3425 migratetype = get_pcppage_migratetype(page);
3426 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3427 migratetype = MIGRATE_MOVABLE;
3429 trace_mm_page_free_batched(page);
3430 free_unref_page_commit(page, migratetype, 0);
3433 * Guard against excessive IRQ disabled times when we get
3434 * a large list of pages to free.
3436 if (++batch_count == SWAP_CLUSTER_MAX) {
3437 local_unlock_irqrestore(&pagesets.lock, flags);
3439 local_lock_irqsave(&pagesets.lock, flags);
3442 local_unlock_irqrestore(&pagesets.lock, flags);
3446 * split_page takes a non-compound higher-order page, and splits it into
3447 * n (1<<order) sub-pages: page[0..n]
3448 * Each sub-page must be freed individually.
3450 * Note: this is probably too low level an operation for use in drivers.
3451 * Please consult with lkml before using this in your driver.
3453 void split_page(struct page *page, unsigned int order)
3457 VM_BUG_ON_PAGE(PageCompound(page), page);
3458 VM_BUG_ON_PAGE(!page_count(page), page);
3460 for (i = 1; i < (1 << order); i++)
3461 set_page_refcounted(page + i);
3462 split_page_owner(page, 1 << order);
3463 split_page_memcg(page, 1 << order);
3465 EXPORT_SYMBOL_GPL(split_page);
3467 int __isolate_free_page(struct page *page, unsigned int order)
3469 unsigned long watermark;
3473 BUG_ON(!PageBuddy(page));
3475 zone = page_zone(page);
3476 mt = get_pageblock_migratetype(page);
3478 if (!is_migrate_isolate(mt)) {
3480 * Obey watermarks as if the page was being allocated. We can
3481 * emulate a high-order watermark check with a raised order-0
3482 * watermark, because we already know our high-order page
3485 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3486 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3489 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3492 /* Remove page from free list */
3494 del_page_from_free_list(page, zone, order);
3497 * Set the pageblock if the isolated page is at least half of a
3500 if (order >= pageblock_order - 1) {
3501 struct page *endpage = page + (1 << order) - 1;
3502 for (; page < endpage; page += pageblock_nr_pages) {
3503 int mt = get_pageblock_migratetype(page);
3505 * Only change normal pageblocks (i.e., they can merge
3508 if (migratetype_is_mergeable(mt))
3509 set_pageblock_migratetype(page,
3515 return 1UL << order;
3519 * __putback_isolated_page - Return a now-isolated page back where we got it
3520 * @page: Page that was isolated
3521 * @order: Order of the isolated page
3522 * @mt: The page's pageblock's migratetype
3524 * This function is meant to return a page pulled from the free lists via
3525 * __isolate_free_page back to the free lists they were pulled from.
3527 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3529 struct zone *zone = page_zone(page);
3531 /* zone lock should be held when this function is called */
3532 lockdep_assert_held(&zone->lock);
3534 /* Return isolated page to tail of freelist. */
3535 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3536 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3540 * Update NUMA hit/miss statistics
3542 * Must be called with interrupts disabled.
3544 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3548 enum numa_stat_item local_stat = NUMA_LOCAL;
3550 /* skip numa counters update if numa stats is disabled */
3551 if (!static_branch_likely(&vm_numa_stat_key))
3554 if (zone_to_nid(z) != numa_node_id())
3555 local_stat = NUMA_OTHER;
3557 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3558 __count_numa_events(z, NUMA_HIT, nr_account);
3560 __count_numa_events(z, NUMA_MISS, nr_account);
3561 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3563 __count_numa_events(z, local_stat, nr_account);
3567 /* Remove page from the per-cpu list, caller must protect the list */
3569 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3571 unsigned int alloc_flags,
3572 struct per_cpu_pages *pcp,
3573 struct list_head *list)
3578 if (list_empty(list)) {
3579 int batch = READ_ONCE(pcp->batch);
3583 * Scale batch relative to order if batch implies
3584 * free pages can be stored on the PCP. Batch can
3585 * be 1 for small zones or for boot pagesets which
3586 * should never store free pages as the pages may
3587 * belong to arbitrary zones.
3590 batch = max(batch >> order, 2);
3591 alloced = rmqueue_bulk(zone, order,
3593 migratetype, alloc_flags);
3595 pcp->count += alloced << order;
3596 if (unlikely(list_empty(list)))
3600 page = list_first_entry(list, struct page, lru);
3601 list_del(&page->lru);
3602 pcp->count -= 1 << order;
3603 } while (check_new_pcp(page, order));
3608 /* Lock and remove page from the per-cpu list */
3609 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3610 struct zone *zone, unsigned int order,
3611 gfp_t gfp_flags, int migratetype,
3612 unsigned int alloc_flags)
3614 struct per_cpu_pages *pcp;
3615 struct list_head *list;
3617 unsigned long flags;
3619 local_lock_irqsave(&pagesets.lock, flags);
3622 * On allocation, reduce the number of pages that are batch freed.
3623 * See nr_pcp_free() where free_factor is increased for subsequent
3626 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3627 pcp->free_factor >>= 1;
3628 list = &pcp->lists[order_to_pindex(migratetype, order)];
3629 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3630 local_unlock_irqrestore(&pagesets.lock, flags);
3632 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3633 zone_statistics(preferred_zone, zone, 1);
3639 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3642 struct page *rmqueue(struct zone *preferred_zone,
3643 struct zone *zone, unsigned int order,
3644 gfp_t gfp_flags, unsigned int alloc_flags,
3647 unsigned long flags;
3650 if (likely(pcp_allowed_order(order))) {
3652 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3653 * we need to skip it when CMA area isn't allowed.
3655 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3656 migratetype != MIGRATE_MOVABLE) {
3657 page = rmqueue_pcplist(preferred_zone, zone, order,
3658 gfp_flags, migratetype, alloc_flags);
3664 * We most definitely don't want callers attempting to
3665 * allocate greater than order-1 page units with __GFP_NOFAIL.
3667 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3671 spin_lock_irqsave(&zone->lock, flags);
3673 * order-0 request can reach here when the pcplist is skipped
3674 * due to non-CMA allocation context. HIGHATOMIC area is
3675 * reserved for high-order atomic allocation, so order-0
3676 * request should skip it.
3678 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3679 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3681 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3684 page = __rmqueue(zone, order, migratetype, alloc_flags);
3688 __mod_zone_freepage_state(zone, -(1 << order),
3689 get_pcppage_migratetype(page));
3690 spin_unlock_irqrestore(&zone->lock, flags);
3691 } while (check_new_pages(page, order));
3693 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3694 zone_statistics(preferred_zone, zone, 1);
3697 /* Separate test+clear to avoid unnecessary atomics */
3698 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3699 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3700 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3703 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3707 spin_unlock_irqrestore(&zone->lock, flags);
3711 #ifdef CONFIG_FAIL_PAGE_ALLOC
3714 struct fault_attr attr;
3716 bool ignore_gfp_highmem;
3717 bool ignore_gfp_reclaim;
3719 } fail_page_alloc = {
3720 .attr = FAULT_ATTR_INITIALIZER,
3721 .ignore_gfp_reclaim = true,
3722 .ignore_gfp_highmem = true,
3726 static int __init setup_fail_page_alloc(char *str)
3728 return setup_fault_attr(&fail_page_alloc.attr, str);
3730 __setup("fail_page_alloc=", setup_fail_page_alloc);
3732 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3734 if (order < fail_page_alloc.min_order)
3736 if (gfp_mask & __GFP_NOFAIL)
3738 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3740 if (fail_page_alloc.ignore_gfp_reclaim &&
3741 (gfp_mask & __GFP_DIRECT_RECLAIM))
3744 return should_fail(&fail_page_alloc.attr, 1 << order);
3747 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3749 static int __init fail_page_alloc_debugfs(void)
3751 umode_t mode = S_IFREG | 0600;
3754 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3755 &fail_page_alloc.attr);
3757 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3758 &fail_page_alloc.ignore_gfp_reclaim);
3759 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3760 &fail_page_alloc.ignore_gfp_highmem);
3761 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3766 late_initcall(fail_page_alloc_debugfs);
3768 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3770 #else /* CONFIG_FAIL_PAGE_ALLOC */
3772 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3777 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3779 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3781 return __should_fail_alloc_page(gfp_mask, order);
3783 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3785 static inline long __zone_watermark_unusable_free(struct zone *z,
3786 unsigned int order, unsigned int alloc_flags)
3788 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3789 long unusable_free = (1 << order) - 1;
3792 * If the caller does not have rights to ALLOC_HARDER then subtract
3793 * the high-atomic reserves. This will over-estimate the size of the
3794 * atomic reserve but it avoids a search.
3796 if (likely(!alloc_harder))
3797 unusable_free += z->nr_reserved_highatomic;
3800 /* If allocation can't use CMA areas don't use free CMA pages */
3801 if (!(alloc_flags & ALLOC_CMA))
3802 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3805 return unusable_free;
3809 * Return true if free base pages are above 'mark'. For high-order checks it
3810 * will return true of the order-0 watermark is reached and there is at least
3811 * one free page of a suitable size. Checking now avoids taking the zone lock
3812 * to check in the allocation paths if no pages are free.
3814 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3815 int highest_zoneidx, unsigned int alloc_flags,
3820 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3822 /* free_pages may go negative - that's OK */
3823 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3825 if (alloc_flags & ALLOC_HIGH)
3828 if (unlikely(alloc_harder)) {
3830 * OOM victims can try even harder than normal ALLOC_HARDER
3831 * users on the grounds that it's definitely going to be in
3832 * the exit path shortly and free memory. Any allocation it
3833 * makes during the free path will be small and short-lived.
3835 if (alloc_flags & ALLOC_OOM)
3842 * Check watermarks for an order-0 allocation request. If these
3843 * are not met, then a high-order request also cannot go ahead
3844 * even if a suitable page happened to be free.
3846 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3849 /* If this is an order-0 request then the watermark is fine */
3853 /* For a high-order request, check at least one suitable page is free */
3854 for (o = order; o < MAX_ORDER; o++) {
3855 struct free_area *area = &z->free_area[o];
3861 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3862 if (!free_area_empty(area, mt))
3867 if ((alloc_flags & ALLOC_CMA) &&
3868 !free_area_empty(area, MIGRATE_CMA)) {
3872 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3878 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3879 int highest_zoneidx, unsigned int alloc_flags)
3881 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3882 zone_page_state(z, NR_FREE_PAGES));
3885 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3886 unsigned long mark, int highest_zoneidx,
3887 unsigned int alloc_flags, gfp_t gfp_mask)
3891 free_pages = zone_page_state(z, NR_FREE_PAGES);
3894 * Fast check for order-0 only. If this fails then the reserves
3895 * need to be calculated.
3900 fast_free = free_pages;
3901 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3902 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3906 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3910 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3911 * when checking the min watermark. The min watermark is the
3912 * point where boosting is ignored so that kswapd is woken up
3913 * when below the low watermark.
3915 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3916 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3917 mark = z->_watermark[WMARK_MIN];
3918 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3919 alloc_flags, free_pages);
3925 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3926 unsigned long mark, int highest_zoneidx)
3928 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3930 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3931 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3933 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3938 int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3940 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3942 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3943 node_reclaim_distance;
3945 #else /* CONFIG_NUMA */
3946 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3950 #endif /* CONFIG_NUMA */
3953 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3954 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3955 * premature use of a lower zone may cause lowmem pressure problems that
3956 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3957 * probably too small. It only makes sense to spread allocations to avoid
3958 * fragmentation between the Normal and DMA32 zones.
3960 static inline unsigned int
3961 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3963 unsigned int alloc_flags;
3966 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3969 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3971 #ifdef CONFIG_ZONE_DMA32
3975 if (zone_idx(zone) != ZONE_NORMAL)
3979 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3980 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3981 * on UMA that if Normal is populated then so is DMA32.
3983 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3984 if (nr_online_nodes > 1 && !populated_zone(--zone))
3987 alloc_flags |= ALLOC_NOFRAGMENT;
3988 #endif /* CONFIG_ZONE_DMA32 */
3992 /* Must be called after current_gfp_context() which can change gfp_mask */
3993 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3994 unsigned int alloc_flags)
3997 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3998 alloc_flags |= ALLOC_CMA;
4004 * get_page_from_freelist goes through the zonelist trying to allocate
4007 static struct page *
4008 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4009 const struct alloc_context *ac)
4013 struct pglist_data *last_pgdat_dirty_limit = NULL;
4018 * Scan zonelist, looking for a zone with enough free.
4019 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4021 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4022 z = ac->preferred_zoneref;
4023 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4028 if (cpusets_enabled() &&
4029 (alloc_flags & ALLOC_CPUSET) &&
4030 !__cpuset_zone_allowed(zone, gfp_mask))
4033 * When allocating a page cache page for writing, we
4034 * want to get it from a node that is within its dirty
4035 * limit, such that no single node holds more than its
4036 * proportional share of globally allowed dirty pages.
4037 * The dirty limits take into account the node's
4038 * lowmem reserves and high watermark so that kswapd
4039 * should be able to balance it without having to
4040 * write pages from its LRU list.
4042 * XXX: For now, allow allocations to potentially
4043 * exceed the per-node dirty limit in the slowpath
4044 * (spread_dirty_pages unset) before going into reclaim,
4045 * which is important when on a NUMA setup the allowed
4046 * nodes are together not big enough to reach the
4047 * global limit. The proper fix for these situations
4048 * will require awareness of nodes in the
4049 * dirty-throttling and the flusher threads.
4051 if (ac->spread_dirty_pages) {
4052 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4055 if (!node_dirty_ok(zone->zone_pgdat)) {
4056 last_pgdat_dirty_limit = zone->zone_pgdat;
4061 if (no_fallback && nr_online_nodes > 1 &&
4062 zone != ac->preferred_zoneref->zone) {
4066 * If moving to a remote node, retry but allow
4067 * fragmenting fallbacks. Locality is more important
4068 * than fragmentation avoidance.
4070 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4071 if (zone_to_nid(zone) != local_nid) {
4072 alloc_flags &= ~ALLOC_NOFRAGMENT;
4077 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4078 if (!zone_watermark_fast(zone, order, mark,
4079 ac->highest_zoneidx, alloc_flags,
4083 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4085 * Watermark failed for this zone, but see if we can
4086 * grow this zone if it contains deferred pages.
4088 if (static_branch_unlikely(&deferred_pages)) {
4089 if (_deferred_grow_zone(zone, order))
4093 /* Checked here to keep the fast path fast */
4094 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4095 if (alloc_flags & ALLOC_NO_WATERMARKS)
4098 if (!node_reclaim_enabled() ||
4099 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4102 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4104 case NODE_RECLAIM_NOSCAN:
4107 case NODE_RECLAIM_FULL:
4108 /* scanned but unreclaimable */
4111 /* did we reclaim enough */
4112 if (zone_watermark_ok(zone, order, mark,
4113 ac->highest_zoneidx, alloc_flags))
4121 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4122 gfp_mask, alloc_flags, ac->migratetype);
4124 prep_new_page(page, order, gfp_mask, alloc_flags);
4127 * If this is a high-order atomic allocation then check
4128 * if the pageblock should be reserved for the future
4130 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4131 reserve_highatomic_pageblock(page, zone, order);
4135 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4136 /* Try again if zone has deferred pages */
4137 if (static_branch_unlikely(&deferred_pages)) {
4138 if (_deferred_grow_zone(zone, order))
4146 * It's possible on a UMA machine to get through all zones that are
4147 * fragmented. If avoiding fragmentation, reset and try again.
4150 alloc_flags &= ~ALLOC_NOFRAGMENT;
4157 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4159 unsigned int filter = SHOW_MEM_FILTER_NODES;
4162 * This documents exceptions given to allocations in certain
4163 * contexts that are allowed to allocate outside current's set
4166 if (!(gfp_mask & __GFP_NOMEMALLOC))
4167 if (tsk_is_oom_victim(current) ||
4168 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4169 filter &= ~SHOW_MEM_FILTER_NODES;
4170 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4171 filter &= ~SHOW_MEM_FILTER_NODES;
4173 show_mem(filter, nodemask);
4176 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4178 struct va_format vaf;
4180 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4182 if ((gfp_mask & __GFP_NOWARN) ||
4183 !__ratelimit(&nopage_rs) ||
4184 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4187 va_start(args, fmt);
4190 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4191 current->comm, &vaf, gfp_mask, &gfp_mask,
4192 nodemask_pr_args(nodemask));
4195 cpuset_print_current_mems_allowed();
4198 warn_alloc_show_mem(gfp_mask, nodemask);
4201 static inline struct page *
4202 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4203 unsigned int alloc_flags,
4204 const struct alloc_context *ac)
4208 page = get_page_from_freelist(gfp_mask, order,
4209 alloc_flags|ALLOC_CPUSET, ac);
4211 * fallback to ignore cpuset restriction if our nodes
4215 page = get_page_from_freelist(gfp_mask, order,
4221 static inline struct page *
4222 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4223 const struct alloc_context *ac, unsigned long *did_some_progress)
4225 struct oom_control oc = {
4226 .zonelist = ac->zonelist,
4227 .nodemask = ac->nodemask,
4229 .gfp_mask = gfp_mask,
4234 *did_some_progress = 0;
4237 * Acquire the oom lock. If that fails, somebody else is
4238 * making progress for us.
4240 if (!mutex_trylock(&oom_lock)) {
4241 *did_some_progress = 1;
4242 schedule_timeout_uninterruptible(1);
4247 * Go through the zonelist yet one more time, keep very high watermark
4248 * here, this is only to catch a parallel oom killing, we must fail if
4249 * we're still under heavy pressure. But make sure that this reclaim
4250 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4251 * allocation which will never fail due to oom_lock already held.
4253 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4254 ~__GFP_DIRECT_RECLAIM, order,
4255 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4259 /* Coredumps can quickly deplete all memory reserves */
4260 if (current->flags & PF_DUMPCORE)
4262 /* The OOM killer will not help higher order allocs */
4263 if (order > PAGE_ALLOC_COSTLY_ORDER)
4266 * We have already exhausted all our reclaim opportunities without any
4267 * success so it is time to admit defeat. We will skip the OOM killer
4268 * because it is very likely that the caller has a more reasonable
4269 * fallback than shooting a random task.
4271 * The OOM killer may not free memory on a specific node.
4273 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4275 /* The OOM killer does not needlessly kill tasks for lowmem */
4276 if (ac->highest_zoneidx < ZONE_NORMAL)
4278 if (pm_suspended_storage())
4281 * XXX: GFP_NOFS allocations should rather fail than rely on
4282 * other request to make a forward progress.
4283 * We are in an unfortunate situation where out_of_memory cannot
4284 * do much for this context but let's try it to at least get
4285 * access to memory reserved if the current task is killed (see
4286 * out_of_memory). Once filesystems are ready to handle allocation
4287 * failures more gracefully we should just bail out here.
4290 /* Exhausted what can be done so it's blame time */
4291 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4292 *did_some_progress = 1;
4295 * Help non-failing allocations by giving them access to memory
4298 if (gfp_mask & __GFP_NOFAIL)
4299 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4300 ALLOC_NO_WATERMARKS, ac);
4303 mutex_unlock(&oom_lock);
4308 * Maximum number of compaction retries with a progress before OOM
4309 * killer is consider as the only way to move forward.
4311 #define MAX_COMPACT_RETRIES 16
4313 #ifdef CONFIG_COMPACTION
4314 /* Try memory compaction for high-order allocations before reclaim */
4315 static struct page *
4316 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4317 unsigned int alloc_flags, const struct alloc_context *ac,
4318 enum compact_priority prio, enum compact_result *compact_result)
4320 struct page *page = NULL;
4321 unsigned long pflags;
4322 unsigned int noreclaim_flag;
4327 psi_memstall_enter(&pflags);
4328 delayacct_compact_start();
4329 noreclaim_flag = memalloc_noreclaim_save();
4331 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4334 memalloc_noreclaim_restore(noreclaim_flag);
4335 psi_memstall_leave(&pflags);
4336 delayacct_compact_end();
4338 if (*compact_result == COMPACT_SKIPPED)
4341 * At least in one zone compaction wasn't deferred or skipped, so let's
4342 * count a compaction stall
4344 count_vm_event(COMPACTSTALL);
4346 /* Prep a captured page if available */
4348 prep_new_page(page, order, gfp_mask, alloc_flags);
4350 /* Try get a page from the freelist if available */
4352 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4355 struct zone *zone = page_zone(page);
4357 zone->compact_blockskip_flush = false;
4358 compaction_defer_reset(zone, order, true);
4359 count_vm_event(COMPACTSUCCESS);
4364 * It's bad if compaction run occurs and fails. The most likely reason
4365 * is that pages exist, but not enough to satisfy watermarks.
4367 count_vm_event(COMPACTFAIL);
4375 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4376 enum compact_result compact_result,
4377 enum compact_priority *compact_priority,
4378 int *compaction_retries)
4380 int max_retries = MAX_COMPACT_RETRIES;
4383 int retries = *compaction_retries;
4384 enum compact_priority priority = *compact_priority;
4389 if (fatal_signal_pending(current))
4392 if (compaction_made_progress(compact_result))
4393 (*compaction_retries)++;
4396 * compaction considers all the zone as desperately out of memory
4397 * so it doesn't really make much sense to retry except when the
4398 * failure could be caused by insufficient priority
4400 if (compaction_failed(compact_result))
4401 goto check_priority;
4404 * compaction was skipped because there are not enough order-0 pages
4405 * to work with, so we retry only if it looks like reclaim can help.
4407 if (compaction_needs_reclaim(compact_result)) {
4408 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4413 * make sure the compaction wasn't deferred or didn't bail out early
4414 * due to locks contention before we declare that we should give up.
4415 * But the next retry should use a higher priority if allowed, so
4416 * we don't just keep bailing out endlessly.
4418 if (compaction_withdrawn(compact_result)) {
4419 goto check_priority;
4423 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4424 * costly ones because they are de facto nofail and invoke OOM
4425 * killer to move on while costly can fail and users are ready
4426 * to cope with that. 1/4 retries is rather arbitrary but we
4427 * would need much more detailed feedback from compaction to
4428 * make a better decision.
4430 if (order > PAGE_ALLOC_COSTLY_ORDER)
4432 if (*compaction_retries <= max_retries) {
4438 * Make sure there are attempts at the highest priority if we exhausted
4439 * all retries or failed at the lower priorities.
4442 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4443 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4445 if (*compact_priority > min_priority) {
4446 (*compact_priority)--;
4447 *compaction_retries = 0;
4451 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4455 static inline struct page *
4456 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4457 unsigned int alloc_flags, const struct alloc_context *ac,
4458 enum compact_priority prio, enum compact_result *compact_result)
4460 *compact_result = COMPACT_SKIPPED;
4465 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4466 enum compact_result compact_result,
4467 enum compact_priority *compact_priority,
4468 int *compaction_retries)
4473 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4477 * There are setups with compaction disabled which would prefer to loop
4478 * inside the allocator rather than hit the oom killer prematurely.
4479 * Let's give them a good hope and keep retrying while the order-0
4480 * watermarks are OK.
4482 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4483 ac->highest_zoneidx, ac->nodemask) {
4484 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4485 ac->highest_zoneidx, alloc_flags))
4490 #endif /* CONFIG_COMPACTION */
4492 #ifdef CONFIG_LOCKDEP
4493 static struct lockdep_map __fs_reclaim_map =
4494 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4496 static bool __need_reclaim(gfp_t gfp_mask)
4498 /* no reclaim without waiting on it */
4499 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4502 /* this guy won't enter reclaim */
4503 if (current->flags & PF_MEMALLOC)
4506 if (gfp_mask & __GFP_NOLOCKDEP)
4512 void __fs_reclaim_acquire(unsigned long ip)
4514 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4517 void __fs_reclaim_release(unsigned long ip)
4519 lock_release(&__fs_reclaim_map, ip);
4522 void fs_reclaim_acquire(gfp_t gfp_mask)
4524 gfp_mask = current_gfp_context(gfp_mask);
4526 if (__need_reclaim(gfp_mask)) {
4527 if (gfp_mask & __GFP_FS)
4528 __fs_reclaim_acquire(_RET_IP_);
4530 #ifdef CONFIG_MMU_NOTIFIER
4531 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4532 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4537 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4539 void fs_reclaim_release(gfp_t gfp_mask)
4541 gfp_mask = current_gfp_context(gfp_mask);
4543 if (__need_reclaim(gfp_mask)) {
4544 if (gfp_mask & __GFP_FS)
4545 __fs_reclaim_release(_RET_IP_);
4548 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4551 /* Perform direct synchronous page reclaim */
4552 static unsigned long
4553 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4554 const struct alloc_context *ac)
4556 unsigned int noreclaim_flag;
4557 unsigned long progress;
4561 /* We now go into synchronous reclaim */
4562 cpuset_memory_pressure_bump();
4563 fs_reclaim_acquire(gfp_mask);
4564 noreclaim_flag = memalloc_noreclaim_save();
4566 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4569 memalloc_noreclaim_restore(noreclaim_flag);
4570 fs_reclaim_release(gfp_mask);
4577 /* The really slow allocator path where we enter direct reclaim */
4578 static inline struct page *
4579 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4580 unsigned int alloc_flags, const struct alloc_context *ac,
4581 unsigned long *did_some_progress)
4583 struct page *page = NULL;
4584 unsigned long pflags;
4585 bool drained = false;
4587 psi_memstall_enter(&pflags);
4588 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4589 if (unlikely(!(*did_some_progress)))
4593 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4596 * If an allocation failed after direct reclaim, it could be because
4597 * pages are pinned on the per-cpu lists or in high alloc reserves.
4598 * Shrink them and try again
4600 if (!page && !drained) {
4601 unreserve_highatomic_pageblock(ac, false);
4602 drain_all_pages(NULL);
4607 psi_memstall_leave(&pflags);
4612 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4613 const struct alloc_context *ac)
4617 pg_data_t *last_pgdat = NULL;
4618 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4620 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4622 if (last_pgdat != zone->zone_pgdat)
4623 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4624 last_pgdat = zone->zone_pgdat;
4628 static inline unsigned int
4629 gfp_to_alloc_flags(gfp_t gfp_mask)
4631 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4634 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4635 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4636 * to save two branches.
4638 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4639 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4642 * The caller may dip into page reserves a bit more if the caller
4643 * cannot run direct reclaim, or if the caller has realtime scheduling
4644 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4645 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4647 alloc_flags |= (__force int)
4648 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4650 if (gfp_mask & __GFP_ATOMIC) {
4652 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4653 * if it can't schedule.
4655 if (!(gfp_mask & __GFP_NOMEMALLOC))
4656 alloc_flags |= ALLOC_HARDER;
4658 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4659 * comment for __cpuset_node_allowed().
4661 alloc_flags &= ~ALLOC_CPUSET;
4662 } else if (unlikely(rt_task(current)) && in_task())
4663 alloc_flags |= ALLOC_HARDER;
4665 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4670 static bool oom_reserves_allowed(struct task_struct *tsk)
4672 if (!tsk_is_oom_victim(tsk))
4676 * !MMU doesn't have oom reaper so give access to memory reserves
4677 * only to the thread with TIF_MEMDIE set
4679 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4686 * Distinguish requests which really need access to full memory
4687 * reserves from oom victims which can live with a portion of it
4689 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4691 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4693 if (gfp_mask & __GFP_MEMALLOC)
4694 return ALLOC_NO_WATERMARKS;
4695 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4696 return ALLOC_NO_WATERMARKS;
4697 if (!in_interrupt()) {
4698 if (current->flags & PF_MEMALLOC)
4699 return ALLOC_NO_WATERMARKS;
4700 else if (oom_reserves_allowed(current))
4707 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4709 return !!__gfp_pfmemalloc_flags(gfp_mask);
4713 * Checks whether it makes sense to retry the reclaim to make a forward progress
4714 * for the given allocation request.
4716 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4717 * without success, or when we couldn't even meet the watermark if we
4718 * reclaimed all remaining pages on the LRU lists.
4720 * Returns true if a retry is viable or false to enter the oom path.
4723 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4724 struct alloc_context *ac, int alloc_flags,
4725 bool did_some_progress, int *no_progress_loops)
4732 * Costly allocations might have made a progress but this doesn't mean
4733 * their order will become available due to high fragmentation so
4734 * always increment the no progress counter for them
4736 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4737 *no_progress_loops = 0;
4739 (*no_progress_loops)++;
4742 * Make sure we converge to OOM if we cannot make any progress
4743 * several times in the row.
4745 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4746 /* Before OOM, exhaust highatomic_reserve */
4747 return unreserve_highatomic_pageblock(ac, true);
4751 * Keep reclaiming pages while there is a chance this will lead
4752 * somewhere. If none of the target zones can satisfy our allocation
4753 * request even if all reclaimable pages are considered then we are
4754 * screwed and have to go OOM.
4756 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4757 ac->highest_zoneidx, ac->nodemask) {
4758 unsigned long available;
4759 unsigned long reclaimable;
4760 unsigned long min_wmark = min_wmark_pages(zone);
4763 available = reclaimable = zone_reclaimable_pages(zone);
4764 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4767 * Would the allocation succeed if we reclaimed all
4768 * reclaimable pages?
4770 wmark = __zone_watermark_ok(zone, order, min_wmark,
4771 ac->highest_zoneidx, alloc_flags, available);
4772 trace_reclaim_retry_zone(z, order, reclaimable,
4773 available, min_wmark, *no_progress_loops, wmark);
4781 * Memory allocation/reclaim might be called from a WQ context and the
4782 * current implementation of the WQ concurrency control doesn't
4783 * recognize that a particular WQ is congested if the worker thread is
4784 * looping without ever sleeping. Therefore we have to do a short sleep
4785 * here rather than calling cond_resched().
4787 if (current->flags & PF_WQ_WORKER)
4788 schedule_timeout_uninterruptible(1);
4795 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4798 * It's possible that cpuset's mems_allowed and the nodemask from
4799 * mempolicy don't intersect. This should be normally dealt with by
4800 * policy_nodemask(), but it's possible to race with cpuset update in
4801 * such a way the check therein was true, and then it became false
4802 * before we got our cpuset_mems_cookie here.
4803 * This assumes that for all allocations, ac->nodemask can come only
4804 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4805 * when it does not intersect with the cpuset restrictions) or the
4806 * caller can deal with a violated nodemask.
4808 if (cpusets_enabled() && ac->nodemask &&
4809 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4810 ac->nodemask = NULL;
4815 * When updating a task's mems_allowed or mempolicy nodemask, it is
4816 * possible to race with parallel threads in such a way that our
4817 * allocation can fail while the mask is being updated. If we are about
4818 * to fail, check if the cpuset changed during allocation and if so,
4821 if (read_mems_allowed_retry(cpuset_mems_cookie))
4827 static inline struct page *
4828 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4829 struct alloc_context *ac)
4831 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4832 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4833 struct page *page = NULL;
4834 unsigned int alloc_flags;
4835 unsigned long did_some_progress;
4836 enum compact_priority compact_priority;
4837 enum compact_result compact_result;
4838 int compaction_retries;
4839 int no_progress_loops;
4840 unsigned int cpuset_mems_cookie;
4844 * We also sanity check to catch abuse of atomic reserves being used by
4845 * callers that are not in atomic context.
4847 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4848 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4849 gfp_mask &= ~__GFP_ATOMIC;
4852 compaction_retries = 0;
4853 no_progress_loops = 0;
4854 compact_priority = DEF_COMPACT_PRIORITY;
4855 cpuset_mems_cookie = read_mems_allowed_begin();
4858 * The fast path uses conservative alloc_flags to succeed only until
4859 * kswapd needs to be woken up, and to avoid the cost of setting up
4860 * alloc_flags precisely. So we do that now.
4862 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4865 * We need to recalculate the starting point for the zonelist iterator
4866 * because we might have used different nodemask in the fast path, or
4867 * there was a cpuset modification and we are retrying - otherwise we
4868 * could end up iterating over non-eligible zones endlessly.
4870 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4871 ac->highest_zoneidx, ac->nodemask);
4872 if (!ac->preferred_zoneref->zone)
4876 * Check for insane configurations where the cpuset doesn't contain
4877 * any suitable zone to satisfy the request - e.g. non-movable
4878 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4880 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4881 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4882 ac->highest_zoneidx,
4883 &cpuset_current_mems_allowed);
4888 if (alloc_flags & ALLOC_KSWAPD)
4889 wake_all_kswapds(order, gfp_mask, ac);
4892 * The adjusted alloc_flags might result in immediate success, so try
4895 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4900 * For costly allocations, try direct compaction first, as it's likely
4901 * that we have enough base pages and don't need to reclaim. For non-
4902 * movable high-order allocations, do that as well, as compaction will
4903 * try prevent permanent fragmentation by migrating from blocks of the
4905 * Don't try this for allocations that are allowed to ignore
4906 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4908 if (can_direct_reclaim &&
4910 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4911 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4912 page = __alloc_pages_direct_compact(gfp_mask, order,
4914 INIT_COMPACT_PRIORITY,
4920 * Checks for costly allocations with __GFP_NORETRY, which
4921 * includes some THP page fault allocations
4923 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4925 * If allocating entire pageblock(s) and compaction
4926 * failed because all zones are below low watermarks
4927 * or is prohibited because it recently failed at this
4928 * order, fail immediately unless the allocator has
4929 * requested compaction and reclaim retry.
4932 * - potentially very expensive because zones are far
4933 * below their low watermarks or this is part of very
4934 * bursty high order allocations,
4935 * - not guaranteed to help because isolate_freepages()
4936 * may not iterate over freed pages as part of its
4938 * - unlikely to make entire pageblocks free on its
4941 if (compact_result == COMPACT_SKIPPED ||
4942 compact_result == COMPACT_DEFERRED)
4946 * Looks like reclaim/compaction is worth trying, but
4947 * sync compaction could be very expensive, so keep
4948 * using async compaction.
4950 compact_priority = INIT_COMPACT_PRIORITY;
4955 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4956 if (alloc_flags & ALLOC_KSWAPD)
4957 wake_all_kswapds(order, gfp_mask, ac);
4959 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4961 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
4964 * Reset the nodemask and zonelist iterators if memory policies can be
4965 * ignored. These allocations are high priority and system rather than
4968 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4969 ac->nodemask = NULL;
4970 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4971 ac->highest_zoneidx, ac->nodemask);
4974 /* Attempt with potentially adjusted zonelist and alloc_flags */
4975 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4979 /* Caller is not willing to reclaim, we can't balance anything */
4980 if (!can_direct_reclaim)
4983 /* Avoid recursion of direct reclaim */
4984 if (current->flags & PF_MEMALLOC)
4987 /* Try direct reclaim and then allocating */
4988 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4989 &did_some_progress);
4993 /* Try direct compaction and then allocating */
4994 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4995 compact_priority, &compact_result);
4999 /* Do not loop if specifically requested */
5000 if (gfp_mask & __GFP_NORETRY)
5004 * Do not retry costly high order allocations unless they are
5005 * __GFP_RETRY_MAYFAIL
5007 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5010 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5011 did_some_progress > 0, &no_progress_loops))
5015 * It doesn't make any sense to retry for the compaction if the order-0
5016 * reclaim is not able to make any progress because the current
5017 * implementation of the compaction depends on the sufficient amount
5018 * of free memory (see __compaction_suitable)
5020 if (did_some_progress > 0 &&
5021 should_compact_retry(ac, order, alloc_flags,
5022 compact_result, &compact_priority,
5023 &compaction_retries))
5027 /* Deal with possible cpuset update races before we start OOM killing */
5028 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5031 /* Reclaim has failed us, start killing things */
5032 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5036 /* Avoid allocations with no watermarks from looping endlessly */
5037 if (tsk_is_oom_victim(current) &&
5038 (alloc_flags & ALLOC_OOM ||
5039 (gfp_mask & __GFP_NOMEMALLOC)))
5042 /* Retry as long as the OOM killer is making progress */
5043 if (did_some_progress) {
5044 no_progress_loops = 0;
5049 /* Deal with possible cpuset update races before we fail */
5050 if (check_retry_cpuset(cpuset_mems_cookie, ac))
5054 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5057 if (gfp_mask & __GFP_NOFAIL) {
5059 * All existing users of the __GFP_NOFAIL are blockable, so warn
5060 * of any new users that actually require GFP_NOWAIT
5062 if (WARN_ON_ONCE(!can_direct_reclaim))
5066 * PF_MEMALLOC request from this context is rather bizarre
5067 * because we cannot reclaim anything and only can loop waiting
5068 * for somebody to do a work for us
5070 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5073 * non failing costly orders are a hard requirement which we
5074 * are not prepared for much so let's warn about these users
5075 * so that we can identify them and convert them to something
5078 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5081 * Help non-failing allocations by giving them access to memory
5082 * reserves but do not use ALLOC_NO_WATERMARKS because this
5083 * could deplete whole memory reserves which would just make
5084 * the situation worse
5086 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5094 warn_alloc(gfp_mask, ac->nodemask,
5095 "page allocation failure: order:%u", order);
5100 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5101 int preferred_nid, nodemask_t *nodemask,
5102 struct alloc_context *ac, gfp_t *alloc_gfp,
5103 unsigned int *alloc_flags)
5105 ac->highest_zoneidx = gfp_zone(gfp_mask);
5106 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5107 ac->nodemask = nodemask;
5108 ac->migratetype = gfp_migratetype(gfp_mask);
5110 if (cpusets_enabled()) {
5111 *alloc_gfp |= __GFP_HARDWALL;
5113 * When we are in the interrupt context, it is irrelevant
5114 * to the current task context. It means that any node ok.
5116 if (in_task() && !ac->nodemask)
5117 ac->nodemask = &cpuset_current_mems_allowed;
5119 *alloc_flags |= ALLOC_CPUSET;
5122 fs_reclaim_acquire(gfp_mask);
5123 fs_reclaim_release(gfp_mask);
5125 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5127 if (should_fail_alloc_page(gfp_mask, order))
5130 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5132 /* Dirty zone balancing only done in the fast path */
5133 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5136 * The preferred zone is used for statistics but crucially it is
5137 * also used as the starting point for the zonelist iterator. It
5138 * may get reset for allocations that ignore memory policies.
5140 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5141 ac->highest_zoneidx, ac->nodemask);
5147 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5148 * @gfp: GFP flags for the allocation
5149 * @preferred_nid: The preferred NUMA node ID to allocate from
5150 * @nodemask: Set of nodes to allocate from, may be NULL
5151 * @nr_pages: The number of pages desired on the list or array
5152 * @page_list: Optional list to store the allocated pages
5153 * @page_array: Optional array to store the pages
5155 * This is a batched version of the page allocator that attempts to
5156 * allocate nr_pages quickly. Pages are added to page_list if page_list
5157 * is not NULL, otherwise it is assumed that the page_array is valid.
5159 * For lists, nr_pages is the number of pages that should be allocated.
5161 * For arrays, only NULL elements are populated with pages and nr_pages
5162 * is the maximum number of pages that will be stored in the array.
5164 * Returns the number of pages on the list or array.
5166 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5167 nodemask_t *nodemask, int nr_pages,
5168 struct list_head *page_list,
5169 struct page **page_array)
5172 unsigned long flags;
5175 struct per_cpu_pages *pcp;
5176 struct list_head *pcp_list;
5177 struct alloc_context ac;
5179 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5180 int nr_populated = 0, nr_account = 0;
5183 * Skip populated array elements to determine if any pages need
5184 * to be allocated before disabling IRQs.
5186 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5189 /* No pages requested? */
5190 if (unlikely(nr_pages <= 0))
5193 /* Already populated array? */
5194 if (unlikely(page_array && nr_pages - nr_populated == 0))
5197 /* Bulk allocator does not support memcg accounting. */
5198 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5201 /* Use the single page allocator for one page. */
5202 if (nr_pages - nr_populated == 1)
5205 #ifdef CONFIG_PAGE_OWNER
5207 * PAGE_OWNER may recurse into the allocator to allocate space to
5208 * save the stack with pagesets.lock held. Releasing/reacquiring
5209 * removes much of the performance benefit of bulk allocation so
5210 * force the caller to allocate one page at a time as it'll have
5211 * similar performance to added complexity to the bulk allocator.
5213 if (static_branch_unlikely(&page_owner_inited))
5217 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5218 gfp &= gfp_allowed_mask;
5220 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5224 /* Find an allowed local zone that meets the low watermark. */
5225 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5228 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5229 !__cpuset_zone_allowed(zone, gfp)) {
5233 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5234 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5238 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5239 if (zone_watermark_fast(zone, 0, mark,
5240 zonelist_zone_idx(ac.preferred_zoneref),
5241 alloc_flags, gfp)) {
5247 * If there are no allowed local zones that meets the watermarks then
5248 * try to allocate a single page and reclaim if necessary.
5250 if (unlikely(!zone))
5253 /* Attempt the batch allocation */
5254 local_lock_irqsave(&pagesets.lock, flags);
5255 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5256 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5258 while (nr_populated < nr_pages) {
5260 /* Skip existing pages */
5261 if (page_array && page_array[nr_populated]) {
5266 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5268 if (unlikely(!page)) {
5269 /* Try and get at least one page */
5276 prep_new_page(page, 0, gfp, 0);
5278 list_add(&page->lru, page_list);
5280 page_array[nr_populated] = page;
5284 local_unlock_irqrestore(&pagesets.lock, flags);
5286 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5287 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5290 return nr_populated;
5293 local_unlock_irqrestore(&pagesets.lock, flags);
5296 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5299 list_add(&page->lru, page_list);
5301 page_array[nr_populated] = page;
5307 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5310 * This is the 'heart' of the zoned buddy allocator.
5312 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5313 nodemask_t *nodemask)
5316 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5317 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5318 struct alloc_context ac = { };
5321 * There are several places where we assume that the order value is sane
5322 * so bail out early if the request is out of bound.
5324 if (unlikely(order >= MAX_ORDER)) {
5325 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5329 gfp &= gfp_allowed_mask;
5331 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5332 * resp. GFP_NOIO which has to be inherited for all allocation requests
5333 * from a particular context which has been marked by
5334 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5335 * movable zones are not used during allocation.
5337 gfp = current_gfp_context(gfp);
5339 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5340 &alloc_gfp, &alloc_flags))
5344 * Forbid the first pass from falling back to types that fragment
5345 * memory until all local zones are considered.
5347 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5349 /* First allocation attempt */
5350 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5355 ac.spread_dirty_pages = false;
5358 * Restore the original nodemask if it was potentially replaced with
5359 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5361 ac.nodemask = nodemask;
5363 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5366 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5367 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5368 __free_pages(page, order);
5372 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5376 EXPORT_SYMBOL(__alloc_pages);
5378 struct folio *__folio_alloc(gfp_t gfp, unsigned int order, int preferred_nid,
5379 nodemask_t *nodemask)
5381 struct page *page = __alloc_pages(gfp | __GFP_COMP, order,
5382 preferred_nid, nodemask);
5384 if (page && order > 1)
5385 prep_transhuge_page(page);
5386 return (struct folio *)page;
5388 EXPORT_SYMBOL(__folio_alloc);
5391 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5392 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5393 * you need to access high mem.
5395 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5399 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5402 return (unsigned long) page_address(page);
5404 EXPORT_SYMBOL(__get_free_pages);
5406 unsigned long get_zeroed_page(gfp_t gfp_mask)
5408 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5410 EXPORT_SYMBOL(get_zeroed_page);
5413 * __free_pages - Free pages allocated with alloc_pages().
5414 * @page: The page pointer returned from alloc_pages().
5415 * @order: The order of the allocation.
5417 * This function can free multi-page allocations that are not compound
5418 * pages. It does not check that the @order passed in matches that of
5419 * the allocation, so it is easy to leak memory. Freeing more memory
5420 * than was allocated will probably emit a warning.
5422 * If the last reference to this page is speculative, it will be released
5423 * by put_page() which only frees the first page of a non-compound
5424 * allocation. To prevent the remaining pages from being leaked, we free
5425 * the subsequent pages here. If you want to use the page's reference
5426 * count to decide when to free the allocation, you should allocate a
5427 * compound page, and use put_page() instead of __free_pages().
5429 * Context: May be called in interrupt context or while holding a normal
5430 * spinlock, but not in NMI context or while holding a raw spinlock.
5432 void __free_pages(struct page *page, unsigned int order)
5434 if (put_page_testzero(page))
5435 free_the_page(page, order);
5436 else if (!PageHead(page))
5438 free_the_page(page + (1 << order), order);
5440 EXPORT_SYMBOL(__free_pages);
5442 void free_pages(unsigned long addr, unsigned int order)
5445 VM_BUG_ON(!virt_addr_valid((void *)addr));
5446 __free_pages(virt_to_page((void *)addr), order);
5450 EXPORT_SYMBOL(free_pages);
5454 * An arbitrary-length arbitrary-offset area of memory which resides
5455 * within a 0 or higher order page. Multiple fragments within that page
5456 * are individually refcounted, in the page's reference counter.
5458 * The page_frag functions below provide a simple allocation framework for
5459 * page fragments. This is used by the network stack and network device
5460 * drivers to provide a backing region of memory for use as either an
5461 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5463 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5466 struct page *page = NULL;
5467 gfp_t gfp = gfp_mask;
5469 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5470 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5472 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5473 PAGE_FRAG_CACHE_MAX_ORDER);
5474 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5476 if (unlikely(!page))
5477 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5479 nc->va = page ? page_address(page) : NULL;
5484 void __page_frag_cache_drain(struct page *page, unsigned int count)
5486 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5488 if (page_ref_sub_and_test(page, count))
5489 free_the_page(page, compound_order(page));
5491 EXPORT_SYMBOL(__page_frag_cache_drain);
5493 void *page_frag_alloc_align(struct page_frag_cache *nc,
5494 unsigned int fragsz, gfp_t gfp_mask,
5495 unsigned int align_mask)
5497 unsigned int size = PAGE_SIZE;
5501 if (unlikely(!nc->va)) {
5503 page = __page_frag_cache_refill(nc, gfp_mask);
5507 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5508 /* if size can vary use size else just use PAGE_SIZE */
5511 /* Even if we own the page, we do not use atomic_set().
5512 * This would break get_page_unless_zero() users.
5514 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5516 /* reset page count bias and offset to start of new frag */
5517 nc->pfmemalloc = page_is_pfmemalloc(page);
5518 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5522 offset = nc->offset - fragsz;
5523 if (unlikely(offset < 0)) {
5524 page = virt_to_page(nc->va);
5526 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5529 if (unlikely(nc->pfmemalloc)) {
5530 free_the_page(page, compound_order(page));
5534 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5535 /* if size can vary use size else just use PAGE_SIZE */
5538 /* OK, page count is 0, we can safely set it */
5539 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5541 /* reset page count bias and offset to start of new frag */
5542 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5543 offset = size - fragsz;
5547 offset &= align_mask;
5548 nc->offset = offset;
5550 return nc->va + offset;
5552 EXPORT_SYMBOL(page_frag_alloc_align);
5555 * Frees a page fragment allocated out of either a compound or order 0 page.
5557 void page_frag_free(void *addr)
5559 struct page *page = virt_to_head_page(addr);
5561 if (unlikely(put_page_testzero(page)))
5562 free_the_page(page, compound_order(page));
5564 EXPORT_SYMBOL(page_frag_free);
5566 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5570 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5571 unsigned long used = addr + PAGE_ALIGN(size);
5573 split_page(virt_to_page((void *)addr), order);
5574 while (used < alloc_end) {
5579 return (void *)addr;
5583 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5584 * @size: the number of bytes to allocate
5585 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5587 * This function is similar to alloc_pages(), except that it allocates the
5588 * minimum number of pages to satisfy the request. alloc_pages() can only
5589 * allocate memory in power-of-two pages.
5591 * This function is also limited by MAX_ORDER.
5593 * Memory allocated by this function must be released by free_pages_exact().
5595 * Return: pointer to the allocated area or %NULL in case of error.
5597 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5599 unsigned int order = get_order(size);
5602 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5603 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5605 addr = __get_free_pages(gfp_mask, order);
5606 return make_alloc_exact(addr, order, size);
5608 EXPORT_SYMBOL(alloc_pages_exact);
5611 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5613 * @nid: the preferred node ID where memory should be allocated
5614 * @size: the number of bytes to allocate
5615 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5617 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5620 * Return: pointer to the allocated area or %NULL in case of error.
5622 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5624 unsigned int order = get_order(size);
5627 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
5628 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
5630 p = alloc_pages_node(nid, gfp_mask, order);
5633 return make_alloc_exact((unsigned long)page_address(p), order, size);
5637 * free_pages_exact - release memory allocated via alloc_pages_exact()
5638 * @virt: the value returned by alloc_pages_exact.
5639 * @size: size of allocation, same value as passed to alloc_pages_exact().
5641 * Release the memory allocated by a previous call to alloc_pages_exact.
5643 void free_pages_exact(void *virt, size_t size)
5645 unsigned long addr = (unsigned long)virt;
5646 unsigned long end = addr + PAGE_ALIGN(size);
5648 while (addr < end) {
5653 EXPORT_SYMBOL(free_pages_exact);
5656 * nr_free_zone_pages - count number of pages beyond high watermark
5657 * @offset: The zone index of the highest zone
5659 * nr_free_zone_pages() counts the number of pages which are beyond the
5660 * high watermark within all zones at or below a given zone index. For each
5661 * zone, the number of pages is calculated as:
5663 * nr_free_zone_pages = managed_pages - high_pages
5665 * Return: number of pages beyond high watermark.
5667 static unsigned long nr_free_zone_pages(int offset)
5672 /* Just pick one node, since fallback list is circular */
5673 unsigned long sum = 0;
5675 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5677 for_each_zone_zonelist(zone, z, zonelist, offset) {
5678 unsigned long size = zone_managed_pages(zone);
5679 unsigned long high = high_wmark_pages(zone);
5688 * nr_free_buffer_pages - count number of pages beyond high watermark
5690 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5691 * watermark within ZONE_DMA and ZONE_NORMAL.
5693 * Return: number of pages beyond high watermark within ZONE_DMA and
5696 unsigned long nr_free_buffer_pages(void)
5698 return nr_free_zone_pages(gfp_zone(GFP_USER));
5700 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5702 static inline void show_node(struct zone *zone)
5704 if (IS_ENABLED(CONFIG_NUMA))
5705 printk("Node %d ", zone_to_nid(zone));
5708 long si_mem_available(void)
5711 unsigned long pagecache;
5712 unsigned long wmark_low = 0;
5713 unsigned long pages[NR_LRU_LISTS];
5714 unsigned long reclaimable;
5718 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5719 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5722 wmark_low += low_wmark_pages(zone);
5725 * Estimate the amount of memory available for userspace allocations,
5726 * without causing swapping.
5728 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5731 * Not all the page cache can be freed, otherwise the system will
5732 * start swapping. Assume at least half of the page cache, or the
5733 * low watermark worth of cache, needs to stay.
5735 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5736 pagecache -= min(pagecache / 2, wmark_low);
5737 available += pagecache;
5740 * Part of the reclaimable slab and other kernel memory consists of
5741 * items that are in use, and cannot be freed. Cap this estimate at the
5744 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5745 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5746 available += reclaimable - min(reclaimable / 2, wmark_low);
5752 EXPORT_SYMBOL_GPL(si_mem_available);
5754 void si_meminfo(struct sysinfo *val)
5756 val->totalram = totalram_pages();
5757 val->sharedram = global_node_page_state(NR_SHMEM);
5758 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5759 val->bufferram = nr_blockdev_pages();
5760 val->totalhigh = totalhigh_pages();
5761 val->freehigh = nr_free_highpages();
5762 val->mem_unit = PAGE_SIZE;
5765 EXPORT_SYMBOL(si_meminfo);
5768 void si_meminfo_node(struct sysinfo *val, int nid)
5770 int zone_type; /* needs to be signed */
5771 unsigned long managed_pages = 0;
5772 unsigned long managed_highpages = 0;
5773 unsigned long free_highpages = 0;
5774 pg_data_t *pgdat = NODE_DATA(nid);
5776 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5777 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5778 val->totalram = managed_pages;
5779 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5780 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5781 #ifdef CONFIG_HIGHMEM
5782 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5783 struct zone *zone = &pgdat->node_zones[zone_type];
5785 if (is_highmem(zone)) {
5786 managed_highpages += zone_managed_pages(zone);
5787 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5790 val->totalhigh = managed_highpages;
5791 val->freehigh = free_highpages;
5793 val->totalhigh = managed_highpages;
5794 val->freehigh = free_highpages;
5796 val->mem_unit = PAGE_SIZE;
5801 * Determine whether the node should be displayed or not, depending on whether
5802 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5804 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5806 if (!(flags & SHOW_MEM_FILTER_NODES))
5810 * no node mask - aka implicit memory numa policy. Do not bother with
5811 * the synchronization - read_mems_allowed_begin - because we do not
5812 * have to be precise here.
5815 nodemask = &cpuset_current_mems_allowed;
5817 return !node_isset(nid, *nodemask);
5820 #define K(x) ((x) << (PAGE_SHIFT-10))
5822 static void show_migration_types(unsigned char type)
5824 static const char types[MIGRATE_TYPES] = {
5825 [MIGRATE_UNMOVABLE] = 'U',
5826 [MIGRATE_MOVABLE] = 'M',
5827 [MIGRATE_RECLAIMABLE] = 'E',
5828 [MIGRATE_HIGHATOMIC] = 'H',
5830 [MIGRATE_CMA] = 'C',
5832 #ifdef CONFIG_MEMORY_ISOLATION
5833 [MIGRATE_ISOLATE] = 'I',
5836 char tmp[MIGRATE_TYPES + 1];
5840 for (i = 0; i < MIGRATE_TYPES; i++) {
5841 if (type & (1 << i))
5846 printk(KERN_CONT "(%s) ", tmp);
5850 * Show free area list (used inside shift_scroll-lock stuff)
5851 * We also calculate the percentage fragmentation. We do this by counting the
5852 * memory on each free list with the exception of the first item on the list.
5855 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5858 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5860 unsigned long free_pcp = 0;
5865 for_each_populated_zone(zone) {
5866 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5869 for_each_online_cpu(cpu)
5870 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5873 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5874 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5875 " unevictable:%lu dirty:%lu writeback:%lu\n"
5876 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5877 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5878 " kernel_misc_reclaimable:%lu\n"
5879 " free:%lu free_pcp:%lu free_cma:%lu\n",
5880 global_node_page_state(NR_ACTIVE_ANON),
5881 global_node_page_state(NR_INACTIVE_ANON),
5882 global_node_page_state(NR_ISOLATED_ANON),
5883 global_node_page_state(NR_ACTIVE_FILE),
5884 global_node_page_state(NR_INACTIVE_FILE),
5885 global_node_page_state(NR_ISOLATED_FILE),
5886 global_node_page_state(NR_UNEVICTABLE),
5887 global_node_page_state(NR_FILE_DIRTY),
5888 global_node_page_state(NR_WRITEBACK),
5889 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5890 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5891 global_node_page_state(NR_FILE_MAPPED),
5892 global_node_page_state(NR_SHMEM),
5893 global_node_page_state(NR_PAGETABLE),
5894 global_zone_page_state(NR_BOUNCE),
5895 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5896 global_zone_page_state(NR_FREE_PAGES),
5898 global_zone_page_state(NR_FREE_CMA_PAGES));
5900 for_each_online_pgdat(pgdat) {
5901 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5905 " active_anon:%lukB"
5906 " inactive_anon:%lukB"
5907 " active_file:%lukB"
5908 " inactive_file:%lukB"
5909 " unevictable:%lukB"
5910 " isolated(anon):%lukB"
5911 " isolated(file):%lukB"
5916 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5918 " shmem_pmdmapped: %lukB"
5921 " writeback_tmp:%lukB"
5922 " kernel_stack:%lukB"
5923 #ifdef CONFIG_SHADOW_CALL_STACK
5924 " shadow_call_stack:%lukB"
5927 " all_unreclaimable? %s"
5930 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5931 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5932 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5933 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5934 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5935 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5936 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5937 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5938 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5939 K(node_page_state(pgdat, NR_WRITEBACK)),
5940 K(node_page_state(pgdat, NR_SHMEM)),
5941 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5942 K(node_page_state(pgdat, NR_SHMEM_THPS)),
5943 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
5944 K(node_page_state(pgdat, NR_ANON_THPS)),
5946 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5947 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5948 #ifdef CONFIG_SHADOW_CALL_STACK
5949 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5951 K(node_page_state(pgdat, NR_PAGETABLE)),
5952 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5956 for_each_populated_zone(zone) {
5959 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5963 for_each_online_cpu(cpu)
5964 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5974 " reserved_highatomic:%luKB"
5975 " active_anon:%lukB"
5976 " inactive_anon:%lukB"
5977 " active_file:%lukB"
5978 " inactive_file:%lukB"
5979 " unevictable:%lukB"
5980 " writepending:%lukB"
5990 K(zone_page_state(zone, NR_FREE_PAGES)),
5991 K(zone->watermark_boost),
5992 K(min_wmark_pages(zone)),
5993 K(low_wmark_pages(zone)),
5994 K(high_wmark_pages(zone)),
5995 K(zone->nr_reserved_highatomic),
5996 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5997 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5998 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5999 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6000 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6001 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6002 K(zone->present_pages),
6003 K(zone_managed_pages(zone)),
6004 K(zone_page_state(zone, NR_MLOCK)),
6005 K(zone_page_state(zone, NR_BOUNCE)),
6007 K(this_cpu_read(zone->per_cpu_pageset->count)),
6008 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6009 printk("lowmem_reserve[]:");
6010 for (i = 0; i < MAX_NR_ZONES; i++)
6011 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6012 printk(KERN_CONT "\n");
6015 for_each_populated_zone(zone) {
6017 unsigned long nr[MAX_ORDER], flags, total = 0;
6018 unsigned char types[MAX_ORDER];
6020 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6023 printk(KERN_CONT "%s: ", zone->name);
6025 spin_lock_irqsave(&zone->lock, flags);
6026 for (order = 0; order < MAX_ORDER; order++) {
6027 struct free_area *area = &zone->free_area[order];
6030 nr[order] = area->nr_free;
6031 total += nr[order] << order;
6034 for (type = 0; type < MIGRATE_TYPES; type++) {
6035 if (!free_area_empty(area, type))
6036 types[order] |= 1 << type;
6039 spin_unlock_irqrestore(&zone->lock, flags);
6040 for (order = 0; order < MAX_ORDER; order++) {
6041 printk(KERN_CONT "%lu*%lukB ",
6042 nr[order], K(1UL) << order);
6044 show_migration_types(types[order]);
6046 printk(KERN_CONT "= %lukB\n", K(total));
6049 hugetlb_show_meminfo();
6051 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6053 show_swap_cache_info();
6056 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6058 zoneref->zone = zone;
6059 zoneref->zone_idx = zone_idx(zone);
6063 * Builds allocation fallback zone lists.
6065 * Add all populated zones of a node to the zonelist.
6067 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6070 enum zone_type zone_type = MAX_NR_ZONES;
6075 zone = pgdat->node_zones + zone_type;
6076 if (managed_zone(zone)) {
6077 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6078 check_highest_zone(zone_type);
6080 } while (zone_type);
6087 static int __parse_numa_zonelist_order(char *s)
6090 * We used to support different zonelists modes but they turned
6091 * out to be just not useful. Let's keep the warning in place
6092 * if somebody still use the cmd line parameter so that we do
6093 * not fail it silently
6095 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6096 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6102 char numa_zonelist_order[] = "Node";
6105 * sysctl handler for numa_zonelist_order
6107 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6108 void *buffer, size_t *length, loff_t *ppos)
6111 return __parse_numa_zonelist_order(buffer);
6112 return proc_dostring(table, write, buffer, length, ppos);
6116 #define MAX_NODE_LOAD (nr_online_nodes)
6117 static int node_load[MAX_NUMNODES];
6120 * find_next_best_node - find the next node that should appear in a given node's fallback list
6121 * @node: node whose fallback list we're appending
6122 * @used_node_mask: nodemask_t of already used nodes
6124 * We use a number of factors to determine which is the next node that should
6125 * appear on a given node's fallback list. The node should not have appeared
6126 * already in @node's fallback list, and it should be the next closest node
6127 * according to the distance array (which contains arbitrary distance values
6128 * from each node to each node in the system), and should also prefer nodes
6129 * with no CPUs, since presumably they'll have very little allocation pressure
6130 * on them otherwise.
6132 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6134 int find_next_best_node(int node, nodemask_t *used_node_mask)
6137 int min_val = INT_MAX;
6138 int best_node = NUMA_NO_NODE;
6140 /* Use the local node if we haven't already */
6141 if (!node_isset(node, *used_node_mask)) {
6142 node_set(node, *used_node_mask);
6146 for_each_node_state(n, N_MEMORY) {
6148 /* Don't want a node to appear more than once */
6149 if (node_isset(n, *used_node_mask))
6152 /* Use the distance array to find the distance */
6153 val = node_distance(node, n);
6155 /* Penalize nodes under us ("prefer the next node") */
6158 /* Give preference to headless and unused nodes */
6159 if (!cpumask_empty(cpumask_of_node(n)))
6160 val += PENALTY_FOR_NODE_WITH_CPUS;
6162 /* Slight preference for less loaded node */
6163 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6164 val += node_load[n];
6166 if (val < min_val) {
6173 node_set(best_node, *used_node_mask);
6180 * Build zonelists ordered by node and zones within node.
6181 * This results in maximum locality--normal zone overflows into local
6182 * DMA zone, if any--but risks exhausting DMA zone.
6184 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6187 struct zoneref *zonerefs;
6190 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6192 for (i = 0; i < nr_nodes; i++) {
6195 pg_data_t *node = NODE_DATA(node_order[i]);
6197 nr_zones = build_zonerefs_node(node, zonerefs);
6198 zonerefs += nr_zones;
6200 zonerefs->zone = NULL;
6201 zonerefs->zone_idx = 0;
6205 * Build gfp_thisnode zonelists
6207 static void build_thisnode_zonelists(pg_data_t *pgdat)
6209 struct zoneref *zonerefs;
6212 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6213 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6214 zonerefs += nr_zones;
6215 zonerefs->zone = NULL;
6216 zonerefs->zone_idx = 0;
6220 * Build zonelists ordered by zone and nodes within zones.
6221 * This results in conserving DMA zone[s] until all Normal memory is
6222 * exhausted, but results in overflowing to remote node while memory
6223 * may still exist in local DMA zone.
6226 static void build_zonelists(pg_data_t *pgdat)
6228 static int node_order[MAX_NUMNODES];
6229 int node, load, nr_nodes = 0;
6230 nodemask_t used_mask = NODE_MASK_NONE;
6231 int local_node, prev_node;
6233 /* NUMA-aware ordering of nodes */
6234 local_node = pgdat->node_id;
6235 load = nr_online_nodes;
6236 prev_node = local_node;
6238 memset(node_order, 0, sizeof(node_order));
6239 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6241 * We don't want to pressure a particular node.
6242 * So adding penalty to the first node in same
6243 * distance group to make it round-robin.
6245 if (node_distance(local_node, node) !=
6246 node_distance(local_node, prev_node))
6247 node_load[node] += load;
6249 node_order[nr_nodes++] = node;
6254 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6255 build_thisnode_zonelists(pgdat);
6256 pr_info("Fallback order for Node %d: ", local_node);
6257 for (node = 0; node < nr_nodes; node++)
6258 pr_cont("%d ", node_order[node]);
6262 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6264 * Return node id of node used for "local" allocations.
6265 * I.e., first node id of first zone in arg node's generic zonelist.
6266 * Used for initializing percpu 'numa_mem', which is used primarily
6267 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6269 int local_memory_node(int node)
6273 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6274 gfp_zone(GFP_KERNEL),
6276 return zone_to_nid(z->zone);
6280 static void setup_min_unmapped_ratio(void);
6281 static void setup_min_slab_ratio(void);
6282 #else /* CONFIG_NUMA */
6284 static void build_zonelists(pg_data_t *pgdat)
6286 int node, local_node;
6287 struct zoneref *zonerefs;
6290 local_node = pgdat->node_id;
6292 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6293 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6294 zonerefs += nr_zones;
6297 * Now we build the zonelist so that it contains the zones
6298 * of all the other nodes.
6299 * We don't want to pressure a particular node, so when
6300 * building the zones for node N, we make sure that the
6301 * zones coming right after the local ones are those from
6302 * node N+1 (modulo N)
6304 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6305 if (!node_online(node))
6307 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6308 zonerefs += nr_zones;
6310 for (node = 0; node < local_node; node++) {
6311 if (!node_online(node))
6313 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6314 zonerefs += nr_zones;
6317 zonerefs->zone = NULL;
6318 zonerefs->zone_idx = 0;
6321 #endif /* CONFIG_NUMA */
6324 * Boot pageset table. One per cpu which is going to be used for all
6325 * zones and all nodes. The parameters will be set in such a way
6326 * that an item put on a list will immediately be handed over to
6327 * the buddy list. This is safe since pageset manipulation is done
6328 * with interrupts disabled.
6330 * The boot_pagesets must be kept even after bootup is complete for
6331 * unused processors and/or zones. They do play a role for bootstrapping
6332 * hotplugged processors.
6334 * zoneinfo_show() and maybe other functions do
6335 * not check if the processor is online before following the pageset pointer.
6336 * Other parts of the kernel may not check if the zone is available.
6338 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6339 /* These effectively disable the pcplists in the boot pageset completely */
6340 #define BOOT_PAGESET_HIGH 0
6341 #define BOOT_PAGESET_BATCH 1
6342 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6343 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6344 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6346 static void __build_all_zonelists(void *data)
6349 int __maybe_unused cpu;
6350 pg_data_t *self = data;
6351 static DEFINE_SPINLOCK(lock);
6356 memset(node_load, 0, sizeof(node_load));
6360 * This node is hotadded and no memory is yet present. So just
6361 * building zonelists is fine - no need to touch other nodes.
6363 if (self && !node_online(self->node_id)) {
6364 build_zonelists(self);
6366 for_each_online_node(nid) {
6367 pg_data_t *pgdat = NODE_DATA(nid);
6369 build_zonelists(pgdat);
6372 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6374 * We now know the "local memory node" for each node--
6375 * i.e., the node of the first zone in the generic zonelist.
6376 * Set up numa_mem percpu variable for on-line cpus. During
6377 * boot, only the boot cpu should be on-line; we'll init the
6378 * secondary cpus' numa_mem as they come on-line. During
6379 * node/memory hotplug, we'll fixup all on-line cpus.
6381 for_each_online_cpu(cpu)
6382 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6389 static noinline void __init
6390 build_all_zonelists_init(void)
6394 __build_all_zonelists(NULL);
6397 * Initialize the boot_pagesets that are going to be used
6398 * for bootstrapping processors. The real pagesets for
6399 * each zone will be allocated later when the per cpu
6400 * allocator is available.
6402 * boot_pagesets are used also for bootstrapping offline
6403 * cpus if the system is already booted because the pagesets
6404 * are needed to initialize allocators on a specific cpu too.
6405 * F.e. the percpu allocator needs the page allocator which
6406 * needs the percpu allocator in order to allocate its pagesets
6407 * (a chicken-egg dilemma).
6409 for_each_possible_cpu(cpu)
6410 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6412 mminit_verify_zonelist();
6413 cpuset_init_current_mems_allowed();
6417 * unless system_state == SYSTEM_BOOTING.
6419 * __ref due to call of __init annotated helper build_all_zonelists_init
6420 * [protected by SYSTEM_BOOTING].
6422 void __ref build_all_zonelists(pg_data_t *pgdat)
6424 unsigned long vm_total_pages;
6426 if (system_state == SYSTEM_BOOTING) {
6427 build_all_zonelists_init();
6429 __build_all_zonelists(pgdat);
6430 /* cpuset refresh routine should be here */
6432 /* Get the number of free pages beyond high watermark in all zones. */
6433 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6435 * Disable grouping by mobility if the number of pages in the
6436 * system is too low to allow the mechanism to work. It would be
6437 * more accurate, but expensive to check per-zone. This check is
6438 * made on memory-hotadd so a system can start with mobility
6439 * disabled and enable it later
6441 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6442 page_group_by_mobility_disabled = 1;
6444 page_group_by_mobility_disabled = 0;
6446 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6448 page_group_by_mobility_disabled ? "off" : "on",
6451 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6455 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6456 static bool __meminit
6457 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6459 static struct memblock_region *r;
6461 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6462 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6463 for_each_mem_region(r) {
6464 if (*pfn < memblock_region_memory_end_pfn(r))
6468 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6469 memblock_is_mirror(r)) {
6470 *pfn = memblock_region_memory_end_pfn(r);
6478 * Initially all pages are reserved - free ones are freed
6479 * up by memblock_free_all() once the early boot process is
6480 * done. Non-atomic initialization, single-pass.
6482 * All aligned pageblocks are initialized to the specified migratetype
6483 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6484 * zone stats (e.g., nr_isolate_pageblock) are touched.
6486 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6487 unsigned long start_pfn, unsigned long zone_end_pfn,
6488 enum meminit_context context,
6489 struct vmem_altmap *altmap, int migratetype)
6491 unsigned long pfn, end_pfn = start_pfn + size;
6494 if (highest_memmap_pfn < end_pfn - 1)
6495 highest_memmap_pfn = end_pfn - 1;
6497 #ifdef CONFIG_ZONE_DEVICE
6499 * Honor reservation requested by the driver for this ZONE_DEVICE
6500 * memory. We limit the total number of pages to initialize to just
6501 * those that might contain the memory mapping. We will defer the
6502 * ZONE_DEVICE page initialization until after we have released
6505 if (zone == ZONE_DEVICE) {
6509 if (start_pfn == altmap->base_pfn)
6510 start_pfn += altmap->reserve;
6511 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6515 for (pfn = start_pfn; pfn < end_pfn; ) {
6517 * There can be holes in boot-time mem_map[]s handed to this
6518 * function. They do not exist on hotplugged memory.
6520 if (context == MEMINIT_EARLY) {
6521 if (overlap_memmap_init(zone, &pfn))
6523 if (defer_init(nid, pfn, zone_end_pfn))
6527 page = pfn_to_page(pfn);
6528 __init_single_page(page, pfn, zone, nid);
6529 if (context == MEMINIT_HOTPLUG)
6530 __SetPageReserved(page);
6533 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6534 * such that unmovable allocations won't be scattered all
6535 * over the place during system boot.
6537 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6538 set_pageblock_migratetype(page, migratetype);
6545 #ifdef CONFIG_ZONE_DEVICE
6546 static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
6547 unsigned long zone_idx, int nid,
6548 struct dev_pagemap *pgmap)
6551 __init_single_page(page, pfn, zone_idx, nid);
6554 * Mark page reserved as it will need to wait for onlining
6555 * phase for it to be fully associated with a zone.
6557 * We can use the non-atomic __set_bit operation for setting
6558 * the flag as we are still initializing the pages.
6560 __SetPageReserved(page);
6563 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6564 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6565 * ever freed or placed on a driver-private list.
6567 page->pgmap = pgmap;
6568 page->zone_device_data = NULL;
6571 * Mark the block movable so that blocks are reserved for
6572 * movable at startup. This will force kernel allocations
6573 * to reserve their blocks rather than leaking throughout
6574 * the address space during boot when many long-lived
6575 * kernel allocations are made.
6577 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6578 * because this is done early in section_activate()
6580 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6581 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6586 static void __ref memmap_init_compound(struct page *head,
6587 unsigned long head_pfn,
6588 unsigned long zone_idx, int nid,
6589 struct dev_pagemap *pgmap,
6590 unsigned long nr_pages)
6592 unsigned long pfn, end_pfn = head_pfn + nr_pages;
6593 unsigned int order = pgmap->vmemmap_shift;
6595 __SetPageHead(head);
6596 for (pfn = head_pfn + 1; pfn < end_pfn; pfn++) {
6597 struct page *page = pfn_to_page(pfn);
6599 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6600 prep_compound_tail(head, pfn - head_pfn);
6601 set_page_count(page, 0);
6604 * The first tail page stores compound_mapcount_ptr() and
6605 * compound_order() and the second tail page stores
6606 * compound_pincount_ptr(). Call prep_compound_head() after
6607 * the first and second tail pages have been initialized to
6608 * not have the data overwritten.
6610 if (pfn == head_pfn + 2)
6611 prep_compound_head(head, order);
6615 void __ref memmap_init_zone_device(struct zone *zone,
6616 unsigned long start_pfn,
6617 unsigned long nr_pages,
6618 struct dev_pagemap *pgmap)
6620 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6621 struct pglist_data *pgdat = zone->zone_pgdat;
6622 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6623 unsigned int pfns_per_compound = pgmap_vmemmap_nr(pgmap);
6624 unsigned long zone_idx = zone_idx(zone);
6625 unsigned long start = jiffies;
6626 int nid = pgdat->node_id;
6628 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6632 * The call to memmap_init should have already taken care
6633 * of the pages reserved for the memmap, so we can just jump to
6634 * the end of that region and start processing the device pages.
6637 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6638 nr_pages = end_pfn - start_pfn;
6641 for (pfn = start_pfn; pfn < end_pfn; pfn += pfns_per_compound) {
6642 struct page *page = pfn_to_page(pfn);
6644 __init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
6646 if (pfns_per_compound == 1)
6649 memmap_init_compound(page, pfn, zone_idx, nid, pgmap,
6653 pr_info("%s initialised %lu pages in %ums\n", __func__,
6654 nr_pages, jiffies_to_msecs(jiffies - start));
6658 static void __meminit zone_init_free_lists(struct zone *zone)
6660 unsigned int order, t;
6661 for_each_migratetype_order(order, t) {
6662 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6663 zone->free_area[order].nr_free = 0;
6668 * Only struct pages that correspond to ranges defined by memblock.memory
6669 * are zeroed and initialized by going through __init_single_page() during
6670 * memmap_init_zone_range().
6672 * But, there could be struct pages that correspond to holes in
6673 * memblock.memory. This can happen because of the following reasons:
6674 * - physical memory bank size is not necessarily the exact multiple of the
6675 * arbitrary section size
6676 * - early reserved memory may not be listed in memblock.memory
6677 * - memory layouts defined with memmap= kernel parameter may not align
6678 * nicely with memmap sections
6680 * Explicitly initialize those struct pages so that:
6681 * - PG_Reserved is set
6682 * - zone and node links point to zone and node that span the page if the
6683 * hole is in the middle of a zone
6684 * - zone and node links point to adjacent zone/node if the hole falls on
6685 * the zone boundary; the pages in such holes will be prepended to the
6686 * zone/node above the hole except for the trailing pages in the last
6687 * section that will be appended to the zone/node below.
6689 static void __init init_unavailable_range(unsigned long spfn,
6696 for (pfn = spfn; pfn < epfn; pfn++) {
6697 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6698 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6699 + pageblock_nr_pages - 1;
6702 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6703 __SetPageReserved(pfn_to_page(pfn));
6708 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6709 node, zone_names[zone], pgcnt);
6712 static void __init memmap_init_zone_range(struct zone *zone,
6713 unsigned long start_pfn,
6714 unsigned long end_pfn,
6715 unsigned long *hole_pfn)
6717 unsigned long zone_start_pfn = zone->zone_start_pfn;
6718 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6719 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6721 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6722 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6724 if (start_pfn >= end_pfn)
6727 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6728 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6730 if (*hole_pfn < start_pfn)
6731 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6733 *hole_pfn = end_pfn;
6736 static void __init memmap_init(void)
6738 unsigned long start_pfn, end_pfn;
6739 unsigned long hole_pfn = 0;
6740 int i, j, zone_id = 0, nid;
6742 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6743 struct pglist_data *node = NODE_DATA(nid);
6745 for (j = 0; j < MAX_NR_ZONES; j++) {
6746 struct zone *zone = node->node_zones + j;
6748 if (!populated_zone(zone))
6751 memmap_init_zone_range(zone, start_pfn, end_pfn,
6757 #ifdef CONFIG_SPARSEMEM
6759 * Initialize the memory map for hole in the range [memory_end,
6761 * Append the pages in this hole to the highest zone in the last
6763 * The call to init_unavailable_range() is outside the ifdef to
6764 * silence the compiler warining about zone_id set but not used;
6765 * for FLATMEM it is a nop anyway
6767 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6768 if (hole_pfn < end_pfn)
6770 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6773 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6774 phys_addr_t min_addr, int nid, bool exact_nid)
6779 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6780 MEMBLOCK_ALLOC_ACCESSIBLE,
6783 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6784 MEMBLOCK_ALLOC_ACCESSIBLE,
6787 if (ptr && size > 0)
6788 page_init_poison(ptr, size);
6793 static int zone_batchsize(struct zone *zone)
6799 * The number of pages to batch allocate is either ~0.1%
6800 * of the zone or 1MB, whichever is smaller. The batch
6801 * size is striking a balance between allocation latency
6802 * and zone lock contention.
6804 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6805 batch /= 4; /* We effectively *= 4 below */
6810 * Clamp the batch to a 2^n - 1 value. Having a power
6811 * of 2 value was found to be more likely to have
6812 * suboptimal cache aliasing properties in some cases.
6814 * For example if 2 tasks are alternately allocating
6815 * batches of pages, one task can end up with a lot
6816 * of pages of one half of the possible page colors
6817 * and the other with pages of the other colors.
6819 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6824 /* The deferral and batching of frees should be suppressed under NOMMU
6827 * The problem is that NOMMU needs to be able to allocate large chunks
6828 * of contiguous memory as there's no hardware page translation to
6829 * assemble apparent contiguous memory from discontiguous pages.
6831 * Queueing large contiguous runs of pages for batching, however,
6832 * causes the pages to actually be freed in smaller chunks. As there
6833 * can be a significant delay between the individual batches being
6834 * recycled, this leads to the once large chunks of space being
6835 * fragmented and becoming unavailable for high-order allocations.
6841 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6846 unsigned long total_pages;
6848 if (!percpu_pagelist_high_fraction) {
6850 * By default, the high value of the pcp is based on the zone
6851 * low watermark so that if they are full then background
6852 * reclaim will not be started prematurely.
6854 total_pages = low_wmark_pages(zone);
6857 * If percpu_pagelist_high_fraction is configured, the high
6858 * value is based on a fraction of the managed pages in the
6861 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6865 * Split the high value across all online CPUs local to the zone. Note
6866 * that early in boot that CPUs may not be online yet and that during
6867 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6868 * onlined. For memory nodes that have no CPUs, split pcp->high across
6869 * all online CPUs to mitigate the risk that reclaim is triggered
6870 * prematurely due to pages stored on pcp lists.
6872 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6874 nr_split_cpus = num_online_cpus();
6875 high = total_pages / nr_split_cpus;
6878 * Ensure high is at least batch*4. The multiple is based on the
6879 * historical relationship between high and batch.
6881 high = max(high, batch << 2);
6890 * pcp->high and pcp->batch values are related and generally batch is lower
6891 * than high. They are also related to pcp->count such that count is lower
6892 * than high, and as soon as it reaches high, the pcplist is flushed.
6894 * However, guaranteeing these relations at all times would require e.g. write
6895 * barriers here but also careful usage of read barriers at the read side, and
6896 * thus be prone to error and bad for performance. Thus the update only prevents
6897 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6898 * can cope with those fields changing asynchronously, and fully trust only the
6899 * pcp->count field on the local CPU with interrupts disabled.
6901 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6902 * outside of boot time (or some other assurance that no concurrent updaters
6905 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6906 unsigned long batch)
6908 WRITE_ONCE(pcp->batch, batch);
6909 WRITE_ONCE(pcp->high, high);
6912 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6916 memset(pcp, 0, sizeof(*pcp));
6917 memset(pzstats, 0, sizeof(*pzstats));
6919 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6920 INIT_LIST_HEAD(&pcp->lists[pindex]);
6923 * Set batch and high values safe for a boot pageset. A true percpu
6924 * pageset's initialization will update them subsequently. Here we don't
6925 * need to be as careful as pageset_update() as nobody can access the
6928 pcp->high = BOOT_PAGESET_HIGH;
6929 pcp->batch = BOOT_PAGESET_BATCH;
6930 pcp->free_factor = 0;
6933 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6934 unsigned long batch)
6936 struct per_cpu_pages *pcp;
6939 for_each_possible_cpu(cpu) {
6940 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6941 pageset_update(pcp, high, batch);
6946 * Calculate and set new high and batch values for all per-cpu pagesets of a
6947 * zone based on the zone's size.
6949 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6951 int new_high, new_batch;
6953 new_batch = max(1, zone_batchsize(zone));
6954 new_high = zone_highsize(zone, new_batch, cpu_online);
6956 if (zone->pageset_high == new_high &&
6957 zone->pageset_batch == new_batch)
6960 zone->pageset_high = new_high;
6961 zone->pageset_batch = new_batch;
6963 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6966 void __meminit setup_zone_pageset(struct zone *zone)
6970 /* Size may be 0 on !SMP && !NUMA */
6971 if (sizeof(struct per_cpu_zonestat) > 0)
6972 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
6974 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
6975 for_each_possible_cpu(cpu) {
6976 struct per_cpu_pages *pcp;
6977 struct per_cpu_zonestat *pzstats;
6979 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6980 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6981 per_cpu_pages_init(pcp, pzstats);
6984 zone_set_pageset_high_and_batch(zone, 0);
6988 * Allocate per cpu pagesets and initialize them.
6989 * Before this call only boot pagesets were available.
6991 void __init setup_per_cpu_pageset(void)
6993 struct pglist_data *pgdat;
6995 int __maybe_unused cpu;
6997 for_each_populated_zone(zone)
6998 setup_zone_pageset(zone);
7002 * Unpopulated zones continue using the boot pagesets.
7003 * The numa stats for these pagesets need to be reset.
7004 * Otherwise, they will end up skewing the stats of
7005 * the nodes these zones are associated with.
7007 for_each_possible_cpu(cpu) {
7008 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7009 memset(pzstats->vm_numa_event, 0,
7010 sizeof(pzstats->vm_numa_event));
7014 for_each_online_pgdat(pgdat)
7015 pgdat->per_cpu_nodestats =
7016 alloc_percpu(struct per_cpu_nodestat);
7019 static __meminit void zone_pcp_init(struct zone *zone)
7022 * per cpu subsystem is not up at this point. The following code
7023 * relies on the ability of the linker to provide the
7024 * offset of a (static) per cpu variable into the per cpu area.
7026 zone->per_cpu_pageset = &boot_pageset;
7027 zone->per_cpu_zonestats = &boot_zonestats;
7028 zone->pageset_high = BOOT_PAGESET_HIGH;
7029 zone->pageset_batch = BOOT_PAGESET_BATCH;
7031 if (populated_zone(zone))
7032 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7033 zone->present_pages, zone_batchsize(zone));
7036 void __meminit init_currently_empty_zone(struct zone *zone,
7037 unsigned long zone_start_pfn,
7040 struct pglist_data *pgdat = zone->zone_pgdat;
7041 int zone_idx = zone_idx(zone) + 1;
7043 if (zone_idx > pgdat->nr_zones)
7044 pgdat->nr_zones = zone_idx;
7046 zone->zone_start_pfn = zone_start_pfn;
7048 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7049 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7051 (unsigned long)zone_idx(zone),
7052 zone_start_pfn, (zone_start_pfn + size));
7054 zone_init_free_lists(zone);
7055 zone->initialized = 1;
7059 * get_pfn_range_for_nid - Return the start and end page frames for a node
7060 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7061 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7062 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7064 * It returns the start and end page frame of a node based on information
7065 * provided by memblock_set_node(). If called for a node
7066 * with no available memory, a warning is printed and the start and end
7069 void __init get_pfn_range_for_nid(unsigned int nid,
7070 unsigned long *start_pfn, unsigned long *end_pfn)
7072 unsigned long this_start_pfn, this_end_pfn;
7078 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7079 *start_pfn = min(*start_pfn, this_start_pfn);
7080 *end_pfn = max(*end_pfn, this_end_pfn);
7083 if (*start_pfn == -1UL)
7088 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7089 * assumption is made that zones within a node are ordered in monotonic
7090 * increasing memory addresses so that the "highest" populated zone is used
7092 static void __init find_usable_zone_for_movable(void)
7095 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7096 if (zone_index == ZONE_MOVABLE)
7099 if (arch_zone_highest_possible_pfn[zone_index] >
7100 arch_zone_lowest_possible_pfn[zone_index])
7104 VM_BUG_ON(zone_index == -1);
7105 movable_zone = zone_index;
7109 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7110 * because it is sized independent of architecture. Unlike the other zones,
7111 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7112 * in each node depending on the size of each node and how evenly kernelcore
7113 * is distributed. This helper function adjusts the zone ranges
7114 * provided by the architecture for a given node by using the end of the
7115 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7116 * zones within a node are in order of monotonic increases memory addresses
7118 static void __init adjust_zone_range_for_zone_movable(int nid,
7119 unsigned long zone_type,
7120 unsigned long node_start_pfn,
7121 unsigned long node_end_pfn,
7122 unsigned long *zone_start_pfn,
7123 unsigned long *zone_end_pfn)
7125 /* Only adjust if ZONE_MOVABLE is on this node */
7126 if (zone_movable_pfn[nid]) {
7127 /* Size ZONE_MOVABLE */
7128 if (zone_type == ZONE_MOVABLE) {
7129 *zone_start_pfn = zone_movable_pfn[nid];
7130 *zone_end_pfn = min(node_end_pfn,
7131 arch_zone_highest_possible_pfn[movable_zone]);
7133 /* Adjust for ZONE_MOVABLE starting within this range */
7134 } else if (!mirrored_kernelcore &&
7135 *zone_start_pfn < zone_movable_pfn[nid] &&
7136 *zone_end_pfn > zone_movable_pfn[nid]) {
7137 *zone_end_pfn = zone_movable_pfn[nid];
7139 /* Check if this whole range is within ZONE_MOVABLE */
7140 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7141 *zone_start_pfn = *zone_end_pfn;
7146 * Return the number of pages a zone spans in a node, including holes
7147 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7149 static unsigned long __init zone_spanned_pages_in_node(int nid,
7150 unsigned long zone_type,
7151 unsigned long node_start_pfn,
7152 unsigned long node_end_pfn,
7153 unsigned long *zone_start_pfn,
7154 unsigned long *zone_end_pfn)
7156 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7157 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7158 /* When hotadd a new node from cpu_up(), the node should be empty */
7159 if (!node_start_pfn && !node_end_pfn)
7162 /* Get the start and end of the zone */
7163 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7164 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7165 adjust_zone_range_for_zone_movable(nid, zone_type,
7166 node_start_pfn, node_end_pfn,
7167 zone_start_pfn, zone_end_pfn);
7169 /* Check that this node has pages within the zone's required range */
7170 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7173 /* Move the zone boundaries inside the node if necessary */
7174 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7175 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7177 /* Return the spanned pages */
7178 return *zone_end_pfn - *zone_start_pfn;
7182 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7183 * then all holes in the requested range will be accounted for.
7185 unsigned long __init __absent_pages_in_range(int nid,
7186 unsigned long range_start_pfn,
7187 unsigned long range_end_pfn)
7189 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7190 unsigned long start_pfn, end_pfn;
7193 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7194 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7195 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7196 nr_absent -= end_pfn - start_pfn;
7202 * absent_pages_in_range - Return number of page frames in holes within a range
7203 * @start_pfn: The start PFN to start searching for holes
7204 * @end_pfn: The end PFN to stop searching for holes
7206 * Return: the number of pages frames in memory holes within a range.
7208 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7209 unsigned long end_pfn)
7211 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7214 /* Return the number of page frames in holes in a zone on a node */
7215 static unsigned long __init zone_absent_pages_in_node(int nid,
7216 unsigned long zone_type,
7217 unsigned long node_start_pfn,
7218 unsigned long node_end_pfn)
7220 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7221 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7222 unsigned long zone_start_pfn, zone_end_pfn;
7223 unsigned long nr_absent;
7225 /* When hotadd a new node from cpu_up(), the node should be empty */
7226 if (!node_start_pfn && !node_end_pfn)
7229 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7230 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7232 adjust_zone_range_for_zone_movable(nid, zone_type,
7233 node_start_pfn, node_end_pfn,
7234 &zone_start_pfn, &zone_end_pfn);
7235 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7238 * ZONE_MOVABLE handling.
7239 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7242 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7243 unsigned long start_pfn, end_pfn;
7244 struct memblock_region *r;
7246 for_each_mem_region(r) {
7247 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7248 zone_start_pfn, zone_end_pfn);
7249 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7250 zone_start_pfn, zone_end_pfn);
7252 if (zone_type == ZONE_MOVABLE &&
7253 memblock_is_mirror(r))
7254 nr_absent += end_pfn - start_pfn;
7256 if (zone_type == ZONE_NORMAL &&
7257 !memblock_is_mirror(r))
7258 nr_absent += end_pfn - start_pfn;
7265 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7266 unsigned long node_start_pfn,
7267 unsigned long node_end_pfn)
7269 unsigned long realtotalpages = 0, totalpages = 0;
7272 for (i = 0; i < MAX_NR_ZONES; i++) {
7273 struct zone *zone = pgdat->node_zones + i;
7274 unsigned long zone_start_pfn, zone_end_pfn;
7275 unsigned long spanned, absent;
7276 unsigned long size, real_size;
7278 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7283 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7288 real_size = size - absent;
7291 zone->zone_start_pfn = zone_start_pfn;
7293 zone->zone_start_pfn = 0;
7294 zone->spanned_pages = size;
7295 zone->present_pages = real_size;
7296 #if defined(CONFIG_MEMORY_HOTPLUG)
7297 zone->present_early_pages = real_size;
7301 realtotalpages += real_size;
7304 pgdat->node_spanned_pages = totalpages;
7305 pgdat->node_present_pages = realtotalpages;
7306 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7309 #ifndef CONFIG_SPARSEMEM
7311 * Calculate the size of the zone->blockflags rounded to an unsigned long
7312 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7313 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7314 * round what is now in bits to nearest long in bits, then return it in
7317 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7319 unsigned long usemapsize;
7321 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7322 usemapsize = roundup(zonesize, pageblock_nr_pages);
7323 usemapsize = usemapsize >> pageblock_order;
7324 usemapsize *= NR_PAGEBLOCK_BITS;
7325 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7327 return usemapsize / 8;
7330 static void __ref setup_usemap(struct zone *zone)
7332 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7333 zone->spanned_pages);
7334 zone->pageblock_flags = NULL;
7336 zone->pageblock_flags =
7337 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7339 if (!zone->pageblock_flags)
7340 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7341 usemapsize, zone->name, zone_to_nid(zone));
7345 static inline void setup_usemap(struct zone *zone) {}
7346 #endif /* CONFIG_SPARSEMEM */
7348 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7350 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7351 void __init set_pageblock_order(void)
7353 unsigned int order = MAX_ORDER - 1;
7355 /* Check that pageblock_nr_pages has not already been setup */
7356 if (pageblock_order)
7359 /* Don't let pageblocks exceed the maximum allocation granularity. */
7360 if (HPAGE_SHIFT > PAGE_SHIFT && HUGETLB_PAGE_ORDER < order)
7361 order = HUGETLB_PAGE_ORDER;
7364 * Assume the largest contiguous order of interest is a huge page.
7365 * This value may be variable depending on boot parameters on IA64 and
7368 pageblock_order = order;
7370 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7373 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7374 * is unused as pageblock_order is set at compile-time. See
7375 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7378 void __init set_pageblock_order(void)
7382 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7384 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7385 unsigned long present_pages)
7387 unsigned long pages = spanned_pages;
7390 * Provide a more accurate estimation if there are holes within
7391 * the zone and SPARSEMEM is in use. If there are holes within the
7392 * zone, each populated memory region may cost us one or two extra
7393 * memmap pages due to alignment because memmap pages for each
7394 * populated regions may not be naturally aligned on page boundary.
7395 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7397 if (spanned_pages > present_pages + (present_pages >> 4) &&
7398 IS_ENABLED(CONFIG_SPARSEMEM))
7399 pages = present_pages;
7401 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7404 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7405 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7407 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7409 spin_lock_init(&ds_queue->split_queue_lock);
7410 INIT_LIST_HEAD(&ds_queue->split_queue);
7411 ds_queue->split_queue_len = 0;
7414 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7417 #ifdef CONFIG_COMPACTION
7418 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7420 init_waitqueue_head(&pgdat->kcompactd_wait);
7423 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7426 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7430 pgdat_resize_init(pgdat);
7432 pgdat_init_split_queue(pgdat);
7433 pgdat_init_kcompactd(pgdat);
7435 init_waitqueue_head(&pgdat->kswapd_wait);
7436 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7438 for (i = 0; i < NR_VMSCAN_THROTTLE; i++)
7439 init_waitqueue_head(&pgdat->reclaim_wait[i]);
7441 pgdat_page_ext_init(pgdat);
7442 lruvec_init(&pgdat->__lruvec);
7445 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7446 unsigned long remaining_pages)
7448 atomic_long_set(&zone->managed_pages, remaining_pages);
7449 zone_set_nid(zone, nid);
7450 zone->name = zone_names[idx];
7451 zone->zone_pgdat = NODE_DATA(nid);
7452 spin_lock_init(&zone->lock);
7453 zone_seqlock_init(zone);
7454 zone_pcp_init(zone);
7458 * Set up the zone data structures
7459 * - init pgdat internals
7460 * - init all zones belonging to this node
7462 * NOTE: this function is only called during memory hotplug
7464 #ifdef CONFIG_MEMORY_HOTPLUG
7465 void __ref free_area_init_core_hotplug(int nid)
7468 pg_data_t *pgdat = NODE_DATA(nid);
7470 pgdat_init_internals(pgdat);
7471 for (z = 0; z < MAX_NR_ZONES; z++)
7472 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7477 * Set up the zone data structures:
7478 * - mark all pages reserved
7479 * - mark all memory queues empty
7480 * - clear the memory bitmaps
7482 * NOTE: pgdat should get zeroed by caller.
7483 * NOTE: this function is only called during early init.
7485 static void __init free_area_init_core(struct pglist_data *pgdat)
7488 int nid = pgdat->node_id;
7490 pgdat_init_internals(pgdat);
7491 pgdat->per_cpu_nodestats = &boot_nodestats;
7493 for (j = 0; j < MAX_NR_ZONES; j++) {
7494 struct zone *zone = pgdat->node_zones + j;
7495 unsigned long size, freesize, memmap_pages;
7497 size = zone->spanned_pages;
7498 freesize = zone->present_pages;
7501 * Adjust freesize so that it accounts for how much memory
7502 * is used by this zone for memmap. This affects the watermark
7503 * and per-cpu initialisations
7505 memmap_pages = calc_memmap_size(size, freesize);
7506 if (!is_highmem_idx(j)) {
7507 if (freesize >= memmap_pages) {
7508 freesize -= memmap_pages;
7510 pr_debug(" %s zone: %lu pages used for memmap\n",
7511 zone_names[j], memmap_pages);
7513 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7514 zone_names[j], memmap_pages, freesize);
7517 /* Account for reserved pages */
7518 if (j == 0 && freesize > dma_reserve) {
7519 freesize -= dma_reserve;
7520 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7523 if (!is_highmem_idx(j))
7524 nr_kernel_pages += freesize;
7525 /* Charge for highmem memmap if there are enough kernel pages */
7526 else if (nr_kernel_pages > memmap_pages * 2)
7527 nr_kernel_pages -= memmap_pages;
7528 nr_all_pages += freesize;
7531 * Set an approximate value for lowmem here, it will be adjusted
7532 * when the bootmem allocator frees pages into the buddy system.
7533 * And all highmem pages will be managed by the buddy system.
7535 zone_init_internals(zone, j, nid, freesize);
7540 set_pageblock_order();
7542 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7546 #ifdef CONFIG_FLATMEM
7547 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7549 unsigned long __maybe_unused start = 0;
7550 unsigned long __maybe_unused offset = 0;
7552 /* Skip empty nodes */
7553 if (!pgdat->node_spanned_pages)
7556 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7557 offset = pgdat->node_start_pfn - start;
7558 /* ia64 gets its own node_mem_map, before this, without bootmem */
7559 if (!pgdat->node_mem_map) {
7560 unsigned long size, end;
7564 * The zone's endpoints aren't required to be MAX_ORDER
7565 * aligned but the node_mem_map endpoints must be in order
7566 * for the buddy allocator to function correctly.
7568 end = pgdat_end_pfn(pgdat);
7569 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7570 size = (end - start) * sizeof(struct page);
7571 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7572 pgdat->node_id, false);
7574 panic("Failed to allocate %ld bytes for node %d memory map\n",
7575 size, pgdat->node_id);
7576 pgdat->node_mem_map = map + offset;
7578 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7579 __func__, pgdat->node_id, (unsigned long)pgdat,
7580 (unsigned long)pgdat->node_mem_map);
7583 * With no DISCONTIG, the global mem_map is just set as node 0's
7585 if (pgdat == NODE_DATA(0)) {
7586 mem_map = NODE_DATA(0)->node_mem_map;
7587 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7593 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7594 #endif /* CONFIG_FLATMEM */
7596 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7597 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7599 pgdat->first_deferred_pfn = ULONG_MAX;
7602 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7605 static void __init free_area_init_node(int nid)
7607 pg_data_t *pgdat = NODE_DATA(nid);
7608 unsigned long start_pfn = 0;
7609 unsigned long end_pfn = 0;
7611 /* pg_data_t should be reset to zero when it's allocated */
7612 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7614 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7616 pgdat->node_id = nid;
7617 pgdat->node_start_pfn = start_pfn;
7618 pgdat->per_cpu_nodestats = NULL;
7620 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7621 (u64)start_pfn << PAGE_SHIFT,
7622 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7623 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7625 alloc_node_mem_map(pgdat);
7626 pgdat_set_deferred_range(pgdat);
7628 free_area_init_core(pgdat);
7631 static void __init free_area_init_memoryless_node(int nid)
7633 free_area_init_node(nid);
7636 #if MAX_NUMNODES > 1
7638 * Figure out the number of possible node ids.
7640 void __init setup_nr_node_ids(void)
7642 unsigned int highest;
7644 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7645 nr_node_ids = highest + 1;
7650 * node_map_pfn_alignment - determine the maximum internode alignment
7652 * This function should be called after node map is populated and sorted.
7653 * It calculates the maximum power of two alignment which can distinguish
7656 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7657 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7658 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7659 * shifted, 1GiB is enough and this function will indicate so.
7661 * This is used to test whether pfn -> nid mapping of the chosen memory
7662 * model has fine enough granularity to avoid incorrect mapping for the
7663 * populated node map.
7665 * Return: the determined alignment in pfn's. 0 if there is no alignment
7666 * requirement (single node).
7668 unsigned long __init node_map_pfn_alignment(void)
7670 unsigned long accl_mask = 0, last_end = 0;
7671 unsigned long start, end, mask;
7672 int last_nid = NUMA_NO_NODE;
7675 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7676 if (!start || last_nid < 0 || last_nid == nid) {
7683 * Start with a mask granular enough to pin-point to the
7684 * start pfn and tick off bits one-by-one until it becomes
7685 * too coarse to separate the current node from the last.
7687 mask = ~((1 << __ffs(start)) - 1);
7688 while (mask && last_end <= (start & (mask << 1)))
7691 /* accumulate all internode masks */
7695 /* convert mask to number of pages */
7696 return ~accl_mask + 1;
7700 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7702 * Return: the minimum PFN based on information provided via
7703 * memblock_set_node().
7705 unsigned long __init find_min_pfn_with_active_regions(void)
7707 return PHYS_PFN(memblock_start_of_DRAM());
7711 * early_calculate_totalpages()
7712 * Sum pages in active regions for movable zone.
7713 * Populate N_MEMORY for calculating usable_nodes.
7715 static unsigned long __init early_calculate_totalpages(void)
7717 unsigned long totalpages = 0;
7718 unsigned long start_pfn, end_pfn;
7721 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7722 unsigned long pages = end_pfn - start_pfn;
7724 totalpages += pages;
7726 node_set_state(nid, N_MEMORY);
7732 * Find the PFN the Movable zone begins in each node. Kernel memory
7733 * is spread evenly between nodes as long as the nodes have enough
7734 * memory. When they don't, some nodes will have more kernelcore than
7737 static void __init find_zone_movable_pfns_for_nodes(void)
7740 unsigned long usable_startpfn;
7741 unsigned long kernelcore_node, kernelcore_remaining;
7742 /* save the state before borrow the nodemask */
7743 nodemask_t saved_node_state = node_states[N_MEMORY];
7744 unsigned long totalpages = early_calculate_totalpages();
7745 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7746 struct memblock_region *r;
7748 /* Need to find movable_zone earlier when movable_node is specified. */
7749 find_usable_zone_for_movable();
7752 * If movable_node is specified, ignore kernelcore and movablecore
7755 if (movable_node_is_enabled()) {
7756 for_each_mem_region(r) {
7757 if (!memblock_is_hotpluggable(r))
7760 nid = memblock_get_region_node(r);
7762 usable_startpfn = PFN_DOWN(r->base);
7763 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7764 min(usable_startpfn, zone_movable_pfn[nid]) :
7772 * If kernelcore=mirror is specified, ignore movablecore option
7774 if (mirrored_kernelcore) {
7775 bool mem_below_4gb_not_mirrored = false;
7777 for_each_mem_region(r) {
7778 if (memblock_is_mirror(r))
7781 nid = memblock_get_region_node(r);
7783 usable_startpfn = memblock_region_memory_base_pfn(r);
7785 if (usable_startpfn < 0x100000) {
7786 mem_below_4gb_not_mirrored = true;
7790 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7791 min(usable_startpfn, zone_movable_pfn[nid]) :
7795 if (mem_below_4gb_not_mirrored)
7796 pr_warn("This configuration results in unmirrored kernel memory.\n");
7802 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7803 * amount of necessary memory.
7805 if (required_kernelcore_percent)
7806 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7808 if (required_movablecore_percent)
7809 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7813 * If movablecore= was specified, calculate what size of
7814 * kernelcore that corresponds so that memory usable for
7815 * any allocation type is evenly spread. If both kernelcore
7816 * and movablecore are specified, then the value of kernelcore
7817 * will be used for required_kernelcore if it's greater than
7818 * what movablecore would have allowed.
7820 if (required_movablecore) {
7821 unsigned long corepages;
7824 * Round-up so that ZONE_MOVABLE is at least as large as what
7825 * was requested by the user
7827 required_movablecore =
7828 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7829 required_movablecore = min(totalpages, required_movablecore);
7830 corepages = totalpages - required_movablecore;
7832 required_kernelcore = max(required_kernelcore, corepages);
7836 * If kernelcore was not specified or kernelcore size is larger
7837 * than totalpages, there is no ZONE_MOVABLE.
7839 if (!required_kernelcore || required_kernelcore >= totalpages)
7842 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7843 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7846 /* Spread kernelcore memory as evenly as possible throughout nodes */
7847 kernelcore_node = required_kernelcore / usable_nodes;
7848 for_each_node_state(nid, N_MEMORY) {
7849 unsigned long start_pfn, end_pfn;
7852 * Recalculate kernelcore_node if the division per node
7853 * now exceeds what is necessary to satisfy the requested
7854 * amount of memory for the kernel
7856 if (required_kernelcore < kernelcore_node)
7857 kernelcore_node = required_kernelcore / usable_nodes;
7860 * As the map is walked, we track how much memory is usable
7861 * by the kernel using kernelcore_remaining. When it is
7862 * 0, the rest of the node is usable by ZONE_MOVABLE
7864 kernelcore_remaining = kernelcore_node;
7866 /* Go through each range of PFNs within this node */
7867 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7868 unsigned long size_pages;
7870 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7871 if (start_pfn >= end_pfn)
7874 /* Account for what is only usable for kernelcore */
7875 if (start_pfn < usable_startpfn) {
7876 unsigned long kernel_pages;
7877 kernel_pages = min(end_pfn, usable_startpfn)
7880 kernelcore_remaining -= min(kernel_pages,
7881 kernelcore_remaining);
7882 required_kernelcore -= min(kernel_pages,
7883 required_kernelcore);
7885 /* Continue if range is now fully accounted */
7886 if (end_pfn <= usable_startpfn) {
7889 * Push zone_movable_pfn to the end so
7890 * that if we have to rebalance
7891 * kernelcore across nodes, we will
7892 * not double account here
7894 zone_movable_pfn[nid] = end_pfn;
7897 start_pfn = usable_startpfn;
7901 * The usable PFN range for ZONE_MOVABLE is from
7902 * start_pfn->end_pfn. Calculate size_pages as the
7903 * number of pages used as kernelcore
7905 size_pages = end_pfn - start_pfn;
7906 if (size_pages > kernelcore_remaining)
7907 size_pages = kernelcore_remaining;
7908 zone_movable_pfn[nid] = start_pfn + size_pages;
7911 * Some kernelcore has been met, update counts and
7912 * break if the kernelcore for this node has been
7915 required_kernelcore -= min(required_kernelcore,
7917 kernelcore_remaining -= size_pages;
7918 if (!kernelcore_remaining)
7924 * If there is still required_kernelcore, we do another pass with one
7925 * less node in the count. This will push zone_movable_pfn[nid] further
7926 * along on the nodes that still have memory until kernelcore is
7930 if (usable_nodes && required_kernelcore > usable_nodes)
7934 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7935 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7936 unsigned long start_pfn, end_pfn;
7938 zone_movable_pfn[nid] =
7939 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7941 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7942 if (zone_movable_pfn[nid] >= end_pfn)
7943 zone_movable_pfn[nid] = 0;
7947 /* restore the node_state */
7948 node_states[N_MEMORY] = saved_node_state;
7951 /* Any regular or high memory on that node ? */
7952 static void check_for_memory(pg_data_t *pgdat, int nid)
7954 enum zone_type zone_type;
7956 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7957 struct zone *zone = &pgdat->node_zones[zone_type];
7958 if (populated_zone(zone)) {
7959 if (IS_ENABLED(CONFIG_HIGHMEM))
7960 node_set_state(nid, N_HIGH_MEMORY);
7961 if (zone_type <= ZONE_NORMAL)
7962 node_set_state(nid, N_NORMAL_MEMORY);
7969 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7970 * such cases we allow max_zone_pfn sorted in the descending order
7972 bool __weak arch_has_descending_max_zone_pfns(void)
7978 * free_area_init - Initialise all pg_data_t and zone data
7979 * @max_zone_pfn: an array of max PFNs for each zone
7981 * This will call free_area_init_node() for each active node in the system.
7982 * Using the page ranges provided by memblock_set_node(), the size of each
7983 * zone in each node and their holes is calculated. If the maximum PFN
7984 * between two adjacent zones match, it is assumed that the zone is empty.
7985 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7986 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7987 * starts where the previous one ended. For example, ZONE_DMA32 starts
7988 * at arch_max_dma_pfn.
7990 void __init free_area_init(unsigned long *max_zone_pfn)
7992 unsigned long start_pfn, end_pfn;
7996 /* Record where the zone boundaries are */
7997 memset(arch_zone_lowest_possible_pfn, 0,
7998 sizeof(arch_zone_lowest_possible_pfn));
7999 memset(arch_zone_highest_possible_pfn, 0,
8000 sizeof(arch_zone_highest_possible_pfn));
8002 start_pfn = find_min_pfn_with_active_regions();
8003 descending = arch_has_descending_max_zone_pfns();
8005 for (i = 0; i < MAX_NR_ZONES; i++) {
8007 zone = MAX_NR_ZONES - i - 1;
8011 if (zone == ZONE_MOVABLE)
8014 end_pfn = max(max_zone_pfn[zone], start_pfn);
8015 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8016 arch_zone_highest_possible_pfn[zone] = end_pfn;
8018 start_pfn = end_pfn;
8021 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8022 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8023 find_zone_movable_pfns_for_nodes();
8025 /* Print out the zone ranges */
8026 pr_info("Zone ranges:\n");
8027 for (i = 0; i < MAX_NR_ZONES; i++) {
8028 if (i == ZONE_MOVABLE)
8030 pr_info(" %-8s ", zone_names[i]);
8031 if (arch_zone_lowest_possible_pfn[i] ==
8032 arch_zone_highest_possible_pfn[i])
8035 pr_cont("[mem %#018Lx-%#018Lx]\n",
8036 (u64)arch_zone_lowest_possible_pfn[i]
8038 ((u64)arch_zone_highest_possible_pfn[i]
8039 << PAGE_SHIFT) - 1);
8042 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8043 pr_info("Movable zone start for each node\n");
8044 for (i = 0; i < MAX_NUMNODES; i++) {
8045 if (zone_movable_pfn[i])
8046 pr_info(" Node %d: %#018Lx\n", i,
8047 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8051 * Print out the early node map, and initialize the
8052 * subsection-map relative to active online memory ranges to
8053 * enable future "sub-section" extensions of the memory map.
8055 pr_info("Early memory node ranges\n");
8056 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8057 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8058 (u64)start_pfn << PAGE_SHIFT,
8059 ((u64)end_pfn << PAGE_SHIFT) - 1);
8060 subsection_map_init(start_pfn, end_pfn - start_pfn);
8063 /* Initialise every node */
8064 mminit_verify_pageflags_layout();
8065 setup_nr_node_ids();
8066 for_each_online_node(nid) {
8067 pg_data_t *pgdat = NODE_DATA(nid);
8068 free_area_init_node(nid);
8070 /* Any memory on that node */
8071 if (pgdat->node_present_pages)
8072 node_set_state(nid, N_MEMORY);
8073 check_for_memory(pgdat, nid);
8079 static int __init cmdline_parse_core(char *p, unsigned long *core,
8080 unsigned long *percent)
8082 unsigned long long coremem;
8088 /* Value may be a percentage of total memory, otherwise bytes */
8089 coremem = simple_strtoull(p, &endptr, 0);
8090 if (*endptr == '%') {
8091 /* Paranoid check for percent values greater than 100 */
8092 WARN_ON(coremem > 100);
8096 coremem = memparse(p, &p);
8097 /* Paranoid check that UL is enough for the coremem value */
8098 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8100 *core = coremem >> PAGE_SHIFT;
8107 * kernelcore=size sets the amount of memory for use for allocations that
8108 * cannot be reclaimed or migrated.
8110 static int __init cmdline_parse_kernelcore(char *p)
8112 /* parse kernelcore=mirror */
8113 if (parse_option_str(p, "mirror")) {
8114 mirrored_kernelcore = true;
8118 return cmdline_parse_core(p, &required_kernelcore,
8119 &required_kernelcore_percent);
8123 * movablecore=size sets the amount of memory for use for allocations that
8124 * can be reclaimed or migrated.
8126 static int __init cmdline_parse_movablecore(char *p)
8128 return cmdline_parse_core(p, &required_movablecore,
8129 &required_movablecore_percent);
8132 early_param("kernelcore", cmdline_parse_kernelcore);
8133 early_param("movablecore", cmdline_parse_movablecore);
8135 void adjust_managed_page_count(struct page *page, long count)
8137 atomic_long_add(count, &page_zone(page)->managed_pages);
8138 totalram_pages_add(count);
8139 #ifdef CONFIG_HIGHMEM
8140 if (PageHighMem(page))
8141 totalhigh_pages_add(count);
8144 EXPORT_SYMBOL(adjust_managed_page_count);
8146 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8149 unsigned long pages = 0;
8151 start = (void *)PAGE_ALIGN((unsigned long)start);
8152 end = (void *)((unsigned long)end & PAGE_MASK);
8153 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8154 struct page *page = virt_to_page(pos);
8155 void *direct_map_addr;
8158 * 'direct_map_addr' might be different from 'pos'
8159 * because some architectures' virt_to_page()
8160 * work with aliases. Getting the direct map
8161 * address ensures that we get a _writeable_
8162 * alias for the memset().
8164 direct_map_addr = page_address(page);
8166 * Perform a kasan-unchecked memset() since this memory
8167 * has not been initialized.
8169 direct_map_addr = kasan_reset_tag(direct_map_addr);
8170 if ((unsigned int)poison <= 0xFF)
8171 memset(direct_map_addr, poison, PAGE_SIZE);
8173 free_reserved_page(page);
8177 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
8182 void __init mem_init_print_info(void)
8184 unsigned long physpages, codesize, datasize, rosize, bss_size;
8185 unsigned long init_code_size, init_data_size;
8187 physpages = get_num_physpages();
8188 codesize = _etext - _stext;
8189 datasize = _edata - _sdata;
8190 rosize = __end_rodata - __start_rodata;
8191 bss_size = __bss_stop - __bss_start;
8192 init_data_size = __init_end - __init_begin;
8193 init_code_size = _einittext - _sinittext;
8196 * Detect special cases and adjust section sizes accordingly:
8197 * 1) .init.* may be embedded into .data sections
8198 * 2) .init.text.* may be out of [__init_begin, __init_end],
8199 * please refer to arch/tile/kernel/vmlinux.lds.S.
8200 * 3) .rodata.* may be embedded into .text or .data sections.
8202 #define adj_init_size(start, end, size, pos, adj) \
8204 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8208 adj_init_size(__init_begin, __init_end, init_data_size,
8209 _sinittext, init_code_size);
8210 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8211 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8212 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8213 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8215 #undef adj_init_size
8217 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8218 #ifdef CONFIG_HIGHMEM
8222 K(nr_free_pages()), K(physpages),
8223 codesize >> 10, datasize >> 10, rosize >> 10,
8224 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8225 K(physpages - totalram_pages() - totalcma_pages),
8227 #ifdef CONFIG_HIGHMEM
8228 , K(totalhigh_pages())
8234 * set_dma_reserve - set the specified number of pages reserved in the first zone
8235 * @new_dma_reserve: The number of pages to mark reserved
8237 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8238 * In the DMA zone, a significant percentage may be consumed by kernel image
8239 * and other unfreeable allocations which can skew the watermarks badly. This
8240 * function may optionally be used to account for unfreeable pages in the
8241 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8242 * smaller per-cpu batchsize.
8244 void __init set_dma_reserve(unsigned long new_dma_reserve)
8246 dma_reserve = new_dma_reserve;
8249 static int page_alloc_cpu_dead(unsigned int cpu)
8253 lru_add_drain_cpu(cpu);
8257 * Spill the event counters of the dead processor
8258 * into the current processors event counters.
8259 * This artificially elevates the count of the current
8262 vm_events_fold_cpu(cpu);
8265 * Zero the differential counters of the dead processor
8266 * so that the vm statistics are consistent.
8268 * This is only okay since the processor is dead and cannot
8269 * race with what we are doing.
8271 cpu_vm_stats_fold(cpu);
8273 for_each_populated_zone(zone)
8274 zone_pcp_update(zone, 0);
8279 static int page_alloc_cpu_online(unsigned int cpu)
8283 for_each_populated_zone(zone)
8284 zone_pcp_update(zone, 1);
8289 int hashdist = HASHDIST_DEFAULT;
8291 static int __init set_hashdist(char *str)
8295 hashdist = simple_strtoul(str, &str, 0);
8298 __setup("hashdist=", set_hashdist);
8301 void __init page_alloc_init(void)
8306 if (num_node_state(N_MEMORY) == 1)
8310 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8311 "mm/page_alloc:pcp",
8312 page_alloc_cpu_online,
8313 page_alloc_cpu_dead);
8318 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8319 * or min_free_kbytes changes.
8321 static void calculate_totalreserve_pages(void)
8323 struct pglist_data *pgdat;
8324 unsigned long reserve_pages = 0;
8325 enum zone_type i, j;
8327 for_each_online_pgdat(pgdat) {
8329 pgdat->totalreserve_pages = 0;
8331 for (i = 0; i < MAX_NR_ZONES; i++) {
8332 struct zone *zone = pgdat->node_zones + i;
8334 unsigned long managed_pages = zone_managed_pages(zone);
8336 /* Find valid and maximum lowmem_reserve in the zone */
8337 for (j = i; j < MAX_NR_ZONES; j++) {
8338 if (zone->lowmem_reserve[j] > max)
8339 max = zone->lowmem_reserve[j];
8342 /* we treat the high watermark as reserved pages. */
8343 max += high_wmark_pages(zone);
8345 if (max > managed_pages)
8346 max = managed_pages;
8348 pgdat->totalreserve_pages += max;
8350 reserve_pages += max;
8353 totalreserve_pages = reserve_pages;
8357 * setup_per_zone_lowmem_reserve - called whenever
8358 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8359 * has a correct pages reserved value, so an adequate number of
8360 * pages are left in the zone after a successful __alloc_pages().
8362 static void setup_per_zone_lowmem_reserve(void)
8364 struct pglist_data *pgdat;
8365 enum zone_type i, j;
8367 for_each_online_pgdat(pgdat) {
8368 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8369 struct zone *zone = &pgdat->node_zones[i];
8370 int ratio = sysctl_lowmem_reserve_ratio[i];
8371 bool clear = !ratio || !zone_managed_pages(zone);
8372 unsigned long managed_pages = 0;
8374 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8375 struct zone *upper_zone = &pgdat->node_zones[j];
8377 managed_pages += zone_managed_pages(upper_zone);
8380 zone->lowmem_reserve[j] = 0;
8382 zone->lowmem_reserve[j] = managed_pages / ratio;
8387 /* update totalreserve_pages */
8388 calculate_totalreserve_pages();
8391 static void __setup_per_zone_wmarks(void)
8393 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8394 unsigned long lowmem_pages = 0;
8396 unsigned long flags;
8398 /* Calculate total number of !ZONE_HIGHMEM pages */
8399 for_each_zone(zone) {
8400 if (!is_highmem(zone))
8401 lowmem_pages += zone_managed_pages(zone);
8404 for_each_zone(zone) {
8407 spin_lock_irqsave(&zone->lock, flags);
8408 tmp = (u64)pages_min * zone_managed_pages(zone);
8409 do_div(tmp, lowmem_pages);
8410 if (is_highmem(zone)) {
8412 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8413 * need highmem pages, so cap pages_min to a small
8416 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8417 * deltas control async page reclaim, and so should
8418 * not be capped for highmem.
8420 unsigned long min_pages;
8422 min_pages = zone_managed_pages(zone) / 1024;
8423 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8424 zone->_watermark[WMARK_MIN] = min_pages;
8427 * If it's a lowmem zone, reserve a number of pages
8428 * proportionate to the zone's size.
8430 zone->_watermark[WMARK_MIN] = tmp;
8434 * Set the kswapd watermarks distance according to the
8435 * scale factor in proportion to available memory, but
8436 * ensure a minimum size on small systems.
8438 tmp = max_t(u64, tmp >> 2,
8439 mult_frac(zone_managed_pages(zone),
8440 watermark_scale_factor, 10000));
8442 zone->watermark_boost = 0;
8443 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8444 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8446 spin_unlock_irqrestore(&zone->lock, flags);
8449 /* update totalreserve_pages */
8450 calculate_totalreserve_pages();
8454 * setup_per_zone_wmarks - called when min_free_kbytes changes
8455 * or when memory is hot-{added|removed}
8457 * Ensures that the watermark[min,low,high] values for each zone are set
8458 * correctly with respect to min_free_kbytes.
8460 void setup_per_zone_wmarks(void)
8463 static DEFINE_SPINLOCK(lock);
8466 __setup_per_zone_wmarks();
8470 * The watermark size have changed so update the pcpu batch
8471 * and high limits or the limits may be inappropriate.
8474 zone_pcp_update(zone, 0);
8478 * Initialise min_free_kbytes.
8480 * For small machines we want it small (128k min). For large machines
8481 * we want it large (256MB max). But it is not linear, because network
8482 * bandwidth does not increase linearly with machine size. We use
8484 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8485 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8501 void calculate_min_free_kbytes(void)
8503 unsigned long lowmem_kbytes;
8504 int new_min_free_kbytes;
8506 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8507 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8509 if (new_min_free_kbytes > user_min_free_kbytes)
8510 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
8512 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8513 new_min_free_kbytes, user_min_free_kbytes);
8517 int __meminit init_per_zone_wmark_min(void)
8519 calculate_min_free_kbytes();
8520 setup_per_zone_wmarks();
8521 refresh_zone_stat_thresholds();
8522 setup_per_zone_lowmem_reserve();
8525 setup_min_unmapped_ratio();
8526 setup_min_slab_ratio();
8529 khugepaged_min_free_kbytes_update();
8533 postcore_initcall(init_per_zone_wmark_min)
8536 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8537 * that we can call two helper functions whenever min_free_kbytes
8540 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8541 void *buffer, size_t *length, loff_t *ppos)
8545 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8550 user_min_free_kbytes = min_free_kbytes;
8551 setup_per_zone_wmarks();
8556 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8557 void *buffer, size_t *length, loff_t *ppos)
8561 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8566 setup_per_zone_wmarks();
8572 static void setup_min_unmapped_ratio(void)
8577 for_each_online_pgdat(pgdat)
8578 pgdat->min_unmapped_pages = 0;
8581 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8582 sysctl_min_unmapped_ratio) / 100;
8586 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8587 void *buffer, size_t *length, loff_t *ppos)
8591 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8595 setup_min_unmapped_ratio();
8600 static void setup_min_slab_ratio(void)
8605 for_each_online_pgdat(pgdat)
8606 pgdat->min_slab_pages = 0;
8609 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8610 sysctl_min_slab_ratio) / 100;
8613 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8614 void *buffer, size_t *length, loff_t *ppos)
8618 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8622 setup_min_slab_ratio();
8629 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8630 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8631 * whenever sysctl_lowmem_reserve_ratio changes.
8633 * The reserve ratio obviously has absolutely no relation with the
8634 * minimum watermarks. The lowmem reserve ratio can only make sense
8635 * if in function of the boot time zone sizes.
8637 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8638 void *buffer, size_t *length, loff_t *ppos)
8642 proc_dointvec_minmax(table, write, buffer, length, ppos);
8644 for (i = 0; i < MAX_NR_ZONES; i++) {
8645 if (sysctl_lowmem_reserve_ratio[i] < 1)
8646 sysctl_lowmem_reserve_ratio[i] = 0;
8649 setup_per_zone_lowmem_reserve();
8654 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8655 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8656 * pagelist can have before it gets flushed back to buddy allocator.
8658 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8659 int write, void *buffer, size_t *length, loff_t *ppos)
8662 int old_percpu_pagelist_high_fraction;
8665 mutex_lock(&pcp_batch_high_lock);
8666 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8668 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8669 if (!write || ret < 0)
8672 /* Sanity checking to avoid pcp imbalance */
8673 if (percpu_pagelist_high_fraction &&
8674 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8675 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8681 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8684 for_each_populated_zone(zone)
8685 zone_set_pageset_high_and_batch(zone, 0);
8687 mutex_unlock(&pcp_batch_high_lock);
8691 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8693 * Returns the number of pages that arch has reserved but
8694 * is not known to alloc_large_system_hash().
8696 static unsigned long __init arch_reserved_kernel_pages(void)
8703 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8704 * machines. As memory size is increased the scale is also increased but at
8705 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8706 * quadruples the scale is increased by one, which means the size of hash table
8707 * only doubles, instead of quadrupling as well.
8708 * Because 32-bit systems cannot have large physical memory, where this scaling
8709 * makes sense, it is disabled on such platforms.
8711 #if __BITS_PER_LONG > 32
8712 #define ADAPT_SCALE_BASE (64ul << 30)
8713 #define ADAPT_SCALE_SHIFT 2
8714 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8718 * allocate a large system hash table from bootmem
8719 * - it is assumed that the hash table must contain an exact power-of-2
8720 * quantity of entries
8721 * - limit is the number of hash buckets, not the total allocation size
8723 void *__init alloc_large_system_hash(const char *tablename,
8724 unsigned long bucketsize,
8725 unsigned long numentries,
8728 unsigned int *_hash_shift,
8729 unsigned int *_hash_mask,
8730 unsigned long low_limit,
8731 unsigned long high_limit)
8733 unsigned long long max = high_limit;
8734 unsigned long log2qty, size;
8740 /* allow the kernel cmdline to have a say */
8742 /* round applicable memory size up to nearest megabyte */
8743 numentries = nr_kernel_pages;
8744 numentries -= arch_reserved_kernel_pages();
8746 /* It isn't necessary when PAGE_SIZE >= 1MB */
8747 if (PAGE_SHIFT < 20)
8748 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8750 #if __BITS_PER_LONG > 32
8752 unsigned long adapt;
8754 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8755 adapt <<= ADAPT_SCALE_SHIFT)
8760 /* limit to 1 bucket per 2^scale bytes of low memory */
8761 if (scale > PAGE_SHIFT)
8762 numentries >>= (scale - PAGE_SHIFT);
8764 numentries <<= (PAGE_SHIFT - scale);
8766 /* Make sure we've got at least a 0-order allocation.. */
8767 if (unlikely(flags & HASH_SMALL)) {
8768 /* Makes no sense without HASH_EARLY */
8769 WARN_ON(!(flags & HASH_EARLY));
8770 if (!(numentries >> *_hash_shift)) {
8771 numentries = 1UL << *_hash_shift;
8772 BUG_ON(!numentries);
8774 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8775 numentries = PAGE_SIZE / bucketsize;
8777 numentries = roundup_pow_of_two(numentries);
8779 /* limit allocation size to 1/16 total memory by default */
8781 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8782 do_div(max, bucketsize);
8784 max = min(max, 0x80000000ULL);
8786 if (numentries < low_limit)
8787 numentries = low_limit;
8788 if (numentries > max)
8791 log2qty = ilog2(numentries);
8793 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8796 size = bucketsize << log2qty;
8797 if (flags & HASH_EARLY) {
8798 if (flags & HASH_ZERO)
8799 table = memblock_alloc(size, SMP_CACHE_BYTES);
8801 table = memblock_alloc_raw(size,
8803 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8804 table = __vmalloc(size, gfp_flags);
8807 huge = is_vm_area_hugepages(table);
8810 * If bucketsize is not a power-of-two, we may free
8811 * some pages at the end of hash table which
8812 * alloc_pages_exact() automatically does
8814 table = alloc_pages_exact(size, gfp_flags);
8815 kmemleak_alloc(table, size, 1, gfp_flags);
8817 } while (!table && size > PAGE_SIZE && --log2qty);
8820 panic("Failed to allocate %s hash table\n", tablename);
8822 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8823 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8824 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8827 *_hash_shift = log2qty;
8829 *_hash_mask = (1 << log2qty) - 1;
8835 * This function checks whether pageblock includes unmovable pages or not.
8837 * PageLRU check without isolation or lru_lock could race so that
8838 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8839 * check without lock_page also may miss some movable non-lru pages at
8840 * race condition. So you can't expect this function should be exact.
8842 * Returns a page without holding a reference. If the caller wants to
8843 * dereference that page (e.g., dumping), it has to make sure that it
8844 * cannot get removed (e.g., via memory unplug) concurrently.
8847 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8848 int migratetype, int flags)
8850 unsigned long iter = 0;
8851 unsigned long pfn = page_to_pfn(page);
8852 unsigned long offset = pfn % pageblock_nr_pages;
8854 if (is_migrate_cma_page(page)) {
8856 * CMA allocations (alloc_contig_range) really need to mark
8857 * isolate CMA pageblocks even when they are not movable in fact
8858 * so consider them movable here.
8860 if (is_migrate_cma(migratetype))
8866 for (; iter < pageblock_nr_pages - offset; iter++) {
8867 page = pfn_to_page(pfn + iter);
8870 * Both, bootmem allocations and memory holes are marked
8871 * PG_reserved and are unmovable. We can even have unmovable
8872 * allocations inside ZONE_MOVABLE, for example when
8873 * specifying "movablecore".
8875 if (PageReserved(page))
8879 * If the zone is movable and we have ruled out all reserved
8880 * pages then it should be reasonably safe to assume the rest
8883 if (zone_idx(zone) == ZONE_MOVABLE)
8887 * Hugepages are not in LRU lists, but they're movable.
8888 * THPs are on the LRU, but need to be counted as #small pages.
8889 * We need not scan over tail pages because we don't
8890 * handle each tail page individually in migration.
8892 if (PageHuge(page) || PageTransCompound(page)) {
8893 struct page *head = compound_head(page);
8894 unsigned int skip_pages;
8896 if (PageHuge(page)) {
8897 if (!hugepage_migration_supported(page_hstate(head)))
8899 } else if (!PageLRU(head) && !__PageMovable(head)) {
8903 skip_pages = compound_nr(head) - (page - head);
8904 iter += skip_pages - 1;
8909 * We can't use page_count without pin a page
8910 * because another CPU can free compound page.
8911 * This check already skips compound tails of THP
8912 * because their page->_refcount is zero at all time.
8914 if (!page_ref_count(page)) {
8915 if (PageBuddy(page))
8916 iter += (1 << buddy_order(page)) - 1;
8921 * The HWPoisoned page may be not in buddy system, and
8922 * page_count() is not 0.
8924 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8928 * We treat all PageOffline() pages as movable when offlining
8929 * to give drivers a chance to decrement their reference count
8930 * in MEM_GOING_OFFLINE in order to indicate that these pages
8931 * can be offlined as there are no direct references anymore.
8932 * For actually unmovable PageOffline() where the driver does
8933 * not support this, we will fail later when trying to actually
8934 * move these pages that still have a reference count > 0.
8935 * (false negatives in this function only)
8937 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8940 if (__PageMovable(page) || PageLRU(page))
8944 * If there are RECLAIMABLE pages, we need to check
8945 * it. But now, memory offline itself doesn't call
8946 * shrink_node_slabs() and it still to be fixed.
8953 #ifdef CONFIG_CONTIG_ALLOC
8954 static unsigned long pfn_max_align_down(unsigned long pfn)
8956 return ALIGN_DOWN(pfn, MAX_ORDER_NR_PAGES);
8959 static unsigned long pfn_max_align_up(unsigned long pfn)
8961 return ALIGN(pfn, MAX_ORDER_NR_PAGES);
8964 #if defined(CONFIG_DYNAMIC_DEBUG) || \
8965 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
8966 /* Usage: See admin-guide/dynamic-debug-howto.rst */
8967 static void alloc_contig_dump_pages(struct list_head *page_list)
8969 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
8971 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
8975 list_for_each_entry(page, page_list, lru)
8976 dump_page(page, "migration failure");
8980 static inline void alloc_contig_dump_pages(struct list_head *page_list)
8985 /* [start, end) must belong to a single zone. */
8986 static int __alloc_contig_migrate_range(struct compact_control *cc,
8987 unsigned long start, unsigned long end)
8989 /* This function is based on compact_zone() from compaction.c. */
8990 unsigned int nr_reclaimed;
8991 unsigned long pfn = start;
8992 unsigned int tries = 0;
8994 struct migration_target_control mtc = {
8995 .nid = zone_to_nid(cc->zone),
8996 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8999 lru_cache_disable();
9001 while (pfn < end || !list_empty(&cc->migratepages)) {
9002 if (fatal_signal_pending(current)) {
9007 if (list_empty(&cc->migratepages)) {
9008 cc->nr_migratepages = 0;
9009 ret = isolate_migratepages_range(cc, pfn, end);
9010 if (ret && ret != -EAGAIN)
9012 pfn = cc->migrate_pfn;
9014 } else if (++tries == 5) {
9019 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9021 cc->nr_migratepages -= nr_reclaimed;
9023 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9024 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9027 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9028 * to retry again over this error, so do the same here.
9037 alloc_contig_dump_pages(&cc->migratepages);
9038 putback_movable_pages(&cc->migratepages);
9045 * alloc_contig_range() -- tries to allocate given range of pages
9046 * @start: start PFN to allocate
9047 * @end: one-past-the-last PFN to allocate
9048 * @migratetype: migratetype of the underlying pageblocks (either
9049 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9050 * in range must have the same migratetype and it must
9051 * be either of the two.
9052 * @gfp_mask: GFP mask to use during compaction
9054 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9055 * aligned. The PFN range must belong to a single zone.
9057 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9058 * pageblocks in the range. Once isolated, the pageblocks should not
9059 * be modified by others.
9061 * Return: zero on success or negative error code. On success all
9062 * pages which PFN is in [start, end) are allocated for the caller and
9063 * need to be freed with free_contig_range().
9065 int alloc_contig_range(unsigned long start, unsigned long end,
9066 unsigned migratetype, gfp_t gfp_mask)
9068 unsigned long outer_start, outer_end;
9072 struct compact_control cc = {
9073 .nr_migratepages = 0,
9075 .zone = page_zone(pfn_to_page(start)),
9076 .mode = MIGRATE_SYNC,
9077 .ignore_skip_hint = true,
9078 .no_set_skip_hint = true,
9079 .gfp_mask = current_gfp_context(gfp_mask),
9080 .alloc_contig = true,
9082 INIT_LIST_HEAD(&cc.migratepages);
9085 * What we do here is we mark all pageblocks in range as
9086 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9087 * have different sizes, and due to the way page allocator
9088 * work, we align the range to biggest of the two pages so
9089 * that page allocator won't try to merge buddies from
9090 * different pageblocks and change MIGRATE_ISOLATE to some
9091 * other migration type.
9093 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9094 * migrate the pages from an unaligned range (ie. pages that
9095 * we are interested in). This will put all the pages in
9096 * range back to page allocator as MIGRATE_ISOLATE.
9098 * When this is done, we take the pages in range from page
9099 * allocator removing them from the buddy system. This way
9100 * page allocator will never consider using them.
9102 * This lets us mark the pageblocks back as
9103 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9104 * aligned range but not in the unaligned, original range are
9105 * put back to page allocator so that buddy can use them.
9108 ret = start_isolate_page_range(pfn_max_align_down(start),
9109 pfn_max_align_up(end), migratetype, 0);
9113 drain_all_pages(cc.zone);
9116 * In case of -EBUSY, we'd like to know which page causes problem.
9117 * So, just fall through. test_pages_isolated() has a tracepoint
9118 * which will report the busy page.
9120 * It is possible that busy pages could become available before
9121 * the call to test_pages_isolated, and the range will actually be
9122 * allocated. So, if we fall through be sure to clear ret so that
9123 * -EBUSY is not accidentally used or returned to caller.
9125 ret = __alloc_contig_migrate_range(&cc, start, end);
9126 if (ret && ret != -EBUSY)
9131 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9132 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9133 * more, all pages in [start, end) are free in page allocator.
9134 * What we are going to do is to allocate all pages from
9135 * [start, end) (that is remove them from page allocator).
9137 * The only problem is that pages at the beginning and at the
9138 * end of interesting range may be not aligned with pages that
9139 * page allocator holds, ie. they can be part of higher order
9140 * pages. Because of this, we reserve the bigger range and
9141 * once this is done free the pages we are not interested in.
9143 * We don't have to hold zone->lock here because the pages are
9144 * isolated thus they won't get removed from buddy.
9148 outer_start = start;
9149 while (!PageBuddy(pfn_to_page(outer_start))) {
9150 if (++order >= MAX_ORDER) {
9151 outer_start = start;
9154 outer_start &= ~0UL << order;
9157 if (outer_start != start) {
9158 order = buddy_order(pfn_to_page(outer_start));
9161 * outer_start page could be small order buddy page and
9162 * it doesn't include start page. Adjust outer_start
9163 * in this case to report failed page properly
9164 * on tracepoint in test_pages_isolated()
9166 if (outer_start + (1UL << order) <= start)
9167 outer_start = start;
9170 /* Make sure the range is really isolated. */
9171 if (test_pages_isolated(outer_start, end, 0)) {
9176 /* Grab isolated pages from freelists. */
9177 outer_end = isolate_freepages_range(&cc, outer_start, end);
9183 /* Free head and tail (if any) */
9184 if (start != outer_start)
9185 free_contig_range(outer_start, start - outer_start);
9186 if (end != outer_end)
9187 free_contig_range(end, outer_end - end);
9190 undo_isolate_page_range(pfn_max_align_down(start),
9191 pfn_max_align_up(end), migratetype);
9194 EXPORT_SYMBOL(alloc_contig_range);
9196 static int __alloc_contig_pages(unsigned long start_pfn,
9197 unsigned long nr_pages, gfp_t gfp_mask)
9199 unsigned long end_pfn = start_pfn + nr_pages;
9201 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9205 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9206 unsigned long nr_pages)
9208 unsigned long i, end_pfn = start_pfn + nr_pages;
9211 for (i = start_pfn; i < end_pfn; i++) {
9212 page = pfn_to_online_page(i);
9216 if (page_zone(page) != z)
9219 if (PageReserved(page))
9225 static bool zone_spans_last_pfn(const struct zone *zone,
9226 unsigned long start_pfn, unsigned long nr_pages)
9228 unsigned long last_pfn = start_pfn + nr_pages - 1;
9230 return zone_spans_pfn(zone, last_pfn);
9234 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9235 * @nr_pages: Number of contiguous pages to allocate
9236 * @gfp_mask: GFP mask to limit search and used during compaction
9238 * @nodemask: Mask for other possible nodes
9240 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9241 * on an applicable zonelist to find a contiguous pfn range which can then be
9242 * tried for allocation with alloc_contig_range(). This routine is intended
9243 * for allocation requests which can not be fulfilled with the buddy allocator.
9245 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9246 * power of two, then allocated range is also guaranteed to be aligned to same
9247 * nr_pages (e.g. 1GB request would be aligned to 1GB).
9249 * Allocated pages can be freed with free_contig_range() or by manually calling
9250 * __free_page() on each allocated page.
9252 * Return: pointer to contiguous pages on success, or NULL if not successful.
9254 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9255 int nid, nodemask_t *nodemask)
9257 unsigned long ret, pfn, flags;
9258 struct zonelist *zonelist;
9262 zonelist = node_zonelist(nid, gfp_mask);
9263 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9264 gfp_zone(gfp_mask), nodemask) {
9265 spin_lock_irqsave(&zone->lock, flags);
9267 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9268 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9269 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9271 * We release the zone lock here because
9272 * alloc_contig_range() will also lock the zone
9273 * at some point. If there's an allocation
9274 * spinning on this lock, it may win the race
9275 * and cause alloc_contig_range() to fail...
9277 spin_unlock_irqrestore(&zone->lock, flags);
9278 ret = __alloc_contig_pages(pfn, nr_pages,
9281 return pfn_to_page(pfn);
9282 spin_lock_irqsave(&zone->lock, flags);
9286 spin_unlock_irqrestore(&zone->lock, flags);
9290 #endif /* CONFIG_CONTIG_ALLOC */
9292 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9294 unsigned long count = 0;
9296 for (; nr_pages--; pfn++) {
9297 struct page *page = pfn_to_page(pfn);
9299 count += page_count(page) != 1;
9302 WARN(count != 0, "%lu pages are still in use!\n", count);
9304 EXPORT_SYMBOL(free_contig_range);
9307 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9308 * page high values need to be recalculated.
9310 void zone_pcp_update(struct zone *zone, int cpu_online)
9312 mutex_lock(&pcp_batch_high_lock);
9313 zone_set_pageset_high_and_batch(zone, cpu_online);
9314 mutex_unlock(&pcp_batch_high_lock);
9318 * Effectively disable pcplists for the zone by setting the high limit to 0
9319 * and draining all cpus. A concurrent page freeing on another CPU that's about
9320 * to put the page on pcplist will either finish before the drain and the page
9321 * will be drained, or observe the new high limit and skip the pcplist.
9323 * Must be paired with a call to zone_pcp_enable().
9325 void zone_pcp_disable(struct zone *zone)
9327 mutex_lock(&pcp_batch_high_lock);
9328 __zone_set_pageset_high_and_batch(zone, 0, 1);
9329 __drain_all_pages(zone, true);
9332 void zone_pcp_enable(struct zone *zone)
9334 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9335 mutex_unlock(&pcp_batch_high_lock);
9338 void zone_pcp_reset(struct zone *zone)
9341 struct per_cpu_zonestat *pzstats;
9343 if (zone->per_cpu_pageset != &boot_pageset) {
9344 for_each_online_cpu(cpu) {
9345 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9346 drain_zonestat(zone, pzstats);
9348 free_percpu(zone->per_cpu_pageset);
9349 free_percpu(zone->per_cpu_zonestats);
9350 zone->per_cpu_pageset = &boot_pageset;
9351 zone->per_cpu_zonestats = &boot_zonestats;
9355 #ifdef CONFIG_MEMORY_HOTREMOVE
9357 * All pages in the range must be in a single zone, must not contain holes,
9358 * must span full sections, and must be isolated before calling this function.
9360 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9362 unsigned long pfn = start_pfn;
9366 unsigned long flags;
9368 offline_mem_sections(pfn, end_pfn);
9369 zone = page_zone(pfn_to_page(pfn));
9370 spin_lock_irqsave(&zone->lock, flags);
9371 while (pfn < end_pfn) {
9372 page = pfn_to_page(pfn);
9374 * The HWPoisoned page may be not in buddy system, and
9375 * page_count() is not 0.
9377 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9382 * At this point all remaining PageOffline() pages have a
9383 * reference count of 0 and can simply be skipped.
9385 if (PageOffline(page)) {
9386 BUG_ON(page_count(page));
9387 BUG_ON(PageBuddy(page));
9392 BUG_ON(page_count(page));
9393 BUG_ON(!PageBuddy(page));
9394 order = buddy_order(page);
9395 del_page_from_free_list(page, zone, order);
9396 pfn += (1 << order);
9398 spin_unlock_irqrestore(&zone->lock, flags);
9403 * This function returns a stable result only if called under zone lock.
9405 bool is_free_buddy_page(struct page *page)
9407 unsigned long pfn = page_to_pfn(page);
9410 for (order = 0; order < MAX_ORDER; order++) {
9411 struct page *page_head = page - (pfn & ((1 << order) - 1));
9413 if (PageBuddy(page_head) &&
9414 buddy_order_unsafe(page_head) >= order)
9418 return order < MAX_ORDER;
9420 EXPORT_SYMBOL(is_free_buddy_page);
9422 #ifdef CONFIG_MEMORY_FAILURE
9424 * Break down a higher-order page in sub-pages, and keep our target out of
9427 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9428 struct page *target, int low, int high,
9431 unsigned long size = 1 << high;
9432 struct page *current_buddy, *next_page;
9434 while (high > low) {
9438 if (target >= &page[size]) {
9439 next_page = page + size;
9440 current_buddy = page;
9443 current_buddy = page + size;
9446 if (set_page_guard(zone, current_buddy, high, migratetype))
9449 if (current_buddy != target) {
9450 add_to_free_list(current_buddy, zone, high, migratetype);
9451 set_buddy_order(current_buddy, high);
9458 * Take a page that will be marked as poisoned off the buddy allocator.
9460 bool take_page_off_buddy(struct page *page)
9462 struct zone *zone = page_zone(page);
9463 unsigned long pfn = page_to_pfn(page);
9464 unsigned long flags;
9468 spin_lock_irqsave(&zone->lock, flags);
9469 for (order = 0; order < MAX_ORDER; order++) {
9470 struct page *page_head = page - (pfn & ((1 << order) - 1));
9471 int page_order = buddy_order(page_head);
9473 if (PageBuddy(page_head) && page_order >= order) {
9474 unsigned long pfn_head = page_to_pfn(page_head);
9475 int migratetype = get_pfnblock_migratetype(page_head,
9478 del_page_from_free_list(page_head, zone, page_order);
9479 break_down_buddy_pages(zone, page_head, page, 0,
9480 page_order, migratetype);
9481 SetPageHWPoisonTakenOff(page);
9482 if (!is_migrate_isolate(migratetype))
9483 __mod_zone_freepage_state(zone, -1, migratetype);
9487 if (page_count(page_head) > 0)
9490 spin_unlock_irqrestore(&zone->lock, flags);
9495 * Cancel takeoff done by take_page_off_buddy().
9497 bool put_page_back_buddy(struct page *page)
9499 struct zone *zone = page_zone(page);
9500 unsigned long pfn = page_to_pfn(page);
9501 unsigned long flags;
9502 int migratetype = get_pfnblock_migratetype(page, pfn);
9505 spin_lock_irqsave(&zone->lock, flags);
9506 if (put_page_testzero(page)) {
9507 ClearPageHWPoisonTakenOff(page);
9508 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
9509 if (TestClearPageHWPoison(page)) {
9510 num_poisoned_pages_dec();
9514 spin_unlock_irqrestore(&zone->lock, flags);
9520 #ifdef CONFIG_ZONE_DMA
9521 bool has_managed_dma(void)
9523 struct pglist_data *pgdat;
9525 for_each_online_pgdat(pgdat) {
9526 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9528 if (managed_zone(zone))
9533 #endif /* CONFIG_ZONE_DMA */