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>
76 #include <asm/sections.h>
77 #include <asm/tlbflush.h>
78 #include <asm/div64.h>
81 #include "page_reporting.h"
83 /* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
84 typedef int __bitwise fpi_t;
86 /* No special request */
87 #define FPI_NONE ((__force fpi_t)0)
90 * Skip free page reporting notification for the (possibly merged) page.
91 * This does not hinder free page reporting from grabbing the page,
92 * reporting it and marking it "reported" - it only skips notifying
93 * the free page reporting infrastructure about a newly freed page. For
94 * example, used when temporarily pulling a page from a freelist and
95 * putting it back unmodified.
97 #define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
100 * Place the (possibly merged) page to the tail of the freelist. Will ignore
101 * page shuffling (relevant code - e.g., memory onlining - is expected to
102 * shuffle the whole zone).
104 * Note: No code should rely on this flag for correctness - it's purely
105 * to allow for optimizations when handing back either fresh pages
106 * (memory onlining) or untouched pages (page isolation, free page
109 #define FPI_TO_TAIL ((__force fpi_t)BIT(1))
111 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
112 static DEFINE_MUTEX(pcp_batch_high_lock);
113 #define MIN_PERCPU_PAGELIST_FRACTION (8)
115 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
116 DEFINE_PER_CPU(int, numa_node);
117 EXPORT_PER_CPU_SYMBOL(numa_node);
120 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
122 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
124 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
125 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
126 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
127 * defined in <linux/topology.h>.
129 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
130 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
133 /* work_structs for global per-cpu drains */
136 struct work_struct work;
138 static DEFINE_MUTEX(pcpu_drain_mutex);
139 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
141 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
142 volatile unsigned long latent_entropy __latent_entropy;
143 EXPORT_SYMBOL(latent_entropy);
147 * Array of node states.
149 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
150 [N_POSSIBLE] = NODE_MASK_ALL,
151 [N_ONLINE] = { { [0] = 1UL } },
153 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
154 #ifdef CONFIG_HIGHMEM
155 [N_HIGH_MEMORY] = { { [0] = 1UL } },
157 [N_MEMORY] = { { [0] = 1UL } },
158 [N_CPU] = { { [0] = 1UL } },
161 EXPORT_SYMBOL(node_states);
163 atomic_long_t _totalram_pages __read_mostly;
164 EXPORT_SYMBOL(_totalram_pages);
165 unsigned long totalreserve_pages __read_mostly;
166 unsigned long totalcma_pages __read_mostly;
168 int percpu_pagelist_fraction;
169 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
170 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
171 EXPORT_SYMBOL(init_on_alloc);
173 DEFINE_STATIC_KEY_FALSE(init_on_free);
174 EXPORT_SYMBOL(init_on_free);
176 static bool _init_on_alloc_enabled_early __read_mostly
177 = IS_ENABLED(CONFIG_INIT_ON_ALLOC_DEFAULT_ON);
178 static int __init early_init_on_alloc(char *buf)
181 return kstrtobool(buf, &_init_on_alloc_enabled_early);
183 early_param("init_on_alloc", early_init_on_alloc);
185 static bool _init_on_free_enabled_early __read_mostly
186 = IS_ENABLED(CONFIG_INIT_ON_FREE_DEFAULT_ON);
187 static int __init early_init_on_free(char *buf)
189 return kstrtobool(buf, &_init_on_free_enabled_early);
191 early_param("init_on_free", early_init_on_free);
194 * A cached value of the page's pageblock's migratetype, used when the page is
195 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
196 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
197 * Also the migratetype set in the page does not necessarily match the pcplist
198 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
199 * other index - this ensures that it will be put on the correct CMA freelist.
201 static inline int get_pcppage_migratetype(struct page *page)
206 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
208 page->index = migratetype;
211 #ifdef CONFIG_PM_SLEEP
213 * The following functions are used by the suspend/hibernate code to temporarily
214 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
215 * while devices are suspended. To avoid races with the suspend/hibernate code,
216 * they should always be called with system_transition_mutex held
217 * (gfp_allowed_mask also should only be modified with system_transition_mutex
218 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
219 * with that modification).
222 static gfp_t saved_gfp_mask;
224 void pm_restore_gfp_mask(void)
226 WARN_ON(!mutex_is_locked(&system_transition_mutex));
227 if (saved_gfp_mask) {
228 gfp_allowed_mask = saved_gfp_mask;
233 void pm_restrict_gfp_mask(void)
235 WARN_ON(!mutex_is_locked(&system_transition_mutex));
236 WARN_ON(saved_gfp_mask);
237 saved_gfp_mask = gfp_allowed_mask;
238 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
241 bool pm_suspended_storage(void)
243 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
247 #endif /* CONFIG_PM_SLEEP */
249 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
250 unsigned int pageblock_order __read_mostly;
253 static void __free_pages_ok(struct page *page, unsigned int order,
257 * results with 256, 32 in the lowmem_reserve sysctl:
258 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
259 * 1G machine -> (16M dma, 784M normal, 224M high)
260 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
261 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
262 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
264 * TBD: should special case ZONE_DMA32 machines here - in those we normally
265 * don't need any ZONE_NORMAL reservation
267 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
268 #ifdef CONFIG_ZONE_DMA
271 #ifdef CONFIG_ZONE_DMA32
275 #ifdef CONFIG_HIGHMEM
281 static char * const zone_names[MAX_NR_ZONES] = {
282 #ifdef CONFIG_ZONE_DMA
285 #ifdef CONFIG_ZONE_DMA32
289 #ifdef CONFIG_HIGHMEM
293 #ifdef CONFIG_ZONE_DEVICE
298 const char * const migratetype_names[MIGRATE_TYPES] = {
306 #ifdef CONFIG_MEMORY_ISOLATION
311 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
312 [NULL_COMPOUND_DTOR] = NULL,
313 [COMPOUND_PAGE_DTOR] = free_compound_page,
314 #ifdef CONFIG_HUGETLB_PAGE
315 [HUGETLB_PAGE_DTOR] = free_huge_page,
317 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
318 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
322 int min_free_kbytes = 1024;
323 int user_min_free_kbytes = -1;
324 #ifdef CONFIG_DISCONTIGMEM
326 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
327 * are not on separate NUMA nodes. Functionally this works but with
328 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
329 * quite small. By default, do not boost watermarks on discontigmem as in
330 * many cases very high-order allocations like THP are likely to be
331 * unsupported and the premature reclaim offsets the advantage of long-term
332 * fragmentation avoidance.
334 int watermark_boost_factor __read_mostly;
336 int watermark_boost_factor __read_mostly = 15000;
338 int watermark_scale_factor = 10;
340 static unsigned long nr_kernel_pages __initdata;
341 static unsigned long nr_all_pages __initdata;
342 static unsigned long dma_reserve __initdata;
344 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
345 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
346 static unsigned long required_kernelcore __initdata;
347 static unsigned long required_kernelcore_percent __initdata;
348 static unsigned long required_movablecore __initdata;
349 static unsigned long required_movablecore_percent __initdata;
350 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
351 static bool mirrored_kernelcore __meminitdata;
353 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
355 EXPORT_SYMBOL(movable_zone);
358 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
359 unsigned int nr_online_nodes __read_mostly = 1;
360 EXPORT_SYMBOL(nr_node_ids);
361 EXPORT_SYMBOL(nr_online_nodes);
364 int page_group_by_mobility_disabled __read_mostly;
366 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
368 * During boot we initialize deferred pages on-demand, as needed, but once
369 * page_alloc_init_late() has finished, the deferred pages are all initialized,
370 * and we can permanently disable that path.
372 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
375 * Calling kasan_free_pages() only after deferred memory initialization
376 * has completed. Poisoning pages during deferred memory init will greatly
377 * lengthen the process and cause problem in large memory systems as the
378 * deferred pages initialization is done with interrupt disabled.
380 * Assuming that there will be no reference to those newly initialized
381 * pages before they are ever allocated, this should have no effect on
382 * KASAN memory tracking as the poison will be properly inserted at page
383 * allocation time. The only corner case is when pages are allocated by
384 * on-demand allocation and then freed again before the deferred pages
385 * initialization is done, but this is not likely to happen.
387 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
389 if (!static_branch_unlikely(&deferred_pages))
390 kasan_free_pages(page, order);
393 /* Returns true if the struct page for the pfn is uninitialised */
394 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
396 int nid = early_pfn_to_nid(pfn);
398 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
405 * Returns true when the remaining initialisation should be deferred until
406 * later in the boot cycle when it can be parallelised.
408 static bool __meminit
409 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
411 static unsigned long prev_end_pfn, nr_initialised;
414 * prev_end_pfn static that contains the end of previous zone
415 * No need to protect because called very early in boot before smp_init.
417 if (prev_end_pfn != end_pfn) {
418 prev_end_pfn = end_pfn;
422 /* Always populate low zones for address-constrained allocations */
423 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
427 * We start only with one section of pages, more pages are added as
428 * needed until the rest of deferred pages are initialized.
431 if ((nr_initialised > PAGES_PER_SECTION) &&
432 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
433 NODE_DATA(nid)->first_deferred_pfn = pfn;
439 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
441 static inline bool early_page_uninitialised(unsigned long pfn)
446 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
452 /* Return a pointer to the bitmap storing bits affecting a block of pages */
453 static inline unsigned long *get_pageblock_bitmap(struct page *page,
456 #ifdef CONFIG_SPARSEMEM
457 return section_to_usemap(__pfn_to_section(pfn));
459 return page_zone(page)->pageblock_flags;
460 #endif /* CONFIG_SPARSEMEM */
463 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
465 #ifdef CONFIG_SPARSEMEM
466 pfn &= (PAGES_PER_SECTION-1);
468 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
469 #endif /* CONFIG_SPARSEMEM */
470 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
473 static __always_inline
474 unsigned long __get_pfnblock_flags_mask(struct page *page,
478 unsigned long *bitmap;
479 unsigned long bitidx, word_bitidx;
482 bitmap = get_pageblock_bitmap(page, pfn);
483 bitidx = pfn_to_bitidx(page, pfn);
484 word_bitidx = bitidx / BITS_PER_LONG;
485 bitidx &= (BITS_PER_LONG-1);
487 word = bitmap[word_bitidx];
488 return (word >> bitidx) & mask;
492 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
493 * @page: The page within the block of interest
494 * @pfn: The target page frame number
495 * @mask: mask of bits that the caller is interested in
497 * Return: pageblock_bits flags
499 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
502 return __get_pfnblock_flags_mask(page, pfn, mask);
505 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
507 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
511 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
512 * @page: The page within the block of interest
513 * @flags: The flags to set
514 * @pfn: The target page frame number
515 * @mask: mask of bits that the caller is interested in
517 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
521 unsigned long *bitmap;
522 unsigned long bitidx, word_bitidx;
523 unsigned long old_word, word;
525 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
526 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
528 bitmap = get_pageblock_bitmap(page, pfn);
529 bitidx = pfn_to_bitidx(page, pfn);
530 word_bitidx = bitidx / BITS_PER_LONG;
531 bitidx &= (BITS_PER_LONG-1);
533 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
538 word = READ_ONCE(bitmap[word_bitidx]);
540 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
541 if (word == old_word)
547 void set_pageblock_migratetype(struct page *page, int migratetype)
549 if (unlikely(page_group_by_mobility_disabled &&
550 migratetype < MIGRATE_PCPTYPES))
551 migratetype = MIGRATE_UNMOVABLE;
553 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
554 page_to_pfn(page), MIGRATETYPE_MASK);
557 #ifdef CONFIG_DEBUG_VM
558 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
562 unsigned long pfn = page_to_pfn(page);
563 unsigned long sp, start_pfn;
566 seq = zone_span_seqbegin(zone);
567 start_pfn = zone->zone_start_pfn;
568 sp = zone->spanned_pages;
569 if (!zone_spans_pfn(zone, pfn))
571 } while (zone_span_seqretry(zone, seq));
574 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
575 pfn, zone_to_nid(zone), zone->name,
576 start_pfn, start_pfn + sp);
581 static int page_is_consistent(struct zone *zone, struct page *page)
583 if (!pfn_valid_within(page_to_pfn(page)))
585 if (zone != page_zone(page))
591 * Temporary debugging check for pages not lying within a given zone.
593 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
595 if (page_outside_zone_boundaries(zone, page))
597 if (!page_is_consistent(zone, page))
603 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
609 static void bad_page(struct page *page, const char *reason)
611 static unsigned long resume;
612 static unsigned long nr_shown;
613 static unsigned long nr_unshown;
616 * Allow a burst of 60 reports, then keep quiet for that minute;
617 * or allow a steady drip of one report per second.
619 if (nr_shown == 60) {
620 if (time_before(jiffies, resume)) {
626 "BUG: Bad page state: %lu messages suppressed\n",
633 resume = jiffies + 60 * HZ;
635 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
636 current->comm, page_to_pfn(page));
637 __dump_page(page, reason);
638 dump_page_owner(page);
643 /* Leave bad fields for debug, except PageBuddy could make trouble */
644 page_mapcount_reset(page); /* remove PageBuddy */
645 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
649 * Higher-order pages are called "compound pages". They are structured thusly:
651 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
653 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
654 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
656 * The first tail page's ->compound_dtor holds the offset in array of compound
657 * page destructors. See compound_page_dtors.
659 * The first tail page's ->compound_order holds the order of allocation.
660 * This usage means that zero-order pages may not be compound.
663 void free_compound_page(struct page *page)
665 mem_cgroup_uncharge(page);
666 __free_pages_ok(page, compound_order(page), FPI_NONE);
669 void prep_compound_page(struct page *page, unsigned int order)
672 int nr_pages = 1 << order;
675 for (i = 1; i < nr_pages; i++) {
676 struct page *p = page + i;
677 set_page_count(p, 0);
678 p->mapping = TAIL_MAPPING;
679 set_compound_head(p, page);
682 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
683 set_compound_order(page, order);
684 atomic_set(compound_mapcount_ptr(page), -1);
685 if (hpage_pincount_available(page))
686 atomic_set(compound_pincount_ptr(page), 0);
689 #ifdef CONFIG_DEBUG_PAGEALLOC
690 unsigned int _debug_guardpage_minorder;
692 bool _debug_pagealloc_enabled_early __read_mostly
693 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
694 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
695 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
696 EXPORT_SYMBOL(_debug_pagealloc_enabled);
698 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
700 static int __init early_debug_pagealloc(char *buf)
702 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
704 early_param("debug_pagealloc", early_debug_pagealloc);
706 static int __init debug_guardpage_minorder_setup(char *buf)
710 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
711 pr_err("Bad debug_guardpage_minorder value\n");
714 _debug_guardpage_minorder = res;
715 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
718 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
720 static inline bool set_page_guard(struct zone *zone, struct page *page,
721 unsigned int order, int migratetype)
723 if (!debug_guardpage_enabled())
726 if (order >= debug_guardpage_minorder())
729 __SetPageGuard(page);
730 INIT_LIST_HEAD(&page->lru);
731 set_page_private(page, order);
732 /* Guard pages are not available for any usage */
733 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
738 static inline void clear_page_guard(struct zone *zone, struct page *page,
739 unsigned int order, int migratetype)
741 if (!debug_guardpage_enabled())
744 __ClearPageGuard(page);
746 set_page_private(page, 0);
747 if (!is_migrate_isolate(migratetype))
748 __mod_zone_freepage_state(zone, (1 << order), migratetype);
751 static inline bool set_page_guard(struct zone *zone, struct page *page,
752 unsigned int order, int migratetype) { return false; }
753 static inline void clear_page_guard(struct zone *zone, struct page *page,
754 unsigned int order, int migratetype) {}
758 * Enable static keys related to various memory debugging and hardening options.
759 * Some override others, and depend on early params that are evaluated in the
760 * order of appearance. So we need to first gather the full picture of what was
761 * enabled, and then make decisions.
763 void init_mem_debugging_and_hardening(void)
765 if (_init_on_alloc_enabled_early) {
766 if (page_poisoning_enabled())
767 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
768 "will take precedence over init_on_alloc\n");
770 static_branch_enable(&init_on_alloc);
772 if (_init_on_free_enabled_early) {
773 if (page_poisoning_enabled())
774 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
775 "will take precedence over init_on_free\n");
777 static_branch_enable(&init_on_free);
780 #ifdef CONFIG_PAGE_POISONING
782 * Page poisoning is debug page alloc for some arches. If
783 * either of those options are enabled, enable poisoning.
785 if (page_poisoning_enabled() ||
786 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
787 debug_pagealloc_enabled()))
788 static_branch_enable(&_page_poisoning_enabled);
791 #ifdef CONFIG_DEBUG_PAGEALLOC
792 if (!debug_pagealloc_enabled())
795 static_branch_enable(&_debug_pagealloc_enabled);
797 if (!debug_guardpage_minorder())
800 static_branch_enable(&_debug_guardpage_enabled);
804 static inline void set_buddy_order(struct page *page, unsigned int order)
806 set_page_private(page, order);
807 __SetPageBuddy(page);
811 * This function checks whether a page is free && is the buddy
812 * we can coalesce a page and its buddy if
813 * (a) the buddy is not in a hole (check before calling!) &&
814 * (b) the buddy is in the buddy system &&
815 * (c) a page and its buddy have the same order &&
816 * (d) a page and its buddy are in the same zone.
818 * For recording whether a page is in the buddy system, we set PageBuddy.
819 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
821 * For recording page's order, we use page_private(page).
823 static inline bool page_is_buddy(struct page *page, struct page *buddy,
826 if (!page_is_guard(buddy) && !PageBuddy(buddy))
829 if (buddy_order(buddy) != order)
833 * zone check is done late to avoid uselessly calculating
834 * zone/node ids for pages that could never merge.
836 if (page_zone_id(page) != page_zone_id(buddy))
839 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
844 #ifdef CONFIG_COMPACTION
845 static inline struct capture_control *task_capc(struct zone *zone)
847 struct capture_control *capc = current->capture_control;
849 return unlikely(capc) &&
850 !(current->flags & PF_KTHREAD) &&
852 capc->cc->zone == zone ? capc : NULL;
856 compaction_capture(struct capture_control *capc, struct page *page,
857 int order, int migratetype)
859 if (!capc || order != capc->cc->order)
862 /* Do not accidentally pollute CMA or isolated regions*/
863 if (is_migrate_cma(migratetype) ||
864 is_migrate_isolate(migratetype))
868 * Do not let lower order allocations polluate a movable pageblock.
869 * This might let an unmovable request use a reclaimable pageblock
870 * and vice-versa but no more than normal fallback logic which can
871 * have trouble finding a high-order free page.
873 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
881 static inline struct capture_control *task_capc(struct zone *zone)
887 compaction_capture(struct capture_control *capc, struct page *page,
888 int order, int migratetype)
892 #endif /* CONFIG_COMPACTION */
894 /* Used for pages not on another list */
895 static inline void add_to_free_list(struct page *page, struct zone *zone,
896 unsigned int order, int migratetype)
898 struct free_area *area = &zone->free_area[order];
900 list_add(&page->lru, &area->free_list[migratetype]);
904 /* Used for pages not on another list */
905 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
906 unsigned int order, int migratetype)
908 struct free_area *area = &zone->free_area[order];
910 list_add_tail(&page->lru, &area->free_list[migratetype]);
915 * Used for pages which are on another list. Move the pages to the tail
916 * of the list - so the moved pages won't immediately be considered for
917 * allocation again (e.g., optimization for memory onlining).
919 static inline void move_to_free_list(struct page *page, struct zone *zone,
920 unsigned int order, int migratetype)
922 struct free_area *area = &zone->free_area[order];
924 list_move_tail(&page->lru, &area->free_list[migratetype]);
927 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
930 /* clear reported state and update reported page count */
931 if (page_reported(page))
932 __ClearPageReported(page);
934 list_del(&page->lru);
935 __ClearPageBuddy(page);
936 set_page_private(page, 0);
937 zone->free_area[order].nr_free--;
941 * If this is not the largest possible page, check if the buddy
942 * of the next-highest order is free. If it is, it's possible
943 * that pages are being freed that will coalesce soon. In case,
944 * that is happening, add the free page to the tail of the list
945 * so it's less likely to be used soon and more likely to be merged
946 * as a higher order page
949 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
950 struct page *page, unsigned int order)
952 struct page *higher_page, *higher_buddy;
953 unsigned long combined_pfn;
955 if (order >= MAX_ORDER - 2)
958 if (!pfn_valid_within(buddy_pfn))
961 combined_pfn = buddy_pfn & pfn;
962 higher_page = page + (combined_pfn - pfn);
963 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
964 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
966 return pfn_valid_within(buddy_pfn) &&
967 page_is_buddy(higher_page, higher_buddy, order + 1);
971 * Freeing function for a buddy system allocator.
973 * The concept of a buddy system is to maintain direct-mapped table
974 * (containing bit values) for memory blocks of various "orders".
975 * The bottom level table contains the map for the smallest allocatable
976 * units of memory (here, pages), and each level above it describes
977 * pairs of units from the levels below, hence, "buddies".
978 * At a high level, all that happens here is marking the table entry
979 * at the bottom level available, and propagating the changes upward
980 * as necessary, plus some accounting needed to play nicely with other
981 * parts of the VM system.
982 * At each level, we keep a list of pages, which are heads of continuous
983 * free pages of length of (1 << order) and marked with PageBuddy.
984 * Page's order is recorded in page_private(page) field.
985 * So when we are allocating or freeing one, we can derive the state of the
986 * other. That is, if we allocate a small block, and both were
987 * free, the remainder of the region must be split into blocks.
988 * If a block is freed, and its buddy is also free, then this
989 * triggers coalescing into a block of larger size.
994 static inline void __free_one_page(struct page *page,
996 struct zone *zone, unsigned int order,
997 int migratetype, fpi_t fpi_flags)
999 struct capture_control *capc = task_capc(zone);
1000 unsigned long buddy_pfn;
1001 unsigned long combined_pfn;
1002 unsigned int max_order;
1006 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1008 VM_BUG_ON(!zone_is_initialized(zone));
1009 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1011 VM_BUG_ON(migratetype == -1);
1012 if (likely(!is_migrate_isolate(migratetype)))
1013 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1015 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1016 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1019 while (order < max_order) {
1020 if (compaction_capture(capc, page, order, migratetype)) {
1021 __mod_zone_freepage_state(zone, -(1 << order),
1025 buddy_pfn = __find_buddy_pfn(pfn, order);
1026 buddy = page + (buddy_pfn - pfn);
1028 if (!pfn_valid_within(buddy_pfn))
1030 if (!page_is_buddy(page, buddy, order))
1033 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1034 * merge with it and move up one order.
1036 if (page_is_guard(buddy))
1037 clear_page_guard(zone, buddy, order, migratetype);
1039 del_page_from_free_list(buddy, zone, order);
1040 combined_pfn = buddy_pfn & pfn;
1041 page = page + (combined_pfn - pfn);
1045 if (order < MAX_ORDER - 1) {
1046 /* If we are here, it means order is >= pageblock_order.
1047 * We want to prevent merge between freepages on isolate
1048 * pageblock and normal pageblock. Without this, pageblock
1049 * isolation could cause incorrect freepage or CMA accounting.
1051 * We don't want to hit this code for the more frequent
1052 * low-order merging.
1054 if (unlikely(has_isolate_pageblock(zone))) {
1057 buddy_pfn = __find_buddy_pfn(pfn, order);
1058 buddy = page + (buddy_pfn - pfn);
1059 buddy_mt = get_pageblock_migratetype(buddy);
1061 if (migratetype != buddy_mt
1062 && (is_migrate_isolate(migratetype) ||
1063 is_migrate_isolate(buddy_mt)))
1066 max_order = order + 1;
1067 goto continue_merging;
1071 set_buddy_order(page, order);
1073 if (fpi_flags & FPI_TO_TAIL)
1075 else if (is_shuffle_order(order))
1076 to_tail = shuffle_pick_tail();
1078 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1081 add_to_free_list_tail(page, zone, order, migratetype);
1083 add_to_free_list(page, zone, order, migratetype);
1085 /* Notify page reporting subsystem of freed page */
1086 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1087 page_reporting_notify_free(order);
1091 * A bad page could be due to a number of fields. Instead of multiple branches,
1092 * try and check multiple fields with one check. The caller must do a detailed
1093 * check if necessary.
1095 static inline bool page_expected_state(struct page *page,
1096 unsigned long check_flags)
1098 if (unlikely(atomic_read(&page->_mapcount) != -1))
1101 if (unlikely((unsigned long)page->mapping |
1102 page_ref_count(page) |
1104 (unsigned long)page_memcg(page) |
1106 (page->flags & check_flags)))
1112 static const char *page_bad_reason(struct page *page, unsigned long flags)
1114 const char *bad_reason = NULL;
1116 if (unlikely(atomic_read(&page->_mapcount) != -1))
1117 bad_reason = "nonzero mapcount";
1118 if (unlikely(page->mapping != NULL))
1119 bad_reason = "non-NULL mapping";
1120 if (unlikely(page_ref_count(page) != 0))
1121 bad_reason = "nonzero _refcount";
1122 if (unlikely(page->flags & flags)) {
1123 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1124 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1126 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1129 if (unlikely(page_memcg(page)))
1130 bad_reason = "page still charged to cgroup";
1135 static void check_free_page_bad(struct page *page)
1138 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1141 static inline int check_free_page(struct page *page)
1143 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1146 /* Something has gone sideways, find it */
1147 check_free_page_bad(page);
1151 static int free_tail_pages_check(struct page *head_page, struct page *page)
1156 * We rely page->lru.next never has bit 0 set, unless the page
1157 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1159 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1161 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1165 switch (page - head_page) {
1167 /* the first tail page: ->mapping may be compound_mapcount() */
1168 if (unlikely(compound_mapcount(page))) {
1169 bad_page(page, "nonzero compound_mapcount");
1175 * the second tail page: ->mapping is
1176 * deferred_list.next -- ignore value.
1180 if (page->mapping != TAIL_MAPPING) {
1181 bad_page(page, "corrupted mapping in tail page");
1186 if (unlikely(!PageTail(page))) {
1187 bad_page(page, "PageTail not set");
1190 if (unlikely(compound_head(page) != head_page)) {
1191 bad_page(page, "compound_head not consistent");
1196 page->mapping = NULL;
1197 clear_compound_head(page);
1201 static void kernel_init_free_pages(struct page *page, int numpages)
1205 /* s390's use of memset() could override KASAN redzones. */
1206 kasan_disable_current();
1207 for (i = 0; i < numpages; i++) {
1208 page_kasan_tag_reset(page + i);
1209 clear_highpage(page + i);
1211 kasan_enable_current();
1214 static __always_inline bool free_pages_prepare(struct page *page,
1215 unsigned int order, bool check_free)
1219 VM_BUG_ON_PAGE(PageTail(page), page);
1221 trace_mm_page_free(page, order);
1223 if (unlikely(PageHWPoison(page)) && !order) {
1225 * Do not let hwpoison pages hit pcplists/buddy
1226 * Untie memcg state and reset page's owner
1228 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1229 __memcg_kmem_uncharge_page(page, order);
1230 reset_page_owner(page, order);
1235 * Check tail pages before head page information is cleared to
1236 * avoid checking PageCompound for order-0 pages.
1238 if (unlikely(order)) {
1239 bool compound = PageCompound(page);
1242 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1245 ClearPageDoubleMap(page);
1246 for (i = 1; i < (1 << order); i++) {
1248 bad += free_tail_pages_check(page, page + i);
1249 if (unlikely(check_free_page(page + i))) {
1253 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1256 if (PageMappingFlags(page))
1257 page->mapping = NULL;
1258 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1259 __memcg_kmem_uncharge_page(page, order);
1261 bad += check_free_page(page);
1265 page_cpupid_reset_last(page);
1266 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1267 reset_page_owner(page, order);
1269 if (!PageHighMem(page)) {
1270 debug_check_no_locks_freed(page_address(page),
1271 PAGE_SIZE << order);
1272 debug_check_no_obj_freed(page_address(page),
1273 PAGE_SIZE << order);
1275 if (want_init_on_free())
1276 kernel_init_free_pages(page, 1 << order);
1278 kernel_poison_pages(page, 1 << order);
1281 * arch_free_page() can make the page's contents inaccessible. s390
1282 * does this. So nothing which can access the page's contents should
1283 * happen after this.
1285 arch_free_page(page, order);
1287 debug_pagealloc_unmap_pages(page, 1 << order);
1289 kasan_free_nondeferred_pages(page, order);
1294 #ifdef CONFIG_DEBUG_VM
1296 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1297 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1298 * moved from pcp lists to free lists.
1300 static bool free_pcp_prepare(struct page *page)
1302 return free_pages_prepare(page, 0, true);
1305 static bool bulkfree_pcp_prepare(struct page *page)
1307 if (debug_pagealloc_enabled_static())
1308 return check_free_page(page);
1314 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1315 * moving from pcp lists to free list in order to reduce overhead. With
1316 * debug_pagealloc enabled, they are checked also immediately when being freed
1319 static bool free_pcp_prepare(struct page *page)
1321 if (debug_pagealloc_enabled_static())
1322 return free_pages_prepare(page, 0, true);
1324 return free_pages_prepare(page, 0, false);
1327 static bool bulkfree_pcp_prepare(struct page *page)
1329 return check_free_page(page);
1331 #endif /* CONFIG_DEBUG_VM */
1333 static inline void prefetch_buddy(struct page *page)
1335 unsigned long pfn = page_to_pfn(page);
1336 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1337 struct page *buddy = page + (buddy_pfn - pfn);
1343 * Frees a number of pages from the PCP lists
1344 * Assumes all pages on list are in same zone, and of same order.
1345 * count is the number of pages to free.
1347 * If the zone was previously in an "all pages pinned" state then look to
1348 * see if this freeing clears that state.
1350 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1351 * pinned" detection logic.
1353 static void free_pcppages_bulk(struct zone *zone, int count,
1354 struct per_cpu_pages *pcp)
1356 int migratetype = 0;
1358 int prefetch_nr = READ_ONCE(pcp->batch);
1359 bool isolated_pageblocks;
1360 struct page *page, *tmp;
1364 * Ensure proper count is passed which otherwise would stuck in the
1365 * below while (list_empty(list)) loop.
1367 count = min(pcp->count, count);
1369 struct list_head *list;
1372 * Remove pages from lists in a round-robin fashion. A
1373 * batch_free count is maintained that is incremented when an
1374 * empty list is encountered. This is so more pages are freed
1375 * off fuller lists instead of spinning excessively around empty
1380 if (++migratetype == MIGRATE_PCPTYPES)
1382 list = &pcp->lists[migratetype];
1383 } while (list_empty(list));
1385 /* This is the only non-empty list. Free them all. */
1386 if (batch_free == MIGRATE_PCPTYPES)
1390 page = list_last_entry(list, struct page, lru);
1391 /* must delete to avoid corrupting pcp list */
1392 list_del(&page->lru);
1395 if (bulkfree_pcp_prepare(page))
1398 list_add_tail(&page->lru, &head);
1401 * We are going to put the page back to the global
1402 * pool, prefetch its buddy to speed up later access
1403 * under zone->lock. It is believed the overhead of
1404 * an additional test and calculating buddy_pfn here
1405 * can be offset by reduced memory latency later. To
1406 * avoid excessive prefetching due to large count, only
1407 * prefetch buddy for the first pcp->batch nr of pages.
1410 prefetch_buddy(page);
1413 } while (--count && --batch_free && !list_empty(list));
1416 spin_lock(&zone->lock);
1417 isolated_pageblocks = has_isolate_pageblock(zone);
1420 * Use safe version since after __free_one_page(),
1421 * page->lru.next will not point to original list.
1423 list_for_each_entry_safe(page, tmp, &head, lru) {
1424 int mt = get_pcppage_migratetype(page);
1425 /* MIGRATE_ISOLATE page should not go to pcplists */
1426 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1427 /* Pageblock could have been isolated meanwhile */
1428 if (unlikely(isolated_pageblocks))
1429 mt = get_pageblock_migratetype(page);
1431 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1432 trace_mm_page_pcpu_drain(page, 0, mt);
1434 spin_unlock(&zone->lock);
1437 static void free_one_page(struct zone *zone,
1438 struct page *page, unsigned long pfn,
1440 int migratetype, fpi_t fpi_flags)
1442 spin_lock(&zone->lock);
1443 if (unlikely(has_isolate_pageblock(zone) ||
1444 is_migrate_isolate(migratetype))) {
1445 migratetype = get_pfnblock_migratetype(page, pfn);
1447 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1448 spin_unlock(&zone->lock);
1451 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1452 unsigned long zone, int nid)
1454 mm_zero_struct_page(page);
1455 set_page_links(page, zone, nid, pfn);
1456 init_page_count(page);
1457 page_mapcount_reset(page);
1458 page_cpupid_reset_last(page);
1459 page_kasan_tag_reset(page);
1461 INIT_LIST_HEAD(&page->lru);
1462 #ifdef WANT_PAGE_VIRTUAL
1463 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1464 if (!is_highmem_idx(zone))
1465 set_page_address(page, __va(pfn << PAGE_SHIFT));
1469 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1470 static void __meminit init_reserved_page(unsigned long pfn)
1475 if (!early_page_uninitialised(pfn))
1478 nid = early_pfn_to_nid(pfn);
1479 pgdat = NODE_DATA(nid);
1481 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1482 struct zone *zone = &pgdat->node_zones[zid];
1484 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1487 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1490 static inline void init_reserved_page(unsigned long pfn)
1493 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1496 * Initialised pages do not have PageReserved set. This function is
1497 * called for each range allocated by the bootmem allocator and
1498 * marks the pages PageReserved. The remaining valid pages are later
1499 * sent to the buddy page allocator.
1501 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1503 unsigned long start_pfn = PFN_DOWN(start);
1504 unsigned long end_pfn = PFN_UP(end);
1506 for (; start_pfn < end_pfn; start_pfn++) {
1507 if (pfn_valid(start_pfn)) {
1508 struct page *page = pfn_to_page(start_pfn);
1510 init_reserved_page(start_pfn);
1512 /* Avoid false-positive PageTail() */
1513 INIT_LIST_HEAD(&page->lru);
1516 * no need for atomic set_bit because the struct
1517 * page is not visible yet so nobody should
1520 __SetPageReserved(page);
1525 static void __free_pages_ok(struct page *page, unsigned int order,
1528 unsigned long flags;
1530 unsigned long pfn = page_to_pfn(page);
1532 if (!free_pages_prepare(page, order, true))
1535 migratetype = get_pfnblock_migratetype(page, pfn);
1536 local_irq_save(flags);
1537 __count_vm_events(PGFREE, 1 << order);
1538 free_one_page(page_zone(page), page, pfn, order, migratetype,
1540 local_irq_restore(flags);
1543 void __free_pages_core(struct page *page, unsigned int order)
1545 unsigned int nr_pages = 1 << order;
1546 struct page *p = page;
1550 * When initializing the memmap, __init_single_page() sets the refcount
1551 * of all pages to 1 ("allocated"/"not free"). We have to set the
1552 * refcount of all involved pages to 0.
1555 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1557 __ClearPageReserved(p);
1558 set_page_count(p, 0);
1560 __ClearPageReserved(p);
1561 set_page_count(p, 0);
1563 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1566 * Bypass PCP and place fresh pages right to the tail, primarily
1567 * relevant for memory onlining.
1569 __free_pages_ok(page, order, FPI_TO_TAIL);
1572 #ifdef CONFIG_NEED_MULTIPLE_NODES
1575 * During memory init memblocks map pfns to nids. The search is expensive and
1576 * this caches recent lookups. The implementation of __early_pfn_to_nid
1577 * treats start/end as pfns.
1579 struct mminit_pfnnid_cache {
1580 unsigned long last_start;
1581 unsigned long last_end;
1585 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1588 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1590 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1591 struct mminit_pfnnid_cache *state)
1593 unsigned long start_pfn, end_pfn;
1596 if (state->last_start <= pfn && pfn < state->last_end)
1597 return state->last_nid;
1599 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1600 if (nid != NUMA_NO_NODE) {
1601 state->last_start = start_pfn;
1602 state->last_end = end_pfn;
1603 state->last_nid = nid;
1609 int __meminit early_pfn_to_nid(unsigned long pfn)
1611 static DEFINE_SPINLOCK(early_pfn_lock);
1614 spin_lock(&early_pfn_lock);
1615 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1617 nid = first_online_node;
1618 spin_unlock(&early_pfn_lock);
1622 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1624 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1627 if (early_page_uninitialised(pfn))
1629 __free_pages_core(page, order);
1633 * Check that the whole (or subset of) a pageblock given by the interval of
1634 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1635 * with the migration of free compaction scanner. The scanners then need to
1636 * use only pfn_valid_within() check for arches that allow holes within
1639 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1641 * It's possible on some configurations to have a setup like node0 node1 node0
1642 * i.e. it's possible that all pages within a zones range of pages do not
1643 * belong to a single zone. We assume that a border between node0 and node1
1644 * can occur within a single pageblock, but not a node0 node1 node0
1645 * interleaving within a single pageblock. It is therefore sufficient to check
1646 * the first and last page of a pageblock and avoid checking each individual
1647 * page in a pageblock.
1649 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1650 unsigned long end_pfn, struct zone *zone)
1652 struct page *start_page;
1653 struct page *end_page;
1655 /* end_pfn is one past the range we are checking */
1658 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1661 start_page = pfn_to_online_page(start_pfn);
1665 if (page_zone(start_page) != zone)
1668 end_page = pfn_to_page(end_pfn);
1670 /* This gives a shorter code than deriving page_zone(end_page) */
1671 if (page_zone_id(start_page) != page_zone_id(end_page))
1677 void set_zone_contiguous(struct zone *zone)
1679 unsigned long block_start_pfn = zone->zone_start_pfn;
1680 unsigned long block_end_pfn;
1682 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1683 for (; block_start_pfn < zone_end_pfn(zone);
1684 block_start_pfn = block_end_pfn,
1685 block_end_pfn += pageblock_nr_pages) {
1687 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1689 if (!__pageblock_pfn_to_page(block_start_pfn,
1690 block_end_pfn, zone))
1695 /* We confirm that there is no hole */
1696 zone->contiguous = true;
1699 void clear_zone_contiguous(struct zone *zone)
1701 zone->contiguous = false;
1704 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1705 static void __init deferred_free_range(unsigned long pfn,
1706 unsigned long nr_pages)
1714 page = pfn_to_page(pfn);
1716 /* Free a large naturally-aligned chunk if possible */
1717 if (nr_pages == pageblock_nr_pages &&
1718 (pfn & (pageblock_nr_pages - 1)) == 0) {
1719 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1720 __free_pages_core(page, pageblock_order);
1724 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1725 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1726 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1727 __free_pages_core(page, 0);
1731 /* Completion tracking for deferred_init_memmap() threads */
1732 static atomic_t pgdat_init_n_undone __initdata;
1733 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1735 static inline void __init pgdat_init_report_one_done(void)
1737 if (atomic_dec_and_test(&pgdat_init_n_undone))
1738 complete(&pgdat_init_all_done_comp);
1742 * Returns true if page needs to be initialized or freed to buddy allocator.
1744 * First we check if pfn is valid on architectures where it is possible to have
1745 * holes within pageblock_nr_pages. On systems where it is not possible, this
1746 * function is optimized out.
1748 * Then, we check if a current large page is valid by only checking the validity
1751 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1753 if (!pfn_valid_within(pfn))
1755 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1761 * Free pages to buddy allocator. Try to free aligned pages in
1762 * pageblock_nr_pages sizes.
1764 static void __init deferred_free_pages(unsigned long pfn,
1765 unsigned long end_pfn)
1767 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1768 unsigned long nr_free = 0;
1770 for (; pfn < end_pfn; pfn++) {
1771 if (!deferred_pfn_valid(pfn)) {
1772 deferred_free_range(pfn - nr_free, nr_free);
1774 } else if (!(pfn & nr_pgmask)) {
1775 deferred_free_range(pfn - nr_free, nr_free);
1781 /* Free the last block of pages to allocator */
1782 deferred_free_range(pfn - nr_free, nr_free);
1786 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1787 * by performing it only once every pageblock_nr_pages.
1788 * Return number of pages initialized.
1790 static unsigned long __init deferred_init_pages(struct zone *zone,
1792 unsigned long end_pfn)
1794 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1795 int nid = zone_to_nid(zone);
1796 unsigned long nr_pages = 0;
1797 int zid = zone_idx(zone);
1798 struct page *page = NULL;
1800 for (; pfn < end_pfn; pfn++) {
1801 if (!deferred_pfn_valid(pfn)) {
1804 } else if (!page || !(pfn & nr_pgmask)) {
1805 page = pfn_to_page(pfn);
1809 __init_single_page(page, pfn, zid, nid);
1816 * This function is meant to pre-load the iterator for the zone init.
1817 * Specifically it walks through the ranges until we are caught up to the
1818 * first_init_pfn value and exits there. If we never encounter the value we
1819 * return false indicating there are no valid ranges left.
1822 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1823 unsigned long *spfn, unsigned long *epfn,
1824 unsigned long first_init_pfn)
1829 * Start out by walking through the ranges in this zone that have
1830 * already been initialized. We don't need to do anything with them
1831 * so we just need to flush them out of the system.
1833 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1834 if (*epfn <= first_init_pfn)
1836 if (*spfn < first_init_pfn)
1837 *spfn = first_init_pfn;
1846 * Initialize and free pages. We do it in two loops: first we initialize
1847 * struct page, then free to buddy allocator, because while we are
1848 * freeing pages we can access pages that are ahead (computing buddy
1849 * page in __free_one_page()).
1851 * In order to try and keep some memory in the cache we have the loop
1852 * broken along max page order boundaries. This way we will not cause
1853 * any issues with the buddy page computation.
1855 static unsigned long __init
1856 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1857 unsigned long *end_pfn)
1859 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1860 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1861 unsigned long nr_pages = 0;
1864 /* First we loop through and initialize the page values */
1865 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1868 if (mo_pfn <= *start_pfn)
1871 t = min(mo_pfn, *end_pfn);
1872 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1874 if (mo_pfn < *end_pfn) {
1875 *start_pfn = mo_pfn;
1880 /* Reset values and now loop through freeing pages as needed */
1883 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1889 t = min(mo_pfn, epfn);
1890 deferred_free_pages(spfn, t);
1900 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1903 unsigned long spfn, epfn;
1904 struct zone *zone = arg;
1907 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1910 * Initialize and free pages in MAX_ORDER sized increments so that we
1911 * can avoid introducing any issues with the buddy allocator.
1913 while (spfn < end_pfn) {
1914 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1919 /* An arch may override for more concurrency. */
1921 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1926 /* Initialise remaining memory on a node */
1927 static int __init deferred_init_memmap(void *data)
1929 pg_data_t *pgdat = data;
1930 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1931 unsigned long spfn = 0, epfn = 0;
1932 unsigned long first_init_pfn, flags;
1933 unsigned long start = jiffies;
1935 int zid, max_threads;
1938 /* Bind memory initialisation thread to a local node if possible */
1939 if (!cpumask_empty(cpumask))
1940 set_cpus_allowed_ptr(current, cpumask);
1942 pgdat_resize_lock(pgdat, &flags);
1943 first_init_pfn = pgdat->first_deferred_pfn;
1944 if (first_init_pfn == ULONG_MAX) {
1945 pgdat_resize_unlock(pgdat, &flags);
1946 pgdat_init_report_one_done();
1950 /* Sanity check boundaries */
1951 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1952 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1953 pgdat->first_deferred_pfn = ULONG_MAX;
1956 * Once we unlock here, the zone cannot be grown anymore, thus if an
1957 * interrupt thread must allocate this early in boot, zone must be
1958 * pre-grown prior to start of deferred page initialization.
1960 pgdat_resize_unlock(pgdat, &flags);
1962 /* Only the highest zone is deferred so find it */
1963 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1964 zone = pgdat->node_zones + zid;
1965 if (first_init_pfn < zone_end_pfn(zone))
1969 /* If the zone is empty somebody else may have cleared out the zone */
1970 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1974 max_threads = deferred_page_init_max_threads(cpumask);
1976 while (spfn < epfn) {
1977 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1978 struct padata_mt_job job = {
1979 .thread_fn = deferred_init_memmap_chunk,
1982 .size = epfn_align - spfn,
1983 .align = PAGES_PER_SECTION,
1984 .min_chunk = PAGES_PER_SECTION,
1985 .max_threads = max_threads,
1988 padata_do_multithreaded(&job);
1989 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1993 /* Sanity check that the next zone really is unpopulated */
1994 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1996 pr_info("node %d deferred pages initialised in %ums\n",
1997 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1999 pgdat_init_report_one_done();
2004 * If this zone has deferred pages, try to grow it by initializing enough
2005 * deferred pages to satisfy the allocation specified by order, rounded up to
2006 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2007 * of SECTION_SIZE bytes by initializing struct pages in increments of
2008 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2010 * Return true when zone was grown, otherwise return false. We return true even
2011 * when we grow less than requested, to let the caller decide if there are
2012 * enough pages to satisfy the allocation.
2014 * Note: We use noinline because this function is needed only during boot, and
2015 * it is called from a __ref function _deferred_grow_zone. This way we are
2016 * making sure that it is not inlined into permanent text section.
2018 static noinline bool __init
2019 deferred_grow_zone(struct zone *zone, unsigned int order)
2021 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2022 pg_data_t *pgdat = zone->zone_pgdat;
2023 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2024 unsigned long spfn, epfn, flags;
2025 unsigned long nr_pages = 0;
2028 /* Only the last zone may have deferred pages */
2029 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2032 pgdat_resize_lock(pgdat, &flags);
2035 * If someone grew this zone while we were waiting for spinlock, return
2036 * true, as there might be enough pages already.
2038 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2039 pgdat_resize_unlock(pgdat, &flags);
2043 /* If the zone is empty somebody else may have cleared out the zone */
2044 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2045 first_deferred_pfn)) {
2046 pgdat->first_deferred_pfn = ULONG_MAX;
2047 pgdat_resize_unlock(pgdat, &flags);
2048 /* Retry only once. */
2049 return first_deferred_pfn != ULONG_MAX;
2053 * Initialize and free pages in MAX_ORDER sized increments so
2054 * that we can avoid introducing any issues with the buddy
2057 while (spfn < epfn) {
2058 /* update our first deferred PFN for this section */
2059 first_deferred_pfn = spfn;
2061 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2062 touch_nmi_watchdog();
2064 /* We should only stop along section boundaries */
2065 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2068 /* If our quota has been met we can stop here */
2069 if (nr_pages >= nr_pages_needed)
2073 pgdat->first_deferred_pfn = spfn;
2074 pgdat_resize_unlock(pgdat, &flags);
2076 return nr_pages > 0;
2080 * deferred_grow_zone() is __init, but it is called from
2081 * get_page_from_freelist() during early boot until deferred_pages permanently
2082 * disables this call. This is why we have refdata wrapper to avoid warning,
2083 * and to ensure that the function body gets unloaded.
2086 _deferred_grow_zone(struct zone *zone, unsigned int order)
2088 return deferred_grow_zone(zone, order);
2091 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2093 void __init page_alloc_init_late(void)
2098 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2100 /* There will be num_node_state(N_MEMORY) threads */
2101 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2102 for_each_node_state(nid, N_MEMORY) {
2103 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2106 /* Block until all are initialised */
2107 wait_for_completion(&pgdat_init_all_done_comp);
2110 * The number of managed pages has changed due to the initialisation
2111 * so the pcpu batch and high limits needs to be updated or the limits
2112 * will be artificially small.
2114 for_each_populated_zone(zone)
2115 zone_pcp_update(zone);
2118 * We initialized the rest of the deferred pages. Permanently disable
2119 * on-demand struct page initialization.
2121 static_branch_disable(&deferred_pages);
2123 /* Reinit limits that are based on free pages after the kernel is up */
2124 files_maxfiles_init();
2129 /* Discard memblock private memory */
2132 for_each_node_state(nid, N_MEMORY)
2133 shuffle_free_memory(NODE_DATA(nid));
2135 for_each_populated_zone(zone)
2136 set_zone_contiguous(zone);
2140 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2141 void __init init_cma_reserved_pageblock(struct page *page)
2143 unsigned i = pageblock_nr_pages;
2144 struct page *p = page;
2147 __ClearPageReserved(p);
2148 set_page_count(p, 0);
2151 set_pageblock_migratetype(page, MIGRATE_CMA);
2153 if (pageblock_order >= MAX_ORDER) {
2154 i = pageblock_nr_pages;
2157 set_page_refcounted(p);
2158 __free_pages(p, MAX_ORDER - 1);
2159 p += MAX_ORDER_NR_PAGES;
2160 } while (i -= MAX_ORDER_NR_PAGES);
2162 set_page_refcounted(page);
2163 __free_pages(page, pageblock_order);
2166 adjust_managed_page_count(page, pageblock_nr_pages);
2171 * The order of subdivision here is critical for the IO subsystem.
2172 * Please do not alter this order without good reasons and regression
2173 * testing. Specifically, as large blocks of memory are subdivided,
2174 * the order in which smaller blocks are delivered depends on the order
2175 * they're subdivided in this function. This is the primary factor
2176 * influencing the order in which pages are delivered to the IO
2177 * subsystem according to empirical testing, and this is also justified
2178 * by considering the behavior of a buddy system containing a single
2179 * large block of memory acted on by a series of small allocations.
2180 * This behavior is a critical factor in sglist merging's success.
2184 static inline void expand(struct zone *zone, struct page *page,
2185 int low, int high, int migratetype)
2187 unsigned long size = 1 << high;
2189 while (high > low) {
2192 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2195 * Mark as guard pages (or page), that will allow to
2196 * merge back to allocator when buddy will be freed.
2197 * Corresponding page table entries will not be touched,
2198 * pages will stay not present in virtual address space
2200 if (set_page_guard(zone, &page[size], high, migratetype))
2203 add_to_free_list(&page[size], zone, high, migratetype);
2204 set_buddy_order(&page[size], high);
2208 static void check_new_page_bad(struct page *page)
2210 if (unlikely(page->flags & __PG_HWPOISON)) {
2211 /* Don't complain about hwpoisoned pages */
2212 page_mapcount_reset(page); /* remove PageBuddy */
2217 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2221 * This page is about to be returned from the page allocator
2223 static inline int check_new_page(struct page *page)
2225 if (likely(page_expected_state(page,
2226 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2229 check_new_page_bad(page);
2233 #ifdef CONFIG_DEBUG_VM
2235 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2236 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2237 * also checked when pcp lists are refilled from the free lists.
2239 static inline bool check_pcp_refill(struct page *page)
2241 if (debug_pagealloc_enabled_static())
2242 return check_new_page(page);
2247 static inline bool check_new_pcp(struct page *page)
2249 return check_new_page(page);
2253 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2254 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2255 * enabled, they are also checked when being allocated from the pcp lists.
2257 static inline bool check_pcp_refill(struct page *page)
2259 return check_new_page(page);
2261 static inline bool check_new_pcp(struct page *page)
2263 if (debug_pagealloc_enabled_static())
2264 return check_new_page(page);
2268 #endif /* CONFIG_DEBUG_VM */
2270 static bool check_new_pages(struct page *page, unsigned int order)
2273 for (i = 0; i < (1 << order); i++) {
2274 struct page *p = page + i;
2276 if (unlikely(check_new_page(p)))
2283 inline void post_alloc_hook(struct page *page, unsigned int order,
2286 set_page_private(page, 0);
2287 set_page_refcounted(page);
2289 arch_alloc_page(page, order);
2290 debug_pagealloc_map_pages(page, 1 << order);
2291 kasan_alloc_pages(page, order);
2292 kernel_unpoison_pages(page, 1 << order);
2293 set_page_owner(page, order, gfp_flags);
2295 if (!want_init_on_free() && want_init_on_alloc(gfp_flags))
2296 kernel_init_free_pages(page, 1 << order);
2299 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2300 unsigned int alloc_flags)
2302 post_alloc_hook(page, order, gfp_flags);
2304 if (order && (gfp_flags & __GFP_COMP))
2305 prep_compound_page(page, order);
2308 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2309 * allocate the page. The expectation is that the caller is taking
2310 * steps that will free more memory. The caller should avoid the page
2311 * being used for !PFMEMALLOC purposes.
2313 if (alloc_flags & ALLOC_NO_WATERMARKS)
2314 set_page_pfmemalloc(page);
2316 clear_page_pfmemalloc(page);
2320 * Go through the free lists for the given migratetype and remove
2321 * the smallest available page from the freelists
2323 static __always_inline
2324 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2327 unsigned int current_order;
2328 struct free_area *area;
2331 /* Find a page of the appropriate size in the preferred list */
2332 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2333 area = &(zone->free_area[current_order]);
2334 page = get_page_from_free_area(area, migratetype);
2337 del_page_from_free_list(page, zone, current_order);
2338 expand(zone, page, order, current_order, migratetype);
2339 set_pcppage_migratetype(page, migratetype);
2348 * This array describes the order lists are fallen back to when
2349 * the free lists for the desirable migrate type are depleted
2351 static int fallbacks[MIGRATE_TYPES][3] = {
2352 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2353 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2354 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2356 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2358 #ifdef CONFIG_MEMORY_ISOLATION
2359 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2364 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2367 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2370 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2371 unsigned int order) { return NULL; }
2375 * Move the free pages in a range to the freelist tail of the requested type.
2376 * Note that start_page and end_pages are not aligned on a pageblock
2377 * boundary. If alignment is required, use move_freepages_block()
2379 static int move_freepages(struct zone *zone,
2380 struct page *start_page, struct page *end_page,
2381 int migratetype, int *num_movable)
2385 int pages_moved = 0;
2387 for (page = start_page; page <= end_page;) {
2388 if (!pfn_valid_within(page_to_pfn(page))) {
2393 if (!PageBuddy(page)) {
2395 * We assume that pages that could be isolated for
2396 * migration are movable. But we don't actually try
2397 * isolating, as that would be expensive.
2400 (PageLRU(page) || __PageMovable(page)))
2407 /* Make sure we are not inadvertently changing nodes */
2408 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2409 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2411 order = buddy_order(page);
2412 move_to_free_list(page, zone, order, migratetype);
2414 pages_moved += 1 << order;
2420 int move_freepages_block(struct zone *zone, struct page *page,
2421 int migratetype, int *num_movable)
2423 unsigned long start_pfn, end_pfn;
2424 struct page *start_page, *end_page;
2429 start_pfn = page_to_pfn(page);
2430 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2431 start_page = pfn_to_page(start_pfn);
2432 end_page = start_page + pageblock_nr_pages - 1;
2433 end_pfn = start_pfn + pageblock_nr_pages - 1;
2435 /* Do not cross zone boundaries */
2436 if (!zone_spans_pfn(zone, start_pfn))
2438 if (!zone_spans_pfn(zone, end_pfn))
2441 return move_freepages(zone, start_page, end_page, migratetype,
2445 static void change_pageblock_range(struct page *pageblock_page,
2446 int start_order, int migratetype)
2448 int nr_pageblocks = 1 << (start_order - pageblock_order);
2450 while (nr_pageblocks--) {
2451 set_pageblock_migratetype(pageblock_page, migratetype);
2452 pageblock_page += pageblock_nr_pages;
2457 * When we are falling back to another migratetype during allocation, try to
2458 * steal extra free pages from the same pageblocks to satisfy further
2459 * allocations, instead of polluting multiple pageblocks.
2461 * If we are stealing a relatively large buddy page, it is likely there will
2462 * be more free pages in the pageblock, so try to steal them all. For
2463 * reclaimable and unmovable allocations, we steal regardless of page size,
2464 * as fragmentation caused by those allocations polluting movable pageblocks
2465 * is worse than movable allocations stealing from unmovable and reclaimable
2468 static bool can_steal_fallback(unsigned int order, int start_mt)
2471 * Leaving this order check is intended, although there is
2472 * relaxed order check in next check. The reason is that
2473 * we can actually steal whole pageblock if this condition met,
2474 * but, below check doesn't guarantee it and that is just heuristic
2475 * so could be changed anytime.
2477 if (order >= pageblock_order)
2480 if (order >= pageblock_order / 2 ||
2481 start_mt == MIGRATE_RECLAIMABLE ||
2482 start_mt == MIGRATE_UNMOVABLE ||
2483 page_group_by_mobility_disabled)
2489 static inline bool boost_watermark(struct zone *zone)
2491 unsigned long max_boost;
2493 if (!watermark_boost_factor)
2496 * Don't bother in zones that are unlikely to produce results.
2497 * On small machines, including kdump capture kernels running
2498 * in a small area, boosting the watermark can cause an out of
2499 * memory situation immediately.
2501 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2504 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2505 watermark_boost_factor, 10000);
2508 * high watermark may be uninitialised if fragmentation occurs
2509 * very early in boot so do not boost. We do not fall
2510 * through and boost by pageblock_nr_pages as failing
2511 * allocations that early means that reclaim is not going
2512 * to help and it may even be impossible to reclaim the
2513 * boosted watermark resulting in a hang.
2518 max_boost = max(pageblock_nr_pages, max_boost);
2520 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2527 * This function implements actual steal behaviour. If order is large enough,
2528 * we can steal whole pageblock. If not, we first move freepages in this
2529 * pageblock to our migratetype and determine how many already-allocated pages
2530 * are there in the pageblock with a compatible migratetype. If at least half
2531 * of pages are free or compatible, we can change migratetype of the pageblock
2532 * itself, so pages freed in the future will be put on the correct free list.
2534 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2535 unsigned int alloc_flags, int start_type, bool whole_block)
2537 unsigned int current_order = buddy_order(page);
2538 int free_pages, movable_pages, alike_pages;
2541 old_block_type = get_pageblock_migratetype(page);
2544 * This can happen due to races and we want to prevent broken
2545 * highatomic accounting.
2547 if (is_migrate_highatomic(old_block_type))
2550 /* Take ownership for orders >= pageblock_order */
2551 if (current_order >= pageblock_order) {
2552 change_pageblock_range(page, current_order, start_type);
2557 * Boost watermarks to increase reclaim pressure to reduce the
2558 * likelihood of future fallbacks. Wake kswapd now as the node
2559 * may be balanced overall and kswapd will not wake naturally.
2561 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2562 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2564 /* We are not allowed to try stealing from the whole block */
2568 free_pages = move_freepages_block(zone, page, start_type,
2571 * Determine how many pages are compatible with our allocation.
2572 * For movable allocation, it's the number of movable pages which
2573 * we just obtained. For other types it's a bit more tricky.
2575 if (start_type == MIGRATE_MOVABLE) {
2576 alike_pages = movable_pages;
2579 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2580 * to MOVABLE pageblock, consider all non-movable pages as
2581 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2582 * vice versa, be conservative since we can't distinguish the
2583 * exact migratetype of non-movable pages.
2585 if (old_block_type == MIGRATE_MOVABLE)
2586 alike_pages = pageblock_nr_pages
2587 - (free_pages + movable_pages);
2592 /* moving whole block can fail due to zone boundary conditions */
2597 * If a sufficient number of pages in the block are either free or of
2598 * comparable migratability as our allocation, claim the whole block.
2600 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2601 page_group_by_mobility_disabled)
2602 set_pageblock_migratetype(page, start_type);
2607 move_to_free_list(page, zone, current_order, start_type);
2611 * Check whether there is a suitable fallback freepage with requested order.
2612 * If only_stealable is true, this function returns fallback_mt only if
2613 * we can steal other freepages all together. This would help to reduce
2614 * fragmentation due to mixed migratetype pages in one pageblock.
2616 int find_suitable_fallback(struct free_area *area, unsigned int order,
2617 int migratetype, bool only_stealable, bool *can_steal)
2622 if (area->nr_free == 0)
2627 fallback_mt = fallbacks[migratetype][i];
2628 if (fallback_mt == MIGRATE_TYPES)
2631 if (free_area_empty(area, fallback_mt))
2634 if (can_steal_fallback(order, migratetype))
2637 if (!only_stealable)
2648 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2649 * there are no empty page blocks that contain a page with a suitable order
2651 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2652 unsigned int alloc_order)
2655 unsigned long max_managed, flags;
2658 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2659 * Check is race-prone but harmless.
2661 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2662 if (zone->nr_reserved_highatomic >= max_managed)
2665 spin_lock_irqsave(&zone->lock, flags);
2667 /* Recheck the nr_reserved_highatomic limit under the lock */
2668 if (zone->nr_reserved_highatomic >= max_managed)
2672 mt = get_pageblock_migratetype(page);
2673 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2674 && !is_migrate_cma(mt)) {
2675 zone->nr_reserved_highatomic += pageblock_nr_pages;
2676 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2677 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2681 spin_unlock_irqrestore(&zone->lock, flags);
2685 * Used when an allocation is about to fail under memory pressure. This
2686 * potentially hurts the reliability of high-order allocations when under
2687 * intense memory pressure but failed atomic allocations should be easier
2688 * to recover from than an OOM.
2690 * If @force is true, try to unreserve a pageblock even though highatomic
2691 * pageblock is exhausted.
2693 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2696 struct zonelist *zonelist = ac->zonelist;
2697 unsigned long flags;
2704 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2707 * Preserve at least one pageblock unless memory pressure
2710 if (!force && zone->nr_reserved_highatomic <=
2714 spin_lock_irqsave(&zone->lock, flags);
2715 for (order = 0; order < MAX_ORDER; order++) {
2716 struct free_area *area = &(zone->free_area[order]);
2718 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2723 * In page freeing path, migratetype change is racy so
2724 * we can counter several free pages in a pageblock
2725 * in this loop althoug we changed the pageblock type
2726 * from highatomic to ac->migratetype. So we should
2727 * adjust the count once.
2729 if (is_migrate_highatomic_page(page)) {
2731 * It should never happen but changes to
2732 * locking could inadvertently allow a per-cpu
2733 * drain to add pages to MIGRATE_HIGHATOMIC
2734 * while unreserving so be safe and watch for
2737 zone->nr_reserved_highatomic -= min(
2739 zone->nr_reserved_highatomic);
2743 * Convert to ac->migratetype and avoid the normal
2744 * pageblock stealing heuristics. Minimally, the caller
2745 * is doing the work and needs the pages. More
2746 * importantly, if the block was always converted to
2747 * MIGRATE_UNMOVABLE or another type then the number
2748 * of pageblocks that cannot be completely freed
2751 set_pageblock_migratetype(page, ac->migratetype);
2752 ret = move_freepages_block(zone, page, ac->migratetype,
2755 spin_unlock_irqrestore(&zone->lock, flags);
2759 spin_unlock_irqrestore(&zone->lock, flags);
2766 * Try finding a free buddy page on the fallback list and put it on the free
2767 * list of requested migratetype, possibly along with other pages from the same
2768 * block, depending on fragmentation avoidance heuristics. Returns true if
2769 * fallback was found so that __rmqueue_smallest() can grab it.
2771 * The use of signed ints for order and current_order is a deliberate
2772 * deviation from the rest of this file, to make the for loop
2773 * condition simpler.
2775 static __always_inline bool
2776 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2777 unsigned int alloc_flags)
2779 struct free_area *area;
2781 int min_order = order;
2787 * Do not steal pages from freelists belonging to other pageblocks
2788 * i.e. orders < pageblock_order. If there are no local zones free,
2789 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2791 if (alloc_flags & ALLOC_NOFRAGMENT)
2792 min_order = pageblock_order;
2795 * Find the largest available free page in the other list. This roughly
2796 * approximates finding the pageblock with the most free pages, which
2797 * would be too costly to do exactly.
2799 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2801 area = &(zone->free_area[current_order]);
2802 fallback_mt = find_suitable_fallback(area, current_order,
2803 start_migratetype, false, &can_steal);
2804 if (fallback_mt == -1)
2808 * We cannot steal all free pages from the pageblock and the
2809 * requested migratetype is movable. In that case it's better to
2810 * steal and split the smallest available page instead of the
2811 * largest available page, because even if the next movable
2812 * allocation falls back into a different pageblock than this
2813 * one, it won't cause permanent fragmentation.
2815 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2816 && current_order > order)
2825 for (current_order = order; current_order < MAX_ORDER;
2827 area = &(zone->free_area[current_order]);
2828 fallback_mt = find_suitable_fallback(area, current_order,
2829 start_migratetype, false, &can_steal);
2830 if (fallback_mt != -1)
2835 * This should not happen - we already found a suitable fallback
2836 * when looking for the largest page.
2838 VM_BUG_ON(current_order == MAX_ORDER);
2841 page = get_page_from_free_area(area, fallback_mt);
2843 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2846 trace_mm_page_alloc_extfrag(page, order, current_order,
2847 start_migratetype, fallback_mt);
2854 * Do the hard work of removing an element from the buddy allocator.
2855 * Call me with the zone->lock already held.
2857 static __always_inline struct page *
2858 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2859 unsigned int alloc_flags)
2865 * Balance movable allocations between regular and CMA areas by
2866 * allocating from CMA when over half of the zone's free memory
2867 * is in the CMA area.
2869 if (alloc_flags & ALLOC_CMA &&
2870 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2871 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2872 page = __rmqueue_cma_fallback(zone, order);
2878 page = __rmqueue_smallest(zone, order, migratetype);
2879 if (unlikely(!page)) {
2880 if (alloc_flags & ALLOC_CMA)
2881 page = __rmqueue_cma_fallback(zone, order);
2883 if (!page && __rmqueue_fallback(zone, order, migratetype,
2888 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2893 * Obtain a specified number of elements from the buddy allocator, all under
2894 * a single hold of the lock, for efficiency. Add them to the supplied list.
2895 * Returns the number of new pages which were placed at *list.
2897 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2898 unsigned long count, struct list_head *list,
2899 int migratetype, unsigned int alloc_flags)
2903 spin_lock(&zone->lock);
2904 for (i = 0; i < count; ++i) {
2905 struct page *page = __rmqueue(zone, order, migratetype,
2907 if (unlikely(page == NULL))
2910 if (unlikely(check_pcp_refill(page)))
2914 * Split buddy pages returned by expand() are received here in
2915 * physical page order. The page is added to the tail of
2916 * caller's list. From the callers perspective, the linked list
2917 * is ordered by page number under some conditions. This is
2918 * useful for IO devices that can forward direction from the
2919 * head, thus also in the physical page order. This is useful
2920 * for IO devices that can merge IO requests if the physical
2921 * pages are ordered properly.
2923 list_add_tail(&page->lru, list);
2925 if (is_migrate_cma(get_pcppage_migratetype(page)))
2926 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2931 * i pages were removed from the buddy list even if some leak due
2932 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2933 * on i. Do not confuse with 'alloced' which is the number of
2934 * pages added to the pcp list.
2936 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2937 spin_unlock(&zone->lock);
2943 * Called from the vmstat counter updater to drain pagesets of this
2944 * currently executing processor on remote nodes after they have
2947 * Note that this function must be called with the thread pinned to
2948 * a single processor.
2950 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2952 unsigned long flags;
2953 int to_drain, batch;
2955 local_irq_save(flags);
2956 batch = READ_ONCE(pcp->batch);
2957 to_drain = min(pcp->count, batch);
2959 free_pcppages_bulk(zone, to_drain, pcp);
2960 local_irq_restore(flags);
2965 * Drain pcplists of the indicated processor and zone.
2967 * The processor must either be the current processor and the
2968 * thread pinned to the current processor or a processor that
2971 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2973 unsigned long flags;
2974 struct per_cpu_pageset *pset;
2975 struct per_cpu_pages *pcp;
2977 local_irq_save(flags);
2978 pset = per_cpu_ptr(zone->pageset, cpu);
2982 free_pcppages_bulk(zone, pcp->count, pcp);
2983 local_irq_restore(flags);
2987 * Drain pcplists of all zones on the indicated processor.
2989 * The processor must either be the current processor and the
2990 * thread pinned to the current processor or a processor that
2993 static void drain_pages(unsigned int cpu)
2997 for_each_populated_zone(zone) {
2998 drain_pages_zone(cpu, zone);
3003 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3005 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3006 * the single zone's pages.
3008 void drain_local_pages(struct zone *zone)
3010 int cpu = smp_processor_id();
3013 drain_pages_zone(cpu, zone);
3018 static void drain_local_pages_wq(struct work_struct *work)
3020 struct pcpu_drain *drain;
3022 drain = container_of(work, struct pcpu_drain, work);
3025 * drain_all_pages doesn't use proper cpu hotplug protection so
3026 * we can race with cpu offline when the WQ can move this from
3027 * a cpu pinned worker to an unbound one. We can operate on a different
3028 * cpu which is allright but we also have to make sure to not move to
3032 drain_local_pages(drain->zone);
3037 * The implementation of drain_all_pages(), exposing an extra parameter to
3038 * drain on all cpus.
3040 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3041 * not empty. The check for non-emptiness can however race with a free to
3042 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3043 * that need the guarantee that every CPU has drained can disable the
3044 * optimizing racy check.
3046 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3051 * Allocate in the BSS so we wont require allocation in
3052 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3054 static cpumask_t cpus_with_pcps;
3057 * Make sure nobody triggers this path before mm_percpu_wq is fully
3060 if (WARN_ON_ONCE(!mm_percpu_wq))
3064 * Do not drain if one is already in progress unless it's specific to
3065 * a zone. Such callers are primarily CMA and memory hotplug and need
3066 * the drain to be complete when the call returns.
3068 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3071 mutex_lock(&pcpu_drain_mutex);
3075 * We don't care about racing with CPU hotplug event
3076 * as offline notification will cause the notified
3077 * cpu to drain that CPU pcps and on_each_cpu_mask
3078 * disables preemption as part of its processing
3080 for_each_online_cpu(cpu) {
3081 struct per_cpu_pageset *pcp;
3083 bool has_pcps = false;
3085 if (force_all_cpus) {
3087 * The pcp.count check is racy, some callers need a
3088 * guarantee that no cpu is missed.
3092 pcp = per_cpu_ptr(zone->pageset, cpu);
3096 for_each_populated_zone(z) {
3097 pcp = per_cpu_ptr(z->pageset, cpu);
3098 if (pcp->pcp.count) {
3106 cpumask_set_cpu(cpu, &cpus_with_pcps);
3108 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3111 for_each_cpu(cpu, &cpus_with_pcps) {
3112 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3115 INIT_WORK(&drain->work, drain_local_pages_wq);
3116 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3118 for_each_cpu(cpu, &cpus_with_pcps)
3119 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3121 mutex_unlock(&pcpu_drain_mutex);
3125 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3127 * When zone parameter is non-NULL, spill just the single zone's pages.
3129 * Note that this can be extremely slow as the draining happens in a workqueue.
3131 void drain_all_pages(struct zone *zone)
3133 __drain_all_pages(zone, false);
3136 #ifdef CONFIG_HIBERNATION
3139 * Touch the watchdog for every WD_PAGE_COUNT pages.
3141 #define WD_PAGE_COUNT (128*1024)
3143 void mark_free_pages(struct zone *zone)
3145 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3146 unsigned long flags;
3147 unsigned int order, t;
3150 if (zone_is_empty(zone))
3153 spin_lock_irqsave(&zone->lock, flags);
3155 max_zone_pfn = zone_end_pfn(zone);
3156 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3157 if (pfn_valid(pfn)) {
3158 page = pfn_to_page(pfn);
3160 if (!--page_count) {
3161 touch_nmi_watchdog();
3162 page_count = WD_PAGE_COUNT;
3165 if (page_zone(page) != zone)
3168 if (!swsusp_page_is_forbidden(page))
3169 swsusp_unset_page_free(page);
3172 for_each_migratetype_order(order, t) {
3173 list_for_each_entry(page,
3174 &zone->free_area[order].free_list[t], lru) {
3177 pfn = page_to_pfn(page);
3178 for (i = 0; i < (1UL << order); i++) {
3179 if (!--page_count) {
3180 touch_nmi_watchdog();
3181 page_count = WD_PAGE_COUNT;
3183 swsusp_set_page_free(pfn_to_page(pfn + i));
3187 spin_unlock_irqrestore(&zone->lock, flags);
3189 #endif /* CONFIG_PM */
3191 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3195 if (!free_pcp_prepare(page))
3198 migratetype = get_pfnblock_migratetype(page, pfn);
3199 set_pcppage_migratetype(page, migratetype);
3203 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3205 struct zone *zone = page_zone(page);
3206 struct per_cpu_pages *pcp;
3209 migratetype = get_pcppage_migratetype(page);
3210 __count_vm_event(PGFREE);
3213 * We only track unmovable, reclaimable and movable on pcp lists.
3214 * Free ISOLATE pages back to the allocator because they are being
3215 * offlined but treat HIGHATOMIC as movable pages so we can get those
3216 * areas back if necessary. Otherwise, we may have to free
3217 * excessively into the page allocator
3219 if (migratetype >= MIGRATE_PCPTYPES) {
3220 if (unlikely(is_migrate_isolate(migratetype))) {
3221 free_one_page(zone, page, pfn, 0, migratetype,
3225 migratetype = MIGRATE_MOVABLE;
3228 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3229 list_add(&page->lru, &pcp->lists[migratetype]);
3231 if (pcp->count >= READ_ONCE(pcp->high))
3232 free_pcppages_bulk(zone, READ_ONCE(pcp->batch), pcp);
3236 * Free a 0-order page
3238 void free_unref_page(struct page *page)
3240 unsigned long flags;
3241 unsigned long pfn = page_to_pfn(page);
3243 if (!free_unref_page_prepare(page, pfn))
3246 local_irq_save(flags);
3247 free_unref_page_commit(page, pfn);
3248 local_irq_restore(flags);
3252 * Free a list of 0-order pages
3254 void free_unref_page_list(struct list_head *list)
3256 struct page *page, *next;
3257 unsigned long flags, pfn;
3258 int batch_count = 0;
3260 /* Prepare pages for freeing */
3261 list_for_each_entry_safe(page, next, list, lru) {
3262 pfn = page_to_pfn(page);
3263 if (!free_unref_page_prepare(page, pfn))
3264 list_del(&page->lru);
3265 set_page_private(page, pfn);
3268 local_irq_save(flags);
3269 list_for_each_entry_safe(page, next, list, lru) {
3270 unsigned long pfn = page_private(page);
3272 set_page_private(page, 0);
3273 trace_mm_page_free_batched(page);
3274 free_unref_page_commit(page, pfn);
3277 * Guard against excessive IRQ disabled times when we get
3278 * a large list of pages to free.
3280 if (++batch_count == SWAP_CLUSTER_MAX) {
3281 local_irq_restore(flags);
3283 local_irq_save(flags);
3286 local_irq_restore(flags);
3290 * split_page takes a non-compound higher-order page, and splits it into
3291 * n (1<<order) sub-pages: page[0..n]
3292 * Each sub-page must be freed individually.
3294 * Note: this is probably too low level an operation for use in drivers.
3295 * Please consult with lkml before using this in your driver.
3297 void split_page(struct page *page, unsigned int order)
3301 VM_BUG_ON_PAGE(PageCompound(page), page);
3302 VM_BUG_ON_PAGE(!page_count(page), page);
3304 for (i = 1; i < (1 << order); i++)
3305 set_page_refcounted(page + i);
3306 split_page_owner(page, 1 << order);
3308 EXPORT_SYMBOL_GPL(split_page);
3310 int __isolate_free_page(struct page *page, unsigned int order)
3312 unsigned long watermark;
3316 BUG_ON(!PageBuddy(page));
3318 zone = page_zone(page);
3319 mt = get_pageblock_migratetype(page);
3321 if (!is_migrate_isolate(mt)) {
3323 * Obey watermarks as if the page was being allocated. We can
3324 * emulate a high-order watermark check with a raised order-0
3325 * watermark, because we already know our high-order page
3328 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3329 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3332 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3335 /* Remove page from free list */
3337 del_page_from_free_list(page, zone, order);
3340 * Set the pageblock if the isolated page is at least half of a
3343 if (order >= pageblock_order - 1) {
3344 struct page *endpage = page + (1 << order) - 1;
3345 for (; page < endpage; page += pageblock_nr_pages) {
3346 int mt = get_pageblock_migratetype(page);
3347 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3348 && !is_migrate_highatomic(mt))
3349 set_pageblock_migratetype(page,
3355 return 1UL << order;
3359 * __putback_isolated_page - Return a now-isolated page back where we got it
3360 * @page: Page that was isolated
3361 * @order: Order of the isolated page
3362 * @mt: The page's pageblock's migratetype
3364 * This function is meant to return a page pulled from the free lists via
3365 * __isolate_free_page back to the free lists they were pulled from.
3367 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3369 struct zone *zone = page_zone(page);
3371 /* zone lock should be held when this function is called */
3372 lockdep_assert_held(&zone->lock);
3374 /* Return isolated page to tail of freelist. */
3375 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3376 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3380 * Update NUMA hit/miss statistics
3382 * Must be called with interrupts disabled.
3384 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3387 enum numa_stat_item local_stat = NUMA_LOCAL;
3389 /* skip numa counters update if numa stats is disabled */
3390 if (!static_branch_likely(&vm_numa_stat_key))
3393 if (zone_to_nid(z) != numa_node_id())
3394 local_stat = NUMA_OTHER;
3396 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3397 __inc_numa_state(z, NUMA_HIT);
3399 __inc_numa_state(z, NUMA_MISS);
3400 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3402 __inc_numa_state(z, local_stat);
3406 /* Remove page from the per-cpu list, caller must protect the list */
3407 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3408 unsigned int alloc_flags,
3409 struct per_cpu_pages *pcp,
3410 struct list_head *list)
3415 if (list_empty(list)) {
3416 pcp->count += rmqueue_bulk(zone, 0,
3417 READ_ONCE(pcp->batch), list,
3418 migratetype, alloc_flags);
3419 if (unlikely(list_empty(list)))
3423 page = list_first_entry(list, struct page, lru);
3424 list_del(&page->lru);
3426 } while (check_new_pcp(page));
3431 /* Lock and remove page from the per-cpu list */
3432 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3433 struct zone *zone, gfp_t gfp_flags,
3434 int migratetype, unsigned int alloc_flags)
3436 struct per_cpu_pages *pcp;
3437 struct list_head *list;
3439 unsigned long flags;
3441 local_irq_save(flags);
3442 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3443 list = &pcp->lists[migratetype];
3444 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3446 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3447 zone_statistics(preferred_zone, zone);
3449 local_irq_restore(flags);
3454 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3457 struct page *rmqueue(struct zone *preferred_zone,
3458 struct zone *zone, unsigned int order,
3459 gfp_t gfp_flags, unsigned int alloc_flags,
3462 unsigned long flags;
3465 if (likely(order == 0)) {
3467 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3468 * we need to skip it when CMA area isn't allowed.
3470 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3471 migratetype != MIGRATE_MOVABLE) {
3472 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3473 migratetype, alloc_flags);
3479 * We most definitely don't want callers attempting to
3480 * allocate greater than order-1 page units with __GFP_NOFAIL.
3482 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3483 spin_lock_irqsave(&zone->lock, flags);
3488 * order-0 request can reach here when the pcplist is skipped
3489 * due to non-CMA allocation context. HIGHATOMIC area is
3490 * reserved for high-order atomic allocation, so order-0
3491 * request should skip it.
3493 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3494 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3496 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3499 page = __rmqueue(zone, order, migratetype, alloc_flags);
3500 } while (page && check_new_pages(page, order));
3501 spin_unlock(&zone->lock);
3504 __mod_zone_freepage_state(zone, -(1 << order),
3505 get_pcppage_migratetype(page));
3507 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3508 zone_statistics(preferred_zone, zone);
3509 local_irq_restore(flags);
3512 /* Separate test+clear to avoid unnecessary atomics */
3513 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3514 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3515 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3518 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3522 local_irq_restore(flags);
3526 #ifdef CONFIG_FAIL_PAGE_ALLOC
3529 struct fault_attr attr;
3531 bool ignore_gfp_highmem;
3532 bool ignore_gfp_reclaim;
3534 } fail_page_alloc = {
3535 .attr = FAULT_ATTR_INITIALIZER,
3536 .ignore_gfp_reclaim = true,
3537 .ignore_gfp_highmem = true,
3541 static int __init setup_fail_page_alloc(char *str)
3543 return setup_fault_attr(&fail_page_alloc.attr, str);
3545 __setup("fail_page_alloc=", setup_fail_page_alloc);
3547 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3549 if (order < fail_page_alloc.min_order)
3551 if (gfp_mask & __GFP_NOFAIL)
3553 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3555 if (fail_page_alloc.ignore_gfp_reclaim &&
3556 (gfp_mask & __GFP_DIRECT_RECLAIM))
3559 return should_fail(&fail_page_alloc.attr, 1 << order);
3562 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3564 static int __init fail_page_alloc_debugfs(void)
3566 umode_t mode = S_IFREG | 0600;
3569 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3570 &fail_page_alloc.attr);
3572 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3573 &fail_page_alloc.ignore_gfp_reclaim);
3574 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3575 &fail_page_alloc.ignore_gfp_highmem);
3576 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3581 late_initcall(fail_page_alloc_debugfs);
3583 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3585 #else /* CONFIG_FAIL_PAGE_ALLOC */
3587 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3592 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3594 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3596 return __should_fail_alloc_page(gfp_mask, order);
3598 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3600 static inline long __zone_watermark_unusable_free(struct zone *z,
3601 unsigned int order, unsigned int alloc_flags)
3603 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3604 long unusable_free = (1 << order) - 1;
3607 * If the caller does not have rights to ALLOC_HARDER then subtract
3608 * the high-atomic reserves. This will over-estimate the size of the
3609 * atomic reserve but it avoids a search.
3611 if (likely(!alloc_harder))
3612 unusable_free += z->nr_reserved_highatomic;
3615 /* If allocation can't use CMA areas don't use free CMA pages */
3616 if (!(alloc_flags & ALLOC_CMA))
3617 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3620 return unusable_free;
3624 * Return true if free base pages are above 'mark'. For high-order checks it
3625 * will return true of the order-0 watermark is reached and there is at least
3626 * one free page of a suitable size. Checking now avoids taking the zone lock
3627 * to check in the allocation paths if no pages are free.
3629 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3630 int highest_zoneidx, unsigned int alloc_flags,
3635 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3637 /* free_pages may go negative - that's OK */
3638 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3640 if (alloc_flags & ALLOC_HIGH)
3643 if (unlikely(alloc_harder)) {
3645 * OOM victims can try even harder than normal ALLOC_HARDER
3646 * users on the grounds that it's definitely going to be in
3647 * the exit path shortly and free memory. Any allocation it
3648 * makes during the free path will be small and short-lived.
3650 if (alloc_flags & ALLOC_OOM)
3657 * Check watermarks for an order-0 allocation request. If these
3658 * are not met, then a high-order request also cannot go ahead
3659 * even if a suitable page happened to be free.
3661 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3664 /* If this is an order-0 request then the watermark is fine */
3668 /* For a high-order request, check at least one suitable page is free */
3669 for (o = order; o < MAX_ORDER; o++) {
3670 struct free_area *area = &z->free_area[o];
3676 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3677 if (!free_area_empty(area, mt))
3682 if ((alloc_flags & ALLOC_CMA) &&
3683 !free_area_empty(area, MIGRATE_CMA)) {
3687 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3693 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3694 int highest_zoneidx, unsigned int alloc_flags)
3696 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3697 zone_page_state(z, NR_FREE_PAGES));
3700 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3701 unsigned long mark, int highest_zoneidx,
3702 unsigned int alloc_flags, gfp_t gfp_mask)
3706 free_pages = zone_page_state(z, NR_FREE_PAGES);
3709 * Fast check for order-0 only. If this fails then the reserves
3710 * need to be calculated.
3715 fast_free = free_pages;
3716 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3717 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3721 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3725 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3726 * when checking the min watermark. The min watermark is the
3727 * point where boosting is ignored so that kswapd is woken up
3728 * when below the low watermark.
3730 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3731 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3732 mark = z->_watermark[WMARK_MIN];
3733 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3734 alloc_flags, free_pages);
3740 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3741 unsigned long mark, int highest_zoneidx)
3743 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3745 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3746 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3748 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3753 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3755 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3756 node_reclaim_distance;
3758 #else /* CONFIG_NUMA */
3759 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3763 #endif /* CONFIG_NUMA */
3766 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3767 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3768 * premature use of a lower zone may cause lowmem pressure problems that
3769 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3770 * probably too small. It only makes sense to spread allocations to avoid
3771 * fragmentation between the Normal and DMA32 zones.
3773 static inline unsigned int
3774 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3776 unsigned int alloc_flags;
3779 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3782 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3784 #ifdef CONFIG_ZONE_DMA32
3788 if (zone_idx(zone) != ZONE_NORMAL)
3792 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3793 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3794 * on UMA that if Normal is populated then so is DMA32.
3796 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3797 if (nr_online_nodes > 1 && !populated_zone(--zone))
3800 alloc_flags |= ALLOC_NOFRAGMENT;
3801 #endif /* CONFIG_ZONE_DMA32 */
3805 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3806 unsigned int alloc_flags)
3809 unsigned int pflags = current->flags;
3811 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3812 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3813 alloc_flags |= ALLOC_CMA;
3820 * get_page_from_freelist goes through the zonelist trying to allocate
3823 static struct page *
3824 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3825 const struct alloc_context *ac)
3829 struct pglist_data *last_pgdat_dirty_limit = NULL;
3834 * Scan zonelist, looking for a zone with enough free.
3835 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3837 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3838 z = ac->preferred_zoneref;
3839 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3844 if (cpusets_enabled() &&
3845 (alloc_flags & ALLOC_CPUSET) &&
3846 !__cpuset_zone_allowed(zone, gfp_mask))
3849 * When allocating a page cache page for writing, we
3850 * want to get it from a node that is within its dirty
3851 * limit, such that no single node holds more than its
3852 * proportional share of globally allowed dirty pages.
3853 * The dirty limits take into account the node's
3854 * lowmem reserves and high watermark so that kswapd
3855 * should be able to balance it without having to
3856 * write pages from its LRU list.
3858 * XXX: For now, allow allocations to potentially
3859 * exceed the per-node dirty limit in the slowpath
3860 * (spread_dirty_pages unset) before going into reclaim,
3861 * which is important when on a NUMA setup the allowed
3862 * nodes are together not big enough to reach the
3863 * global limit. The proper fix for these situations
3864 * will require awareness of nodes in the
3865 * dirty-throttling and the flusher threads.
3867 if (ac->spread_dirty_pages) {
3868 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3871 if (!node_dirty_ok(zone->zone_pgdat)) {
3872 last_pgdat_dirty_limit = zone->zone_pgdat;
3877 if (no_fallback && nr_online_nodes > 1 &&
3878 zone != ac->preferred_zoneref->zone) {
3882 * If moving to a remote node, retry but allow
3883 * fragmenting fallbacks. Locality is more important
3884 * than fragmentation avoidance.
3886 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3887 if (zone_to_nid(zone) != local_nid) {
3888 alloc_flags &= ~ALLOC_NOFRAGMENT;
3893 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3894 if (!zone_watermark_fast(zone, order, mark,
3895 ac->highest_zoneidx, alloc_flags,
3899 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3901 * Watermark failed for this zone, but see if we can
3902 * grow this zone if it contains deferred pages.
3904 if (static_branch_unlikely(&deferred_pages)) {
3905 if (_deferred_grow_zone(zone, order))
3909 /* Checked here to keep the fast path fast */
3910 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3911 if (alloc_flags & ALLOC_NO_WATERMARKS)
3914 if (node_reclaim_mode == 0 ||
3915 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3918 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3920 case NODE_RECLAIM_NOSCAN:
3923 case NODE_RECLAIM_FULL:
3924 /* scanned but unreclaimable */
3927 /* did we reclaim enough */
3928 if (zone_watermark_ok(zone, order, mark,
3929 ac->highest_zoneidx, alloc_flags))
3937 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3938 gfp_mask, alloc_flags, ac->migratetype);
3940 prep_new_page(page, order, gfp_mask, alloc_flags);
3943 * If this is a high-order atomic allocation then check
3944 * if the pageblock should be reserved for the future
3946 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3947 reserve_highatomic_pageblock(page, zone, order);
3951 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3952 /* Try again if zone has deferred pages */
3953 if (static_branch_unlikely(&deferred_pages)) {
3954 if (_deferred_grow_zone(zone, order))
3962 * It's possible on a UMA machine to get through all zones that are
3963 * fragmented. If avoiding fragmentation, reset and try again.
3966 alloc_flags &= ~ALLOC_NOFRAGMENT;
3973 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3975 unsigned int filter = SHOW_MEM_FILTER_NODES;
3978 * This documents exceptions given to allocations in certain
3979 * contexts that are allowed to allocate outside current's set
3982 if (!(gfp_mask & __GFP_NOMEMALLOC))
3983 if (tsk_is_oom_victim(current) ||
3984 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3985 filter &= ~SHOW_MEM_FILTER_NODES;
3986 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3987 filter &= ~SHOW_MEM_FILTER_NODES;
3989 show_mem(filter, nodemask);
3992 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3994 struct va_format vaf;
3996 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3998 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4001 va_start(args, fmt);
4004 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4005 current->comm, &vaf, gfp_mask, &gfp_mask,
4006 nodemask_pr_args(nodemask));
4009 cpuset_print_current_mems_allowed();
4012 warn_alloc_show_mem(gfp_mask, nodemask);
4015 static inline struct page *
4016 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4017 unsigned int alloc_flags,
4018 const struct alloc_context *ac)
4022 page = get_page_from_freelist(gfp_mask, order,
4023 alloc_flags|ALLOC_CPUSET, ac);
4025 * fallback to ignore cpuset restriction if our nodes
4029 page = get_page_from_freelist(gfp_mask, order,
4035 static inline struct page *
4036 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4037 const struct alloc_context *ac, unsigned long *did_some_progress)
4039 struct oom_control oc = {
4040 .zonelist = ac->zonelist,
4041 .nodemask = ac->nodemask,
4043 .gfp_mask = gfp_mask,
4048 *did_some_progress = 0;
4051 * Acquire the oom lock. If that fails, somebody else is
4052 * making progress for us.
4054 if (!mutex_trylock(&oom_lock)) {
4055 *did_some_progress = 1;
4056 schedule_timeout_uninterruptible(1);
4061 * Go through the zonelist yet one more time, keep very high watermark
4062 * here, this is only to catch a parallel oom killing, we must fail if
4063 * we're still under heavy pressure. But make sure that this reclaim
4064 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4065 * allocation which will never fail due to oom_lock already held.
4067 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4068 ~__GFP_DIRECT_RECLAIM, order,
4069 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4073 /* Coredumps can quickly deplete all memory reserves */
4074 if (current->flags & PF_DUMPCORE)
4076 /* The OOM killer will not help higher order allocs */
4077 if (order > PAGE_ALLOC_COSTLY_ORDER)
4080 * We have already exhausted all our reclaim opportunities without any
4081 * success so it is time to admit defeat. We will skip the OOM killer
4082 * because it is very likely that the caller has a more reasonable
4083 * fallback than shooting a random task.
4085 * The OOM killer may not free memory on a specific node.
4087 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4089 /* The OOM killer does not needlessly kill tasks for lowmem */
4090 if (ac->highest_zoneidx < ZONE_NORMAL)
4092 if (pm_suspended_storage())
4095 * XXX: GFP_NOFS allocations should rather fail than rely on
4096 * other request to make a forward progress.
4097 * We are in an unfortunate situation where out_of_memory cannot
4098 * do much for this context but let's try it to at least get
4099 * access to memory reserved if the current task is killed (see
4100 * out_of_memory). Once filesystems are ready to handle allocation
4101 * failures more gracefully we should just bail out here.
4104 /* Exhausted what can be done so it's blame time */
4105 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4106 *did_some_progress = 1;
4109 * Help non-failing allocations by giving them access to memory
4112 if (gfp_mask & __GFP_NOFAIL)
4113 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4114 ALLOC_NO_WATERMARKS, ac);
4117 mutex_unlock(&oom_lock);
4122 * Maximum number of compaction retries wit a progress before OOM
4123 * killer is consider as the only way to move forward.
4125 #define MAX_COMPACT_RETRIES 16
4127 #ifdef CONFIG_COMPACTION
4128 /* Try memory compaction for high-order allocations before reclaim */
4129 static struct page *
4130 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4131 unsigned int alloc_flags, const struct alloc_context *ac,
4132 enum compact_priority prio, enum compact_result *compact_result)
4134 struct page *page = NULL;
4135 unsigned long pflags;
4136 unsigned int noreclaim_flag;
4141 psi_memstall_enter(&pflags);
4142 noreclaim_flag = memalloc_noreclaim_save();
4144 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4147 memalloc_noreclaim_restore(noreclaim_flag);
4148 psi_memstall_leave(&pflags);
4151 * At least in one zone compaction wasn't deferred or skipped, so let's
4152 * count a compaction stall
4154 count_vm_event(COMPACTSTALL);
4156 /* Prep a captured page if available */
4158 prep_new_page(page, order, gfp_mask, alloc_flags);
4160 /* Try get a page from the freelist if available */
4162 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4165 struct zone *zone = page_zone(page);
4167 zone->compact_blockskip_flush = false;
4168 compaction_defer_reset(zone, order, true);
4169 count_vm_event(COMPACTSUCCESS);
4174 * It's bad if compaction run occurs and fails. The most likely reason
4175 * is that pages exist, but not enough to satisfy watermarks.
4177 count_vm_event(COMPACTFAIL);
4185 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4186 enum compact_result compact_result,
4187 enum compact_priority *compact_priority,
4188 int *compaction_retries)
4190 int max_retries = MAX_COMPACT_RETRIES;
4193 int retries = *compaction_retries;
4194 enum compact_priority priority = *compact_priority;
4199 if (compaction_made_progress(compact_result))
4200 (*compaction_retries)++;
4203 * compaction considers all the zone as desperately out of memory
4204 * so it doesn't really make much sense to retry except when the
4205 * failure could be caused by insufficient priority
4207 if (compaction_failed(compact_result))
4208 goto check_priority;
4211 * compaction was skipped because there are not enough order-0 pages
4212 * to work with, so we retry only if it looks like reclaim can help.
4214 if (compaction_needs_reclaim(compact_result)) {
4215 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4220 * make sure the compaction wasn't deferred or didn't bail out early
4221 * due to locks contention before we declare that we should give up.
4222 * But the next retry should use a higher priority if allowed, so
4223 * we don't just keep bailing out endlessly.
4225 if (compaction_withdrawn(compact_result)) {
4226 goto check_priority;
4230 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4231 * costly ones because they are de facto nofail and invoke OOM
4232 * killer to move on while costly can fail and users are ready
4233 * to cope with that. 1/4 retries is rather arbitrary but we
4234 * would need much more detailed feedback from compaction to
4235 * make a better decision.
4237 if (order > PAGE_ALLOC_COSTLY_ORDER)
4239 if (*compaction_retries <= max_retries) {
4245 * Make sure there are attempts at the highest priority if we exhausted
4246 * all retries or failed at the lower priorities.
4249 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4250 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4252 if (*compact_priority > min_priority) {
4253 (*compact_priority)--;
4254 *compaction_retries = 0;
4258 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4262 static inline struct page *
4263 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4264 unsigned int alloc_flags, const struct alloc_context *ac,
4265 enum compact_priority prio, enum compact_result *compact_result)
4267 *compact_result = COMPACT_SKIPPED;
4272 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4273 enum compact_result compact_result,
4274 enum compact_priority *compact_priority,
4275 int *compaction_retries)
4280 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4284 * There are setups with compaction disabled which would prefer to loop
4285 * inside the allocator rather than hit the oom killer prematurely.
4286 * Let's give them a good hope and keep retrying while the order-0
4287 * watermarks are OK.
4289 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4290 ac->highest_zoneidx, ac->nodemask) {
4291 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4292 ac->highest_zoneidx, alloc_flags))
4297 #endif /* CONFIG_COMPACTION */
4299 #ifdef CONFIG_LOCKDEP
4300 static struct lockdep_map __fs_reclaim_map =
4301 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4303 static bool __need_reclaim(gfp_t gfp_mask)
4305 /* no reclaim without waiting on it */
4306 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4309 /* this guy won't enter reclaim */
4310 if (current->flags & PF_MEMALLOC)
4313 if (gfp_mask & __GFP_NOLOCKDEP)
4319 void __fs_reclaim_acquire(void)
4321 lock_map_acquire(&__fs_reclaim_map);
4324 void __fs_reclaim_release(void)
4326 lock_map_release(&__fs_reclaim_map);
4329 void fs_reclaim_acquire(gfp_t gfp_mask)
4331 gfp_mask = current_gfp_context(gfp_mask);
4333 if (__need_reclaim(gfp_mask)) {
4334 if (gfp_mask & __GFP_FS)
4335 __fs_reclaim_acquire();
4337 #ifdef CONFIG_MMU_NOTIFIER
4338 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4339 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4344 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4346 void fs_reclaim_release(gfp_t gfp_mask)
4348 gfp_mask = current_gfp_context(gfp_mask);
4350 if (__need_reclaim(gfp_mask)) {
4351 if (gfp_mask & __GFP_FS)
4352 __fs_reclaim_release();
4355 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4358 /* Perform direct synchronous page reclaim */
4359 static unsigned long
4360 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4361 const struct alloc_context *ac)
4363 unsigned int noreclaim_flag;
4364 unsigned long pflags, progress;
4368 /* We now go into synchronous reclaim */
4369 cpuset_memory_pressure_bump();
4370 psi_memstall_enter(&pflags);
4371 fs_reclaim_acquire(gfp_mask);
4372 noreclaim_flag = memalloc_noreclaim_save();
4374 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4377 memalloc_noreclaim_restore(noreclaim_flag);
4378 fs_reclaim_release(gfp_mask);
4379 psi_memstall_leave(&pflags);
4386 /* The really slow allocator path where we enter direct reclaim */
4387 static inline struct page *
4388 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4389 unsigned int alloc_flags, const struct alloc_context *ac,
4390 unsigned long *did_some_progress)
4392 struct page *page = NULL;
4393 bool drained = false;
4395 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4396 if (unlikely(!(*did_some_progress)))
4400 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4403 * If an allocation failed after direct reclaim, it could be because
4404 * pages are pinned on the per-cpu lists or in high alloc reserves.
4405 * Shrink them and try again
4407 if (!page && !drained) {
4408 unreserve_highatomic_pageblock(ac, false);
4409 drain_all_pages(NULL);
4417 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4418 const struct alloc_context *ac)
4422 pg_data_t *last_pgdat = NULL;
4423 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4425 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4427 if (last_pgdat != zone->zone_pgdat)
4428 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4429 last_pgdat = zone->zone_pgdat;
4433 static inline unsigned int
4434 gfp_to_alloc_flags(gfp_t gfp_mask)
4436 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4439 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4440 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4441 * to save two branches.
4443 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4444 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4447 * The caller may dip into page reserves a bit more if the caller
4448 * cannot run direct reclaim, or if the caller has realtime scheduling
4449 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4450 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4452 alloc_flags |= (__force int)
4453 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4455 if (gfp_mask & __GFP_ATOMIC) {
4457 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4458 * if it can't schedule.
4460 if (!(gfp_mask & __GFP_NOMEMALLOC))
4461 alloc_flags |= ALLOC_HARDER;
4463 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4464 * comment for __cpuset_node_allowed().
4466 alloc_flags &= ~ALLOC_CPUSET;
4467 } else if (unlikely(rt_task(current)) && !in_interrupt())
4468 alloc_flags |= ALLOC_HARDER;
4470 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4475 static bool oom_reserves_allowed(struct task_struct *tsk)
4477 if (!tsk_is_oom_victim(tsk))
4481 * !MMU doesn't have oom reaper so give access to memory reserves
4482 * only to the thread with TIF_MEMDIE set
4484 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4491 * Distinguish requests which really need access to full memory
4492 * reserves from oom victims which can live with a portion of it
4494 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4496 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4498 if (gfp_mask & __GFP_MEMALLOC)
4499 return ALLOC_NO_WATERMARKS;
4500 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4501 return ALLOC_NO_WATERMARKS;
4502 if (!in_interrupt()) {
4503 if (current->flags & PF_MEMALLOC)
4504 return ALLOC_NO_WATERMARKS;
4505 else if (oom_reserves_allowed(current))
4512 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4514 return !!__gfp_pfmemalloc_flags(gfp_mask);
4518 * Checks whether it makes sense to retry the reclaim to make a forward progress
4519 * for the given allocation request.
4521 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4522 * without success, or when we couldn't even meet the watermark if we
4523 * reclaimed all remaining pages on the LRU lists.
4525 * Returns true if a retry is viable or false to enter the oom path.
4528 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4529 struct alloc_context *ac, int alloc_flags,
4530 bool did_some_progress, int *no_progress_loops)
4537 * Costly allocations might have made a progress but this doesn't mean
4538 * their order will become available due to high fragmentation so
4539 * always increment the no progress counter for them
4541 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4542 *no_progress_loops = 0;
4544 (*no_progress_loops)++;
4547 * Make sure we converge to OOM if we cannot make any progress
4548 * several times in the row.
4550 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4551 /* Before OOM, exhaust highatomic_reserve */
4552 return unreserve_highatomic_pageblock(ac, true);
4556 * Keep reclaiming pages while there is a chance this will lead
4557 * somewhere. If none of the target zones can satisfy our allocation
4558 * request even if all reclaimable pages are considered then we are
4559 * screwed and have to go OOM.
4561 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4562 ac->highest_zoneidx, ac->nodemask) {
4563 unsigned long available;
4564 unsigned long reclaimable;
4565 unsigned long min_wmark = min_wmark_pages(zone);
4568 available = reclaimable = zone_reclaimable_pages(zone);
4569 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4572 * Would the allocation succeed if we reclaimed all
4573 * reclaimable pages?
4575 wmark = __zone_watermark_ok(zone, order, min_wmark,
4576 ac->highest_zoneidx, alloc_flags, available);
4577 trace_reclaim_retry_zone(z, order, reclaimable,
4578 available, min_wmark, *no_progress_loops, wmark);
4581 * If we didn't make any progress and have a lot of
4582 * dirty + writeback pages then we should wait for
4583 * an IO to complete to slow down the reclaim and
4584 * prevent from pre mature OOM
4586 if (!did_some_progress) {
4587 unsigned long write_pending;
4589 write_pending = zone_page_state_snapshot(zone,
4590 NR_ZONE_WRITE_PENDING);
4592 if (2 * write_pending > reclaimable) {
4593 congestion_wait(BLK_RW_ASYNC, HZ/10);
4605 * Memory allocation/reclaim might be called from a WQ context and the
4606 * current implementation of the WQ concurrency control doesn't
4607 * recognize that a particular WQ is congested if the worker thread is
4608 * looping without ever sleeping. Therefore we have to do a short sleep
4609 * here rather than calling cond_resched().
4611 if (current->flags & PF_WQ_WORKER)
4612 schedule_timeout_uninterruptible(1);
4619 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4622 * It's possible that cpuset's mems_allowed and the nodemask from
4623 * mempolicy don't intersect. This should be normally dealt with by
4624 * policy_nodemask(), but it's possible to race with cpuset update in
4625 * such a way the check therein was true, and then it became false
4626 * before we got our cpuset_mems_cookie here.
4627 * This assumes that for all allocations, ac->nodemask can come only
4628 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4629 * when it does not intersect with the cpuset restrictions) or the
4630 * caller can deal with a violated nodemask.
4632 if (cpusets_enabled() && ac->nodemask &&
4633 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4634 ac->nodemask = NULL;
4639 * When updating a task's mems_allowed or mempolicy nodemask, it is
4640 * possible to race with parallel threads in such a way that our
4641 * allocation can fail while the mask is being updated. If we are about
4642 * to fail, check if the cpuset changed during allocation and if so,
4645 if (read_mems_allowed_retry(cpuset_mems_cookie))
4651 static inline struct page *
4652 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4653 struct alloc_context *ac)
4655 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4656 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4657 struct page *page = NULL;
4658 unsigned int alloc_flags;
4659 unsigned long did_some_progress;
4660 enum compact_priority compact_priority;
4661 enum compact_result compact_result;
4662 int compaction_retries;
4663 int no_progress_loops;
4664 unsigned int cpuset_mems_cookie;
4668 * We also sanity check to catch abuse of atomic reserves being used by
4669 * callers that are not in atomic context.
4671 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4672 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4673 gfp_mask &= ~__GFP_ATOMIC;
4676 compaction_retries = 0;
4677 no_progress_loops = 0;
4678 compact_priority = DEF_COMPACT_PRIORITY;
4679 cpuset_mems_cookie = read_mems_allowed_begin();
4682 * The fast path uses conservative alloc_flags to succeed only until
4683 * kswapd needs to be woken up, and to avoid the cost of setting up
4684 * alloc_flags precisely. So we do that now.
4686 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4689 * We need to recalculate the starting point for the zonelist iterator
4690 * because we might have used different nodemask in the fast path, or
4691 * there was a cpuset modification and we are retrying - otherwise we
4692 * could end up iterating over non-eligible zones endlessly.
4694 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4695 ac->highest_zoneidx, ac->nodemask);
4696 if (!ac->preferred_zoneref->zone)
4699 if (alloc_flags & ALLOC_KSWAPD)
4700 wake_all_kswapds(order, gfp_mask, ac);
4703 * The adjusted alloc_flags might result in immediate success, so try
4706 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4711 * For costly allocations, try direct compaction first, as it's likely
4712 * that we have enough base pages and don't need to reclaim. For non-
4713 * movable high-order allocations, do that as well, as compaction will
4714 * try prevent permanent fragmentation by migrating from blocks of the
4716 * Don't try this for allocations that are allowed to ignore
4717 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4719 if (can_direct_reclaim &&
4721 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4722 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4723 page = __alloc_pages_direct_compact(gfp_mask, order,
4725 INIT_COMPACT_PRIORITY,
4731 * Checks for costly allocations with __GFP_NORETRY, which
4732 * includes some THP page fault allocations
4734 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4736 * If allocating entire pageblock(s) and compaction
4737 * failed because all zones are below low watermarks
4738 * or is prohibited because it recently failed at this
4739 * order, fail immediately unless the allocator has
4740 * requested compaction and reclaim retry.
4743 * - potentially very expensive because zones are far
4744 * below their low watermarks or this is part of very
4745 * bursty high order allocations,
4746 * - not guaranteed to help because isolate_freepages()
4747 * may not iterate over freed pages as part of its
4749 * - unlikely to make entire pageblocks free on its
4752 if (compact_result == COMPACT_SKIPPED ||
4753 compact_result == COMPACT_DEFERRED)
4757 * Looks like reclaim/compaction is worth trying, but
4758 * sync compaction could be very expensive, so keep
4759 * using async compaction.
4761 compact_priority = INIT_COMPACT_PRIORITY;
4766 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4767 if (alloc_flags & ALLOC_KSWAPD)
4768 wake_all_kswapds(order, gfp_mask, ac);
4770 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4772 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4775 * Reset the nodemask and zonelist iterators if memory policies can be
4776 * ignored. These allocations are high priority and system rather than
4779 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4780 ac->nodemask = NULL;
4781 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4782 ac->highest_zoneidx, ac->nodemask);
4785 /* Attempt with potentially adjusted zonelist and alloc_flags */
4786 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4790 /* Caller is not willing to reclaim, we can't balance anything */
4791 if (!can_direct_reclaim)
4794 /* Avoid recursion of direct reclaim */
4795 if (current->flags & PF_MEMALLOC)
4798 /* Try direct reclaim and then allocating */
4799 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4800 &did_some_progress);
4804 /* Try direct compaction and then allocating */
4805 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4806 compact_priority, &compact_result);
4810 /* Do not loop if specifically requested */
4811 if (gfp_mask & __GFP_NORETRY)
4815 * Do not retry costly high order allocations unless they are
4816 * __GFP_RETRY_MAYFAIL
4818 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4821 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4822 did_some_progress > 0, &no_progress_loops))
4826 * It doesn't make any sense to retry for the compaction if the order-0
4827 * reclaim is not able to make any progress because the current
4828 * implementation of the compaction depends on the sufficient amount
4829 * of free memory (see __compaction_suitable)
4831 if (did_some_progress > 0 &&
4832 should_compact_retry(ac, order, alloc_flags,
4833 compact_result, &compact_priority,
4834 &compaction_retries))
4838 /* Deal with possible cpuset update races before we start OOM killing */
4839 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4842 /* Reclaim has failed us, start killing things */
4843 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4847 /* Avoid allocations with no watermarks from looping endlessly */
4848 if (tsk_is_oom_victim(current) &&
4849 (alloc_flags & ALLOC_OOM ||
4850 (gfp_mask & __GFP_NOMEMALLOC)))
4853 /* Retry as long as the OOM killer is making progress */
4854 if (did_some_progress) {
4855 no_progress_loops = 0;
4860 /* Deal with possible cpuset update races before we fail */
4861 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4865 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4868 if (gfp_mask & __GFP_NOFAIL) {
4870 * All existing users of the __GFP_NOFAIL are blockable, so warn
4871 * of any new users that actually require GFP_NOWAIT
4873 if (WARN_ON_ONCE(!can_direct_reclaim))
4877 * PF_MEMALLOC request from this context is rather bizarre
4878 * because we cannot reclaim anything and only can loop waiting
4879 * for somebody to do a work for us
4881 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4884 * non failing costly orders are a hard requirement which we
4885 * are not prepared for much so let's warn about these users
4886 * so that we can identify them and convert them to something
4889 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4892 * Help non-failing allocations by giving them access to memory
4893 * reserves but do not use ALLOC_NO_WATERMARKS because this
4894 * could deplete whole memory reserves which would just make
4895 * the situation worse
4897 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4905 warn_alloc(gfp_mask, ac->nodemask,
4906 "page allocation failure: order:%u", order);
4911 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4912 int preferred_nid, nodemask_t *nodemask,
4913 struct alloc_context *ac, gfp_t *alloc_mask,
4914 unsigned int *alloc_flags)
4916 ac->highest_zoneidx = gfp_zone(gfp_mask);
4917 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4918 ac->nodemask = nodemask;
4919 ac->migratetype = gfp_migratetype(gfp_mask);
4921 if (cpusets_enabled()) {
4922 *alloc_mask |= __GFP_HARDWALL;
4924 * When we are in the interrupt context, it is irrelevant
4925 * to the current task context. It means that any node ok.
4927 if (!in_interrupt() && !ac->nodemask)
4928 ac->nodemask = &cpuset_current_mems_allowed;
4930 *alloc_flags |= ALLOC_CPUSET;
4933 fs_reclaim_acquire(gfp_mask);
4934 fs_reclaim_release(gfp_mask);
4936 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4938 if (should_fail_alloc_page(gfp_mask, order))
4941 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4943 /* Dirty zone balancing only done in the fast path */
4944 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4947 * The preferred zone is used for statistics but crucially it is
4948 * also used as the starting point for the zonelist iterator. It
4949 * may get reset for allocations that ignore memory policies.
4951 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4952 ac->highest_zoneidx, ac->nodemask);
4958 * This is the 'heart' of the zoned buddy allocator.
4961 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4962 nodemask_t *nodemask)
4965 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4966 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4967 struct alloc_context ac = { };
4970 * There are several places where we assume that the order value is sane
4971 * so bail out early if the request is out of bound.
4973 if (unlikely(order >= MAX_ORDER)) {
4974 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4978 gfp_mask &= gfp_allowed_mask;
4979 alloc_mask = gfp_mask;
4980 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4984 * Forbid the first pass from falling back to types that fragment
4985 * memory until all local zones are considered.
4987 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4989 /* First allocation attempt */
4990 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4995 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4996 * resp. GFP_NOIO which has to be inherited for all allocation requests
4997 * from a particular context which has been marked by
4998 * memalloc_no{fs,io}_{save,restore}.
5000 alloc_mask = current_gfp_context(gfp_mask);
5001 ac.spread_dirty_pages = false;
5004 * Restore the original nodemask if it was potentially replaced with
5005 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5007 ac.nodemask = nodemask;
5009 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
5012 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
5013 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
5014 __free_pages(page, order);
5018 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
5022 EXPORT_SYMBOL(__alloc_pages_nodemask);
5025 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5026 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5027 * you need to access high mem.
5029 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5033 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5036 return (unsigned long) page_address(page);
5038 EXPORT_SYMBOL(__get_free_pages);
5040 unsigned long get_zeroed_page(gfp_t gfp_mask)
5042 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5044 EXPORT_SYMBOL(get_zeroed_page);
5046 static inline void free_the_page(struct page *page, unsigned int order)
5048 if (order == 0) /* Via pcp? */
5049 free_unref_page(page);
5051 __free_pages_ok(page, order, FPI_NONE);
5055 * __free_pages - Free pages allocated with alloc_pages().
5056 * @page: The page pointer returned from alloc_pages().
5057 * @order: The order of the allocation.
5059 * This function can free multi-page allocations that are not compound
5060 * pages. It does not check that the @order passed in matches that of
5061 * the allocation, so it is easy to leak memory. Freeing more memory
5062 * than was allocated will probably emit a warning.
5064 * If the last reference to this page is speculative, it will be released
5065 * by put_page() which only frees the first page of a non-compound
5066 * allocation. To prevent the remaining pages from being leaked, we free
5067 * the subsequent pages here. If you want to use the page's reference
5068 * count to decide when to free the allocation, you should allocate a
5069 * compound page, and use put_page() instead of __free_pages().
5071 * Context: May be called in interrupt context or while holding a normal
5072 * spinlock, but not in NMI context or while holding a raw spinlock.
5074 void __free_pages(struct page *page, unsigned int order)
5076 if (put_page_testzero(page))
5077 free_the_page(page, order);
5078 else if (!PageHead(page))
5080 free_the_page(page + (1 << order), order);
5082 EXPORT_SYMBOL(__free_pages);
5084 void free_pages(unsigned long addr, unsigned int order)
5087 VM_BUG_ON(!virt_addr_valid((void *)addr));
5088 __free_pages(virt_to_page((void *)addr), order);
5092 EXPORT_SYMBOL(free_pages);
5096 * An arbitrary-length arbitrary-offset area of memory which resides
5097 * within a 0 or higher order page. Multiple fragments within that page
5098 * are individually refcounted, in the page's reference counter.
5100 * The page_frag functions below provide a simple allocation framework for
5101 * page fragments. This is used by the network stack and network device
5102 * drivers to provide a backing region of memory for use as either an
5103 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5105 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5108 struct page *page = NULL;
5109 gfp_t gfp = gfp_mask;
5111 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5112 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5114 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5115 PAGE_FRAG_CACHE_MAX_ORDER);
5116 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5118 if (unlikely(!page))
5119 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5121 nc->va = page ? page_address(page) : NULL;
5126 void __page_frag_cache_drain(struct page *page, unsigned int count)
5128 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5130 if (page_ref_sub_and_test(page, count))
5131 free_the_page(page, compound_order(page));
5133 EXPORT_SYMBOL(__page_frag_cache_drain);
5135 void *page_frag_alloc(struct page_frag_cache *nc,
5136 unsigned int fragsz, gfp_t gfp_mask)
5138 unsigned int size = PAGE_SIZE;
5142 if (unlikely(!nc->va)) {
5144 page = __page_frag_cache_refill(nc, gfp_mask);
5148 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5149 /* if size can vary use size else just use PAGE_SIZE */
5152 /* Even if we own the page, we do not use atomic_set().
5153 * This would break get_page_unless_zero() users.
5155 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5157 /* reset page count bias and offset to start of new frag */
5158 nc->pfmemalloc = page_is_pfmemalloc(page);
5159 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5163 offset = nc->offset - fragsz;
5164 if (unlikely(offset < 0)) {
5165 page = virt_to_page(nc->va);
5167 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5170 if (unlikely(nc->pfmemalloc)) {
5171 free_the_page(page, compound_order(page));
5175 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5176 /* if size can vary use size else just use PAGE_SIZE */
5179 /* OK, page count is 0, we can safely set it */
5180 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5182 /* reset page count bias and offset to start of new frag */
5183 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5184 offset = size - fragsz;
5188 nc->offset = offset;
5190 return nc->va + offset;
5192 EXPORT_SYMBOL(page_frag_alloc);
5195 * Frees a page fragment allocated out of either a compound or order 0 page.
5197 void page_frag_free(void *addr)
5199 struct page *page = virt_to_head_page(addr);
5201 if (unlikely(put_page_testzero(page)))
5202 free_the_page(page, compound_order(page));
5204 EXPORT_SYMBOL(page_frag_free);
5206 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5210 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5211 unsigned long used = addr + PAGE_ALIGN(size);
5213 split_page(virt_to_page((void *)addr), order);
5214 while (used < alloc_end) {
5219 return (void *)addr;
5223 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5224 * @size: the number of bytes to allocate
5225 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5227 * This function is similar to alloc_pages(), except that it allocates the
5228 * minimum number of pages to satisfy the request. alloc_pages() can only
5229 * allocate memory in power-of-two pages.
5231 * This function is also limited by MAX_ORDER.
5233 * Memory allocated by this function must be released by free_pages_exact().
5235 * Return: pointer to the allocated area or %NULL in case of error.
5237 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5239 unsigned int order = get_order(size);
5242 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5243 gfp_mask &= ~__GFP_COMP;
5245 addr = __get_free_pages(gfp_mask, order);
5246 return make_alloc_exact(addr, order, size);
5248 EXPORT_SYMBOL(alloc_pages_exact);
5251 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5253 * @nid: the preferred node ID where memory should be allocated
5254 * @size: the number of bytes to allocate
5255 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5257 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5260 * Return: pointer to the allocated area or %NULL in case of error.
5262 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5264 unsigned int order = get_order(size);
5267 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5268 gfp_mask &= ~__GFP_COMP;
5270 p = alloc_pages_node(nid, gfp_mask, order);
5273 return make_alloc_exact((unsigned long)page_address(p), order, size);
5277 * free_pages_exact - release memory allocated via alloc_pages_exact()
5278 * @virt: the value returned by alloc_pages_exact.
5279 * @size: size of allocation, same value as passed to alloc_pages_exact().
5281 * Release the memory allocated by a previous call to alloc_pages_exact.
5283 void free_pages_exact(void *virt, size_t size)
5285 unsigned long addr = (unsigned long)virt;
5286 unsigned long end = addr + PAGE_ALIGN(size);
5288 while (addr < end) {
5293 EXPORT_SYMBOL(free_pages_exact);
5296 * nr_free_zone_pages - count number of pages beyond high watermark
5297 * @offset: The zone index of the highest zone
5299 * nr_free_zone_pages() counts the number of pages which are beyond the
5300 * high watermark within all zones at or below a given zone index. For each
5301 * zone, the number of pages is calculated as:
5303 * nr_free_zone_pages = managed_pages - high_pages
5305 * Return: number of pages beyond high watermark.
5307 static unsigned long nr_free_zone_pages(int offset)
5312 /* Just pick one node, since fallback list is circular */
5313 unsigned long sum = 0;
5315 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5317 for_each_zone_zonelist(zone, z, zonelist, offset) {
5318 unsigned long size = zone_managed_pages(zone);
5319 unsigned long high = high_wmark_pages(zone);
5328 * nr_free_buffer_pages - count number of pages beyond high watermark
5330 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5331 * watermark within ZONE_DMA and ZONE_NORMAL.
5333 * Return: number of pages beyond high watermark within ZONE_DMA and
5336 unsigned long nr_free_buffer_pages(void)
5338 return nr_free_zone_pages(gfp_zone(GFP_USER));
5340 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5342 static inline void show_node(struct zone *zone)
5344 if (IS_ENABLED(CONFIG_NUMA))
5345 printk("Node %d ", zone_to_nid(zone));
5348 long si_mem_available(void)
5351 unsigned long pagecache;
5352 unsigned long wmark_low = 0;
5353 unsigned long pages[NR_LRU_LISTS];
5354 unsigned long reclaimable;
5358 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5359 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5362 wmark_low += low_wmark_pages(zone);
5365 * Estimate the amount of memory available for userspace allocations,
5366 * without causing swapping.
5368 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5371 * Not all the page cache can be freed, otherwise the system will
5372 * start swapping. Assume at least half of the page cache, or the
5373 * low watermark worth of cache, needs to stay.
5375 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5376 pagecache -= min(pagecache / 2, wmark_low);
5377 available += pagecache;
5380 * Part of the reclaimable slab and other kernel memory consists of
5381 * items that are in use, and cannot be freed. Cap this estimate at the
5384 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5385 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5386 available += reclaimable - min(reclaimable / 2, wmark_low);
5392 EXPORT_SYMBOL_GPL(si_mem_available);
5394 void si_meminfo(struct sysinfo *val)
5396 val->totalram = totalram_pages();
5397 val->sharedram = global_node_page_state(NR_SHMEM);
5398 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5399 val->bufferram = nr_blockdev_pages();
5400 val->totalhigh = totalhigh_pages();
5401 val->freehigh = nr_free_highpages();
5402 val->mem_unit = PAGE_SIZE;
5405 EXPORT_SYMBOL(si_meminfo);
5408 void si_meminfo_node(struct sysinfo *val, int nid)
5410 int zone_type; /* needs to be signed */
5411 unsigned long managed_pages = 0;
5412 unsigned long managed_highpages = 0;
5413 unsigned long free_highpages = 0;
5414 pg_data_t *pgdat = NODE_DATA(nid);
5416 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5417 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5418 val->totalram = managed_pages;
5419 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5420 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5421 #ifdef CONFIG_HIGHMEM
5422 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5423 struct zone *zone = &pgdat->node_zones[zone_type];
5425 if (is_highmem(zone)) {
5426 managed_highpages += zone_managed_pages(zone);
5427 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5430 val->totalhigh = managed_highpages;
5431 val->freehigh = free_highpages;
5433 val->totalhigh = managed_highpages;
5434 val->freehigh = free_highpages;
5436 val->mem_unit = PAGE_SIZE;
5441 * Determine whether the node should be displayed or not, depending on whether
5442 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5444 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5446 if (!(flags & SHOW_MEM_FILTER_NODES))
5450 * no node mask - aka implicit memory numa policy. Do not bother with
5451 * the synchronization - read_mems_allowed_begin - because we do not
5452 * have to be precise here.
5455 nodemask = &cpuset_current_mems_allowed;
5457 return !node_isset(nid, *nodemask);
5460 #define K(x) ((x) << (PAGE_SHIFT-10))
5462 static void show_migration_types(unsigned char type)
5464 static const char types[MIGRATE_TYPES] = {
5465 [MIGRATE_UNMOVABLE] = 'U',
5466 [MIGRATE_MOVABLE] = 'M',
5467 [MIGRATE_RECLAIMABLE] = 'E',
5468 [MIGRATE_HIGHATOMIC] = 'H',
5470 [MIGRATE_CMA] = 'C',
5472 #ifdef CONFIG_MEMORY_ISOLATION
5473 [MIGRATE_ISOLATE] = 'I',
5476 char tmp[MIGRATE_TYPES + 1];
5480 for (i = 0; i < MIGRATE_TYPES; i++) {
5481 if (type & (1 << i))
5486 printk(KERN_CONT "(%s) ", tmp);
5490 * Show free area list (used inside shift_scroll-lock stuff)
5491 * We also calculate the percentage fragmentation. We do this by counting the
5492 * memory on each free list with the exception of the first item on the list.
5495 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5498 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5500 unsigned long free_pcp = 0;
5505 for_each_populated_zone(zone) {
5506 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5509 for_each_online_cpu(cpu)
5510 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5513 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5514 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5515 " unevictable:%lu dirty:%lu writeback:%lu\n"
5516 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5517 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5518 " free:%lu free_pcp:%lu free_cma:%lu\n",
5519 global_node_page_state(NR_ACTIVE_ANON),
5520 global_node_page_state(NR_INACTIVE_ANON),
5521 global_node_page_state(NR_ISOLATED_ANON),
5522 global_node_page_state(NR_ACTIVE_FILE),
5523 global_node_page_state(NR_INACTIVE_FILE),
5524 global_node_page_state(NR_ISOLATED_FILE),
5525 global_node_page_state(NR_UNEVICTABLE),
5526 global_node_page_state(NR_FILE_DIRTY),
5527 global_node_page_state(NR_WRITEBACK),
5528 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5529 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5530 global_node_page_state(NR_FILE_MAPPED),
5531 global_node_page_state(NR_SHMEM),
5532 global_node_page_state(NR_PAGETABLE),
5533 global_zone_page_state(NR_BOUNCE),
5534 global_zone_page_state(NR_FREE_PAGES),
5536 global_zone_page_state(NR_FREE_CMA_PAGES));
5538 for_each_online_pgdat(pgdat) {
5539 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5543 " active_anon:%lukB"
5544 " inactive_anon:%lukB"
5545 " active_file:%lukB"
5546 " inactive_file:%lukB"
5547 " unevictable:%lukB"
5548 " isolated(anon):%lukB"
5549 " isolated(file):%lukB"
5554 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5556 " shmem_pmdmapped: %lukB"
5559 " writeback_tmp:%lukB"
5560 " kernel_stack:%lukB"
5561 #ifdef CONFIG_SHADOW_CALL_STACK
5562 " shadow_call_stack:%lukB"
5565 " all_unreclaimable? %s"
5568 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5569 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5570 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5571 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5572 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5573 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5574 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5575 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5576 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5577 K(node_page_state(pgdat, NR_WRITEBACK)),
5578 K(node_page_state(pgdat, NR_SHMEM)),
5579 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5580 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5581 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5583 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5585 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5586 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5587 #ifdef CONFIG_SHADOW_CALL_STACK
5588 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5590 K(node_page_state(pgdat, NR_PAGETABLE)),
5591 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5595 for_each_populated_zone(zone) {
5598 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5602 for_each_online_cpu(cpu)
5603 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5612 " reserved_highatomic:%luKB"
5613 " active_anon:%lukB"
5614 " inactive_anon:%lukB"
5615 " active_file:%lukB"
5616 " inactive_file:%lukB"
5617 " unevictable:%lukB"
5618 " writepending:%lukB"
5628 K(zone_page_state(zone, NR_FREE_PAGES)),
5629 K(min_wmark_pages(zone)),
5630 K(low_wmark_pages(zone)),
5631 K(high_wmark_pages(zone)),
5632 K(zone->nr_reserved_highatomic),
5633 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5634 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5635 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5636 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5637 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5638 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5639 K(zone->present_pages),
5640 K(zone_managed_pages(zone)),
5641 K(zone_page_state(zone, NR_MLOCK)),
5642 K(zone_page_state(zone, NR_BOUNCE)),
5644 K(this_cpu_read(zone->pageset->pcp.count)),
5645 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5646 printk("lowmem_reserve[]:");
5647 for (i = 0; i < MAX_NR_ZONES; i++)
5648 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5649 printk(KERN_CONT "\n");
5652 for_each_populated_zone(zone) {
5654 unsigned long nr[MAX_ORDER], flags, total = 0;
5655 unsigned char types[MAX_ORDER];
5657 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5660 printk(KERN_CONT "%s: ", zone->name);
5662 spin_lock_irqsave(&zone->lock, flags);
5663 for (order = 0; order < MAX_ORDER; order++) {
5664 struct free_area *area = &zone->free_area[order];
5667 nr[order] = area->nr_free;
5668 total += nr[order] << order;
5671 for (type = 0; type < MIGRATE_TYPES; type++) {
5672 if (!free_area_empty(area, type))
5673 types[order] |= 1 << type;
5676 spin_unlock_irqrestore(&zone->lock, flags);
5677 for (order = 0; order < MAX_ORDER; order++) {
5678 printk(KERN_CONT "%lu*%lukB ",
5679 nr[order], K(1UL) << order);
5681 show_migration_types(types[order]);
5683 printk(KERN_CONT "= %lukB\n", K(total));
5686 hugetlb_show_meminfo();
5688 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5690 show_swap_cache_info();
5693 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5695 zoneref->zone = zone;
5696 zoneref->zone_idx = zone_idx(zone);
5700 * Builds allocation fallback zone lists.
5702 * Add all populated zones of a node to the zonelist.
5704 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5707 enum zone_type zone_type = MAX_NR_ZONES;
5712 zone = pgdat->node_zones + zone_type;
5713 if (managed_zone(zone)) {
5714 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5715 check_highest_zone(zone_type);
5717 } while (zone_type);
5724 static int __parse_numa_zonelist_order(char *s)
5727 * We used to support different zonlists modes but they turned
5728 * out to be just not useful. Let's keep the warning in place
5729 * if somebody still use the cmd line parameter so that we do
5730 * not fail it silently
5732 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5733 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5739 char numa_zonelist_order[] = "Node";
5742 * sysctl handler for numa_zonelist_order
5744 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5745 void *buffer, size_t *length, loff_t *ppos)
5748 return __parse_numa_zonelist_order(buffer);
5749 return proc_dostring(table, write, buffer, length, ppos);
5753 #define MAX_NODE_LOAD (nr_online_nodes)
5754 static int node_load[MAX_NUMNODES];
5757 * find_next_best_node - find the next node that should appear in a given node's fallback list
5758 * @node: node whose fallback list we're appending
5759 * @used_node_mask: nodemask_t of already used nodes
5761 * We use a number of factors to determine which is the next node that should
5762 * appear on a given node's fallback list. The node should not have appeared
5763 * already in @node's fallback list, and it should be the next closest node
5764 * according to the distance array (which contains arbitrary distance values
5765 * from each node to each node in the system), and should also prefer nodes
5766 * with no CPUs, since presumably they'll have very little allocation pressure
5767 * on them otherwise.
5769 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5771 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5774 int min_val = INT_MAX;
5775 int best_node = NUMA_NO_NODE;
5777 /* Use the local node if we haven't already */
5778 if (!node_isset(node, *used_node_mask)) {
5779 node_set(node, *used_node_mask);
5783 for_each_node_state(n, N_MEMORY) {
5785 /* Don't want a node to appear more than once */
5786 if (node_isset(n, *used_node_mask))
5789 /* Use the distance array to find the distance */
5790 val = node_distance(node, n);
5792 /* Penalize nodes under us ("prefer the next node") */
5795 /* Give preference to headless and unused nodes */
5796 if (!cpumask_empty(cpumask_of_node(n)))
5797 val += PENALTY_FOR_NODE_WITH_CPUS;
5799 /* Slight preference for less loaded node */
5800 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5801 val += node_load[n];
5803 if (val < min_val) {
5810 node_set(best_node, *used_node_mask);
5817 * Build zonelists ordered by node and zones within node.
5818 * This results in maximum locality--normal zone overflows into local
5819 * DMA zone, if any--but risks exhausting DMA zone.
5821 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5824 struct zoneref *zonerefs;
5827 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5829 for (i = 0; i < nr_nodes; i++) {
5832 pg_data_t *node = NODE_DATA(node_order[i]);
5834 nr_zones = build_zonerefs_node(node, zonerefs);
5835 zonerefs += nr_zones;
5837 zonerefs->zone = NULL;
5838 zonerefs->zone_idx = 0;
5842 * Build gfp_thisnode zonelists
5844 static void build_thisnode_zonelists(pg_data_t *pgdat)
5846 struct zoneref *zonerefs;
5849 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5850 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5851 zonerefs += nr_zones;
5852 zonerefs->zone = NULL;
5853 zonerefs->zone_idx = 0;
5857 * Build zonelists ordered by zone and nodes within zones.
5858 * This results in conserving DMA zone[s] until all Normal memory is
5859 * exhausted, but results in overflowing to remote node while memory
5860 * may still exist in local DMA zone.
5863 static void build_zonelists(pg_data_t *pgdat)
5865 static int node_order[MAX_NUMNODES];
5866 int node, load, nr_nodes = 0;
5867 nodemask_t used_mask = NODE_MASK_NONE;
5868 int local_node, prev_node;
5870 /* NUMA-aware ordering of nodes */
5871 local_node = pgdat->node_id;
5872 load = nr_online_nodes;
5873 prev_node = local_node;
5875 memset(node_order, 0, sizeof(node_order));
5876 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5878 * We don't want to pressure a particular node.
5879 * So adding penalty to the first node in same
5880 * distance group to make it round-robin.
5882 if (node_distance(local_node, node) !=
5883 node_distance(local_node, prev_node))
5884 node_load[node] = load;
5886 node_order[nr_nodes++] = node;
5891 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5892 build_thisnode_zonelists(pgdat);
5895 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5897 * Return node id of node used for "local" allocations.
5898 * I.e., first node id of first zone in arg node's generic zonelist.
5899 * Used for initializing percpu 'numa_mem', which is used primarily
5900 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5902 int local_memory_node(int node)
5906 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5907 gfp_zone(GFP_KERNEL),
5909 return zone_to_nid(z->zone);
5913 static void setup_min_unmapped_ratio(void);
5914 static void setup_min_slab_ratio(void);
5915 #else /* CONFIG_NUMA */
5917 static void build_zonelists(pg_data_t *pgdat)
5919 int node, local_node;
5920 struct zoneref *zonerefs;
5923 local_node = pgdat->node_id;
5925 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5926 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5927 zonerefs += nr_zones;
5930 * Now we build the zonelist so that it contains the zones
5931 * of all the other nodes.
5932 * We don't want to pressure a particular node, so when
5933 * building the zones for node N, we make sure that the
5934 * zones coming right after the local ones are those from
5935 * node N+1 (modulo N)
5937 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5938 if (!node_online(node))
5940 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5941 zonerefs += nr_zones;
5943 for (node = 0; node < local_node; node++) {
5944 if (!node_online(node))
5946 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5947 zonerefs += nr_zones;
5950 zonerefs->zone = NULL;
5951 zonerefs->zone_idx = 0;
5954 #endif /* CONFIG_NUMA */
5957 * Boot pageset table. One per cpu which is going to be used for all
5958 * zones and all nodes. The parameters will be set in such a way
5959 * that an item put on a list will immediately be handed over to
5960 * the buddy list. This is safe since pageset manipulation is done
5961 * with interrupts disabled.
5963 * The boot_pagesets must be kept even after bootup is complete for
5964 * unused processors and/or zones. They do play a role for bootstrapping
5965 * hotplugged processors.
5967 * zoneinfo_show() and maybe other functions do
5968 * not check if the processor is online before following the pageset pointer.
5969 * Other parts of the kernel may not check if the zone is available.
5971 static void pageset_init(struct per_cpu_pageset *p);
5972 /* These effectively disable the pcplists in the boot pageset completely */
5973 #define BOOT_PAGESET_HIGH 0
5974 #define BOOT_PAGESET_BATCH 1
5975 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5976 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5978 static void __build_all_zonelists(void *data)
5981 int __maybe_unused cpu;
5982 pg_data_t *self = data;
5983 static DEFINE_SPINLOCK(lock);
5988 memset(node_load, 0, sizeof(node_load));
5992 * This node is hotadded and no memory is yet present. So just
5993 * building zonelists is fine - no need to touch other nodes.
5995 if (self && !node_online(self->node_id)) {
5996 build_zonelists(self);
5998 for_each_online_node(nid) {
5999 pg_data_t *pgdat = NODE_DATA(nid);
6001 build_zonelists(pgdat);
6004 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6006 * We now know the "local memory node" for each node--
6007 * i.e., the node of the first zone in the generic zonelist.
6008 * Set up numa_mem percpu variable for on-line cpus. During
6009 * boot, only the boot cpu should be on-line; we'll init the
6010 * secondary cpus' numa_mem as they come on-line. During
6011 * node/memory hotplug, we'll fixup all on-line cpus.
6013 for_each_online_cpu(cpu)
6014 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6021 static noinline void __init
6022 build_all_zonelists_init(void)
6026 __build_all_zonelists(NULL);
6029 * Initialize the boot_pagesets that are going to be used
6030 * for bootstrapping processors. The real pagesets for
6031 * each zone will be allocated later when the per cpu
6032 * allocator is available.
6034 * boot_pagesets are used also for bootstrapping offline
6035 * cpus if the system is already booted because the pagesets
6036 * are needed to initialize allocators on a specific cpu too.
6037 * F.e. the percpu allocator needs the page allocator which
6038 * needs the percpu allocator in order to allocate its pagesets
6039 * (a chicken-egg dilemma).
6041 for_each_possible_cpu(cpu)
6042 pageset_init(&per_cpu(boot_pageset, cpu));
6044 mminit_verify_zonelist();
6045 cpuset_init_current_mems_allowed();
6049 * unless system_state == SYSTEM_BOOTING.
6051 * __ref due to call of __init annotated helper build_all_zonelists_init
6052 * [protected by SYSTEM_BOOTING].
6054 void __ref build_all_zonelists(pg_data_t *pgdat)
6056 unsigned long vm_total_pages;
6058 if (system_state == SYSTEM_BOOTING) {
6059 build_all_zonelists_init();
6061 __build_all_zonelists(pgdat);
6062 /* cpuset refresh routine should be here */
6064 /* Get the number of free pages beyond high watermark in all zones. */
6065 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6067 * Disable grouping by mobility if the number of pages in the
6068 * system is too low to allow the mechanism to work. It would be
6069 * more accurate, but expensive to check per-zone. This check is
6070 * made on memory-hotadd so a system can start with mobility
6071 * disabled and enable it later
6073 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6074 page_group_by_mobility_disabled = 1;
6076 page_group_by_mobility_disabled = 0;
6078 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6080 page_group_by_mobility_disabled ? "off" : "on",
6083 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6087 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6088 static bool __meminit
6089 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6091 static struct memblock_region *r;
6093 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6094 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6095 for_each_mem_region(r) {
6096 if (*pfn < memblock_region_memory_end_pfn(r))
6100 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6101 memblock_is_mirror(r)) {
6102 *pfn = memblock_region_memory_end_pfn(r);
6110 * Initially all pages are reserved - free ones are freed
6111 * up by memblock_free_all() once the early boot process is
6112 * done. Non-atomic initialization, single-pass.
6114 * All aligned pageblocks are initialized to the specified migratetype
6115 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6116 * zone stats (e.g., nr_isolate_pageblock) are touched.
6118 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
6119 unsigned long start_pfn,
6120 enum meminit_context context,
6121 struct vmem_altmap *altmap, int migratetype)
6123 unsigned long pfn, end_pfn = start_pfn + size;
6126 if (highest_memmap_pfn < end_pfn - 1)
6127 highest_memmap_pfn = end_pfn - 1;
6129 #ifdef CONFIG_ZONE_DEVICE
6131 * Honor reservation requested by the driver for this ZONE_DEVICE
6132 * memory. We limit the total number of pages to initialize to just
6133 * those that might contain the memory mapping. We will defer the
6134 * ZONE_DEVICE page initialization until after we have released
6137 if (zone == ZONE_DEVICE) {
6141 if (start_pfn == altmap->base_pfn)
6142 start_pfn += altmap->reserve;
6143 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6147 for (pfn = start_pfn; pfn < end_pfn; ) {
6149 * There can be holes in boot-time mem_map[]s handed to this
6150 * function. They do not exist on hotplugged memory.
6152 if (context == MEMINIT_EARLY) {
6153 if (overlap_memmap_init(zone, &pfn))
6155 if (defer_init(nid, pfn, end_pfn))
6159 page = pfn_to_page(pfn);
6160 __init_single_page(page, pfn, zone, nid);
6161 if (context == MEMINIT_HOTPLUG)
6162 __SetPageReserved(page);
6165 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6166 * such that unmovable allocations won't be scattered all
6167 * over the place during system boot.
6169 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6170 set_pageblock_migratetype(page, migratetype);
6177 #ifdef CONFIG_ZONE_DEVICE
6178 void __ref memmap_init_zone_device(struct zone *zone,
6179 unsigned long start_pfn,
6180 unsigned long nr_pages,
6181 struct dev_pagemap *pgmap)
6183 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6184 struct pglist_data *pgdat = zone->zone_pgdat;
6185 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6186 unsigned long zone_idx = zone_idx(zone);
6187 unsigned long start = jiffies;
6188 int nid = pgdat->node_id;
6190 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6194 * The call to memmap_init_zone should have already taken care
6195 * of the pages reserved for the memmap, so we can just jump to
6196 * the end of that region and start processing the device pages.
6199 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6200 nr_pages = end_pfn - start_pfn;
6203 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6204 struct page *page = pfn_to_page(pfn);
6206 __init_single_page(page, pfn, zone_idx, nid);
6209 * Mark page reserved as it will need to wait for onlining
6210 * phase for it to be fully associated with a zone.
6212 * We can use the non-atomic __set_bit operation for setting
6213 * the flag as we are still initializing the pages.
6215 __SetPageReserved(page);
6218 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6219 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6220 * ever freed or placed on a driver-private list.
6222 page->pgmap = pgmap;
6223 page->zone_device_data = NULL;
6226 * Mark the block movable so that blocks are reserved for
6227 * movable at startup. This will force kernel allocations
6228 * to reserve their blocks rather than leaking throughout
6229 * the address space during boot when many long-lived
6230 * kernel allocations are made.
6232 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6233 * because this is done early in section_activate()
6235 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6236 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6241 pr_info("%s initialised %lu pages in %ums\n", __func__,
6242 nr_pages, jiffies_to_msecs(jiffies - start));
6246 static void __meminit zone_init_free_lists(struct zone *zone)
6248 unsigned int order, t;
6249 for_each_migratetype_order(order, t) {
6250 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6251 zone->free_area[order].nr_free = 0;
6255 void __meminit __weak memmap_init(unsigned long size, int nid,
6257 unsigned long range_start_pfn)
6259 unsigned long start_pfn, end_pfn;
6260 unsigned long range_end_pfn = range_start_pfn + size;
6263 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6264 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6265 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6267 if (end_pfn > start_pfn) {
6268 size = end_pfn - start_pfn;
6269 memmap_init_zone(size, nid, zone, start_pfn,
6270 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6275 static int zone_batchsize(struct zone *zone)
6281 * The per-cpu-pages pools are set to around 1000th of the
6284 batch = zone_managed_pages(zone) / 1024;
6285 /* But no more than a meg. */
6286 if (batch * PAGE_SIZE > 1024 * 1024)
6287 batch = (1024 * 1024) / PAGE_SIZE;
6288 batch /= 4; /* We effectively *= 4 below */
6293 * Clamp the batch to a 2^n - 1 value. Having a power
6294 * of 2 value was found to be more likely to have
6295 * suboptimal cache aliasing properties in some cases.
6297 * For example if 2 tasks are alternately allocating
6298 * batches of pages, one task can end up with a lot
6299 * of pages of one half of the possible page colors
6300 * and the other with pages of the other colors.
6302 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6307 /* The deferral and batching of frees should be suppressed under NOMMU
6310 * The problem is that NOMMU needs to be able to allocate large chunks
6311 * of contiguous memory as there's no hardware page translation to
6312 * assemble apparent contiguous memory from discontiguous pages.
6314 * Queueing large contiguous runs of pages for batching, however,
6315 * causes the pages to actually be freed in smaller chunks. As there
6316 * can be a significant delay between the individual batches being
6317 * recycled, this leads to the once large chunks of space being
6318 * fragmented and becoming unavailable for high-order allocations.
6325 * pcp->high and pcp->batch values are related and generally batch is lower
6326 * than high. They are also related to pcp->count such that count is lower
6327 * than high, and as soon as it reaches high, the pcplist is flushed.
6329 * However, guaranteeing these relations at all times would require e.g. write
6330 * barriers here but also careful usage of read barriers at the read side, and
6331 * thus be prone to error and bad for performance. Thus the update only prevents
6332 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6333 * can cope with those fields changing asynchronously, and fully trust only the
6334 * pcp->count field on the local CPU with interrupts disabled.
6336 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6337 * outside of boot time (or some other assurance that no concurrent updaters
6340 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6341 unsigned long batch)
6343 WRITE_ONCE(pcp->batch, batch);
6344 WRITE_ONCE(pcp->high, high);
6347 static void pageset_init(struct per_cpu_pageset *p)
6349 struct per_cpu_pages *pcp;
6352 memset(p, 0, sizeof(*p));
6355 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6356 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6359 * Set batch and high values safe for a boot pageset. A true percpu
6360 * pageset's initialization will update them subsequently. Here we don't
6361 * need to be as careful as pageset_update() as nobody can access the
6364 pcp->high = BOOT_PAGESET_HIGH;
6365 pcp->batch = BOOT_PAGESET_BATCH;
6368 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6369 unsigned long batch)
6371 struct per_cpu_pageset *p;
6374 for_each_possible_cpu(cpu) {
6375 p = per_cpu_ptr(zone->pageset, cpu);
6376 pageset_update(&p->pcp, high, batch);
6381 * Calculate and set new high and batch values for all per-cpu pagesets of a
6382 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6384 static void zone_set_pageset_high_and_batch(struct zone *zone)
6386 unsigned long new_high, new_batch;
6388 if (percpu_pagelist_fraction) {
6389 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6390 new_batch = max(1UL, new_high / 4);
6391 if ((new_high / 4) > (PAGE_SHIFT * 8))
6392 new_batch = PAGE_SHIFT * 8;
6394 new_batch = zone_batchsize(zone);
6395 new_high = 6 * new_batch;
6396 new_batch = max(1UL, 1 * new_batch);
6399 if (zone->pageset_high == new_high &&
6400 zone->pageset_batch == new_batch)
6403 zone->pageset_high = new_high;
6404 zone->pageset_batch = new_batch;
6406 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6409 void __meminit setup_zone_pageset(struct zone *zone)
6411 struct per_cpu_pageset *p;
6414 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6415 for_each_possible_cpu(cpu) {
6416 p = per_cpu_ptr(zone->pageset, cpu);
6420 zone_set_pageset_high_and_batch(zone);
6424 * Allocate per cpu pagesets and initialize them.
6425 * Before this call only boot pagesets were available.
6427 void __init setup_per_cpu_pageset(void)
6429 struct pglist_data *pgdat;
6431 int __maybe_unused cpu;
6433 for_each_populated_zone(zone)
6434 setup_zone_pageset(zone);
6438 * Unpopulated zones continue using the boot pagesets.
6439 * The numa stats for these pagesets need to be reset.
6440 * Otherwise, they will end up skewing the stats of
6441 * the nodes these zones are associated with.
6443 for_each_possible_cpu(cpu) {
6444 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6445 memset(pcp->vm_numa_stat_diff, 0,
6446 sizeof(pcp->vm_numa_stat_diff));
6450 for_each_online_pgdat(pgdat)
6451 pgdat->per_cpu_nodestats =
6452 alloc_percpu(struct per_cpu_nodestat);
6455 static __meminit void zone_pcp_init(struct zone *zone)
6458 * per cpu subsystem is not up at this point. The following code
6459 * relies on the ability of the linker to provide the
6460 * offset of a (static) per cpu variable into the per cpu area.
6462 zone->pageset = &boot_pageset;
6463 zone->pageset_high = BOOT_PAGESET_HIGH;
6464 zone->pageset_batch = BOOT_PAGESET_BATCH;
6466 if (populated_zone(zone))
6467 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6468 zone->name, zone->present_pages,
6469 zone_batchsize(zone));
6472 void __meminit init_currently_empty_zone(struct zone *zone,
6473 unsigned long zone_start_pfn,
6476 struct pglist_data *pgdat = zone->zone_pgdat;
6477 int zone_idx = zone_idx(zone) + 1;
6479 if (zone_idx > pgdat->nr_zones)
6480 pgdat->nr_zones = zone_idx;
6482 zone->zone_start_pfn = zone_start_pfn;
6484 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6485 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6487 (unsigned long)zone_idx(zone),
6488 zone_start_pfn, (zone_start_pfn + size));
6490 zone_init_free_lists(zone);
6491 zone->initialized = 1;
6495 * get_pfn_range_for_nid - Return the start and end page frames for a node
6496 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6497 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6498 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6500 * It returns the start and end page frame of a node based on information
6501 * provided by memblock_set_node(). If called for a node
6502 * with no available memory, a warning is printed and the start and end
6505 void __init get_pfn_range_for_nid(unsigned int nid,
6506 unsigned long *start_pfn, unsigned long *end_pfn)
6508 unsigned long this_start_pfn, this_end_pfn;
6514 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6515 *start_pfn = min(*start_pfn, this_start_pfn);
6516 *end_pfn = max(*end_pfn, this_end_pfn);
6519 if (*start_pfn == -1UL)
6524 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6525 * assumption is made that zones within a node are ordered in monotonic
6526 * increasing memory addresses so that the "highest" populated zone is used
6528 static void __init find_usable_zone_for_movable(void)
6531 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6532 if (zone_index == ZONE_MOVABLE)
6535 if (arch_zone_highest_possible_pfn[zone_index] >
6536 arch_zone_lowest_possible_pfn[zone_index])
6540 VM_BUG_ON(zone_index == -1);
6541 movable_zone = zone_index;
6545 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6546 * because it is sized independent of architecture. Unlike the other zones,
6547 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6548 * in each node depending on the size of each node and how evenly kernelcore
6549 * is distributed. This helper function adjusts the zone ranges
6550 * provided by the architecture for a given node by using the end of the
6551 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6552 * zones within a node are in order of monotonic increases memory addresses
6554 static void __init adjust_zone_range_for_zone_movable(int nid,
6555 unsigned long zone_type,
6556 unsigned long node_start_pfn,
6557 unsigned long node_end_pfn,
6558 unsigned long *zone_start_pfn,
6559 unsigned long *zone_end_pfn)
6561 /* Only adjust if ZONE_MOVABLE is on this node */
6562 if (zone_movable_pfn[nid]) {
6563 /* Size ZONE_MOVABLE */
6564 if (zone_type == ZONE_MOVABLE) {
6565 *zone_start_pfn = zone_movable_pfn[nid];
6566 *zone_end_pfn = min(node_end_pfn,
6567 arch_zone_highest_possible_pfn[movable_zone]);
6569 /* Adjust for ZONE_MOVABLE starting within this range */
6570 } else if (!mirrored_kernelcore &&
6571 *zone_start_pfn < zone_movable_pfn[nid] &&
6572 *zone_end_pfn > zone_movable_pfn[nid]) {
6573 *zone_end_pfn = zone_movable_pfn[nid];
6575 /* Check if this whole range is within ZONE_MOVABLE */
6576 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6577 *zone_start_pfn = *zone_end_pfn;
6582 * Return the number of pages a zone spans in a node, including holes
6583 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6585 static unsigned long __init zone_spanned_pages_in_node(int nid,
6586 unsigned long zone_type,
6587 unsigned long node_start_pfn,
6588 unsigned long node_end_pfn,
6589 unsigned long *zone_start_pfn,
6590 unsigned long *zone_end_pfn)
6592 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6593 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6594 /* When hotadd a new node from cpu_up(), the node should be empty */
6595 if (!node_start_pfn && !node_end_pfn)
6598 /* Get the start and end of the zone */
6599 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6600 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6601 adjust_zone_range_for_zone_movable(nid, zone_type,
6602 node_start_pfn, node_end_pfn,
6603 zone_start_pfn, zone_end_pfn);
6605 /* Check that this node has pages within the zone's required range */
6606 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6609 /* Move the zone boundaries inside the node if necessary */
6610 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6611 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6613 /* Return the spanned pages */
6614 return *zone_end_pfn - *zone_start_pfn;
6618 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6619 * then all holes in the requested range will be accounted for.
6621 unsigned long __init __absent_pages_in_range(int nid,
6622 unsigned long range_start_pfn,
6623 unsigned long range_end_pfn)
6625 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6626 unsigned long start_pfn, end_pfn;
6629 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6630 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6631 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6632 nr_absent -= end_pfn - start_pfn;
6638 * absent_pages_in_range - Return number of page frames in holes within a range
6639 * @start_pfn: The start PFN to start searching for holes
6640 * @end_pfn: The end PFN to stop searching for holes
6642 * Return: the number of pages frames in memory holes within a range.
6644 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6645 unsigned long end_pfn)
6647 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6650 /* Return the number of page frames in holes in a zone on a node */
6651 static unsigned long __init zone_absent_pages_in_node(int nid,
6652 unsigned long zone_type,
6653 unsigned long node_start_pfn,
6654 unsigned long node_end_pfn)
6656 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6657 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6658 unsigned long zone_start_pfn, zone_end_pfn;
6659 unsigned long nr_absent;
6661 /* When hotadd a new node from cpu_up(), the node should be empty */
6662 if (!node_start_pfn && !node_end_pfn)
6665 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6666 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6668 adjust_zone_range_for_zone_movable(nid, zone_type,
6669 node_start_pfn, node_end_pfn,
6670 &zone_start_pfn, &zone_end_pfn);
6671 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6674 * ZONE_MOVABLE handling.
6675 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6678 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6679 unsigned long start_pfn, end_pfn;
6680 struct memblock_region *r;
6682 for_each_mem_region(r) {
6683 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6684 zone_start_pfn, zone_end_pfn);
6685 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6686 zone_start_pfn, zone_end_pfn);
6688 if (zone_type == ZONE_MOVABLE &&
6689 memblock_is_mirror(r))
6690 nr_absent += end_pfn - start_pfn;
6692 if (zone_type == ZONE_NORMAL &&
6693 !memblock_is_mirror(r))
6694 nr_absent += end_pfn - start_pfn;
6701 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6702 unsigned long node_start_pfn,
6703 unsigned long node_end_pfn)
6705 unsigned long realtotalpages = 0, totalpages = 0;
6708 for (i = 0; i < MAX_NR_ZONES; i++) {
6709 struct zone *zone = pgdat->node_zones + i;
6710 unsigned long zone_start_pfn, zone_end_pfn;
6711 unsigned long spanned, absent;
6712 unsigned long size, real_size;
6714 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6719 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6724 real_size = size - absent;
6727 zone->zone_start_pfn = zone_start_pfn;
6729 zone->zone_start_pfn = 0;
6730 zone->spanned_pages = size;
6731 zone->present_pages = real_size;
6734 realtotalpages += real_size;
6737 pgdat->node_spanned_pages = totalpages;
6738 pgdat->node_present_pages = realtotalpages;
6739 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6743 #ifndef CONFIG_SPARSEMEM
6745 * Calculate the size of the zone->blockflags rounded to an unsigned long
6746 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6747 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6748 * round what is now in bits to nearest long in bits, then return it in
6751 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6753 unsigned long usemapsize;
6755 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6756 usemapsize = roundup(zonesize, pageblock_nr_pages);
6757 usemapsize = usemapsize >> pageblock_order;
6758 usemapsize *= NR_PAGEBLOCK_BITS;
6759 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6761 return usemapsize / 8;
6764 static void __ref setup_usemap(struct pglist_data *pgdat,
6766 unsigned long zone_start_pfn,
6767 unsigned long zonesize)
6769 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6770 zone->pageblock_flags = NULL;
6772 zone->pageblock_flags =
6773 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6775 if (!zone->pageblock_flags)
6776 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6777 usemapsize, zone->name, pgdat->node_id);
6781 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6782 unsigned long zone_start_pfn, unsigned long zonesize) {}
6783 #endif /* CONFIG_SPARSEMEM */
6785 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6787 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6788 void __init set_pageblock_order(void)
6792 /* Check that pageblock_nr_pages has not already been setup */
6793 if (pageblock_order)
6796 if (HPAGE_SHIFT > PAGE_SHIFT)
6797 order = HUGETLB_PAGE_ORDER;
6799 order = MAX_ORDER - 1;
6802 * Assume the largest contiguous order of interest is a huge page.
6803 * This value may be variable depending on boot parameters on IA64 and
6806 pageblock_order = order;
6808 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6811 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6812 * is unused as pageblock_order is set at compile-time. See
6813 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6816 void __init set_pageblock_order(void)
6820 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6822 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6823 unsigned long present_pages)
6825 unsigned long pages = spanned_pages;
6828 * Provide a more accurate estimation if there are holes within
6829 * the zone and SPARSEMEM is in use. If there are holes within the
6830 * zone, each populated memory region may cost us one or two extra
6831 * memmap pages due to alignment because memmap pages for each
6832 * populated regions may not be naturally aligned on page boundary.
6833 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6835 if (spanned_pages > present_pages + (present_pages >> 4) &&
6836 IS_ENABLED(CONFIG_SPARSEMEM))
6837 pages = present_pages;
6839 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6842 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6843 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6845 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6847 spin_lock_init(&ds_queue->split_queue_lock);
6848 INIT_LIST_HEAD(&ds_queue->split_queue);
6849 ds_queue->split_queue_len = 0;
6852 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6855 #ifdef CONFIG_COMPACTION
6856 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6858 init_waitqueue_head(&pgdat->kcompactd_wait);
6861 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6864 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6866 pgdat_resize_init(pgdat);
6868 pgdat_init_split_queue(pgdat);
6869 pgdat_init_kcompactd(pgdat);
6871 init_waitqueue_head(&pgdat->kswapd_wait);
6872 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6874 pgdat_page_ext_init(pgdat);
6875 lruvec_init(&pgdat->__lruvec);
6878 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6879 unsigned long remaining_pages)
6881 atomic_long_set(&zone->managed_pages, remaining_pages);
6882 zone_set_nid(zone, nid);
6883 zone->name = zone_names[idx];
6884 zone->zone_pgdat = NODE_DATA(nid);
6885 spin_lock_init(&zone->lock);
6886 zone_seqlock_init(zone);
6887 zone_pcp_init(zone);
6891 * Set up the zone data structures
6892 * - init pgdat internals
6893 * - init all zones belonging to this node
6895 * NOTE: this function is only called during memory hotplug
6897 #ifdef CONFIG_MEMORY_HOTPLUG
6898 void __ref free_area_init_core_hotplug(int nid)
6901 pg_data_t *pgdat = NODE_DATA(nid);
6903 pgdat_init_internals(pgdat);
6904 for (z = 0; z < MAX_NR_ZONES; z++)
6905 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6910 * Set up the zone data structures:
6911 * - mark all pages reserved
6912 * - mark all memory queues empty
6913 * - clear the memory bitmaps
6915 * NOTE: pgdat should get zeroed by caller.
6916 * NOTE: this function is only called during early init.
6918 static void __init free_area_init_core(struct pglist_data *pgdat)
6921 int nid = pgdat->node_id;
6923 pgdat_init_internals(pgdat);
6924 pgdat->per_cpu_nodestats = &boot_nodestats;
6926 for (j = 0; j < MAX_NR_ZONES; j++) {
6927 struct zone *zone = pgdat->node_zones + j;
6928 unsigned long size, freesize, memmap_pages;
6929 unsigned long zone_start_pfn = zone->zone_start_pfn;
6931 size = zone->spanned_pages;
6932 freesize = zone->present_pages;
6935 * Adjust freesize so that it accounts for how much memory
6936 * is used by this zone for memmap. This affects the watermark
6937 * and per-cpu initialisations
6939 memmap_pages = calc_memmap_size(size, freesize);
6940 if (!is_highmem_idx(j)) {
6941 if (freesize >= memmap_pages) {
6942 freesize -= memmap_pages;
6945 " %s zone: %lu pages used for memmap\n",
6946 zone_names[j], memmap_pages);
6948 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6949 zone_names[j], memmap_pages, freesize);
6952 /* Account for reserved pages */
6953 if (j == 0 && freesize > dma_reserve) {
6954 freesize -= dma_reserve;
6955 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6956 zone_names[0], dma_reserve);
6959 if (!is_highmem_idx(j))
6960 nr_kernel_pages += freesize;
6961 /* Charge for highmem memmap if there are enough kernel pages */
6962 else if (nr_kernel_pages > memmap_pages * 2)
6963 nr_kernel_pages -= memmap_pages;
6964 nr_all_pages += freesize;
6967 * Set an approximate value for lowmem here, it will be adjusted
6968 * when the bootmem allocator frees pages into the buddy system.
6969 * And all highmem pages will be managed by the buddy system.
6971 zone_init_internals(zone, j, nid, freesize);
6976 set_pageblock_order();
6977 setup_usemap(pgdat, zone, zone_start_pfn, size);
6978 init_currently_empty_zone(zone, zone_start_pfn, size);
6979 memmap_init(size, nid, j, zone_start_pfn);
6983 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6984 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6986 unsigned long __maybe_unused start = 0;
6987 unsigned long __maybe_unused offset = 0;
6989 /* Skip empty nodes */
6990 if (!pgdat->node_spanned_pages)
6993 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6994 offset = pgdat->node_start_pfn - start;
6995 /* ia64 gets its own node_mem_map, before this, without bootmem */
6996 if (!pgdat->node_mem_map) {
6997 unsigned long size, end;
7001 * The zone's endpoints aren't required to be MAX_ORDER
7002 * aligned but the node_mem_map endpoints must be in order
7003 * for the buddy allocator to function correctly.
7005 end = pgdat_end_pfn(pgdat);
7006 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7007 size = (end - start) * sizeof(struct page);
7008 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7011 panic("Failed to allocate %ld bytes for node %d memory map\n",
7012 size, pgdat->node_id);
7013 pgdat->node_mem_map = map + offset;
7015 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7016 __func__, pgdat->node_id, (unsigned long)pgdat,
7017 (unsigned long)pgdat->node_mem_map);
7018 #ifndef CONFIG_NEED_MULTIPLE_NODES
7020 * With no DISCONTIG, the global mem_map is just set as node 0's
7022 if (pgdat == NODE_DATA(0)) {
7023 mem_map = NODE_DATA(0)->node_mem_map;
7024 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7030 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7031 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7033 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7034 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7036 pgdat->first_deferred_pfn = ULONG_MAX;
7039 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7042 static void __init free_area_init_node(int nid)
7044 pg_data_t *pgdat = NODE_DATA(nid);
7045 unsigned long start_pfn = 0;
7046 unsigned long end_pfn = 0;
7048 /* pg_data_t should be reset to zero when it's allocated */
7049 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7051 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7053 pgdat->node_id = nid;
7054 pgdat->node_start_pfn = start_pfn;
7055 pgdat->per_cpu_nodestats = NULL;
7057 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7058 (u64)start_pfn << PAGE_SHIFT,
7059 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7060 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7062 alloc_node_mem_map(pgdat);
7063 pgdat_set_deferred_range(pgdat);
7065 free_area_init_core(pgdat);
7068 void __init free_area_init_memoryless_node(int nid)
7070 free_area_init_node(nid);
7073 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
7075 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
7076 * PageReserved(). Return the number of struct pages that were initialized.
7078 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
7083 for (pfn = spfn; pfn < epfn; pfn++) {
7084 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
7085 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
7086 + pageblock_nr_pages - 1;
7090 * Use a fake node/zone (0) for now. Some of these pages
7091 * (in memblock.reserved but not in memblock.memory) will
7092 * get re-initialized via reserve_bootmem_region() later.
7094 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7095 __SetPageReserved(pfn_to_page(pfn));
7103 * Only struct pages that are backed by physical memory are zeroed and
7104 * initialized by going through __init_single_page(). But, there are some
7105 * struct pages which are reserved in memblock allocator and their fields
7106 * may be accessed (for example page_to_pfn() on some configuration accesses
7107 * flags). We must explicitly initialize those struct pages.
7109 * This function also addresses a similar issue where struct pages are left
7110 * uninitialized because the physical address range is not covered by
7111 * memblock.memory or memblock.reserved. That could happen when memblock
7112 * layout is manually configured via memmap=, or when the highest physical
7113 * address (max_pfn) does not end on a section boundary.
7115 static void __init init_unavailable_mem(void)
7117 phys_addr_t start, end;
7119 phys_addr_t next = 0;
7122 * Loop through unavailable ranges not covered by memblock.memory.
7125 for_each_mem_range(i, &start, &end) {
7127 pgcnt += init_unavailable_range(PFN_DOWN(next),
7133 * Early sections always have a fully populated memmap for the whole
7134 * section - see pfn_valid(). If the last section has holes at the
7135 * end and that section is marked "online", the memmap will be
7136 * considered initialized. Make sure that memmap has a well defined
7139 pgcnt += init_unavailable_range(PFN_DOWN(next),
7140 round_up(max_pfn, PAGES_PER_SECTION));
7143 * Struct pages that do not have backing memory. This could be because
7144 * firmware is using some of this memory, or for some other reasons.
7147 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7150 static inline void __init init_unavailable_mem(void)
7153 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7155 #if MAX_NUMNODES > 1
7157 * Figure out the number of possible node ids.
7159 void __init setup_nr_node_ids(void)
7161 unsigned int highest;
7163 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7164 nr_node_ids = highest + 1;
7169 * node_map_pfn_alignment - determine the maximum internode alignment
7171 * This function should be called after node map is populated and sorted.
7172 * It calculates the maximum power of two alignment which can distinguish
7175 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7176 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7177 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7178 * shifted, 1GiB is enough and this function will indicate so.
7180 * This is used to test whether pfn -> nid mapping of the chosen memory
7181 * model has fine enough granularity to avoid incorrect mapping for the
7182 * populated node map.
7184 * Return: the determined alignment in pfn's. 0 if there is no alignment
7185 * requirement (single node).
7187 unsigned long __init node_map_pfn_alignment(void)
7189 unsigned long accl_mask = 0, last_end = 0;
7190 unsigned long start, end, mask;
7191 int last_nid = NUMA_NO_NODE;
7194 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7195 if (!start || last_nid < 0 || last_nid == nid) {
7202 * Start with a mask granular enough to pin-point to the
7203 * start pfn and tick off bits one-by-one until it becomes
7204 * too coarse to separate the current node from the last.
7206 mask = ~((1 << __ffs(start)) - 1);
7207 while (mask && last_end <= (start & (mask << 1)))
7210 /* accumulate all internode masks */
7214 /* convert mask to number of pages */
7215 return ~accl_mask + 1;
7219 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7221 * Return: the minimum PFN based on information provided via
7222 * memblock_set_node().
7224 unsigned long __init find_min_pfn_with_active_regions(void)
7226 return PHYS_PFN(memblock_start_of_DRAM());
7230 * early_calculate_totalpages()
7231 * Sum pages in active regions for movable zone.
7232 * Populate N_MEMORY for calculating usable_nodes.
7234 static unsigned long __init early_calculate_totalpages(void)
7236 unsigned long totalpages = 0;
7237 unsigned long start_pfn, end_pfn;
7240 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7241 unsigned long pages = end_pfn - start_pfn;
7243 totalpages += pages;
7245 node_set_state(nid, N_MEMORY);
7251 * Find the PFN the Movable zone begins in each node. Kernel memory
7252 * is spread evenly between nodes as long as the nodes have enough
7253 * memory. When they don't, some nodes will have more kernelcore than
7256 static void __init find_zone_movable_pfns_for_nodes(void)
7259 unsigned long usable_startpfn;
7260 unsigned long kernelcore_node, kernelcore_remaining;
7261 /* save the state before borrow the nodemask */
7262 nodemask_t saved_node_state = node_states[N_MEMORY];
7263 unsigned long totalpages = early_calculate_totalpages();
7264 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7265 struct memblock_region *r;
7267 /* Need to find movable_zone earlier when movable_node is specified. */
7268 find_usable_zone_for_movable();
7271 * If movable_node is specified, ignore kernelcore and movablecore
7274 if (movable_node_is_enabled()) {
7275 for_each_mem_region(r) {
7276 if (!memblock_is_hotpluggable(r))
7279 nid = memblock_get_region_node(r);
7281 usable_startpfn = PFN_DOWN(r->base);
7282 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7283 min(usable_startpfn, zone_movable_pfn[nid]) :
7291 * If kernelcore=mirror is specified, ignore movablecore option
7293 if (mirrored_kernelcore) {
7294 bool mem_below_4gb_not_mirrored = false;
7296 for_each_mem_region(r) {
7297 if (memblock_is_mirror(r))
7300 nid = memblock_get_region_node(r);
7302 usable_startpfn = memblock_region_memory_base_pfn(r);
7304 if (usable_startpfn < 0x100000) {
7305 mem_below_4gb_not_mirrored = true;
7309 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7310 min(usable_startpfn, zone_movable_pfn[nid]) :
7314 if (mem_below_4gb_not_mirrored)
7315 pr_warn("This configuration results in unmirrored kernel memory.\n");
7321 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7322 * amount of necessary memory.
7324 if (required_kernelcore_percent)
7325 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7327 if (required_movablecore_percent)
7328 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7332 * If movablecore= was specified, calculate what size of
7333 * kernelcore that corresponds so that memory usable for
7334 * any allocation type is evenly spread. If both kernelcore
7335 * and movablecore are specified, then the value of kernelcore
7336 * will be used for required_kernelcore if it's greater than
7337 * what movablecore would have allowed.
7339 if (required_movablecore) {
7340 unsigned long corepages;
7343 * Round-up so that ZONE_MOVABLE is at least as large as what
7344 * was requested by the user
7346 required_movablecore =
7347 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7348 required_movablecore = min(totalpages, required_movablecore);
7349 corepages = totalpages - required_movablecore;
7351 required_kernelcore = max(required_kernelcore, corepages);
7355 * If kernelcore was not specified or kernelcore size is larger
7356 * than totalpages, there is no ZONE_MOVABLE.
7358 if (!required_kernelcore || required_kernelcore >= totalpages)
7361 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7362 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7365 /* Spread kernelcore memory as evenly as possible throughout nodes */
7366 kernelcore_node = required_kernelcore / usable_nodes;
7367 for_each_node_state(nid, N_MEMORY) {
7368 unsigned long start_pfn, end_pfn;
7371 * Recalculate kernelcore_node if the division per node
7372 * now exceeds what is necessary to satisfy the requested
7373 * amount of memory for the kernel
7375 if (required_kernelcore < kernelcore_node)
7376 kernelcore_node = required_kernelcore / usable_nodes;
7379 * As the map is walked, we track how much memory is usable
7380 * by the kernel using kernelcore_remaining. When it is
7381 * 0, the rest of the node is usable by ZONE_MOVABLE
7383 kernelcore_remaining = kernelcore_node;
7385 /* Go through each range of PFNs within this node */
7386 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7387 unsigned long size_pages;
7389 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7390 if (start_pfn >= end_pfn)
7393 /* Account for what is only usable for kernelcore */
7394 if (start_pfn < usable_startpfn) {
7395 unsigned long kernel_pages;
7396 kernel_pages = min(end_pfn, usable_startpfn)
7399 kernelcore_remaining -= min(kernel_pages,
7400 kernelcore_remaining);
7401 required_kernelcore -= min(kernel_pages,
7402 required_kernelcore);
7404 /* Continue if range is now fully accounted */
7405 if (end_pfn <= usable_startpfn) {
7408 * Push zone_movable_pfn to the end so
7409 * that if we have to rebalance
7410 * kernelcore across nodes, we will
7411 * not double account here
7413 zone_movable_pfn[nid] = end_pfn;
7416 start_pfn = usable_startpfn;
7420 * The usable PFN range for ZONE_MOVABLE is from
7421 * start_pfn->end_pfn. Calculate size_pages as the
7422 * number of pages used as kernelcore
7424 size_pages = end_pfn - start_pfn;
7425 if (size_pages > kernelcore_remaining)
7426 size_pages = kernelcore_remaining;
7427 zone_movable_pfn[nid] = start_pfn + size_pages;
7430 * Some kernelcore has been met, update counts and
7431 * break if the kernelcore for this node has been
7434 required_kernelcore -= min(required_kernelcore,
7436 kernelcore_remaining -= size_pages;
7437 if (!kernelcore_remaining)
7443 * If there is still required_kernelcore, we do another pass with one
7444 * less node in the count. This will push zone_movable_pfn[nid] further
7445 * along on the nodes that still have memory until kernelcore is
7449 if (usable_nodes && required_kernelcore > usable_nodes)
7453 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7454 for (nid = 0; nid < MAX_NUMNODES; nid++)
7455 zone_movable_pfn[nid] =
7456 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7459 /* restore the node_state */
7460 node_states[N_MEMORY] = saved_node_state;
7463 /* Any regular or high memory on that node ? */
7464 static void check_for_memory(pg_data_t *pgdat, int nid)
7466 enum zone_type zone_type;
7468 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7469 struct zone *zone = &pgdat->node_zones[zone_type];
7470 if (populated_zone(zone)) {
7471 if (IS_ENABLED(CONFIG_HIGHMEM))
7472 node_set_state(nid, N_HIGH_MEMORY);
7473 if (zone_type <= ZONE_NORMAL)
7474 node_set_state(nid, N_NORMAL_MEMORY);
7481 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7482 * such cases we allow max_zone_pfn sorted in the descending order
7484 bool __weak arch_has_descending_max_zone_pfns(void)
7490 * free_area_init - Initialise all pg_data_t and zone data
7491 * @max_zone_pfn: an array of max PFNs for each zone
7493 * This will call free_area_init_node() for each active node in the system.
7494 * Using the page ranges provided by memblock_set_node(), the size of each
7495 * zone in each node and their holes is calculated. If the maximum PFN
7496 * between two adjacent zones match, it is assumed that the zone is empty.
7497 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7498 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7499 * starts where the previous one ended. For example, ZONE_DMA32 starts
7500 * at arch_max_dma_pfn.
7502 void __init free_area_init(unsigned long *max_zone_pfn)
7504 unsigned long start_pfn, end_pfn;
7508 /* Record where the zone boundaries are */
7509 memset(arch_zone_lowest_possible_pfn, 0,
7510 sizeof(arch_zone_lowest_possible_pfn));
7511 memset(arch_zone_highest_possible_pfn, 0,
7512 sizeof(arch_zone_highest_possible_pfn));
7514 start_pfn = find_min_pfn_with_active_regions();
7515 descending = arch_has_descending_max_zone_pfns();
7517 for (i = 0; i < MAX_NR_ZONES; i++) {
7519 zone = MAX_NR_ZONES - i - 1;
7523 if (zone == ZONE_MOVABLE)
7526 end_pfn = max(max_zone_pfn[zone], start_pfn);
7527 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7528 arch_zone_highest_possible_pfn[zone] = end_pfn;
7530 start_pfn = end_pfn;
7533 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7534 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7535 find_zone_movable_pfns_for_nodes();
7537 /* Print out the zone ranges */
7538 pr_info("Zone ranges:\n");
7539 for (i = 0; i < MAX_NR_ZONES; i++) {
7540 if (i == ZONE_MOVABLE)
7542 pr_info(" %-8s ", zone_names[i]);
7543 if (arch_zone_lowest_possible_pfn[i] ==
7544 arch_zone_highest_possible_pfn[i])
7547 pr_cont("[mem %#018Lx-%#018Lx]\n",
7548 (u64)arch_zone_lowest_possible_pfn[i]
7550 ((u64)arch_zone_highest_possible_pfn[i]
7551 << PAGE_SHIFT) - 1);
7554 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7555 pr_info("Movable zone start for each node\n");
7556 for (i = 0; i < MAX_NUMNODES; i++) {
7557 if (zone_movable_pfn[i])
7558 pr_info(" Node %d: %#018Lx\n", i,
7559 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7563 * Print out the early node map, and initialize the
7564 * subsection-map relative to active online memory ranges to
7565 * enable future "sub-section" extensions of the memory map.
7567 pr_info("Early memory node ranges\n");
7568 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7569 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7570 (u64)start_pfn << PAGE_SHIFT,
7571 ((u64)end_pfn << PAGE_SHIFT) - 1);
7572 subsection_map_init(start_pfn, end_pfn - start_pfn);
7575 /* Initialise every node */
7576 mminit_verify_pageflags_layout();
7577 setup_nr_node_ids();
7578 init_unavailable_mem();
7579 for_each_online_node(nid) {
7580 pg_data_t *pgdat = NODE_DATA(nid);
7581 free_area_init_node(nid);
7583 /* Any memory on that node */
7584 if (pgdat->node_present_pages)
7585 node_set_state(nid, N_MEMORY);
7586 check_for_memory(pgdat, nid);
7590 static int __init cmdline_parse_core(char *p, unsigned long *core,
7591 unsigned long *percent)
7593 unsigned long long coremem;
7599 /* Value may be a percentage of total memory, otherwise bytes */
7600 coremem = simple_strtoull(p, &endptr, 0);
7601 if (*endptr == '%') {
7602 /* Paranoid check for percent values greater than 100 */
7603 WARN_ON(coremem > 100);
7607 coremem = memparse(p, &p);
7608 /* Paranoid check that UL is enough for the coremem value */
7609 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7611 *core = coremem >> PAGE_SHIFT;
7618 * kernelcore=size sets the amount of memory for use for allocations that
7619 * cannot be reclaimed or migrated.
7621 static int __init cmdline_parse_kernelcore(char *p)
7623 /* parse kernelcore=mirror */
7624 if (parse_option_str(p, "mirror")) {
7625 mirrored_kernelcore = true;
7629 return cmdline_parse_core(p, &required_kernelcore,
7630 &required_kernelcore_percent);
7634 * movablecore=size sets the amount of memory for use for allocations that
7635 * can be reclaimed or migrated.
7637 static int __init cmdline_parse_movablecore(char *p)
7639 return cmdline_parse_core(p, &required_movablecore,
7640 &required_movablecore_percent);
7643 early_param("kernelcore", cmdline_parse_kernelcore);
7644 early_param("movablecore", cmdline_parse_movablecore);
7646 void adjust_managed_page_count(struct page *page, long count)
7648 atomic_long_add(count, &page_zone(page)->managed_pages);
7649 totalram_pages_add(count);
7650 #ifdef CONFIG_HIGHMEM
7651 if (PageHighMem(page))
7652 totalhigh_pages_add(count);
7655 EXPORT_SYMBOL(adjust_managed_page_count);
7657 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7660 unsigned long pages = 0;
7662 start = (void *)PAGE_ALIGN((unsigned long)start);
7663 end = (void *)((unsigned long)end & PAGE_MASK);
7664 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7665 struct page *page = virt_to_page(pos);
7666 void *direct_map_addr;
7669 * 'direct_map_addr' might be different from 'pos'
7670 * because some architectures' virt_to_page()
7671 * work with aliases. Getting the direct map
7672 * address ensures that we get a _writeable_
7673 * alias for the memset().
7675 direct_map_addr = page_address(page);
7677 * Perform a kasan-unchecked memset() since this memory
7678 * has not been initialized.
7680 direct_map_addr = kasan_reset_tag(direct_map_addr);
7681 if ((unsigned int)poison <= 0xFF)
7682 memset(direct_map_addr, poison, PAGE_SIZE);
7684 free_reserved_page(page);
7688 pr_info("Freeing %s memory: %ldK\n",
7689 s, pages << (PAGE_SHIFT - 10));
7694 #ifdef CONFIG_HIGHMEM
7695 void free_highmem_page(struct page *page)
7697 __free_reserved_page(page);
7698 totalram_pages_inc();
7699 atomic_long_inc(&page_zone(page)->managed_pages);
7700 totalhigh_pages_inc();
7705 void __init mem_init_print_info(const char *str)
7707 unsigned long physpages, codesize, datasize, rosize, bss_size;
7708 unsigned long init_code_size, init_data_size;
7710 physpages = get_num_physpages();
7711 codesize = _etext - _stext;
7712 datasize = _edata - _sdata;
7713 rosize = __end_rodata - __start_rodata;
7714 bss_size = __bss_stop - __bss_start;
7715 init_data_size = __init_end - __init_begin;
7716 init_code_size = _einittext - _sinittext;
7719 * Detect special cases and adjust section sizes accordingly:
7720 * 1) .init.* may be embedded into .data sections
7721 * 2) .init.text.* may be out of [__init_begin, __init_end],
7722 * please refer to arch/tile/kernel/vmlinux.lds.S.
7723 * 3) .rodata.* may be embedded into .text or .data sections.
7725 #define adj_init_size(start, end, size, pos, adj) \
7727 if (start <= pos && pos < end && size > adj) \
7731 adj_init_size(__init_begin, __init_end, init_data_size,
7732 _sinittext, init_code_size);
7733 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7734 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7735 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7736 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7738 #undef adj_init_size
7740 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7741 #ifdef CONFIG_HIGHMEM
7745 nr_free_pages() << (PAGE_SHIFT - 10),
7746 physpages << (PAGE_SHIFT - 10),
7747 codesize >> 10, datasize >> 10, rosize >> 10,
7748 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7749 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7750 totalcma_pages << (PAGE_SHIFT - 10),
7751 #ifdef CONFIG_HIGHMEM
7752 totalhigh_pages() << (PAGE_SHIFT - 10),
7754 str ? ", " : "", str ? str : "");
7758 * set_dma_reserve - set the specified number of pages reserved in the first zone
7759 * @new_dma_reserve: The number of pages to mark reserved
7761 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7762 * In the DMA zone, a significant percentage may be consumed by kernel image
7763 * and other unfreeable allocations which can skew the watermarks badly. This
7764 * function may optionally be used to account for unfreeable pages in the
7765 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7766 * smaller per-cpu batchsize.
7768 void __init set_dma_reserve(unsigned long new_dma_reserve)
7770 dma_reserve = new_dma_reserve;
7773 static int page_alloc_cpu_dead(unsigned int cpu)
7776 lru_add_drain_cpu(cpu);
7780 * Spill the event counters of the dead processor
7781 * into the current processors event counters.
7782 * This artificially elevates the count of the current
7785 vm_events_fold_cpu(cpu);
7788 * Zero the differential counters of the dead processor
7789 * so that the vm statistics are consistent.
7791 * This is only okay since the processor is dead and cannot
7792 * race with what we are doing.
7794 cpu_vm_stats_fold(cpu);
7799 int hashdist = HASHDIST_DEFAULT;
7801 static int __init set_hashdist(char *str)
7805 hashdist = simple_strtoul(str, &str, 0);
7808 __setup("hashdist=", set_hashdist);
7811 void __init page_alloc_init(void)
7816 if (num_node_state(N_MEMORY) == 1)
7820 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7821 "mm/page_alloc:dead", NULL,
7822 page_alloc_cpu_dead);
7827 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7828 * or min_free_kbytes changes.
7830 static void calculate_totalreserve_pages(void)
7832 struct pglist_data *pgdat;
7833 unsigned long reserve_pages = 0;
7834 enum zone_type i, j;
7836 for_each_online_pgdat(pgdat) {
7838 pgdat->totalreserve_pages = 0;
7840 for (i = 0; i < MAX_NR_ZONES; i++) {
7841 struct zone *zone = pgdat->node_zones + i;
7843 unsigned long managed_pages = zone_managed_pages(zone);
7845 /* Find valid and maximum lowmem_reserve in the zone */
7846 for (j = i; j < MAX_NR_ZONES; j++) {
7847 if (zone->lowmem_reserve[j] > max)
7848 max = zone->lowmem_reserve[j];
7851 /* we treat the high watermark as reserved pages. */
7852 max += high_wmark_pages(zone);
7854 if (max > managed_pages)
7855 max = managed_pages;
7857 pgdat->totalreserve_pages += max;
7859 reserve_pages += max;
7862 totalreserve_pages = reserve_pages;
7866 * setup_per_zone_lowmem_reserve - called whenever
7867 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7868 * has a correct pages reserved value, so an adequate number of
7869 * pages are left in the zone after a successful __alloc_pages().
7871 static void setup_per_zone_lowmem_reserve(void)
7873 struct pglist_data *pgdat;
7874 enum zone_type i, j;
7876 for_each_online_pgdat(pgdat) {
7877 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
7878 struct zone *zone = &pgdat->node_zones[i];
7879 int ratio = sysctl_lowmem_reserve_ratio[i];
7880 bool clear = !ratio || !zone_managed_pages(zone);
7881 unsigned long managed_pages = 0;
7883 for (j = i + 1; j < MAX_NR_ZONES; j++) {
7885 zone->lowmem_reserve[j] = 0;
7887 struct zone *upper_zone = &pgdat->node_zones[j];
7889 managed_pages += zone_managed_pages(upper_zone);
7890 zone->lowmem_reserve[j] = managed_pages / ratio;
7896 /* update totalreserve_pages */
7897 calculate_totalreserve_pages();
7900 static void __setup_per_zone_wmarks(void)
7902 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7903 unsigned long lowmem_pages = 0;
7905 unsigned long flags;
7907 /* Calculate total number of !ZONE_HIGHMEM pages */
7908 for_each_zone(zone) {
7909 if (!is_highmem(zone))
7910 lowmem_pages += zone_managed_pages(zone);
7913 for_each_zone(zone) {
7916 spin_lock_irqsave(&zone->lock, flags);
7917 tmp = (u64)pages_min * zone_managed_pages(zone);
7918 do_div(tmp, lowmem_pages);
7919 if (is_highmem(zone)) {
7921 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7922 * need highmem pages, so cap pages_min to a small
7925 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7926 * deltas control async page reclaim, and so should
7927 * not be capped for highmem.
7929 unsigned long min_pages;
7931 min_pages = zone_managed_pages(zone) / 1024;
7932 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7933 zone->_watermark[WMARK_MIN] = min_pages;
7936 * If it's a lowmem zone, reserve a number of pages
7937 * proportionate to the zone's size.
7939 zone->_watermark[WMARK_MIN] = tmp;
7943 * Set the kswapd watermarks distance according to the
7944 * scale factor in proportion to available memory, but
7945 * ensure a minimum size on small systems.
7947 tmp = max_t(u64, tmp >> 2,
7948 mult_frac(zone_managed_pages(zone),
7949 watermark_scale_factor, 10000));
7951 zone->watermark_boost = 0;
7952 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7953 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7955 spin_unlock_irqrestore(&zone->lock, flags);
7958 /* update totalreserve_pages */
7959 calculate_totalreserve_pages();
7963 * setup_per_zone_wmarks - called when min_free_kbytes changes
7964 * or when memory is hot-{added|removed}
7966 * Ensures that the watermark[min,low,high] values for each zone are set
7967 * correctly with respect to min_free_kbytes.
7969 void setup_per_zone_wmarks(void)
7971 static DEFINE_SPINLOCK(lock);
7974 __setup_per_zone_wmarks();
7979 * Initialise min_free_kbytes.
7981 * For small machines we want it small (128k min). For large machines
7982 * we want it large (256MB max). But it is not linear, because network
7983 * bandwidth does not increase linearly with machine size. We use
7985 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7986 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8002 int __meminit init_per_zone_wmark_min(void)
8004 unsigned long lowmem_kbytes;
8005 int new_min_free_kbytes;
8007 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8008 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8010 if (new_min_free_kbytes > user_min_free_kbytes) {
8011 min_free_kbytes = new_min_free_kbytes;
8012 if (min_free_kbytes < 128)
8013 min_free_kbytes = 128;
8014 if (min_free_kbytes > 262144)
8015 min_free_kbytes = 262144;
8017 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8018 new_min_free_kbytes, user_min_free_kbytes);
8020 setup_per_zone_wmarks();
8021 refresh_zone_stat_thresholds();
8022 setup_per_zone_lowmem_reserve();
8025 setup_min_unmapped_ratio();
8026 setup_min_slab_ratio();
8029 khugepaged_min_free_kbytes_update();
8033 postcore_initcall(init_per_zone_wmark_min)
8036 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8037 * that we can call two helper functions whenever min_free_kbytes
8040 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8041 void *buffer, size_t *length, loff_t *ppos)
8045 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8050 user_min_free_kbytes = min_free_kbytes;
8051 setup_per_zone_wmarks();
8056 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8057 void *buffer, size_t *length, loff_t *ppos)
8061 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8066 setup_per_zone_wmarks();
8072 static void setup_min_unmapped_ratio(void)
8077 for_each_online_pgdat(pgdat)
8078 pgdat->min_unmapped_pages = 0;
8081 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8082 sysctl_min_unmapped_ratio) / 100;
8086 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8087 void *buffer, size_t *length, loff_t *ppos)
8091 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8095 setup_min_unmapped_ratio();
8100 static void setup_min_slab_ratio(void)
8105 for_each_online_pgdat(pgdat)
8106 pgdat->min_slab_pages = 0;
8109 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8110 sysctl_min_slab_ratio) / 100;
8113 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8114 void *buffer, size_t *length, loff_t *ppos)
8118 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8122 setup_min_slab_ratio();
8129 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8130 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8131 * whenever sysctl_lowmem_reserve_ratio changes.
8133 * The reserve ratio obviously has absolutely no relation with the
8134 * minimum watermarks. The lowmem reserve ratio can only make sense
8135 * if in function of the boot time zone sizes.
8137 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8138 void *buffer, size_t *length, loff_t *ppos)
8142 proc_dointvec_minmax(table, write, buffer, length, ppos);
8144 for (i = 0; i < MAX_NR_ZONES; i++) {
8145 if (sysctl_lowmem_reserve_ratio[i] < 1)
8146 sysctl_lowmem_reserve_ratio[i] = 0;
8149 setup_per_zone_lowmem_reserve();
8154 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8155 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8156 * pagelist can have before it gets flushed back to buddy allocator.
8158 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8159 void *buffer, size_t *length, loff_t *ppos)
8162 int old_percpu_pagelist_fraction;
8165 mutex_lock(&pcp_batch_high_lock);
8166 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8168 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8169 if (!write || ret < 0)
8172 /* Sanity checking to avoid pcp imbalance */
8173 if (percpu_pagelist_fraction &&
8174 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8175 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8181 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8184 for_each_populated_zone(zone)
8185 zone_set_pageset_high_and_batch(zone);
8187 mutex_unlock(&pcp_batch_high_lock);
8191 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8193 * Returns the number of pages that arch has reserved but
8194 * is not known to alloc_large_system_hash().
8196 static unsigned long __init arch_reserved_kernel_pages(void)
8203 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8204 * machines. As memory size is increased the scale is also increased but at
8205 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8206 * quadruples the scale is increased by one, which means the size of hash table
8207 * only doubles, instead of quadrupling as well.
8208 * Because 32-bit systems cannot have large physical memory, where this scaling
8209 * makes sense, it is disabled on such platforms.
8211 #if __BITS_PER_LONG > 32
8212 #define ADAPT_SCALE_BASE (64ul << 30)
8213 #define ADAPT_SCALE_SHIFT 2
8214 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8218 * allocate a large system hash table from bootmem
8219 * - it is assumed that the hash table must contain an exact power-of-2
8220 * quantity of entries
8221 * - limit is the number of hash buckets, not the total allocation size
8223 void *__init alloc_large_system_hash(const char *tablename,
8224 unsigned long bucketsize,
8225 unsigned long numentries,
8228 unsigned int *_hash_shift,
8229 unsigned int *_hash_mask,
8230 unsigned long low_limit,
8231 unsigned long high_limit)
8233 unsigned long long max = high_limit;
8234 unsigned long log2qty, size;
8239 /* allow the kernel cmdline to have a say */
8241 /* round applicable memory size up to nearest megabyte */
8242 numentries = nr_kernel_pages;
8243 numentries -= arch_reserved_kernel_pages();
8245 /* It isn't necessary when PAGE_SIZE >= 1MB */
8246 if (PAGE_SHIFT < 20)
8247 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8249 #if __BITS_PER_LONG > 32
8251 unsigned long adapt;
8253 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8254 adapt <<= ADAPT_SCALE_SHIFT)
8259 /* limit to 1 bucket per 2^scale bytes of low memory */
8260 if (scale > PAGE_SHIFT)
8261 numentries >>= (scale - PAGE_SHIFT);
8263 numentries <<= (PAGE_SHIFT - scale);
8265 /* Make sure we've got at least a 0-order allocation.. */
8266 if (unlikely(flags & HASH_SMALL)) {
8267 /* Makes no sense without HASH_EARLY */
8268 WARN_ON(!(flags & HASH_EARLY));
8269 if (!(numentries >> *_hash_shift)) {
8270 numentries = 1UL << *_hash_shift;
8271 BUG_ON(!numentries);
8273 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8274 numentries = PAGE_SIZE / bucketsize;
8276 numentries = roundup_pow_of_two(numentries);
8278 /* limit allocation size to 1/16 total memory by default */
8280 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8281 do_div(max, bucketsize);
8283 max = min(max, 0x80000000ULL);
8285 if (numentries < low_limit)
8286 numentries = low_limit;
8287 if (numentries > max)
8290 log2qty = ilog2(numentries);
8292 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8295 size = bucketsize << log2qty;
8296 if (flags & HASH_EARLY) {
8297 if (flags & HASH_ZERO)
8298 table = memblock_alloc(size, SMP_CACHE_BYTES);
8300 table = memblock_alloc_raw(size,
8302 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8303 table = __vmalloc(size, gfp_flags);
8307 * If bucketsize is not a power-of-two, we may free
8308 * some pages at the end of hash table which
8309 * alloc_pages_exact() automatically does
8311 table = alloc_pages_exact(size, gfp_flags);
8312 kmemleak_alloc(table, size, 1, gfp_flags);
8314 } while (!table && size > PAGE_SIZE && --log2qty);
8317 panic("Failed to allocate %s hash table\n", tablename);
8319 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8320 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8321 virt ? "vmalloc" : "linear");
8324 *_hash_shift = log2qty;
8326 *_hash_mask = (1 << log2qty) - 1;
8332 * This function checks whether pageblock includes unmovable pages or not.
8334 * PageLRU check without isolation or lru_lock could race so that
8335 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8336 * check without lock_page also may miss some movable non-lru pages at
8337 * race condition. So you can't expect this function should be exact.
8339 * Returns a page without holding a reference. If the caller wants to
8340 * dereference that page (e.g., dumping), it has to make sure that it
8341 * cannot get removed (e.g., via memory unplug) concurrently.
8344 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8345 int migratetype, int flags)
8347 unsigned long iter = 0;
8348 unsigned long pfn = page_to_pfn(page);
8349 unsigned long offset = pfn % pageblock_nr_pages;
8351 if (is_migrate_cma_page(page)) {
8353 * CMA allocations (alloc_contig_range) really need to mark
8354 * isolate CMA pageblocks even when they are not movable in fact
8355 * so consider them movable here.
8357 if (is_migrate_cma(migratetype))
8363 for (; iter < pageblock_nr_pages - offset; iter++) {
8364 if (!pfn_valid_within(pfn + iter))
8367 page = pfn_to_page(pfn + iter);
8370 * Both, bootmem allocations and memory holes are marked
8371 * PG_reserved and are unmovable. We can even have unmovable
8372 * allocations inside ZONE_MOVABLE, for example when
8373 * specifying "movablecore".
8375 if (PageReserved(page))
8379 * If the zone is movable and we have ruled out all reserved
8380 * pages then it should be reasonably safe to assume the rest
8383 if (zone_idx(zone) == ZONE_MOVABLE)
8387 * Hugepages are not in LRU lists, but they're movable.
8388 * THPs are on the LRU, but need to be counted as #small pages.
8389 * We need not scan over tail pages because we don't
8390 * handle each tail page individually in migration.
8392 if (PageHuge(page) || PageTransCompound(page)) {
8393 struct page *head = compound_head(page);
8394 unsigned int skip_pages;
8396 if (PageHuge(page)) {
8397 if (!hugepage_migration_supported(page_hstate(head)))
8399 } else if (!PageLRU(head) && !__PageMovable(head)) {
8403 skip_pages = compound_nr(head) - (page - head);
8404 iter += skip_pages - 1;
8409 * We can't use page_count without pin a page
8410 * because another CPU can free compound page.
8411 * This check already skips compound tails of THP
8412 * because their page->_refcount is zero at all time.
8414 if (!page_ref_count(page)) {
8415 if (PageBuddy(page))
8416 iter += (1 << buddy_order(page)) - 1;
8421 * The HWPoisoned page may be not in buddy system, and
8422 * page_count() is not 0.
8424 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8428 * We treat all PageOffline() pages as movable when offlining
8429 * to give drivers a chance to decrement their reference count
8430 * in MEM_GOING_OFFLINE in order to indicate that these pages
8431 * can be offlined as there are no direct references anymore.
8432 * For actually unmovable PageOffline() where the driver does
8433 * not support this, we will fail later when trying to actually
8434 * move these pages that still have a reference count > 0.
8435 * (false negatives in this function only)
8437 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8440 if (__PageMovable(page) || PageLRU(page))
8444 * If there are RECLAIMABLE pages, we need to check
8445 * it. But now, memory offline itself doesn't call
8446 * shrink_node_slabs() and it still to be fixed.
8453 #ifdef CONFIG_CONTIG_ALLOC
8454 static unsigned long pfn_max_align_down(unsigned long pfn)
8456 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8457 pageblock_nr_pages) - 1);
8460 static unsigned long pfn_max_align_up(unsigned long pfn)
8462 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8463 pageblock_nr_pages));
8466 /* [start, end) must belong to a single zone. */
8467 static int __alloc_contig_migrate_range(struct compact_control *cc,
8468 unsigned long start, unsigned long end)
8470 /* This function is based on compact_zone() from compaction.c. */
8471 unsigned int nr_reclaimed;
8472 unsigned long pfn = start;
8473 unsigned int tries = 0;
8475 struct migration_target_control mtc = {
8476 .nid = zone_to_nid(cc->zone),
8477 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8482 while (pfn < end || !list_empty(&cc->migratepages)) {
8483 if (fatal_signal_pending(current)) {
8488 if (list_empty(&cc->migratepages)) {
8489 cc->nr_migratepages = 0;
8490 pfn = isolate_migratepages_range(cc, pfn, end);
8496 } else if (++tries == 5) {
8497 ret = ret < 0 ? ret : -EBUSY;
8501 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8503 cc->nr_migratepages -= nr_reclaimed;
8505 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8506 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8509 putback_movable_pages(&cc->migratepages);
8516 * alloc_contig_range() -- tries to allocate given range of pages
8517 * @start: start PFN to allocate
8518 * @end: one-past-the-last PFN to allocate
8519 * @migratetype: migratetype of the underlaying pageblocks (either
8520 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8521 * in range must have the same migratetype and it must
8522 * be either of the two.
8523 * @gfp_mask: GFP mask to use during compaction
8525 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8526 * aligned. The PFN range must belong to a single zone.
8528 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8529 * pageblocks in the range. Once isolated, the pageblocks should not
8530 * be modified by others.
8532 * Return: zero on success or negative error code. On success all
8533 * pages which PFN is in [start, end) are allocated for the caller and
8534 * need to be freed with free_contig_range().
8536 int alloc_contig_range(unsigned long start, unsigned long end,
8537 unsigned migratetype, gfp_t gfp_mask)
8539 unsigned long outer_start, outer_end;
8543 struct compact_control cc = {
8544 .nr_migratepages = 0,
8546 .zone = page_zone(pfn_to_page(start)),
8547 .mode = MIGRATE_SYNC,
8548 .ignore_skip_hint = true,
8549 .no_set_skip_hint = true,
8550 .gfp_mask = current_gfp_context(gfp_mask),
8551 .alloc_contig = true,
8553 INIT_LIST_HEAD(&cc.migratepages);
8556 * What we do here is we mark all pageblocks in range as
8557 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8558 * have different sizes, and due to the way page allocator
8559 * work, we align the range to biggest of the two pages so
8560 * that page allocator won't try to merge buddies from
8561 * different pageblocks and change MIGRATE_ISOLATE to some
8562 * other migration type.
8564 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8565 * migrate the pages from an unaligned range (ie. pages that
8566 * we are interested in). This will put all the pages in
8567 * range back to page allocator as MIGRATE_ISOLATE.
8569 * When this is done, we take the pages in range from page
8570 * allocator removing them from the buddy system. This way
8571 * page allocator will never consider using them.
8573 * This lets us mark the pageblocks back as
8574 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8575 * aligned range but not in the unaligned, original range are
8576 * put back to page allocator so that buddy can use them.
8579 ret = start_isolate_page_range(pfn_max_align_down(start),
8580 pfn_max_align_up(end), migratetype, 0);
8584 drain_all_pages(cc.zone);
8587 * In case of -EBUSY, we'd like to know which page causes problem.
8588 * So, just fall through. test_pages_isolated() has a tracepoint
8589 * which will report the busy page.
8591 * It is possible that busy pages could become available before
8592 * the call to test_pages_isolated, and the range will actually be
8593 * allocated. So, if we fall through be sure to clear ret so that
8594 * -EBUSY is not accidentally used or returned to caller.
8596 ret = __alloc_contig_migrate_range(&cc, start, end);
8597 if (ret && ret != -EBUSY)
8602 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8603 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8604 * more, all pages in [start, end) are free in page allocator.
8605 * What we are going to do is to allocate all pages from
8606 * [start, end) (that is remove them from page allocator).
8608 * The only problem is that pages at the beginning and at the
8609 * end of interesting range may be not aligned with pages that
8610 * page allocator holds, ie. they can be part of higher order
8611 * pages. Because of this, we reserve the bigger range and
8612 * once this is done free the pages we are not interested in.
8614 * We don't have to hold zone->lock here because the pages are
8615 * isolated thus they won't get removed from buddy.
8618 lru_add_drain_all();
8621 outer_start = start;
8622 while (!PageBuddy(pfn_to_page(outer_start))) {
8623 if (++order >= MAX_ORDER) {
8624 outer_start = start;
8627 outer_start &= ~0UL << order;
8630 if (outer_start != start) {
8631 order = buddy_order(pfn_to_page(outer_start));
8634 * outer_start page could be small order buddy page and
8635 * it doesn't include start page. Adjust outer_start
8636 * in this case to report failed page properly
8637 * on tracepoint in test_pages_isolated()
8639 if (outer_start + (1UL << order) <= start)
8640 outer_start = start;
8643 /* Make sure the range is really isolated. */
8644 if (test_pages_isolated(outer_start, end, 0)) {
8645 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8646 __func__, outer_start, end);
8651 /* Grab isolated pages from freelists. */
8652 outer_end = isolate_freepages_range(&cc, outer_start, end);
8658 /* Free head and tail (if any) */
8659 if (start != outer_start)
8660 free_contig_range(outer_start, start - outer_start);
8661 if (end != outer_end)
8662 free_contig_range(end, outer_end - end);
8665 undo_isolate_page_range(pfn_max_align_down(start),
8666 pfn_max_align_up(end), migratetype);
8669 EXPORT_SYMBOL(alloc_contig_range);
8671 static int __alloc_contig_pages(unsigned long start_pfn,
8672 unsigned long nr_pages, gfp_t gfp_mask)
8674 unsigned long end_pfn = start_pfn + nr_pages;
8676 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8680 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8681 unsigned long nr_pages)
8683 unsigned long i, end_pfn = start_pfn + nr_pages;
8686 for (i = start_pfn; i < end_pfn; i++) {
8687 page = pfn_to_online_page(i);
8691 if (page_zone(page) != z)
8694 if (PageReserved(page))
8697 if (page_count(page) > 0)
8706 static bool zone_spans_last_pfn(const struct zone *zone,
8707 unsigned long start_pfn, unsigned long nr_pages)
8709 unsigned long last_pfn = start_pfn + nr_pages - 1;
8711 return zone_spans_pfn(zone, last_pfn);
8715 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8716 * @nr_pages: Number of contiguous pages to allocate
8717 * @gfp_mask: GFP mask to limit search and used during compaction
8719 * @nodemask: Mask for other possible nodes
8721 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8722 * on an applicable zonelist to find a contiguous pfn range which can then be
8723 * tried for allocation with alloc_contig_range(). This routine is intended
8724 * for allocation requests which can not be fulfilled with the buddy allocator.
8726 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8727 * power of two then the alignment is guaranteed to be to the given nr_pages
8728 * (e.g. 1GB request would be aligned to 1GB).
8730 * Allocated pages can be freed with free_contig_range() or by manually calling
8731 * __free_page() on each allocated page.
8733 * Return: pointer to contiguous pages on success, or NULL if not successful.
8735 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8736 int nid, nodemask_t *nodemask)
8738 unsigned long ret, pfn, flags;
8739 struct zonelist *zonelist;
8743 zonelist = node_zonelist(nid, gfp_mask);
8744 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8745 gfp_zone(gfp_mask), nodemask) {
8746 spin_lock_irqsave(&zone->lock, flags);
8748 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8749 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8750 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8752 * We release the zone lock here because
8753 * alloc_contig_range() will also lock the zone
8754 * at some point. If there's an allocation
8755 * spinning on this lock, it may win the race
8756 * and cause alloc_contig_range() to fail...
8758 spin_unlock_irqrestore(&zone->lock, flags);
8759 ret = __alloc_contig_pages(pfn, nr_pages,
8762 return pfn_to_page(pfn);
8763 spin_lock_irqsave(&zone->lock, flags);
8767 spin_unlock_irqrestore(&zone->lock, flags);
8771 #endif /* CONFIG_CONTIG_ALLOC */
8773 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8775 unsigned int count = 0;
8777 for (; nr_pages--; pfn++) {
8778 struct page *page = pfn_to_page(pfn);
8780 count += page_count(page) != 1;
8783 WARN(count != 0, "%d pages are still in use!\n", count);
8785 EXPORT_SYMBOL(free_contig_range);
8788 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8789 * page high values need to be recalulated.
8791 void __meminit zone_pcp_update(struct zone *zone)
8793 mutex_lock(&pcp_batch_high_lock);
8794 zone_set_pageset_high_and_batch(zone);
8795 mutex_unlock(&pcp_batch_high_lock);
8799 * Effectively disable pcplists for the zone by setting the high limit to 0
8800 * and draining all cpus. A concurrent page freeing on another CPU that's about
8801 * to put the page on pcplist will either finish before the drain and the page
8802 * will be drained, or observe the new high limit and skip the pcplist.
8804 * Must be paired with a call to zone_pcp_enable().
8806 void zone_pcp_disable(struct zone *zone)
8808 mutex_lock(&pcp_batch_high_lock);
8809 __zone_set_pageset_high_and_batch(zone, 0, 1);
8810 __drain_all_pages(zone, true);
8813 void zone_pcp_enable(struct zone *zone)
8815 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
8816 mutex_unlock(&pcp_batch_high_lock);
8819 void zone_pcp_reset(struct zone *zone)
8821 unsigned long flags;
8823 struct per_cpu_pageset *pset;
8825 /* avoid races with drain_pages() */
8826 local_irq_save(flags);
8827 if (zone->pageset != &boot_pageset) {
8828 for_each_online_cpu(cpu) {
8829 pset = per_cpu_ptr(zone->pageset, cpu);
8830 drain_zonestat(zone, pset);
8832 free_percpu(zone->pageset);
8833 zone->pageset = &boot_pageset;
8835 local_irq_restore(flags);
8838 #ifdef CONFIG_MEMORY_HOTREMOVE
8840 * All pages in the range must be in a single zone, must not contain holes,
8841 * must span full sections, and must be isolated before calling this function.
8843 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8845 unsigned long pfn = start_pfn;
8849 unsigned long flags;
8851 offline_mem_sections(pfn, end_pfn);
8852 zone = page_zone(pfn_to_page(pfn));
8853 spin_lock_irqsave(&zone->lock, flags);
8854 while (pfn < end_pfn) {
8855 page = pfn_to_page(pfn);
8857 * The HWPoisoned page may be not in buddy system, and
8858 * page_count() is not 0.
8860 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8865 * At this point all remaining PageOffline() pages have a
8866 * reference count of 0 and can simply be skipped.
8868 if (PageOffline(page)) {
8869 BUG_ON(page_count(page));
8870 BUG_ON(PageBuddy(page));
8875 BUG_ON(page_count(page));
8876 BUG_ON(!PageBuddy(page));
8877 order = buddy_order(page);
8878 del_page_from_free_list(page, zone, order);
8879 pfn += (1 << order);
8881 spin_unlock_irqrestore(&zone->lock, flags);
8885 bool is_free_buddy_page(struct page *page)
8887 struct zone *zone = page_zone(page);
8888 unsigned long pfn = page_to_pfn(page);
8889 unsigned long flags;
8892 spin_lock_irqsave(&zone->lock, flags);
8893 for (order = 0; order < MAX_ORDER; order++) {
8894 struct page *page_head = page - (pfn & ((1 << order) - 1));
8896 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
8899 spin_unlock_irqrestore(&zone->lock, flags);
8901 return order < MAX_ORDER;
8904 #ifdef CONFIG_MEMORY_FAILURE
8906 * Break down a higher-order page in sub-pages, and keep our target out of
8909 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8910 struct page *target, int low, int high,
8913 unsigned long size = 1 << high;
8914 struct page *current_buddy, *next_page;
8916 while (high > low) {
8920 if (target >= &page[size]) {
8921 next_page = page + size;
8922 current_buddy = page;
8925 current_buddy = page + size;
8928 if (set_page_guard(zone, current_buddy, high, migratetype))
8931 if (current_buddy != target) {
8932 add_to_free_list(current_buddy, zone, high, migratetype);
8933 set_buddy_order(current_buddy, high);
8940 * Take a page that will be marked as poisoned off the buddy allocator.
8942 bool take_page_off_buddy(struct page *page)
8944 struct zone *zone = page_zone(page);
8945 unsigned long pfn = page_to_pfn(page);
8946 unsigned long flags;
8950 spin_lock_irqsave(&zone->lock, flags);
8951 for (order = 0; order < MAX_ORDER; order++) {
8952 struct page *page_head = page - (pfn & ((1 << order) - 1));
8953 int page_order = buddy_order(page_head);
8955 if (PageBuddy(page_head) && page_order >= order) {
8956 unsigned long pfn_head = page_to_pfn(page_head);
8957 int migratetype = get_pfnblock_migratetype(page_head,
8960 del_page_from_free_list(page_head, zone, page_order);
8961 break_down_buddy_pages(zone, page_head, page, 0,
8962 page_order, migratetype);
8966 if (page_count(page_head) > 0)
8969 spin_unlock_irqrestore(&zone->lock, flags);