1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/mm/page_alloc.c
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
18 #include <linux/stddef.h>
20 #include <linux/highmem.h>
21 #include <linux/swap.h>
22 #include <linux/interrupt.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/memblock.h>
26 #include <linux/compiler.h>
27 #include <linux/kernel.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/random.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/mmu_notifier.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
71 #include <linux/psi.h>
72 #include <linux/padata.h>
73 #include <linux/khugepaged.h>
74 #include <linux/buffer_head.h>
75 #include <asm/sections.h>
76 #include <asm/tlbflush.h>
77 #include <asm/div64.h>
80 #include "page_reporting.h"
82 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
83 typedef int __bitwise fpi_t;
85 /* No special request */
86 #define FPI_NONE ((__force fpi_t)0)
89 * Skip free page reporting notification for the (possibly merged) page.
90 * This does not hinder free page reporting from grabbing the page,
91 * reporting it and marking it "reported" - it only skips notifying
92 * the free page reporting infrastructure about a newly freed page. For
93 * example, used when temporarily pulling a page from a freelist and
94 * putting it back unmodified.
96 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
99 * Place the (possibly merged) page to the tail of the freelist. Will ignore
100 * page shuffling (relevant code - e.g., memory onlining - is expected to
101 * shuffle the whole zone).
103 * Note: No code should rely on this flag for correctness - it's purely
104 * to allow for optimizations when handing back either fresh pages
105 * (memory onlining) or untouched pages (page isolation, free page
108 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
111 * Don't poison memory with KASAN (only for the tag-based modes).
112 * During boot, all non-reserved memblock memory is exposed to page_alloc.
113 * Poisoning all that memory lengthens boot time, especially on systems with
114 * large amount of RAM. This flag is used to skip that poisoning.
115 * This is only done for the tag-based KASAN modes, as those are able to
116 * detect memory corruptions with the memory tags assigned by default.
117 * All memory allocated normally after boot gets poisoned as usual.
119 #define FPI_SKIP_KASAN_POISON ((__force fpi_t)BIT(2))
121 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
122 static DEFINE_MUTEX(pcp_batch_high_lock);
123 #define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
128 static DEFINE_PER_CPU(struct pagesets, pagesets) = {
129 .lock = INIT_LOCAL_LOCK(lock),
132 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
133 DEFINE_PER_CPU(int, numa_node);
134 EXPORT_PER_CPU_SYMBOL(numa_node);
137 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
139 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
141 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
142 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
143 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
144 * defined in <linux/topology.h>.
146 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
147 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
150 /* work_structs for global per-cpu drains */
153 struct work_struct work;
155 static DEFINE_MUTEX(pcpu_drain_mutex);
156 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
158 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
159 volatile unsigned long latent_entropy __latent_entropy;
160 EXPORT_SYMBOL(latent_entropy);
164 * Array of node states.
166 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
167 [N_POSSIBLE] = NODE_MASK_ALL,
168 [N_ONLINE] = { { [0] = 1UL } },
170 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
171 #ifdef CONFIG_HIGHMEM
172 [N_HIGH_MEMORY] = { { [0] = 1UL } },
174 [N_MEMORY] = { { [0] = 1UL } },
175 [N_CPU] = { { [0] = 1UL } },
178 EXPORT_SYMBOL(node_states);
180 atomic_long_t _totalram_pages __read_mostly;
181 EXPORT_SYMBOL(_totalram_pages);
182 unsigned long totalreserve_pages __read_mostly;
183 unsigned long totalcma_pages __read_mostly;
185 int percpu_pagelist_high_fraction;
186 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
187 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
188 EXPORT_SYMBOL(init_on_alloc);
190 DEFINE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_FREE_DEFAULT_ON, init_on_free);
191 EXPORT_SYMBOL(init_on_free);
193 #define ALLOC_IN_CMA_THRESHOLD_MAX 16
194 #define ALLOC_IN_CMA_THRESHOLD_DEFAULT 12
196 static unsigned long _alloc_in_cma_threshold __read_mostly
197 = ALLOC_IN_CMA_THRESHOLD_DEFAULT;
199 static int __init alloc_in_cma_threshold_setup(char *buf)
203 if (kstrtoul(buf, 10, &res) < 0 ||
204 res > ALLOC_IN_CMA_THRESHOLD_MAX) {
205 pr_err("Bad alloc_cma_threshold value\n");
208 _alloc_in_cma_threshold = res;
209 pr_info("Setting alloc_in_cma_threshold to %lu\n", res);
212 early_param("alloc_in_cma_threshold", alloc_in_cma_threshold_setup);
214 static bool _init_on_alloc_enabled_early __read_mostly
215 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
216 static int __init early_init_on_alloc(char *buf)
219 return kstrtobool(buf, &_init_on_alloc_enabled_early);
221 early_param("init_on_alloc", early_init_on_alloc);
223 static bool _init_on_free_enabled_early __read_mostly
224 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
225 static int __init early_init_on_free(char *buf)
227 return kstrtobool(buf, &_init_on_free_enabled_early);
229 early_param("init_on_free", early_init_on_free);
232 * A cached value of the page's pageblock's migratetype, used when the page is
233 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
234 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
235 * Also the migratetype set in the page does not necessarily match the pcplist
236 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
237 * other index - this ensures that it will be put on the correct CMA freelist.
239 static inline int get_pcppage_migratetype(struct page *page)
244 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
246 page->index = migratetype;
249 #ifdef CONFIG_PM_SLEEP
251 * The following functions are used by the suspend/hibernate code to temporarily
252 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
253 * while devices are suspended. To avoid races with the suspend/hibernate code,
254 * they should always be called with system_transition_mutex held
255 * (gfp_allowed_mask also should only be modified with system_transition_mutex
256 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
257 * with that modification).
260 static gfp_t saved_gfp_mask;
262 void pm_restore_gfp_mask(void)
264 WARN_ON(!mutex_is_locked(&system_transition_mutex));
265 if (saved_gfp_mask) {
266 gfp_allowed_mask = saved_gfp_mask;
271 void pm_restrict_gfp_mask(void)
273 WARN_ON(!mutex_is_locked(&system_transition_mutex));
274 WARN_ON(saved_gfp_mask);
275 saved_gfp_mask = gfp_allowed_mask;
276 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
279 bool pm_suspended_storage(void)
281 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
285 #endif /* CONFIG_PM_SLEEP */
287 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
288 unsigned int pageblock_order __read_mostly;
291 static void __free_pages_ok(struct page *page, unsigned int order,
295 * results with 256, 32 in the lowmem_reserve sysctl:
296 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
297 * 1G machine -> (16M dma, 784M normal, 224M high)
298 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
299 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
300 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
302 * TBD: should special case ZONE_DMA32 machines here - in those we normally
303 * don't need any ZONE_NORMAL reservation
305 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
306 #ifdef CONFIG_ZONE_DMA
309 #ifdef CONFIG_ZONE_DMA32
313 #ifdef CONFIG_HIGHMEM
319 static char * const zone_names[MAX_NR_ZONES] = {
320 #ifdef CONFIG_ZONE_DMA
323 #ifdef CONFIG_ZONE_DMA32
327 #ifdef CONFIG_HIGHMEM
331 #ifdef CONFIG_ZONE_DEVICE
336 const char * const migratetype_names[MIGRATE_TYPES] = {
344 #ifdef CONFIG_MEMORY_ISOLATION
349 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
350 [NULL_COMPOUND_DTOR] = NULL,
351 [COMPOUND_PAGE_DTOR] = free_compound_page,
352 #ifdef CONFIG_HUGETLB_PAGE
353 [HUGETLB_PAGE_DTOR] = free_huge_page,
355 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
356 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
360 int min_free_kbytes = 1024;
361 int user_min_free_kbytes = -1;
362 int watermark_boost_factor __read_mostly = 15000;
363 int watermark_scale_factor = 10;
365 static unsigned long nr_kernel_pages __initdata;
366 static unsigned long nr_all_pages __initdata;
367 static unsigned long dma_reserve __initdata;
369 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
370 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
371 static unsigned long required_kernelcore __initdata;
372 static unsigned long required_kernelcore_percent __initdata;
373 static unsigned long required_movablecore __initdata;
374 static unsigned long required_movablecore_percent __initdata;
375 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
376 static bool mirrored_kernelcore __meminitdata;
378 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
380 EXPORT_SYMBOL(movable_zone);
383 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
384 unsigned int nr_online_nodes __read_mostly = 1;
385 EXPORT_SYMBOL(nr_node_ids);
386 EXPORT_SYMBOL(nr_online_nodes);
389 int page_group_by_mobility_disabled __read_mostly;
391 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
393 * During boot we initialize deferred pages on-demand, as needed, but once
394 * page_alloc_init_late() has finished, the deferred pages are all initialized,
395 * and we can permanently disable that path.
397 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
400 * Calling kasan_poison_pages() only after deferred memory initialization
401 * has completed. Poisoning pages during deferred memory init will greatly
402 * lengthen the process and cause problem in large memory systems as the
403 * deferred pages initialization is done with interrupt disabled.
405 * Assuming that there will be no reference to those newly initialized
406 * pages before they are ever allocated, this should have no effect on
407 * KASAN memory tracking as the poison will be properly inserted at page
408 * allocation time. The only corner case is when pages are allocated by
409 * on-demand allocation and then freed again before the deferred pages
410 * initialization is done, but this is not likely to happen.
412 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
414 return static_branch_unlikely(&deferred_pages) ||
415 (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
416 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
417 PageSkipKASanPoison(page);
420 /* Returns true if the struct page for the pfn is uninitialised */
421 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
423 int nid = early_pfn_to_nid(pfn);
425 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
432 * Returns true when the remaining initialisation should be deferred until
433 * later in the boot cycle when it can be parallelised.
435 static bool __meminit
436 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
438 static unsigned long prev_end_pfn, nr_initialised;
441 * prev_end_pfn static that contains the end of previous zone
442 * No need to protect because called very early in boot before smp_init.
444 if (prev_end_pfn != end_pfn) {
445 prev_end_pfn = end_pfn;
449 /* Always populate low zones for address-constrained allocations */
450 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
453 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
456 * We start only with one section of pages, more pages are added as
457 * needed until the rest of deferred pages are initialized.
460 if ((nr_initialised > PAGES_PER_SECTION) &&
461 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
462 NODE_DATA(nid)->first_deferred_pfn = pfn;
468 static inline bool should_skip_kasan_poison(struct page *page, fpi_t fpi_flags)
470 return (!IS_ENABLED(CONFIG_KASAN_GENERIC) &&
471 (fpi_flags & FPI_SKIP_KASAN_POISON)) ||
472 PageSkipKASanPoison(page);
475 static inline bool early_page_uninitialised(unsigned long pfn)
480 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
486 /* Return a pointer to the bitmap storing bits affecting a block of pages */
487 static inline unsigned long *get_pageblock_bitmap(const struct page *page,
490 #ifdef CONFIG_SPARSEMEM
491 return section_to_usemap(__pfn_to_section(pfn));
493 return page_zone(page)->pageblock_flags;
494 #endif /* CONFIG_SPARSEMEM */
497 static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
499 #ifdef CONFIG_SPARSEMEM
500 pfn &= (PAGES_PER_SECTION-1);
502 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
503 #endif /* CONFIG_SPARSEMEM */
504 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
507 static __always_inline
508 unsigned long __get_pfnblock_flags_mask(const struct page *page,
512 unsigned long *bitmap;
513 unsigned long bitidx, word_bitidx;
516 bitmap = get_pageblock_bitmap(page, pfn);
517 bitidx = pfn_to_bitidx(page, pfn);
518 word_bitidx = bitidx / BITS_PER_LONG;
519 bitidx &= (BITS_PER_LONG-1);
521 word = bitmap[word_bitidx];
522 return (word >> bitidx) & mask;
526 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
527 * @page: The page within the block of interest
528 * @pfn: The target page frame number
529 * @mask: mask of bits that the caller is interested in
531 * Return: pageblock_bits flags
533 unsigned long get_pfnblock_flags_mask(const struct page *page,
534 unsigned long pfn, unsigned long mask)
536 return __get_pfnblock_flags_mask(page, pfn, mask);
539 static __always_inline int get_pfnblock_migratetype(const struct page *page,
542 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
546 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
547 * @page: The page within the block of interest
548 * @flags: The flags to set
549 * @pfn: The target page frame number
550 * @mask: mask of bits that the caller is interested in
552 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
556 unsigned long *bitmap;
557 unsigned long bitidx, word_bitidx;
558 unsigned long old_word, word;
560 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
561 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
563 bitmap = get_pageblock_bitmap(page, pfn);
564 bitidx = pfn_to_bitidx(page, pfn);
565 word_bitidx = bitidx / BITS_PER_LONG;
566 bitidx &= (BITS_PER_LONG-1);
568 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
573 word = READ_ONCE(bitmap[word_bitidx]);
575 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
576 if (word == old_word)
582 void set_pageblock_migratetype(struct page *page, int migratetype)
584 if (unlikely(page_group_by_mobility_disabled &&
585 migratetype < MIGRATE_PCPTYPES))
586 migratetype = MIGRATE_UNMOVABLE;
588 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
589 page_to_pfn(page), MIGRATETYPE_MASK);
592 #ifdef CONFIG_DEBUG_VM
593 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
597 unsigned long pfn = page_to_pfn(page);
598 unsigned long sp, start_pfn;
601 seq = zone_span_seqbegin(zone);
602 start_pfn = zone->zone_start_pfn;
603 sp = zone->spanned_pages;
604 if (!zone_spans_pfn(zone, pfn))
606 } while (zone_span_seqretry(zone, seq));
609 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
610 pfn, zone_to_nid(zone), zone->name,
611 start_pfn, start_pfn + sp);
616 static int page_is_consistent(struct zone *zone, struct page *page)
618 if (zone != page_zone(page))
624 * Temporary debugging check for pages not lying within a given zone.
626 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
628 if (page_outside_zone_boundaries(zone, page))
630 if (!page_is_consistent(zone, page))
636 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
642 static void bad_page(struct page *page, const char *reason)
644 static unsigned long resume;
645 static unsigned long nr_shown;
646 static unsigned long nr_unshown;
649 * Allow a burst of 60 reports, then keep quiet for that minute;
650 * or allow a steady drip of one report per second.
652 if (nr_shown == 60) {
653 if (time_before(jiffies, resume)) {
659 "BUG: Bad page state: %lu messages suppressed\n",
666 resume = jiffies + 60 * HZ;
668 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
669 current->comm, page_to_pfn(page));
670 dump_page(page, reason);
675 /* Leave bad fields for debug, except PageBuddy could make trouble */
676 page_mapcount_reset(page); /* remove PageBuddy */
677 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
680 static inline unsigned int order_to_pindex(int migratetype, int order)
684 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
685 if (order > PAGE_ALLOC_COSTLY_ORDER) {
686 VM_BUG_ON(order != pageblock_order);
687 base = PAGE_ALLOC_COSTLY_ORDER + 1;
690 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
693 return (MIGRATE_PCPTYPES * base) + migratetype;
696 static inline int pindex_to_order(unsigned int pindex)
698 int order = pindex / MIGRATE_PCPTYPES;
700 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
701 if (order > PAGE_ALLOC_COSTLY_ORDER) {
702 order = pageblock_order;
703 VM_BUG_ON(order != pageblock_order);
706 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
712 static inline bool pcp_allowed_order(unsigned int order)
714 if (order <= PAGE_ALLOC_COSTLY_ORDER)
716 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
717 if (order == pageblock_order)
723 static inline void free_the_page(struct page *page, unsigned int order)
725 if (pcp_allowed_order(order)) /* Via pcp? */
726 free_unref_page(page, order);
728 __free_pages_ok(page, order, FPI_NONE);
732 * Higher-order pages are called "compound pages". They are structured thusly:
734 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
736 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
737 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
739 * The first tail page's ->compound_dtor holds the offset in array of compound
740 * page destructors. See compound_page_dtors.
742 * The first tail page's ->compound_order holds the order of allocation.
743 * This usage means that zero-order pages may not be compound.
746 void free_compound_page(struct page *page)
748 mem_cgroup_uncharge(page);
749 free_the_page(page, compound_order(page));
752 void prep_compound_page(struct page *page, unsigned int order)
755 int nr_pages = 1 << order;
758 for (i = 1; i < nr_pages; i++) {
759 struct page *p = page + i;
760 p->mapping = TAIL_MAPPING;
761 set_compound_head(p, page);
764 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
765 set_compound_order(page, order);
766 atomic_set(compound_mapcount_ptr(page), -1);
767 if (hpage_pincount_available(page))
768 atomic_set(compound_pincount_ptr(page), 0);
771 #ifdef CONFIG_DEBUG_PAGEALLOC
772 unsigned int _debug_guardpage_minorder;
774 bool _debug_pagealloc_enabled_early __read_mostly
775 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
776 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
777 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
778 EXPORT_SYMBOL(_debug_pagealloc_enabled);
780 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
782 static int __init early_debug_pagealloc(char *buf)
784 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
786 early_param("debug_pagealloc", early_debug_pagealloc);
788 static int __init debug_guardpage_minorder_setup(char *buf)
792 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
793 pr_err("Bad debug_guardpage_minorder value\n");
796 _debug_guardpage_minorder = res;
797 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
800 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
802 static inline bool set_page_guard(struct zone *zone, struct page *page,
803 unsigned int order, int migratetype)
805 if (!debug_guardpage_enabled())
808 if (order >= debug_guardpage_minorder())
811 __SetPageGuard(page);
812 INIT_LIST_HEAD(&page->lru);
813 set_page_private(page, order);
814 /* Guard pages are not available for any usage */
815 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
820 static inline void clear_page_guard(struct zone *zone, struct page *page,
821 unsigned int order, int migratetype)
823 if (!debug_guardpage_enabled())
826 __ClearPageGuard(page);
828 set_page_private(page, 0);
829 if (!is_migrate_isolate(migratetype))
830 __mod_zone_freepage_state(zone, (1 << order), migratetype);
833 static inline bool set_page_guard(struct zone *zone, struct page *page,
834 unsigned int order, int migratetype) { return false; }
835 static inline void clear_page_guard(struct zone *zone, struct page *page,
836 unsigned int order, int migratetype) {}
840 * Enable static keys related to various memory debugging and hardening options.
841 * Some override others, and depend on early params that are evaluated in the
842 * order of appearance. So we need to first gather the full picture of what was
843 * enabled, and then make decisions.
845 void init_mem_debugging_and_hardening(void)
847 bool page_poisoning_requested = false;
849 #ifdef CONFIG_PAGE_POISONING
851 * Page poisoning is debug page alloc for some arches. If
852 * either of those options are enabled, enable poisoning.
854 if (page_poisoning_enabled() ||
855 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
856 debug_pagealloc_enabled())) {
857 static_branch_enable(&_page_poisoning_enabled);
858 page_poisoning_requested = true;
862 if ((_init_on_alloc_enabled_early || _init_on_free_enabled_early) &&
863 page_poisoning_requested) {
864 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
865 "will take precedence over init_on_alloc and init_on_free\n");
866 _init_on_alloc_enabled_early = false;
867 _init_on_free_enabled_early = false;
870 if (_init_on_alloc_enabled_early)
871 static_branch_enable(&init_on_alloc);
873 static_branch_disable(&init_on_alloc);
875 if (_init_on_free_enabled_early)
876 static_branch_enable(&init_on_free);
878 static_branch_disable(&init_on_free);
880 #ifdef CONFIG_DEBUG_PAGEALLOC
881 if (!debug_pagealloc_enabled())
884 static_branch_enable(&_debug_pagealloc_enabled);
886 if (!debug_guardpage_minorder())
889 static_branch_enable(&_debug_guardpage_enabled);
893 static inline void set_buddy_order(struct page *page, unsigned int order)
895 set_page_private(page, order);
896 __SetPageBuddy(page);
900 * This function checks whether a page is free && is the buddy
901 * we can coalesce a page and its buddy if
902 * (a) the buddy is not in a hole (check before calling!) &&
903 * (b) the buddy is in the buddy system &&
904 * (c) a page and its buddy have the same order &&
905 * (d) a page and its buddy are in the same zone.
907 * For recording whether a page is in the buddy system, we set PageBuddy.
908 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
910 * For recording page's order, we use page_private(page).
912 static inline bool page_is_buddy(struct page *page, struct page *buddy,
915 if (!page_is_guard(buddy) && !PageBuddy(buddy))
918 if (buddy_order(buddy) != order)
922 * zone check is done late to avoid uselessly calculating
923 * zone/node ids for pages that could never merge.
925 if (page_zone_id(page) != page_zone_id(buddy))
928 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
933 #ifdef CONFIG_COMPACTION
934 static inline struct capture_control *task_capc(struct zone *zone)
936 struct capture_control *capc = current->capture_control;
938 return unlikely(capc) &&
939 !(current->flags & PF_KTHREAD) &&
941 capc->cc->zone == zone ? capc : NULL;
945 compaction_capture(struct capture_control *capc, struct page *page,
946 int order, int migratetype)
948 if (!capc || order != capc->cc->order)
951 /* Do not accidentally pollute CMA or isolated regions*/
952 if (is_migrate_cma(migratetype) ||
953 is_migrate_isolate(migratetype))
957 * Do not let lower order allocations pollute a movable pageblock.
958 * This might let an unmovable request use a reclaimable pageblock
959 * and vice-versa but no more than normal fallback logic which can
960 * have trouble finding a high-order free page.
962 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
970 static inline struct capture_control *task_capc(struct zone *zone)
976 compaction_capture(struct capture_control *capc, struct page *page,
977 int order, int migratetype)
981 #endif /* CONFIG_COMPACTION */
983 /* Used for pages not on another list */
984 static inline void add_to_free_list(struct page *page, struct zone *zone,
985 unsigned int order, int migratetype)
987 struct free_area *area = &zone->free_area[order];
989 list_add(&page->lru, &area->free_list[migratetype]);
993 /* Used for pages not on another list */
994 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
995 unsigned int order, int migratetype)
997 struct free_area *area = &zone->free_area[order];
999 list_add_tail(&page->lru, &area->free_list[migratetype]);
1004 * Used for pages which are on another list. Move the pages to the tail
1005 * of the list - so the moved pages won't immediately be considered for
1006 * allocation again (e.g., optimization for memory onlining).
1008 static inline void move_to_free_list(struct page *page, struct zone *zone,
1009 unsigned int order, int migratetype)
1011 struct free_area *area = &zone->free_area[order];
1013 list_move_tail(&page->lru, &area->free_list[migratetype]);
1016 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
1019 /* clear reported state and update reported page count */
1020 if (page_reported(page))
1021 __ClearPageReported(page);
1023 list_del(&page->lru);
1024 __ClearPageBuddy(page);
1025 set_page_private(page, 0);
1026 zone->free_area[order].nr_free--;
1030 * If this is not the largest possible page, check if the buddy
1031 * of the next-highest order is free. If it is, it's possible
1032 * that pages are being freed that will coalesce soon. In case,
1033 * that is happening, add the free page to the tail of the list
1034 * so it's less likely to be used soon and more likely to be merged
1035 * as a higher order page
1038 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
1039 struct page *page, unsigned int order)
1041 struct page *higher_page, *higher_buddy;
1042 unsigned long combined_pfn;
1044 if (order >= MAX_ORDER - 2)
1047 combined_pfn = buddy_pfn & pfn;
1048 higher_page = page + (combined_pfn - pfn);
1049 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
1050 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
1052 return page_is_buddy(higher_page, higher_buddy, order + 1);
1056 * Freeing function for a buddy system allocator.
1058 * The concept of a buddy system is to maintain direct-mapped table
1059 * (containing bit values) for memory blocks of various "orders".
1060 * The bottom level table contains the map for the smallest allocatable
1061 * units of memory (here, pages), and each level above it describes
1062 * pairs of units from the levels below, hence, "buddies".
1063 * At a high level, all that happens here is marking the table entry
1064 * at the bottom level available, and propagating the changes upward
1065 * as necessary, plus some accounting needed to play nicely with other
1066 * parts of the VM system.
1067 * At each level, we keep a list of pages, which are heads of continuous
1068 * free pages of length of (1 << order) and marked with PageBuddy.
1069 * Page's order is recorded in page_private(page) field.
1070 * So when we are allocating or freeing one, we can derive the state of the
1071 * other. That is, if we allocate a small block, and both were
1072 * free, the remainder of the region must be split into blocks.
1073 * If a block is freed, and its buddy is also free, then this
1074 * triggers coalescing into a block of larger size.
1079 static inline void __free_one_page(struct page *page,
1081 struct zone *zone, unsigned int order,
1082 int migratetype, fpi_t fpi_flags)
1084 struct capture_control *capc = task_capc(zone);
1085 unsigned long buddy_pfn;
1086 unsigned long combined_pfn;
1087 unsigned int max_order;
1091 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1093 VM_BUG_ON(!zone_is_initialized(zone));
1094 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1096 VM_BUG_ON(migratetype == -1);
1097 if (likely(!is_migrate_isolate(migratetype)))
1098 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1100 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1101 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1104 while (order < max_order) {
1105 if (compaction_capture(capc, page, order, migratetype)) {
1106 __mod_zone_freepage_state(zone, -(1 << order),
1110 buddy_pfn = __find_buddy_pfn(pfn, order);
1111 buddy = page + (buddy_pfn - pfn);
1113 if (!page_is_buddy(page, buddy, order))
1116 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1117 * merge with it and move up one order.
1119 if (page_is_guard(buddy))
1120 clear_page_guard(zone, buddy, order, migratetype);
1122 del_page_from_free_list(buddy, zone, order);
1123 combined_pfn = buddy_pfn & pfn;
1124 page = page + (combined_pfn - pfn);
1128 if (order < MAX_ORDER - 1) {
1129 /* If we are here, it means order is >= pageblock_order.
1130 * We want to prevent merge between freepages on isolate
1131 * pageblock and normal pageblock. Without this, pageblock
1132 * isolation could cause incorrect freepage or CMA accounting.
1134 * We don't want to hit this code for the more frequent
1135 * low-order merging.
1137 if (unlikely(has_isolate_pageblock(zone))) {
1140 buddy_pfn = __find_buddy_pfn(pfn, order);
1141 buddy = page + (buddy_pfn - pfn);
1142 buddy_mt = get_pageblock_migratetype(buddy);
1144 if (migratetype != buddy_mt
1145 && (is_migrate_isolate(migratetype) ||
1146 is_migrate_isolate(buddy_mt)))
1149 max_order = order + 1;
1150 goto continue_merging;
1154 set_buddy_order(page, order);
1156 if (fpi_flags & FPI_TO_TAIL)
1158 else if (is_shuffle_order(order))
1159 to_tail = shuffle_pick_tail();
1161 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1164 add_to_free_list_tail(page, zone, order, migratetype);
1166 add_to_free_list(page, zone, order, migratetype);
1168 /* Notify page reporting subsystem of freed page */
1169 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1170 page_reporting_notify_free(order);
1174 * A bad page could be due to a number of fields. Instead of multiple branches,
1175 * try and check multiple fields with one check. The caller must do a detailed
1176 * check if necessary.
1178 static inline bool page_expected_state(struct page *page,
1179 unsigned long check_flags)
1181 if (unlikely(atomic_read(&page->_mapcount) != -1))
1184 if (unlikely((unsigned long)page->mapping |
1185 page_ref_count(page) |
1189 (page->flags & check_flags)))
1195 static const char *page_bad_reason(struct page *page, unsigned long flags)
1197 const char *bad_reason = NULL;
1199 if (unlikely(atomic_read(&page->_mapcount) != -1))
1200 bad_reason = "nonzero mapcount";
1201 if (unlikely(page->mapping != NULL))
1202 bad_reason = "non-NULL mapping";
1203 if (unlikely(page_ref_count(page) != 0))
1204 bad_reason = "nonzero _refcount";
1205 if (unlikely(page->flags & flags)) {
1206 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1207 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1209 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1212 if (unlikely(page->memcg_data))
1213 bad_reason = "page still charged to cgroup";
1218 static void check_free_page_bad(struct page *page)
1221 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1224 static inline int check_free_page(struct page *page)
1226 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1229 /* Something has gone sideways, find it */
1230 check_free_page_bad(page);
1234 static int free_tail_pages_check(struct page *head_page, struct page *page)
1239 * We rely page->lru.next never has bit 0 set, unless the page
1240 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1242 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1244 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1248 switch (page - head_page) {
1250 /* the first tail page: ->mapping may be compound_mapcount() */
1251 if (unlikely(compound_mapcount(page))) {
1252 bad_page(page, "nonzero compound_mapcount");
1258 * the second tail page: ->mapping is
1259 * deferred_list.next -- ignore value.
1263 if (page->mapping != TAIL_MAPPING) {
1264 bad_page(page, "corrupted mapping in tail page");
1269 if (unlikely(!PageTail(page))) {
1270 bad_page(page, "PageTail not set");
1273 if (unlikely(compound_head(page) != head_page)) {
1274 bad_page(page, "compound_head not consistent");
1279 page->mapping = NULL;
1280 clear_compound_head(page);
1284 static void kernel_init_free_pages(struct page *page, int numpages, bool zero_tags)
1289 for (i = 0; i < numpages; i++)
1290 tag_clear_highpage(page + i);
1294 /* s390's use of memset() could override KASAN redzones. */
1295 kasan_disable_current();
1296 for (i = 0; i < numpages; i++) {
1297 u8 tag = page_kasan_tag(page + i);
1298 page_kasan_tag_reset(page + i);
1299 clear_highpage(page + i);
1300 page_kasan_tag_set(page + i, tag);
1302 kasan_enable_current();
1305 static __always_inline bool free_pages_prepare(struct page *page,
1306 unsigned int order, bool check_free, fpi_t fpi_flags)
1309 bool skip_kasan_poison = should_skip_kasan_poison(page, fpi_flags);
1311 VM_BUG_ON_PAGE(PageTail(page), page);
1313 trace_mm_page_free(page, order);
1315 if (unlikely(PageHWPoison(page)) && !order) {
1317 * Do not let hwpoison pages hit pcplists/buddy
1318 * Untie memcg state and reset page's owner
1320 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1321 __memcg_kmem_uncharge_page(page, order);
1322 reset_page_owner(page, order);
1327 * Check tail pages before head page information is cleared to
1328 * avoid checking PageCompound for order-0 pages.
1330 if (unlikely(order)) {
1331 bool compound = PageCompound(page);
1334 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1337 ClearPageDoubleMap(page);
1338 ClearPageHasHWPoisoned(page);
1340 for (i = 1; i < (1 << order); i++) {
1342 bad += free_tail_pages_check(page, page + i);
1343 if (unlikely(check_free_page(page + i))) {
1347 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1350 if (PageMappingFlags(page))
1351 page->mapping = NULL;
1352 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1353 __memcg_kmem_uncharge_page(page, order);
1355 bad += check_free_page(page);
1359 page_cpupid_reset_last(page);
1360 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1361 reset_page_owner(page, order);
1363 if (!PageHighMem(page)) {
1364 debug_check_no_locks_freed(page_address(page),
1365 PAGE_SIZE << order);
1366 debug_check_no_obj_freed(page_address(page),
1367 PAGE_SIZE << order);
1370 kernel_poison_pages(page, 1 << order);
1373 * As memory initialization might be integrated into KASAN,
1374 * kasan_free_pages and kernel_init_free_pages must be
1375 * kept together to avoid discrepancies in behavior.
1377 * With hardware tag-based KASAN, memory tags must be set before the
1378 * page becomes unavailable via debug_pagealloc or arch_free_page.
1380 if (kasan_has_integrated_init()) {
1381 if (!skip_kasan_poison)
1382 kasan_free_pages(page, order);
1384 bool init = want_init_on_free();
1387 kernel_init_free_pages(page, 1 << order, false);
1388 if (!skip_kasan_poison)
1389 kasan_poison_pages(page, order, init);
1393 * arch_free_page() can make the page's contents inaccessible. s390
1394 * does this. So nothing which can access the page's contents should
1395 * happen after this.
1397 arch_free_page(page, order);
1399 debug_pagealloc_unmap_pages(page, 1 << order);
1404 #ifdef CONFIG_DEBUG_VM
1406 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1407 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1408 * moved from pcp lists to free lists.
1410 static bool free_pcp_prepare(struct page *page, unsigned int order)
1412 return free_pages_prepare(page, order, true, FPI_NONE);
1415 static bool bulkfree_pcp_prepare(struct page *page)
1417 if (debug_pagealloc_enabled_static())
1418 return check_free_page(page);
1424 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1425 * moving from pcp lists to free list in order to reduce overhead. With
1426 * debug_pagealloc enabled, they are checked also immediately when being freed
1429 static bool free_pcp_prepare(struct page *page, unsigned int order)
1431 if (debug_pagealloc_enabled_static())
1432 return free_pages_prepare(page, order, true, FPI_NONE);
1434 return free_pages_prepare(page, order, false, FPI_NONE);
1437 static bool bulkfree_pcp_prepare(struct page *page)
1439 return check_free_page(page);
1441 #endif /* CONFIG_DEBUG_VM */
1443 static inline void prefetch_buddy(struct page *page)
1445 unsigned long pfn = page_to_pfn(page);
1446 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1447 struct page *buddy = page + (buddy_pfn - pfn);
1453 * Frees a number of pages from the PCP lists
1454 * Assumes all pages on list are in same zone, and of same order.
1455 * count is the number of pages to free.
1457 * If the zone was previously in an "all pages pinned" state then look to
1458 * see if this freeing clears that state.
1460 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1461 * pinned" detection logic.
1463 static void free_pcppages_bulk(struct zone *zone, int count,
1464 struct per_cpu_pages *pcp)
1470 int prefetch_nr = READ_ONCE(pcp->batch);
1471 bool isolated_pageblocks;
1472 struct page *page, *tmp;
1476 * Ensure proper count is passed which otherwise would stuck in the
1477 * below while (list_empty(list)) loop.
1479 count = min(pcp->count, count);
1481 struct list_head *list;
1484 * Remove pages from lists in a round-robin fashion. A
1485 * batch_free count is maintained that is incremented when an
1486 * empty list is encountered. This is so more pages are freed
1487 * off fuller lists instead of spinning excessively around empty
1492 if (++pindex == NR_PCP_LISTS)
1494 list = &pcp->lists[pindex];
1495 } while (list_empty(list));
1497 /* This is the only non-empty list. Free them all. */
1498 if (batch_free == NR_PCP_LISTS)
1501 order = pindex_to_order(pindex);
1502 BUILD_BUG_ON(MAX_ORDER >= (1<<NR_PCP_ORDER_WIDTH));
1504 page = list_last_entry(list, struct page, lru);
1505 /* must delete to avoid corrupting pcp list */
1506 list_del(&page->lru);
1507 nr_freed += 1 << order;
1508 count -= 1 << order;
1510 if (bulkfree_pcp_prepare(page))
1513 /* Encode order with the migratetype */
1514 page->index <<= NR_PCP_ORDER_WIDTH;
1515 page->index |= order;
1517 list_add_tail(&page->lru, &head);
1520 * We are going to put the page back to the global
1521 * pool, prefetch its buddy to speed up later access
1522 * under zone->lock. It is believed the overhead of
1523 * an additional test and calculating buddy_pfn here
1524 * can be offset by reduced memory latency later. To
1525 * avoid excessive prefetching due to large count, only
1526 * prefetch buddy for the first pcp->batch nr of pages.
1529 prefetch_buddy(page);
1532 } while (count > 0 && --batch_free && !list_empty(list));
1534 pcp->count -= nr_freed;
1537 * local_lock_irq held so equivalent to spin_lock_irqsave for
1538 * both PREEMPT_RT and non-PREEMPT_RT configurations.
1540 spin_lock(&zone->lock);
1541 isolated_pageblocks = has_isolate_pageblock(zone);
1544 * Use safe version since after __free_one_page(),
1545 * page->lru.next will not point to original list.
1547 list_for_each_entry_safe(page, tmp, &head, lru) {
1548 int mt = get_pcppage_migratetype(page);
1550 /* mt has been encoded with the order (see above) */
1551 order = mt & NR_PCP_ORDER_MASK;
1552 mt >>= NR_PCP_ORDER_WIDTH;
1554 /* MIGRATE_ISOLATE page should not go to pcplists */
1555 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1556 /* Pageblock could have been isolated meanwhile */
1557 if (unlikely(isolated_pageblocks))
1558 mt = get_pageblock_migratetype(page);
1560 __free_one_page(page, page_to_pfn(page), zone, order, mt, FPI_NONE);
1561 trace_mm_page_pcpu_drain(page, order, mt);
1563 spin_unlock(&zone->lock);
1566 static void free_one_page(struct zone *zone,
1567 struct page *page, unsigned long pfn,
1569 int migratetype, fpi_t fpi_flags)
1571 unsigned long flags;
1573 spin_lock_irqsave(&zone->lock, flags);
1574 if (unlikely(has_isolate_pageblock(zone) ||
1575 is_migrate_isolate(migratetype))) {
1576 migratetype = get_pfnblock_migratetype(page, pfn);
1578 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1579 spin_unlock_irqrestore(&zone->lock, flags);
1582 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1583 unsigned long zone, int nid)
1585 mm_zero_struct_page(page);
1586 set_page_links(page, zone, nid, pfn);
1587 init_page_count(page);
1588 page_mapcount_reset(page);
1589 page_cpupid_reset_last(page);
1590 page_kasan_tag_reset(page);
1592 INIT_LIST_HEAD(&page->lru);
1593 #ifdef WANT_PAGE_VIRTUAL
1594 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1595 if (!is_highmem_idx(zone))
1596 set_page_address(page, __va(pfn << PAGE_SHIFT));
1600 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1601 static void __meminit init_reserved_page(unsigned long pfn)
1606 if (!early_page_uninitialised(pfn))
1609 nid = early_pfn_to_nid(pfn);
1610 pgdat = NODE_DATA(nid);
1612 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1613 struct zone *zone = &pgdat->node_zones[zid];
1615 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1618 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1621 static inline void init_reserved_page(unsigned long pfn)
1624 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1627 * Initialised pages do not have PageReserved set. This function is
1628 * called for each range allocated by the bootmem allocator and
1629 * marks the pages PageReserved. The remaining valid pages are later
1630 * sent to the buddy page allocator.
1632 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1634 unsigned long start_pfn = PFN_DOWN(start);
1635 unsigned long end_pfn = PFN_UP(end);
1637 for (; start_pfn < end_pfn; start_pfn++) {
1638 if (pfn_valid(start_pfn)) {
1639 struct page *page = pfn_to_page(start_pfn);
1641 init_reserved_page(start_pfn);
1643 /* Avoid false-positive PageTail() */
1644 INIT_LIST_HEAD(&page->lru);
1647 * no need for atomic set_bit because the struct
1648 * page is not visible yet so nobody should
1651 __SetPageReserved(page);
1656 static void __free_pages_ok(struct page *page, unsigned int order,
1659 unsigned long flags;
1661 unsigned long pfn = page_to_pfn(page);
1662 struct zone *zone = page_zone(page);
1664 if (!free_pages_prepare(page, order, true, fpi_flags))
1667 migratetype = get_pfnblock_migratetype(page, pfn);
1669 spin_lock_irqsave(&zone->lock, flags);
1670 if (unlikely(has_isolate_pageblock(zone) ||
1671 is_migrate_isolate(migratetype))) {
1672 migratetype = get_pfnblock_migratetype(page, pfn);
1674 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1675 spin_unlock_irqrestore(&zone->lock, flags);
1677 __count_vm_events(PGFREE, 1 << order);
1680 void __free_pages_core(struct page *page, unsigned int order)
1682 unsigned int nr_pages = 1 << order;
1683 struct page *p = page;
1687 * When initializing the memmap, __init_single_page() sets the refcount
1688 * of all pages to 1 ("allocated"/"not free"). We have to set the
1689 * refcount of all involved pages to 0.
1692 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1694 __ClearPageReserved(p);
1695 set_page_count(p, 0);
1697 __ClearPageReserved(p);
1698 set_page_count(p, 0);
1700 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1703 * Bypass PCP and place fresh pages right to the tail, primarily
1704 * relevant for memory onlining.
1706 __free_pages_ok(page, order, FPI_TO_TAIL | FPI_SKIP_KASAN_POISON);
1712 * During memory init memblocks map pfns to nids. The search is expensive and
1713 * this caches recent lookups. The implementation of __early_pfn_to_nid
1714 * treats start/end as pfns.
1716 struct mminit_pfnnid_cache {
1717 unsigned long last_start;
1718 unsigned long last_end;
1722 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1725 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1727 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1728 struct mminit_pfnnid_cache *state)
1730 unsigned long start_pfn, end_pfn;
1733 if (state->last_start <= pfn && pfn < state->last_end)
1734 return state->last_nid;
1736 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1737 if (nid != NUMA_NO_NODE) {
1738 state->last_start = start_pfn;
1739 state->last_end = end_pfn;
1740 state->last_nid = nid;
1746 int __meminit early_pfn_to_nid(unsigned long pfn)
1748 static DEFINE_SPINLOCK(early_pfn_lock);
1751 spin_lock(&early_pfn_lock);
1752 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1754 nid = first_online_node;
1755 spin_unlock(&early_pfn_lock);
1759 #endif /* CONFIG_NUMA */
1761 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1764 if (early_page_uninitialised(pfn))
1766 __free_pages_core(page, order);
1770 * Check that the whole (or subset of) a pageblock given by the interval of
1771 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1772 * with the migration of free compaction scanner.
1774 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1776 * It's possible on some configurations to have a setup like node0 node1 node0
1777 * i.e. it's possible that all pages within a zones range of pages do not
1778 * belong to a single zone. We assume that a border between node0 and node1
1779 * can occur within a single pageblock, but not a node0 node1 node0
1780 * interleaving within a single pageblock. It is therefore sufficient to check
1781 * the first and last page of a pageblock and avoid checking each individual
1782 * page in a pageblock.
1784 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1785 unsigned long end_pfn, struct zone *zone)
1787 struct page *start_page;
1788 struct page *end_page;
1790 /* end_pfn is one past the range we are checking */
1793 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1796 start_page = pfn_to_online_page(start_pfn);
1800 if (page_zone(start_page) != zone)
1803 end_page = pfn_to_page(end_pfn);
1805 /* This gives a shorter code than deriving page_zone(end_page) */
1806 if (page_zone_id(start_page) != page_zone_id(end_page))
1812 void set_zone_contiguous(struct zone *zone)
1814 unsigned long block_start_pfn = zone->zone_start_pfn;
1815 unsigned long block_end_pfn;
1817 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1818 for (; block_start_pfn < zone_end_pfn(zone);
1819 block_start_pfn = block_end_pfn,
1820 block_end_pfn += pageblock_nr_pages) {
1822 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1824 if (!__pageblock_pfn_to_page(block_start_pfn,
1825 block_end_pfn, zone))
1830 /* We confirm that there is no hole */
1831 zone->contiguous = true;
1834 void clear_zone_contiguous(struct zone *zone)
1836 zone->contiguous = false;
1839 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1840 static void __init deferred_free_range(unsigned long pfn,
1841 unsigned long nr_pages)
1849 page = pfn_to_page(pfn);
1851 /* Free a large naturally-aligned chunk if possible */
1852 if (nr_pages == pageblock_nr_pages &&
1853 (pfn & (pageblock_nr_pages - 1)) == 0) {
1854 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1855 __free_pages_core(page, pageblock_order);
1859 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1860 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1861 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1862 __free_pages_core(page, 0);
1866 /* Completion tracking for deferred_init_memmap() threads */
1867 static atomic_t pgdat_init_n_undone __initdata;
1868 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1870 static inline void __init pgdat_init_report_one_done(void)
1872 if (atomic_dec_and_test(&pgdat_init_n_undone))
1873 complete(&pgdat_init_all_done_comp);
1877 * Returns true if page needs to be initialized or freed to buddy allocator.
1879 * First we check if pfn is valid on architectures where it is possible to have
1880 * holes within pageblock_nr_pages. On systems where it is not possible, this
1881 * function is optimized out.
1883 * Then, we check if a current large page is valid by only checking the validity
1886 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1888 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1894 * Free pages to buddy allocator. Try to free aligned pages in
1895 * pageblock_nr_pages sizes.
1897 static void __init deferred_free_pages(unsigned long pfn,
1898 unsigned long end_pfn)
1900 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1901 unsigned long nr_free = 0;
1903 for (; pfn < end_pfn; pfn++) {
1904 if (!deferred_pfn_valid(pfn)) {
1905 deferred_free_range(pfn - nr_free, nr_free);
1907 } else if (!(pfn & nr_pgmask)) {
1908 deferred_free_range(pfn - nr_free, nr_free);
1914 /* Free the last block of pages to allocator */
1915 deferred_free_range(pfn - nr_free, nr_free);
1919 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1920 * by performing it only once every pageblock_nr_pages.
1921 * Return number of pages initialized.
1923 static unsigned long __init deferred_init_pages(struct zone *zone,
1925 unsigned long end_pfn)
1927 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1928 int nid = zone_to_nid(zone);
1929 unsigned long nr_pages = 0;
1930 int zid = zone_idx(zone);
1931 struct page *page = NULL;
1933 for (; pfn < end_pfn; pfn++) {
1934 if (!deferred_pfn_valid(pfn)) {
1937 } else if (!page || !(pfn & nr_pgmask)) {
1938 page = pfn_to_page(pfn);
1942 __init_single_page(page, pfn, zid, nid);
1949 * This function is meant to pre-load the iterator for the zone init.
1950 * Specifically it walks through the ranges until we are caught up to the
1951 * first_init_pfn value and exits there. If we never encounter the value we
1952 * return false indicating there are no valid ranges left.
1955 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1956 unsigned long *spfn, unsigned long *epfn,
1957 unsigned long first_init_pfn)
1962 * Start out by walking through the ranges in this zone that have
1963 * already been initialized. We don't need to do anything with them
1964 * so we just need to flush them out of the system.
1966 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1967 if (*epfn <= first_init_pfn)
1969 if (*spfn < first_init_pfn)
1970 *spfn = first_init_pfn;
1979 * Initialize and free pages. We do it in two loops: first we initialize
1980 * struct page, then free to buddy allocator, because while we are
1981 * freeing pages we can access pages that are ahead (computing buddy
1982 * page in __free_one_page()).
1984 * In order to try and keep some memory in the cache we have the loop
1985 * broken along max page order boundaries. This way we will not cause
1986 * any issues with the buddy page computation.
1988 static unsigned long __init
1989 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1990 unsigned long *end_pfn)
1992 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1993 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1994 unsigned long nr_pages = 0;
1997 /* First we loop through and initialize the page values */
1998 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
2001 if (mo_pfn <= *start_pfn)
2004 t = min(mo_pfn, *end_pfn);
2005 nr_pages += deferred_init_pages(zone, *start_pfn, t);
2007 if (mo_pfn < *end_pfn) {
2008 *start_pfn = mo_pfn;
2013 /* Reset values and now loop through freeing pages as needed */
2016 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
2022 t = min(mo_pfn, epfn);
2023 deferred_free_pages(spfn, t);
2033 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
2036 unsigned long spfn, epfn;
2037 struct zone *zone = arg;
2040 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
2043 * Initialize and free pages in MAX_ORDER sized increments so that we
2044 * can avoid introducing any issues with the buddy allocator.
2046 while (spfn < end_pfn) {
2047 deferred_init_maxorder(&i, zone, &spfn, &epfn);
2052 /* An arch may override for more concurrency. */
2054 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
2059 /* Initialise remaining memory on a node */
2060 static int __init deferred_init_memmap(void *data)
2062 pg_data_t *pgdat = data;
2063 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
2064 unsigned long spfn = 0, epfn = 0;
2065 unsigned long first_init_pfn, flags;
2066 unsigned long start = jiffies;
2068 int zid, max_threads;
2071 /* Bind memory initialisation thread to a local node if possible */
2072 if (!cpumask_empty(cpumask))
2073 set_cpus_allowed_ptr(current, cpumask);
2075 pgdat_resize_lock(pgdat, &flags);
2076 first_init_pfn = pgdat->first_deferred_pfn;
2077 if (first_init_pfn == ULONG_MAX) {
2078 pgdat_resize_unlock(pgdat, &flags);
2079 pgdat_init_report_one_done();
2083 /* Sanity check boundaries */
2084 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
2085 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
2086 pgdat->first_deferred_pfn = ULONG_MAX;
2089 * Once we unlock here, the zone cannot be grown anymore, thus if an
2090 * interrupt thread must allocate this early in boot, zone must be
2091 * pre-grown prior to start of deferred page initialization.
2093 pgdat_resize_unlock(pgdat, &flags);
2095 /* Only the highest zone is deferred so find it */
2096 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2097 zone = pgdat->node_zones + zid;
2098 if (first_init_pfn < zone_end_pfn(zone))
2102 /* If the zone is empty somebody else may have cleared out the zone */
2103 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2107 max_threads = deferred_page_init_max_threads(cpumask);
2109 while (spfn < epfn) {
2110 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
2111 struct padata_mt_job job = {
2112 .thread_fn = deferred_init_memmap_chunk,
2115 .size = epfn_align - spfn,
2116 .align = PAGES_PER_SECTION,
2117 .min_chunk = PAGES_PER_SECTION,
2118 .max_threads = max_threads,
2121 padata_do_multithreaded(&job);
2122 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2126 /* Sanity check that the next zone really is unpopulated */
2127 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2129 pr_info("node %d deferred pages initialised in %ums\n",
2130 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2132 pgdat_init_report_one_done();
2137 * If this zone has deferred pages, try to grow it by initializing enough
2138 * deferred pages to satisfy the allocation specified by order, rounded up to
2139 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2140 * of SECTION_SIZE bytes by initializing struct pages in increments of
2141 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2143 * Return true when zone was grown, otherwise return false. We return true even
2144 * when we grow less than requested, to let the caller decide if there are
2145 * enough pages to satisfy the allocation.
2147 * Note: We use noinline because this function is needed only during boot, and
2148 * it is called from a __ref function _deferred_grow_zone. This way we are
2149 * making sure that it is not inlined into permanent text section.
2151 static noinline bool __init
2152 deferred_grow_zone(struct zone *zone, unsigned int order)
2154 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2155 pg_data_t *pgdat = zone->zone_pgdat;
2156 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2157 unsigned long spfn, epfn, flags;
2158 unsigned long nr_pages = 0;
2161 /* Only the last zone may have deferred pages */
2162 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2165 pgdat_resize_lock(pgdat, &flags);
2168 * If someone grew this zone while we were waiting for spinlock, return
2169 * true, as there might be enough pages already.
2171 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2172 pgdat_resize_unlock(pgdat, &flags);
2176 /* If the zone is empty somebody else may have cleared out the zone */
2177 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2178 first_deferred_pfn)) {
2179 pgdat->first_deferred_pfn = ULONG_MAX;
2180 pgdat_resize_unlock(pgdat, &flags);
2181 /* Retry only once. */
2182 return first_deferred_pfn != ULONG_MAX;
2186 * Initialize and free pages in MAX_ORDER sized increments so
2187 * that we can avoid introducing any issues with the buddy
2190 while (spfn < epfn) {
2191 /* update our first deferred PFN for this section */
2192 first_deferred_pfn = spfn;
2194 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2195 touch_nmi_watchdog();
2197 /* We should only stop along section boundaries */
2198 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2201 /* If our quota has been met we can stop here */
2202 if (nr_pages >= nr_pages_needed)
2206 pgdat->first_deferred_pfn = spfn;
2207 pgdat_resize_unlock(pgdat, &flags);
2209 return nr_pages > 0;
2213 * deferred_grow_zone() is __init, but it is called from
2214 * get_page_from_freelist() during early boot until deferred_pages permanently
2215 * disables this call. This is why we have refdata wrapper to avoid warning,
2216 * and to ensure that the function body gets unloaded.
2219 _deferred_grow_zone(struct zone *zone, unsigned int order)
2221 return deferred_grow_zone(zone, order);
2224 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2226 void __init page_alloc_init_late(void)
2231 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2233 /* There will be num_node_state(N_MEMORY) threads */
2234 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2235 for_each_node_state(nid, N_MEMORY) {
2236 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2239 /* Block until all are initialised */
2240 wait_for_completion(&pgdat_init_all_done_comp);
2243 * We initialized the rest of the deferred pages. Permanently disable
2244 * on-demand struct page initialization.
2246 static_branch_disable(&deferred_pages);
2248 /* Reinit limits that are based on free pages after the kernel is up */
2249 files_maxfiles_init();
2254 /* Discard memblock private memory */
2257 for_each_node_state(nid, N_MEMORY)
2258 shuffle_free_memory(NODE_DATA(nid));
2260 for_each_populated_zone(zone)
2261 set_zone_contiguous(zone);
2265 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2266 void __init init_cma_reserved_pageblock(struct page *page)
2268 unsigned i = pageblock_nr_pages;
2269 struct page *p = page;
2272 __ClearPageReserved(p);
2273 set_page_count(p, 0);
2276 set_pageblock_migratetype(page, MIGRATE_CMA);
2278 if (pageblock_order >= MAX_ORDER) {
2279 i = pageblock_nr_pages;
2282 set_page_refcounted(p);
2283 __free_pages(p, MAX_ORDER - 1);
2284 p += MAX_ORDER_NR_PAGES;
2285 } while (i -= MAX_ORDER_NR_PAGES);
2287 set_page_refcounted(page);
2288 __free_pages(page, pageblock_order);
2291 adjust_managed_page_count(page, pageblock_nr_pages);
2292 page_zone(page)->cma_pages += pageblock_nr_pages;
2297 * The order of subdivision here is critical for the IO subsystem.
2298 * Please do not alter this order without good reasons and regression
2299 * testing. Specifically, as large blocks of memory are subdivided,
2300 * the order in which smaller blocks are delivered depends on the order
2301 * they're subdivided in this function. This is the primary factor
2302 * influencing the order in which pages are delivered to the IO
2303 * subsystem according to empirical testing, and this is also justified
2304 * by considering the behavior of a buddy system containing a single
2305 * large block of memory acted on by a series of small allocations.
2306 * This behavior is a critical factor in sglist merging's success.
2310 static inline void expand(struct zone *zone, struct page *page,
2311 int low, int high, int migratetype)
2313 unsigned long size = 1 << high;
2315 while (high > low) {
2318 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2321 * Mark as guard pages (or page), that will allow to
2322 * merge back to allocator when buddy will be freed.
2323 * Corresponding page table entries will not be touched,
2324 * pages will stay not present in virtual address space
2326 if (set_page_guard(zone, &page[size], high, migratetype))
2329 add_to_free_list(&page[size], zone, high, migratetype);
2330 set_buddy_order(&page[size], high);
2334 static void check_new_page_bad(struct page *page)
2336 if (unlikely(page->flags & __PG_HWPOISON)) {
2337 /* Don't complain about hwpoisoned pages */
2338 page_mapcount_reset(page); /* remove PageBuddy */
2343 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2347 * This page is about to be returned from the page allocator
2349 static inline int check_new_page(struct page *page)
2351 if (likely(page_expected_state(page,
2352 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2355 check_new_page_bad(page);
2359 #ifdef CONFIG_DEBUG_VM
2361 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2362 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2363 * also checked when pcp lists are refilled from the free lists.
2365 static inline bool check_pcp_refill(struct page *page)
2367 if (debug_pagealloc_enabled_static())
2368 return check_new_page(page);
2373 static inline bool check_new_pcp(struct page *page)
2375 return check_new_page(page);
2379 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2380 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2381 * enabled, they are also checked when being allocated from the pcp lists.
2383 static inline bool check_pcp_refill(struct page *page)
2385 return check_new_page(page);
2387 static inline bool check_new_pcp(struct page *page)
2389 if (debug_pagealloc_enabled_static())
2390 return check_new_page(page);
2394 #endif /* CONFIG_DEBUG_VM */
2396 static bool check_new_pages(struct page *page, unsigned int order)
2399 for (i = 0; i < (1 << order); i++) {
2400 struct page *p = page + i;
2402 if (unlikely(check_new_page(p)))
2409 inline void post_alloc_hook(struct page *page, unsigned int order,
2412 set_page_private(page, 0);
2413 set_page_refcounted(page);
2415 arch_alloc_page(page, order);
2416 debug_pagealloc_map_pages(page, 1 << order);
2419 * Page unpoisoning must happen before memory initialization.
2420 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
2421 * allocations and the page unpoisoning code will complain.
2423 kernel_unpoison_pages(page, 1 << order);
2426 * As memory initialization might be integrated into KASAN,
2427 * kasan_alloc_pages and kernel_init_free_pages must be
2428 * kept together to avoid discrepancies in behavior.
2430 if (kasan_has_integrated_init()) {
2431 kasan_alloc_pages(page, order, gfp_flags);
2433 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags);
2435 kasan_unpoison_pages(page, order, init);
2437 kernel_init_free_pages(page, 1 << order,
2438 gfp_flags & __GFP_ZEROTAGS);
2441 set_page_owner(page, order, gfp_flags);
2444 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2445 unsigned int alloc_flags)
2447 post_alloc_hook(page, order, gfp_flags);
2449 if (order && (gfp_flags & __GFP_COMP))
2450 prep_compound_page(page, order);
2453 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2454 * allocate the page. The expectation is that the caller is taking
2455 * steps that will free more memory. The caller should avoid the page
2456 * being used for !PFMEMALLOC purposes.
2458 if (alloc_flags & ALLOC_NO_WATERMARKS)
2459 set_page_pfmemalloc(page);
2461 clear_page_pfmemalloc(page);
2465 * Go through the free lists for the given migratetype and remove
2466 * the smallest available page from the freelists
2468 static __always_inline
2469 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2472 unsigned int current_order;
2473 struct free_area *area;
2476 /* Find a page of the appropriate size in the preferred list */
2477 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2478 area = &(zone->free_area[current_order]);
2479 page = get_page_from_free_area(area, migratetype);
2482 del_page_from_free_list(page, zone, current_order);
2483 expand(zone, page, order, current_order, migratetype);
2484 set_pcppage_migratetype(page, migratetype);
2493 * This array describes the order lists are fallen back to when
2494 * the free lists for the desirable migrate type are depleted
2496 static int fallbacks[MIGRATE_TYPES][3] = {
2497 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2498 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2499 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2501 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2503 #ifdef CONFIG_MEMORY_ISOLATION
2504 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2509 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2512 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2515 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2516 unsigned int order) { return NULL; }
2520 * Move the free pages in a range to the freelist tail of the requested type.
2521 * Note that start_page and end_pages are not aligned on a pageblock
2522 * boundary. If alignment is required, use move_freepages_block()
2524 static int move_freepages(struct zone *zone,
2525 unsigned long start_pfn, unsigned long end_pfn,
2526 int migratetype, int *num_movable)
2531 int pages_moved = 0;
2533 for (pfn = start_pfn; pfn <= end_pfn;) {
2534 page = pfn_to_page(pfn);
2535 if (!PageBuddy(page)) {
2537 * We assume that pages that could be isolated for
2538 * migration are movable. But we don't actually try
2539 * isolating, as that would be expensive.
2542 (PageLRU(page) || __PageMovable(page)))
2548 /* Make sure we are not inadvertently changing nodes */
2549 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2550 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2552 order = buddy_order(page);
2553 move_to_free_list(page, zone, order, migratetype);
2555 pages_moved += 1 << order;
2561 int move_freepages_block(struct zone *zone, struct page *page,
2562 int migratetype, int *num_movable)
2564 unsigned long start_pfn, end_pfn, pfn;
2569 pfn = page_to_pfn(page);
2570 start_pfn = pfn & ~(pageblock_nr_pages - 1);
2571 end_pfn = start_pfn + pageblock_nr_pages - 1;
2573 /* Do not cross zone boundaries */
2574 if (!zone_spans_pfn(zone, start_pfn))
2576 if (!zone_spans_pfn(zone, end_pfn))
2579 return move_freepages(zone, start_pfn, end_pfn, migratetype,
2583 static void change_pageblock_range(struct page *pageblock_page,
2584 int start_order, int migratetype)
2586 int nr_pageblocks = 1 << (start_order - pageblock_order);
2588 while (nr_pageblocks--) {
2589 set_pageblock_migratetype(pageblock_page, migratetype);
2590 pageblock_page += pageblock_nr_pages;
2595 * When we are falling back to another migratetype during allocation, try to
2596 * steal extra free pages from the same pageblocks to satisfy further
2597 * allocations, instead of polluting multiple pageblocks.
2599 * If we are stealing a relatively large buddy page, it is likely there will
2600 * be more free pages in the pageblock, so try to steal them all. For
2601 * reclaimable and unmovable allocations, we steal regardless of page size,
2602 * as fragmentation caused by those allocations polluting movable pageblocks
2603 * is worse than movable allocations stealing from unmovable and reclaimable
2606 static bool can_steal_fallback(unsigned int order, int start_mt)
2609 * Leaving this order check is intended, although there is
2610 * relaxed order check in next check. The reason is that
2611 * we can actually steal whole pageblock if this condition met,
2612 * but, below check doesn't guarantee it and that is just heuristic
2613 * so could be changed anytime.
2615 if (order >= pageblock_order)
2618 if (order >= pageblock_order / 2 ||
2619 start_mt == MIGRATE_RECLAIMABLE ||
2620 start_mt == MIGRATE_UNMOVABLE ||
2621 page_group_by_mobility_disabled)
2627 static inline bool boost_watermark(struct zone *zone)
2629 unsigned long max_boost;
2631 if (!watermark_boost_factor)
2634 * Don't bother in zones that are unlikely to produce results.
2635 * On small machines, including kdump capture kernels running
2636 * in a small area, boosting the watermark can cause an out of
2637 * memory situation immediately.
2639 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2642 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2643 watermark_boost_factor, 10000);
2646 * high watermark may be uninitialised if fragmentation occurs
2647 * very early in boot so do not boost. We do not fall
2648 * through and boost by pageblock_nr_pages as failing
2649 * allocations that early means that reclaim is not going
2650 * to help and it may even be impossible to reclaim the
2651 * boosted watermark resulting in a hang.
2656 max_boost = max(pageblock_nr_pages, max_boost);
2658 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2665 * This function implements actual steal behaviour. If order is large enough,
2666 * we can steal whole pageblock. If not, we first move freepages in this
2667 * pageblock to our migratetype and determine how many already-allocated pages
2668 * are there in the pageblock with a compatible migratetype. If at least half
2669 * of pages are free or compatible, we can change migratetype of the pageblock
2670 * itself, so pages freed in the future will be put on the correct free list.
2672 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2673 unsigned int alloc_flags, int start_type, bool whole_block)
2675 unsigned int current_order = buddy_order(page);
2676 int free_pages, movable_pages, alike_pages;
2679 old_block_type = get_pageblock_migratetype(page);
2682 * This can happen due to races and we want to prevent broken
2683 * highatomic accounting.
2685 if (is_migrate_highatomic(old_block_type))
2688 /* Take ownership for orders >= pageblock_order */
2689 if (current_order >= pageblock_order) {
2690 change_pageblock_range(page, current_order, start_type);
2695 * Boost watermarks to increase reclaim pressure to reduce the
2696 * likelihood of future fallbacks. Wake kswapd now as the node
2697 * may be balanced overall and kswapd will not wake naturally.
2699 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2700 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2702 /* We are not allowed to try stealing from the whole block */
2706 free_pages = move_freepages_block(zone, page, start_type,
2709 * Determine how many pages are compatible with our allocation.
2710 * For movable allocation, it's the number of movable pages which
2711 * we just obtained. For other types it's a bit more tricky.
2713 if (start_type == MIGRATE_MOVABLE) {
2714 alike_pages = movable_pages;
2717 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2718 * to MOVABLE pageblock, consider all non-movable pages as
2719 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2720 * vice versa, be conservative since we can't distinguish the
2721 * exact migratetype of non-movable pages.
2723 if (old_block_type == MIGRATE_MOVABLE)
2724 alike_pages = pageblock_nr_pages
2725 - (free_pages + movable_pages);
2730 /* moving whole block can fail due to zone boundary conditions */
2735 * If a sufficient number of pages in the block are either free or of
2736 * comparable migratability as our allocation, claim the whole block.
2738 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2739 page_group_by_mobility_disabled)
2740 set_pageblock_migratetype(page, start_type);
2745 move_to_free_list(page, zone, current_order, start_type);
2749 * Check whether there is a suitable fallback freepage with requested order.
2750 * If only_stealable is true, this function returns fallback_mt only if
2751 * we can steal other freepages all together. This would help to reduce
2752 * fragmentation due to mixed migratetype pages in one pageblock.
2754 int find_suitable_fallback(struct free_area *area, unsigned int order,
2755 int migratetype, bool only_stealable, bool *can_steal)
2760 if (area->nr_free == 0)
2765 fallback_mt = fallbacks[migratetype][i];
2766 if (fallback_mt == MIGRATE_TYPES)
2769 if (free_area_empty(area, fallback_mt))
2772 if (can_steal_fallback(order, migratetype))
2775 if (!only_stealable)
2786 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2787 * there are no empty page blocks that contain a page with a suitable order
2789 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2790 unsigned int alloc_order)
2793 unsigned long max_managed, flags;
2796 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2797 * Check is race-prone but harmless.
2799 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2800 if (zone->nr_reserved_highatomic >= max_managed)
2803 spin_lock_irqsave(&zone->lock, flags);
2805 /* Recheck the nr_reserved_highatomic limit under the lock */
2806 if (zone->nr_reserved_highatomic >= max_managed)
2810 mt = get_pageblock_migratetype(page);
2811 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2812 && !is_migrate_cma(mt)) {
2813 zone->nr_reserved_highatomic += pageblock_nr_pages;
2814 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2815 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2819 spin_unlock_irqrestore(&zone->lock, flags);
2823 * Used when an allocation is about to fail under memory pressure. This
2824 * potentially hurts the reliability of high-order allocations when under
2825 * intense memory pressure but failed atomic allocations should be easier
2826 * to recover from than an OOM.
2828 * If @force is true, try to unreserve a pageblock even though highatomic
2829 * pageblock is exhausted.
2831 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2834 struct zonelist *zonelist = ac->zonelist;
2835 unsigned long flags;
2842 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2845 * Preserve at least one pageblock unless memory pressure
2848 if (!force && zone->nr_reserved_highatomic <=
2852 spin_lock_irqsave(&zone->lock, flags);
2853 for (order = 0; order < MAX_ORDER; order++) {
2854 struct free_area *area = &(zone->free_area[order]);
2856 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2861 * In page freeing path, migratetype change is racy so
2862 * we can counter several free pages in a pageblock
2863 * in this loop although we changed the pageblock type
2864 * from highatomic to ac->migratetype. So we should
2865 * adjust the count once.
2867 if (is_migrate_highatomic_page(page)) {
2869 * It should never happen but changes to
2870 * locking could inadvertently allow a per-cpu
2871 * drain to add pages to MIGRATE_HIGHATOMIC
2872 * while unreserving so be safe and watch for
2875 zone->nr_reserved_highatomic -= min(
2877 zone->nr_reserved_highatomic);
2881 * Convert to ac->migratetype and avoid the normal
2882 * pageblock stealing heuristics. Minimally, the caller
2883 * is doing the work and needs the pages. More
2884 * importantly, if the block was always converted to
2885 * MIGRATE_UNMOVABLE or another type then the number
2886 * of pageblocks that cannot be completely freed
2889 set_pageblock_migratetype(page, ac->migratetype);
2890 ret = move_freepages_block(zone, page, ac->migratetype,
2893 spin_unlock_irqrestore(&zone->lock, flags);
2897 spin_unlock_irqrestore(&zone->lock, flags);
2904 * Try finding a free buddy page on the fallback list and put it on the free
2905 * list of requested migratetype, possibly along with other pages from the same
2906 * block, depending on fragmentation avoidance heuristics. Returns true if
2907 * fallback was found so that __rmqueue_smallest() can grab it.
2909 * The use of signed ints for order and current_order is a deliberate
2910 * deviation from the rest of this file, to make the for loop
2911 * condition simpler.
2913 static __always_inline bool
2914 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2915 unsigned int alloc_flags)
2917 struct free_area *area;
2919 int min_order = order;
2925 * Do not steal pages from freelists belonging to other pageblocks
2926 * i.e. orders < pageblock_order. If there are no local zones free,
2927 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2929 if (alloc_flags & ALLOC_NOFRAGMENT)
2930 min_order = pageblock_order;
2933 * Find the largest available free page in the other list. This roughly
2934 * approximates finding the pageblock with the most free pages, which
2935 * would be too costly to do exactly.
2937 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2939 area = &(zone->free_area[current_order]);
2940 fallback_mt = find_suitable_fallback(area, current_order,
2941 start_migratetype, false, &can_steal);
2942 if (fallback_mt == -1)
2946 * We cannot steal all free pages from the pageblock and the
2947 * requested migratetype is movable. In that case it's better to
2948 * steal and split the smallest available page instead of the
2949 * largest available page, because even if the next movable
2950 * allocation falls back into a different pageblock than this
2951 * one, it won't cause permanent fragmentation.
2953 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2954 && current_order > order)
2963 for (current_order = order; current_order < MAX_ORDER;
2965 area = &(zone->free_area[current_order]);
2966 fallback_mt = find_suitable_fallback(area, current_order,
2967 start_migratetype, false, &can_steal);
2968 if (fallback_mt != -1)
2973 * This should not happen - we already found a suitable fallback
2974 * when looking for the largest page.
2976 VM_BUG_ON(current_order == MAX_ORDER);
2979 page = get_page_from_free_area(area, fallback_mt);
2981 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2984 trace_mm_page_alloc_extfrag(page, order, current_order,
2985 start_migratetype, fallback_mt);
2992 * Do the hard work of removing an element from the buddy allocator.
2993 * Call me with the zone->lock already held.
2995 static __always_inline struct page *
2996 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2997 unsigned int alloc_flags)
3001 if (IS_ENABLED(CONFIG_CMA)) {
3003 * Balance movable allocations between regular and CMA areas by
3004 * allocating from CMA when over more than a given proportion of
3005 * the zone's free memory is in the CMA area.
3007 if (alloc_flags & ALLOC_CMA &&
3008 zone_page_state(zone, NR_FREE_CMA_PAGES) >
3009 zone_page_state(zone, NR_FREE_PAGES) / ALLOC_IN_CMA_THRESHOLD_MAX
3010 * _alloc_in_cma_threshold) {
3011 page = __rmqueue_cma_fallback(zone, order);
3017 page = __rmqueue_smallest(zone, order, migratetype);
3018 if (unlikely(!page)) {
3019 if (alloc_flags & ALLOC_CMA)
3020 page = __rmqueue_cma_fallback(zone, order);
3022 if (!page && __rmqueue_fallback(zone, order, migratetype,
3028 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3033 * Obtain a specified number of elements from the buddy allocator, all under
3034 * a single hold of the lock, for efficiency. Add them to the supplied list.
3035 * Returns the number of new pages which were placed at *list.
3037 static int rmqueue_bulk(struct zone *zone, unsigned int order,
3038 unsigned long count, struct list_head *list,
3039 int migratetype, unsigned int alloc_flags)
3041 int i, allocated = 0;
3044 * local_lock_irq held so equivalent to spin_lock_irqsave for
3045 * both PREEMPT_RT and non-PREEMPT_RT configurations.
3047 spin_lock(&zone->lock);
3048 for (i = 0; i < count; ++i) {
3049 struct page *page = __rmqueue(zone, order, migratetype,
3051 if (unlikely(page == NULL))
3054 if (unlikely(check_pcp_refill(page)))
3058 * Split buddy pages returned by expand() are received here in
3059 * physical page order. The page is added to the tail of
3060 * caller's list. From the callers perspective, the linked list
3061 * is ordered by page number under some conditions. This is
3062 * useful for IO devices that can forward direction from the
3063 * head, thus also in the physical page order. This is useful
3064 * for IO devices that can merge IO requests if the physical
3065 * pages are ordered properly.
3067 list_add_tail(&page->lru, list);
3069 if (is_migrate_cma(get_pcppage_migratetype(page)))
3070 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
3075 * i pages were removed from the buddy list even if some leak due
3076 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
3077 * on i. Do not confuse with 'allocated' which is the number of
3078 * pages added to the pcp list.
3080 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
3081 spin_unlock(&zone->lock);
3087 * Called from the vmstat counter updater to drain pagesets of this
3088 * currently executing processor on remote nodes after they have
3091 * Note that this function must be called with the thread pinned to
3092 * a single processor.
3094 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
3096 unsigned long flags;
3097 int to_drain, batch;
3099 local_lock_irqsave(&pagesets.lock, flags);
3100 batch = READ_ONCE(pcp->batch);
3101 to_drain = min(pcp->count, batch);
3103 free_pcppages_bulk(zone, to_drain, pcp);
3104 local_unlock_irqrestore(&pagesets.lock, flags);
3109 * Drain pcplists of the indicated processor and zone.
3111 * The processor must either be the current processor and the
3112 * thread pinned to the current processor or a processor that
3115 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
3117 unsigned long flags;
3118 struct per_cpu_pages *pcp;
3120 local_lock_irqsave(&pagesets.lock, flags);
3122 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3124 free_pcppages_bulk(zone, pcp->count, pcp);
3126 local_unlock_irqrestore(&pagesets.lock, flags);
3130 * Drain pcplists of all zones on the indicated processor.
3132 * The processor must either be the current processor and the
3133 * thread pinned to the current processor or a processor that
3136 static void drain_pages(unsigned int cpu)
3140 for_each_populated_zone(zone) {
3141 drain_pages_zone(cpu, zone);
3146 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3148 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3149 * the single zone's pages.
3151 void drain_local_pages(struct zone *zone)
3153 int cpu = smp_processor_id();
3156 drain_pages_zone(cpu, zone);
3161 static void drain_local_pages_wq(struct work_struct *work)
3163 struct pcpu_drain *drain;
3165 drain = container_of(work, struct pcpu_drain, work);
3168 * drain_all_pages doesn't use proper cpu hotplug protection so
3169 * we can race with cpu offline when the WQ can move this from
3170 * a cpu pinned worker to an unbound one. We can operate on a different
3171 * cpu which is alright but we also have to make sure to not move to
3175 drain_local_pages(drain->zone);
3180 * The implementation of drain_all_pages(), exposing an extra parameter to
3181 * drain on all cpus.
3183 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3184 * not empty. The check for non-emptiness can however race with a free to
3185 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3186 * that need the guarantee that every CPU has drained can disable the
3187 * optimizing racy check.
3189 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3194 * Allocate in the BSS so we won't require allocation in
3195 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3197 static cpumask_t cpus_with_pcps;
3200 * Make sure nobody triggers this path before mm_percpu_wq is fully
3203 if (WARN_ON_ONCE(!mm_percpu_wq))
3207 * Do not drain if one is already in progress unless it's specific to
3208 * a zone. Such callers are primarily CMA and memory hotplug and need
3209 * the drain to be complete when the call returns.
3211 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3214 mutex_lock(&pcpu_drain_mutex);
3218 * We don't care about racing with CPU hotplug event
3219 * as offline notification will cause the notified
3220 * cpu to drain that CPU pcps and on_each_cpu_mask
3221 * disables preemption as part of its processing
3223 for_each_online_cpu(cpu) {
3224 struct per_cpu_pages *pcp;
3226 bool has_pcps = false;
3228 if (force_all_cpus) {
3230 * The pcp.count check is racy, some callers need a
3231 * guarantee that no cpu is missed.
3235 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
3239 for_each_populated_zone(z) {
3240 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
3249 cpumask_set_cpu(cpu, &cpus_with_pcps);
3251 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3254 for_each_cpu(cpu, &cpus_with_pcps) {
3255 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3258 INIT_WORK(&drain->work, drain_local_pages_wq);
3259 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3261 for_each_cpu(cpu, &cpus_with_pcps)
3262 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3264 mutex_unlock(&pcpu_drain_mutex);
3268 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3270 * When zone parameter is non-NULL, spill just the single zone's pages.
3272 * Note that this can be extremely slow as the draining happens in a workqueue.
3274 void drain_all_pages(struct zone *zone)
3276 __drain_all_pages(zone, false);
3279 #ifdef CONFIG_HIBERNATION
3282 * Touch the watchdog for every WD_PAGE_COUNT pages.
3284 #define WD_PAGE_COUNT (128*1024)
3286 void mark_free_pages(struct zone *zone)
3288 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3289 unsigned long flags;
3290 unsigned int order, t;
3293 if (zone_is_empty(zone))
3296 spin_lock_irqsave(&zone->lock, flags);
3298 max_zone_pfn = zone_end_pfn(zone);
3299 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3300 if (pfn_valid(pfn)) {
3301 page = pfn_to_page(pfn);
3303 if (!--page_count) {
3304 touch_nmi_watchdog();
3305 page_count = WD_PAGE_COUNT;
3308 if (page_zone(page) != zone)
3311 if (!swsusp_page_is_forbidden(page))
3312 swsusp_unset_page_free(page);
3315 for_each_migratetype_order(order, t) {
3316 list_for_each_entry(page,
3317 &zone->free_area[order].free_list[t], lru) {
3320 pfn = page_to_pfn(page);
3321 for (i = 0; i < (1UL << order); i++) {
3322 if (!--page_count) {
3323 touch_nmi_watchdog();
3324 page_count = WD_PAGE_COUNT;
3326 swsusp_set_page_free(pfn_to_page(pfn + i));
3330 spin_unlock_irqrestore(&zone->lock, flags);
3332 #endif /* CONFIG_PM */
3334 static bool free_unref_page_prepare(struct page *page, unsigned long pfn,
3339 if (!free_pcp_prepare(page, order))
3342 migratetype = get_pfnblock_migratetype(page, pfn);
3343 set_pcppage_migratetype(page, migratetype);
3347 static int nr_pcp_free(struct per_cpu_pages *pcp, int high, int batch)
3349 int min_nr_free, max_nr_free;
3351 /* Check for PCP disabled or boot pageset */
3352 if (unlikely(high < batch))
3355 /* Leave at least pcp->batch pages on the list */
3356 min_nr_free = batch;
3357 max_nr_free = high - batch;
3360 * Double the number of pages freed each time there is subsequent
3361 * freeing of pages without any allocation.
3363 batch <<= pcp->free_factor;
3364 if (batch < max_nr_free)
3366 batch = clamp(batch, min_nr_free, max_nr_free);
3371 static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone)
3373 int high = READ_ONCE(pcp->high);
3375 if (unlikely(!high))
3378 if (!test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags))
3382 * If reclaim is active, limit the number of pages that can be
3383 * stored on pcp lists
3385 return min(READ_ONCE(pcp->batch) << 2, high);
3388 static void free_unref_page_commit(struct page *page, unsigned long pfn,
3389 int migratetype, unsigned int order)
3391 struct zone *zone = page_zone(page);
3392 struct per_cpu_pages *pcp;
3396 __count_vm_event(PGFREE);
3397 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3398 pindex = order_to_pindex(migratetype, order);
3399 list_add(&page->lru, &pcp->lists[pindex]);
3400 pcp->count += 1 << order;
3401 high = nr_pcp_high(pcp, zone);
3402 if (pcp->count >= high) {
3403 int batch = READ_ONCE(pcp->batch);
3405 free_pcppages_bulk(zone, nr_pcp_free(pcp, high, batch), pcp);
3412 void free_unref_page(struct page *page, unsigned int order)
3414 unsigned long flags;
3415 unsigned long pfn = page_to_pfn(page);
3418 if (!free_unref_page_prepare(page, pfn, order))
3422 * We only track unmovable, reclaimable and movable on pcp lists.
3423 * Place ISOLATE pages on the isolated list because they are being
3424 * offlined but treat HIGHATOMIC as movable pages so we can get those
3425 * areas back if necessary. Otherwise, we may have to free
3426 * excessively into the page allocator
3428 migratetype = get_pcppage_migratetype(page);
3429 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
3430 if (unlikely(is_migrate_isolate(migratetype))) {
3431 free_one_page(page_zone(page), page, pfn, order, migratetype, FPI_NONE);
3434 migratetype = MIGRATE_MOVABLE;
3437 local_lock_irqsave(&pagesets.lock, flags);
3438 free_unref_page_commit(page, pfn, migratetype, order);
3439 local_unlock_irqrestore(&pagesets.lock, flags);
3443 * Free a list of 0-order pages
3445 void free_unref_page_list(struct list_head *list)
3447 struct page *page, *next;
3448 unsigned long flags, pfn;
3449 int batch_count = 0;
3452 /* Prepare pages for freeing */
3453 list_for_each_entry_safe(page, next, list, lru) {
3454 pfn = page_to_pfn(page);
3455 if (!free_unref_page_prepare(page, pfn, 0)) {
3456 list_del(&page->lru);
3461 * Free isolated pages directly to the allocator, see
3462 * comment in free_unref_page.
3464 migratetype = get_pcppage_migratetype(page);
3465 if (unlikely(is_migrate_isolate(migratetype))) {
3466 list_del(&page->lru);
3467 free_one_page(page_zone(page), page, pfn, 0, migratetype, FPI_NONE);
3471 set_page_private(page, pfn);
3474 local_lock_irqsave(&pagesets.lock, flags);
3475 list_for_each_entry_safe(page, next, list, lru) {
3476 pfn = page_private(page);
3477 set_page_private(page, 0);
3480 * Non-isolated types over MIGRATE_PCPTYPES get added
3481 * to the MIGRATE_MOVABLE pcp list.
3483 migratetype = get_pcppage_migratetype(page);
3484 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
3485 migratetype = MIGRATE_MOVABLE;
3487 trace_mm_page_free_batched(page);
3488 free_unref_page_commit(page, pfn, migratetype, 0);
3491 * Guard against excessive IRQ disabled times when we get
3492 * a large list of pages to free.
3494 if (++batch_count == SWAP_CLUSTER_MAX) {
3495 local_unlock_irqrestore(&pagesets.lock, flags);
3497 local_lock_irqsave(&pagesets.lock, flags);
3500 local_unlock_irqrestore(&pagesets.lock, flags);
3504 * split_page takes a non-compound higher-order page, and splits it into
3505 * n (1<<order) sub-pages: page[0..n]
3506 * Each sub-page must be freed individually.
3508 * Note: this is probably too low level an operation for use in drivers.
3509 * Please consult with lkml before using this in your driver.
3511 void split_page(struct page *page, unsigned int order)
3515 VM_BUG_ON_PAGE(PageCompound(page), page);
3516 VM_BUG_ON_PAGE(!page_count(page), page);
3518 for (i = 1; i < (1 << order); i++)
3519 set_page_refcounted(page + i);
3520 split_page_owner(page, 1 << order);
3521 split_page_memcg(page, 1 << order);
3523 EXPORT_SYMBOL_GPL(split_page);
3525 int __isolate_free_page(struct page *page, unsigned int order)
3527 unsigned long watermark;
3531 BUG_ON(!PageBuddy(page));
3533 zone = page_zone(page);
3534 mt = get_pageblock_migratetype(page);
3536 if (!is_migrate_isolate(mt)) {
3538 * Obey watermarks as if the page was being allocated. We can
3539 * emulate a high-order watermark check with a raised order-0
3540 * watermark, because we already know our high-order page
3543 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3544 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3547 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3550 /* Remove page from free list */
3552 del_page_from_free_list(page, zone, order);
3555 * Set the pageblock if the isolated page is at least half of a
3558 if (order >= pageblock_order - 1) {
3559 struct page *endpage = page + (1 << order) - 1;
3560 for (; page < endpage; page += pageblock_nr_pages) {
3561 int mt = get_pageblock_migratetype(page);
3562 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3563 && !is_migrate_highatomic(mt))
3564 set_pageblock_migratetype(page,
3570 return 1UL << order;
3574 * __putback_isolated_page - Return a now-isolated page back where we got it
3575 * @page: Page that was isolated
3576 * @order: Order of the isolated page
3577 * @mt: The page's pageblock's migratetype
3579 * This function is meant to return a page pulled from the free lists via
3580 * __isolate_free_page back to the free lists they were pulled from.
3582 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3584 struct zone *zone = page_zone(page);
3586 /* zone lock should be held when this function is called */
3587 lockdep_assert_held(&zone->lock);
3589 /* Return isolated page to tail of freelist. */
3590 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3591 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3595 * Update NUMA hit/miss statistics
3597 * Must be called with interrupts disabled.
3599 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
3603 enum numa_stat_item local_stat = NUMA_LOCAL;
3605 /* skip numa counters update if numa stats is disabled */
3606 if (!static_branch_likely(&vm_numa_stat_key))
3609 if (zone_to_nid(z) != numa_node_id())
3610 local_stat = NUMA_OTHER;
3612 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3613 __count_numa_events(z, NUMA_HIT, nr_account);
3615 __count_numa_events(z, NUMA_MISS, nr_account);
3616 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
3618 __count_numa_events(z, local_stat, nr_account);
3622 /* Remove page from the per-cpu list, caller must protect the list */
3624 struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
3626 unsigned int alloc_flags,
3627 struct per_cpu_pages *pcp,
3628 struct list_head *list)
3633 if (list_empty(list)) {
3634 int batch = READ_ONCE(pcp->batch);
3638 * Scale batch relative to order if batch implies
3639 * free pages can be stored on the PCP. Batch can
3640 * be 1 for small zones or for boot pagesets which
3641 * should never store free pages as the pages may
3642 * belong to arbitrary zones.
3645 batch = max(batch >> order, 2);
3646 alloced = rmqueue_bulk(zone, order,
3648 migratetype, alloc_flags);
3650 pcp->count += alloced << order;
3651 if (unlikely(list_empty(list)))
3655 page = list_first_entry(list, struct page, lru);
3656 list_del(&page->lru);
3657 pcp->count -= 1 << order;
3658 } while (check_new_pcp(page));
3663 /* Lock and remove page from the per-cpu list */
3664 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3665 struct zone *zone, unsigned int order,
3666 gfp_t gfp_flags, int migratetype,
3667 unsigned int alloc_flags)
3669 struct per_cpu_pages *pcp;
3670 struct list_head *list;
3672 unsigned long flags;
3674 local_lock_irqsave(&pagesets.lock, flags);
3677 * On allocation, reduce the number of pages that are batch freed.
3678 * See nr_pcp_free() where free_factor is increased for subsequent
3681 pcp = this_cpu_ptr(zone->per_cpu_pageset);
3682 pcp->free_factor >>= 1;
3683 list = &pcp->lists[order_to_pindex(migratetype, order)];
3684 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3685 local_unlock_irqrestore(&pagesets.lock, flags);
3687 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3688 zone_statistics(preferred_zone, zone, 1);
3694 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3697 struct page *rmqueue(struct zone *preferred_zone,
3698 struct zone *zone, unsigned int order,
3699 gfp_t gfp_flags, unsigned int alloc_flags,
3702 unsigned long flags;
3705 if (likely(pcp_allowed_order(order))) {
3707 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3708 * we need to skip it when CMA area isn't allowed.
3710 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3711 migratetype != MIGRATE_MOVABLE) {
3712 page = rmqueue_pcplist(preferred_zone, zone, order,
3713 gfp_flags, migratetype, alloc_flags);
3719 * We most definitely don't want callers attempting to
3720 * allocate greater than order-1 page units with __GFP_NOFAIL.
3722 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3723 spin_lock_irqsave(&zone->lock, flags);
3728 * order-0 request can reach here when the pcplist is skipped
3729 * due to non-CMA allocation context. HIGHATOMIC area is
3730 * reserved for high-order atomic allocation, so order-0
3731 * request should skip it.
3733 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3734 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3736 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3739 page = __rmqueue(zone, order, migratetype, alloc_flags);
3740 } while (page && check_new_pages(page, order));
3744 __mod_zone_freepage_state(zone, -(1 << order),
3745 get_pcppage_migratetype(page));
3746 spin_unlock_irqrestore(&zone->lock, flags);
3748 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3749 zone_statistics(preferred_zone, zone, 1);
3752 /* Separate test+clear to avoid unnecessary atomics */
3753 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3754 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3755 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3758 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3762 spin_unlock_irqrestore(&zone->lock, flags);
3766 #ifdef CONFIG_FAIL_PAGE_ALLOC
3769 struct fault_attr attr;
3771 bool ignore_gfp_highmem;
3772 bool ignore_gfp_reclaim;
3774 } fail_page_alloc = {
3775 .attr = FAULT_ATTR_INITIALIZER,
3776 .ignore_gfp_reclaim = true,
3777 .ignore_gfp_highmem = true,
3781 static int __init setup_fail_page_alloc(char *str)
3783 return setup_fault_attr(&fail_page_alloc.attr, str);
3785 __setup("fail_page_alloc=", setup_fail_page_alloc);
3787 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3789 if (order < fail_page_alloc.min_order)
3791 if (gfp_mask & __GFP_NOFAIL)
3793 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3795 if (fail_page_alloc.ignore_gfp_reclaim &&
3796 (gfp_mask & __GFP_DIRECT_RECLAIM))
3799 return should_fail(&fail_page_alloc.attr, 1 << order);
3802 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3804 static int __init fail_page_alloc_debugfs(void)
3806 umode_t mode = S_IFREG | 0600;
3809 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3810 &fail_page_alloc.attr);
3812 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3813 &fail_page_alloc.ignore_gfp_reclaim);
3814 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3815 &fail_page_alloc.ignore_gfp_highmem);
3816 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3821 late_initcall(fail_page_alloc_debugfs);
3823 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3825 #else /* CONFIG_FAIL_PAGE_ALLOC */
3827 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3832 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3834 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3836 return __should_fail_alloc_page(gfp_mask, order);
3838 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3840 static inline long __zone_watermark_unusable_free(struct zone *z,
3841 unsigned int order, unsigned int alloc_flags)
3843 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3844 long unusable_free = (1 << order) - 1;
3847 * If the caller does not have rights to ALLOC_HARDER then subtract
3848 * the high-atomic reserves. This will over-estimate the size of the
3849 * atomic reserve but it avoids a search.
3851 if (likely(!alloc_harder))
3852 unusable_free += z->nr_reserved_highatomic;
3855 /* If allocation can't use CMA areas don't use free CMA pages */
3856 if (!(alloc_flags & ALLOC_CMA))
3857 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3860 return unusable_free;
3864 * Return true if free base pages are above 'mark'. For high-order checks it
3865 * will return true of the order-0 watermark is reached and there is at least
3866 * one free page of a suitable size. Checking now avoids taking the zone lock
3867 * to check in the allocation paths if no pages are free.
3869 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3870 int highest_zoneidx, unsigned int alloc_flags,
3875 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3877 /* free_pages may go negative - that's OK */
3878 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3880 if (alloc_flags & ALLOC_HIGH)
3883 if (unlikely(alloc_harder)) {
3885 * OOM victims can try even harder than normal ALLOC_HARDER
3886 * users on the grounds that it's definitely going to be in
3887 * the exit path shortly and free memory. Any allocation it
3888 * makes during the free path will be small and short-lived.
3890 if (alloc_flags & ALLOC_OOM)
3897 * Check watermarks for an order-0 allocation request. If these
3898 * are not met, then a high-order request also cannot go ahead
3899 * even if a suitable page happened to be free.
3901 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3904 /* If this is an order-0 request then the watermark is fine */
3908 /* For a high-order request, check at least one suitable page is free */
3909 for (o = order; o < MAX_ORDER; o++) {
3910 struct free_area *area = &z->free_area[o];
3916 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3917 if (!free_area_empty(area, mt))
3922 if ((alloc_flags & ALLOC_CMA) &&
3923 !free_area_empty(area, MIGRATE_CMA)) {
3927 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3933 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3934 int highest_zoneidx, unsigned int alloc_flags)
3936 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3937 zone_page_state(z, NR_FREE_PAGES));
3940 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3941 unsigned long mark, int highest_zoneidx,
3942 unsigned int alloc_flags, gfp_t gfp_mask)
3946 free_pages = zone_page_state(z, NR_FREE_PAGES);
3949 * Fast check for order-0 only. If this fails then the reserves
3950 * need to be calculated.
3956 usable_free = free_pages;
3957 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3959 /* reserved may over estimate high-atomic reserves. */
3960 usable_free -= min(usable_free, reserved);
3961 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3965 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3969 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3970 * when checking the min watermark. The min watermark is the
3971 * point where boosting is ignored so that kswapd is woken up
3972 * when below the low watermark.
3974 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3975 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3976 mark = z->_watermark[WMARK_MIN];
3977 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3978 alloc_flags, free_pages);
3984 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3985 unsigned long mark, int highest_zoneidx)
3987 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3989 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3990 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3992 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3997 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3999 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
4000 node_reclaim_distance;
4002 #else /* CONFIG_NUMA */
4003 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
4007 #endif /* CONFIG_NUMA */
4010 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
4011 * fragmentation is subtle. If the preferred zone was HIGHMEM then
4012 * premature use of a lower zone may cause lowmem pressure problems that
4013 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
4014 * probably too small. It only makes sense to spread allocations to avoid
4015 * fragmentation between the Normal and DMA32 zones.
4017 static inline unsigned int
4018 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
4020 unsigned int alloc_flags;
4023 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4026 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
4028 #ifdef CONFIG_ZONE_DMA32
4032 if (zone_idx(zone) != ZONE_NORMAL)
4036 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
4037 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
4038 * on UMA that if Normal is populated then so is DMA32.
4040 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
4041 if (nr_online_nodes > 1 && !populated_zone(--zone))
4044 alloc_flags |= ALLOC_NOFRAGMENT;
4045 #endif /* CONFIG_ZONE_DMA32 */
4049 /* Must be called after current_gfp_context() which can change gfp_mask */
4050 static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
4051 unsigned int alloc_flags)
4054 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4055 alloc_flags |= ALLOC_CMA;
4061 * get_page_from_freelist goes through the zonelist trying to allocate
4064 static struct page *
4065 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
4066 const struct alloc_context *ac)
4070 struct pglist_data *last_pgdat_dirty_limit = NULL;
4075 * Scan zonelist, looking for a zone with enough free.
4076 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
4078 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
4079 z = ac->preferred_zoneref;
4080 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
4085 if (cpusets_enabled() &&
4086 (alloc_flags & ALLOC_CPUSET) &&
4087 !__cpuset_zone_allowed(zone, gfp_mask))
4090 * When allocating a page cache page for writing, we
4091 * want to get it from a node that is within its dirty
4092 * limit, such that no single node holds more than its
4093 * proportional share of globally allowed dirty pages.
4094 * The dirty limits take into account the node's
4095 * lowmem reserves and high watermark so that kswapd
4096 * should be able to balance it without having to
4097 * write pages from its LRU list.
4099 * XXX: For now, allow allocations to potentially
4100 * exceed the per-node dirty limit in the slowpath
4101 * (spread_dirty_pages unset) before going into reclaim,
4102 * which is important when on a NUMA setup the allowed
4103 * nodes are together not big enough to reach the
4104 * global limit. The proper fix for these situations
4105 * will require awareness of nodes in the
4106 * dirty-throttling and the flusher threads.
4108 if (ac->spread_dirty_pages) {
4109 if (last_pgdat_dirty_limit == zone->zone_pgdat)
4112 if (!node_dirty_ok(zone->zone_pgdat)) {
4113 last_pgdat_dirty_limit = zone->zone_pgdat;
4118 if (no_fallback && nr_online_nodes > 1 &&
4119 zone != ac->preferred_zoneref->zone) {
4123 * If moving to a remote node, retry but allow
4124 * fragmenting fallbacks. Locality is more important
4125 * than fragmentation avoidance.
4127 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
4128 if (zone_to_nid(zone) != local_nid) {
4129 alloc_flags &= ~ALLOC_NOFRAGMENT;
4134 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
4135 if (!zone_watermark_fast(zone, order, mark,
4136 ac->highest_zoneidx, alloc_flags,
4140 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4142 * Watermark failed for this zone, but see if we can
4143 * grow this zone if it contains deferred pages.
4145 if (static_branch_unlikely(&deferred_pages)) {
4146 if (_deferred_grow_zone(zone, order))
4150 /* Checked here to keep the fast path fast */
4151 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
4152 if (alloc_flags & ALLOC_NO_WATERMARKS)
4155 if (!node_reclaim_enabled() ||
4156 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
4159 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
4161 case NODE_RECLAIM_NOSCAN:
4164 case NODE_RECLAIM_FULL:
4165 /* scanned but unreclaimable */
4168 /* did we reclaim enough */
4169 if (zone_watermark_ok(zone, order, mark,
4170 ac->highest_zoneidx, alloc_flags))
4178 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
4179 gfp_mask, alloc_flags, ac->migratetype);
4181 prep_new_page(page, order, gfp_mask, alloc_flags);
4184 * If this is a high-order atomic allocation then check
4185 * if the pageblock should be reserved for the future
4187 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
4188 reserve_highatomic_pageblock(page, zone, order);
4192 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
4193 /* Try again if zone has deferred pages */
4194 if (static_branch_unlikely(&deferred_pages)) {
4195 if (_deferred_grow_zone(zone, order))
4203 * It's possible on a UMA machine to get through all zones that are
4204 * fragmented. If avoiding fragmentation, reset and try again.
4207 alloc_flags &= ~ALLOC_NOFRAGMENT;
4214 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
4216 unsigned int filter = SHOW_MEM_FILTER_NODES;
4219 * This documents exceptions given to allocations in certain
4220 * contexts that are allowed to allocate outside current's set
4223 if (!(gfp_mask & __GFP_NOMEMALLOC))
4224 if (tsk_is_oom_victim(current) ||
4225 (current->flags & (PF_MEMALLOC | PF_EXITING)))
4226 filter &= ~SHOW_MEM_FILTER_NODES;
4227 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
4228 filter &= ~SHOW_MEM_FILTER_NODES;
4230 show_mem(filter, nodemask);
4233 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
4235 struct va_format vaf;
4237 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4239 if ((gfp_mask & __GFP_NOWARN) ||
4240 !__ratelimit(&nopage_rs) ||
4241 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
4244 va_start(args, fmt);
4247 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4248 current->comm, &vaf, gfp_mask, &gfp_mask,
4249 nodemask_pr_args(nodemask));
4252 cpuset_print_current_mems_allowed();
4255 warn_alloc_show_mem(gfp_mask, nodemask);
4258 static inline struct page *
4259 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4260 unsigned int alloc_flags,
4261 const struct alloc_context *ac)
4265 page = get_page_from_freelist(gfp_mask, order,
4266 alloc_flags|ALLOC_CPUSET, ac);
4268 * fallback to ignore cpuset restriction if our nodes
4272 page = get_page_from_freelist(gfp_mask, order,
4278 static inline struct page *
4279 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4280 const struct alloc_context *ac, unsigned long *did_some_progress)
4282 struct oom_control oc = {
4283 .zonelist = ac->zonelist,
4284 .nodemask = ac->nodemask,
4286 .gfp_mask = gfp_mask,
4291 *did_some_progress = 0;
4294 * Acquire the oom lock. If that fails, somebody else is
4295 * making progress for us.
4297 if (!mutex_trylock(&oom_lock)) {
4298 *did_some_progress = 1;
4299 schedule_timeout_uninterruptible(1);
4304 * Go through the zonelist yet one more time, keep very high watermark
4305 * here, this is only to catch a parallel oom killing, we must fail if
4306 * we're still under heavy pressure. But make sure that this reclaim
4307 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4308 * allocation which will never fail due to oom_lock already held.
4310 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4311 ~__GFP_DIRECT_RECLAIM, order,
4312 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4316 /* Coredumps can quickly deplete all memory reserves */
4317 if (current->flags & PF_DUMPCORE)
4319 /* The OOM killer will not help higher order allocs */
4320 if (order > PAGE_ALLOC_COSTLY_ORDER)
4323 * We have already exhausted all our reclaim opportunities without any
4324 * success so it is time to admit defeat. We will skip the OOM killer
4325 * because it is very likely that the caller has a more reasonable
4326 * fallback than shooting a random task.
4328 * The OOM killer may not free memory on a specific node.
4330 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4332 /* The OOM killer does not needlessly kill tasks for lowmem */
4333 if (ac->highest_zoneidx < ZONE_NORMAL)
4335 if (pm_suspended_storage())
4338 * XXX: GFP_NOFS allocations should rather fail than rely on
4339 * other request to make a forward progress.
4340 * We are in an unfortunate situation where out_of_memory cannot
4341 * do much for this context but let's try it to at least get
4342 * access to memory reserved if the current task is killed (see
4343 * out_of_memory). Once filesystems are ready to handle allocation
4344 * failures more gracefully we should just bail out here.
4347 /* Exhausted what can be done so it's blame time */
4348 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4349 *did_some_progress = 1;
4352 * Help non-failing allocations by giving them access to memory
4355 if (gfp_mask & __GFP_NOFAIL)
4356 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4357 ALLOC_NO_WATERMARKS, ac);
4360 mutex_unlock(&oom_lock);
4365 * Maximum number of compaction retries with a progress before OOM
4366 * killer is consider as the only way to move forward.
4368 #define MAX_COMPACT_RETRIES 16
4370 #ifdef CONFIG_COMPACTION
4371 /* Try memory compaction for high-order allocations before reclaim */
4372 static struct page *
4373 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4374 unsigned int alloc_flags, const struct alloc_context *ac,
4375 enum compact_priority prio, enum compact_result *compact_result)
4377 struct page *page = NULL;
4378 unsigned long pflags;
4379 unsigned int noreclaim_flag;
4384 psi_memstall_enter(&pflags);
4385 noreclaim_flag = memalloc_noreclaim_save();
4387 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4390 memalloc_noreclaim_restore(noreclaim_flag);
4391 psi_memstall_leave(&pflags);
4393 if (*compact_result == COMPACT_SKIPPED)
4396 * At least in one zone compaction wasn't deferred or skipped, so let's
4397 * count a compaction stall
4399 count_vm_event(COMPACTSTALL);
4401 /* Prep a captured page if available */
4403 prep_new_page(page, order, gfp_mask, alloc_flags);
4405 /* Try get a page from the freelist if available */
4407 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4410 struct zone *zone = page_zone(page);
4412 zone->compact_blockskip_flush = false;
4413 compaction_defer_reset(zone, order, true);
4414 count_vm_event(COMPACTSUCCESS);
4419 * It's bad if compaction run occurs and fails. The most likely reason
4420 * is that pages exist, but not enough to satisfy watermarks.
4422 count_vm_event(COMPACTFAIL);
4430 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4431 enum compact_result compact_result,
4432 enum compact_priority *compact_priority,
4433 int *compaction_retries)
4435 int max_retries = MAX_COMPACT_RETRIES;
4438 int retries = *compaction_retries;
4439 enum compact_priority priority = *compact_priority;
4444 if (fatal_signal_pending(current))
4447 if (compaction_made_progress(compact_result))
4448 (*compaction_retries)++;
4451 * compaction considers all the zone as desperately out of memory
4452 * so it doesn't really make much sense to retry except when the
4453 * failure could be caused by insufficient priority
4455 if (compaction_failed(compact_result))
4456 goto check_priority;
4459 * compaction was skipped because there are not enough order-0 pages
4460 * to work with, so we retry only if it looks like reclaim can help.
4462 if (compaction_needs_reclaim(compact_result)) {
4463 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4468 * make sure the compaction wasn't deferred or didn't bail out early
4469 * due to locks contention before we declare that we should give up.
4470 * But the next retry should use a higher priority if allowed, so
4471 * we don't just keep bailing out endlessly.
4473 if (compaction_withdrawn(compact_result)) {
4474 goto check_priority;
4478 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4479 * costly ones because they are de facto nofail and invoke OOM
4480 * killer to move on while costly can fail and users are ready
4481 * to cope with that. 1/4 retries is rather arbitrary but we
4482 * would need much more detailed feedback from compaction to
4483 * make a better decision.
4485 if (order > PAGE_ALLOC_COSTLY_ORDER)
4487 if (*compaction_retries <= max_retries) {
4493 * Make sure there are attempts at the highest priority if we exhausted
4494 * all retries or failed at the lower priorities.
4497 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4498 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4500 if (*compact_priority > min_priority) {
4501 (*compact_priority)--;
4502 *compaction_retries = 0;
4506 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4510 static inline struct page *
4511 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4512 unsigned int alloc_flags, const struct alloc_context *ac,
4513 enum compact_priority prio, enum compact_result *compact_result)
4515 *compact_result = COMPACT_SKIPPED;
4520 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4521 enum compact_result compact_result,
4522 enum compact_priority *compact_priority,
4523 int *compaction_retries)
4528 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4532 * There are setups with compaction disabled which would prefer to loop
4533 * inside the allocator rather than hit the oom killer prematurely.
4534 * Let's give them a good hope and keep retrying while the order-0
4535 * watermarks are OK.
4537 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4538 ac->highest_zoneidx, ac->nodemask) {
4539 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4540 ac->highest_zoneidx, alloc_flags))
4545 #endif /* CONFIG_COMPACTION */
4547 #ifdef CONFIG_LOCKDEP
4548 static struct lockdep_map __fs_reclaim_map =
4549 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4551 static bool __need_reclaim(gfp_t gfp_mask)
4553 /* no reclaim without waiting on it */
4554 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4557 /* this guy won't enter reclaim */
4558 if (current->flags & PF_MEMALLOC)
4561 if (gfp_mask & __GFP_NOLOCKDEP)
4567 void __fs_reclaim_acquire(unsigned long ip)
4569 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
4572 void __fs_reclaim_release(unsigned long ip)
4574 lock_release(&__fs_reclaim_map, ip);
4577 void fs_reclaim_acquire(gfp_t gfp_mask)
4579 gfp_mask = current_gfp_context(gfp_mask);
4581 if (__need_reclaim(gfp_mask)) {
4582 if (gfp_mask & __GFP_FS)
4583 __fs_reclaim_acquire(_RET_IP_);
4585 #ifdef CONFIG_MMU_NOTIFIER
4586 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4587 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4592 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4594 void fs_reclaim_release(gfp_t gfp_mask)
4596 gfp_mask = current_gfp_context(gfp_mask);
4598 if (__need_reclaim(gfp_mask)) {
4599 if (gfp_mask & __GFP_FS)
4600 __fs_reclaim_release(_RET_IP_);
4603 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4607 * Zonelists may change due to hotplug during allocation. Detect when zonelists
4608 * have been rebuilt so allocation retries. Reader side does not lock and
4609 * retries the allocation if zonelist changes. Writer side is protected by the
4610 * embedded spin_lock.
4612 static DEFINE_SEQLOCK(zonelist_update_seq);
4614 static unsigned int zonelist_iter_begin(void)
4616 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4617 return read_seqbegin(&zonelist_update_seq);
4622 static unsigned int check_retry_zonelist(unsigned int seq)
4624 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
4625 return read_seqretry(&zonelist_update_seq, seq);
4630 /* Perform direct synchronous page reclaim */
4631 static unsigned long
4632 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4633 const struct alloc_context *ac)
4635 unsigned int noreclaim_flag;
4636 unsigned long pflags, progress;
4640 /* We now go into synchronous reclaim */
4641 cpuset_memory_pressure_bump();
4642 psi_memstall_enter(&pflags);
4643 fs_reclaim_acquire(gfp_mask);
4644 noreclaim_flag = memalloc_noreclaim_save();
4646 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4649 memalloc_noreclaim_restore(noreclaim_flag);
4650 fs_reclaim_release(gfp_mask);
4651 psi_memstall_leave(&pflags);
4658 /* The really slow allocator path where we enter direct reclaim */
4659 static inline struct page *
4660 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4661 unsigned int alloc_flags, const struct alloc_context *ac,
4662 unsigned long *did_some_progress)
4664 struct page *page = NULL;
4665 bool drained = false;
4667 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4668 if (unlikely(!(*did_some_progress)))
4672 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4675 * If an allocation failed after direct reclaim, it could be because
4676 * pages are pinned on the per-cpu lists or in high alloc reserves.
4677 * Shrink them and try again
4679 if (!page && !drained) {
4680 unreserve_highatomic_pageblock(ac, false);
4681 drain_all_pages(NULL);
4689 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4690 const struct alloc_context *ac)
4694 pg_data_t *last_pgdat = NULL;
4695 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4697 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4699 if (last_pgdat != zone->zone_pgdat)
4700 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4701 last_pgdat = zone->zone_pgdat;
4705 static inline unsigned int
4706 gfp_to_alloc_flags(gfp_t gfp_mask)
4708 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4711 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4712 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4713 * to save two branches.
4715 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4716 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4719 * The caller may dip into page reserves a bit more if the caller
4720 * cannot run direct reclaim, or if the caller has realtime scheduling
4721 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4722 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4724 alloc_flags |= (__force int)
4725 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4727 if (gfp_mask & __GFP_ATOMIC) {
4729 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4730 * if it can't schedule.
4732 if (!(gfp_mask & __GFP_NOMEMALLOC))
4733 alloc_flags |= ALLOC_HARDER;
4735 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4736 * comment for __cpuset_node_allowed().
4738 alloc_flags &= ~ALLOC_CPUSET;
4739 } else if (unlikely(rt_task(current)) && in_task())
4740 alloc_flags |= ALLOC_HARDER;
4742 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4747 static bool oom_reserves_allowed(struct task_struct *tsk)
4749 if (!tsk_is_oom_victim(tsk))
4753 * !MMU doesn't have oom reaper so give access to memory reserves
4754 * only to the thread with TIF_MEMDIE set
4756 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4763 * Distinguish requests which really need access to full memory
4764 * reserves from oom victims which can live with a portion of it
4766 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4768 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4770 if (gfp_mask & __GFP_MEMALLOC)
4771 return ALLOC_NO_WATERMARKS;
4772 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4773 return ALLOC_NO_WATERMARKS;
4774 if (!in_interrupt()) {
4775 if (current->flags & PF_MEMALLOC)
4776 return ALLOC_NO_WATERMARKS;
4777 else if (oom_reserves_allowed(current))
4784 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4786 return !!__gfp_pfmemalloc_flags(gfp_mask);
4790 * Checks whether it makes sense to retry the reclaim to make a forward progress
4791 * for the given allocation request.
4793 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4794 * without success, or when we couldn't even meet the watermark if we
4795 * reclaimed all remaining pages on the LRU lists.
4797 * Returns true if a retry is viable or false to enter the oom path.
4800 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4801 struct alloc_context *ac, int alloc_flags,
4802 bool did_some_progress, int *no_progress_loops)
4809 * Costly allocations might have made a progress but this doesn't mean
4810 * their order will become available due to high fragmentation so
4811 * always increment the no progress counter for them
4813 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4814 *no_progress_loops = 0;
4816 (*no_progress_loops)++;
4819 * Make sure we converge to OOM if we cannot make any progress
4820 * several times in the row.
4822 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4823 /* Before OOM, exhaust highatomic_reserve */
4824 return unreserve_highatomic_pageblock(ac, true);
4828 * Keep reclaiming pages while there is a chance this will lead
4829 * somewhere. If none of the target zones can satisfy our allocation
4830 * request even if all reclaimable pages are considered then we are
4831 * screwed and have to go OOM.
4833 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4834 ac->highest_zoneidx, ac->nodemask) {
4835 unsigned long available;
4836 unsigned long reclaimable;
4837 unsigned long min_wmark = min_wmark_pages(zone);
4840 available = reclaimable = zone_reclaimable_pages(zone);
4841 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4844 * Would the allocation succeed if we reclaimed all
4845 * reclaimable pages?
4847 wmark = __zone_watermark_ok(zone, order, min_wmark,
4848 ac->highest_zoneidx, alloc_flags, available);
4849 trace_reclaim_retry_zone(z, order, reclaimable,
4850 available, min_wmark, *no_progress_loops, wmark);
4853 * If we didn't make any progress and have a lot of
4854 * dirty + writeback pages then we should wait for
4855 * an IO to complete to slow down the reclaim and
4856 * prevent from pre mature OOM
4858 if (!did_some_progress) {
4859 unsigned long write_pending;
4861 write_pending = zone_page_state_snapshot(zone,
4862 NR_ZONE_WRITE_PENDING);
4864 if (2 * write_pending > reclaimable) {
4865 congestion_wait(BLK_RW_ASYNC, HZ/10);
4877 * Memory allocation/reclaim might be called from a WQ context and the
4878 * current implementation of the WQ concurrency control doesn't
4879 * recognize that a particular WQ is congested if the worker thread is
4880 * looping without ever sleeping. Therefore we have to do a short sleep
4881 * here rather than calling cond_resched().
4883 if (current->flags & PF_WQ_WORKER)
4884 schedule_timeout_uninterruptible(1);
4891 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4894 * It's possible that cpuset's mems_allowed and the nodemask from
4895 * mempolicy don't intersect. This should be normally dealt with by
4896 * policy_nodemask(), but it's possible to race with cpuset update in
4897 * such a way the check therein was true, and then it became false
4898 * before we got our cpuset_mems_cookie here.
4899 * This assumes that for all allocations, ac->nodemask can come only
4900 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4901 * when it does not intersect with the cpuset restrictions) or the
4902 * caller can deal with a violated nodemask.
4904 if (cpusets_enabled() && ac->nodemask &&
4905 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4906 ac->nodemask = NULL;
4911 * When updating a task's mems_allowed or mempolicy nodemask, it is
4912 * possible to race with parallel threads in such a way that our
4913 * allocation can fail while the mask is being updated. If we are about
4914 * to fail, check if the cpuset changed during allocation and if so,
4917 if (read_mems_allowed_retry(cpuset_mems_cookie))
4923 static inline struct page *
4924 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4925 struct alloc_context *ac)
4927 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4928 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4929 struct page *page = NULL;
4930 unsigned int alloc_flags;
4931 unsigned long did_some_progress;
4932 enum compact_priority compact_priority;
4933 enum compact_result compact_result;
4934 int compaction_retries;
4935 int no_progress_loops;
4936 unsigned int cpuset_mems_cookie;
4937 unsigned int zonelist_iter_cookie;
4941 * We also sanity check to catch abuse of atomic reserves being used by
4942 * callers that are not in atomic context.
4944 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4945 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4946 gfp_mask &= ~__GFP_ATOMIC;
4949 compaction_retries = 0;
4950 no_progress_loops = 0;
4951 compact_priority = DEF_COMPACT_PRIORITY;
4952 cpuset_mems_cookie = read_mems_allowed_begin();
4953 zonelist_iter_cookie = zonelist_iter_begin();
4956 * The fast path uses conservative alloc_flags to succeed only until
4957 * kswapd needs to be woken up, and to avoid the cost of setting up
4958 * alloc_flags precisely. So we do that now.
4960 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4963 * We need to recalculate the starting point for the zonelist iterator
4964 * because we might have used different nodemask in the fast path, or
4965 * there was a cpuset modification and we are retrying - otherwise we
4966 * could end up iterating over non-eligible zones endlessly.
4968 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4969 ac->highest_zoneidx, ac->nodemask);
4970 if (!ac->preferred_zoneref->zone)
4973 if (alloc_flags & ALLOC_KSWAPD)
4974 wake_all_kswapds(order, gfp_mask, ac);
4977 * The adjusted alloc_flags might result in immediate success, so try
4980 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4985 * For costly allocations, try direct compaction first, as it's likely
4986 * that we have enough base pages and don't need to reclaim. For non-
4987 * movable high-order allocations, do that as well, as compaction will
4988 * try prevent permanent fragmentation by migrating from blocks of the
4990 * Don't try this for allocations that are allowed to ignore
4991 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4993 if (can_direct_reclaim &&
4995 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4996 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4997 page = __alloc_pages_direct_compact(gfp_mask, order,
4999 INIT_COMPACT_PRIORITY,
5005 * Checks for costly allocations with __GFP_NORETRY, which
5006 * includes some THP page fault allocations
5008 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
5010 * If allocating entire pageblock(s) and compaction
5011 * failed because all zones are below low watermarks
5012 * or is prohibited because it recently failed at this
5013 * order, fail immediately unless the allocator has
5014 * requested compaction and reclaim retry.
5017 * - potentially very expensive because zones are far
5018 * below their low watermarks or this is part of very
5019 * bursty high order allocations,
5020 * - not guaranteed to help because isolate_freepages()
5021 * may not iterate over freed pages as part of its
5023 * - unlikely to make entire pageblocks free on its
5026 if (compact_result == COMPACT_SKIPPED ||
5027 compact_result == COMPACT_DEFERRED)
5031 * Looks like reclaim/compaction is worth trying, but
5032 * sync compaction could be very expensive, so keep
5033 * using async compaction.
5035 compact_priority = INIT_COMPACT_PRIORITY;
5040 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
5041 if (alloc_flags & ALLOC_KSWAPD)
5042 wake_all_kswapds(order, gfp_mask, ac);
5044 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
5046 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags);
5049 * Reset the nodemask and zonelist iterators if memory policies can be
5050 * ignored. These allocations are high priority and system rather than
5053 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
5054 ac->nodemask = NULL;
5055 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5056 ac->highest_zoneidx, ac->nodemask);
5059 /* Attempt with potentially adjusted zonelist and alloc_flags */
5060 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
5064 /* Caller is not willing to reclaim, we can't balance anything */
5065 if (!can_direct_reclaim)
5068 /* Avoid recursion of direct reclaim */
5069 if (current->flags & PF_MEMALLOC)
5072 /* Try direct reclaim and then allocating */
5073 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
5074 &did_some_progress);
5078 /* Try direct compaction and then allocating */
5079 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
5080 compact_priority, &compact_result);
5084 /* Do not loop if specifically requested */
5085 if (gfp_mask & __GFP_NORETRY)
5089 * Do not retry costly high order allocations unless they are
5090 * __GFP_RETRY_MAYFAIL
5092 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
5095 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
5096 did_some_progress > 0, &no_progress_loops))
5100 * It doesn't make any sense to retry for the compaction if the order-0
5101 * reclaim is not able to make any progress because the current
5102 * implementation of the compaction depends on the sufficient amount
5103 * of free memory (see __compaction_suitable)
5105 if (did_some_progress > 0 &&
5106 should_compact_retry(ac, order, alloc_flags,
5107 compact_result, &compact_priority,
5108 &compaction_retries))
5113 * Deal with possible cpuset update races or zonelist updates to avoid
5114 * a unnecessary OOM kill.
5116 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5117 check_retry_zonelist(zonelist_iter_cookie))
5120 /* Reclaim has failed us, start killing things */
5121 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
5125 /* Avoid allocations with no watermarks from looping endlessly */
5126 if (tsk_is_oom_victim(current) &&
5127 (alloc_flags & ALLOC_OOM ||
5128 (gfp_mask & __GFP_NOMEMALLOC)))
5131 /* Retry as long as the OOM killer is making progress */
5132 if (did_some_progress) {
5133 no_progress_loops = 0;
5139 * Deal with possible cpuset update races or zonelist updates to avoid
5140 * a unnecessary OOM kill.
5142 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
5143 check_retry_zonelist(zonelist_iter_cookie))
5147 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
5150 if (gfp_mask & __GFP_NOFAIL) {
5152 * All existing users of the __GFP_NOFAIL are blockable, so warn
5153 * of any new users that actually require GFP_NOWAIT
5155 if (WARN_ON_ONCE(!can_direct_reclaim))
5159 * PF_MEMALLOC request from this context is rather bizarre
5160 * because we cannot reclaim anything and only can loop waiting
5161 * for somebody to do a work for us
5163 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
5166 * non failing costly orders are a hard requirement which we
5167 * are not prepared for much so let's warn about these users
5168 * so that we can identify them and convert them to something
5171 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
5174 * Help non-failing allocations by giving them access to memory
5175 * reserves but do not use ALLOC_NO_WATERMARKS because this
5176 * could deplete whole memory reserves which would just make
5177 * the situation worse
5179 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
5187 warn_alloc(gfp_mask, ac->nodemask,
5188 "page allocation failure: order:%u", order);
5193 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
5194 int preferred_nid, nodemask_t *nodemask,
5195 struct alloc_context *ac, gfp_t *alloc_gfp,
5196 unsigned int *alloc_flags)
5198 ac->highest_zoneidx = gfp_zone(gfp_mask);
5199 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
5200 ac->nodemask = nodemask;
5201 ac->migratetype = gfp_migratetype(gfp_mask);
5203 if (cpusets_enabled()) {
5204 *alloc_gfp |= __GFP_HARDWALL;
5206 * When we are in the interrupt context, it is irrelevant
5207 * to the current task context. It means that any node ok.
5209 if (in_task() && !ac->nodemask)
5210 ac->nodemask = &cpuset_current_mems_allowed;
5212 *alloc_flags |= ALLOC_CPUSET;
5215 fs_reclaim_acquire(gfp_mask);
5216 fs_reclaim_release(gfp_mask);
5218 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
5220 if (should_fail_alloc_page(gfp_mask, order))
5223 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
5225 /* Dirty zone balancing only done in the fast path */
5226 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
5229 * The preferred zone is used for statistics but crucially it is
5230 * also used as the starting point for the zonelist iterator. It
5231 * may get reset for allocations that ignore memory policies.
5233 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
5234 ac->highest_zoneidx, ac->nodemask);
5240 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
5241 * @gfp: GFP flags for the allocation
5242 * @preferred_nid: The preferred NUMA node ID to allocate from
5243 * @nodemask: Set of nodes to allocate from, may be NULL
5244 * @nr_pages: The number of pages desired on the list or array
5245 * @page_list: Optional list to store the allocated pages
5246 * @page_array: Optional array to store the pages
5248 * This is a batched version of the page allocator that attempts to
5249 * allocate nr_pages quickly. Pages are added to page_list if page_list
5250 * is not NULL, otherwise it is assumed that the page_array is valid.
5252 * For lists, nr_pages is the number of pages that should be allocated.
5254 * For arrays, only NULL elements are populated with pages and nr_pages
5255 * is the maximum number of pages that will be stored in the array.
5257 * Returns the number of pages on the list or array.
5259 unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
5260 nodemask_t *nodemask, int nr_pages,
5261 struct list_head *page_list,
5262 struct page **page_array)
5265 unsigned long flags;
5268 struct per_cpu_pages *pcp;
5269 struct list_head *pcp_list;
5270 struct alloc_context ac;
5272 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5273 int nr_populated = 0, nr_account = 0;
5276 * Skip populated array elements to determine if any pages need
5277 * to be allocated before disabling IRQs.
5279 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
5282 /* No pages requested? */
5283 if (unlikely(nr_pages <= 0))
5286 /* Already populated array? */
5287 if (unlikely(page_array && nr_pages - nr_populated == 0))
5290 /* Bulk allocator does not support memcg accounting. */
5291 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT))
5294 /* Use the single page allocator for one page. */
5295 if (nr_pages - nr_populated == 1)
5298 #ifdef CONFIG_PAGE_OWNER
5300 * PAGE_OWNER may recurse into the allocator to allocate space to
5301 * save the stack with pagesets.lock held. Releasing/reacquiring
5302 * removes much of the performance benefit of bulk allocation so
5303 * force the caller to allocate one page at a time as it'll have
5304 * similar performance to added complexity to the bulk allocator.
5306 if (static_branch_unlikely(&page_owner_inited))
5310 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
5311 gfp &= gfp_allowed_mask;
5313 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
5317 /* Find an allowed local zone that meets the low watermark. */
5318 for_each_zone_zonelist_nodemask(zone, z, ac.zonelist, ac.highest_zoneidx, ac.nodemask) {
5321 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
5322 !__cpuset_zone_allowed(zone, gfp)) {
5326 if (nr_online_nodes > 1 && zone != ac.preferred_zoneref->zone &&
5327 zone_to_nid(zone) != zone_to_nid(ac.preferred_zoneref->zone)) {
5331 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
5332 if (zone_watermark_fast(zone, 0, mark,
5333 zonelist_zone_idx(ac.preferred_zoneref),
5334 alloc_flags, gfp)) {
5340 * If there are no allowed local zones that meets the watermarks then
5341 * try to allocate a single page and reclaim if necessary.
5343 if (unlikely(!zone))
5346 /* Attempt the batch allocation */
5347 local_lock_irqsave(&pagesets.lock, flags);
5348 pcp = this_cpu_ptr(zone->per_cpu_pageset);
5349 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
5351 while (nr_populated < nr_pages) {
5353 /* Skip existing pages */
5354 if (page_array && page_array[nr_populated]) {
5359 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
5361 if (unlikely(!page)) {
5362 /* Try and allocate at least one page */
5369 prep_new_page(page, 0, gfp, 0);
5371 list_add(&page->lru, page_list);
5373 page_array[nr_populated] = page;
5377 local_unlock_irqrestore(&pagesets.lock, flags);
5379 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
5380 zone_statistics(ac.preferred_zoneref->zone, zone, nr_account);
5383 return nr_populated;
5386 local_unlock_irqrestore(&pagesets.lock, flags);
5389 page = __alloc_pages(gfp, 0, preferred_nid, nodemask);
5392 list_add(&page->lru, page_list);
5394 page_array[nr_populated] = page;
5400 EXPORT_SYMBOL_GPL(__alloc_pages_bulk);
5403 * This is the 'heart' of the zoned buddy allocator.
5405 struct page *__alloc_pages(gfp_t gfp, unsigned int order, int preferred_nid,
5406 nodemask_t *nodemask)
5409 unsigned int alloc_flags = ALLOC_WMARK_LOW;
5410 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
5411 struct alloc_context ac = { };
5414 * There are several places where we assume that the order value is sane
5415 * so bail out early if the request is out of bound.
5417 if (unlikely(order >= MAX_ORDER)) {
5418 WARN_ON_ONCE(!(gfp & __GFP_NOWARN));
5422 gfp &= gfp_allowed_mask;
5424 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5425 * resp. GFP_NOIO which has to be inherited for all allocation requests
5426 * from a particular context which has been marked by
5427 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
5428 * movable zones are not used during allocation.
5430 gfp = current_gfp_context(gfp);
5432 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
5433 &alloc_gfp, &alloc_flags))
5437 * Forbid the first pass from falling back to types that fragment
5438 * memory until all local zones are considered.
5440 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp);
5442 /* First allocation attempt */
5443 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
5448 ac.spread_dirty_pages = false;
5451 * Restore the original nodemask if it was potentially replaced with
5452 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5454 ac.nodemask = nodemask;
5456 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
5459 if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
5460 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
5461 __free_pages(page, order);
5465 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
5469 EXPORT_SYMBOL(__alloc_pages);
5472 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5473 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5474 * you need to access high mem.
5476 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5480 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5483 return (unsigned long) page_address(page);
5485 EXPORT_SYMBOL(__get_free_pages);
5487 unsigned long get_zeroed_page(gfp_t gfp_mask)
5489 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5491 EXPORT_SYMBOL(get_zeroed_page);
5494 * __free_pages - Free pages allocated with alloc_pages().
5495 * @page: The page pointer returned from alloc_pages().
5496 * @order: The order of the allocation.
5498 * This function can free multi-page allocations that are not compound
5499 * pages. It does not check that the @order passed in matches that of
5500 * the allocation, so it is easy to leak memory. Freeing more memory
5501 * than was allocated will probably emit a warning.
5503 * If the last reference to this page is speculative, it will be released
5504 * by put_page() which only frees the first page of a non-compound
5505 * allocation. To prevent the remaining pages from being leaked, we free
5506 * the subsequent pages here. If you want to use the page's reference
5507 * count to decide when to free the allocation, you should allocate a
5508 * compound page, and use put_page() instead of __free_pages().
5510 * Context: May be called in interrupt context or while holding a normal
5511 * spinlock, but not in NMI context or while holding a raw spinlock.
5513 void __free_pages(struct page *page, unsigned int order)
5515 if (put_page_testzero(page))
5516 free_the_page(page, order);
5517 else if (!PageHead(page))
5519 free_the_page(page + (1 << order), order);
5521 EXPORT_SYMBOL(__free_pages);
5523 void free_pages(unsigned long addr, unsigned int order)
5526 VM_BUG_ON(!virt_addr_valid((void *)addr));
5527 __free_pages(virt_to_page((void *)addr), order);
5531 EXPORT_SYMBOL(free_pages);
5535 * An arbitrary-length arbitrary-offset area of memory which resides
5536 * within a 0 or higher order page. Multiple fragments within that page
5537 * are individually refcounted, in the page's reference counter.
5539 * The page_frag functions below provide a simple allocation framework for
5540 * page fragments. This is used by the network stack and network device
5541 * drivers to provide a backing region of memory for use as either an
5542 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5544 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5547 struct page *page = NULL;
5548 gfp_t gfp = gfp_mask;
5550 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5551 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5553 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5554 PAGE_FRAG_CACHE_MAX_ORDER);
5555 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5557 if (unlikely(!page))
5558 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5560 nc->va = page ? page_address(page) : NULL;
5565 void __page_frag_cache_drain(struct page *page, unsigned int count)
5567 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5569 if (page_ref_sub_and_test(page, count))
5570 free_the_page(page, compound_order(page));
5572 EXPORT_SYMBOL(__page_frag_cache_drain);
5574 void *page_frag_alloc_align(struct page_frag_cache *nc,
5575 unsigned int fragsz, gfp_t gfp_mask,
5576 unsigned int align_mask)
5578 unsigned int size = PAGE_SIZE;
5582 if (unlikely(!nc->va)) {
5584 page = __page_frag_cache_refill(nc, gfp_mask);
5588 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5589 /* if size can vary use size else just use PAGE_SIZE */
5592 /* Even if we own the page, we do not use atomic_set().
5593 * This would break get_page_unless_zero() users.
5595 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5597 /* reset page count bias and offset to start of new frag */
5598 nc->pfmemalloc = page_is_pfmemalloc(page);
5599 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5603 offset = nc->offset - fragsz;
5604 if (unlikely(offset < 0)) {
5605 page = virt_to_page(nc->va);
5607 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5610 if (unlikely(nc->pfmemalloc)) {
5611 free_the_page(page, compound_order(page));
5615 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5616 /* if size can vary use size else just use PAGE_SIZE */
5619 /* OK, page count is 0, we can safely set it */
5620 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5622 /* reset page count bias and offset to start of new frag */
5623 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5624 offset = size - fragsz;
5625 if (unlikely(offset < 0)) {
5627 * The caller is trying to allocate a fragment
5628 * with fragsz > PAGE_SIZE but the cache isn't big
5629 * enough to satisfy the request, this may
5630 * happen in low memory conditions.
5631 * We don't release the cache page because
5632 * it could make memory pressure worse
5633 * so we simply return NULL here.
5640 offset &= align_mask;
5641 nc->offset = offset;
5643 return nc->va + offset;
5645 EXPORT_SYMBOL(page_frag_alloc_align);
5648 * Frees a page fragment allocated out of either a compound or order 0 page.
5650 void page_frag_free(void *addr)
5652 struct page *page = virt_to_head_page(addr);
5654 if (unlikely(put_page_testzero(page)))
5655 free_the_page(page, compound_order(page));
5657 EXPORT_SYMBOL(page_frag_free);
5659 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5663 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5664 unsigned long used = addr + PAGE_ALIGN(size);
5666 split_page(virt_to_page((void *)addr), order);
5667 while (used < alloc_end) {
5672 return (void *)addr;
5676 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5677 * @size: the number of bytes to allocate
5678 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5680 * This function is similar to alloc_pages(), except that it allocates the
5681 * minimum number of pages to satisfy the request. alloc_pages() can only
5682 * allocate memory in power-of-two pages.
5684 * This function is also limited by MAX_ORDER.
5686 * Memory allocated by this function must be released by free_pages_exact().
5688 * Return: pointer to the allocated area or %NULL in case of error.
5690 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5692 unsigned int order = get_order(size);
5695 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5696 gfp_mask &= ~__GFP_COMP;
5698 addr = __get_free_pages(gfp_mask, order);
5699 return make_alloc_exact(addr, order, size);
5701 EXPORT_SYMBOL(alloc_pages_exact);
5704 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5706 * @nid: the preferred node ID where memory should be allocated
5707 * @size: the number of bytes to allocate
5708 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5710 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5713 * Return: pointer to the allocated area or %NULL in case of error.
5715 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5717 unsigned int order = get_order(size);
5720 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5721 gfp_mask &= ~__GFP_COMP;
5723 p = alloc_pages_node(nid, gfp_mask, order);
5726 return make_alloc_exact((unsigned long)page_address(p), order, size);
5730 * free_pages_exact - release memory allocated via alloc_pages_exact()
5731 * @virt: the value returned by alloc_pages_exact.
5732 * @size: size of allocation, same value as passed to alloc_pages_exact().
5734 * Release the memory allocated by a previous call to alloc_pages_exact.
5736 void free_pages_exact(void *virt, size_t size)
5738 unsigned long addr = (unsigned long)virt;
5739 unsigned long end = addr + PAGE_ALIGN(size);
5741 while (addr < end) {
5746 EXPORT_SYMBOL(free_pages_exact);
5749 * nr_free_zone_pages - count number of pages beyond high watermark
5750 * @offset: The zone index of the highest zone
5752 * nr_free_zone_pages() counts the number of pages which are beyond the
5753 * high watermark within all zones at or below a given zone index. For each
5754 * zone, the number of pages is calculated as:
5756 * nr_free_zone_pages = managed_pages - high_pages
5758 * Return: number of pages beyond high watermark.
5760 static unsigned long nr_free_zone_pages(int offset)
5765 /* Just pick one node, since fallback list is circular */
5766 unsigned long sum = 0;
5768 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5770 for_each_zone_zonelist(zone, z, zonelist, offset) {
5771 unsigned long size = zone_managed_pages(zone);
5772 unsigned long high = high_wmark_pages(zone);
5781 * nr_free_buffer_pages - count number of pages beyond high watermark
5783 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5784 * watermark within ZONE_DMA and ZONE_NORMAL.
5786 * Return: number of pages beyond high watermark within ZONE_DMA and
5789 unsigned long nr_free_buffer_pages(void)
5791 return nr_free_zone_pages(gfp_zone(GFP_USER));
5793 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5795 static inline void show_node(struct zone *zone)
5797 if (IS_ENABLED(CONFIG_NUMA))
5798 printk("Node %d ", zone_to_nid(zone));
5801 long si_mem_available(void)
5804 unsigned long pagecache;
5805 unsigned long wmark_low = 0;
5806 unsigned long pages[NR_LRU_LISTS];
5807 unsigned long reclaimable;
5811 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5812 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5815 wmark_low += low_wmark_pages(zone);
5818 * Estimate the amount of memory available for userspace allocations,
5819 * without causing swapping.
5821 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5824 * Not all the page cache can be freed, otherwise the system will
5825 * start swapping. Assume at least half of the page cache, or the
5826 * low watermark worth of cache, needs to stay.
5828 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5829 pagecache -= min(pagecache / 2, wmark_low);
5830 available += pagecache;
5833 * Part of the reclaimable slab and other kernel memory consists of
5834 * items that are in use, and cannot be freed. Cap this estimate at the
5837 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5838 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5839 available += reclaimable - min(reclaimable / 2, wmark_low);
5845 EXPORT_SYMBOL_GPL(si_mem_available);
5847 void si_meminfo(struct sysinfo *val)
5849 val->totalram = totalram_pages();
5850 val->sharedram = global_node_page_state(NR_SHMEM);
5851 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5852 val->bufferram = nr_blockdev_pages();
5853 val->totalhigh = totalhigh_pages();
5854 val->freehigh = nr_free_highpages();
5855 val->mem_unit = PAGE_SIZE;
5858 EXPORT_SYMBOL(si_meminfo);
5861 void si_meminfo_node(struct sysinfo *val, int nid)
5863 int zone_type; /* needs to be signed */
5864 unsigned long managed_pages = 0;
5865 unsigned long managed_highpages = 0;
5866 unsigned long free_highpages = 0;
5867 pg_data_t *pgdat = NODE_DATA(nid);
5869 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5870 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5871 val->totalram = managed_pages;
5872 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5873 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5874 #ifdef CONFIG_HIGHMEM
5875 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5876 struct zone *zone = &pgdat->node_zones[zone_type];
5878 if (is_highmem(zone)) {
5879 managed_highpages += zone_managed_pages(zone);
5880 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5883 val->totalhigh = managed_highpages;
5884 val->freehigh = free_highpages;
5886 val->totalhigh = managed_highpages;
5887 val->freehigh = free_highpages;
5889 val->mem_unit = PAGE_SIZE;
5894 * Determine whether the node should be displayed or not, depending on whether
5895 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5897 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5899 if (!(flags & SHOW_MEM_FILTER_NODES))
5903 * no node mask - aka implicit memory numa policy. Do not bother with
5904 * the synchronization - read_mems_allowed_begin - because we do not
5905 * have to be precise here.
5908 nodemask = &cpuset_current_mems_allowed;
5910 return !node_isset(nid, *nodemask);
5913 #define K(x) ((x) << (PAGE_SHIFT-10))
5915 static void show_migration_types(unsigned char type)
5917 static const char types[MIGRATE_TYPES] = {
5918 [MIGRATE_UNMOVABLE] = 'U',
5919 [MIGRATE_MOVABLE] = 'M',
5920 [MIGRATE_RECLAIMABLE] = 'E',
5921 [MIGRATE_HIGHATOMIC] = 'H',
5923 [MIGRATE_CMA] = 'C',
5925 #ifdef CONFIG_MEMORY_ISOLATION
5926 [MIGRATE_ISOLATE] = 'I',
5929 char tmp[MIGRATE_TYPES + 1];
5933 for (i = 0; i < MIGRATE_TYPES; i++) {
5934 if (type & (1 << i))
5939 printk(KERN_CONT "(%s) ", tmp);
5943 * Show free area list (used inside shift_scroll-lock stuff)
5944 * We also calculate the percentage fragmentation. We do this by counting the
5945 * memory on each free list with the exception of the first item on the list.
5948 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5951 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5953 unsigned long free_pcp = 0;
5958 for_each_populated_zone(zone) {
5959 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5962 for_each_online_cpu(cpu)
5963 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
5966 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5967 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5968 " unevictable:%lu dirty:%lu writeback:%lu\n"
5969 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5970 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5971 " kernel_misc_reclaimable:%lu\n"
5972 " free:%lu free_pcp:%lu free_cma:%lu\n",
5973 global_node_page_state(NR_ACTIVE_ANON),
5974 global_node_page_state(NR_INACTIVE_ANON),
5975 global_node_page_state(NR_ISOLATED_ANON),
5976 global_node_page_state(NR_ACTIVE_FILE),
5977 global_node_page_state(NR_INACTIVE_FILE),
5978 global_node_page_state(NR_ISOLATED_FILE),
5979 global_node_page_state(NR_UNEVICTABLE),
5980 global_node_page_state(NR_FILE_DIRTY),
5981 global_node_page_state(NR_WRITEBACK),
5982 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5983 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5984 global_node_page_state(NR_FILE_MAPPED),
5985 global_node_page_state(NR_SHMEM),
5986 global_node_page_state(NR_PAGETABLE),
5987 global_zone_page_state(NR_BOUNCE),
5988 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE),
5989 global_zone_page_state(NR_FREE_PAGES),
5991 global_zone_page_state(NR_FREE_CMA_PAGES));
5993 for_each_online_pgdat(pgdat) {
5994 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5998 " active_anon:%lukB"
5999 " inactive_anon:%lukB"
6000 " active_file:%lukB"
6001 " inactive_file:%lukB"
6002 " unevictable:%lukB"
6003 " isolated(anon):%lukB"
6004 " isolated(file):%lukB"
6009 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6011 " shmem_pmdmapped: %lukB"
6014 " writeback_tmp:%lukB"
6015 " kernel_stack:%lukB"
6016 #ifdef CONFIG_SHADOW_CALL_STACK
6017 " shadow_call_stack:%lukB"
6020 " all_unreclaimable? %s"
6023 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
6024 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
6025 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
6026 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
6027 K(node_page_state(pgdat, NR_UNEVICTABLE)),
6028 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
6029 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
6030 K(node_page_state(pgdat, NR_FILE_MAPPED)),
6031 K(node_page_state(pgdat, NR_FILE_DIRTY)),
6032 K(node_page_state(pgdat, NR_WRITEBACK)),
6033 K(node_page_state(pgdat, NR_SHMEM)),
6034 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6035 K(node_page_state(pgdat, NR_SHMEM_THPS)),
6036 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)),
6037 K(node_page_state(pgdat, NR_ANON_THPS)),
6039 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
6040 node_page_state(pgdat, NR_KERNEL_STACK_KB),
6041 #ifdef CONFIG_SHADOW_CALL_STACK
6042 node_page_state(pgdat, NR_KERNEL_SCS_KB),
6044 K(node_page_state(pgdat, NR_PAGETABLE)),
6045 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
6049 for_each_populated_zone(zone) {
6052 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6056 for_each_online_cpu(cpu)
6057 free_pcp += per_cpu_ptr(zone->per_cpu_pageset, cpu)->count;
6066 " reserved_highatomic:%luKB"
6067 " active_anon:%lukB"
6068 " inactive_anon:%lukB"
6069 " active_file:%lukB"
6070 " inactive_file:%lukB"
6071 " unevictable:%lukB"
6072 " writepending:%lukB"
6082 K(zone_page_state(zone, NR_FREE_PAGES)),
6083 K(min_wmark_pages(zone)),
6084 K(low_wmark_pages(zone)),
6085 K(high_wmark_pages(zone)),
6086 K(zone->nr_reserved_highatomic),
6087 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
6088 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
6089 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
6090 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
6091 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
6092 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
6093 K(zone->present_pages),
6094 K(zone_managed_pages(zone)),
6095 K(zone_page_state(zone, NR_MLOCK)),
6096 K(zone_page_state(zone, NR_BOUNCE)),
6098 K(this_cpu_read(zone->per_cpu_pageset->count)),
6099 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
6100 printk("lowmem_reserve[]:");
6101 for (i = 0; i < MAX_NR_ZONES; i++)
6102 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
6103 printk(KERN_CONT "\n");
6106 for_each_populated_zone(zone) {
6108 unsigned long nr[MAX_ORDER], flags, total = 0;
6109 unsigned char types[MAX_ORDER];
6111 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
6114 printk(KERN_CONT "%s: ", zone->name);
6116 spin_lock_irqsave(&zone->lock, flags);
6117 for (order = 0; order < MAX_ORDER; order++) {
6118 struct free_area *area = &zone->free_area[order];
6121 nr[order] = area->nr_free;
6122 total += nr[order] << order;
6125 for (type = 0; type < MIGRATE_TYPES; type++) {
6126 if (!free_area_empty(area, type))
6127 types[order] |= 1 << type;
6130 spin_unlock_irqrestore(&zone->lock, flags);
6131 for (order = 0; order < MAX_ORDER; order++) {
6132 printk(KERN_CONT "%lu*%lukB ",
6133 nr[order], K(1UL) << order);
6135 show_migration_types(types[order]);
6137 printk(KERN_CONT "= %lukB\n", K(total));
6140 hugetlb_show_meminfo();
6142 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
6144 show_swap_cache_info();
6147 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
6149 zoneref->zone = zone;
6150 zoneref->zone_idx = zone_idx(zone);
6154 * Builds allocation fallback zone lists.
6156 * Add all populated zones of a node to the zonelist.
6158 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
6161 enum zone_type zone_type = MAX_NR_ZONES;
6166 zone = pgdat->node_zones + zone_type;
6167 if (populated_zone(zone)) {
6168 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
6169 check_highest_zone(zone_type);
6171 } while (zone_type);
6178 static int __parse_numa_zonelist_order(char *s)
6181 * We used to support different zonelists modes but they turned
6182 * out to be just not useful. Let's keep the warning in place
6183 * if somebody still use the cmd line parameter so that we do
6184 * not fail it silently
6186 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
6187 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
6193 char numa_zonelist_order[] = "Node";
6196 * sysctl handler for numa_zonelist_order
6198 int numa_zonelist_order_handler(struct ctl_table *table, int write,
6199 void *buffer, size_t *length, loff_t *ppos)
6202 return __parse_numa_zonelist_order(buffer);
6203 return proc_dostring(table, write, buffer, length, ppos);
6207 #define MAX_NODE_LOAD (nr_online_nodes)
6208 static int node_load[MAX_NUMNODES];
6211 * find_next_best_node - find the next node that should appear in a given node's fallback list
6212 * @node: node whose fallback list we're appending
6213 * @used_node_mask: nodemask_t of already used nodes
6215 * We use a number of factors to determine which is the next node that should
6216 * appear on a given node's fallback list. The node should not have appeared
6217 * already in @node's fallback list, and it should be the next closest node
6218 * according to the distance array (which contains arbitrary distance values
6219 * from each node to each node in the system), and should also prefer nodes
6220 * with no CPUs, since presumably they'll have very little allocation pressure
6221 * on them otherwise.
6223 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
6225 int find_next_best_node(int node, nodemask_t *used_node_mask)
6228 int min_val = INT_MAX;
6229 int best_node = NUMA_NO_NODE;
6231 /* Use the local node if we haven't already */
6232 if (!node_isset(node, *used_node_mask)) {
6233 node_set(node, *used_node_mask);
6237 for_each_node_state(n, N_MEMORY) {
6239 /* Don't want a node to appear more than once */
6240 if (node_isset(n, *used_node_mask))
6243 /* Use the distance array to find the distance */
6244 val = node_distance(node, n);
6246 /* Penalize nodes under us ("prefer the next node") */
6249 /* Give preference to headless and unused nodes */
6250 if (!cpumask_empty(cpumask_of_node(n)))
6251 val += PENALTY_FOR_NODE_WITH_CPUS;
6253 /* Slight preference for less loaded node */
6254 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
6255 val += node_load[n];
6257 if (val < min_val) {
6264 node_set(best_node, *used_node_mask);
6271 * Build zonelists ordered by node and zones within node.
6272 * This results in maximum locality--normal zone overflows into local
6273 * DMA zone, if any--but risks exhausting DMA zone.
6275 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
6278 struct zoneref *zonerefs;
6281 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6283 for (i = 0; i < nr_nodes; i++) {
6286 pg_data_t *node = NODE_DATA(node_order[i]);
6288 nr_zones = build_zonerefs_node(node, zonerefs);
6289 zonerefs += nr_zones;
6291 zonerefs->zone = NULL;
6292 zonerefs->zone_idx = 0;
6296 * Build gfp_thisnode zonelists
6298 static void build_thisnode_zonelists(pg_data_t *pgdat)
6300 struct zoneref *zonerefs;
6303 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
6304 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6305 zonerefs += nr_zones;
6306 zonerefs->zone = NULL;
6307 zonerefs->zone_idx = 0;
6311 * Build zonelists ordered by zone and nodes within zones.
6312 * This results in conserving DMA zone[s] until all Normal memory is
6313 * exhausted, but results in overflowing to remote node while memory
6314 * may still exist in local DMA zone.
6317 static void build_zonelists(pg_data_t *pgdat)
6319 static int node_order[MAX_NUMNODES];
6320 int node, load, nr_nodes = 0;
6321 nodemask_t used_mask = NODE_MASK_NONE;
6322 int local_node, prev_node;
6324 /* NUMA-aware ordering of nodes */
6325 local_node = pgdat->node_id;
6326 load = nr_online_nodes;
6327 prev_node = local_node;
6329 memset(node_order, 0, sizeof(node_order));
6330 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
6332 * We don't want to pressure a particular node.
6333 * So adding penalty to the first node in same
6334 * distance group to make it round-robin.
6336 if (node_distance(local_node, node) !=
6337 node_distance(local_node, prev_node))
6338 node_load[node] = load;
6340 node_order[nr_nodes++] = node;
6345 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
6346 build_thisnode_zonelists(pgdat);
6349 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6351 * Return node id of node used for "local" allocations.
6352 * I.e., first node id of first zone in arg node's generic zonelist.
6353 * Used for initializing percpu 'numa_mem', which is used primarily
6354 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
6356 int local_memory_node(int node)
6360 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
6361 gfp_zone(GFP_KERNEL),
6363 return zone_to_nid(z->zone);
6367 static void setup_min_unmapped_ratio(void);
6368 static void setup_min_slab_ratio(void);
6369 #else /* CONFIG_NUMA */
6371 static void build_zonelists(pg_data_t *pgdat)
6373 int node, local_node;
6374 struct zoneref *zonerefs;
6377 local_node = pgdat->node_id;
6379 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
6380 nr_zones = build_zonerefs_node(pgdat, zonerefs);
6381 zonerefs += nr_zones;
6384 * Now we build the zonelist so that it contains the zones
6385 * of all the other nodes.
6386 * We don't want to pressure a particular node, so when
6387 * building the zones for node N, we make sure that the
6388 * zones coming right after the local ones are those from
6389 * node N+1 (modulo N)
6391 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
6392 if (!node_online(node))
6394 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6395 zonerefs += nr_zones;
6397 for (node = 0; node < local_node; node++) {
6398 if (!node_online(node))
6400 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
6401 zonerefs += nr_zones;
6404 zonerefs->zone = NULL;
6405 zonerefs->zone_idx = 0;
6408 #endif /* CONFIG_NUMA */
6411 * Boot pageset table. One per cpu which is going to be used for all
6412 * zones and all nodes. The parameters will be set in such a way
6413 * that an item put on a list will immediately be handed over to
6414 * the buddy list. This is safe since pageset manipulation is done
6415 * with interrupts disabled.
6417 * The boot_pagesets must be kept even after bootup is complete for
6418 * unused processors and/or zones. They do play a role for bootstrapping
6419 * hotplugged processors.
6421 * zoneinfo_show() and maybe other functions do
6422 * not check if the processor is online before following the pageset pointer.
6423 * Other parts of the kernel may not check if the zone is available.
6425 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
6426 /* These effectively disable the pcplists in the boot pageset completely */
6427 #define BOOT_PAGESET_HIGH 0
6428 #define BOOT_PAGESET_BATCH 1
6429 static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
6430 static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
6431 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
6433 static void __build_all_zonelists(void *data)
6436 int __maybe_unused cpu;
6437 pg_data_t *self = data;
6439 write_seqlock(&zonelist_update_seq);
6442 memset(node_load, 0, sizeof(node_load));
6446 * This node is hotadded and no memory is yet present. So just
6447 * building zonelists is fine - no need to touch other nodes.
6449 if (self && !node_online(self->node_id)) {
6450 build_zonelists(self);
6452 for_each_online_node(nid) {
6453 pg_data_t *pgdat = NODE_DATA(nid);
6455 build_zonelists(pgdat);
6458 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6460 * We now know the "local memory node" for each node--
6461 * i.e., the node of the first zone in the generic zonelist.
6462 * Set up numa_mem percpu variable for on-line cpus. During
6463 * boot, only the boot cpu should be on-line; we'll init the
6464 * secondary cpus' numa_mem as they come on-line. During
6465 * node/memory hotplug, we'll fixup all on-line cpus.
6467 for_each_online_cpu(cpu)
6468 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6472 write_sequnlock(&zonelist_update_seq);
6475 static noinline void __init
6476 build_all_zonelists_init(void)
6480 __build_all_zonelists(NULL);
6483 * Initialize the boot_pagesets that are going to be used
6484 * for bootstrapping processors. The real pagesets for
6485 * each zone will be allocated later when the per cpu
6486 * allocator is available.
6488 * boot_pagesets are used also for bootstrapping offline
6489 * cpus if the system is already booted because the pagesets
6490 * are needed to initialize allocators on a specific cpu too.
6491 * F.e. the percpu allocator needs the page allocator which
6492 * needs the percpu allocator in order to allocate its pagesets
6493 * (a chicken-egg dilemma).
6495 for_each_possible_cpu(cpu)
6496 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
6498 mminit_verify_zonelist();
6499 cpuset_init_current_mems_allowed();
6503 * unless system_state == SYSTEM_BOOTING.
6505 * __ref due to call of __init annotated helper build_all_zonelists_init
6506 * [protected by SYSTEM_BOOTING].
6508 void __ref build_all_zonelists(pg_data_t *pgdat)
6510 unsigned long vm_total_pages;
6512 if (system_state == SYSTEM_BOOTING) {
6513 build_all_zonelists_init();
6515 __build_all_zonelists(pgdat);
6516 /* cpuset refresh routine should be here */
6518 /* Get the number of free pages beyond high watermark in all zones. */
6519 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6521 * Disable grouping by mobility if the number of pages in the
6522 * system is too low to allow the mechanism to work. It would be
6523 * more accurate, but expensive to check per-zone. This check is
6524 * made on memory-hotadd so a system can start with mobility
6525 * disabled and enable it later
6527 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6528 page_group_by_mobility_disabled = 1;
6530 page_group_by_mobility_disabled = 0;
6532 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6534 page_group_by_mobility_disabled ? "off" : "on",
6537 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6541 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6542 static bool __meminit
6543 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6545 static struct memblock_region *r;
6547 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6548 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6549 for_each_mem_region(r) {
6550 if (*pfn < memblock_region_memory_end_pfn(r))
6554 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6555 memblock_is_mirror(r)) {
6556 *pfn = memblock_region_memory_end_pfn(r);
6564 * Initially all pages are reserved - free ones are freed
6565 * up by memblock_free_all() once the early boot process is
6566 * done. Non-atomic initialization, single-pass.
6568 * All aligned pageblocks are initialized to the specified migratetype
6569 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6570 * zone stats (e.g., nr_isolate_pageblock) are touched.
6572 void __meminit memmap_init_range(unsigned long size, int nid, unsigned long zone,
6573 unsigned long start_pfn, unsigned long zone_end_pfn,
6574 enum meminit_context context,
6575 struct vmem_altmap *altmap, int migratetype)
6577 unsigned long pfn, end_pfn = start_pfn + size;
6580 if (highest_memmap_pfn < end_pfn - 1)
6581 highest_memmap_pfn = end_pfn - 1;
6583 #ifdef CONFIG_ZONE_DEVICE
6585 * Honor reservation requested by the driver for this ZONE_DEVICE
6586 * memory. We limit the total number of pages to initialize to just
6587 * those that might contain the memory mapping. We will defer the
6588 * ZONE_DEVICE page initialization until after we have released
6591 if (zone == ZONE_DEVICE) {
6595 if (start_pfn == altmap->base_pfn)
6596 start_pfn += altmap->reserve;
6597 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6601 for (pfn = start_pfn; pfn < end_pfn; ) {
6603 * There can be holes in boot-time mem_map[]s handed to this
6604 * function. They do not exist on hotplugged memory.
6606 if (context == MEMINIT_EARLY) {
6607 if (overlap_memmap_init(zone, &pfn))
6609 if (defer_init(nid, pfn, zone_end_pfn))
6613 page = pfn_to_page(pfn);
6614 __init_single_page(page, pfn, zone, nid);
6615 if (context == MEMINIT_HOTPLUG)
6616 __SetPageReserved(page);
6619 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6620 * such that unmovable allocations won't be scattered all
6621 * over the place during system boot.
6623 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6624 set_pageblock_migratetype(page, migratetype);
6631 #ifdef CONFIG_ZONE_DEVICE
6632 void __ref memmap_init_zone_device(struct zone *zone,
6633 unsigned long start_pfn,
6634 unsigned long nr_pages,
6635 struct dev_pagemap *pgmap)
6637 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6638 struct pglist_data *pgdat = zone->zone_pgdat;
6639 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6640 unsigned long zone_idx = zone_idx(zone);
6641 unsigned long start = jiffies;
6642 int nid = pgdat->node_id;
6644 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6648 * The call to memmap_init should have already taken care
6649 * of the pages reserved for the memmap, so we can just jump to
6650 * the end of that region and start processing the device pages.
6653 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6654 nr_pages = end_pfn - start_pfn;
6657 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6658 struct page *page = pfn_to_page(pfn);
6660 __init_single_page(page, pfn, zone_idx, nid);
6663 * Mark page reserved as it will need to wait for onlining
6664 * phase for it to be fully associated with a zone.
6666 * We can use the non-atomic __set_bit operation for setting
6667 * the flag as we are still initializing the pages.
6669 __SetPageReserved(page);
6672 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6673 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6674 * ever freed or placed on a driver-private list.
6676 page->pgmap = pgmap;
6677 page->zone_device_data = NULL;
6680 * Mark the block movable so that blocks are reserved for
6681 * movable at startup. This will force kernel allocations
6682 * to reserve their blocks rather than leaking throughout
6683 * the address space during boot when many long-lived
6684 * kernel allocations are made.
6686 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6687 * because this is done early in section_activate()
6689 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6690 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6695 pr_info("%s initialised %lu pages in %ums\n", __func__,
6696 nr_pages, jiffies_to_msecs(jiffies - start));
6700 static void __meminit zone_init_free_lists(struct zone *zone)
6702 unsigned int order, t;
6703 for_each_migratetype_order(order, t) {
6704 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6705 zone->free_area[order].nr_free = 0;
6710 * Only struct pages that correspond to ranges defined by memblock.memory
6711 * are zeroed and initialized by going through __init_single_page() during
6712 * memmap_init_zone_range().
6714 * But, there could be struct pages that correspond to holes in
6715 * memblock.memory. This can happen because of the following reasons:
6716 * - physical memory bank size is not necessarily the exact multiple of the
6717 * arbitrary section size
6718 * - early reserved memory may not be listed in memblock.memory
6719 * - memory layouts defined with memmap= kernel parameter may not align
6720 * nicely with memmap sections
6722 * Explicitly initialize those struct pages so that:
6723 * - PG_Reserved is set
6724 * - zone and node links point to zone and node that span the page if the
6725 * hole is in the middle of a zone
6726 * - zone and node links point to adjacent zone/node if the hole falls on
6727 * the zone boundary; the pages in such holes will be prepended to the
6728 * zone/node above the hole except for the trailing pages in the last
6729 * section that will be appended to the zone/node below.
6731 static void __init init_unavailable_range(unsigned long spfn,
6738 for (pfn = spfn; pfn < epfn; pfn++) {
6739 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6740 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6741 + pageblock_nr_pages - 1;
6744 __init_single_page(pfn_to_page(pfn), pfn, zone, node);
6745 __SetPageReserved(pfn_to_page(pfn));
6750 pr_info("On node %d, zone %s: %lld pages in unavailable ranges",
6751 node, zone_names[zone], pgcnt);
6754 static void __init memmap_init_zone_range(struct zone *zone,
6755 unsigned long start_pfn,
6756 unsigned long end_pfn,
6757 unsigned long *hole_pfn)
6759 unsigned long zone_start_pfn = zone->zone_start_pfn;
6760 unsigned long zone_end_pfn = zone_start_pfn + zone->spanned_pages;
6761 int nid = zone_to_nid(zone), zone_id = zone_idx(zone);
6763 start_pfn = clamp(start_pfn, zone_start_pfn, zone_end_pfn);
6764 end_pfn = clamp(end_pfn, zone_start_pfn, zone_end_pfn);
6766 if (start_pfn >= end_pfn)
6769 memmap_init_range(end_pfn - start_pfn, nid, zone_id, start_pfn,
6770 zone_end_pfn, MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6772 if (*hole_pfn < start_pfn)
6773 init_unavailable_range(*hole_pfn, start_pfn, zone_id, nid);
6775 *hole_pfn = end_pfn;
6778 static void __init memmap_init(void)
6780 unsigned long start_pfn, end_pfn;
6781 unsigned long hole_pfn = 0;
6782 int i, j, zone_id = 0, nid;
6784 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6785 struct pglist_data *node = NODE_DATA(nid);
6787 for (j = 0; j < MAX_NR_ZONES; j++) {
6788 struct zone *zone = node->node_zones + j;
6790 if (!populated_zone(zone))
6793 memmap_init_zone_range(zone, start_pfn, end_pfn,
6799 #ifdef CONFIG_SPARSEMEM
6801 * Initialize the memory map for hole in the range [memory_end,
6803 * Append the pages in this hole to the highest zone in the last
6805 * The call to init_unavailable_range() is outside the ifdef to
6806 * silence the compiler warining about zone_id set but not used;
6807 * for FLATMEM it is a nop anyway
6809 end_pfn = round_up(end_pfn, PAGES_PER_SECTION);
6810 if (hole_pfn < end_pfn)
6812 init_unavailable_range(hole_pfn, end_pfn, zone_id, nid);
6815 void __init *memmap_alloc(phys_addr_t size, phys_addr_t align,
6816 phys_addr_t min_addr, int nid, bool exact_nid)
6821 ptr = memblock_alloc_exact_nid_raw(size, align, min_addr,
6822 MEMBLOCK_ALLOC_ACCESSIBLE,
6825 ptr = memblock_alloc_try_nid_raw(size, align, min_addr,
6826 MEMBLOCK_ALLOC_ACCESSIBLE,
6829 if (ptr && size > 0)
6830 page_init_poison(ptr, size);
6835 static int zone_batchsize(struct zone *zone)
6841 * The number of pages to batch allocate is either ~0.1%
6842 * of the zone or 1MB, whichever is smaller. The batch
6843 * size is striking a balance between allocation latency
6844 * and zone lock contention.
6846 batch = min(zone_managed_pages(zone) >> 10, (1024 * 1024) / PAGE_SIZE);
6847 batch /= 4; /* We effectively *= 4 below */
6852 * Clamp the batch to a 2^n - 1 value. Having a power
6853 * of 2 value was found to be more likely to have
6854 * suboptimal cache aliasing properties in some cases.
6856 * For example if 2 tasks are alternately allocating
6857 * batches of pages, one task can end up with a lot
6858 * of pages of one half of the possible page colors
6859 * and the other with pages of the other colors.
6861 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6866 /* The deferral and batching of frees should be suppressed under NOMMU
6869 * The problem is that NOMMU needs to be able to allocate large chunks
6870 * of contiguous memory as there's no hardware page translation to
6871 * assemble apparent contiguous memory from discontiguous pages.
6873 * Queueing large contiguous runs of pages for batching, however,
6874 * causes the pages to actually be freed in smaller chunks. As there
6875 * can be a significant delay between the individual batches being
6876 * recycled, this leads to the once large chunks of space being
6877 * fragmented and becoming unavailable for high-order allocations.
6883 static int zone_highsize(struct zone *zone, int batch, int cpu_online)
6888 unsigned long total_pages;
6890 if (!percpu_pagelist_high_fraction) {
6892 * By default, the high value of the pcp is based on the zone
6893 * low watermark so that if they are full then background
6894 * reclaim will not be started prematurely.
6896 total_pages = low_wmark_pages(zone);
6899 * If percpu_pagelist_high_fraction is configured, the high
6900 * value is based on a fraction of the managed pages in the
6903 total_pages = zone_managed_pages(zone) / percpu_pagelist_high_fraction;
6907 * Split the high value across all online CPUs local to the zone. Note
6908 * that early in boot that CPUs may not be online yet and that during
6909 * CPU hotplug that the cpumask is not yet updated when a CPU is being
6910 * onlined. For memory nodes that have no CPUs, split pcp->high across
6911 * all online CPUs to mitigate the risk that reclaim is triggered
6912 * prematurely due to pages stored on pcp lists.
6914 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
6916 nr_split_cpus = num_online_cpus();
6917 high = total_pages / nr_split_cpus;
6920 * Ensure high is at least batch*4. The multiple is based on the
6921 * historical relationship between high and batch.
6923 high = max(high, batch << 2);
6932 * pcp->high and pcp->batch values are related and generally batch is lower
6933 * than high. They are also related to pcp->count such that count is lower
6934 * than high, and as soon as it reaches high, the pcplist is flushed.
6936 * However, guaranteeing these relations at all times would require e.g. write
6937 * barriers here but also careful usage of read barriers at the read side, and
6938 * thus be prone to error and bad for performance. Thus the update only prevents
6939 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6940 * can cope with those fields changing asynchronously, and fully trust only the
6941 * pcp->count field on the local CPU with interrupts disabled.
6943 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6944 * outside of boot time (or some other assurance that no concurrent updaters
6947 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6948 unsigned long batch)
6950 WRITE_ONCE(pcp->batch, batch);
6951 WRITE_ONCE(pcp->high, high);
6954 static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
6958 memset(pcp, 0, sizeof(*pcp));
6959 memset(pzstats, 0, sizeof(*pzstats));
6961 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
6962 INIT_LIST_HEAD(&pcp->lists[pindex]);
6965 * Set batch and high values safe for a boot pageset. A true percpu
6966 * pageset's initialization will update them subsequently. Here we don't
6967 * need to be as careful as pageset_update() as nobody can access the
6970 pcp->high = BOOT_PAGESET_HIGH;
6971 pcp->batch = BOOT_PAGESET_BATCH;
6972 pcp->free_factor = 0;
6975 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6976 unsigned long batch)
6978 struct per_cpu_pages *pcp;
6981 for_each_possible_cpu(cpu) {
6982 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
6983 pageset_update(pcp, high, batch);
6988 * Calculate and set new high and batch values for all per-cpu pagesets of a
6989 * zone based on the zone's size.
6991 static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
6993 int new_high, new_batch;
6995 new_batch = max(1, zone_batchsize(zone));
6996 new_high = zone_highsize(zone, new_batch, cpu_online);
6998 if (zone->pageset_high == new_high &&
6999 zone->pageset_batch == new_batch)
7002 zone->pageset_high = new_high;
7003 zone->pageset_batch = new_batch;
7005 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
7008 void __meminit setup_zone_pageset(struct zone *zone)
7012 /* Size may be 0 on !SMP && !NUMA */
7013 if (sizeof(struct per_cpu_zonestat) > 0)
7014 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
7016 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
7017 for_each_possible_cpu(cpu) {
7018 struct per_cpu_pages *pcp;
7019 struct per_cpu_zonestat *pzstats;
7021 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
7022 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
7023 per_cpu_pages_init(pcp, pzstats);
7026 zone_set_pageset_high_and_batch(zone, 0);
7030 * Allocate per cpu pagesets and initialize them.
7031 * Before this call only boot pagesets were available.
7033 void __init setup_per_cpu_pageset(void)
7035 struct pglist_data *pgdat;
7037 int __maybe_unused cpu;
7039 for_each_populated_zone(zone)
7040 setup_zone_pageset(zone);
7044 * Unpopulated zones continue using the boot pagesets.
7045 * The numa stats for these pagesets need to be reset.
7046 * Otherwise, they will end up skewing the stats of
7047 * the nodes these zones are associated with.
7049 for_each_possible_cpu(cpu) {
7050 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
7051 memset(pzstats->vm_numa_event, 0,
7052 sizeof(pzstats->vm_numa_event));
7056 for_each_online_pgdat(pgdat)
7057 pgdat->per_cpu_nodestats =
7058 alloc_percpu(struct per_cpu_nodestat);
7061 static __meminit void zone_pcp_init(struct zone *zone)
7064 * per cpu subsystem is not up at this point. The following code
7065 * relies on the ability of the linker to provide the
7066 * offset of a (static) per cpu variable into the per cpu area.
7068 zone->per_cpu_pageset = &boot_pageset;
7069 zone->per_cpu_zonestats = &boot_zonestats;
7070 zone->pageset_high = BOOT_PAGESET_HIGH;
7071 zone->pageset_batch = BOOT_PAGESET_BATCH;
7073 if (populated_zone(zone))
7074 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
7075 zone->present_pages, zone_batchsize(zone));
7078 void __meminit init_currently_empty_zone(struct zone *zone,
7079 unsigned long zone_start_pfn,
7082 struct pglist_data *pgdat = zone->zone_pgdat;
7083 int zone_idx = zone_idx(zone) + 1;
7085 if (zone_idx > pgdat->nr_zones)
7086 pgdat->nr_zones = zone_idx;
7088 zone->zone_start_pfn = zone_start_pfn;
7090 mminit_dprintk(MMINIT_TRACE, "memmap_init",
7091 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
7093 (unsigned long)zone_idx(zone),
7094 zone_start_pfn, (zone_start_pfn + size));
7096 zone_init_free_lists(zone);
7097 zone->initialized = 1;
7101 * get_pfn_range_for_nid - Return the start and end page frames for a node
7102 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
7103 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
7104 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
7106 * It returns the start and end page frame of a node based on information
7107 * provided by memblock_set_node(). If called for a node
7108 * with no available memory, a warning is printed and the start and end
7111 void __init get_pfn_range_for_nid(unsigned int nid,
7112 unsigned long *start_pfn, unsigned long *end_pfn)
7114 unsigned long this_start_pfn, this_end_pfn;
7120 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
7121 *start_pfn = min(*start_pfn, this_start_pfn);
7122 *end_pfn = max(*end_pfn, this_end_pfn);
7125 if (*start_pfn == -1UL)
7130 * This finds a zone that can be used for ZONE_MOVABLE pages. The
7131 * assumption is made that zones within a node are ordered in monotonic
7132 * increasing memory addresses so that the "highest" populated zone is used
7134 static void __init find_usable_zone_for_movable(void)
7137 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
7138 if (zone_index == ZONE_MOVABLE)
7141 if (arch_zone_highest_possible_pfn[zone_index] >
7142 arch_zone_lowest_possible_pfn[zone_index])
7146 VM_BUG_ON(zone_index == -1);
7147 movable_zone = zone_index;
7151 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
7152 * because it is sized independent of architecture. Unlike the other zones,
7153 * the starting point for ZONE_MOVABLE is not fixed. It may be different
7154 * in each node depending on the size of each node and how evenly kernelcore
7155 * is distributed. This helper function adjusts the zone ranges
7156 * provided by the architecture for a given node by using the end of the
7157 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
7158 * zones within a node are in order of monotonic increases memory addresses
7160 static void __init adjust_zone_range_for_zone_movable(int nid,
7161 unsigned long zone_type,
7162 unsigned long node_start_pfn,
7163 unsigned long node_end_pfn,
7164 unsigned long *zone_start_pfn,
7165 unsigned long *zone_end_pfn)
7167 /* Only adjust if ZONE_MOVABLE is on this node */
7168 if (zone_movable_pfn[nid]) {
7169 /* Size ZONE_MOVABLE */
7170 if (zone_type == ZONE_MOVABLE) {
7171 *zone_start_pfn = zone_movable_pfn[nid];
7172 *zone_end_pfn = min(node_end_pfn,
7173 arch_zone_highest_possible_pfn[movable_zone]);
7175 /* Adjust for ZONE_MOVABLE starting within this range */
7176 } else if (!mirrored_kernelcore &&
7177 *zone_start_pfn < zone_movable_pfn[nid] &&
7178 *zone_end_pfn > zone_movable_pfn[nid]) {
7179 *zone_end_pfn = zone_movable_pfn[nid];
7181 /* Check if this whole range is within ZONE_MOVABLE */
7182 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
7183 *zone_start_pfn = *zone_end_pfn;
7188 * Return the number of pages a zone spans in a node, including holes
7189 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
7191 static unsigned long __init zone_spanned_pages_in_node(int nid,
7192 unsigned long zone_type,
7193 unsigned long node_start_pfn,
7194 unsigned long node_end_pfn,
7195 unsigned long *zone_start_pfn,
7196 unsigned long *zone_end_pfn)
7198 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7199 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7200 /* When hotadd a new node from cpu_up(), the node should be empty */
7201 if (!node_start_pfn && !node_end_pfn)
7204 /* Get the start and end of the zone */
7205 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7206 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7207 adjust_zone_range_for_zone_movable(nid, zone_type,
7208 node_start_pfn, node_end_pfn,
7209 zone_start_pfn, zone_end_pfn);
7211 /* Check that this node has pages within the zone's required range */
7212 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
7215 /* Move the zone boundaries inside the node if necessary */
7216 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
7217 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
7219 /* Return the spanned pages */
7220 return *zone_end_pfn - *zone_start_pfn;
7224 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
7225 * then all holes in the requested range will be accounted for.
7227 unsigned long __init __absent_pages_in_range(int nid,
7228 unsigned long range_start_pfn,
7229 unsigned long range_end_pfn)
7231 unsigned long nr_absent = range_end_pfn - range_start_pfn;
7232 unsigned long start_pfn, end_pfn;
7235 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7236 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
7237 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
7238 nr_absent -= end_pfn - start_pfn;
7244 * absent_pages_in_range - Return number of page frames in holes within a range
7245 * @start_pfn: The start PFN to start searching for holes
7246 * @end_pfn: The end PFN to stop searching for holes
7248 * Return: the number of pages frames in memory holes within a range.
7250 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
7251 unsigned long end_pfn)
7253 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
7256 /* Return the number of page frames in holes in a zone on a node */
7257 static unsigned long __init zone_absent_pages_in_node(int nid,
7258 unsigned long zone_type,
7259 unsigned long node_start_pfn,
7260 unsigned long node_end_pfn)
7262 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
7263 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
7264 unsigned long zone_start_pfn, zone_end_pfn;
7265 unsigned long nr_absent;
7267 /* When hotadd a new node from cpu_up(), the node should be empty */
7268 if (!node_start_pfn && !node_end_pfn)
7271 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
7272 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
7274 adjust_zone_range_for_zone_movable(nid, zone_type,
7275 node_start_pfn, node_end_pfn,
7276 &zone_start_pfn, &zone_end_pfn);
7277 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
7280 * ZONE_MOVABLE handling.
7281 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
7284 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
7285 unsigned long start_pfn, end_pfn;
7286 struct memblock_region *r;
7288 for_each_mem_region(r) {
7289 start_pfn = clamp(memblock_region_memory_base_pfn(r),
7290 zone_start_pfn, zone_end_pfn);
7291 end_pfn = clamp(memblock_region_memory_end_pfn(r),
7292 zone_start_pfn, zone_end_pfn);
7294 if (zone_type == ZONE_MOVABLE &&
7295 memblock_is_mirror(r))
7296 nr_absent += end_pfn - start_pfn;
7298 if (zone_type == ZONE_NORMAL &&
7299 !memblock_is_mirror(r))
7300 nr_absent += end_pfn - start_pfn;
7307 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
7308 unsigned long node_start_pfn,
7309 unsigned long node_end_pfn)
7311 unsigned long realtotalpages = 0, totalpages = 0;
7314 for (i = 0; i < MAX_NR_ZONES; i++) {
7315 struct zone *zone = pgdat->node_zones + i;
7316 unsigned long zone_start_pfn, zone_end_pfn;
7317 unsigned long spanned, absent;
7318 unsigned long size, real_size;
7320 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
7325 absent = zone_absent_pages_in_node(pgdat->node_id, i,
7330 real_size = size - absent;
7333 zone->zone_start_pfn = zone_start_pfn;
7335 zone->zone_start_pfn = 0;
7336 zone->spanned_pages = size;
7337 zone->present_pages = real_size;
7338 #if defined(CONFIG_MEMORY_HOTPLUG)
7339 zone->present_early_pages = real_size;
7343 realtotalpages += real_size;
7346 pgdat->node_spanned_pages = totalpages;
7347 pgdat->node_present_pages = realtotalpages;
7348 pr_debug("On node %d totalpages: %lu\n", pgdat->node_id, realtotalpages);
7351 #ifndef CONFIG_SPARSEMEM
7353 * Calculate the size of the zone->blockflags rounded to an unsigned long
7354 * Start by making sure zonesize is a multiple of pageblock_order by rounding
7355 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
7356 * round what is now in bits to nearest long in bits, then return it in
7359 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
7361 unsigned long usemapsize;
7363 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
7364 usemapsize = roundup(zonesize, pageblock_nr_pages);
7365 usemapsize = usemapsize >> pageblock_order;
7366 usemapsize *= NR_PAGEBLOCK_BITS;
7367 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
7369 return usemapsize / 8;
7372 static void __ref setup_usemap(struct zone *zone)
7374 unsigned long usemapsize = usemap_size(zone->zone_start_pfn,
7375 zone->spanned_pages);
7376 zone->pageblock_flags = NULL;
7378 zone->pageblock_flags =
7379 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
7381 if (!zone->pageblock_flags)
7382 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
7383 usemapsize, zone->name, zone_to_nid(zone));
7387 static inline void setup_usemap(struct zone *zone) {}
7388 #endif /* CONFIG_SPARSEMEM */
7390 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
7392 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
7393 void __init set_pageblock_order(void)
7397 /* Check that pageblock_nr_pages has not already been setup */
7398 if (pageblock_order)
7401 if (HPAGE_SHIFT > PAGE_SHIFT)
7402 order = HUGETLB_PAGE_ORDER;
7404 order = MAX_ORDER - 1;
7407 * Assume the largest contiguous order of interest is a huge page.
7408 * This value may be variable depending on boot parameters on IA64 and
7411 pageblock_order = order;
7413 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7416 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
7417 * is unused as pageblock_order is set at compile-time. See
7418 * include/linux/pageblock-flags.h for the values of pageblock_order based on
7421 void __init set_pageblock_order(void)
7425 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
7427 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
7428 unsigned long present_pages)
7430 unsigned long pages = spanned_pages;
7433 * Provide a more accurate estimation if there are holes within
7434 * the zone and SPARSEMEM is in use. If there are holes within the
7435 * zone, each populated memory region may cost us one or two extra
7436 * memmap pages due to alignment because memmap pages for each
7437 * populated regions may not be naturally aligned on page boundary.
7438 * So the (present_pages >> 4) heuristic is a tradeoff for that.
7440 if (spanned_pages > present_pages + (present_pages >> 4) &&
7441 IS_ENABLED(CONFIG_SPARSEMEM))
7442 pages = present_pages;
7444 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
7447 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
7448 static void pgdat_init_split_queue(struct pglist_data *pgdat)
7450 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
7452 spin_lock_init(&ds_queue->split_queue_lock);
7453 INIT_LIST_HEAD(&ds_queue->split_queue);
7454 ds_queue->split_queue_len = 0;
7457 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
7460 #ifdef CONFIG_COMPACTION
7461 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
7463 init_waitqueue_head(&pgdat->kcompactd_wait);
7466 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
7469 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
7471 pgdat_resize_init(pgdat);
7473 pgdat_init_split_queue(pgdat);
7474 pgdat_init_kcompactd(pgdat);
7476 init_waitqueue_head(&pgdat->kswapd_wait);
7477 init_waitqueue_head(&pgdat->pfmemalloc_wait);
7479 pgdat_page_ext_init(pgdat);
7480 lruvec_init(&pgdat->__lruvec);
7483 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
7484 unsigned long remaining_pages)
7486 atomic_long_set(&zone->managed_pages, remaining_pages);
7487 zone_set_nid(zone, nid);
7488 zone->name = zone_names[idx];
7489 zone->zone_pgdat = NODE_DATA(nid);
7490 spin_lock_init(&zone->lock);
7491 zone_seqlock_init(zone);
7492 zone_pcp_init(zone);
7496 * Set up the zone data structures
7497 * - init pgdat internals
7498 * - init all zones belonging to this node
7500 * NOTE: this function is only called during memory hotplug
7502 #ifdef CONFIG_MEMORY_HOTPLUG
7503 void __ref free_area_init_core_hotplug(int nid)
7506 pg_data_t *pgdat = NODE_DATA(nid);
7508 pgdat_init_internals(pgdat);
7509 for (z = 0; z < MAX_NR_ZONES; z++)
7510 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
7515 * Set up the zone data structures:
7516 * - mark all pages reserved
7517 * - mark all memory queues empty
7518 * - clear the memory bitmaps
7520 * NOTE: pgdat should get zeroed by caller.
7521 * NOTE: this function is only called during early init.
7523 static void __init free_area_init_core(struct pglist_data *pgdat)
7526 int nid = pgdat->node_id;
7528 pgdat_init_internals(pgdat);
7529 pgdat->per_cpu_nodestats = &boot_nodestats;
7531 for (j = 0; j < MAX_NR_ZONES; j++) {
7532 struct zone *zone = pgdat->node_zones + j;
7533 unsigned long size, freesize, memmap_pages;
7535 size = zone->spanned_pages;
7536 freesize = zone->present_pages;
7539 * Adjust freesize so that it accounts for how much memory
7540 * is used by this zone for memmap. This affects the watermark
7541 * and per-cpu initialisations
7543 memmap_pages = calc_memmap_size(size, freesize);
7544 if (!is_highmem_idx(j)) {
7545 if (freesize >= memmap_pages) {
7546 freesize -= memmap_pages;
7548 pr_debug(" %s zone: %lu pages used for memmap\n",
7549 zone_names[j], memmap_pages);
7551 pr_warn(" %s zone: %lu memmap pages exceeds freesize %lu\n",
7552 zone_names[j], memmap_pages, freesize);
7555 /* Account for reserved pages */
7556 if (j == 0 && freesize > dma_reserve) {
7557 freesize -= dma_reserve;
7558 pr_debug(" %s zone: %lu pages reserved\n", zone_names[0], dma_reserve);
7561 if (!is_highmem_idx(j))
7562 nr_kernel_pages += freesize;
7563 /* Charge for highmem memmap if there are enough kernel pages */
7564 else if (nr_kernel_pages > memmap_pages * 2)
7565 nr_kernel_pages -= memmap_pages;
7566 nr_all_pages += freesize;
7569 * Set an approximate value for lowmem here, it will be adjusted
7570 * when the bootmem allocator frees pages into the buddy system.
7571 * And all highmem pages will be managed by the buddy system.
7573 zone_init_internals(zone, j, nid, freesize);
7578 set_pageblock_order();
7580 init_currently_empty_zone(zone, zone->zone_start_pfn, size);
7584 #ifdef CONFIG_FLATMEM
7585 static void __init alloc_node_mem_map(struct pglist_data *pgdat)
7587 unsigned long __maybe_unused start = 0;
7588 unsigned long __maybe_unused offset = 0;
7590 /* Skip empty nodes */
7591 if (!pgdat->node_spanned_pages)
7594 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
7595 offset = pgdat->node_start_pfn - start;
7596 /* ia64 gets its own node_mem_map, before this, without bootmem */
7597 if (!pgdat->node_mem_map) {
7598 unsigned long size, end;
7602 * The zone's endpoints aren't required to be MAX_ORDER
7603 * aligned but the node_mem_map endpoints must be in order
7604 * for the buddy allocator to function correctly.
7606 end = pgdat_end_pfn(pgdat);
7607 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7608 size = (end - start) * sizeof(struct page);
7609 map = memmap_alloc(size, SMP_CACHE_BYTES, MEMBLOCK_LOW_LIMIT,
7610 pgdat->node_id, false);
7612 panic("Failed to allocate %ld bytes for node %d memory map\n",
7613 size, pgdat->node_id);
7614 pgdat->node_mem_map = map + offset;
7616 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7617 __func__, pgdat->node_id, (unsigned long)pgdat,
7618 (unsigned long)pgdat->node_mem_map);
7621 * With no DISCONTIG, the global mem_map is just set as node 0's
7623 if (pgdat == NODE_DATA(0)) {
7624 mem_map = NODE_DATA(0)->node_mem_map;
7625 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7631 static inline void alloc_node_mem_map(struct pglist_data *pgdat) { }
7632 #endif /* CONFIG_FLATMEM */
7634 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7635 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7637 pgdat->first_deferred_pfn = ULONG_MAX;
7640 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7643 static void __init free_area_init_node(int nid)
7645 pg_data_t *pgdat = NODE_DATA(nid);
7646 unsigned long start_pfn = 0;
7647 unsigned long end_pfn = 0;
7649 /* pg_data_t should be reset to zero when it's allocated */
7650 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7652 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7654 pgdat->node_id = nid;
7655 pgdat->node_start_pfn = start_pfn;
7656 pgdat->per_cpu_nodestats = NULL;
7658 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7659 (u64)start_pfn << PAGE_SHIFT,
7660 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7661 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7663 alloc_node_mem_map(pgdat);
7664 pgdat_set_deferred_range(pgdat);
7666 free_area_init_core(pgdat);
7669 void __init free_area_init_memoryless_node(int nid)
7671 free_area_init_node(nid);
7674 #if MAX_NUMNODES > 1
7676 * Figure out the number of possible node ids.
7678 void __init setup_nr_node_ids(void)
7680 unsigned int highest;
7682 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7683 nr_node_ids = highest + 1;
7688 * node_map_pfn_alignment - determine the maximum internode alignment
7690 * This function should be called after node map is populated and sorted.
7691 * It calculates the maximum power of two alignment which can distinguish
7694 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7695 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7696 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7697 * shifted, 1GiB is enough and this function will indicate so.
7699 * This is used to test whether pfn -> nid mapping of the chosen memory
7700 * model has fine enough granularity to avoid incorrect mapping for the
7701 * populated node map.
7703 * Return: the determined alignment in pfn's. 0 if there is no alignment
7704 * requirement (single node).
7706 unsigned long __init node_map_pfn_alignment(void)
7708 unsigned long accl_mask = 0, last_end = 0;
7709 unsigned long start, end, mask;
7710 int last_nid = NUMA_NO_NODE;
7713 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7714 if (!start || last_nid < 0 || last_nid == nid) {
7721 * Start with a mask granular enough to pin-point to the
7722 * start pfn and tick off bits one-by-one until it becomes
7723 * too coarse to separate the current node from the last.
7725 mask = ~((1 << __ffs(start)) - 1);
7726 while (mask && last_end <= (start & (mask << 1)))
7729 /* accumulate all internode masks */
7733 /* convert mask to number of pages */
7734 return ~accl_mask + 1;
7738 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7740 * Return: the minimum PFN based on information provided via
7741 * memblock_set_node().
7743 unsigned long __init find_min_pfn_with_active_regions(void)
7745 return PHYS_PFN(memblock_start_of_DRAM());
7749 * early_calculate_totalpages()
7750 * Sum pages in active regions for movable zone.
7751 * Populate N_MEMORY for calculating usable_nodes.
7753 static unsigned long __init early_calculate_totalpages(void)
7755 unsigned long totalpages = 0;
7756 unsigned long start_pfn, end_pfn;
7759 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7760 unsigned long pages = end_pfn - start_pfn;
7762 totalpages += pages;
7764 node_set_state(nid, N_MEMORY);
7770 * Find the PFN the Movable zone begins in each node. Kernel memory
7771 * is spread evenly between nodes as long as the nodes have enough
7772 * memory. When they don't, some nodes will have more kernelcore than
7775 static void __init find_zone_movable_pfns_for_nodes(void)
7778 unsigned long usable_startpfn;
7779 unsigned long kernelcore_node, kernelcore_remaining;
7780 /* save the state before borrow the nodemask */
7781 nodemask_t saved_node_state = node_states[N_MEMORY];
7782 unsigned long totalpages = early_calculate_totalpages();
7783 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7784 struct memblock_region *r;
7786 /* Need to find movable_zone earlier when movable_node is specified. */
7787 find_usable_zone_for_movable();
7790 * If movable_node is specified, ignore kernelcore and movablecore
7793 if (movable_node_is_enabled()) {
7794 for_each_mem_region(r) {
7795 if (!memblock_is_hotpluggable(r))
7798 nid = memblock_get_region_node(r);
7800 usable_startpfn = PFN_DOWN(r->base);
7801 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7802 min(usable_startpfn, zone_movable_pfn[nid]) :
7810 * If kernelcore=mirror is specified, ignore movablecore option
7812 if (mirrored_kernelcore) {
7813 bool mem_below_4gb_not_mirrored = false;
7815 for_each_mem_region(r) {
7816 if (memblock_is_mirror(r))
7819 nid = memblock_get_region_node(r);
7821 usable_startpfn = memblock_region_memory_base_pfn(r);
7823 if (usable_startpfn < 0x100000) {
7824 mem_below_4gb_not_mirrored = true;
7828 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7829 min(usable_startpfn, zone_movable_pfn[nid]) :
7833 if (mem_below_4gb_not_mirrored)
7834 pr_warn("This configuration results in unmirrored kernel memory.\n");
7840 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7841 * amount of necessary memory.
7843 if (required_kernelcore_percent)
7844 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7846 if (required_movablecore_percent)
7847 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7851 * If movablecore= was specified, calculate what size of
7852 * kernelcore that corresponds so that memory usable for
7853 * any allocation type is evenly spread. If both kernelcore
7854 * and movablecore are specified, then the value of kernelcore
7855 * will be used for required_kernelcore if it's greater than
7856 * what movablecore would have allowed.
7858 if (required_movablecore) {
7859 unsigned long corepages;
7862 * Round-up so that ZONE_MOVABLE is at least as large as what
7863 * was requested by the user
7865 required_movablecore =
7866 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7867 required_movablecore = min(totalpages, required_movablecore);
7868 corepages = totalpages - required_movablecore;
7870 required_kernelcore = max(required_kernelcore, corepages);
7874 * If kernelcore was not specified or kernelcore size is larger
7875 * than totalpages, there is no ZONE_MOVABLE.
7877 if (!required_kernelcore || required_kernelcore >= totalpages)
7880 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7881 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7884 /* Spread kernelcore memory as evenly as possible throughout nodes */
7885 kernelcore_node = required_kernelcore / usable_nodes;
7886 for_each_node_state(nid, N_MEMORY) {
7887 unsigned long start_pfn, end_pfn;
7890 * Recalculate kernelcore_node if the division per node
7891 * now exceeds what is necessary to satisfy the requested
7892 * amount of memory for the kernel
7894 if (required_kernelcore < kernelcore_node)
7895 kernelcore_node = required_kernelcore / usable_nodes;
7898 * As the map is walked, we track how much memory is usable
7899 * by the kernel using kernelcore_remaining. When it is
7900 * 0, the rest of the node is usable by ZONE_MOVABLE
7902 kernelcore_remaining = kernelcore_node;
7904 /* Go through each range of PFNs within this node */
7905 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7906 unsigned long size_pages;
7908 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7909 if (start_pfn >= end_pfn)
7912 /* Account for what is only usable for kernelcore */
7913 if (start_pfn < usable_startpfn) {
7914 unsigned long kernel_pages;
7915 kernel_pages = min(end_pfn, usable_startpfn)
7918 kernelcore_remaining -= min(kernel_pages,
7919 kernelcore_remaining);
7920 required_kernelcore -= min(kernel_pages,
7921 required_kernelcore);
7923 /* Continue if range is now fully accounted */
7924 if (end_pfn <= usable_startpfn) {
7927 * Push zone_movable_pfn to the end so
7928 * that if we have to rebalance
7929 * kernelcore across nodes, we will
7930 * not double account here
7932 zone_movable_pfn[nid] = end_pfn;
7935 start_pfn = usable_startpfn;
7939 * The usable PFN range for ZONE_MOVABLE is from
7940 * start_pfn->end_pfn. Calculate size_pages as the
7941 * number of pages used as kernelcore
7943 size_pages = end_pfn - start_pfn;
7944 if (size_pages > kernelcore_remaining)
7945 size_pages = kernelcore_remaining;
7946 zone_movable_pfn[nid] = start_pfn + size_pages;
7949 * Some kernelcore has been met, update counts and
7950 * break if the kernelcore for this node has been
7953 required_kernelcore -= min(required_kernelcore,
7955 kernelcore_remaining -= size_pages;
7956 if (!kernelcore_remaining)
7962 * If there is still required_kernelcore, we do another pass with one
7963 * less node in the count. This will push zone_movable_pfn[nid] further
7964 * along on the nodes that still have memory until kernelcore is
7968 if (usable_nodes && required_kernelcore > usable_nodes)
7972 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7973 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7974 unsigned long start_pfn, end_pfn;
7976 zone_movable_pfn[nid] =
7977 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7979 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7980 if (zone_movable_pfn[nid] >= end_pfn)
7981 zone_movable_pfn[nid] = 0;
7985 /* restore the node_state */
7986 node_states[N_MEMORY] = saved_node_state;
7989 /* Any regular or high memory on that node ? */
7990 static void check_for_memory(pg_data_t *pgdat, int nid)
7992 enum zone_type zone_type;
7994 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7995 struct zone *zone = &pgdat->node_zones[zone_type];
7996 if (populated_zone(zone)) {
7997 if (IS_ENABLED(CONFIG_HIGHMEM))
7998 node_set_state(nid, N_HIGH_MEMORY);
7999 if (zone_type <= ZONE_NORMAL)
8000 node_set_state(nid, N_NORMAL_MEMORY);
8007 * Some architectures, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
8008 * such cases we allow max_zone_pfn sorted in the descending order
8010 bool __weak arch_has_descending_max_zone_pfns(void)
8016 * free_area_init - Initialise all pg_data_t and zone data
8017 * @max_zone_pfn: an array of max PFNs for each zone
8019 * This will call free_area_init_node() for each active node in the system.
8020 * Using the page ranges provided by memblock_set_node(), the size of each
8021 * zone in each node and their holes is calculated. If the maximum PFN
8022 * between two adjacent zones match, it is assumed that the zone is empty.
8023 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
8024 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
8025 * starts where the previous one ended. For example, ZONE_DMA32 starts
8026 * at arch_max_dma_pfn.
8028 void __init free_area_init(unsigned long *max_zone_pfn)
8030 unsigned long start_pfn, end_pfn;
8034 /* Record where the zone boundaries are */
8035 memset(arch_zone_lowest_possible_pfn, 0,
8036 sizeof(arch_zone_lowest_possible_pfn));
8037 memset(arch_zone_highest_possible_pfn, 0,
8038 sizeof(arch_zone_highest_possible_pfn));
8040 start_pfn = find_min_pfn_with_active_regions();
8041 descending = arch_has_descending_max_zone_pfns();
8043 for (i = 0; i < MAX_NR_ZONES; i++) {
8045 zone = MAX_NR_ZONES - i - 1;
8049 if (zone == ZONE_MOVABLE)
8052 end_pfn = max(max_zone_pfn[zone], start_pfn);
8053 arch_zone_lowest_possible_pfn[zone] = start_pfn;
8054 arch_zone_highest_possible_pfn[zone] = end_pfn;
8056 start_pfn = end_pfn;
8059 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
8060 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
8061 find_zone_movable_pfns_for_nodes();
8063 /* Print out the zone ranges */
8064 pr_info("Zone ranges:\n");
8065 for (i = 0; i < MAX_NR_ZONES; i++) {
8066 if (i == ZONE_MOVABLE)
8068 pr_info(" %-8s ", zone_names[i]);
8069 if (arch_zone_lowest_possible_pfn[i] ==
8070 arch_zone_highest_possible_pfn[i])
8073 pr_cont("[mem %#018Lx-%#018Lx]\n",
8074 (u64)arch_zone_lowest_possible_pfn[i]
8076 ((u64)arch_zone_highest_possible_pfn[i]
8077 << PAGE_SHIFT) - 1);
8080 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
8081 pr_info("Movable zone start for each node\n");
8082 for (i = 0; i < MAX_NUMNODES; i++) {
8083 if (zone_movable_pfn[i])
8084 pr_info(" Node %d: %#018Lx\n", i,
8085 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
8089 * Print out the early node map, and initialize the
8090 * subsection-map relative to active online memory ranges to
8091 * enable future "sub-section" extensions of the memory map.
8093 pr_info("Early memory node ranges\n");
8094 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
8095 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
8096 (u64)start_pfn << PAGE_SHIFT,
8097 ((u64)end_pfn << PAGE_SHIFT) - 1);
8098 subsection_map_init(start_pfn, end_pfn - start_pfn);
8101 /* Initialise every node */
8102 mminit_verify_pageflags_layout();
8103 setup_nr_node_ids();
8104 for_each_online_node(nid) {
8105 pg_data_t *pgdat = NODE_DATA(nid);
8106 free_area_init_node(nid);
8108 /* Any memory on that node */
8109 if (pgdat->node_present_pages)
8110 node_set_state(nid, N_MEMORY);
8111 check_for_memory(pgdat, nid);
8117 static int __init cmdline_parse_core(char *p, unsigned long *core,
8118 unsigned long *percent)
8120 unsigned long long coremem;
8126 /* Value may be a percentage of total memory, otherwise bytes */
8127 coremem = simple_strtoull(p, &endptr, 0);
8128 if (*endptr == '%') {
8129 /* Paranoid check for percent values greater than 100 */
8130 WARN_ON(coremem > 100);
8134 coremem = memparse(p, &p);
8135 /* Paranoid check that UL is enough for the coremem value */
8136 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
8138 *core = coremem >> PAGE_SHIFT;
8145 * kernelcore=size sets the amount of memory for use for allocations that
8146 * cannot be reclaimed or migrated.
8148 static int __init cmdline_parse_kernelcore(char *p)
8150 /* parse kernelcore=mirror */
8151 if (parse_option_str(p, "mirror")) {
8152 mirrored_kernelcore = true;
8156 return cmdline_parse_core(p, &required_kernelcore,
8157 &required_kernelcore_percent);
8161 * movablecore=size sets the amount of memory for use for allocations that
8162 * can be reclaimed or migrated.
8164 static int __init cmdline_parse_movablecore(char *p)
8166 return cmdline_parse_core(p, &required_movablecore,
8167 &required_movablecore_percent);
8170 early_param("kernelcore", cmdline_parse_kernelcore);
8171 early_param("movablecore", cmdline_parse_movablecore);
8173 void adjust_managed_page_count(struct page *page, long count)
8175 atomic_long_add(count, &page_zone(page)->managed_pages);
8176 totalram_pages_add(count);
8177 #ifdef CONFIG_HIGHMEM
8178 if (PageHighMem(page))
8179 totalhigh_pages_add(count);
8182 EXPORT_SYMBOL(adjust_managed_page_count);
8184 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
8187 unsigned long pages = 0;
8189 start = (void *)PAGE_ALIGN((unsigned long)start);
8190 end = (void *)((unsigned long)end & PAGE_MASK);
8191 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
8192 struct page *page = virt_to_page(pos);
8193 void *direct_map_addr;
8196 * 'direct_map_addr' might be different from 'pos'
8197 * because some architectures' virt_to_page()
8198 * work with aliases. Getting the direct map
8199 * address ensures that we get a _writeable_
8200 * alias for the memset().
8202 direct_map_addr = page_address(page);
8204 * Perform a kasan-unchecked memset() since this memory
8205 * has not been initialized.
8207 direct_map_addr = kasan_reset_tag(direct_map_addr);
8208 if ((unsigned int)poison <= 0xFF)
8209 memset(direct_map_addr, poison, PAGE_SIZE);
8211 free_reserved_page(page);
8215 pr_info("Freeing %s memory: %ldK\n",
8216 s, pages << (PAGE_SHIFT - 10));
8221 void __init mem_init_print_info(void)
8223 unsigned long physpages, codesize, datasize, rosize, bss_size;
8224 unsigned long init_code_size, init_data_size;
8226 physpages = get_num_physpages();
8227 codesize = _etext - _stext;
8228 datasize = _edata - _sdata;
8229 rosize = __end_rodata - __start_rodata;
8230 bss_size = __bss_stop - __bss_start;
8231 init_data_size = __init_end - __init_begin;
8232 init_code_size = _einittext - _sinittext;
8235 * Detect special cases and adjust section sizes accordingly:
8236 * 1) .init.* may be embedded into .data sections
8237 * 2) .init.text.* may be out of [__init_begin, __init_end],
8238 * please refer to arch/tile/kernel/vmlinux.lds.S.
8239 * 3) .rodata.* may be embedded into .text or .data sections.
8241 #define adj_init_size(start, end, size, pos, adj) \
8243 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
8247 adj_init_size(__init_begin, __init_end, init_data_size,
8248 _sinittext, init_code_size);
8249 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
8250 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
8251 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
8252 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
8254 #undef adj_init_size
8256 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
8257 #ifdef CONFIG_HIGHMEM
8261 nr_free_pages() << (PAGE_SHIFT - 10),
8262 physpages << (PAGE_SHIFT - 10),
8263 codesize >> 10, datasize >> 10, rosize >> 10,
8264 (init_data_size + init_code_size) >> 10, bss_size >> 10,
8265 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
8266 totalcma_pages << (PAGE_SHIFT - 10)
8267 #ifdef CONFIG_HIGHMEM
8268 , totalhigh_pages() << (PAGE_SHIFT - 10)
8274 * set_dma_reserve - set the specified number of pages reserved in the first zone
8275 * @new_dma_reserve: The number of pages to mark reserved
8277 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
8278 * In the DMA zone, a significant percentage may be consumed by kernel image
8279 * and other unfreeable allocations which can skew the watermarks badly. This
8280 * function may optionally be used to account for unfreeable pages in the
8281 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
8282 * smaller per-cpu batchsize.
8284 void __init set_dma_reserve(unsigned long new_dma_reserve)
8286 dma_reserve = new_dma_reserve;
8289 static int page_alloc_cpu_dead(unsigned int cpu)
8293 lru_add_drain_cpu(cpu);
8297 * Spill the event counters of the dead processor
8298 * into the current processors event counters.
8299 * This artificially elevates the count of the current
8302 vm_events_fold_cpu(cpu);
8305 * Zero the differential counters of the dead processor
8306 * so that the vm statistics are consistent.
8308 * This is only okay since the processor is dead and cannot
8309 * race with what we are doing.
8311 cpu_vm_stats_fold(cpu);
8313 for_each_populated_zone(zone)
8314 zone_pcp_update(zone, 0);
8319 static int page_alloc_cpu_online(unsigned int cpu)
8323 for_each_populated_zone(zone)
8324 zone_pcp_update(zone, 1);
8329 int hashdist = HASHDIST_DEFAULT;
8331 static int __init set_hashdist(char *str)
8335 hashdist = simple_strtoul(str, &str, 0);
8338 __setup("hashdist=", set_hashdist);
8341 void __init page_alloc_init(void)
8346 if (num_node_state(N_MEMORY) == 1)
8350 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
8351 "mm/page_alloc:pcp",
8352 page_alloc_cpu_online,
8353 page_alloc_cpu_dead);
8358 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
8359 * or min_free_kbytes changes.
8361 static void calculate_totalreserve_pages(void)
8363 struct pglist_data *pgdat;
8364 unsigned long reserve_pages = 0;
8365 enum zone_type i, j;
8367 for_each_online_pgdat(pgdat) {
8369 pgdat->totalreserve_pages = 0;
8371 for (i = 0; i < MAX_NR_ZONES; i++) {
8372 struct zone *zone = pgdat->node_zones + i;
8374 unsigned long managed_pages = zone_managed_pages(zone);
8376 /* Find valid and maximum lowmem_reserve in the zone */
8377 for (j = i; j < MAX_NR_ZONES; j++) {
8378 if (zone->lowmem_reserve[j] > max)
8379 max = zone->lowmem_reserve[j];
8382 /* we treat the high watermark as reserved pages. */
8383 max += high_wmark_pages(zone);
8385 if (max > managed_pages)
8386 max = managed_pages;
8388 pgdat->totalreserve_pages += max;
8390 reserve_pages += max;
8393 totalreserve_pages = reserve_pages;
8397 * setup_per_zone_lowmem_reserve - called whenever
8398 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
8399 * has a correct pages reserved value, so an adequate number of
8400 * pages are left in the zone after a successful __alloc_pages().
8402 static void setup_per_zone_lowmem_reserve(void)
8404 struct pglist_data *pgdat;
8405 enum zone_type i, j;
8407 for_each_online_pgdat(pgdat) {
8408 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
8409 struct zone *zone = &pgdat->node_zones[i];
8410 int ratio = sysctl_lowmem_reserve_ratio[i];
8411 bool clear = !ratio || !zone_managed_pages(zone);
8412 unsigned long managed_pages = 0;
8414 for (j = i + 1; j < MAX_NR_ZONES; j++) {
8415 struct zone *upper_zone = &pgdat->node_zones[j];
8417 managed_pages += zone_managed_pages(upper_zone);
8420 zone->lowmem_reserve[j] = 0;
8422 zone->lowmem_reserve[j] = managed_pages / ratio;
8427 /* update totalreserve_pages */
8428 calculate_totalreserve_pages();
8431 static void __setup_per_zone_wmarks(void)
8433 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
8434 unsigned long lowmem_pages = 0;
8436 unsigned long flags;
8438 /* Calculate total number of !ZONE_HIGHMEM pages */
8439 for_each_zone(zone) {
8440 if (!is_highmem(zone))
8441 lowmem_pages += zone_managed_pages(zone);
8444 for_each_zone(zone) {
8447 spin_lock_irqsave(&zone->lock, flags);
8448 tmp = (u64)pages_min * zone_managed_pages(zone);
8449 do_div(tmp, lowmem_pages);
8450 if (is_highmem(zone)) {
8452 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
8453 * need highmem pages, so cap pages_min to a small
8456 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
8457 * deltas control async page reclaim, and so should
8458 * not be capped for highmem.
8460 unsigned long min_pages;
8462 min_pages = zone_managed_pages(zone) / 1024;
8463 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
8464 zone->_watermark[WMARK_MIN] = min_pages;
8467 * If it's a lowmem zone, reserve a number of pages
8468 * proportionate to the zone's size.
8470 zone->_watermark[WMARK_MIN] = tmp;
8474 * Set the kswapd watermarks distance according to the
8475 * scale factor in proportion to available memory, but
8476 * ensure a minimum size on small systems.
8478 tmp = max_t(u64, tmp >> 2,
8479 mult_frac(zone_managed_pages(zone),
8480 watermark_scale_factor, 10000));
8482 zone->watermark_boost = 0;
8483 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
8484 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
8486 spin_unlock_irqrestore(&zone->lock, flags);
8489 /* update totalreserve_pages */
8490 calculate_totalreserve_pages();
8494 * setup_per_zone_wmarks - called when min_free_kbytes changes
8495 * or when memory is hot-{added|removed}
8497 * Ensures that the watermark[min,low,high] values for each zone are set
8498 * correctly with respect to min_free_kbytes.
8500 void setup_per_zone_wmarks(void)
8503 static DEFINE_SPINLOCK(lock);
8506 __setup_per_zone_wmarks();
8510 * The watermark size have changed so update the pcpu batch
8511 * and high limits or the limits may be inappropriate.
8514 zone_pcp_update(zone, 0);
8518 * Initialise min_free_kbytes.
8520 * For small machines we want it small (128k min). For large machines
8521 * we want it large (256MB max). But it is not linear, because network
8522 * bandwidth does not increase linearly with machine size. We use
8524 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
8525 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8541 int __meminit init_per_zone_wmark_min(void)
8543 unsigned long lowmem_kbytes;
8544 int new_min_free_kbytes;
8546 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8547 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8549 if (new_min_free_kbytes > user_min_free_kbytes) {
8550 min_free_kbytes = new_min_free_kbytes;
8551 if (min_free_kbytes < 128)
8552 min_free_kbytes = 128;
8553 if (min_free_kbytes > 262144)
8554 min_free_kbytes = 262144;
8556 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8557 new_min_free_kbytes, user_min_free_kbytes);
8559 setup_per_zone_wmarks();
8560 refresh_zone_stat_thresholds();
8561 setup_per_zone_lowmem_reserve();
8564 setup_min_unmapped_ratio();
8565 setup_min_slab_ratio();
8568 khugepaged_min_free_kbytes_update();
8572 postcore_initcall(init_per_zone_wmark_min)
8575 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8576 * that we can call two helper functions whenever min_free_kbytes
8579 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8580 void *buffer, size_t *length, loff_t *ppos)
8584 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8589 user_min_free_kbytes = min_free_kbytes;
8590 setup_per_zone_wmarks();
8595 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8596 void *buffer, size_t *length, loff_t *ppos)
8600 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8605 setup_per_zone_wmarks();
8611 static void setup_min_unmapped_ratio(void)
8616 for_each_online_pgdat(pgdat)
8617 pgdat->min_unmapped_pages = 0;
8620 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8621 sysctl_min_unmapped_ratio) / 100;
8625 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8626 void *buffer, size_t *length, loff_t *ppos)
8630 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8634 setup_min_unmapped_ratio();
8639 static void setup_min_slab_ratio(void)
8644 for_each_online_pgdat(pgdat)
8645 pgdat->min_slab_pages = 0;
8648 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8649 sysctl_min_slab_ratio) / 100;
8652 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8653 void *buffer, size_t *length, loff_t *ppos)
8657 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8661 setup_min_slab_ratio();
8668 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8669 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8670 * whenever sysctl_lowmem_reserve_ratio changes.
8672 * The reserve ratio obviously has absolutely no relation with the
8673 * minimum watermarks. The lowmem reserve ratio can only make sense
8674 * if in function of the boot time zone sizes.
8676 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8677 void *buffer, size_t *length, loff_t *ppos)
8681 proc_dointvec_minmax(table, write, buffer, length, ppos);
8683 for (i = 0; i < MAX_NR_ZONES; i++) {
8684 if (sysctl_lowmem_reserve_ratio[i] < 1)
8685 sysctl_lowmem_reserve_ratio[i] = 0;
8688 setup_per_zone_lowmem_reserve();
8693 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
8694 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8695 * pagelist can have before it gets flushed back to buddy allocator.
8697 int percpu_pagelist_high_fraction_sysctl_handler(struct ctl_table *table,
8698 int write, void *buffer, size_t *length, loff_t *ppos)
8701 int old_percpu_pagelist_high_fraction;
8704 mutex_lock(&pcp_batch_high_lock);
8705 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
8707 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8708 if (!write || ret < 0)
8711 /* Sanity checking to avoid pcp imbalance */
8712 if (percpu_pagelist_high_fraction &&
8713 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
8714 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
8720 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
8723 for_each_populated_zone(zone)
8724 zone_set_pageset_high_and_batch(zone, 0);
8726 mutex_unlock(&pcp_batch_high_lock);
8730 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8732 * Returns the number of pages that arch has reserved but
8733 * is not known to alloc_large_system_hash().
8735 static unsigned long __init arch_reserved_kernel_pages(void)
8742 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8743 * machines. As memory size is increased the scale is also increased but at
8744 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8745 * quadruples the scale is increased by one, which means the size of hash table
8746 * only doubles, instead of quadrupling as well.
8747 * Because 32-bit systems cannot have large physical memory, where this scaling
8748 * makes sense, it is disabled on such platforms.
8750 #if __BITS_PER_LONG > 32
8751 #define ADAPT_SCALE_BASE (64ul << 30)
8752 #define ADAPT_SCALE_SHIFT 2
8753 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8757 * allocate a large system hash table from bootmem
8758 * - it is assumed that the hash table must contain an exact power-of-2
8759 * quantity of entries
8760 * - limit is the number of hash buckets, not the total allocation size
8762 void *__init alloc_large_system_hash(const char *tablename,
8763 unsigned long bucketsize,
8764 unsigned long numentries,
8767 unsigned int *_hash_shift,
8768 unsigned int *_hash_mask,
8769 unsigned long low_limit,
8770 unsigned long high_limit)
8772 unsigned long long max = high_limit;
8773 unsigned long log2qty, size;
8779 /* allow the kernel cmdline to have a say */
8781 /* round applicable memory size up to nearest megabyte */
8782 numentries = nr_kernel_pages;
8783 numentries -= arch_reserved_kernel_pages();
8785 /* It isn't necessary when PAGE_SIZE >= 1MB */
8786 if (PAGE_SHIFT < 20)
8787 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8789 #if __BITS_PER_LONG > 32
8791 unsigned long adapt;
8793 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8794 adapt <<= ADAPT_SCALE_SHIFT)
8799 /* limit to 1 bucket per 2^scale bytes of low memory */
8800 if (scale > PAGE_SHIFT)
8801 numentries >>= (scale - PAGE_SHIFT);
8803 numentries <<= (PAGE_SHIFT - scale);
8805 /* Make sure we've got at least a 0-order allocation.. */
8806 if (unlikely(flags & HASH_SMALL)) {
8807 /* Makes no sense without HASH_EARLY */
8808 WARN_ON(!(flags & HASH_EARLY));
8809 if (!(numentries >> *_hash_shift)) {
8810 numentries = 1UL << *_hash_shift;
8811 BUG_ON(!numentries);
8813 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8814 numentries = PAGE_SIZE / bucketsize;
8816 numentries = roundup_pow_of_two(numentries);
8818 /* limit allocation size to 1/16 total memory by default */
8820 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8821 do_div(max, bucketsize);
8823 max = min(max, 0x80000000ULL);
8825 if (numentries < low_limit)
8826 numentries = low_limit;
8827 if (numentries > max)
8830 log2qty = ilog2(numentries);
8832 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8835 size = bucketsize << log2qty;
8836 if (flags & HASH_EARLY) {
8837 if (flags & HASH_ZERO)
8838 table = memblock_alloc(size, SMP_CACHE_BYTES);
8840 table = memblock_alloc_raw(size,
8842 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8843 table = __vmalloc(size, gfp_flags);
8845 huge = is_vm_area_hugepages(table);
8848 * If bucketsize is not a power-of-two, we may free
8849 * some pages at the end of hash table which
8850 * alloc_pages_exact() automatically does
8852 table = alloc_pages_exact(size, gfp_flags);
8853 kmemleak_alloc(table, size, 1, gfp_flags);
8855 } while (!table && size > PAGE_SIZE && --log2qty);
8858 panic("Failed to allocate %s hash table\n", tablename);
8860 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8861 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8862 virt ? (huge ? "vmalloc hugepage" : "vmalloc") : "linear");
8865 *_hash_shift = log2qty;
8867 *_hash_mask = (1 << log2qty) - 1;
8873 * This function checks whether pageblock includes unmovable pages or not.
8875 * PageLRU check without isolation or lru_lock could race so that
8876 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8877 * check without lock_page also may miss some movable non-lru pages at
8878 * race condition. So you can't expect this function should be exact.
8880 * Returns a page without holding a reference. If the caller wants to
8881 * dereference that page (e.g., dumping), it has to make sure that it
8882 * cannot get removed (e.g., via memory unplug) concurrently.
8885 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8886 int migratetype, int flags)
8888 unsigned long iter = 0;
8889 unsigned long pfn = page_to_pfn(page);
8890 unsigned long offset = pfn % pageblock_nr_pages;
8892 if (is_migrate_cma_page(page)) {
8894 * CMA allocations (alloc_contig_range) really need to mark
8895 * isolate CMA pageblocks even when they are not movable in fact
8896 * so consider them movable here.
8898 if (is_migrate_cma(migratetype))
8904 for (; iter < pageblock_nr_pages - offset; iter++) {
8905 page = pfn_to_page(pfn + iter);
8908 * Both, bootmem allocations and memory holes are marked
8909 * PG_reserved and are unmovable. We can even have unmovable
8910 * allocations inside ZONE_MOVABLE, for example when
8911 * specifying "movablecore".
8913 if (PageReserved(page))
8917 * If the zone is movable and we have ruled out all reserved
8918 * pages then it should be reasonably safe to assume the rest
8921 if (zone_idx(zone) == ZONE_MOVABLE)
8925 * Hugepages are not in LRU lists, but they're movable.
8926 * THPs are on the LRU, but need to be counted as #small pages.
8927 * We need not scan over tail pages because we don't
8928 * handle each tail page individually in migration.
8930 if (PageHuge(page) || PageTransCompound(page)) {
8931 struct page *head = compound_head(page);
8932 unsigned int skip_pages;
8934 if (PageHuge(page)) {
8935 if (!hugepage_migration_supported(page_hstate(head)))
8937 } else if (!PageLRU(head) && !__PageMovable(head)) {
8941 skip_pages = compound_nr(head) - (page - head);
8942 iter += skip_pages - 1;
8947 * We can't use page_count without pin a page
8948 * because another CPU can free compound page.
8949 * This check already skips compound tails of THP
8950 * because their page->_refcount is zero at all time.
8952 if (!page_ref_count(page)) {
8953 if (PageBuddy(page))
8954 iter += (1 << buddy_order(page)) - 1;
8959 * The HWPoisoned page may be not in buddy system, and
8960 * page_count() is not 0.
8962 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8966 * We treat all PageOffline() pages as movable when offlining
8967 * to give drivers a chance to decrement their reference count
8968 * in MEM_GOING_OFFLINE in order to indicate that these pages
8969 * can be offlined as there are no direct references anymore.
8970 * For actually unmovable PageOffline() where the driver does
8971 * not support this, we will fail later when trying to actually
8972 * move these pages that still have a reference count > 0.
8973 * (false negatives in this function only)
8975 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8978 if (__PageMovable(page) || PageLRU(page))
8982 * If there are RECLAIMABLE pages, we need to check
8983 * it. But now, memory offline itself doesn't call
8984 * shrink_node_slabs() and it still to be fixed.
8991 #ifdef CONFIG_CONTIG_ALLOC
8992 static unsigned long pfn_max_align_down(unsigned long pfn)
8994 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8995 pageblock_nr_pages) - 1);
8998 static unsigned long pfn_max_align_up(unsigned long pfn)
9000 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
9001 pageblock_nr_pages));
9004 #if defined(CONFIG_DYNAMIC_DEBUG) || \
9005 (defined(CONFIG_DYNAMIC_DEBUG_CORE) && defined(DYNAMIC_DEBUG_MODULE))
9006 /* Usage: See admin-guide/dynamic-debug-howto.rst */
9007 static void alloc_contig_dump_pages(struct list_head *page_list)
9009 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
9011 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
9015 list_for_each_entry(page, page_list, lru)
9016 dump_page(page, "migration failure");
9020 static inline void alloc_contig_dump_pages(struct list_head *page_list)
9025 /* [start, end) must belong to a single zone. */
9026 static int __alloc_contig_migrate_range(struct compact_control *cc,
9027 unsigned long start, unsigned long end)
9029 /* This function is based on compact_zone() from compaction.c. */
9030 unsigned int nr_reclaimed;
9031 unsigned long pfn = start;
9032 unsigned int tries = 0;
9034 struct migration_target_control mtc = {
9035 .nid = zone_to_nid(cc->zone),
9036 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
9039 lru_cache_disable();
9041 while (pfn < end || !list_empty(&cc->migratepages)) {
9042 if (fatal_signal_pending(current)) {
9043 pr_err("Fatal signal pending task: %s(%d), %#lx\n",
9044 current->comm, current->pid,
9045 current->pending.signal.sig[0]);
9050 if (list_empty(&cc->migratepages)) {
9051 cc->nr_migratepages = 0;
9052 ret = isolate_migratepages_range(cc, pfn, end);
9053 if (ret && ret != -EAGAIN)
9055 pfn = cc->migrate_pfn;
9057 } else if (++tries == 5) {
9062 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
9064 cc->nr_migratepages -= nr_reclaimed;
9066 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
9067 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
9070 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
9071 * to retry again over this error, so do the same here.
9080 alloc_contig_dump_pages(&cc->migratepages);
9081 putback_movable_pages(&cc->migratepages);
9088 * alloc_contig_range() -- tries to allocate given range of pages
9089 * @start: start PFN to allocate
9090 * @end: one-past-the-last PFN to allocate
9091 * @migratetype: migratetype of the underlying pageblocks (either
9092 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
9093 * in range must have the same migratetype and it must
9094 * be either of the two.
9095 * @gfp_mask: GFP mask to use during compaction
9097 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
9098 * aligned. The PFN range must belong to a single zone.
9100 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
9101 * pageblocks in the range. Once isolated, the pageblocks should not
9102 * be modified by others.
9104 * Return: zero on success or negative error code. On success all
9105 * pages which PFN is in [start, end) are allocated for the caller and
9106 * need to be freed with free_contig_range().
9108 int alloc_contig_range(unsigned long start, unsigned long end,
9109 unsigned migratetype, gfp_t gfp_mask)
9111 unsigned long outer_start, outer_end;
9115 struct compact_control cc = {
9116 .nr_migratepages = 0,
9118 .zone = page_zone(pfn_to_page(start)),
9119 .mode = MIGRATE_SYNC,
9120 .ignore_skip_hint = true,
9121 .no_set_skip_hint = true,
9122 .gfp_mask = current_gfp_context(gfp_mask),
9123 .alloc_contig = true,
9125 INIT_LIST_HEAD(&cc.migratepages);
9128 * What we do here is we mark all pageblocks in range as
9129 * MIGRATE_ISOLATE. Because pageblock and max order pages may
9130 * have different sizes, and due to the way page allocator
9131 * work, we align the range to biggest of the two pages so
9132 * that page allocator won't try to merge buddies from
9133 * different pageblocks and change MIGRATE_ISOLATE to some
9134 * other migration type.
9136 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
9137 * migrate the pages from an unaligned range (ie. pages that
9138 * we are interested in). This will put all the pages in
9139 * range back to page allocator as MIGRATE_ISOLATE.
9141 * When this is done, we take the pages in range from page
9142 * allocator removing them from the buddy system. This way
9143 * page allocator will never consider using them.
9145 * This lets us mark the pageblocks back as
9146 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
9147 * aligned range but not in the unaligned, original range are
9148 * put back to page allocator so that buddy can use them.
9151 ret = start_isolate_page_range(pfn_max_align_down(start),
9152 pfn_max_align_up(end), migratetype, 0);
9156 drain_all_pages(cc.zone);
9159 * In case of -EBUSY, we'd like to know which page causes problem.
9160 * So, just fall through. test_pages_isolated() has a tracepoint
9161 * which will report the busy page.
9163 * It is possible that busy pages could become available before
9164 * the call to test_pages_isolated, and the range will actually be
9165 * allocated. So, if we fall through be sure to clear ret so that
9166 * -EBUSY is not accidentally used or returned to caller.
9168 ret = __alloc_contig_migrate_range(&cc, start, end);
9169 if (ret && ret != -EBUSY)
9174 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
9175 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
9176 * more, all pages in [start, end) are free in page allocator.
9177 * What we are going to do is to allocate all pages from
9178 * [start, end) (that is remove them from page allocator).
9180 * The only problem is that pages at the beginning and at the
9181 * end of interesting range may be not aligned with pages that
9182 * page allocator holds, ie. they can be part of higher order
9183 * pages. Because of this, we reserve the bigger range and
9184 * once this is done free the pages we are not interested in.
9186 * We don't have to hold zone->lock here because the pages are
9187 * isolated thus they won't get removed from buddy.
9191 outer_start = start;
9192 while (!PageBuddy(pfn_to_page(outer_start))) {
9193 if (++order >= MAX_ORDER) {
9194 outer_start = start;
9197 outer_start &= ~0UL << order;
9200 if (outer_start != start) {
9201 order = buddy_order(pfn_to_page(outer_start));
9204 * outer_start page could be small order buddy page and
9205 * it doesn't include start page. Adjust outer_start
9206 * in this case to report failed page properly
9207 * on tracepoint in test_pages_isolated()
9209 if (outer_start + (1UL << order) <= start)
9210 outer_start = start;
9213 /* Make sure the range is really isolated. */
9214 if (test_pages_isolated(outer_start, end, 0)) {
9219 /* Grab isolated pages from freelists. */
9220 outer_end = isolate_freepages_range(&cc, outer_start, end);
9226 /* Free head and tail (if any) */
9227 if (start != outer_start)
9228 free_contig_range(outer_start, start - outer_start);
9229 if (end != outer_end)
9230 free_contig_range(end, outer_end - end);
9233 undo_isolate_page_range(pfn_max_align_down(start),
9234 pfn_max_align_up(end), migratetype);
9237 EXPORT_SYMBOL(alloc_contig_range);
9239 static int __alloc_contig_pages(unsigned long start_pfn,
9240 unsigned long nr_pages, gfp_t gfp_mask)
9242 unsigned long end_pfn = start_pfn + nr_pages;
9244 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
9248 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
9249 unsigned long nr_pages)
9251 unsigned long i, end_pfn = start_pfn + nr_pages;
9254 for (i = start_pfn; i < end_pfn; i++) {
9255 page = pfn_to_online_page(i);
9259 if (page_zone(page) != z)
9262 if (PageReserved(page))
9268 static bool zone_spans_last_pfn(const struct zone *zone,
9269 unsigned long start_pfn, unsigned long nr_pages)
9271 unsigned long last_pfn = start_pfn + nr_pages - 1;
9273 return zone_spans_pfn(zone, last_pfn);
9277 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
9278 * @nr_pages: Number of contiguous pages to allocate
9279 * @gfp_mask: GFP mask to limit search and used during compaction
9281 * @nodemask: Mask for other possible nodes
9283 * This routine is a wrapper around alloc_contig_range(). It scans over zones
9284 * on an applicable zonelist to find a contiguous pfn range which can then be
9285 * tried for allocation with alloc_contig_range(). This routine is intended
9286 * for allocation requests which can not be fulfilled with the buddy allocator.
9288 * The allocated memory is always aligned to a page boundary. If nr_pages is a
9289 * power of two then the alignment is guaranteed to be to the given nr_pages
9290 * (e.g. 1GB request would be aligned to 1GB).
9292 * Allocated pages can be freed with free_contig_range() or by manually calling
9293 * __free_page() on each allocated page.
9295 * Return: pointer to contiguous pages on success, or NULL if not successful.
9297 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
9298 int nid, nodemask_t *nodemask)
9300 unsigned long ret, pfn, flags;
9301 struct zonelist *zonelist;
9305 zonelist = node_zonelist(nid, gfp_mask);
9306 for_each_zone_zonelist_nodemask(zone, z, zonelist,
9307 gfp_zone(gfp_mask), nodemask) {
9308 spin_lock_irqsave(&zone->lock, flags);
9310 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
9311 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
9312 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
9314 * We release the zone lock here because
9315 * alloc_contig_range() will also lock the zone
9316 * at some point. If there's an allocation
9317 * spinning on this lock, it may win the race
9318 * and cause alloc_contig_range() to fail...
9320 spin_unlock_irqrestore(&zone->lock, flags);
9321 ret = __alloc_contig_pages(pfn, nr_pages,
9324 return pfn_to_page(pfn);
9325 spin_lock_irqsave(&zone->lock, flags);
9329 spin_unlock_irqrestore(&zone->lock, flags);
9333 #endif /* CONFIG_CONTIG_ALLOC */
9335 void free_contig_range(unsigned long pfn, unsigned long nr_pages)
9337 unsigned long count = 0;
9339 for (; nr_pages--; pfn++) {
9340 struct page *page = pfn_to_page(pfn);
9342 count += page_count(page) != 1;
9345 WARN(count != 0, "%lu pages are still in use!\n", count);
9347 EXPORT_SYMBOL(free_contig_range);
9350 * The zone indicated has a new number of managed_pages; batch sizes and percpu
9351 * page high values need to be recalculated.
9353 void zone_pcp_update(struct zone *zone, int cpu_online)
9355 mutex_lock(&pcp_batch_high_lock);
9356 zone_set_pageset_high_and_batch(zone, cpu_online);
9357 mutex_unlock(&pcp_batch_high_lock);
9361 * Effectively disable pcplists for the zone by setting the high limit to 0
9362 * and draining all cpus. A concurrent page freeing on another CPU that's about
9363 * to put the page on pcplist will either finish before the drain and the page
9364 * will be drained, or observe the new high limit and skip the pcplist.
9366 * Must be paired with a call to zone_pcp_enable().
9368 void zone_pcp_disable(struct zone *zone)
9370 mutex_lock(&pcp_batch_high_lock);
9371 __zone_set_pageset_high_and_batch(zone, 0, 1);
9372 __drain_all_pages(zone, true);
9375 void zone_pcp_enable(struct zone *zone)
9377 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
9378 mutex_unlock(&pcp_batch_high_lock);
9381 void zone_pcp_reset(struct zone *zone)
9384 struct per_cpu_zonestat *pzstats;
9386 if (zone->per_cpu_pageset != &boot_pageset) {
9387 for_each_online_cpu(cpu) {
9388 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
9389 drain_zonestat(zone, pzstats);
9391 free_percpu(zone->per_cpu_pageset);
9392 free_percpu(zone->per_cpu_zonestats);
9393 zone->per_cpu_pageset = &boot_pageset;
9394 zone->per_cpu_zonestats = &boot_zonestats;
9398 #ifdef CONFIG_MEMORY_HOTREMOVE
9400 * All pages in the range must be in a single zone, must not contain holes,
9401 * must span full sections, and must be isolated before calling this function.
9403 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
9405 unsigned long pfn = start_pfn;
9409 unsigned long flags;
9411 offline_mem_sections(pfn, end_pfn);
9412 zone = page_zone(pfn_to_page(pfn));
9413 spin_lock_irqsave(&zone->lock, flags);
9414 while (pfn < end_pfn) {
9415 page = pfn_to_page(pfn);
9417 * The HWPoisoned page may be not in buddy system, and
9418 * page_count() is not 0.
9420 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
9425 * At this point all remaining PageOffline() pages have a
9426 * reference count of 0 and can simply be skipped.
9428 if (PageOffline(page)) {
9429 BUG_ON(page_count(page));
9430 BUG_ON(PageBuddy(page));
9435 BUG_ON(page_count(page));
9436 BUG_ON(!PageBuddy(page));
9437 order = buddy_order(page);
9438 del_page_from_free_list(page, zone, order);
9439 pfn += (1 << order);
9441 spin_unlock_irqrestore(&zone->lock, flags);
9445 bool is_free_buddy_page(struct page *page)
9447 struct zone *zone = page_zone(page);
9448 unsigned long pfn = page_to_pfn(page);
9449 unsigned long flags;
9452 spin_lock_irqsave(&zone->lock, flags);
9453 for (order = 0; order < MAX_ORDER; order++) {
9454 struct page *page_head = page - (pfn & ((1 << order) - 1));
9456 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
9459 spin_unlock_irqrestore(&zone->lock, flags);
9461 return order < MAX_ORDER;
9464 #ifdef CONFIG_MEMORY_FAILURE
9466 * Break down a higher-order page in sub-pages, and keep our target out of
9469 static void break_down_buddy_pages(struct zone *zone, struct page *page,
9470 struct page *target, int low, int high,
9473 unsigned long size = 1 << high;
9474 struct page *current_buddy, *next_page;
9476 while (high > low) {
9480 if (target >= &page[size]) {
9481 next_page = page + size;
9482 current_buddy = page;
9485 current_buddy = page + size;
9488 if (set_page_guard(zone, current_buddy, high, migratetype))
9491 if (current_buddy != target) {
9492 add_to_free_list(current_buddy, zone, high, migratetype);
9493 set_buddy_order(current_buddy, high);
9500 * Take a page that will be marked as poisoned off the buddy allocator.
9502 bool take_page_off_buddy(struct page *page)
9504 struct zone *zone = page_zone(page);
9505 unsigned long pfn = page_to_pfn(page);
9506 unsigned long flags;
9510 spin_lock_irqsave(&zone->lock, flags);
9511 for (order = 0; order < MAX_ORDER; order++) {
9512 struct page *page_head = page - (pfn & ((1 << order) - 1));
9513 int page_order = buddy_order(page_head);
9515 if (PageBuddy(page_head) && page_order >= order) {
9516 unsigned long pfn_head = page_to_pfn(page_head);
9517 int migratetype = get_pfnblock_migratetype(page_head,
9520 del_page_from_free_list(page_head, zone, page_order);
9521 break_down_buddy_pages(zone, page_head, page, 0,
9522 page_order, migratetype);
9523 if (!is_migrate_isolate(migratetype))
9524 __mod_zone_freepage_state(zone, -1, migratetype);
9528 if (page_count(page_head) > 0)
9531 spin_unlock_irqrestore(&zone->lock, flags);
9536 #ifdef CONFIG_ZONE_DMA
9537 bool has_managed_dma(void)
9539 struct pglist_data *pgdat;
9541 for_each_online_pgdat(pgdat) {
9542 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
9544 if (managed_zone(zone))
9549 #endif /* CONFIG_ZONE_DMA */