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)))
426 if (NODE_DATA(nid)->first_deferred_pfn != ULONG_MAX)
429 * We start only with one section of pages, more pages are added as
430 * needed until the rest of deferred pages are initialized.
433 if ((nr_initialised > PAGES_PER_SECTION) &&
434 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
435 NODE_DATA(nid)->first_deferred_pfn = pfn;
441 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
443 static inline bool early_page_uninitialised(unsigned long pfn)
448 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
454 /* Return a pointer to the bitmap storing bits affecting a block of pages */
455 static inline unsigned long *get_pageblock_bitmap(struct page *page,
458 #ifdef CONFIG_SPARSEMEM
459 return section_to_usemap(__pfn_to_section(pfn));
461 return page_zone(page)->pageblock_flags;
462 #endif /* CONFIG_SPARSEMEM */
465 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
467 #ifdef CONFIG_SPARSEMEM
468 pfn &= (PAGES_PER_SECTION-1);
470 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
471 #endif /* CONFIG_SPARSEMEM */
472 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
475 static __always_inline
476 unsigned long __get_pfnblock_flags_mask(struct page *page,
480 unsigned long *bitmap;
481 unsigned long bitidx, word_bitidx;
484 bitmap = get_pageblock_bitmap(page, pfn);
485 bitidx = pfn_to_bitidx(page, pfn);
486 word_bitidx = bitidx / BITS_PER_LONG;
487 bitidx &= (BITS_PER_LONG-1);
489 word = bitmap[word_bitidx];
490 return (word >> bitidx) & mask;
494 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
495 * @page: The page within the block of interest
496 * @pfn: The target page frame number
497 * @mask: mask of bits that the caller is interested in
499 * Return: pageblock_bits flags
501 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
504 return __get_pfnblock_flags_mask(page, pfn, mask);
507 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
509 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
513 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
514 * @page: The page within the block of interest
515 * @flags: The flags to set
516 * @pfn: The target page frame number
517 * @mask: mask of bits that the caller is interested in
519 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
523 unsigned long *bitmap;
524 unsigned long bitidx, word_bitidx;
525 unsigned long old_word, word;
527 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
528 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
530 bitmap = get_pageblock_bitmap(page, pfn);
531 bitidx = pfn_to_bitidx(page, pfn);
532 word_bitidx = bitidx / BITS_PER_LONG;
533 bitidx &= (BITS_PER_LONG-1);
535 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
540 word = READ_ONCE(bitmap[word_bitidx]);
542 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
543 if (word == old_word)
549 void set_pageblock_migratetype(struct page *page, int migratetype)
551 if (unlikely(page_group_by_mobility_disabled &&
552 migratetype < MIGRATE_PCPTYPES))
553 migratetype = MIGRATE_UNMOVABLE;
555 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
556 page_to_pfn(page), MIGRATETYPE_MASK);
559 #ifdef CONFIG_DEBUG_VM
560 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
564 unsigned long pfn = page_to_pfn(page);
565 unsigned long sp, start_pfn;
568 seq = zone_span_seqbegin(zone);
569 start_pfn = zone->zone_start_pfn;
570 sp = zone->spanned_pages;
571 if (!zone_spans_pfn(zone, pfn))
573 } while (zone_span_seqretry(zone, seq));
576 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
577 pfn, zone_to_nid(zone), zone->name,
578 start_pfn, start_pfn + sp);
583 static int page_is_consistent(struct zone *zone, struct page *page)
585 if (!pfn_valid_within(page_to_pfn(page)))
587 if (zone != page_zone(page))
593 * Temporary debugging check for pages not lying within a given zone.
595 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
597 if (page_outside_zone_boundaries(zone, page))
599 if (!page_is_consistent(zone, page))
605 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
611 static void bad_page(struct page *page, const char *reason)
613 static unsigned long resume;
614 static unsigned long nr_shown;
615 static unsigned long nr_unshown;
618 * Allow a burst of 60 reports, then keep quiet for that minute;
619 * or allow a steady drip of one report per second.
621 if (nr_shown == 60) {
622 if (time_before(jiffies, resume)) {
628 "BUG: Bad page state: %lu messages suppressed\n",
635 resume = jiffies + 60 * HZ;
637 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
638 current->comm, page_to_pfn(page));
639 __dump_page(page, reason);
640 dump_page_owner(page);
645 /* Leave bad fields for debug, except PageBuddy could make trouble */
646 page_mapcount_reset(page); /* remove PageBuddy */
647 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
651 * Higher-order pages are called "compound pages". They are structured thusly:
653 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
655 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
656 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
658 * The first tail page's ->compound_dtor holds the offset in array of compound
659 * page destructors. See compound_page_dtors.
661 * The first tail page's ->compound_order holds the order of allocation.
662 * This usage means that zero-order pages may not be compound.
665 void free_compound_page(struct page *page)
667 mem_cgroup_uncharge(page);
668 __free_pages_ok(page, compound_order(page), FPI_NONE);
671 void prep_compound_page(struct page *page, unsigned int order)
674 int nr_pages = 1 << order;
677 for (i = 1; i < nr_pages; i++) {
678 struct page *p = page + i;
679 set_page_count(p, 0);
680 p->mapping = TAIL_MAPPING;
681 set_compound_head(p, page);
684 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
685 set_compound_order(page, order);
686 atomic_set(compound_mapcount_ptr(page), -1);
687 if (hpage_pincount_available(page))
688 atomic_set(compound_pincount_ptr(page), 0);
691 #ifdef CONFIG_DEBUG_PAGEALLOC
692 unsigned int _debug_guardpage_minorder;
694 bool _debug_pagealloc_enabled_early __read_mostly
695 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
696 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
697 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
698 EXPORT_SYMBOL(_debug_pagealloc_enabled);
700 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
702 static int __init early_debug_pagealloc(char *buf)
704 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
706 early_param("debug_pagealloc", early_debug_pagealloc);
708 static int __init debug_guardpage_minorder_setup(char *buf)
712 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
713 pr_err("Bad debug_guardpage_minorder value\n");
716 _debug_guardpage_minorder = res;
717 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
720 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
722 static inline bool set_page_guard(struct zone *zone, struct page *page,
723 unsigned int order, int migratetype)
725 if (!debug_guardpage_enabled())
728 if (order >= debug_guardpage_minorder())
731 __SetPageGuard(page);
732 INIT_LIST_HEAD(&page->lru);
733 set_page_private(page, order);
734 /* Guard pages are not available for any usage */
735 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
740 static inline void clear_page_guard(struct zone *zone, struct page *page,
741 unsigned int order, int migratetype)
743 if (!debug_guardpage_enabled())
746 __ClearPageGuard(page);
748 set_page_private(page, 0);
749 if (!is_migrate_isolate(migratetype))
750 __mod_zone_freepage_state(zone, (1 << order), migratetype);
753 static inline bool set_page_guard(struct zone *zone, struct page *page,
754 unsigned int order, int migratetype) { return false; }
755 static inline void clear_page_guard(struct zone *zone, struct page *page,
756 unsigned int order, int migratetype) {}
760 * Enable static keys related to various memory debugging and hardening options.
761 * Some override others, and depend on early params that are evaluated in the
762 * order of appearance. So we need to first gather the full picture of what was
763 * enabled, and then make decisions.
765 void init_mem_debugging_and_hardening(void)
767 if (_init_on_alloc_enabled_early) {
768 if (page_poisoning_enabled())
769 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
770 "will take precedence over init_on_alloc\n");
772 static_branch_enable(&init_on_alloc);
774 if (_init_on_free_enabled_early) {
775 if (page_poisoning_enabled())
776 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, "
777 "will take precedence over init_on_free\n");
779 static_branch_enable(&init_on_free);
782 #ifdef CONFIG_PAGE_POISONING
784 * Page poisoning is debug page alloc for some arches. If
785 * either of those options are enabled, enable poisoning.
787 if (page_poisoning_enabled() ||
788 (!IS_ENABLED(CONFIG_ARCH_SUPPORTS_DEBUG_PAGEALLOC) &&
789 debug_pagealloc_enabled()))
790 static_branch_enable(&_page_poisoning_enabled);
793 #ifdef CONFIG_DEBUG_PAGEALLOC
794 if (!debug_pagealloc_enabled())
797 static_branch_enable(&_debug_pagealloc_enabled);
799 if (!debug_guardpage_minorder())
802 static_branch_enable(&_debug_guardpage_enabled);
806 static inline void set_buddy_order(struct page *page, unsigned int order)
808 set_page_private(page, order);
809 __SetPageBuddy(page);
813 * This function checks whether a page is free && is the buddy
814 * we can coalesce a page and its buddy if
815 * (a) the buddy is not in a hole (check before calling!) &&
816 * (b) the buddy is in the buddy system &&
817 * (c) a page and its buddy have the same order &&
818 * (d) a page and its buddy are in the same zone.
820 * For recording whether a page is in the buddy system, we set PageBuddy.
821 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
823 * For recording page's order, we use page_private(page).
825 static inline bool page_is_buddy(struct page *page, struct page *buddy,
828 if (!page_is_guard(buddy) && !PageBuddy(buddy))
831 if (buddy_order(buddy) != order)
835 * zone check is done late to avoid uselessly calculating
836 * zone/node ids for pages that could never merge.
838 if (page_zone_id(page) != page_zone_id(buddy))
841 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
846 #ifdef CONFIG_COMPACTION
847 static inline struct capture_control *task_capc(struct zone *zone)
849 struct capture_control *capc = current->capture_control;
851 return unlikely(capc) &&
852 !(current->flags & PF_KTHREAD) &&
854 capc->cc->zone == zone ? capc : NULL;
858 compaction_capture(struct capture_control *capc, struct page *page,
859 int order, int migratetype)
861 if (!capc || order != capc->cc->order)
864 /* Do not accidentally pollute CMA or isolated regions*/
865 if (is_migrate_cma(migratetype) ||
866 is_migrate_isolate(migratetype))
870 * Do not let lower order allocations polluate a movable pageblock.
871 * This might let an unmovable request use a reclaimable pageblock
872 * and vice-versa but no more than normal fallback logic which can
873 * have trouble finding a high-order free page.
875 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
883 static inline struct capture_control *task_capc(struct zone *zone)
889 compaction_capture(struct capture_control *capc, struct page *page,
890 int order, int migratetype)
894 #endif /* CONFIG_COMPACTION */
896 /* Used for pages not on another list */
897 static inline void add_to_free_list(struct page *page, struct zone *zone,
898 unsigned int order, int migratetype)
900 struct free_area *area = &zone->free_area[order];
902 list_add(&page->lru, &area->free_list[migratetype]);
906 /* Used for pages not on another list */
907 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
908 unsigned int order, int migratetype)
910 struct free_area *area = &zone->free_area[order];
912 list_add_tail(&page->lru, &area->free_list[migratetype]);
917 * Used for pages which are on another list. Move the pages to the tail
918 * of the list - so the moved pages won't immediately be considered for
919 * allocation again (e.g., optimization for memory onlining).
921 static inline void move_to_free_list(struct page *page, struct zone *zone,
922 unsigned int order, int migratetype)
924 struct free_area *area = &zone->free_area[order];
926 list_move_tail(&page->lru, &area->free_list[migratetype]);
929 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
932 /* clear reported state and update reported page count */
933 if (page_reported(page))
934 __ClearPageReported(page);
936 list_del(&page->lru);
937 __ClearPageBuddy(page);
938 set_page_private(page, 0);
939 zone->free_area[order].nr_free--;
943 * If this is not the largest possible page, check if the buddy
944 * of the next-highest order is free. If it is, it's possible
945 * that pages are being freed that will coalesce soon. In case,
946 * that is happening, add the free page to the tail of the list
947 * so it's less likely to be used soon and more likely to be merged
948 * as a higher order page
951 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
952 struct page *page, unsigned int order)
954 struct page *higher_page, *higher_buddy;
955 unsigned long combined_pfn;
957 if (order >= MAX_ORDER - 2)
960 if (!pfn_valid_within(buddy_pfn))
963 combined_pfn = buddy_pfn & pfn;
964 higher_page = page + (combined_pfn - pfn);
965 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
966 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
968 return pfn_valid_within(buddy_pfn) &&
969 page_is_buddy(higher_page, higher_buddy, order + 1);
973 * Freeing function for a buddy system allocator.
975 * The concept of a buddy system is to maintain direct-mapped table
976 * (containing bit values) for memory blocks of various "orders".
977 * The bottom level table contains the map for the smallest allocatable
978 * units of memory (here, pages), and each level above it describes
979 * pairs of units from the levels below, hence, "buddies".
980 * At a high level, all that happens here is marking the table entry
981 * at the bottom level available, and propagating the changes upward
982 * as necessary, plus some accounting needed to play nicely with other
983 * parts of the VM system.
984 * At each level, we keep a list of pages, which are heads of continuous
985 * free pages of length of (1 << order) and marked with PageBuddy.
986 * Page's order is recorded in page_private(page) field.
987 * So when we are allocating or freeing one, we can derive the state of the
988 * other. That is, if we allocate a small block, and both were
989 * free, the remainder of the region must be split into blocks.
990 * If a block is freed, and its buddy is also free, then this
991 * triggers coalescing into a block of larger size.
996 static inline void __free_one_page(struct page *page,
998 struct zone *zone, unsigned int order,
999 int migratetype, fpi_t fpi_flags)
1001 struct capture_control *capc = task_capc(zone);
1002 unsigned long buddy_pfn;
1003 unsigned long combined_pfn;
1004 unsigned int max_order;
1008 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
1010 VM_BUG_ON(!zone_is_initialized(zone));
1011 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
1013 VM_BUG_ON(migratetype == -1);
1014 if (likely(!is_migrate_isolate(migratetype)))
1015 __mod_zone_freepage_state(zone, 1 << order, migratetype);
1017 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
1018 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1021 while (order < max_order) {
1022 if (compaction_capture(capc, page, order, migratetype)) {
1023 __mod_zone_freepage_state(zone, -(1 << order),
1027 buddy_pfn = __find_buddy_pfn(pfn, order);
1028 buddy = page + (buddy_pfn - pfn);
1030 if (!pfn_valid_within(buddy_pfn))
1032 if (!page_is_buddy(page, buddy, order))
1035 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
1036 * merge with it and move up one order.
1038 if (page_is_guard(buddy))
1039 clear_page_guard(zone, buddy, order, migratetype);
1041 del_page_from_free_list(buddy, zone, order);
1042 combined_pfn = buddy_pfn & pfn;
1043 page = page + (combined_pfn - pfn);
1047 if (order < MAX_ORDER - 1) {
1048 /* If we are here, it means order is >= pageblock_order.
1049 * We want to prevent merge between freepages on isolate
1050 * pageblock and normal pageblock. Without this, pageblock
1051 * isolation could cause incorrect freepage or CMA accounting.
1053 * We don't want to hit this code for the more frequent
1054 * low-order merging.
1056 if (unlikely(has_isolate_pageblock(zone))) {
1059 buddy_pfn = __find_buddy_pfn(pfn, order);
1060 buddy = page + (buddy_pfn - pfn);
1061 buddy_mt = get_pageblock_migratetype(buddy);
1063 if (migratetype != buddy_mt
1064 && (is_migrate_isolate(migratetype) ||
1065 is_migrate_isolate(buddy_mt)))
1068 max_order = order + 1;
1069 goto continue_merging;
1073 set_buddy_order(page, order);
1075 if (fpi_flags & FPI_TO_TAIL)
1077 else if (is_shuffle_order(order))
1078 to_tail = shuffle_pick_tail();
1080 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1083 add_to_free_list_tail(page, zone, order, migratetype);
1085 add_to_free_list(page, zone, order, migratetype);
1087 /* Notify page reporting subsystem of freed page */
1088 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
1089 page_reporting_notify_free(order);
1093 * A bad page could be due to a number of fields. Instead of multiple branches,
1094 * try and check multiple fields with one check. The caller must do a detailed
1095 * check if necessary.
1097 static inline bool page_expected_state(struct page *page,
1098 unsigned long check_flags)
1100 if (unlikely(atomic_read(&page->_mapcount) != -1))
1103 if (unlikely((unsigned long)page->mapping |
1104 page_ref_count(page) |
1106 (unsigned long)page_memcg(page) |
1108 (page->flags & check_flags)))
1114 static const char *page_bad_reason(struct page *page, unsigned long flags)
1116 const char *bad_reason = NULL;
1118 if (unlikely(atomic_read(&page->_mapcount) != -1))
1119 bad_reason = "nonzero mapcount";
1120 if (unlikely(page->mapping != NULL))
1121 bad_reason = "non-NULL mapping";
1122 if (unlikely(page_ref_count(page) != 0))
1123 bad_reason = "nonzero _refcount";
1124 if (unlikely(page->flags & flags)) {
1125 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1126 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1128 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1131 if (unlikely(page_memcg(page)))
1132 bad_reason = "page still charged to cgroup";
1137 static void check_free_page_bad(struct page *page)
1140 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1143 static inline int check_free_page(struct page *page)
1145 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1148 /* Something has gone sideways, find it */
1149 check_free_page_bad(page);
1153 static int free_tail_pages_check(struct page *head_page, struct page *page)
1158 * We rely page->lru.next never has bit 0 set, unless the page
1159 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1161 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1163 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1167 switch (page - head_page) {
1169 /* the first tail page: ->mapping may be compound_mapcount() */
1170 if (unlikely(compound_mapcount(page))) {
1171 bad_page(page, "nonzero compound_mapcount");
1177 * the second tail page: ->mapping is
1178 * deferred_list.next -- ignore value.
1182 if (page->mapping != TAIL_MAPPING) {
1183 bad_page(page, "corrupted mapping in tail page");
1188 if (unlikely(!PageTail(page))) {
1189 bad_page(page, "PageTail not set");
1192 if (unlikely(compound_head(page) != head_page)) {
1193 bad_page(page, "compound_head not consistent");
1198 page->mapping = NULL;
1199 clear_compound_head(page);
1203 static void kernel_init_free_pages(struct page *page, int numpages)
1207 /* s390's use of memset() could override KASAN redzones. */
1208 kasan_disable_current();
1209 for (i = 0; i < numpages; i++) {
1210 page_kasan_tag_reset(page + i);
1211 clear_highpage(page + i);
1213 kasan_enable_current();
1216 static __always_inline bool free_pages_prepare(struct page *page,
1217 unsigned int order, bool check_free)
1221 VM_BUG_ON_PAGE(PageTail(page), page);
1223 trace_mm_page_free(page, order);
1225 if (unlikely(PageHWPoison(page)) && !order) {
1227 * Do not let hwpoison pages hit pcplists/buddy
1228 * Untie memcg state and reset page's owner
1230 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1231 __memcg_kmem_uncharge_page(page, order);
1232 reset_page_owner(page, order);
1237 * Check tail pages before head page information is cleared to
1238 * avoid checking PageCompound for order-0 pages.
1240 if (unlikely(order)) {
1241 bool compound = PageCompound(page);
1244 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1247 ClearPageDoubleMap(page);
1248 for (i = 1; i < (1 << order); i++) {
1250 bad += free_tail_pages_check(page, page + i);
1251 if (unlikely(check_free_page(page + i))) {
1255 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1258 if (PageMappingFlags(page))
1259 page->mapping = NULL;
1260 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1261 __memcg_kmem_uncharge_page(page, order);
1263 bad += check_free_page(page);
1267 page_cpupid_reset_last(page);
1268 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1269 reset_page_owner(page, order);
1271 if (!PageHighMem(page)) {
1272 debug_check_no_locks_freed(page_address(page),
1273 PAGE_SIZE << order);
1274 debug_check_no_obj_freed(page_address(page),
1275 PAGE_SIZE << order);
1277 if (want_init_on_free())
1278 kernel_init_free_pages(page, 1 << order);
1280 kernel_poison_pages(page, 1 << order);
1283 * arch_free_page() can make the page's contents inaccessible. s390
1284 * does this. So nothing which can access the page's contents should
1285 * happen after this.
1287 arch_free_page(page, order);
1289 debug_pagealloc_unmap_pages(page, 1 << order);
1291 kasan_free_nondeferred_pages(page, order);
1296 #ifdef CONFIG_DEBUG_VM
1298 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1299 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1300 * moved from pcp lists to free lists.
1302 static bool free_pcp_prepare(struct page *page)
1304 return free_pages_prepare(page, 0, true);
1307 static bool bulkfree_pcp_prepare(struct page *page)
1309 if (debug_pagealloc_enabled_static())
1310 return check_free_page(page);
1316 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1317 * moving from pcp lists to free list in order to reduce overhead. With
1318 * debug_pagealloc enabled, they are checked also immediately when being freed
1321 static bool free_pcp_prepare(struct page *page)
1323 if (debug_pagealloc_enabled_static())
1324 return free_pages_prepare(page, 0, true);
1326 return free_pages_prepare(page, 0, false);
1329 static bool bulkfree_pcp_prepare(struct page *page)
1331 return check_free_page(page);
1333 #endif /* CONFIG_DEBUG_VM */
1335 static inline void prefetch_buddy(struct page *page)
1337 unsigned long pfn = page_to_pfn(page);
1338 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1339 struct page *buddy = page + (buddy_pfn - pfn);
1345 * Frees a number of pages from the PCP lists
1346 * Assumes all pages on list are in same zone, and of same order.
1347 * count is the number of pages to free.
1349 * If the zone was previously in an "all pages pinned" state then look to
1350 * see if this freeing clears that state.
1352 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1353 * pinned" detection logic.
1355 static void free_pcppages_bulk(struct zone *zone, int count,
1356 struct per_cpu_pages *pcp)
1358 int migratetype = 0;
1360 int prefetch_nr = READ_ONCE(pcp->batch);
1361 bool isolated_pageblocks;
1362 struct page *page, *tmp;
1366 * Ensure proper count is passed which otherwise would stuck in the
1367 * below while (list_empty(list)) loop.
1369 count = min(pcp->count, count);
1371 struct list_head *list;
1374 * Remove pages from lists in a round-robin fashion. A
1375 * batch_free count is maintained that is incremented when an
1376 * empty list is encountered. This is so more pages are freed
1377 * off fuller lists instead of spinning excessively around empty
1382 if (++migratetype == MIGRATE_PCPTYPES)
1384 list = &pcp->lists[migratetype];
1385 } while (list_empty(list));
1387 /* This is the only non-empty list. Free them all. */
1388 if (batch_free == MIGRATE_PCPTYPES)
1392 page = list_last_entry(list, struct page, lru);
1393 /* must delete to avoid corrupting pcp list */
1394 list_del(&page->lru);
1397 if (bulkfree_pcp_prepare(page))
1400 list_add_tail(&page->lru, &head);
1403 * We are going to put the page back to the global
1404 * pool, prefetch its buddy to speed up later access
1405 * under zone->lock. It is believed the overhead of
1406 * an additional test and calculating buddy_pfn here
1407 * can be offset by reduced memory latency later. To
1408 * avoid excessive prefetching due to large count, only
1409 * prefetch buddy for the first pcp->batch nr of pages.
1412 prefetch_buddy(page);
1415 } while (--count && --batch_free && !list_empty(list));
1418 spin_lock(&zone->lock);
1419 isolated_pageblocks = has_isolate_pageblock(zone);
1422 * Use safe version since after __free_one_page(),
1423 * page->lru.next will not point to original list.
1425 list_for_each_entry_safe(page, tmp, &head, lru) {
1426 int mt = get_pcppage_migratetype(page);
1427 /* MIGRATE_ISOLATE page should not go to pcplists */
1428 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1429 /* Pageblock could have been isolated meanwhile */
1430 if (unlikely(isolated_pageblocks))
1431 mt = get_pageblock_migratetype(page);
1433 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1434 trace_mm_page_pcpu_drain(page, 0, mt);
1436 spin_unlock(&zone->lock);
1439 static void free_one_page(struct zone *zone,
1440 struct page *page, unsigned long pfn,
1442 int migratetype, fpi_t fpi_flags)
1444 spin_lock(&zone->lock);
1445 if (unlikely(has_isolate_pageblock(zone) ||
1446 is_migrate_isolate(migratetype))) {
1447 migratetype = get_pfnblock_migratetype(page, pfn);
1449 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1450 spin_unlock(&zone->lock);
1453 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1454 unsigned long zone, int nid)
1456 mm_zero_struct_page(page);
1457 set_page_links(page, zone, nid, pfn);
1458 init_page_count(page);
1459 page_mapcount_reset(page);
1460 page_cpupid_reset_last(page);
1461 page_kasan_tag_reset(page);
1463 INIT_LIST_HEAD(&page->lru);
1464 #ifdef WANT_PAGE_VIRTUAL
1465 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1466 if (!is_highmem_idx(zone))
1467 set_page_address(page, __va(pfn << PAGE_SHIFT));
1471 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1472 static void __meminit init_reserved_page(unsigned long pfn)
1477 if (!early_page_uninitialised(pfn))
1480 nid = early_pfn_to_nid(pfn);
1481 pgdat = NODE_DATA(nid);
1483 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1484 struct zone *zone = &pgdat->node_zones[zid];
1486 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1489 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1492 static inline void init_reserved_page(unsigned long pfn)
1495 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1498 * Initialised pages do not have PageReserved set. This function is
1499 * called for each range allocated by the bootmem allocator and
1500 * marks the pages PageReserved. The remaining valid pages are later
1501 * sent to the buddy page allocator.
1503 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1505 unsigned long start_pfn = PFN_DOWN(start);
1506 unsigned long end_pfn = PFN_UP(end);
1508 for (; start_pfn < end_pfn; start_pfn++) {
1509 if (pfn_valid(start_pfn)) {
1510 struct page *page = pfn_to_page(start_pfn);
1512 init_reserved_page(start_pfn);
1514 /* Avoid false-positive PageTail() */
1515 INIT_LIST_HEAD(&page->lru);
1518 * no need for atomic set_bit because the struct
1519 * page is not visible yet so nobody should
1522 __SetPageReserved(page);
1527 static void __free_pages_ok(struct page *page, unsigned int order,
1530 unsigned long flags;
1532 unsigned long pfn = page_to_pfn(page);
1534 if (!free_pages_prepare(page, order, true))
1537 migratetype = get_pfnblock_migratetype(page, pfn);
1538 local_irq_save(flags);
1539 __count_vm_events(PGFREE, 1 << order);
1540 free_one_page(page_zone(page), page, pfn, order, migratetype,
1542 local_irq_restore(flags);
1545 void __free_pages_core(struct page *page, unsigned int order)
1547 unsigned int nr_pages = 1 << order;
1548 struct page *p = page;
1552 * When initializing the memmap, __init_single_page() sets the refcount
1553 * of all pages to 1 ("allocated"/"not free"). We have to set the
1554 * refcount of all involved pages to 0.
1557 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1559 __ClearPageReserved(p);
1560 set_page_count(p, 0);
1562 __ClearPageReserved(p);
1563 set_page_count(p, 0);
1565 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1568 * Bypass PCP and place fresh pages right to the tail, primarily
1569 * relevant for memory onlining.
1571 __free_pages_ok(page, order, FPI_TO_TAIL);
1574 #ifdef CONFIG_NEED_MULTIPLE_NODES
1577 * During memory init memblocks map pfns to nids. The search is expensive and
1578 * this caches recent lookups. The implementation of __early_pfn_to_nid
1579 * treats start/end as pfns.
1581 struct mminit_pfnnid_cache {
1582 unsigned long last_start;
1583 unsigned long last_end;
1587 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1590 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1592 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1593 struct mminit_pfnnid_cache *state)
1595 unsigned long start_pfn, end_pfn;
1598 if (state->last_start <= pfn && pfn < state->last_end)
1599 return state->last_nid;
1601 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1602 if (nid != NUMA_NO_NODE) {
1603 state->last_start = start_pfn;
1604 state->last_end = end_pfn;
1605 state->last_nid = nid;
1611 int __meminit early_pfn_to_nid(unsigned long pfn)
1613 static DEFINE_SPINLOCK(early_pfn_lock);
1616 spin_lock(&early_pfn_lock);
1617 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1619 nid = first_online_node;
1620 spin_unlock(&early_pfn_lock);
1624 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1626 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1629 if (early_page_uninitialised(pfn))
1631 __free_pages_core(page, order);
1635 * Check that the whole (or subset of) a pageblock given by the interval of
1636 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1637 * with the migration of free compaction scanner. The scanners then need to
1638 * use only pfn_valid_within() check for arches that allow holes within
1641 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1643 * It's possible on some configurations to have a setup like node0 node1 node0
1644 * i.e. it's possible that all pages within a zones range of pages do not
1645 * belong to a single zone. We assume that a border between node0 and node1
1646 * can occur within a single pageblock, but not a node0 node1 node0
1647 * interleaving within a single pageblock. It is therefore sufficient to check
1648 * the first and last page of a pageblock and avoid checking each individual
1649 * page in a pageblock.
1651 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1652 unsigned long end_pfn, struct zone *zone)
1654 struct page *start_page;
1655 struct page *end_page;
1657 /* end_pfn is one past the range we are checking */
1660 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1663 start_page = pfn_to_online_page(start_pfn);
1667 if (page_zone(start_page) != zone)
1670 end_page = pfn_to_page(end_pfn);
1672 /* This gives a shorter code than deriving page_zone(end_page) */
1673 if (page_zone_id(start_page) != page_zone_id(end_page))
1679 void set_zone_contiguous(struct zone *zone)
1681 unsigned long block_start_pfn = zone->zone_start_pfn;
1682 unsigned long block_end_pfn;
1684 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1685 for (; block_start_pfn < zone_end_pfn(zone);
1686 block_start_pfn = block_end_pfn,
1687 block_end_pfn += pageblock_nr_pages) {
1689 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1691 if (!__pageblock_pfn_to_page(block_start_pfn,
1692 block_end_pfn, zone))
1697 /* We confirm that there is no hole */
1698 zone->contiguous = true;
1701 void clear_zone_contiguous(struct zone *zone)
1703 zone->contiguous = false;
1706 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1707 static void __init deferred_free_range(unsigned long pfn,
1708 unsigned long nr_pages)
1716 page = pfn_to_page(pfn);
1718 /* Free a large naturally-aligned chunk if possible */
1719 if (nr_pages == pageblock_nr_pages &&
1720 (pfn & (pageblock_nr_pages - 1)) == 0) {
1721 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1722 __free_pages_core(page, pageblock_order);
1726 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1727 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1728 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1729 __free_pages_core(page, 0);
1733 /* Completion tracking for deferred_init_memmap() threads */
1734 static atomic_t pgdat_init_n_undone __initdata;
1735 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1737 static inline void __init pgdat_init_report_one_done(void)
1739 if (atomic_dec_and_test(&pgdat_init_n_undone))
1740 complete(&pgdat_init_all_done_comp);
1744 * Returns true if page needs to be initialized or freed to buddy allocator.
1746 * First we check if pfn is valid on architectures where it is possible to have
1747 * holes within pageblock_nr_pages. On systems where it is not possible, this
1748 * function is optimized out.
1750 * Then, we check if a current large page is valid by only checking the validity
1753 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1755 if (!pfn_valid_within(pfn))
1757 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1763 * Free pages to buddy allocator. Try to free aligned pages in
1764 * pageblock_nr_pages sizes.
1766 static void __init deferred_free_pages(unsigned long pfn,
1767 unsigned long end_pfn)
1769 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1770 unsigned long nr_free = 0;
1772 for (; pfn < end_pfn; pfn++) {
1773 if (!deferred_pfn_valid(pfn)) {
1774 deferred_free_range(pfn - nr_free, nr_free);
1776 } else if (!(pfn & nr_pgmask)) {
1777 deferred_free_range(pfn - nr_free, nr_free);
1783 /* Free the last block of pages to allocator */
1784 deferred_free_range(pfn - nr_free, nr_free);
1788 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1789 * by performing it only once every pageblock_nr_pages.
1790 * Return number of pages initialized.
1792 static unsigned long __init deferred_init_pages(struct zone *zone,
1794 unsigned long end_pfn)
1796 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1797 int nid = zone_to_nid(zone);
1798 unsigned long nr_pages = 0;
1799 int zid = zone_idx(zone);
1800 struct page *page = NULL;
1802 for (; pfn < end_pfn; pfn++) {
1803 if (!deferred_pfn_valid(pfn)) {
1806 } else if (!page || !(pfn & nr_pgmask)) {
1807 page = pfn_to_page(pfn);
1811 __init_single_page(page, pfn, zid, nid);
1818 * This function is meant to pre-load the iterator for the zone init.
1819 * Specifically it walks through the ranges until we are caught up to the
1820 * first_init_pfn value and exits there. If we never encounter the value we
1821 * return false indicating there are no valid ranges left.
1824 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1825 unsigned long *spfn, unsigned long *epfn,
1826 unsigned long first_init_pfn)
1831 * Start out by walking through the ranges in this zone that have
1832 * already been initialized. We don't need to do anything with them
1833 * so we just need to flush them out of the system.
1835 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1836 if (*epfn <= first_init_pfn)
1838 if (*spfn < first_init_pfn)
1839 *spfn = first_init_pfn;
1848 * Initialize and free pages. We do it in two loops: first we initialize
1849 * struct page, then free to buddy allocator, because while we are
1850 * freeing pages we can access pages that are ahead (computing buddy
1851 * page in __free_one_page()).
1853 * In order to try and keep some memory in the cache we have the loop
1854 * broken along max page order boundaries. This way we will not cause
1855 * any issues with the buddy page computation.
1857 static unsigned long __init
1858 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1859 unsigned long *end_pfn)
1861 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1862 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1863 unsigned long nr_pages = 0;
1866 /* First we loop through and initialize the page values */
1867 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1870 if (mo_pfn <= *start_pfn)
1873 t = min(mo_pfn, *end_pfn);
1874 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1876 if (mo_pfn < *end_pfn) {
1877 *start_pfn = mo_pfn;
1882 /* Reset values and now loop through freeing pages as needed */
1885 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1891 t = min(mo_pfn, epfn);
1892 deferred_free_pages(spfn, t);
1902 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1905 unsigned long spfn, epfn;
1906 struct zone *zone = arg;
1909 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1912 * Initialize and free pages in MAX_ORDER sized increments so that we
1913 * can avoid introducing any issues with the buddy allocator.
1915 while (spfn < end_pfn) {
1916 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1921 /* An arch may override for more concurrency. */
1923 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1928 /* Initialise remaining memory on a node */
1929 static int __init deferred_init_memmap(void *data)
1931 pg_data_t *pgdat = data;
1932 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1933 unsigned long spfn = 0, epfn = 0;
1934 unsigned long first_init_pfn, flags;
1935 unsigned long start = jiffies;
1937 int zid, max_threads;
1940 /* Bind memory initialisation thread to a local node if possible */
1941 if (!cpumask_empty(cpumask))
1942 set_cpus_allowed_ptr(current, cpumask);
1944 pgdat_resize_lock(pgdat, &flags);
1945 first_init_pfn = pgdat->first_deferred_pfn;
1946 if (first_init_pfn == ULONG_MAX) {
1947 pgdat_resize_unlock(pgdat, &flags);
1948 pgdat_init_report_one_done();
1952 /* Sanity check boundaries */
1953 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1954 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1955 pgdat->first_deferred_pfn = ULONG_MAX;
1958 * Once we unlock here, the zone cannot be grown anymore, thus if an
1959 * interrupt thread must allocate this early in boot, zone must be
1960 * pre-grown prior to start of deferred page initialization.
1962 pgdat_resize_unlock(pgdat, &flags);
1964 /* Only the highest zone is deferred so find it */
1965 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1966 zone = pgdat->node_zones + zid;
1967 if (first_init_pfn < zone_end_pfn(zone))
1971 /* If the zone is empty somebody else may have cleared out the zone */
1972 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1976 max_threads = deferred_page_init_max_threads(cpumask);
1978 while (spfn < epfn) {
1979 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1980 struct padata_mt_job job = {
1981 .thread_fn = deferred_init_memmap_chunk,
1984 .size = epfn_align - spfn,
1985 .align = PAGES_PER_SECTION,
1986 .min_chunk = PAGES_PER_SECTION,
1987 .max_threads = max_threads,
1990 padata_do_multithreaded(&job);
1991 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1995 /* Sanity check that the next zone really is unpopulated */
1996 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1998 pr_info("node %d deferred pages initialised in %ums\n",
1999 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2001 pgdat_init_report_one_done();
2006 * If this zone has deferred pages, try to grow it by initializing enough
2007 * deferred pages to satisfy the allocation specified by order, rounded up to
2008 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2009 * of SECTION_SIZE bytes by initializing struct pages in increments of
2010 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2012 * Return true when zone was grown, otherwise return false. We return true even
2013 * when we grow less than requested, to let the caller decide if there are
2014 * enough pages to satisfy the allocation.
2016 * Note: We use noinline because this function is needed only during boot, and
2017 * it is called from a __ref function _deferred_grow_zone. This way we are
2018 * making sure that it is not inlined into permanent text section.
2020 static noinline bool __init
2021 deferred_grow_zone(struct zone *zone, unsigned int order)
2023 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2024 pg_data_t *pgdat = zone->zone_pgdat;
2025 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2026 unsigned long spfn, epfn, flags;
2027 unsigned long nr_pages = 0;
2030 /* Only the last zone may have deferred pages */
2031 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2034 pgdat_resize_lock(pgdat, &flags);
2037 * If someone grew this zone while we were waiting for spinlock, return
2038 * true, as there might be enough pages already.
2040 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2041 pgdat_resize_unlock(pgdat, &flags);
2045 /* If the zone is empty somebody else may have cleared out the zone */
2046 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2047 first_deferred_pfn)) {
2048 pgdat->first_deferred_pfn = ULONG_MAX;
2049 pgdat_resize_unlock(pgdat, &flags);
2050 /* Retry only once. */
2051 return first_deferred_pfn != ULONG_MAX;
2055 * Initialize and free pages in MAX_ORDER sized increments so
2056 * that we can avoid introducing any issues with the buddy
2059 while (spfn < epfn) {
2060 /* update our first deferred PFN for this section */
2061 first_deferred_pfn = spfn;
2063 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2064 touch_nmi_watchdog();
2066 /* We should only stop along section boundaries */
2067 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2070 /* If our quota has been met we can stop here */
2071 if (nr_pages >= nr_pages_needed)
2075 pgdat->first_deferred_pfn = spfn;
2076 pgdat_resize_unlock(pgdat, &flags);
2078 return nr_pages > 0;
2082 * deferred_grow_zone() is __init, but it is called from
2083 * get_page_from_freelist() during early boot until deferred_pages permanently
2084 * disables this call. This is why we have refdata wrapper to avoid warning,
2085 * and to ensure that the function body gets unloaded.
2088 _deferred_grow_zone(struct zone *zone, unsigned int order)
2090 return deferred_grow_zone(zone, order);
2093 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2095 void __init page_alloc_init_late(void)
2100 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2102 /* There will be num_node_state(N_MEMORY) threads */
2103 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2104 for_each_node_state(nid, N_MEMORY) {
2105 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2108 /* Block until all are initialised */
2109 wait_for_completion(&pgdat_init_all_done_comp);
2112 * The number of managed pages has changed due to the initialisation
2113 * so the pcpu batch and high limits needs to be updated or the limits
2114 * will be artificially small.
2116 for_each_populated_zone(zone)
2117 zone_pcp_update(zone);
2120 * We initialized the rest of the deferred pages. Permanently disable
2121 * on-demand struct page initialization.
2123 static_branch_disable(&deferred_pages);
2125 /* Reinit limits that are based on free pages after the kernel is up */
2126 files_maxfiles_init();
2131 /* Discard memblock private memory */
2134 for_each_node_state(nid, N_MEMORY)
2135 shuffle_free_memory(NODE_DATA(nid));
2137 for_each_populated_zone(zone)
2138 set_zone_contiguous(zone);
2142 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2143 void __init init_cma_reserved_pageblock(struct page *page)
2145 unsigned i = pageblock_nr_pages;
2146 struct page *p = page;
2149 __ClearPageReserved(p);
2150 set_page_count(p, 0);
2153 set_pageblock_migratetype(page, MIGRATE_CMA);
2155 if (pageblock_order >= MAX_ORDER) {
2156 i = pageblock_nr_pages;
2159 set_page_refcounted(p);
2160 __free_pages(p, MAX_ORDER - 1);
2161 p += MAX_ORDER_NR_PAGES;
2162 } while (i -= MAX_ORDER_NR_PAGES);
2164 set_page_refcounted(page);
2165 __free_pages(page, pageblock_order);
2168 adjust_managed_page_count(page, pageblock_nr_pages);
2173 * The order of subdivision here is critical for the IO subsystem.
2174 * Please do not alter this order without good reasons and regression
2175 * testing. Specifically, as large blocks of memory are subdivided,
2176 * the order in which smaller blocks are delivered depends on the order
2177 * they're subdivided in this function. This is the primary factor
2178 * influencing the order in which pages are delivered to the IO
2179 * subsystem according to empirical testing, and this is also justified
2180 * by considering the behavior of a buddy system containing a single
2181 * large block of memory acted on by a series of small allocations.
2182 * This behavior is a critical factor in sglist merging's success.
2186 static inline void expand(struct zone *zone, struct page *page,
2187 int low, int high, int migratetype)
2189 unsigned long size = 1 << high;
2191 while (high > low) {
2194 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2197 * Mark as guard pages (or page), that will allow to
2198 * merge back to allocator when buddy will be freed.
2199 * Corresponding page table entries will not be touched,
2200 * pages will stay not present in virtual address space
2202 if (set_page_guard(zone, &page[size], high, migratetype))
2205 add_to_free_list(&page[size], zone, high, migratetype);
2206 set_buddy_order(&page[size], high);
2210 static void check_new_page_bad(struct page *page)
2212 if (unlikely(page->flags & __PG_HWPOISON)) {
2213 /* Don't complain about hwpoisoned pages */
2214 page_mapcount_reset(page); /* remove PageBuddy */
2219 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2223 * This page is about to be returned from the page allocator
2225 static inline int check_new_page(struct page *page)
2227 if (likely(page_expected_state(page,
2228 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2231 check_new_page_bad(page);
2235 #ifdef CONFIG_DEBUG_VM
2237 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2238 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2239 * also checked when pcp lists are refilled from the free lists.
2241 static inline bool check_pcp_refill(struct page *page)
2243 if (debug_pagealloc_enabled_static())
2244 return check_new_page(page);
2249 static inline bool check_new_pcp(struct page *page)
2251 return check_new_page(page);
2255 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2256 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2257 * enabled, they are also checked when being allocated from the pcp lists.
2259 static inline bool check_pcp_refill(struct page *page)
2261 return check_new_page(page);
2263 static inline bool check_new_pcp(struct page *page)
2265 if (debug_pagealloc_enabled_static())
2266 return check_new_page(page);
2270 #endif /* CONFIG_DEBUG_VM */
2272 static bool check_new_pages(struct page *page, unsigned int order)
2275 for (i = 0; i < (1 << order); i++) {
2276 struct page *p = page + i;
2278 if (unlikely(check_new_page(p)))
2285 inline void post_alloc_hook(struct page *page, unsigned int order,
2288 set_page_private(page, 0);
2289 set_page_refcounted(page);
2291 arch_alloc_page(page, order);
2292 debug_pagealloc_map_pages(page, 1 << order);
2293 kasan_alloc_pages(page, order);
2294 kernel_unpoison_pages(page, 1 << order);
2295 set_page_owner(page, order, gfp_flags);
2297 if (!want_init_on_free() && want_init_on_alloc(gfp_flags))
2298 kernel_init_free_pages(page, 1 << order);
2301 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2302 unsigned int alloc_flags)
2304 post_alloc_hook(page, order, gfp_flags);
2306 if (order && (gfp_flags & __GFP_COMP))
2307 prep_compound_page(page, order);
2310 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2311 * allocate the page. The expectation is that the caller is taking
2312 * steps that will free more memory. The caller should avoid the page
2313 * being used for !PFMEMALLOC purposes.
2315 if (alloc_flags & ALLOC_NO_WATERMARKS)
2316 set_page_pfmemalloc(page);
2318 clear_page_pfmemalloc(page);
2322 * Go through the free lists for the given migratetype and remove
2323 * the smallest available page from the freelists
2325 static __always_inline
2326 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2329 unsigned int current_order;
2330 struct free_area *area;
2333 /* Find a page of the appropriate size in the preferred list */
2334 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2335 area = &(zone->free_area[current_order]);
2336 page = get_page_from_free_area(area, migratetype);
2339 del_page_from_free_list(page, zone, current_order);
2340 expand(zone, page, order, current_order, migratetype);
2341 set_pcppage_migratetype(page, migratetype);
2350 * This array describes the order lists are fallen back to when
2351 * the free lists for the desirable migrate type are depleted
2353 static int fallbacks[MIGRATE_TYPES][3] = {
2354 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2355 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2356 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2358 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2360 #ifdef CONFIG_MEMORY_ISOLATION
2361 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2366 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2369 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2372 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2373 unsigned int order) { return NULL; }
2377 * Move the free pages in a range to the freelist tail of the requested type.
2378 * Note that start_page and end_pages are not aligned on a pageblock
2379 * boundary. If alignment is required, use move_freepages_block()
2381 static int move_freepages(struct zone *zone,
2382 struct page *start_page, struct page *end_page,
2383 int migratetype, int *num_movable)
2387 int pages_moved = 0;
2389 for (page = start_page; page <= end_page;) {
2390 if (!pfn_valid_within(page_to_pfn(page))) {
2395 if (!PageBuddy(page)) {
2397 * We assume that pages that could be isolated for
2398 * migration are movable. But we don't actually try
2399 * isolating, as that would be expensive.
2402 (PageLRU(page) || __PageMovable(page)))
2409 /* Make sure we are not inadvertently changing nodes */
2410 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2411 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2413 order = buddy_order(page);
2414 move_to_free_list(page, zone, order, migratetype);
2416 pages_moved += 1 << order;
2422 int move_freepages_block(struct zone *zone, struct page *page,
2423 int migratetype, int *num_movable)
2425 unsigned long start_pfn, end_pfn;
2426 struct page *start_page, *end_page;
2431 start_pfn = page_to_pfn(page);
2432 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2433 start_page = pfn_to_page(start_pfn);
2434 end_page = start_page + pageblock_nr_pages - 1;
2435 end_pfn = start_pfn + pageblock_nr_pages - 1;
2437 /* Do not cross zone boundaries */
2438 if (!zone_spans_pfn(zone, start_pfn))
2440 if (!zone_spans_pfn(zone, end_pfn))
2443 return move_freepages(zone, start_page, end_page, migratetype,
2447 static void change_pageblock_range(struct page *pageblock_page,
2448 int start_order, int migratetype)
2450 int nr_pageblocks = 1 << (start_order - pageblock_order);
2452 while (nr_pageblocks--) {
2453 set_pageblock_migratetype(pageblock_page, migratetype);
2454 pageblock_page += pageblock_nr_pages;
2459 * When we are falling back to another migratetype during allocation, try to
2460 * steal extra free pages from the same pageblocks to satisfy further
2461 * allocations, instead of polluting multiple pageblocks.
2463 * If we are stealing a relatively large buddy page, it is likely there will
2464 * be more free pages in the pageblock, so try to steal them all. For
2465 * reclaimable and unmovable allocations, we steal regardless of page size,
2466 * as fragmentation caused by those allocations polluting movable pageblocks
2467 * is worse than movable allocations stealing from unmovable and reclaimable
2470 static bool can_steal_fallback(unsigned int order, int start_mt)
2473 * Leaving this order check is intended, although there is
2474 * relaxed order check in next check. The reason is that
2475 * we can actually steal whole pageblock if this condition met,
2476 * but, below check doesn't guarantee it and that is just heuristic
2477 * so could be changed anytime.
2479 if (order >= pageblock_order)
2482 if (order >= pageblock_order / 2 ||
2483 start_mt == MIGRATE_RECLAIMABLE ||
2484 start_mt == MIGRATE_UNMOVABLE ||
2485 page_group_by_mobility_disabled)
2491 static inline bool boost_watermark(struct zone *zone)
2493 unsigned long max_boost;
2495 if (!watermark_boost_factor)
2498 * Don't bother in zones that are unlikely to produce results.
2499 * On small machines, including kdump capture kernels running
2500 * in a small area, boosting the watermark can cause an out of
2501 * memory situation immediately.
2503 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2506 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2507 watermark_boost_factor, 10000);
2510 * high watermark may be uninitialised if fragmentation occurs
2511 * very early in boot so do not boost. We do not fall
2512 * through and boost by pageblock_nr_pages as failing
2513 * allocations that early means that reclaim is not going
2514 * to help and it may even be impossible to reclaim the
2515 * boosted watermark resulting in a hang.
2520 max_boost = max(pageblock_nr_pages, max_boost);
2522 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2529 * This function implements actual steal behaviour. If order is large enough,
2530 * we can steal whole pageblock. If not, we first move freepages in this
2531 * pageblock to our migratetype and determine how many already-allocated pages
2532 * are there in the pageblock with a compatible migratetype. If at least half
2533 * of pages are free or compatible, we can change migratetype of the pageblock
2534 * itself, so pages freed in the future will be put on the correct free list.
2536 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2537 unsigned int alloc_flags, int start_type, bool whole_block)
2539 unsigned int current_order = buddy_order(page);
2540 int free_pages, movable_pages, alike_pages;
2543 old_block_type = get_pageblock_migratetype(page);
2546 * This can happen due to races and we want to prevent broken
2547 * highatomic accounting.
2549 if (is_migrate_highatomic(old_block_type))
2552 /* Take ownership for orders >= pageblock_order */
2553 if (current_order >= pageblock_order) {
2554 change_pageblock_range(page, current_order, start_type);
2559 * Boost watermarks to increase reclaim pressure to reduce the
2560 * likelihood of future fallbacks. Wake kswapd now as the node
2561 * may be balanced overall and kswapd will not wake naturally.
2563 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2564 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2566 /* We are not allowed to try stealing from the whole block */
2570 free_pages = move_freepages_block(zone, page, start_type,
2573 * Determine how many pages are compatible with our allocation.
2574 * For movable allocation, it's the number of movable pages which
2575 * we just obtained. For other types it's a bit more tricky.
2577 if (start_type == MIGRATE_MOVABLE) {
2578 alike_pages = movable_pages;
2581 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2582 * to MOVABLE pageblock, consider all non-movable pages as
2583 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2584 * vice versa, be conservative since we can't distinguish the
2585 * exact migratetype of non-movable pages.
2587 if (old_block_type == MIGRATE_MOVABLE)
2588 alike_pages = pageblock_nr_pages
2589 - (free_pages + movable_pages);
2594 /* moving whole block can fail due to zone boundary conditions */
2599 * If a sufficient number of pages in the block are either free or of
2600 * comparable migratability as our allocation, claim the whole block.
2602 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2603 page_group_by_mobility_disabled)
2604 set_pageblock_migratetype(page, start_type);
2609 move_to_free_list(page, zone, current_order, start_type);
2613 * Check whether there is a suitable fallback freepage with requested order.
2614 * If only_stealable is true, this function returns fallback_mt only if
2615 * we can steal other freepages all together. This would help to reduce
2616 * fragmentation due to mixed migratetype pages in one pageblock.
2618 int find_suitable_fallback(struct free_area *area, unsigned int order,
2619 int migratetype, bool only_stealable, bool *can_steal)
2624 if (area->nr_free == 0)
2629 fallback_mt = fallbacks[migratetype][i];
2630 if (fallback_mt == MIGRATE_TYPES)
2633 if (free_area_empty(area, fallback_mt))
2636 if (can_steal_fallback(order, migratetype))
2639 if (!only_stealable)
2650 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2651 * there are no empty page blocks that contain a page with a suitable order
2653 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2654 unsigned int alloc_order)
2657 unsigned long max_managed, flags;
2660 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2661 * Check is race-prone but harmless.
2663 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2664 if (zone->nr_reserved_highatomic >= max_managed)
2667 spin_lock_irqsave(&zone->lock, flags);
2669 /* Recheck the nr_reserved_highatomic limit under the lock */
2670 if (zone->nr_reserved_highatomic >= max_managed)
2674 mt = get_pageblock_migratetype(page);
2675 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2676 && !is_migrate_cma(mt)) {
2677 zone->nr_reserved_highatomic += pageblock_nr_pages;
2678 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2679 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2683 spin_unlock_irqrestore(&zone->lock, flags);
2687 * Used when an allocation is about to fail under memory pressure. This
2688 * potentially hurts the reliability of high-order allocations when under
2689 * intense memory pressure but failed atomic allocations should be easier
2690 * to recover from than an OOM.
2692 * If @force is true, try to unreserve a pageblock even though highatomic
2693 * pageblock is exhausted.
2695 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2698 struct zonelist *zonelist = ac->zonelist;
2699 unsigned long flags;
2706 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2709 * Preserve at least one pageblock unless memory pressure
2712 if (!force && zone->nr_reserved_highatomic <=
2716 spin_lock_irqsave(&zone->lock, flags);
2717 for (order = 0; order < MAX_ORDER; order++) {
2718 struct free_area *area = &(zone->free_area[order]);
2720 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2725 * In page freeing path, migratetype change is racy so
2726 * we can counter several free pages in a pageblock
2727 * in this loop althoug we changed the pageblock type
2728 * from highatomic to ac->migratetype. So we should
2729 * adjust the count once.
2731 if (is_migrate_highatomic_page(page)) {
2733 * It should never happen but changes to
2734 * locking could inadvertently allow a per-cpu
2735 * drain to add pages to MIGRATE_HIGHATOMIC
2736 * while unreserving so be safe and watch for
2739 zone->nr_reserved_highatomic -= min(
2741 zone->nr_reserved_highatomic);
2745 * Convert to ac->migratetype and avoid the normal
2746 * pageblock stealing heuristics. Minimally, the caller
2747 * is doing the work and needs the pages. More
2748 * importantly, if the block was always converted to
2749 * MIGRATE_UNMOVABLE or another type then the number
2750 * of pageblocks that cannot be completely freed
2753 set_pageblock_migratetype(page, ac->migratetype);
2754 ret = move_freepages_block(zone, page, ac->migratetype,
2757 spin_unlock_irqrestore(&zone->lock, flags);
2761 spin_unlock_irqrestore(&zone->lock, flags);
2768 * Try finding a free buddy page on the fallback list and put it on the free
2769 * list of requested migratetype, possibly along with other pages from the same
2770 * block, depending on fragmentation avoidance heuristics. Returns true if
2771 * fallback was found so that __rmqueue_smallest() can grab it.
2773 * The use of signed ints for order and current_order is a deliberate
2774 * deviation from the rest of this file, to make the for loop
2775 * condition simpler.
2777 static __always_inline bool
2778 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2779 unsigned int alloc_flags)
2781 struct free_area *area;
2783 int min_order = order;
2789 * Do not steal pages from freelists belonging to other pageblocks
2790 * i.e. orders < pageblock_order. If there are no local zones free,
2791 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2793 if (alloc_flags & ALLOC_NOFRAGMENT)
2794 min_order = pageblock_order;
2797 * Find the largest available free page in the other list. This roughly
2798 * approximates finding the pageblock with the most free pages, which
2799 * would be too costly to do exactly.
2801 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2803 area = &(zone->free_area[current_order]);
2804 fallback_mt = find_suitable_fallback(area, current_order,
2805 start_migratetype, false, &can_steal);
2806 if (fallback_mt == -1)
2810 * We cannot steal all free pages from the pageblock and the
2811 * requested migratetype is movable. In that case it's better to
2812 * steal and split the smallest available page instead of the
2813 * largest available page, because even if the next movable
2814 * allocation falls back into a different pageblock than this
2815 * one, it won't cause permanent fragmentation.
2817 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2818 && current_order > order)
2827 for (current_order = order; current_order < MAX_ORDER;
2829 area = &(zone->free_area[current_order]);
2830 fallback_mt = find_suitable_fallback(area, current_order,
2831 start_migratetype, false, &can_steal);
2832 if (fallback_mt != -1)
2837 * This should not happen - we already found a suitable fallback
2838 * when looking for the largest page.
2840 VM_BUG_ON(current_order == MAX_ORDER);
2843 page = get_page_from_free_area(area, fallback_mt);
2845 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2848 trace_mm_page_alloc_extfrag(page, order, current_order,
2849 start_migratetype, fallback_mt);
2856 * Do the hard work of removing an element from the buddy allocator.
2857 * Call me with the zone->lock already held.
2859 static __always_inline struct page *
2860 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2861 unsigned int alloc_flags)
2865 if (IS_ENABLED(CONFIG_CMA)) {
2867 * Balance movable allocations between regular and CMA areas by
2868 * allocating from CMA when over half of the zone's free memory
2869 * is in the CMA area.
2871 if (alloc_flags & ALLOC_CMA &&
2872 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2873 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2874 page = __rmqueue_cma_fallback(zone, order);
2880 page = __rmqueue_smallest(zone, order, migratetype);
2881 if (unlikely(!page)) {
2882 if (alloc_flags & ALLOC_CMA)
2883 page = __rmqueue_cma_fallback(zone, order);
2885 if (!page && __rmqueue_fallback(zone, order, migratetype,
2891 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2896 * Obtain a specified number of elements from the buddy allocator, all under
2897 * a single hold of the lock, for efficiency. Add them to the supplied list.
2898 * Returns the number of new pages which were placed at *list.
2900 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2901 unsigned long count, struct list_head *list,
2902 int migratetype, unsigned int alloc_flags)
2906 spin_lock(&zone->lock);
2907 for (i = 0; i < count; ++i) {
2908 struct page *page = __rmqueue(zone, order, migratetype,
2910 if (unlikely(page == NULL))
2913 if (unlikely(check_pcp_refill(page)))
2917 * Split buddy pages returned by expand() are received here in
2918 * physical page order. The page is added to the tail of
2919 * caller's list. From the callers perspective, the linked list
2920 * is ordered by page number under some conditions. This is
2921 * useful for IO devices that can forward direction from the
2922 * head, thus also in the physical page order. This is useful
2923 * for IO devices that can merge IO requests if the physical
2924 * pages are ordered properly.
2926 list_add_tail(&page->lru, list);
2928 if (is_migrate_cma(get_pcppage_migratetype(page)))
2929 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2934 * i pages were removed from the buddy list even if some leak due
2935 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2936 * on i. Do not confuse with 'alloced' which is the number of
2937 * pages added to the pcp list.
2939 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2940 spin_unlock(&zone->lock);
2946 * Called from the vmstat counter updater to drain pagesets of this
2947 * currently executing processor on remote nodes after they have
2950 * Note that this function must be called with the thread pinned to
2951 * a single processor.
2953 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2955 unsigned long flags;
2956 int to_drain, batch;
2958 local_irq_save(flags);
2959 batch = READ_ONCE(pcp->batch);
2960 to_drain = min(pcp->count, batch);
2962 free_pcppages_bulk(zone, to_drain, pcp);
2963 local_irq_restore(flags);
2968 * Drain pcplists of the indicated processor and zone.
2970 * The processor must either be the current processor and the
2971 * thread pinned to the current processor or a processor that
2974 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2976 unsigned long flags;
2977 struct per_cpu_pageset *pset;
2978 struct per_cpu_pages *pcp;
2980 local_irq_save(flags);
2981 pset = per_cpu_ptr(zone->pageset, cpu);
2985 free_pcppages_bulk(zone, pcp->count, pcp);
2986 local_irq_restore(flags);
2990 * Drain pcplists of all zones on the indicated processor.
2992 * The processor must either be the current processor and the
2993 * thread pinned to the current processor or a processor that
2996 static void drain_pages(unsigned int cpu)
3000 for_each_populated_zone(zone) {
3001 drain_pages_zone(cpu, zone);
3006 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3008 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3009 * the single zone's pages.
3011 void drain_local_pages(struct zone *zone)
3013 int cpu = smp_processor_id();
3016 drain_pages_zone(cpu, zone);
3021 static void drain_local_pages_wq(struct work_struct *work)
3023 struct pcpu_drain *drain;
3025 drain = container_of(work, struct pcpu_drain, work);
3028 * drain_all_pages doesn't use proper cpu hotplug protection so
3029 * we can race with cpu offline when the WQ can move this from
3030 * a cpu pinned worker to an unbound one. We can operate on a different
3031 * cpu which is allright but we also have to make sure to not move to
3035 drain_local_pages(drain->zone);
3040 * The implementation of drain_all_pages(), exposing an extra parameter to
3041 * drain on all cpus.
3043 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3044 * not empty. The check for non-emptiness can however race with a free to
3045 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3046 * that need the guarantee that every CPU has drained can disable the
3047 * optimizing racy check.
3049 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3054 * Allocate in the BSS so we wont require allocation in
3055 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3057 static cpumask_t cpus_with_pcps;
3060 * Make sure nobody triggers this path before mm_percpu_wq is fully
3063 if (WARN_ON_ONCE(!mm_percpu_wq))
3067 * Do not drain if one is already in progress unless it's specific to
3068 * a zone. Such callers are primarily CMA and memory hotplug and need
3069 * the drain to be complete when the call returns.
3071 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3074 mutex_lock(&pcpu_drain_mutex);
3078 * We don't care about racing with CPU hotplug event
3079 * as offline notification will cause the notified
3080 * cpu to drain that CPU pcps and on_each_cpu_mask
3081 * disables preemption as part of its processing
3083 for_each_online_cpu(cpu) {
3084 struct per_cpu_pageset *pcp;
3086 bool has_pcps = false;
3088 if (force_all_cpus) {
3090 * The pcp.count check is racy, some callers need a
3091 * guarantee that no cpu is missed.
3095 pcp = per_cpu_ptr(zone->pageset, cpu);
3099 for_each_populated_zone(z) {
3100 pcp = per_cpu_ptr(z->pageset, cpu);
3101 if (pcp->pcp.count) {
3109 cpumask_set_cpu(cpu, &cpus_with_pcps);
3111 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3114 for_each_cpu(cpu, &cpus_with_pcps) {
3115 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3118 INIT_WORK(&drain->work, drain_local_pages_wq);
3119 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3121 for_each_cpu(cpu, &cpus_with_pcps)
3122 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3124 mutex_unlock(&pcpu_drain_mutex);
3128 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3130 * When zone parameter is non-NULL, spill just the single zone's pages.
3132 * Note that this can be extremely slow as the draining happens in a workqueue.
3134 void drain_all_pages(struct zone *zone)
3136 __drain_all_pages(zone, false);
3139 #ifdef CONFIG_HIBERNATION
3142 * Touch the watchdog for every WD_PAGE_COUNT pages.
3144 #define WD_PAGE_COUNT (128*1024)
3146 void mark_free_pages(struct zone *zone)
3148 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3149 unsigned long flags;
3150 unsigned int order, t;
3153 if (zone_is_empty(zone))
3156 spin_lock_irqsave(&zone->lock, flags);
3158 max_zone_pfn = zone_end_pfn(zone);
3159 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3160 if (pfn_valid(pfn)) {
3161 page = pfn_to_page(pfn);
3163 if (!--page_count) {
3164 touch_nmi_watchdog();
3165 page_count = WD_PAGE_COUNT;
3168 if (page_zone(page) != zone)
3171 if (!swsusp_page_is_forbidden(page))
3172 swsusp_unset_page_free(page);
3175 for_each_migratetype_order(order, t) {
3176 list_for_each_entry(page,
3177 &zone->free_area[order].free_list[t], lru) {
3180 pfn = page_to_pfn(page);
3181 for (i = 0; i < (1UL << order); i++) {
3182 if (!--page_count) {
3183 touch_nmi_watchdog();
3184 page_count = WD_PAGE_COUNT;
3186 swsusp_set_page_free(pfn_to_page(pfn + i));
3190 spin_unlock_irqrestore(&zone->lock, flags);
3192 #endif /* CONFIG_PM */
3194 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3198 if (!free_pcp_prepare(page))
3201 migratetype = get_pfnblock_migratetype(page, pfn);
3202 set_pcppage_migratetype(page, migratetype);
3206 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3208 struct zone *zone = page_zone(page);
3209 struct per_cpu_pages *pcp;
3212 migratetype = get_pcppage_migratetype(page);
3213 __count_vm_event(PGFREE);
3216 * We only track unmovable, reclaimable and movable on pcp lists.
3217 * Free ISOLATE pages back to the allocator because they are being
3218 * offlined but treat HIGHATOMIC as movable pages so we can get those
3219 * areas back if necessary. Otherwise, we may have to free
3220 * excessively into the page allocator
3222 if (migratetype >= MIGRATE_PCPTYPES) {
3223 if (unlikely(is_migrate_isolate(migratetype))) {
3224 free_one_page(zone, page, pfn, 0, migratetype,
3228 migratetype = MIGRATE_MOVABLE;
3231 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3232 list_add(&page->lru, &pcp->lists[migratetype]);
3234 if (pcp->count >= READ_ONCE(pcp->high))
3235 free_pcppages_bulk(zone, READ_ONCE(pcp->batch), pcp);
3239 * Free a 0-order page
3241 void free_unref_page(struct page *page)
3243 unsigned long flags;
3244 unsigned long pfn = page_to_pfn(page);
3246 if (!free_unref_page_prepare(page, pfn))
3249 local_irq_save(flags);
3250 free_unref_page_commit(page, pfn);
3251 local_irq_restore(flags);
3255 * Free a list of 0-order pages
3257 void free_unref_page_list(struct list_head *list)
3259 struct page *page, *next;
3260 unsigned long flags, pfn;
3261 int batch_count = 0;
3263 /* Prepare pages for freeing */
3264 list_for_each_entry_safe(page, next, list, lru) {
3265 pfn = page_to_pfn(page);
3266 if (!free_unref_page_prepare(page, pfn))
3267 list_del(&page->lru);
3268 set_page_private(page, pfn);
3271 local_irq_save(flags);
3272 list_for_each_entry_safe(page, next, list, lru) {
3273 unsigned long pfn = page_private(page);
3275 set_page_private(page, 0);
3276 trace_mm_page_free_batched(page);
3277 free_unref_page_commit(page, pfn);
3280 * Guard against excessive IRQ disabled times when we get
3281 * a large list of pages to free.
3283 if (++batch_count == SWAP_CLUSTER_MAX) {
3284 local_irq_restore(flags);
3286 local_irq_save(flags);
3289 local_irq_restore(flags);
3293 * split_page takes a non-compound higher-order page, and splits it into
3294 * n (1<<order) sub-pages: page[0..n]
3295 * Each sub-page must be freed individually.
3297 * Note: this is probably too low level an operation for use in drivers.
3298 * Please consult with lkml before using this in your driver.
3300 void split_page(struct page *page, unsigned int order)
3304 VM_BUG_ON_PAGE(PageCompound(page), page);
3305 VM_BUG_ON_PAGE(!page_count(page), page);
3307 for (i = 1; i < (1 << order); i++)
3308 set_page_refcounted(page + i);
3309 split_page_owner(page, 1 << order);
3311 EXPORT_SYMBOL_GPL(split_page);
3313 int __isolate_free_page(struct page *page, unsigned int order)
3315 unsigned long watermark;
3319 BUG_ON(!PageBuddy(page));
3321 zone = page_zone(page);
3322 mt = get_pageblock_migratetype(page);
3324 if (!is_migrate_isolate(mt)) {
3326 * Obey watermarks as if the page was being allocated. We can
3327 * emulate a high-order watermark check with a raised order-0
3328 * watermark, because we already know our high-order page
3331 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3332 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3335 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3338 /* Remove page from free list */
3340 del_page_from_free_list(page, zone, order);
3343 * Set the pageblock if the isolated page is at least half of a
3346 if (order >= pageblock_order - 1) {
3347 struct page *endpage = page + (1 << order) - 1;
3348 for (; page < endpage; page += pageblock_nr_pages) {
3349 int mt = get_pageblock_migratetype(page);
3350 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3351 && !is_migrate_highatomic(mt))
3352 set_pageblock_migratetype(page,
3358 return 1UL << order;
3362 * __putback_isolated_page - Return a now-isolated page back where we got it
3363 * @page: Page that was isolated
3364 * @order: Order of the isolated page
3365 * @mt: The page's pageblock's migratetype
3367 * This function is meant to return a page pulled from the free lists via
3368 * __isolate_free_page back to the free lists they were pulled from.
3370 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3372 struct zone *zone = page_zone(page);
3374 /* zone lock should be held when this function is called */
3375 lockdep_assert_held(&zone->lock);
3377 /* Return isolated page to tail of freelist. */
3378 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3379 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3383 * Update NUMA hit/miss statistics
3385 * Must be called with interrupts disabled.
3387 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3390 enum numa_stat_item local_stat = NUMA_LOCAL;
3392 /* skip numa counters update if numa stats is disabled */
3393 if (!static_branch_likely(&vm_numa_stat_key))
3396 if (zone_to_nid(z) != numa_node_id())
3397 local_stat = NUMA_OTHER;
3399 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3400 __inc_numa_state(z, NUMA_HIT);
3402 __inc_numa_state(z, NUMA_MISS);
3403 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3405 __inc_numa_state(z, local_stat);
3409 /* Remove page from the per-cpu list, caller must protect the list */
3410 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3411 unsigned int alloc_flags,
3412 struct per_cpu_pages *pcp,
3413 struct list_head *list)
3418 if (list_empty(list)) {
3419 pcp->count += rmqueue_bulk(zone, 0,
3420 READ_ONCE(pcp->batch), list,
3421 migratetype, alloc_flags);
3422 if (unlikely(list_empty(list)))
3426 page = list_first_entry(list, struct page, lru);
3427 list_del(&page->lru);
3429 } while (check_new_pcp(page));
3434 /* Lock and remove page from the per-cpu list */
3435 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3436 struct zone *zone, gfp_t gfp_flags,
3437 int migratetype, unsigned int alloc_flags)
3439 struct per_cpu_pages *pcp;
3440 struct list_head *list;
3442 unsigned long flags;
3444 local_irq_save(flags);
3445 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3446 list = &pcp->lists[migratetype];
3447 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3449 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3450 zone_statistics(preferred_zone, zone);
3452 local_irq_restore(flags);
3457 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3460 struct page *rmqueue(struct zone *preferred_zone,
3461 struct zone *zone, unsigned int order,
3462 gfp_t gfp_flags, unsigned int alloc_flags,
3465 unsigned long flags;
3468 if (likely(order == 0)) {
3470 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3471 * we need to skip it when CMA area isn't allowed.
3473 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3474 migratetype != MIGRATE_MOVABLE) {
3475 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3476 migratetype, alloc_flags);
3482 * We most definitely don't want callers attempting to
3483 * allocate greater than order-1 page units with __GFP_NOFAIL.
3485 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3486 spin_lock_irqsave(&zone->lock, flags);
3491 * order-0 request can reach here when the pcplist is skipped
3492 * due to non-CMA allocation context. HIGHATOMIC area is
3493 * reserved for high-order atomic allocation, so order-0
3494 * request should skip it.
3496 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3497 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3499 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3502 page = __rmqueue(zone, order, migratetype, alloc_flags);
3503 } while (page && check_new_pages(page, order));
3504 spin_unlock(&zone->lock);
3507 __mod_zone_freepage_state(zone, -(1 << order),
3508 get_pcppage_migratetype(page));
3510 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3511 zone_statistics(preferred_zone, zone);
3512 local_irq_restore(flags);
3515 /* Separate test+clear to avoid unnecessary atomics */
3516 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3517 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3518 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3521 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3525 local_irq_restore(flags);
3529 #ifdef CONFIG_FAIL_PAGE_ALLOC
3532 struct fault_attr attr;
3534 bool ignore_gfp_highmem;
3535 bool ignore_gfp_reclaim;
3537 } fail_page_alloc = {
3538 .attr = FAULT_ATTR_INITIALIZER,
3539 .ignore_gfp_reclaim = true,
3540 .ignore_gfp_highmem = true,
3544 static int __init setup_fail_page_alloc(char *str)
3546 return setup_fault_attr(&fail_page_alloc.attr, str);
3548 __setup("fail_page_alloc=", setup_fail_page_alloc);
3550 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3552 if (order < fail_page_alloc.min_order)
3554 if (gfp_mask & __GFP_NOFAIL)
3556 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3558 if (fail_page_alloc.ignore_gfp_reclaim &&
3559 (gfp_mask & __GFP_DIRECT_RECLAIM))
3562 return should_fail(&fail_page_alloc.attr, 1 << order);
3565 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3567 static int __init fail_page_alloc_debugfs(void)
3569 umode_t mode = S_IFREG | 0600;
3572 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3573 &fail_page_alloc.attr);
3575 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3576 &fail_page_alloc.ignore_gfp_reclaim);
3577 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3578 &fail_page_alloc.ignore_gfp_highmem);
3579 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3584 late_initcall(fail_page_alloc_debugfs);
3586 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3588 #else /* CONFIG_FAIL_PAGE_ALLOC */
3590 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3595 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3597 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3599 return __should_fail_alloc_page(gfp_mask, order);
3601 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3603 static inline long __zone_watermark_unusable_free(struct zone *z,
3604 unsigned int order, unsigned int alloc_flags)
3606 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3607 long unusable_free = (1 << order) - 1;
3610 * If the caller does not have rights to ALLOC_HARDER then subtract
3611 * the high-atomic reserves. This will over-estimate the size of the
3612 * atomic reserve but it avoids a search.
3614 if (likely(!alloc_harder))
3615 unusable_free += z->nr_reserved_highatomic;
3618 /* If allocation can't use CMA areas don't use free CMA pages */
3619 if (!(alloc_flags & ALLOC_CMA))
3620 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3623 return unusable_free;
3627 * Return true if free base pages are above 'mark'. For high-order checks it
3628 * will return true of the order-0 watermark is reached and there is at least
3629 * one free page of a suitable size. Checking now avoids taking the zone lock
3630 * to check in the allocation paths if no pages are free.
3632 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3633 int highest_zoneidx, unsigned int alloc_flags,
3638 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3640 /* free_pages may go negative - that's OK */
3641 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3643 if (alloc_flags & ALLOC_HIGH)
3646 if (unlikely(alloc_harder)) {
3648 * OOM victims can try even harder than normal ALLOC_HARDER
3649 * users on the grounds that it's definitely going to be in
3650 * the exit path shortly and free memory. Any allocation it
3651 * makes during the free path will be small and short-lived.
3653 if (alloc_flags & ALLOC_OOM)
3660 * Check watermarks for an order-0 allocation request. If these
3661 * are not met, then a high-order request also cannot go ahead
3662 * even if a suitable page happened to be free.
3664 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3667 /* If this is an order-0 request then the watermark is fine */
3671 /* For a high-order request, check at least one suitable page is free */
3672 for (o = order; o < MAX_ORDER; o++) {
3673 struct free_area *area = &z->free_area[o];
3679 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3680 if (!free_area_empty(area, mt))
3685 if ((alloc_flags & ALLOC_CMA) &&
3686 !free_area_empty(area, MIGRATE_CMA)) {
3690 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3696 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3697 int highest_zoneidx, unsigned int alloc_flags)
3699 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3700 zone_page_state(z, NR_FREE_PAGES));
3703 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3704 unsigned long mark, int highest_zoneidx,
3705 unsigned int alloc_flags, gfp_t gfp_mask)
3709 free_pages = zone_page_state(z, NR_FREE_PAGES);
3712 * Fast check for order-0 only. If this fails then the reserves
3713 * need to be calculated.
3718 fast_free = free_pages;
3719 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3720 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3724 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3728 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3729 * when checking the min watermark. The min watermark is the
3730 * point where boosting is ignored so that kswapd is woken up
3731 * when below the low watermark.
3733 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3734 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3735 mark = z->_watermark[WMARK_MIN];
3736 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3737 alloc_flags, free_pages);
3743 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3744 unsigned long mark, int highest_zoneidx)
3746 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3748 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3749 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3751 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3756 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3758 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3759 node_reclaim_distance;
3761 #else /* CONFIG_NUMA */
3762 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3766 #endif /* CONFIG_NUMA */
3769 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3770 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3771 * premature use of a lower zone may cause lowmem pressure problems that
3772 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3773 * probably too small. It only makes sense to spread allocations to avoid
3774 * fragmentation between the Normal and DMA32 zones.
3776 static inline unsigned int
3777 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3779 unsigned int alloc_flags;
3782 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3785 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3787 #ifdef CONFIG_ZONE_DMA32
3791 if (zone_idx(zone) != ZONE_NORMAL)
3795 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3796 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3797 * on UMA that if Normal is populated then so is DMA32.
3799 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3800 if (nr_online_nodes > 1 && !populated_zone(--zone))
3803 alloc_flags |= ALLOC_NOFRAGMENT;
3804 #endif /* CONFIG_ZONE_DMA32 */
3808 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3809 unsigned int alloc_flags)
3812 unsigned int pflags = current->flags;
3814 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3815 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3816 alloc_flags |= ALLOC_CMA;
3823 * get_page_from_freelist goes through the zonelist trying to allocate
3826 static struct page *
3827 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3828 const struct alloc_context *ac)
3832 struct pglist_data *last_pgdat_dirty_limit = NULL;
3837 * Scan zonelist, looking for a zone with enough free.
3838 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3840 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3841 z = ac->preferred_zoneref;
3842 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3847 if (cpusets_enabled() &&
3848 (alloc_flags & ALLOC_CPUSET) &&
3849 !__cpuset_zone_allowed(zone, gfp_mask))
3852 * When allocating a page cache page for writing, we
3853 * want to get it from a node that is within its dirty
3854 * limit, such that no single node holds more than its
3855 * proportional share of globally allowed dirty pages.
3856 * The dirty limits take into account the node's
3857 * lowmem reserves and high watermark so that kswapd
3858 * should be able to balance it without having to
3859 * write pages from its LRU list.
3861 * XXX: For now, allow allocations to potentially
3862 * exceed the per-node dirty limit in the slowpath
3863 * (spread_dirty_pages unset) before going into reclaim,
3864 * which is important when on a NUMA setup the allowed
3865 * nodes are together not big enough to reach the
3866 * global limit. The proper fix for these situations
3867 * will require awareness of nodes in the
3868 * dirty-throttling and the flusher threads.
3870 if (ac->spread_dirty_pages) {
3871 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3874 if (!node_dirty_ok(zone->zone_pgdat)) {
3875 last_pgdat_dirty_limit = zone->zone_pgdat;
3880 if (no_fallback && nr_online_nodes > 1 &&
3881 zone != ac->preferred_zoneref->zone) {
3885 * If moving to a remote node, retry but allow
3886 * fragmenting fallbacks. Locality is more important
3887 * than fragmentation avoidance.
3889 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3890 if (zone_to_nid(zone) != local_nid) {
3891 alloc_flags &= ~ALLOC_NOFRAGMENT;
3896 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3897 if (!zone_watermark_fast(zone, order, mark,
3898 ac->highest_zoneidx, alloc_flags,
3902 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3904 * Watermark failed for this zone, but see if we can
3905 * grow this zone if it contains deferred pages.
3907 if (static_branch_unlikely(&deferred_pages)) {
3908 if (_deferred_grow_zone(zone, order))
3912 /* Checked here to keep the fast path fast */
3913 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3914 if (alloc_flags & ALLOC_NO_WATERMARKS)
3917 if (node_reclaim_mode == 0 ||
3918 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3921 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3923 case NODE_RECLAIM_NOSCAN:
3926 case NODE_RECLAIM_FULL:
3927 /* scanned but unreclaimable */
3930 /* did we reclaim enough */
3931 if (zone_watermark_ok(zone, order, mark,
3932 ac->highest_zoneidx, alloc_flags))
3940 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3941 gfp_mask, alloc_flags, ac->migratetype);
3943 prep_new_page(page, order, gfp_mask, alloc_flags);
3946 * If this is a high-order atomic allocation then check
3947 * if the pageblock should be reserved for the future
3949 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3950 reserve_highatomic_pageblock(page, zone, order);
3954 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3955 /* Try again if zone has deferred pages */
3956 if (static_branch_unlikely(&deferred_pages)) {
3957 if (_deferred_grow_zone(zone, order))
3965 * It's possible on a UMA machine to get through all zones that are
3966 * fragmented. If avoiding fragmentation, reset and try again.
3969 alloc_flags &= ~ALLOC_NOFRAGMENT;
3976 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3978 unsigned int filter = SHOW_MEM_FILTER_NODES;
3981 * This documents exceptions given to allocations in certain
3982 * contexts that are allowed to allocate outside current's set
3985 if (!(gfp_mask & __GFP_NOMEMALLOC))
3986 if (tsk_is_oom_victim(current) ||
3987 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3988 filter &= ~SHOW_MEM_FILTER_NODES;
3989 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3990 filter &= ~SHOW_MEM_FILTER_NODES;
3992 show_mem(filter, nodemask);
3995 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3997 struct va_format vaf;
3999 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4001 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4004 va_start(args, fmt);
4007 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4008 current->comm, &vaf, gfp_mask, &gfp_mask,
4009 nodemask_pr_args(nodemask));
4012 cpuset_print_current_mems_allowed();
4015 warn_alloc_show_mem(gfp_mask, nodemask);
4018 static inline struct page *
4019 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4020 unsigned int alloc_flags,
4021 const struct alloc_context *ac)
4025 page = get_page_from_freelist(gfp_mask, order,
4026 alloc_flags|ALLOC_CPUSET, ac);
4028 * fallback to ignore cpuset restriction if our nodes
4032 page = get_page_from_freelist(gfp_mask, order,
4038 static inline struct page *
4039 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4040 const struct alloc_context *ac, unsigned long *did_some_progress)
4042 struct oom_control oc = {
4043 .zonelist = ac->zonelist,
4044 .nodemask = ac->nodemask,
4046 .gfp_mask = gfp_mask,
4051 *did_some_progress = 0;
4054 * Acquire the oom lock. If that fails, somebody else is
4055 * making progress for us.
4057 if (!mutex_trylock(&oom_lock)) {
4058 *did_some_progress = 1;
4059 schedule_timeout_uninterruptible(1);
4064 * Go through the zonelist yet one more time, keep very high watermark
4065 * here, this is only to catch a parallel oom killing, we must fail if
4066 * we're still under heavy pressure. But make sure that this reclaim
4067 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4068 * allocation which will never fail due to oom_lock already held.
4070 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4071 ~__GFP_DIRECT_RECLAIM, order,
4072 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4076 /* Coredumps can quickly deplete all memory reserves */
4077 if (current->flags & PF_DUMPCORE)
4079 /* The OOM killer will not help higher order allocs */
4080 if (order > PAGE_ALLOC_COSTLY_ORDER)
4083 * We have already exhausted all our reclaim opportunities without any
4084 * success so it is time to admit defeat. We will skip the OOM killer
4085 * because it is very likely that the caller has a more reasonable
4086 * fallback than shooting a random task.
4088 * The OOM killer may not free memory on a specific node.
4090 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4092 /* The OOM killer does not needlessly kill tasks for lowmem */
4093 if (ac->highest_zoneidx < ZONE_NORMAL)
4095 if (pm_suspended_storage())
4098 * XXX: GFP_NOFS allocations should rather fail than rely on
4099 * other request to make a forward progress.
4100 * We are in an unfortunate situation where out_of_memory cannot
4101 * do much for this context but let's try it to at least get
4102 * access to memory reserved if the current task is killed (see
4103 * out_of_memory). Once filesystems are ready to handle allocation
4104 * failures more gracefully we should just bail out here.
4107 /* Exhausted what can be done so it's blame time */
4108 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4109 *did_some_progress = 1;
4112 * Help non-failing allocations by giving them access to memory
4115 if (gfp_mask & __GFP_NOFAIL)
4116 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4117 ALLOC_NO_WATERMARKS, ac);
4120 mutex_unlock(&oom_lock);
4125 * Maximum number of compaction retries wit a progress before OOM
4126 * killer is consider as the only way to move forward.
4128 #define MAX_COMPACT_RETRIES 16
4130 #ifdef CONFIG_COMPACTION
4131 /* Try memory compaction for high-order allocations before reclaim */
4132 static struct page *
4133 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4134 unsigned int alloc_flags, const struct alloc_context *ac,
4135 enum compact_priority prio, enum compact_result *compact_result)
4137 struct page *page = NULL;
4138 unsigned long pflags;
4139 unsigned int noreclaim_flag;
4144 psi_memstall_enter(&pflags);
4145 noreclaim_flag = memalloc_noreclaim_save();
4147 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4150 memalloc_noreclaim_restore(noreclaim_flag);
4151 psi_memstall_leave(&pflags);
4154 * At least in one zone compaction wasn't deferred or skipped, so let's
4155 * count a compaction stall
4157 count_vm_event(COMPACTSTALL);
4159 /* Prep a captured page if available */
4161 prep_new_page(page, order, gfp_mask, alloc_flags);
4163 /* Try get a page from the freelist if available */
4165 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4168 struct zone *zone = page_zone(page);
4170 zone->compact_blockskip_flush = false;
4171 compaction_defer_reset(zone, order, true);
4172 count_vm_event(COMPACTSUCCESS);
4177 * It's bad if compaction run occurs and fails. The most likely reason
4178 * is that pages exist, but not enough to satisfy watermarks.
4180 count_vm_event(COMPACTFAIL);
4188 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4189 enum compact_result compact_result,
4190 enum compact_priority *compact_priority,
4191 int *compaction_retries)
4193 int max_retries = MAX_COMPACT_RETRIES;
4196 int retries = *compaction_retries;
4197 enum compact_priority priority = *compact_priority;
4202 if (compaction_made_progress(compact_result))
4203 (*compaction_retries)++;
4206 * compaction considers all the zone as desperately out of memory
4207 * so it doesn't really make much sense to retry except when the
4208 * failure could be caused by insufficient priority
4210 if (compaction_failed(compact_result))
4211 goto check_priority;
4214 * compaction was skipped because there are not enough order-0 pages
4215 * to work with, so we retry only if it looks like reclaim can help.
4217 if (compaction_needs_reclaim(compact_result)) {
4218 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4223 * make sure the compaction wasn't deferred or didn't bail out early
4224 * due to locks contention before we declare that we should give up.
4225 * But the next retry should use a higher priority if allowed, so
4226 * we don't just keep bailing out endlessly.
4228 if (compaction_withdrawn(compact_result)) {
4229 goto check_priority;
4233 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4234 * costly ones because they are de facto nofail and invoke OOM
4235 * killer to move on while costly can fail and users are ready
4236 * to cope with that. 1/4 retries is rather arbitrary but we
4237 * would need much more detailed feedback from compaction to
4238 * make a better decision.
4240 if (order > PAGE_ALLOC_COSTLY_ORDER)
4242 if (*compaction_retries <= max_retries) {
4248 * Make sure there are attempts at the highest priority if we exhausted
4249 * all retries or failed at the lower priorities.
4252 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4253 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4255 if (*compact_priority > min_priority) {
4256 (*compact_priority)--;
4257 *compaction_retries = 0;
4261 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4265 static inline struct page *
4266 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4267 unsigned int alloc_flags, const struct alloc_context *ac,
4268 enum compact_priority prio, enum compact_result *compact_result)
4270 *compact_result = COMPACT_SKIPPED;
4275 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4276 enum compact_result compact_result,
4277 enum compact_priority *compact_priority,
4278 int *compaction_retries)
4283 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4287 * There are setups with compaction disabled which would prefer to loop
4288 * inside the allocator rather than hit the oom killer prematurely.
4289 * Let's give them a good hope and keep retrying while the order-0
4290 * watermarks are OK.
4292 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4293 ac->highest_zoneidx, ac->nodemask) {
4294 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4295 ac->highest_zoneidx, alloc_flags))
4300 #endif /* CONFIG_COMPACTION */
4302 #ifdef CONFIG_LOCKDEP
4303 static struct lockdep_map __fs_reclaim_map =
4304 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4306 static bool __need_reclaim(gfp_t gfp_mask)
4308 /* no reclaim without waiting on it */
4309 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4312 /* this guy won't enter reclaim */
4313 if (current->flags & PF_MEMALLOC)
4316 if (gfp_mask & __GFP_NOLOCKDEP)
4322 void __fs_reclaim_acquire(void)
4324 lock_map_acquire(&__fs_reclaim_map);
4327 void __fs_reclaim_release(void)
4329 lock_map_release(&__fs_reclaim_map);
4332 void fs_reclaim_acquire(gfp_t gfp_mask)
4334 gfp_mask = current_gfp_context(gfp_mask);
4336 if (__need_reclaim(gfp_mask)) {
4337 if (gfp_mask & __GFP_FS)
4338 __fs_reclaim_acquire();
4340 #ifdef CONFIG_MMU_NOTIFIER
4341 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4342 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4347 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4349 void fs_reclaim_release(gfp_t gfp_mask)
4351 gfp_mask = current_gfp_context(gfp_mask);
4353 if (__need_reclaim(gfp_mask)) {
4354 if (gfp_mask & __GFP_FS)
4355 __fs_reclaim_release();
4358 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4361 /* Perform direct synchronous page reclaim */
4362 static unsigned long
4363 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4364 const struct alloc_context *ac)
4366 unsigned int noreclaim_flag;
4367 unsigned long pflags, progress;
4371 /* We now go into synchronous reclaim */
4372 cpuset_memory_pressure_bump();
4373 psi_memstall_enter(&pflags);
4374 fs_reclaim_acquire(gfp_mask);
4375 noreclaim_flag = memalloc_noreclaim_save();
4377 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4380 memalloc_noreclaim_restore(noreclaim_flag);
4381 fs_reclaim_release(gfp_mask);
4382 psi_memstall_leave(&pflags);
4389 /* The really slow allocator path where we enter direct reclaim */
4390 static inline struct page *
4391 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4392 unsigned int alloc_flags, const struct alloc_context *ac,
4393 unsigned long *did_some_progress)
4395 struct page *page = NULL;
4396 bool drained = false;
4398 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4399 if (unlikely(!(*did_some_progress)))
4403 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4406 * If an allocation failed after direct reclaim, it could be because
4407 * pages are pinned on the per-cpu lists or in high alloc reserves.
4408 * Shrink them and try again
4410 if (!page && !drained) {
4411 unreserve_highatomic_pageblock(ac, false);
4412 drain_all_pages(NULL);
4420 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4421 const struct alloc_context *ac)
4425 pg_data_t *last_pgdat = NULL;
4426 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4428 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4430 if (last_pgdat != zone->zone_pgdat)
4431 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4432 last_pgdat = zone->zone_pgdat;
4436 static inline unsigned int
4437 gfp_to_alloc_flags(gfp_t gfp_mask)
4439 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4442 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4443 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4444 * to save two branches.
4446 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4447 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4450 * The caller may dip into page reserves a bit more if the caller
4451 * cannot run direct reclaim, or if the caller has realtime scheduling
4452 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4453 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4455 alloc_flags |= (__force int)
4456 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4458 if (gfp_mask & __GFP_ATOMIC) {
4460 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4461 * if it can't schedule.
4463 if (!(gfp_mask & __GFP_NOMEMALLOC))
4464 alloc_flags |= ALLOC_HARDER;
4466 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4467 * comment for __cpuset_node_allowed().
4469 alloc_flags &= ~ALLOC_CPUSET;
4470 } else if (unlikely(rt_task(current)) && !in_interrupt())
4471 alloc_flags |= ALLOC_HARDER;
4473 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4478 static bool oom_reserves_allowed(struct task_struct *tsk)
4480 if (!tsk_is_oom_victim(tsk))
4484 * !MMU doesn't have oom reaper so give access to memory reserves
4485 * only to the thread with TIF_MEMDIE set
4487 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4494 * Distinguish requests which really need access to full memory
4495 * reserves from oom victims which can live with a portion of it
4497 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4499 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4501 if (gfp_mask & __GFP_MEMALLOC)
4502 return ALLOC_NO_WATERMARKS;
4503 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4504 return ALLOC_NO_WATERMARKS;
4505 if (!in_interrupt()) {
4506 if (current->flags & PF_MEMALLOC)
4507 return ALLOC_NO_WATERMARKS;
4508 else if (oom_reserves_allowed(current))
4515 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4517 return !!__gfp_pfmemalloc_flags(gfp_mask);
4521 * Checks whether it makes sense to retry the reclaim to make a forward progress
4522 * for the given allocation request.
4524 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4525 * without success, or when we couldn't even meet the watermark if we
4526 * reclaimed all remaining pages on the LRU lists.
4528 * Returns true if a retry is viable or false to enter the oom path.
4531 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4532 struct alloc_context *ac, int alloc_flags,
4533 bool did_some_progress, int *no_progress_loops)
4540 * Costly allocations might have made a progress but this doesn't mean
4541 * their order will become available due to high fragmentation so
4542 * always increment the no progress counter for them
4544 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4545 *no_progress_loops = 0;
4547 (*no_progress_loops)++;
4550 * Make sure we converge to OOM if we cannot make any progress
4551 * several times in the row.
4553 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4554 /* Before OOM, exhaust highatomic_reserve */
4555 return unreserve_highatomic_pageblock(ac, true);
4559 * Keep reclaiming pages while there is a chance this will lead
4560 * somewhere. If none of the target zones can satisfy our allocation
4561 * request even if all reclaimable pages are considered then we are
4562 * screwed and have to go OOM.
4564 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4565 ac->highest_zoneidx, ac->nodemask) {
4566 unsigned long available;
4567 unsigned long reclaimable;
4568 unsigned long min_wmark = min_wmark_pages(zone);
4571 available = reclaimable = zone_reclaimable_pages(zone);
4572 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4575 * Would the allocation succeed if we reclaimed all
4576 * reclaimable pages?
4578 wmark = __zone_watermark_ok(zone, order, min_wmark,
4579 ac->highest_zoneidx, alloc_flags, available);
4580 trace_reclaim_retry_zone(z, order, reclaimable,
4581 available, min_wmark, *no_progress_loops, wmark);
4584 * If we didn't make any progress and have a lot of
4585 * dirty + writeback pages then we should wait for
4586 * an IO to complete to slow down the reclaim and
4587 * prevent from pre mature OOM
4589 if (!did_some_progress) {
4590 unsigned long write_pending;
4592 write_pending = zone_page_state_snapshot(zone,
4593 NR_ZONE_WRITE_PENDING);
4595 if (2 * write_pending > reclaimable) {
4596 congestion_wait(BLK_RW_ASYNC, HZ/10);
4608 * Memory allocation/reclaim might be called from a WQ context and the
4609 * current implementation of the WQ concurrency control doesn't
4610 * recognize that a particular WQ is congested if the worker thread is
4611 * looping without ever sleeping. Therefore we have to do a short sleep
4612 * here rather than calling cond_resched().
4614 if (current->flags & PF_WQ_WORKER)
4615 schedule_timeout_uninterruptible(1);
4622 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4625 * It's possible that cpuset's mems_allowed and the nodemask from
4626 * mempolicy don't intersect. This should be normally dealt with by
4627 * policy_nodemask(), but it's possible to race with cpuset update in
4628 * such a way the check therein was true, and then it became false
4629 * before we got our cpuset_mems_cookie here.
4630 * This assumes that for all allocations, ac->nodemask can come only
4631 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4632 * when it does not intersect with the cpuset restrictions) or the
4633 * caller can deal with a violated nodemask.
4635 if (cpusets_enabled() && ac->nodemask &&
4636 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4637 ac->nodemask = NULL;
4642 * When updating a task's mems_allowed or mempolicy nodemask, it is
4643 * possible to race with parallel threads in such a way that our
4644 * allocation can fail while the mask is being updated. If we are about
4645 * to fail, check if the cpuset changed during allocation and if so,
4648 if (read_mems_allowed_retry(cpuset_mems_cookie))
4654 static inline struct page *
4655 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4656 struct alloc_context *ac)
4658 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4659 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4660 struct page *page = NULL;
4661 unsigned int alloc_flags;
4662 unsigned long did_some_progress;
4663 enum compact_priority compact_priority;
4664 enum compact_result compact_result;
4665 int compaction_retries;
4666 int no_progress_loops;
4667 unsigned int cpuset_mems_cookie;
4671 * We also sanity check to catch abuse of atomic reserves being used by
4672 * callers that are not in atomic context.
4674 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4675 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4676 gfp_mask &= ~__GFP_ATOMIC;
4679 compaction_retries = 0;
4680 no_progress_loops = 0;
4681 compact_priority = DEF_COMPACT_PRIORITY;
4682 cpuset_mems_cookie = read_mems_allowed_begin();
4685 * The fast path uses conservative alloc_flags to succeed only until
4686 * kswapd needs to be woken up, and to avoid the cost of setting up
4687 * alloc_flags precisely. So we do that now.
4689 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4692 * We need to recalculate the starting point for the zonelist iterator
4693 * because we might have used different nodemask in the fast path, or
4694 * there was a cpuset modification and we are retrying - otherwise we
4695 * could end up iterating over non-eligible zones endlessly.
4697 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4698 ac->highest_zoneidx, ac->nodemask);
4699 if (!ac->preferred_zoneref->zone)
4702 if (alloc_flags & ALLOC_KSWAPD)
4703 wake_all_kswapds(order, gfp_mask, ac);
4706 * The adjusted alloc_flags might result in immediate success, so try
4709 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4714 * For costly allocations, try direct compaction first, as it's likely
4715 * that we have enough base pages and don't need to reclaim. For non-
4716 * movable high-order allocations, do that as well, as compaction will
4717 * try prevent permanent fragmentation by migrating from blocks of the
4719 * Don't try this for allocations that are allowed to ignore
4720 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4722 if (can_direct_reclaim &&
4724 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4725 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4726 page = __alloc_pages_direct_compact(gfp_mask, order,
4728 INIT_COMPACT_PRIORITY,
4734 * Checks for costly allocations with __GFP_NORETRY, which
4735 * includes some THP page fault allocations
4737 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4739 * If allocating entire pageblock(s) and compaction
4740 * failed because all zones are below low watermarks
4741 * or is prohibited because it recently failed at this
4742 * order, fail immediately unless the allocator has
4743 * requested compaction and reclaim retry.
4746 * - potentially very expensive because zones are far
4747 * below their low watermarks or this is part of very
4748 * bursty high order allocations,
4749 * - not guaranteed to help because isolate_freepages()
4750 * may not iterate over freed pages as part of its
4752 * - unlikely to make entire pageblocks free on its
4755 if (compact_result == COMPACT_SKIPPED ||
4756 compact_result == COMPACT_DEFERRED)
4760 * Looks like reclaim/compaction is worth trying, but
4761 * sync compaction could be very expensive, so keep
4762 * using async compaction.
4764 compact_priority = INIT_COMPACT_PRIORITY;
4769 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4770 if (alloc_flags & ALLOC_KSWAPD)
4771 wake_all_kswapds(order, gfp_mask, ac);
4773 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4775 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4778 * Reset the nodemask and zonelist iterators if memory policies can be
4779 * ignored. These allocations are high priority and system rather than
4782 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4783 ac->nodemask = NULL;
4784 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4785 ac->highest_zoneidx, ac->nodemask);
4788 /* Attempt with potentially adjusted zonelist and alloc_flags */
4789 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4793 /* Caller is not willing to reclaim, we can't balance anything */
4794 if (!can_direct_reclaim)
4797 /* Avoid recursion of direct reclaim */
4798 if (current->flags & PF_MEMALLOC)
4801 /* Try direct reclaim and then allocating */
4802 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4803 &did_some_progress);
4807 /* Try direct compaction and then allocating */
4808 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4809 compact_priority, &compact_result);
4813 /* Do not loop if specifically requested */
4814 if (gfp_mask & __GFP_NORETRY)
4818 * Do not retry costly high order allocations unless they are
4819 * __GFP_RETRY_MAYFAIL
4821 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4824 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4825 did_some_progress > 0, &no_progress_loops))
4829 * It doesn't make any sense to retry for the compaction if the order-0
4830 * reclaim is not able to make any progress because the current
4831 * implementation of the compaction depends on the sufficient amount
4832 * of free memory (see __compaction_suitable)
4834 if (did_some_progress > 0 &&
4835 should_compact_retry(ac, order, alloc_flags,
4836 compact_result, &compact_priority,
4837 &compaction_retries))
4841 /* Deal with possible cpuset update races before we start OOM killing */
4842 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4845 /* Reclaim has failed us, start killing things */
4846 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4850 /* Avoid allocations with no watermarks from looping endlessly */
4851 if (tsk_is_oom_victim(current) &&
4852 (alloc_flags & ALLOC_OOM ||
4853 (gfp_mask & __GFP_NOMEMALLOC)))
4856 /* Retry as long as the OOM killer is making progress */
4857 if (did_some_progress) {
4858 no_progress_loops = 0;
4863 /* Deal with possible cpuset update races before we fail */
4864 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4868 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4871 if (gfp_mask & __GFP_NOFAIL) {
4873 * All existing users of the __GFP_NOFAIL are blockable, so warn
4874 * of any new users that actually require GFP_NOWAIT
4876 if (WARN_ON_ONCE(!can_direct_reclaim))
4880 * PF_MEMALLOC request from this context is rather bizarre
4881 * because we cannot reclaim anything and only can loop waiting
4882 * for somebody to do a work for us
4884 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4887 * non failing costly orders are a hard requirement which we
4888 * are not prepared for much so let's warn about these users
4889 * so that we can identify them and convert them to something
4892 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4895 * Help non-failing allocations by giving them access to memory
4896 * reserves but do not use ALLOC_NO_WATERMARKS because this
4897 * could deplete whole memory reserves which would just make
4898 * the situation worse
4900 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4908 warn_alloc(gfp_mask, ac->nodemask,
4909 "page allocation failure: order:%u", order);
4914 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4915 int preferred_nid, nodemask_t *nodemask,
4916 struct alloc_context *ac, gfp_t *alloc_mask,
4917 unsigned int *alloc_flags)
4919 ac->highest_zoneidx = gfp_zone(gfp_mask);
4920 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4921 ac->nodemask = nodemask;
4922 ac->migratetype = gfp_migratetype(gfp_mask);
4924 if (cpusets_enabled()) {
4925 *alloc_mask |= __GFP_HARDWALL;
4927 * When we are in the interrupt context, it is irrelevant
4928 * to the current task context. It means that any node ok.
4930 if (!in_interrupt() && !ac->nodemask)
4931 ac->nodemask = &cpuset_current_mems_allowed;
4933 *alloc_flags |= ALLOC_CPUSET;
4936 fs_reclaim_acquire(gfp_mask);
4937 fs_reclaim_release(gfp_mask);
4939 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4941 if (should_fail_alloc_page(gfp_mask, order))
4944 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4946 /* Dirty zone balancing only done in the fast path */
4947 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4950 * The preferred zone is used for statistics but crucially it is
4951 * also used as the starting point for the zonelist iterator. It
4952 * may get reset for allocations that ignore memory policies.
4954 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4955 ac->highest_zoneidx, ac->nodemask);
4961 * This is the 'heart' of the zoned buddy allocator.
4964 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4965 nodemask_t *nodemask)
4968 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4969 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4970 struct alloc_context ac = { };
4973 * There are several places where we assume that the order value is sane
4974 * so bail out early if the request is out of bound.
4976 if (unlikely(order >= MAX_ORDER)) {
4977 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4981 gfp_mask &= gfp_allowed_mask;
4982 alloc_mask = gfp_mask;
4983 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4987 * Forbid the first pass from falling back to types that fragment
4988 * memory until all local zones are considered.
4990 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4992 /* First allocation attempt */
4993 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4998 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4999 * resp. GFP_NOIO which has to be inherited for all allocation requests
5000 * from a particular context which has been marked by
5001 * memalloc_no{fs,io}_{save,restore}.
5003 alloc_mask = current_gfp_context(gfp_mask);
5004 ac.spread_dirty_pages = false;
5007 * Restore the original nodemask if it was potentially replaced with
5008 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5010 ac.nodemask = nodemask;
5012 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
5015 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
5016 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
5017 __free_pages(page, order);
5021 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
5025 EXPORT_SYMBOL(__alloc_pages_nodemask);
5028 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5029 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5030 * you need to access high mem.
5032 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5036 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5039 return (unsigned long) page_address(page);
5041 EXPORT_SYMBOL(__get_free_pages);
5043 unsigned long get_zeroed_page(gfp_t gfp_mask)
5045 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5047 EXPORT_SYMBOL(get_zeroed_page);
5049 static inline void free_the_page(struct page *page, unsigned int order)
5051 if (order == 0) /* Via pcp? */
5052 free_unref_page(page);
5054 __free_pages_ok(page, order, FPI_NONE);
5058 * __free_pages - Free pages allocated with alloc_pages().
5059 * @page: The page pointer returned from alloc_pages().
5060 * @order: The order of the allocation.
5062 * This function can free multi-page allocations that are not compound
5063 * pages. It does not check that the @order passed in matches that of
5064 * the allocation, so it is easy to leak memory. Freeing more memory
5065 * than was allocated will probably emit a warning.
5067 * If the last reference to this page is speculative, it will be released
5068 * by put_page() which only frees the first page of a non-compound
5069 * allocation. To prevent the remaining pages from being leaked, we free
5070 * the subsequent pages here. If you want to use the page's reference
5071 * count to decide when to free the allocation, you should allocate a
5072 * compound page, and use put_page() instead of __free_pages().
5074 * Context: May be called in interrupt context or while holding a normal
5075 * spinlock, but not in NMI context or while holding a raw spinlock.
5077 void __free_pages(struct page *page, unsigned int order)
5079 if (put_page_testzero(page))
5080 free_the_page(page, order);
5081 else if (!PageHead(page))
5083 free_the_page(page + (1 << order), order);
5085 EXPORT_SYMBOL(__free_pages);
5087 void free_pages(unsigned long addr, unsigned int order)
5090 VM_BUG_ON(!virt_addr_valid((void *)addr));
5091 __free_pages(virt_to_page((void *)addr), order);
5095 EXPORT_SYMBOL(free_pages);
5099 * An arbitrary-length arbitrary-offset area of memory which resides
5100 * within a 0 or higher order page. Multiple fragments within that page
5101 * are individually refcounted, in the page's reference counter.
5103 * The page_frag functions below provide a simple allocation framework for
5104 * page fragments. This is used by the network stack and network device
5105 * drivers to provide a backing region of memory for use as either an
5106 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5108 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5111 struct page *page = NULL;
5112 gfp_t gfp = gfp_mask;
5114 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5115 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5117 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5118 PAGE_FRAG_CACHE_MAX_ORDER);
5119 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5121 if (unlikely(!page))
5122 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5124 nc->va = page ? page_address(page) : NULL;
5129 void __page_frag_cache_drain(struct page *page, unsigned int count)
5131 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5133 if (page_ref_sub_and_test(page, count))
5134 free_the_page(page, compound_order(page));
5136 EXPORT_SYMBOL(__page_frag_cache_drain);
5138 void *page_frag_alloc(struct page_frag_cache *nc,
5139 unsigned int fragsz, gfp_t gfp_mask)
5141 unsigned int size = PAGE_SIZE;
5145 if (unlikely(!nc->va)) {
5147 page = __page_frag_cache_refill(nc, gfp_mask);
5151 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5152 /* if size can vary use size else just use PAGE_SIZE */
5155 /* Even if we own the page, we do not use atomic_set().
5156 * This would break get_page_unless_zero() users.
5158 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5160 /* reset page count bias and offset to start of new frag */
5161 nc->pfmemalloc = page_is_pfmemalloc(page);
5162 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5166 offset = nc->offset - fragsz;
5167 if (unlikely(offset < 0)) {
5168 page = virt_to_page(nc->va);
5170 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5173 if (unlikely(nc->pfmemalloc)) {
5174 free_the_page(page, compound_order(page));
5178 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5179 /* if size can vary use size else just use PAGE_SIZE */
5182 /* OK, page count is 0, we can safely set it */
5183 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5185 /* reset page count bias and offset to start of new frag */
5186 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5187 offset = size - fragsz;
5191 nc->offset = offset;
5193 return nc->va + offset;
5195 EXPORT_SYMBOL(page_frag_alloc);
5198 * Frees a page fragment allocated out of either a compound or order 0 page.
5200 void page_frag_free(void *addr)
5202 struct page *page = virt_to_head_page(addr);
5204 if (unlikely(put_page_testzero(page)))
5205 free_the_page(page, compound_order(page));
5207 EXPORT_SYMBOL(page_frag_free);
5209 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5213 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5214 unsigned long used = addr + PAGE_ALIGN(size);
5216 split_page(virt_to_page((void *)addr), order);
5217 while (used < alloc_end) {
5222 return (void *)addr;
5226 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5227 * @size: the number of bytes to allocate
5228 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5230 * This function is similar to alloc_pages(), except that it allocates the
5231 * minimum number of pages to satisfy the request. alloc_pages() can only
5232 * allocate memory in power-of-two pages.
5234 * This function is also limited by MAX_ORDER.
5236 * Memory allocated by this function must be released by free_pages_exact().
5238 * Return: pointer to the allocated area or %NULL in case of error.
5240 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5242 unsigned int order = get_order(size);
5245 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5246 gfp_mask &= ~__GFP_COMP;
5248 addr = __get_free_pages(gfp_mask, order);
5249 return make_alloc_exact(addr, order, size);
5251 EXPORT_SYMBOL(alloc_pages_exact);
5254 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5256 * @nid: the preferred node ID where memory should be allocated
5257 * @size: the number of bytes to allocate
5258 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5260 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5263 * Return: pointer to the allocated area or %NULL in case of error.
5265 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5267 unsigned int order = get_order(size);
5270 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5271 gfp_mask &= ~__GFP_COMP;
5273 p = alloc_pages_node(nid, gfp_mask, order);
5276 return make_alloc_exact((unsigned long)page_address(p), order, size);
5280 * free_pages_exact - release memory allocated via alloc_pages_exact()
5281 * @virt: the value returned by alloc_pages_exact.
5282 * @size: size of allocation, same value as passed to alloc_pages_exact().
5284 * Release the memory allocated by a previous call to alloc_pages_exact.
5286 void free_pages_exact(void *virt, size_t size)
5288 unsigned long addr = (unsigned long)virt;
5289 unsigned long end = addr + PAGE_ALIGN(size);
5291 while (addr < end) {
5296 EXPORT_SYMBOL(free_pages_exact);
5299 * nr_free_zone_pages - count number of pages beyond high watermark
5300 * @offset: The zone index of the highest zone
5302 * nr_free_zone_pages() counts the number of pages which are beyond the
5303 * high watermark within all zones at or below a given zone index. For each
5304 * zone, the number of pages is calculated as:
5306 * nr_free_zone_pages = managed_pages - high_pages
5308 * Return: number of pages beyond high watermark.
5310 static unsigned long nr_free_zone_pages(int offset)
5315 /* Just pick one node, since fallback list is circular */
5316 unsigned long sum = 0;
5318 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5320 for_each_zone_zonelist(zone, z, zonelist, offset) {
5321 unsigned long size = zone_managed_pages(zone);
5322 unsigned long high = high_wmark_pages(zone);
5331 * nr_free_buffer_pages - count number of pages beyond high watermark
5333 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5334 * watermark within ZONE_DMA and ZONE_NORMAL.
5336 * Return: number of pages beyond high watermark within ZONE_DMA and
5339 unsigned long nr_free_buffer_pages(void)
5341 return nr_free_zone_pages(gfp_zone(GFP_USER));
5343 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5345 static inline void show_node(struct zone *zone)
5347 if (IS_ENABLED(CONFIG_NUMA))
5348 printk("Node %d ", zone_to_nid(zone));
5351 long si_mem_available(void)
5354 unsigned long pagecache;
5355 unsigned long wmark_low = 0;
5356 unsigned long pages[NR_LRU_LISTS];
5357 unsigned long reclaimable;
5361 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5362 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5365 wmark_low += low_wmark_pages(zone);
5368 * Estimate the amount of memory available for userspace allocations,
5369 * without causing swapping.
5371 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5374 * Not all the page cache can be freed, otherwise the system will
5375 * start swapping. Assume at least half of the page cache, or the
5376 * low watermark worth of cache, needs to stay.
5378 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5379 pagecache -= min(pagecache / 2, wmark_low);
5380 available += pagecache;
5383 * Part of the reclaimable slab and other kernel memory consists of
5384 * items that are in use, and cannot be freed. Cap this estimate at the
5387 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5388 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5389 available += reclaimable - min(reclaimable / 2, wmark_low);
5395 EXPORT_SYMBOL_GPL(si_mem_available);
5397 void si_meminfo(struct sysinfo *val)
5399 val->totalram = totalram_pages();
5400 val->sharedram = global_node_page_state(NR_SHMEM);
5401 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5402 val->bufferram = nr_blockdev_pages();
5403 val->totalhigh = totalhigh_pages();
5404 val->freehigh = nr_free_highpages();
5405 val->mem_unit = PAGE_SIZE;
5408 EXPORT_SYMBOL(si_meminfo);
5411 void si_meminfo_node(struct sysinfo *val, int nid)
5413 int zone_type; /* needs to be signed */
5414 unsigned long managed_pages = 0;
5415 unsigned long managed_highpages = 0;
5416 unsigned long free_highpages = 0;
5417 pg_data_t *pgdat = NODE_DATA(nid);
5419 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5420 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5421 val->totalram = managed_pages;
5422 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5423 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5424 #ifdef CONFIG_HIGHMEM
5425 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5426 struct zone *zone = &pgdat->node_zones[zone_type];
5428 if (is_highmem(zone)) {
5429 managed_highpages += zone_managed_pages(zone);
5430 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5433 val->totalhigh = managed_highpages;
5434 val->freehigh = free_highpages;
5436 val->totalhigh = managed_highpages;
5437 val->freehigh = free_highpages;
5439 val->mem_unit = PAGE_SIZE;
5444 * Determine whether the node should be displayed or not, depending on whether
5445 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5447 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5449 if (!(flags & SHOW_MEM_FILTER_NODES))
5453 * no node mask - aka implicit memory numa policy. Do not bother with
5454 * the synchronization - read_mems_allowed_begin - because we do not
5455 * have to be precise here.
5458 nodemask = &cpuset_current_mems_allowed;
5460 return !node_isset(nid, *nodemask);
5463 #define K(x) ((x) << (PAGE_SHIFT-10))
5465 static void show_migration_types(unsigned char type)
5467 static const char types[MIGRATE_TYPES] = {
5468 [MIGRATE_UNMOVABLE] = 'U',
5469 [MIGRATE_MOVABLE] = 'M',
5470 [MIGRATE_RECLAIMABLE] = 'E',
5471 [MIGRATE_HIGHATOMIC] = 'H',
5473 [MIGRATE_CMA] = 'C',
5475 #ifdef CONFIG_MEMORY_ISOLATION
5476 [MIGRATE_ISOLATE] = 'I',
5479 char tmp[MIGRATE_TYPES + 1];
5483 for (i = 0; i < MIGRATE_TYPES; i++) {
5484 if (type & (1 << i))
5489 printk(KERN_CONT "(%s) ", tmp);
5493 * Show free area list (used inside shift_scroll-lock stuff)
5494 * We also calculate the percentage fragmentation. We do this by counting the
5495 * memory on each free list with the exception of the first item on the list.
5498 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5501 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5503 unsigned long free_pcp = 0;
5508 for_each_populated_zone(zone) {
5509 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5512 for_each_online_cpu(cpu)
5513 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5516 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5517 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5518 " unevictable:%lu dirty:%lu writeback:%lu\n"
5519 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5520 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5521 " free:%lu free_pcp:%lu free_cma:%lu\n",
5522 global_node_page_state(NR_ACTIVE_ANON),
5523 global_node_page_state(NR_INACTIVE_ANON),
5524 global_node_page_state(NR_ISOLATED_ANON),
5525 global_node_page_state(NR_ACTIVE_FILE),
5526 global_node_page_state(NR_INACTIVE_FILE),
5527 global_node_page_state(NR_ISOLATED_FILE),
5528 global_node_page_state(NR_UNEVICTABLE),
5529 global_node_page_state(NR_FILE_DIRTY),
5530 global_node_page_state(NR_WRITEBACK),
5531 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5532 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5533 global_node_page_state(NR_FILE_MAPPED),
5534 global_node_page_state(NR_SHMEM),
5535 global_node_page_state(NR_PAGETABLE),
5536 global_zone_page_state(NR_BOUNCE),
5537 global_zone_page_state(NR_FREE_PAGES),
5539 global_zone_page_state(NR_FREE_CMA_PAGES));
5541 for_each_online_pgdat(pgdat) {
5542 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5546 " active_anon:%lukB"
5547 " inactive_anon:%lukB"
5548 " active_file:%lukB"
5549 " inactive_file:%lukB"
5550 " unevictable:%lukB"
5551 " isolated(anon):%lukB"
5552 " isolated(file):%lukB"
5557 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5559 " shmem_pmdmapped: %lukB"
5562 " writeback_tmp:%lukB"
5563 " kernel_stack:%lukB"
5564 #ifdef CONFIG_SHADOW_CALL_STACK
5565 " shadow_call_stack:%lukB"
5568 " all_unreclaimable? %s"
5571 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5572 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5573 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5574 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5575 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5576 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5577 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5578 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5579 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5580 K(node_page_state(pgdat, NR_WRITEBACK)),
5581 K(node_page_state(pgdat, NR_SHMEM)),
5582 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5583 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5584 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5586 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5588 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5589 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5590 #ifdef CONFIG_SHADOW_CALL_STACK
5591 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5593 K(node_page_state(pgdat, NR_PAGETABLE)),
5594 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5598 for_each_populated_zone(zone) {
5601 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5605 for_each_online_cpu(cpu)
5606 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5615 " reserved_highatomic:%luKB"
5616 " active_anon:%lukB"
5617 " inactive_anon:%lukB"
5618 " active_file:%lukB"
5619 " inactive_file:%lukB"
5620 " unevictable:%lukB"
5621 " writepending:%lukB"
5631 K(zone_page_state(zone, NR_FREE_PAGES)),
5632 K(min_wmark_pages(zone)),
5633 K(low_wmark_pages(zone)),
5634 K(high_wmark_pages(zone)),
5635 K(zone->nr_reserved_highatomic),
5636 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5637 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5638 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5639 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5640 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5641 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5642 K(zone->present_pages),
5643 K(zone_managed_pages(zone)),
5644 K(zone_page_state(zone, NR_MLOCK)),
5645 K(zone_page_state(zone, NR_BOUNCE)),
5647 K(this_cpu_read(zone->pageset->pcp.count)),
5648 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5649 printk("lowmem_reserve[]:");
5650 for (i = 0; i < MAX_NR_ZONES; i++)
5651 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5652 printk(KERN_CONT "\n");
5655 for_each_populated_zone(zone) {
5657 unsigned long nr[MAX_ORDER], flags, total = 0;
5658 unsigned char types[MAX_ORDER];
5660 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5663 printk(KERN_CONT "%s: ", zone->name);
5665 spin_lock_irqsave(&zone->lock, flags);
5666 for (order = 0; order < MAX_ORDER; order++) {
5667 struct free_area *area = &zone->free_area[order];
5670 nr[order] = area->nr_free;
5671 total += nr[order] << order;
5674 for (type = 0; type < MIGRATE_TYPES; type++) {
5675 if (!free_area_empty(area, type))
5676 types[order] |= 1 << type;
5679 spin_unlock_irqrestore(&zone->lock, flags);
5680 for (order = 0; order < MAX_ORDER; order++) {
5681 printk(KERN_CONT "%lu*%lukB ",
5682 nr[order], K(1UL) << order);
5684 show_migration_types(types[order]);
5686 printk(KERN_CONT "= %lukB\n", K(total));
5689 hugetlb_show_meminfo();
5691 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5693 show_swap_cache_info();
5696 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5698 zoneref->zone = zone;
5699 zoneref->zone_idx = zone_idx(zone);
5703 * Builds allocation fallback zone lists.
5705 * Add all populated zones of a node to the zonelist.
5707 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5710 enum zone_type zone_type = MAX_NR_ZONES;
5715 zone = pgdat->node_zones + zone_type;
5716 if (managed_zone(zone)) {
5717 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5718 check_highest_zone(zone_type);
5720 } while (zone_type);
5727 static int __parse_numa_zonelist_order(char *s)
5730 * We used to support different zonlists modes but they turned
5731 * out to be just not useful. Let's keep the warning in place
5732 * if somebody still use the cmd line parameter so that we do
5733 * not fail it silently
5735 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5736 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5742 char numa_zonelist_order[] = "Node";
5745 * sysctl handler for numa_zonelist_order
5747 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5748 void *buffer, size_t *length, loff_t *ppos)
5751 return __parse_numa_zonelist_order(buffer);
5752 return proc_dostring(table, write, buffer, length, ppos);
5756 #define MAX_NODE_LOAD (nr_online_nodes)
5757 static int node_load[MAX_NUMNODES];
5760 * find_next_best_node - find the next node that should appear in a given node's fallback list
5761 * @node: node whose fallback list we're appending
5762 * @used_node_mask: nodemask_t of already used nodes
5764 * We use a number of factors to determine which is the next node that should
5765 * appear on a given node's fallback list. The node should not have appeared
5766 * already in @node's fallback list, and it should be the next closest node
5767 * according to the distance array (which contains arbitrary distance values
5768 * from each node to each node in the system), and should also prefer nodes
5769 * with no CPUs, since presumably they'll have very little allocation pressure
5770 * on them otherwise.
5772 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5774 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5777 int min_val = INT_MAX;
5778 int best_node = NUMA_NO_NODE;
5780 /* Use the local node if we haven't already */
5781 if (!node_isset(node, *used_node_mask)) {
5782 node_set(node, *used_node_mask);
5786 for_each_node_state(n, N_MEMORY) {
5788 /* Don't want a node to appear more than once */
5789 if (node_isset(n, *used_node_mask))
5792 /* Use the distance array to find the distance */
5793 val = node_distance(node, n);
5795 /* Penalize nodes under us ("prefer the next node") */
5798 /* Give preference to headless and unused nodes */
5799 if (!cpumask_empty(cpumask_of_node(n)))
5800 val += PENALTY_FOR_NODE_WITH_CPUS;
5802 /* Slight preference for less loaded node */
5803 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5804 val += node_load[n];
5806 if (val < min_val) {
5813 node_set(best_node, *used_node_mask);
5820 * Build zonelists ordered by node and zones within node.
5821 * This results in maximum locality--normal zone overflows into local
5822 * DMA zone, if any--but risks exhausting DMA zone.
5824 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5827 struct zoneref *zonerefs;
5830 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5832 for (i = 0; i < nr_nodes; i++) {
5835 pg_data_t *node = NODE_DATA(node_order[i]);
5837 nr_zones = build_zonerefs_node(node, zonerefs);
5838 zonerefs += nr_zones;
5840 zonerefs->zone = NULL;
5841 zonerefs->zone_idx = 0;
5845 * Build gfp_thisnode zonelists
5847 static void build_thisnode_zonelists(pg_data_t *pgdat)
5849 struct zoneref *zonerefs;
5852 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5853 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5854 zonerefs += nr_zones;
5855 zonerefs->zone = NULL;
5856 zonerefs->zone_idx = 0;
5860 * Build zonelists ordered by zone and nodes within zones.
5861 * This results in conserving DMA zone[s] until all Normal memory is
5862 * exhausted, but results in overflowing to remote node while memory
5863 * may still exist in local DMA zone.
5866 static void build_zonelists(pg_data_t *pgdat)
5868 static int node_order[MAX_NUMNODES];
5869 int node, load, nr_nodes = 0;
5870 nodemask_t used_mask = NODE_MASK_NONE;
5871 int local_node, prev_node;
5873 /* NUMA-aware ordering of nodes */
5874 local_node = pgdat->node_id;
5875 load = nr_online_nodes;
5876 prev_node = local_node;
5878 memset(node_order, 0, sizeof(node_order));
5879 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5881 * We don't want to pressure a particular node.
5882 * So adding penalty to the first node in same
5883 * distance group to make it round-robin.
5885 if (node_distance(local_node, node) !=
5886 node_distance(local_node, prev_node))
5887 node_load[node] = load;
5889 node_order[nr_nodes++] = node;
5894 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5895 build_thisnode_zonelists(pgdat);
5898 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5900 * Return node id of node used for "local" allocations.
5901 * I.e., first node id of first zone in arg node's generic zonelist.
5902 * Used for initializing percpu 'numa_mem', which is used primarily
5903 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5905 int local_memory_node(int node)
5909 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5910 gfp_zone(GFP_KERNEL),
5912 return zone_to_nid(z->zone);
5916 static void setup_min_unmapped_ratio(void);
5917 static void setup_min_slab_ratio(void);
5918 #else /* CONFIG_NUMA */
5920 static void build_zonelists(pg_data_t *pgdat)
5922 int node, local_node;
5923 struct zoneref *zonerefs;
5926 local_node = pgdat->node_id;
5928 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5929 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5930 zonerefs += nr_zones;
5933 * Now we build the zonelist so that it contains the zones
5934 * of all the other nodes.
5935 * We don't want to pressure a particular node, so when
5936 * building the zones for node N, we make sure that the
5937 * zones coming right after the local ones are those from
5938 * node N+1 (modulo N)
5940 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5941 if (!node_online(node))
5943 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5944 zonerefs += nr_zones;
5946 for (node = 0; node < local_node; node++) {
5947 if (!node_online(node))
5949 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5950 zonerefs += nr_zones;
5953 zonerefs->zone = NULL;
5954 zonerefs->zone_idx = 0;
5957 #endif /* CONFIG_NUMA */
5960 * Boot pageset table. One per cpu which is going to be used for all
5961 * zones and all nodes. The parameters will be set in such a way
5962 * that an item put on a list will immediately be handed over to
5963 * the buddy list. This is safe since pageset manipulation is done
5964 * with interrupts disabled.
5966 * The boot_pagesets must be kept even after bootup is complete for
5967 * unused processors and/or zones. They do play a role for bootstrapping
5968 * hotplugged processors.
5970 * zoneinfo_show() and maybe other functions do
5971 * not check if the processor is online before following the pageset pointer.
5972 * Other parts of the kernel may not check if the zone is available.
5974 static void pageset_init(struct per_cpu_pageset *p);
5975 /* These effectively disable the pcplists in the boot pageset completely */
5976 #define BOOT_PAGESET_HIGH 0
5977 #define BOOT_PAGESET_BATCH 1
5978 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5979 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5981 static void __build_all_zonelists(void *data)
5984 int __maybe_unused cpu;
5985 pg_data_t *self = data;
5986 static DEFINE_SPINLOCK(lock);
5991 memset(node_load, 0, sizeof(node_load));
5995 * This node is hotadded and no memory is yet present. So just
5996 * building zonelists is fine - no need to touch other nodes.
5998 if (self && !node_online(self->node_id)) {
5999 build_zonelists(self);
6001 for_each_online_node(nid) {
6002 pg_data_t *pgdat = NODE_DATA(nid);
6004 build_zonelists(pgdat);
6007 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6009 * We now know the "local memory node" for each node--
6010 * i.e., the node of the first zone in the generic zonelist.
6011 * Set up numa_mem percpu variable for on-line cpus. During
6012 * boot, only the boot cpu should be on-line; we'll init the
6013 * secondary cpus' numa_mem as they come on-line. During
6014 * node/memory hotplug, we'll fixup all on-line cpus.
6016 for_each_online_cpu(cpu)
6017 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6024 static noinline void __init
6025 build_all_zonelists_init(void)
6029 __build_all_zonelists(NULL);
6032 * Initialize the boot_pagesets that are going to be used
6033 * for bootstrapping processors. The real pagesets for
6034 * each zone will be allocated later when the per cpu
6035 * allocator is available.
6037 * boot_pagesets are used also for bootstrapping offline
6038 * cpus if the system is already booted because the pagesets
6039 * are needed to initialize allocators on a specific cpu too.
6040 * F.e. the percpu allocator needs the page allocator which
6041 * needs the percpu allocator in order to allocate its pagesets
6042 * (a chicken-egg dilemma).
6044 for_each_possible_cpu(cpu)
6045 pageset_init(&per_cpu(boot_pageset, cpu));
6047 mminit_verify_zonelist();
6048 cpuset_init_current_mems_allowed();
6052 * unless system_state == SYSTEM_BOOTING.
6054 * __ref due to call of __init annotated helper build_all_zonelists_init
6055 * [protected by SYSTEM_BOOTING].
6057 void __ref build_all_zonelists(pg_data_t *pgdat)
6059 unsigned long vm_total_pages;
6061 if (system_state == SYSTEM_BOOTING) {
6062 build_all_zonelists_init();
6064 __build_all_zonelists(pgdat);
6065 /* cpuset refresh routine should be here */
6067 /* Get the number of free pages beyond high watermark in all zones. */
6068 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6070 * Disable grouping by mobility if the number of pages in the
6071 * system is too low to allow the mechanism to work. It would be
6072 * more accurate, but expensive to check per-zone. This check is
6073 * made on memory-hotadd so a system can start with mobility
6074 * disabled and enable it later
6076 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6077 page_group_by_mobility_disabled = 1;
6079 page_group_by_mobility_disabled = 0;
6081 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6083 page_group_by_mobility_disabled ? "off" : "on",
6086 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6090 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6091 static bool __meminit
6092 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6094 static struct memblock_region *r;
6096 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6097 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6098 for_each_mem_region(r) {
6099 if (*pfn < memblock_region_memory_end_pfn(r))
6103 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6104 memblock_is_mirror(r)) {
6105 *pfn = memblock_region_memory_end_pfn(r);
6113 * Initially all pages are reserved - free ones are freed
6114 * up by memblock_free_all() once the early boot process is
6115 * done. Non-atomic initialization, single-pass.
6117 * All aligned pageblocks are initialized to the specified migratetype
6118 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6119 * zone stats (e.g., nr_isolate_pageblock) are touched.
6121 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
6122 unsigned long start_pfn, unsigned long zone_end_pfn,
6123 enum meminit_context context,
6124 struct vmem_altmap *altmap, int migratetype)
6126 unsigned long pfn, end_pfn = start_pfn + size;
6129 if (highest_memmap_pfn < end_pfn - 1)
6130 highest_memmap_pfn = end_pfn - 1;
6132 #ifdef CONFIG_ZONE_DEVICE
6134 * Honor reservation requested by the driver for this ZONE_DEVICE
6135 * memory. We limit the total number of pages to initialize to just
6136 * those that might contain the memory mapping. We will defer the
6137 * ZONE_DEVICE page initialization until after we have released
6140 if (zone == ZONE_DEVICE) {
6144 if (start_pfn == altmap->base_pfn)
6145 start_pfn += altmap->reserve;
6146 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6150 for (pfn = start_pfn; pfn < end_pfn; ) {
6152 * There can be holes in boot-time mem_map[]s handed to this
6153 * function. They do not exist on hotplugged memory.
6155 if (context == MEMINIT_EARLY) {
6156 if (overlap_memmap_init(zone, &pfn))
6158 if (defer_init(nid, pfn, zone_end_pfn))
6162 page = pfn_to_page(pfn);
6163 __init_single_page(page, pfn, zone, nid);
6164 if (context == MEMINIT_HOTPLUG)
6165 __SetPageReserved(page);
6168 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6169 * such that unmovable allocations won't be scattered all
6170 * over the place during system boot.
6172 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6173 set_pageblock_migratetype(page, migratetype);
6180 #ifdef CONFIG_ZONE_DEVICE
6181 void __ref memmap_init_zone_device(struct zone *zone,
6182 unsigned long start_pfn,
6183 unsigned long nr_pages,
6184 struct dev_pagemap *pgmap)
6186 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6187 struct pglist_data *pgdat = zone->zone_pgdat;
6188 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6189 unsigned long zone_idx = zone_idx(zone);
6190 unsigned long start = jiffies;
6191 int nid = pgdat->node_id;
6193 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6197 * The call to memmap_init_zone should have already taken care
6198 * of the pages reserved for the memmap, so we can just jump to
6199 * the end of that region and start processing the device pages.
6202 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6203 nr_pages = end_pfn - start_pfn;
6206 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6207 struct page *page = pfn_to_page(pfn);
6209 __init_single_page(page, pfn, zone_idx, nid);
6212 * Mark page reserved as it will need to wait for onlining
6213 * phase for it to be fully associated with a zone.
6215 * We can use the non-atomic __set_bit operation for setting
6216 * the flag as we are still initializing the pages.
6218 __SetPageReserved(page);
6221 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6222 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6223 * ever freed or placed on a driver-private list.
6225 page->pgmap = pgmap;
6226 page->zone_device_data = NULL;
6229 * Mark the block movable so that blocks are reserved for
6230 * movable at startup. This will force kernel allocations
6231 * to reserve their blocks rather than leaking throughout
6232 * the address space during boot when many long-lived
6233 * kernel allocations are made.
6235 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6236 * because this is done early in section_activate()
6238 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6239 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6244 pr_info("%s initialised %lu pages in %ums\n", __func__,
6245 nr_pages, jiffies_to_msecs(jiffies - start));
6249 static void __meminit zone_init_free_lists(struct zone *zone)
6251 unsigned int order, t;
6252 for_each_migratetype_order(order, t) {
6253 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6254 zone->free_area[order].nr_free = 0;
6258 void __meminit __weak memmap_init(unsigned long size, int nid,
6260 unsigned long range_start_pfn)
6262 unsigned long start_pfn, end_pfn;
6263 unsigned long range_end_pfn = range_start_pfn + size;
6266 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6267 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6268 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6270 if (end_pfn > start_pfn) {
6271 size = end_pfn - start_pfn;
6272 memmap_init_zone(size, nid, zone, start_pfn, range_end_pfn,
6273 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6278 static int zone_batchsize(struct zone *zone)
6284 * The per-cpu-pages pools are set to around 1000th of the
6287 batch = zone_managed_pages(zone) / 1024;
6288 /* But no more than a meg. */
6289 if (batch * PAGE_SIZE > 1024 * 1024)
6290 batch = (1024 * 1024) / PAGE_SIZE;
6291 batch /= 4; /* We effectively *= 4 below */
6296 * Clamp the batch to a 2^n - 1 value. Having a power
6297 * of 2 value was found to be more likely to have
6298 * suboptimal cache aliasing properties in some cases.
6300 * For example if 2 tasks are alternately allocating
6301 * batches of pages, one task can end up with a lot
6302 * of pages of one half of the possible page colors
6303 * and the other with pages of the other colors.
6305 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6310 /* The deferral and batching of frees should be suppressed under NOMMU
6313 * The problem is that NOMMU needs to be able to allocate large chunks
6314 * of contiguous memory as there's no hardware page translation to
6315 * assemble apparent contiguous memory from discontiguous pages.
6317 * Queueing large contiguous runs of pages for batching, however,
6318 * causes the pages to actually be freed in smaller chunks. As there
6319 * can be a significant delay between the individual batches being
6320 * recycled, this leads to the once large chunks of space being
6321 * fragmented and becoming unavailable for high-order allocations.
6328 * pcp->high and pcp->batch values are related and generally batch is lower
6329 * than high. They are also related to pcp->count such that count is lower
6330 * than high, and as soon as it reaches high, the pcplist is flushed.
6332 * However, guaranteeing these relations at all times would require e.g. write
6333 * barriers here but also careful usage of read barriers at the read side, and
6334 * thus be prone to error and bad for performance. Thus the update only prevents
6335 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6336 * can cope with those fields changing asynchronously, and fully trust only the
6337 * pcp->count field on the local CPU with interrupts disabled.
6339 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6340 * outside of boot time (or some other assurance that no concurrent updaters
6343 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6344 unsigned long batch)
6346 WRITE_ONCE(pcp->batch, batch);
6347 WRITE_ONCE(pcp->high, high);
6350 static void pageset_init(struct per_cpu_pageset *p)
6352 struct per_cpu_pages *pcp;
6355 memset(p, 0, sizeof(*p));
6358 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6359 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6362 * Set batch and high values safe for a boot pageset. A true percpu
6363 * pageset's initialization will update them subsequently. Here we don't
6364 * need to be as careful as pageset_update() as nobody can access the
6367 pcp->high = BOOT_PAGESET_HIGH;
6368 pcp->batch = BOOT_PAGESET_BATCH;
6371 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6372 unsigned long batch)
6374 struct per_cpu_pageset *p;
6377 for_each_possible_cpu(cpu) {
6378 p = per_cpu_ptr(zone->pageset, cpu);
6379 pageset_update(&p->pcp, high, batch);
6384 * Calculate and set new high and batch values for all per-cpu pagesets of a
6385 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6387 static void zone_set_pageset_high_and_batch(struct zone *zone)
6389 unsigned long new_high, new_batch;
6391 if (percpu_pagelist_fraction) {
6392 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6393 new_batch = max(1UL, new_high / 4);
6394 if ((new_high / 4) > (PAGE_SHIFT * 8))
6395 new_batch = PAGE_SHIFT * 8;
6397 new_batch = zone_batchsize(zone);
6398 new_high = 6 * new_batch;
6399 new_batch = max(1UL, 1 * new_batch);
6402 if (zone->pageset_high == new_high &&
6403 zone->pageset_batch == new_batch)
6406 zone->pageset_high = new_high;
6407 zone->pageset_batch = new_batch;
6409 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6412 void __meminit setup_zone_pageset(struct zone *zone)
6414 struct per_cpu_pageset *p;
6417 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6418 for_each_possible_cpu(cpu) {
6419 p = per_cpu_ptr(zone->pageset, cpu);
6423 zone_set_pageset_high_and_batch(zone);
6427 * Allocate per cpu pagesets and initialize them.
6428 * Before this call only boot pagesets were available.
6430 void __init setup_per_cpu_pageset(void)
6432 struct pglist_data *pgdat;
6434 int __maybe_unused cpu;
6436 for_each_populated_zone(zone)
6437 setup_zone_pageset(zone);
6441 * Unpopulated zones continue using the boot pagesets.
6442 * The numa stats for these pagesets need to be reset.
6443 * Otherwise, they will end up skewing the stats of
6444 * the nodes these zones are associated with.
6446 for_each_possible_cpu(cpu) {
6447 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6448 memset(pcp->vm_numa_stat_diff, 0,
6449 sizeof(pcp->vm_numa_stat_diff));
6453 for_each_online_pgdat(pgdat)
6454 pgdat->per_cpu_nodestats =
6455 alloc_percpu(struct per_cpu_nodestat);
6458 static __meminit void zone_pcp_init(struct zone *zone)
6461 * per cpu subsystem is not up at this point. The following code
6462 * relies on the ability of the linker to provide the
6463 * offset of a (static) per cpu variable into the per cpu area.
6465 zone->pageset = &boot_pageset;
6466 zone->pageset_high = BOOT_PAGESET_HIGH;
6467 zone->pageset_batch = BOOT_PAGESET_BATCH;
6469 if (populated_zone(zone))
6470 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6471 zone->name, zone->present_pages,
6472 zone_batchsize(zone));
6475 void __meminit init_currently_empty_zone(struct zone *zone,
6476 unsigned long zone_start_pfn,
6479 struct pglist_data *pgdat = zone->zone_pgdat;
6480 int zone_idx = zone_idx(zone) + 1;
6482 if (zone_idx > pgdat->nr_zones)
6483 pgdat->nr_zones = zone_idx;
6485 zone->zone_start_pfn = zone_start_pfn;
6487 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6488 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6490 (unsigned long)zone_idx(zone),
6491 zone_start_pfn, (zone_start_pfn + size));
6493 zone_init_free_lists(zone);
6494 zone->initialized = 1;
6498 * get_pfn_range_for_nid - Return the start and end page frames for a node
6499 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6500 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6501 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6503 * It returns the start and end page frame of a node based on information
6504 * provided by memblock_set_node(). If called for a node
6505 * with no available memory, a warning is printed and the start and end
6508 void __init get_pfn_range_for_nid(unsigned int nid,
6509 unsigned long *start_pfn, unsigned long *end_pfn)
6511 unsigned long this_start_pfn, this_end_pfn;
6517 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6518 *start_pfn = min(*start_pfn, this_start_pfn);
6519 *end_pfn = max(*end_pfn, this_end_pfn);
6522 if (*start_pfn == -1UL)
6527 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6528 * assumption is made that zones within a node are ordered in monotonic
6529 * increasing memory addresses so that the "highest" populated zone is used
6531 static void __init find_usable_zone_for_movable(void)
6534 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6535 if (zone_index == ZONE_MOVABLE)
6538 if (arch_zone_highest_possible_pfn[zone_index] >
6539 arch_zone_lowest_possible_pfn[zone_index])
6543 VM_BUG_ON(zone_index == -1);
6544 movable_zone = zone_index;
6548 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6549 * because it is sized independent of architecture. Unlike the other zones,
6550 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6551 * in each node depending on the size of each node and how evenly kernelcore
6552 * is distributed. This helper function adjusts the zone ranges
6553 * provided by the architecture for a given node by using the end of the
6554 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6555 * zones within a node are in order of monotonic increases memory addresses
6557 static void __init adjust_zone_range_for_zone_movable(int nid,
6558 unsigned long zone_type,
6559 unsigned long node_start_pfn,
6560 unsigned long node_end_pfn,
6561 unsigned long *zone_start_pfn,
6562 unsigned long *zone_end_pfn)
6564 /* Only adjust if ZONE_MOVABLE is on this node */
6565 if (zone_movable_pfn[nid]) {
6566 /* Size ZONE_MOVABLE */
6567 if (zone_type == ZONE_MOVABLE) {
6568 *zone_start_pfn = zone_movable_pfn[nid];
6569 *zone_end_pfn = min(node_end_pfn,
6570 arch_zone_highest_possible_pfn[movable_zone]);
6572 /* Adjust for ZONE_MOVABLE starting within this range */
6573 } else if (!mirrored_kernelcore &&
6574 *zone_start_pfn < zone_movable_pfn[nid] &&
6575 *zone_end_pfn > zone_movable_pfn[nid]) {
6576 *zone_end_pfn = zone_movable_pfn[nid];
6578 /* Check if this whole range is within ZONE_MOVABLE */
6579 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6580 *zone_start_pfn = *zone_end_pfn;
6585 * Return the number of pages a zone spans in a node, including holes
6586 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6588 static unsigned long __init zone_spanned_pages_in_node(int nid,
6589 unsigned long zone_type,
6590 unsigned long node_start_pfn,
6591 unsigned long node_end_pfn,
6592 unsigned long *zone_start_pfn,
6593 unsigned long *zone_end_pfn)
6595 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6596 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6597 /* When hotadd a new node from cpu_up(), the node should be empty */
6598 if (!node_start_pfn && !node_end_pfn)
6601 /* Get the start and end of the zone */
6602 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6603 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6604 adjust_zone_range_for_zone_movable(nid, zone_type,
6605 node_start_pfn, node_end_pfn,
6606 zone_start_pfn, zone_end_pfn);
6608 /* Check that this node has pages within the zone's required range */
6609 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6612 /* Move the zone boundaries inside the node if necessary */
6613 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6614 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6616 /* Return the spanned pages */
6617 return *zone_end_pfn - *zone_start_pfn;
6621 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6622 * then all holes in the requested range will be accounted for.
6624 unsigned long __init __absent_pages_in_range(int nid,
6625 unsigned long range_start_pfn,
6626 unsigned long range_end_pfn)
6628 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6629 unsigned long start_pfn, end_pfn;
6632 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6633 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6634 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6635 nr_absent -= end_pfn - start_pfn;
6641 * absent_pages_in_range - Return number of page frames in holes within a range
6642 * @start_pfn: The start PFN to start searching for holes
6643 * @end_pfn: The end PFN to stop searching for holes
6645 * Return: the number of pages frames in memory holes within a range.
6647 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6648 unsigned long end_pfn)
6650 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6653 /* Return the number of page frames in holes in a zone on a node */
6654 static unsigned long __init zone_absent_pages_in_node(int nid,
6655 unsigned long zone_type,
6656 unsigned long node_start_pfn,
6657 unsigned long node_end_pfn)
6659 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6660 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6661 unsigned long zone_start_pfn, zone_end_pfn;
6662 unsigned long nr_absent;
6664 /* When hotadd a new node from cpu_up(), the node should be empty */
6665 if (!node_start_pfn && !node_end_pfn)
6668 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6669 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6671 adjust_zone_range_for_zone_movable(nid, zone_type,
6672 node_start_pfn, node_end_pfn,
6673 &zone_start_pfn, &zone_end_pfn);
6674 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6677 * ZONE_MOVABLE handling.
6678 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6681 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6682 unsigned long start_pfn, end_pfn;
6683 struct memblock_region *r;
6685 for_each_mem_region(r) {
6686 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6687 zone_start_pfn, zone_end_pfn);
6688 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6689 zone_start_pfn, zone_end_pfn);
6691 if (zone_type == ZONE_MOVABLE &&
6692 memblock_is_mirror(r))
6693 nr_absent += end_pfn - start_pfn;
6695 if (zone_type == ZONE_NORMAL &&
6696 !memblock_is_mirror(r))
6697 nr_absent += end_pfn - start_pfn;
6704 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6705 unsigned long node_start_pfn,
6706 unsigned long node_end_pfn)
6708 unsigned long realtotalpages = 0, totalpages = 0;
6711 for (i = 0; i < MAX_NR_ZONES; i++) {
6712 struct zone *zone = pgdat->node_zones + i;
6713 unsigned long zone_start_pfn, zone_end_pfn;
6714 unsigned long spanned, absent;
6715 unsigned long size, real_size;
6717 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6722 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6727 real_size = size - absent;
6730 zone->zone_start_pfn = zone_start_pfn;
6732 zone->zone_start_pfn = 0;
6733 zone->spanned_pages = size;
6734 zone->present_pages = real_size;
6737 realtotalpages += real_size;
6740 pgdat->node_spanned_pages = totalpages;
6741 pgdat->node_present_pages = realtotalpages;
6742 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6746 #ifndef CONFIG_SPARSEMEM
6748 * Calculate the size of the zone->blockflags rounded to an unsigned long
6749 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6750 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6751 * round what is now in bits to nearest long in bits, then return it in
6754 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6756 unsigned long usemapsize;
6758 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6759 usemapsize = roundup(zonesize, pageblock_nr_pages);
6760 usemapsize = usemapsize >> pageblock_order;
6761 usemapsize *= NR_PAGEBLOCK_BITS;
6762 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6764 return usemapsize / 8;
6767 static void __ref setup_usemap(struct pglist_data *pgdat,
6769 unsigned long zone_start_pfn,
6770 unsigned long zonesize)
6772 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6773 zone->pageblock_flags = NULL;
6775 zone->pageblock_flags =
6776 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6778 if (!zone->pageblock_flags)
6779 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6780 usemapsize, zone->name, pgdat->node_id);
6784 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6785 unsigned long zone_start_pfn, unsigned long zonesize) {}
6786 #endif /* CONFIG_SPARSEMEM */
6788 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6790 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6791 void __init set_pageblock_order(void)
6795 /* Check that pageblock_nr_pages has not already been setup */
6796 if (pageblock_order)
6799 if (HPAGE_SHIFT > PAGE_SHIFT)
6800 order = HUGETLB_PAGE_ORDER;
6802 order = MAX_ORDER - 1;
6805 * Assume the largest contiguous order of interest is a huge page.
6806 * This value may be variable depending on boot parameters on IA64 and
6809 pageblock_order = order;
6811 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6814 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6815 * is unused as pageblock_order is set at compile-time. See
6816 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6819 void __init set_pageblock_order(void)
6823 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6825 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6826 unsigned long present_pages)
6828 unsigned long pages = spanned_pages;
6831 * Provide a more accurate estimation if there are holes within
6832 * the zone and SPARSEMEM is in use. If there are holes within the
6833 * zone, each populated memory region may cost us one or two extra
6834 * memmap pages due to alignment because memmap pages for each
6835 * populated regions may not be naturally aligned on page boundary.
6836 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6838 if (spanned_pages > present_pages + (present_pages >> 4) &&
6839 IS_ENABLED(CONFIG_SPARSEMEM))
6840 pages = present_pages;
6842 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6845 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6846 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6848 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6850 spin_lock_init(&ds_queue->split_queue_lock);
6851 INIT_LIST_HEAD(&ds_queue->split_queue);
6852 ds_queue->split_queue_len = 0;
6855 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6858 #ifdef CONFIG_COMPACTION
6859 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6861 init_waitqueue_head(&pgdat->kcompactd_wait);
6864 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6867 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6869 pgdat_resize_init(pgdat);
6871 pgdat_init_split_queue(pgdat);
6872 pgdat_init_kcompactd(pgdat);
6874 init_waitqueue_head(&pgdat->kswapd_wait);
6875 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6877 pgdat_page_ext_init(pgdat);
6878 lruvec_init(&pgdat->__lruvec);
6881 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6882 unsigned long remaining_pages)
6884 atomic_long_set(&zone->managed_pages, remaining_pages);
6885 zone_set_nid(zone, nid);
6886 zone->name = zone_names[idx];
6887 zone->zone_pgdat = NODE_DATA(nid);
6888 spin_lock_init(&zone->lock);
6889 zone_seqlock_init(zone);
6890 zone_pcp_init(zone);
6894 * Set up the zone data structures
6895 * - init pgdat internals
6896 * - init all zones belonging to this node
6898 * NOTE: this function is only called during memory hotplug
6900 #ifdef CONFIG_MEMORY_HOTPLUG
6901 void __ref free_area_init_core_hotplug(int nid)
6904 pg_data_t *pgdat = NODE_DATA(nid);
6906 pgdat_init_internals(pgdat);
6907 for (z = 0; z < MAX_NR_ZONES; z++)
6908 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6913 * Set up the zone data structures:
6914 * - mark all pages reserved
6915 * - mark all memory queues empty
6916 * - clear the memory bitmaps
6918 * NOTE: pgdat should get zeroed by caller.
6919 * NOTE: this function is only called during early init.
6921 static void __init free_area_init_core(struct pglist_data *pgdat)
6924 int nid = pgdat->node_id;
6926 pgdat_init_internals(pgdat);
6927 pgdat->per_cpu_nodestats = &boot_nodestats;
6929 for (j = 0; j < MAX_NR_ZONES; j++) {
6930 struct zone *zone = pgdat->node_zones + j;
6931 unsigned long size, freesize, memmap_pages;
6932 unsigned long zone_start_pfn = zone->zone_start_pfn;
6934 size = zone->spanned_pages;
6935 freesize = zone->present_pages;
6938 * Adjust freesize so that it accounts for how much memory
6939 * is used by this zone for memmap. This affects the watermark
6940 * and per-cpu initialisations
6942 memmap_pages = calc_memmap_size(size, freesize);
6943 if (!is_highmem_idx(j)) {
6944 if (freesize >= memmap_pages) {
6945 freesize -= memmap_pages;
6948 " %s zone: %lu pages used for memmap\n",
6949 zone_names[j], memmap_pages);
6951 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6952 zone_names[j], memmap_pages, freesize);
6955 /* Account for reserved pages */
6956 if (j == 0 && freesize > dma_reserve) {
6957 freesize -= dma_reserve;
6958 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6959 zone_names[0], dma_reserve);
6962 if (!is_highmem_idx(j))
6963 nr_kernel_pages += freesize;
6964 /* Charge for highmem memmap if there are enough kernel pages */
6965 else if (nr_kernel_pages > memmap_pages * 2)
6966 nr_kernel_pages -= memmap_pages;
6967 nr_all_pages += freesize;
6970 * Set an approximate value for lowmem here, it will be adjusted
6971 * when the bootmem allocator frees pages into the buddy system.
6972 * And all highmem pages will be managed by the buddy system.
6974 zone_init_internals(zone, j, nid, freesize);
6979 set_pageblock_order();
6980 setup_usemap(pgdat, zone, zone_start_pfn, size);
6981 init_currently_empty_zone(zone, zone_start_pfn, size);
6982 memmap_init(size, nid, j, zone_start_pfn);
6986 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6987 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6989 unsigned long __maybe_unused start = 0;
6990 unsigned long __maybe_unused offset = 0;
6992 /* Skip empty nodes */
6993 if (!pgdat->node_spanned_pages)
6996 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6997 offset = pgdat->node_start_pfn - start;
6998 /* ia64 gets its own node_mem_map, before this, without bootmem */
6999 if (!pgdat->node_mem_map) {
7000 unsigned long size, end;
7004 * The zone's endpoints aren't required to be MAX_ORDER
7005 * aligned but the node_mem_map endpoints must be in order
7006 * for the buddy allocator to function correctly.
7008 end = pgdat_end_pfn(pgdat);
7009 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7010 size = (end - start) * sizeof(struct page);
7011 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7014 panic("Failed to allocate %ld bytes for node %d memory map\n",
7015 size, pgdat->node_id);
7016 pgdat->node_mem_map = map + offset;
7018 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7019 __func__, pgdat->node_id, (unsigned long)pgdat,
7020 (unsigned long)pgdat->node_mem_map);
7021 #ifndef CONFIG_NEED_MULTIPLE_NODES
7023 * With no DISCONTIG, the global mem_map is just set as node 0's
7025 if (pgdat == NODE_DATA(0)) {
7026 mem_map = NODE_DATA(0)->node_mem_map;
7027 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7033 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7034 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7036 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7037 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7039 pgdat->first_deferred_pfn = ULONG_MAX;
7042 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7045 static void __init free_area_init_node(int nid)
7047 pg_data_t *pgdat = NODE_DATA(nid);
7048 unsigned long start_pfn = 0;
7049 unsigned long end_pfn = 0;
7051 /* pg_data_t should be reset to zero when it's allocated */
7052 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7054 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7056 pgdat->node_id = nid;
7057 pgdat->node_start_pfn = start_pfn;
7058 pgdat->per_cpu_nodestats = NULL;
7060 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7061 (u64)start_pfn << PAGE_SHIFT,
7062 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7063 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7065 alloc_node_mem_map(pgdat);
7066 pgdat_set_deferred_range(pgdat);
7068 free_area_init_core(pgdat);
7071 void __init free_area_init_memoryless_node(int nid)
7073 free_area_init_node(nid);
7076 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
7078 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
7079 * PageReserved(). Return the number of struct pages that were initialized.
7081 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
7086 for (pfn = spfn; pfn < epfn; pfn++) {
7087 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
7088 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
7089 + pageblock_nr_pages - 1;
7093 * Use a fake node/zone (0) for now. Some of these pages
7094 * (in memblock.reserved but not in memblock.memory) will
7095 * get re-initialized via reserve_bootmem_region() later.
7097 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7098 __SetPageReserved(pfn_to_page(pfn));
7106 * Only struct pages that are backed by physical memory are zeroed and
7107 * initialized by going through __init_single_page(). But, there are some
7108 * struct pages which are reserved in memblock allocator and their fields
7109 * may be accessed (for example page_to_pfn() on some configuration accesses
7110 * flags). We must explicitly initialize those struct pages.
7112 * This function also addresses a similar issue where struct pages are left
7113 * uninitialized because the physical address range is not covered by
7114 * memblock.memory or memblock.reserved. That could happen when memblock
7115 * layout is manually configured via memmap=, or when the highest physical
7116 * address (max_pfn) does not end on a section boundary.
7118 static void __init init_unavailable_mem(void)
7120 phys_addr_t start, end;
7122 phys_addr_t next = 0;
7125 * Loop through unavailable ranges not covered by memblock.memory.
7128 for_each_mem_range(i, &start, &end) {
7130 pgcnt += init_unavailable_range(PFN_DOWN(next),
7136 * Early sections always have a fully populated memmap for the whole
7137 * section - see pfn_valid(). If the last section has holes at the
7138 * end and that section is marked "online", the memmap will be
7139 * considered initialized. Make sure that memmap has a well defined
7142 pgcnt += init_unavailable_range(PFN_DOWN(next),
7143 round_up(max_pfn, PAGES_PER_SECTION));
7146 * Struct pages that do not have backing memory. This could be because
7147 * firmware is using some of this memory, or for some other reasons.
7150 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7153 static inline void __init init_unavailable_mem(void)
7156 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7158 #if MAX_NUMNODES > 1
7160 * Figure out the number of possible node ids.
7162 void __init setup_nr_node_ids(void)
7164 unsigned int highest;
7166 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7167 nr_node_ids = highest + 1;
7172 * node_map_pfn_alignment - determine the maximum internode alignment
7174 * This function should be called after node map is populated and sorted.
7175 * It calculates the maximum power of two alignment which can distinguish
7178 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7179 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7180 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7181 * shifted, 1GiB is enough and this function will indicate so.
7183 * This is used to test whether pfn -> nid mapping of the chosen memory
7184 * model has fine enough granularity to avoid incorrect mapping for the
7185 * populated node map.
7187 * Return: the determined alignment in pfn's. 0 if there is no alignment
7188 * requirement (single node).
7190 unsigned long __init node_map_pfn_alignment(void)
7192 unsigned long accl_mask = 0, last_end = 0;
7193 unsigned long start, end, mask;
7194 int last_nid = NUMA_NO_NODE;
7197 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7198 if (!start || last_nid < 0 || last_nid == nid) {
7205 * Start with a mask granular enough to pin-point to the
7206 * start pfn and tick off bits one-by-one until it becomes
7207 * too coarse to separate the current node from the last.
7209 mask = ~((1 << __ffs(start)) - 1);
7210 while (mask && last_end <= (start & (mask << 1)))
7213 /* accumulate all internode masks */
7217 /* convert mask to number of pages */
7218 return ~accl_mask + 1;
7222 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7224 * Return: the minimum PFN based on information provided via
7225 * memblock_set_node().
7227 unsigned long __init find_min_pfn_with_active_regions(void)
7229 return PHYS_PFN(memblock_start_of_DRAM());
7233 * early_calculate_totalpages()
7234 * Sum pages in active regions for movable zone.
7235 * Populate N_MEMORY for calculating usable_nodes.
7237 static unsigned long __init early_calculate_totalpages(void)
7239 unsigned long totalpages = 0;
7240 unsigned long start_pfn, end_pfn;
7243 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7244 unsigned long pages = end_pfn - start_pfn;
7246 totalpages += pages;
7248 node_set_state(nid, N_MEMORY);
7254 * Find the PFN the Movable zone begins in each node. Kernel memory
7255 * is spread evenly between nodes as long as the nodes have enough
7256 * memory. When they don't, some nodes will have more kernelcore than
7259 static void __init find_zone_movable_pfns_for_nodes(void)
7262 unsigned long usable_startpfn;
7263 unsigned long kernelcore_node, kernelcore_remaining;
7264 /* save the state before borrow the nodemask */
7265 nodemask_t saved_node_state = node_states[N_MEMORY];
7266 unsigned long totalpages = early_calculate_totalpages();
7267 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7268 struct memblock_region *r;
7270 /* Need to find movable_zone earlier when movable_node is specified. */
7271 find_usable_zone_for_movable();
7274 * If movable_node is specified, ignore kernelcore and movablecore
7277 if (movable_node_is_enabled()) {
7278 for_each_mem_region(r) {
7279 if (!memblock_is_hotpluggable(r))
7282 nid = memblock_get_region_node(r);
7284 usable_startpfn = PFN_DOWN(r->base);
7285 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7286 min(usable_startpfn, zone_movable_pfn[nid]) :
7294 * If kernelcore=mirror is specified, ignore movablecore option
7296 if (mirrored_kernelcore) {
7297 bool mem_below_4gb_not_mirrored = false;
7299 for_each_mem_region(r) {
7300 if (memblock_is_mirror(r))
7303 nid = memblock_get_region_node(r);
7305 usable_startpfn = memblock_region_memory_base_pfn(r);
7307 if (usable_startpfn < 0x100000) {
7308 mem_below_4gb_not_mirrored = true;
7312 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7313 min(usable_startpfn, zone_movable_pfn[nid]) :
7317 if (mem_below_4gb_not_mirrored)
7318 pr_warn("This configuration results in unmirrored kernel memory.\n");
7324 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7325 * amount of necessary memory.
7327 if (required_kernelcore_percent)
7328 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7330 if (required_movablecore_percent)
7331 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7335 * If movablecore= was specified, calculate what size of
7336 * kernelcore that corresponds so that memory usable for
7337 * any allocation type is evenly spread. If both kernelcore
7338 * and movablecore are specified, then the value of kernelcore
7339 * will be used for required_kernelcore if it's greater than
7340 * what movablecore would have allowed.
7342 if (required_movablecore) {
7343 unsigned long corepages;
7346 * Round-up so that ZONE_MOVABLE is at least as large as what
7347 * was requested by the user
7349 required_movablecore =
7350 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7351 required_movablecore = min(totalpages, required_movablecore);
7352 corepages = totalpages - required_movablecore;
7354 required_kernelcore = max(required_kernelcore, corepages);
7358 * If kernelcore was not specified or kernelcore size is larger
7359 * than totalpages, there is no ZONE_MOVABLE.
7361 if (!required_kernelcore || required_kernelcore >= totalpages)
7364 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7365 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7368 /* Spread kernelcore memory as evenly as possible throughout nodes */
7369 kernelcore_node = required_kernelcore / usable_nodes;
7370 for_each_node_state(nid, N_MEMORY) {
7371 unsigned long start_pfn, end_pfn;
7374 * Recalculate kernelcore_node if the division per node
7375 * now exceeds what is necessary to satisfy the requested
7376 * amount of memory for the kernel
7378 if (required_kernelcore < kernelcore_node)
7379 kernelcore_node = required_kernelcore / usable_nodes;
7382 * As the map is walked, we track how much memory is usable
7383 * by the kernel using kernelcore_remaining. When it is
7384 * 0, the rest of the node is usable by ZONE_MOVABLE
7386 kernelcore_remaining = kernelcore_node;
7388 /* Go through each range of PFNs within this node */
7389 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7390 unsigned long size_pages;
7392 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7393 if (start_pfn >= end_pfn)
7396 /* Account for what is only usable for kernelcore */
7397 if (start_pfn < usable_startpfn) {
7398 unsigned long kernel_pages;
7399 kernel_pages = min(end_pfn, usable_startpfn)
7402 kernelcore_remaining -= min(kernel_pages,
7403 kernelcore_remaining);
7404 required_kernelcore -= min(kernel_pages,
7405 required_kernelcore);
7407 /* Continue if range is now fully accounted */
7408 if (end_pfn <= usable_startpfn) {
7411 * Push zone_movable_pfn to the end so
7412 * that if we have to rebalance
7413 * kernelcore across nodes, we will
7414 * not double account here
7416 zone_movable_pfn[nid] = end_pfn;
7419 start_pfn = usable_startpfn;
7423 * The usable PFN range for ZONE_MOVABLE is from
7424 * start_pfn->end_pfn. Calculate size_pages as the
7425 * number of pages used as kernelcore
7427 size_pages = end_pfn - start_pfn;
7428 if (size_pages > kernelcore_remaining)
7429 size_pages = kernelcore_remaining;
7430 zone_movable_pfn[nid] = start_pfn + size_pages;
7433 * Some kernelcore has been met, update counts and
7434 * break if the kernelcore for this node has been
7437 required_kernelcore -= min(required_kernelcore,
7439 kernelcore_remaining -= size_pages;
7440 if (!kernelcore_remaining)
7446 * If there is still required_kernelcore, we do another pass with one
7447 * less node in the count. This will push zone_movable_pfn[nid] further
7448 * along on the nodes that still have memory until kernelcore is
7452 if (usable_nodes && required_kernelcore > usable_nodes)
7456 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7457 for (nid = 0; nid < MAX_NUMNODES; nid++)
7458 zone_movable_pfn[nid] =
7459 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7462 /* restore the node_state */
7463 node_states[N_MEMORY] = saved_node_state;
7466 /* Any regular or high memory on that node ? */
7467 static void check_for_memory(pg_data_t *pgdat, int nid)
7469 enum zone_type zone_type;
7471 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7472 struct zone *zone = &pgdat->node_zones[zone_type];
7473 if (populated_zone(zone)) {
7474 if (IS_ENABLED(CONFIG_HIGHMEM))
7475 node_set_state(nid, N_HIGH_MEMORY);
7476 if (zone_type <= ZONE_NORMAL)
7477 node_set_state(nid, N_NORMAL_MEMORY);
7484 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7485 * such cases we allow max_zone_pfn sorted in the descending order
7487 bool __weak arch_has_descending_max_zone_pfns(void)
7493 * free_area_init - Initialise all pg_data_t and zone data
7494 * @max_zone_pfn: an array of max PFNs for each zone
7496 * This will call free_area_init_node() for each active node in the system.
7497 * Using the page ranges provided by memblock_set_node(), the size of each
7498 * zone in each node and their holes is calculated. If the maximum PFN
7499 * between two adjacent zones match, it is assumed that the zone is empty.
7500 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7501 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7502 * starts where the previous one ended. For example, ZONE_DMA32 starts
7503 * at arch_max_dma_pfn.
7505 void __init free_area_init(unsigned long *max_zone_pfn)
7507 unsigned long start_pfn, end_pfn;
7511 /* Record where the zone boundaries are */
7512 memset(arch_zone_lowest_possible_pfn, 0,
7513 sizeof(arch_zone_lowest_possible_pfn));
7514 memset(arch_zone_highest_possible_pfn, 0,
7515 sizeof(arch_zone_highest_possible_pfn));
7517 start_pfn = find_min_pfn_with_active_regions();
7518 descending = arch_has_descending_max_zone_pfns();
7520 for (i = 0; i < MAX_NR_ZONES; i++) {
7522 zone = MAX_NR_ZONES - i - 1;
7526 if (zone == ZONE_MOVABLE)
7529 end_pfn = max(max_zone_pfn[zone], start_pfn);
7530 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7531 arch_zone_highest_possible_pfn[zone] = end_pfn;
7533 start_pfn = end_pfn;
7536 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7537 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7538 find_zone_movable_pfns_for_nodes();
7540 /* Print out the zone ranges */
7541 pr_info("Zone ranges:\n");
7542 for (i = 0; i < MAX_NR_ZONES; i++) {
7543 if (i == ZONE_MOVABLE)
7545 pr_info(" %-8s ", zone_names[i]);
7546 if (arch_zone_lowest_possible_pfn[i] ==
7547 arch_zone_highest_possible_pfn[i])
7550 pr_cont("[mem %#018Lx-%#018Lx]\n",
7551 (u64)arch_zone_lowest_possible_pfn[i]
7553 ((u64)arch_zone_highest_possible_pfn[i]
7554 << PAGE_SHIFT) - 1);
7557 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7558 pr_info("Movable zone start for each node\n");
7559 for (i = 0; i < MAX_NUMNODES; i++) {
7560 if (zone_movable_pfn[i])
7561 pr_info(" Node %d: %#018Lx\n", i,
7562 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7566 * Print out the early node map, and initialize the
7567 * subsection-map relative to active online memory ranges to
7568 * enable future "sub-section" extensions of the memory map.
7570 pr_info("Early memory node ranges\n");
7571 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7572 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7573 (u64)start_pfn << PAGE_SHIFT,
7574 ((u64)end_pfn << PAGE_SHIFT) - 1);
7575 subsection_map_init(start_pfn, end_pfn - start_pfn);
7578 /* Initialise every node */
7579 mminit_verify_pageflags_layout();
7580 setup_nr_node_ids();
7581 init_unavailable_mem();
7582 for_each_online_node(nid) {
7583 pg_data_t *pgdat = NODE_DATA(nid);
7584 free_area_init_node(nid);
7586 /* Any memory on that node */
7587 if (pgdat->node_present_pages)
7588 node_set_state(nid, N_MEMORY);
7589 check_for_memory(pgdat, nid);
7593 static int __init cmdline_parse_core(char *p, unsigned long *core,
7594 unsigned long *percent)
7596 unsigned long long coremem;
7602 /* Value may be a percentage of total memory, otherwise bytes */
7603 coremem = simple_strtoull(p, &endptr, 0);
7604 if (*endptr == '%') {
7605 /* Paranoid check for percent values greater than 100 */
7606 WARN_ON(coremem > 100);
7610 coremem = memparse(p, &p);
7611 /* Paranoid check that UL is enough for the coremem value */
7612 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7614 *core = coremem >> PAGE_SHIFT;
7621 * kernelcore=size sets the amount of memory for use for allocations that
7622 * cannot be reclaimed or migrated.
7624 static int __init cmdline_parse_kernelcore(char *p)
7626 /* parse kernelcore=mirror */
7627 if (parse_option_str(p, "mirror")) {
7628 mirrored_kernelcore = true;
7632 return cmdline_parse_core(p, &required_kernelcore,
7633 &required_kernelcore_percent);
7637 * movablecore=size sets the amount of memory for use for allocations that
7638 * can be reclaimed or migrated.
7640 static int __init cmdline_parse_movablecore(char *p)
7642 return cmdline_parse_core(p, &required_movablecore,
7643 &required_movablecore_percent);
7646 early_param("kernelcore", cmdline_parse_kernelcore);
7647 early_param("movablecore", cmdline_parse_movablecore);
7649 void adjust_managed_page_count(struct page *page, long count)
7651 atomic_long_add(count, &page_zone(page)->managed_pages);
7652 totalram_pages_add(count);
7653 #ifdef CONFIG_HIGHMEM
7654 if (PageHighMem(page))
7655 totalhigh_pages_add(count);
7658 EXPORT_SYMBOL(adjust_managed_page_count);
7660 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7663 unsigned long pages = 0;
7665 start = (void *)PAGE_ALIGN((unsigned long)start);
7666 end = (void *)((unsigned long)end & PAGE_MASK);
7667 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7668 struct page *page = virt_to_page(pos);
7669 void *direct_map_addr;
7672 * 'direct_map_addr' might be different from 'pos'
7673 * because some architectures' virt_to_page()
7674 * work with aliases. Getting the direct map
7675 * address ensures that we get a _writeable_
7676 * alias for the memset().
7678 direct_map_addr = page_address(page);
7680 * Perform a kasan-unchecked memset() since this memory
7681 * has not been initialized.
7683 direct_map_addr = kasan_reset_tag(direct_map_addr);
7684 if ((unsigned int)poison <= 0xFF)
7685 memset(direct_map_addr, poison, PAGE_SIZE);
7687 free_reserved_page(page);
7691 pr_info("Freeing %s memory: %ldK\n",
7692 s, pages << (PAGE_SHIFT - 10));
7697 #ifdef CONFIG_HIGHMEM
7698 void free_highmem_page(struct page *page)
7700 __free_reserved_page(page);
7701 totalram_pages_inc();
7702 atomic_long_inc(&page_zone(page)->managed_pages);
7703 totalhigh_pages_inc();
7708 void __init mem_init_print_info(const char *str)
7710 unsigned long physpages, codesize, datasize, rosize, bss_size;
7711 unsigned long init_code_size, init_data_size;
7713 physpages = get_num_physpages();
7714 codesize = _etext - _stext;
7715 datasize = _edata - _sdata;
7716 rosize = __end_rodata - __start_rodata;
7717 bss_size = __bss_stop - __bss_start;
7718 init_data_size = __init_end - __init_begin;
7719 init_code_size = _einittext - _sinittext;
7722 * Detect special cases and adjust section sizes accordingly:
7723 * 1) .init.* may be embedded into .data sections
7724 * 2) .init.text.* may be out of [__init_begin, __init_end],
7725 * please refer to arch/tile/kernel/vmlinux.lds.S.
7726 * 3) .rodata.* may be embedded into .text or .data sections.
7728 #define adj_init_size(start, end, size, pos, adj) \
7730 if (start <= pos && pos < end && size > adj) \
7734 adj_init_size(__init_begin, __init_end, init_data_size,
7735 _sinittext, init_code_size);
7736 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7737 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7738 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7739 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7741 #undef adj_init_size
7743 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7744 #ifdef CONFIG_HIGHMEM
7748 nr_free_pages() << (PAGE_SHIFT - 10),
7749 physpages << (PAGE_SHIFT - 10),
7750 codesize >> 10, datasize >> 10, rosize >> 10,
7751 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7752 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7753 totalcma_pages << (PAGE_SHIFT - 10),
7754 #ifdef CONFIG_HIGHMEM
7755 totalhigh_pages() << (PAGE_SHIFT - 10),
7757 str ? ", " : "", str ? str : "");
7761 * set_dma_reserve - set the specified number of pages reserved in the first zone
7762 * @new_dma_reserve: The number of pages to mark reserved
7764 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7765 * In the DMA zone, a significant percentage may be consumed by kernel image
7766 * and other unfreeable allocations which can skew the watermarks badly. This
7767 * function may optionally be used to account for unfreeable pages in the
7768 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7769 * smaller per-cpu batchsize.
7771 void __init set_dma_reserve(unsigned long new_dma_reserve)
7773 dma_reserve = new_dma_reserve;
7776 static int page_alloc_cpu_dead(unsigned int cpu)
7779 lru_add_drain_cpu(cpu);
7783 * Spill the event counters of the dead processor
7784 * into the current processors event counters.
7785 * This artificially elevates the count of the current
7788 vm_events_fold_cpu(cpu);
7791 * Zero the differential counters of the dead processor
7792 * so that the vm statistics are consistent.
7794 * This is only okay since the processor is dead and cannot
7795 * race with what we are doing.
7797 cpu_vm_stats_fold(cpu);
7802 int hashdist = HASHDIST_DEFAULT;
7804 static int __init set_hashdist(char *str)
7808 hashdist = simple_strtoul(str, &str, 0);
7811 __setup("hashdist=", set_hashdist);
7814 void __init page_alloc_init(void)
7819 if (num_node_state(N_MEMORY) == 1)
7823 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7824 "mm/page_alloc:dead", NULL,
7825 page_alloc_cpu_dead);
7830 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7831 * or min_free_kbytes changes.
7833 static void calculate_totalreserve_pages(void)
7835 struct pglist_data *pgdat;
7836 unsigned long reserve_pages = 0;
7837 enum zone_type i, j;
7839 for_each_online_pgdat(pgdat) {
7841 pgdat->totalreserve_pages = 0;
7843 for (i = 0; i < MAX_NR_ZONES; i++) {
7844 struct zone *zone = pgdat->node_zones + i;
7846 unsigned long managed_pages = zone_managed_pages(zone);
7848 /* Find valid and maximum lowmem_reserve in the zone */
7849 for (j = i; j < MAX_NR_ZONES; j++) {
7850 if (zone->lowmem_reserve[j] > max)
7851 max = zone->lowmem_reserve[j];
7854 /* we treat the high watermark as reserved pages. */
7855 max += high_wmark_pages(zone);
7857 if (max > managed_pages)
7858 max = managed_pages;
7860 pgdat->totalreserve_pages += max;
7862 reserve_pages += max;
7865 totalreserve_pages = reserve_pages;
7869 * setup_per_zone_lowmem_reserve - called whenever
7870 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7871 * has a correct pages reserved value, so an adequate number of
7872 * pages are left in the zone after a successful __alloc_pages().
7874 static void setup_per_zone_lowmem_reserve(void)
7876 struct pglist_data *pgdat;
7877 enum zone_type i, j;
7879 for_each_online_pgdat(pgdat) {
7880 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
7881 struct zone *zone = &pgdat->node_zones[i];
7882 int ratio = sysctl_lowmem_reserve_ratio[i];
7883 bool clear = !ratio || !zone_managed_pages(zone);
7884 unsigned long managed_pages = 0;
7886 for (j = i + 1; j < MAX_NR_ZONES; j++) {
7888 zone->lowmem_reserve[j] = 0;
7890 struct zone *upper_zone = &pgdat->node_zones[j];
7892 managed_pages += zone_managed_pages(upper_zone);
7893 zone->lowmem_reserve[j] = managed_pages / ratio;
7899 /* update totalreserve_pages */
7900 calculate_totalreserve_pages();
7903 static void __setup_per_zone_wmarks(void)
7905 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7906 unsigned long lowmem_pages = 0;
7908 unsigned long flags;
7910 /* Calculate total number of !ZONE_HIGHMEM pages */
7911 for_each_zone(zone) {
7912 if (!is_highmem(zone))
7913 lowmem_pages += zone_managed_pages(zone);
7916 for_each_zone(zone) {
7919 spin_lock_irqsave(&zone->lock, flags);
7920 tmp = (u64)pages_min * zone_managed_pages(zone);
7921 do_div(tmp, lowmem_pages);
7922 if (is_highmem(zone)) {
7924 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7925 * need highmem pages, so cap pages_min to a small
7928 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7929 * deltas control async page reclaim, and so should
7930 * not be capped for highmem.
7932 unsigned long min_pages;
7934 min_pages = zone_managed_pages(zone) / 1024;
7935 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7936 zone->_watermark[WMARK_MIN] = min_pages;
7939 * If it's a lowmem zone, reserve a number of pages
7940 * proportionate to the zone's size.
7942 zone->_watermark[WMARK_MIN] = tmp;
7946 * Set the kswapd watermarks distance according to the
7947 * scale factor in proportion to available memory, but
7948 * ensure a minimum size on small systems.
7950 tmp = max_t(u64, tmp >> 2,
7951 mult_frac(zone_managed_pages(zone),
7952 watermark_scale_factor, 10000));
7954 zone->watermark_boost = 0;
7955 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7956 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7958 spin_unlock_irqrestore(&zone->lock, flags);
7961 /* update totalreserve_pages */
7962 calculate_totalreserve_pages();
7966 * setup_per_zone_wmarks - called when min_free_kbytes changes
7967 * or when memory is hot-{added|removed}
7969 * Ensures that the watermark[min,low,high] values for each zone are set
7970 * correctly with respect to min_free_kbytes.
7972 void setup_per_zone_wmarks(void)
7974 static DEFINE_SPINLOCK(lock);
7977 __setup_per_zone_wmarks();
7982 * Initialise min_free_kbytes.
7984 * For small machines we want it small (128k min). For large machines
7985 * we want it large (256MB max). But it is not linear, because network
7986 * bandwidth does not increase linearly with machine size. We use
7988 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7989 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8005 int __meminit init_per_zone_wmark_min(void)
8007 unsigned long lowmem_kbytes;
8008 int new_min_free_kbytes;
8010 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8011 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8013 if (new_min_free_kbytes > user_min_free_kbytes) {
8014 min_free_kbytes = new_min_free_kbytes;
8015 if (min_free_kbytes < 128)
8016 min_free_kbytes = 128;
8017 if (min_free_kbytes > 262144)
8018 min_free_kbytes = 262144;
8020 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8021 new_min_free_kbytes, user_min_free_kbytes);
8023 setup_per_zone_wmarks();
8024 refresh_zone_stat_thresholds();
8025 setup_per_zone_lowmem_reserve();
8028 setup_min_unmapped_ratio();
8029 setup_min_slab_ratio();
8032 khugepaged_min_free_kbytes_update();
8036 postcore_initcall(init_per_zone_wmark_min)
8039 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8040 * that we can call two helper functions whenever min_free_kbytes
8043 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8044 void *buffer, size_t *length, loff_t *ppos)
8048 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8053 user_min_free_kbytes = min_free_kbytes;
8054 setup_per_zone_wmarks();
8059 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8060 void *buffer, size_t *length, loff_t *ppos)
8064 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8069 setup_per_zone_wmarks();
8075 static void setup_min_unmapped_ratio(void)
8080 for_each_online_pgdat(pgdat)
8081 pgdat->min_unmapped_pages = 0;
8084 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8085 sysctl_min_unmapped_ratio) / 100;
8089 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8090 void *buffer, size_t *length, loff_t *ppos)
8094 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8098 setup_min_unmapped_ratio();
8103 static void setup_min_slab_ratio(void)
8108 for_each_online_pgdat(pgdat)
8109 pgdat->min_slab_pages = 0;
8112 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8113 sysctl_min_slab_ratio) / 100;
8116 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8117 void *buffer, size_t *length, loff_t *ppos)
8121 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8125 setup_min_slab_ratio();
8132 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8133 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8134 * whenever sysctl_lowmem_reserve_ratio changes.
8136 * The reserve ratio obviously has absolutely no relation with the
8137 * minimum watermarks. The lowmem reserve ratio can only make sense
8138 * if in function of the boot time zone sizes.
8140 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8141 void *buffer, size_t *length, loff_t *ppos)
8145 proc_dointvec_minmax(table, write, buffer, length, ppos);
8147 for (i = 0; i < MAX_NR_ZONES; i++) {
8148 if (sysctl_lowmem_reserve_ratio[i] < 1)
8149 sysctl_lowmem_reserve_ratio[i] = 0;
8152 setup_per_zone_lowmem_reserve();
8157 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8158 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8159 * pagelist can have before it gets flushed back to buddy allocator.
8161 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8162 void *buffer, size_t *length, loff_t *ppos)
8165 int old_percpu_pagelist_fraction;
8168 mutex_lock(&pcp_batch_high_lock);
8169 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8171 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8172 if (!write || ret < 0)
8175 /* Sanity checking to avoid pcp imbalance */
8176 if (percpu_pagelist_fraction &&
8177 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8178 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8184 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8187 for_each_populated_zone(zone)
8188 zone_set_pageset_high_and_batch(zone);
8190 mutex_unlock(&pcp_batch_high_lock);
8194 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8196 * Returns the number of pages that arch has reserved but
8197 * is not known to alloc_large_system_hash().
8199 static unsigned long __init arch_reserved_kernel_pages(void)
8206 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8207 * machines. As memory size is increased the scale is also increased but at
8208 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8209 * quadruples the scale is increased by one, which means the size of hash table
8210 * only doubles, instead of quadrupling as well.
8211 * Because 32-bit systems cannot have large physical memory, where this scaling
8212 * makes sense, it is disabled on such platforms.
8214 #if __BITS_PER_LONG > 32
8215 #define ADAPT_SCALE_BASE (64ul << 30)
8216 #define ADAPT_SCALE_SHIFT 2
8217 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8221 * allocate a large system hash table from bootmem
8222 * - it is assumed that the hash table must contain an exact power-of-2
8223 * quantity of entries
8224 * - limit is the number of hash buckets, not the total allocation size
8226 void *__init alloc_large_system_hash(const char *tablename,
8227 unsigned long bucketsize,
8228 unsigned long numentries,
8231 unsigned int *_hash_shift,
8232 unsigned int *_hash_mask,
8233 unsigned long low_limit,
8234 unsigned long high_limit)
8236 unsigned long long max = high_limit;
8237 unsigned long log2qty, size;
8242 /* allow the kernel cmdline to have a say */
8244 /* round applicable memory size up to nearest megabyte */
8245 numentries = nr_kernel_pages;
8246 numentries -= arch_reserved_kernel_pages();
8248 /* It isn't necessary when PAGE_SIZE >= 1MB */
8249 if (PAGE_SHIFT < 20)
8250 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8252 #if __BITS_PER_LONG > 32
8254 unsigned long adapt;
8256 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8257 adapt <<= ADAPT_SCALE_SHIFT)
8262 /* limit to 1 bucket per 2^scale bytes of low memory */
8263 if (scale > PAGE_SHIFT)
8264 numentries >>= (scale - PAGE_SHIFT);
8266 numentries <<= (PAGE_SHIFT - scale);
8268 /* Make sure we've got at least a 0-order allocation.. */
8269 if (unlikely(flags & HASH_SMALL)) {
8270 /* Makes no sense without HASH_EARLY */
8271 WARN_ON(!(flags & HASH_EARLY));
8272 if (!(numentries >> *_hash_shift)) {
8273 numentries = 1UL << *_hash_shift;
8274 BUG_ON(!numentries);
8276 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8277 numentries = PAGE_SIZE / bucketsize;
8279 numentries = roundup_pow_of_two(numentries);
8281 /* limit allocation size to 1/16 total memory by default */
8283 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8284 do_div(max, bucketsize);
8286 max = min(max, 0x80000000ULL);
8288 if (numentries < low_limit)
8289 numentries = low_limit;
8290 if (numentries > max)
8293 log2qty = ilog2(numentries);
8295 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8298 size = bucketsize << log2qty;
8299 if (flags & HASH_EARLY) {
8300 if (flags & HASH_ZERO)
8301 table = memblock_alloc(size, SMP_CACHE_BYTES);
8303 table = memblock_alloc_raw(size,
8305 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8306 table = __vmalloc(size, gfp_flags);
8310 * If bucketsize is not a power-of-two, we may free
8311 * some pages at the end of hash table which
8312 * alloc_pages_exact() automatically does
8314 table = alloc_pages_exact(size, gfp_flags);
8315 kmemleak_alloc(table, size, 1, gfp_flags);
8317 } while (!table && size > PAGE_SIZE && --log2qty);
8320 panic("Failed to allocate %s hash table\n", tablename);
8322 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8323 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8324 virt ? "vmalloc" : "linear");
8327 *_hash_shift = log2qty;
8329 *_hash_mask = (1 << log2qty) - 1;
8335 * This function checks whether pageblock includes unmovable pages or not.
8337 * PageLRU check without isolation or lru_lock could race so that
8338 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8339 * check without lock_page also may miss some movable non-lru pages at
8340 * race condition. So you can't expect this function should be exact.
8342 * Returns a page without holding a reference. If the caller wants to
8343 * dereference that page (e.g., dumping), it has to make sure that it
8344 * cannot get removed (e.g., via memory unplug) concurrently.
8347 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8348 int migratetype, int flags)
8350 unsigned long iter = 0;
8351 unsigned long pfn = page_to_pfn(page);
8352 unsigned long offset = pfn % pageblock_nr_pages;
8354 if (is_migrate_cma_page(page)) {
8356 * CMA allocations (alloc_contig_range) really need to mark
8357 * isolate CMA pageblocks even when they are not movable in fact
8358 * so consider them movable here.
8360 if (is_migrate_cma(migratetype))
8366 for (; iter < pageblock_nr_pages - offset; iter++) {
8367 if (!pfn_valid_within(pfn + iter))
8370 page = pfn_to_page(pfn + iter);
8373 * Both, bootmem allocations and memory holes are marked
8374 * PG_reserved and are unmovable. We can even have unmovable
8375 * allocations inside ZONE_MOVABLE, for example when
8376 * specifying "movablecore".
8378 if (PageReserved(page))
8382 * If the zone is movable and we have ruled out all reserved
8383 * pages then it should be reasonably safe to assume the rest
8386 if (zone_idx(zone) == ZONE_MOVABLE)
8390 * Hugepages are not in LRU lists, but they're movable.
8391 * THPs are on the LRU, but need to be counted as #small pages.
8392 * We need not scan over tail pages because we don't
8393 * handle each tail page individually in migration.
8395 if (PageHuge(page) || PageTransCompound(page)) {
8396 struct page *head = compound_head(page);
8397 unsigned int skip_pages;
8399 if (PageHuge(page)) {
8400 if (!hugepage_migration_supported(page_hstate(head)))
8402 } else if (!PageLRU(head) && !__PageMovable(head)) {
8406 skip_pages = compound_nr(head) - (page - head);
8407 iter += skip_pages - 1;
8412 * We can't use page_count without pin a page
8413 * because another CPU can free compound page.
8414 * This check already skips compound tails of THP
8415 * because their page->_refcount is zero at all time.
8417 if (!page_ref_count(page)) {
8418 if (PageBuddy(page))
8419 iter += (1 << buddy_order(page)) - 1;
8424 * The HWPoisoned page may be not in buddy system, and
8425 * page_count() is not 0.
8427 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8431 * We treat all PageOffline() pages as movable when offlining
8432 * to give drivers a chance to decrement their reference count
8433 * in MEM_GOING_OFFLINE in order to indicate that these pages
8434 * can be offlined as there are no direct references anymore.
8435 * For actually unmovable PageOffline() where the driver does
8436 * not support this, we will fail later when trying to actually
8437 * move these pages that still have a reference count > 0.
8438 * (false negatives in this function only)
8440 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8443 if (__PageMovable(page) || PageLRU(page))
8447 * If there are RECLAIMABLE pages, we need to check
8448 * it. But now, memory offline itself doesn't call
8449 * shrink_node_slabs() and it still to be fixed.
8456 #ifdef CONFIG_CONTIG_ALLOC
8457 static unsigned long pfn_max_align_down(unsigned long pfn)
8459 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8460 pageblock_nr_pages) - 1);
8463 static unsigned long pfn_max_align_up(unsigned long pfn)
8465 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8466 pageblock_nr_pages));
8469 /* [start, end) must belong to a single zone. */
8470 static int __alloc_contig_migrate_range(struct compact_control *cc,
8471 unsigned long start, unsigned long end)
8473 /* This function is based on compact_zone() from compaction.c. */
8474 unsigned int nr_reclaimed;
8475 unsigned long pfn = start;
8476 unsigned int tries = 0;
8478 struct migration_target_control mtc = {
8479 .nid = zone_to_nid(cc->zone),
8480 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8485 while (pfn < end || !list_empty(&cc->migratepages)) {
8486 if (fatal_signal_pending(current)) {
8491 if (list_empty(&cc->migratepages)) {
8492 cc->nr_migratepages = 0;
8493 pfn = isolate_migratepages_range(cc, pfn, end);
8499 } else if (++tries == 5) {
8500 ret = ret < 0 ? ret : -EBUSY;
8504 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8506 cc->nr_migratepages -= nr_reclaimed;
8508 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8509 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8512 putback_movable_pages(&cc->migratepages);
8519 * alloc_contig_range() -- tries to allocate given range of pages
8520 * @start: start PFN to allocate
8521 * @end: one-past-the-last PFN to allocate
8522 * @migratetype: migratetype of the underlaying pageblocks (either
8523 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8524 * in range must have the same migratetype and it must
8525 * be either of the two.
8526 * @gfp_mask: GFP mask to use during compaction
8528 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8529 * aligned. The PFN range must belong to a single zone.
8531 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8532 * pageblocks in the range. Once isolated, the pageblocks should not
8533 * be modified by others.
8535 * Return: zero on success or negative error code. On success all
8536 * pages which PFN is in [start, end) are allocated for the caller and
8537 * need to be freed with free_contig_range().
8539 int alloc_contig_range(unsigned long start, unsigned long end,
8540 unsigned migratetype, gfp_t gfp_mask)
8542 unsigned long outer_start, outer_end;
8546 struct compact_control cc = {
8547 .nr_migratepages = 0,
8549 .zone = page_zone(pfn_to_page(start)),
8550 .mode = MIGRATE_SYNC,
8551 .ignore_skip_hint = true,
8552 .no_set_skip_hint = true,
8553 .gfp_mask = current_gfp_context(gfp_mask),
8554 .alloc_contig = true,
8556 INIT_LIST_HEAD(&cc.migratepages);
8559 * What we do here is we mark all pageblocks in range as
8560 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8561 * have different sizes, and due to the way page allocator
8562 * work, we align the range to biggest of the two pages so
8563 * that page allocator won't try to merge buddies from
8564 * different pageblocks and change MIGRATE_ISOLATE to some
8565 * other migration type.
8567 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8568 * migrate the pages from an unaligned range (ie. pages that
8569 * we are interested in). This will put all the pages in
8570 * range back to page allocator as MIGRATE_ISOLATE.
8572 * When this is done, we take the pages in range from page
8573 * allocator removing them from the buddy system. This way
8574 * page allocator will never consider using them.
8576 * This lets us mark the pageblocks back as
8577 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8578 * aligned range but not in the unaligned, original range are
8579 * put back to page allocator so that buddy can use them.
8582 ret = start_isolate_page_range(pfn_max_align_down(start),
8583 pfn_max_align_up(end), migratetype, 0);
8587 drain_all_pages(cc.zone);
8590 * In case of -EBUSY, we'd like to know which page causes problem.
8591 * So, just fall through. test_pages_isolated() has a tracepoint
8592 * which will report the busy page.
8594 * It is possible that busy pages could become available before
8595 * the call to test_pages_isolated, and the range will actually be
8596 * allocated. So, if we fall through be sure to clear ret so that
8597 * -EBUSY is not accidentally used or returned to caller.
8599 ret = __alloc_contig_migrate_range(&cc, start, end);
8600 if (ret && ret != -EBUSY)
8605 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8606 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8607 * more, all pages in [start, end) are free in page allocator.
8608 * What we are going to do is to allocate all pages from
8609 * [start, end) (that is remove them from page allocator).
8611 * The only problem is that pages at the beginning and at the
8612 * end of interesting range may be not aligned with pages that
8613 * page allocator holds, ie. they can be part of higher order
8614 * pages. Because of this, we reserve the bigger range and
8615 * once this is done free the pages we are not interested in.
8617 * We don't have to hold zone->lock here because the pages are
8618 * isolated thus they won't get removed from buddy.
8621 lru_add_drain_all();
8624 outer_start = start;
8625 while (!PageBuddy(pfn_to_page(outer_start))) {
8626 if (++order >= MAX_ORDER) {
8627 outer_start = start;
8630 outer_start &= ~0UL << order;
8633 if (outer_start != start) {
8634 order = buddy_order(pfn_to_page(outer_start));
8637 * outer_start page could be small order buddy page and
8638 * it doesn't include start page. Adjust outer_start
8639 * in this case to report failed page properly
8640 * on tracepoint in test_pages_isolated()
8642 if (outer_start + (1UL << order) <= start)
8643 outer_start = start;
8646 /* Make sure the range is really isolated. */
8647 if (test_pages_isolated(outer_start, end, 0)) {
8648 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8649 __func__, outer_start, end);
8654 /* Grab isolated pages from freelists. */
8655 outer_end = isolate_freepages_range(&cc, outer_start, end);
8661 /* Free head and tail (if any) */
8662 if (start != outer_start)
8663 free_contig_range(outer_start, start - outer_start);
8664 if (end != outer_end)
8665 free_contig_range(end, outer_end - end);
8668 undo_isolate_page_range(pfn_max_align_down(start),
8669 pfn_max_align_up(end), migratetype);
8672 EXPORT_SYMBOL(alloc_contig_range);
8674 static int __alloc_contig_pages(unsigned long start_pfn,
8675 unsigned long nr_pages, gfp_t gfp_mask)
8677 unsigned long end_pfn = start_pfn + nr_pages;
8679 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8683 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8684 unsigned long nr_pages)
8686 unsigned long i, end_pfn = start_pfn + nr_pages;
8689 for (i = start_pfn; i < end_pfn; i++) {
8690 page = pfn_to_online_page(i);
8694 if (page_zone(page) != z)
8697 if (PageReserved(page))
8700 if (page_count(page) > 0)
8709 static bool zone_spans_last_pfn(const struct zone *zone,
8710 unsigned long start_pfn, unsigned long nr_pages)
8712 unsigned long last_pfn = start_pfn + nr_pages - 1;
8714 return zone_spans_pfn(zone, last_pfn);
8718 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8719 * @nr_pages: Number of contiguous pages to allocate
8720 * @gfp_mask: GFP mask to limit search and used during compaction
8722 * @nodemask: Mask for other possible nodes
8724 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8725 * on an applicable zonelist to find a contiguous pfn range which can then be
8726 * tried for allocation with alloc_contig_range(). This routine is intended
8727 * for allocation requests which can not be fulfilled with the buddy allocator.
8729 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8730 * power of two then the alignment is guaranteed to be to the given nr_pages
8731 * (e.g. 1GB request would be aligned to 1GB).
8733 * Allocated pages can be freed with free_contig_range() or by manually calling
8734 * __free_page() on each allocated page.
8736 * Return: pointer to contiguous pages on success, or NULL if not successful.
8738 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8739 int nid, nodemask_t *nodemask)
8741 unsigned long ret, pfn, flags;
8742 struct zonelist *zonelist;
8746 zonelist = node_zonelist(nid, gfp_mask);
8747 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8748 gfp_zone(gfp_mask), nodemask) {
8749 spin_lock_irqsave(&zone->lock, flags);
8751 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8752 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8753 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8755 * We release the zone lock here because
8756 * alloc_contig_range() will also lock the zone
8757 * at some point. If there's an allocation
8758 * spinning on this lock, it may win the race
8759 * and cause alloc_contig_range() to fail...
8761 spin_unlock_irqrestore(&zone->lock, flags);
8762 ret = __alloc_contig_pages(pfn, nr_pages,
8765 return pfn_to_page(pfn);
8766 spin_lock_irqsave(&zone->lock, flags);
8770 spin_unlock_irqrestore(&zone->lock, flags);
8774 #endif /* CONFIG_CONTIG_ALLOC */
8776 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8778 unsigned int count = 0;
8780 for (; nr_pages--; pfn++) {
8781 struct page *page = pfn_to_page(pfn);
8783 count += page_count(page) != 1;
8786 WARN(count != 0, "%d pages are still in use!\n", count);
8788 EXPORT_SYMBOL(free_contig_range);
8791 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8792 * page high values need to be recalulated.
8794 void __meminit zone_pcp_update(struct zone *zone)
8796 mutex_lock(&pcp_batch_high_lock);
8797 zone_set_pageset_high_and_batch(zone);
8798 mutex_unlock(&pcp_batch_high_lock);
8802 * Effectively disable pcplists for the zone by setting the high limit to 0
8803 * and draining all cpus. A concurrent page freeing on another CPU that's about
8804 * to put the page on pcplist will either finish before the drain and the page
8805 * will be drained, or observe the new high limit and skip the pcplist.
8807 * Must be paired with a call to zone_pcp_enable().
8809 void zone_pcp_disable(struct zone *zone)
8811 mutex_lock(&pcp_batch_high_lock);
8812 __zone_set_pageset_high_and_batch(zone, 0, 1);
8813 __drain_all_pages(zone, true);
8816 void zone_pcp_enable(struct zone *zone)
8818 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
8819 mutex_unlock(&pcp_batch_high_lock);
8822 void zone_pcp_reset(struct zone *zone)
8824 unsigned long flags;
8826 struct per_cpu_pageset *pset;
8828 /* avoid races with drain_pages() */
8829 local_irq_save(flags);
8830 if (zone->pageset != &boot_pageset) {
8831 for_each_online_cpu(cpu) {
8832 pset = per_cpu_ptr(zone->pageset, cpu);
8833 drain_zonestat(zone, pset);
8835 free_percpu(zone->pageset);
8836 zone->pageset = &boot_pageset;
8838 local_irq_restore(flags);
8841 #ifdef CONFIG_MEMORY_HOTREMOVE
8843 * All pages in the range must be in a single zone, must not contain holes,
8844 * must span full sections, and must be isolated before calling this function.
8846 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8848 unsigned long pfn = start_pfn;
8852 unsigned long flags;
8854 offline_mem_sections(pfn, end_pfn);
8855 zone = page_zone(pfn_to_page(pfn));
8856 spin_lock_irqsave(&zone->lock, flags);
8857 while (pfn < end_pfn) {
8858 page = pfn_to_page(pfn);
8860 * The HWPoisoned page may be not in buddy system, and
8861 * page_count() is not 0.
8863 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8868 * At this point all remaining PageOffline() pages have a
8869 * reference count of 0 and can simply be skipped.
8871 if (PageOffline(page)) {
8872 BUG_ON(page_count(page));
8873 BUG_ON(PageBuddy(page));
8878 BUG_ON(page_count(page));
8879 BUG_ON(!PageBuddy(page));
8880 order = buddy_order(page);
8881 del_page_from_free_list(page, zone, order);
8882 pfn += (1 << order);
8884 spin_unlock_irqrestore(&zone->lock, flags);
8888 bool is_free_buddy_page(struct page *page)
8890 struct zone *zone = page_zone(page);
8891 unsigned long pfn = page_to_pfn(page);
8892 unsigned long flags;
8895 spin_lock_irqsave(&zone->lock, flags);
8896 for (order = 0; order < MAX_ORDER; order++) {
8897 struct page *page_head = page - (pfn & ((1 << order) - 1));
8899 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
8902 spin_unlock_irqrestore(&zone->lock, flags);
8904 return order < MAX_ORDER;
8907 #ifdef CONFIG_MEMORY_FAILURE
8909 * Break down a higher-order page in sub-pages, and keep our target out of
8912 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8913 struct page *target, int low, int high,
8916 unsigned long size = 1 << high;
8917 struct page *current_buddy, *next_page;
8919 while (high > low) {
8923 if (target >= &page[size]) {
8924 next_page = page + size;
8925 current_buddy = page;
8928 current_buddy = page + size;
8931 if (set_page_guard(zone, current_buddy, high, migratetype))
8934 if (current_buddy != target) {
8935 add_to_free_list(current_buddy, zone, high, migratetype);
8936 set_buddy_order(current_buddy, high);
8943 * Take a page that will be marked as poisoned off the buddy allocator.
8945 bool take_page_off_buddy(struct page *page)
8947 struct zone *zone = page_zone(page);
8948 unsigned long pfn = page_to_pfn(page);
8949 unsigned long flags;
8953 spin_lock_irqsave(&zone->lock, flags);
8954 for (order = 0; order < MAX_ORDER; order++) {
8955 struct page *page_head = page - (pfn & ((1 << order) - 1));
8956 int page_order = buddy_order(page_head);
8958 if (PageBuddy(page_head) && page_order >= order) {
8959 unsigned long pfn_head = page_to_pfn(page_head);
8960 int migratetype = get_pfnblock_migratetype(page_head,
8963 del_page_from_free_list(page_head, zone, page_order);
8964 break_down_buddy_pages(zone, page_head, page, 0,
8965 page_order, migratetype);
8969 if (page_count(page_head) > 0)
8972 spin_unlock_irqrestore(&zone->lock, flags);