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 u8 tag = page_kasan_tag(page + i);
1211 page_kasan_tag_reset(page + i);
1212 clear_highpage(page + i);
1213 page_kasan_tag_set(page + i, tag);
1215 kasan_enable_current();
1218 static __always_inline bool free_pages_prepare(struct page *page,
1219 unsigned int order, bool check_free)
1223 VM_BUG_ON_PAGE(PageTail(page), page);
1225 trace_mm_page_free(page, order);
1227 if (unlikely(PageHWPoison(page)) && !order) {
1229 * Do not let hwpoison pages hit pcplists/buddy
1230 * Untie memcg state and reset page's owner
1232 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1233 __memcg_kmem_uncharge_page(page, order);
1234 reset_page_owner(page, order);
1239 * Check tail pages before head page information is cleared to
1240 * avoid checking PageCompound for order-0 pages.
1242 if (unlikely(order)) {
1243 bool compound = PageCompound(page);
1246 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1249 ClearPageDoubleMap(page);
1250 for (i = 1; i < (1 << order); i++) {
1252 bad += free_tail_pages_check(page, page + i);
1253 if (unlikely(check_free_page(page + i))) {
1257 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1260 if (PageMappingFlags(page))
1261 page->mapping = NULL;
1262 if (memcg_kmem_enabled() && PageMemcgKmem(page))
1263 __memcg_kmem_uncharge_page(page, order);
1265 bad += check_free_page(page);
1269 page_cpupid_reset_last(page);
1270 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1271 reset_page_owner(page, order);
1273 if (!PageHighMem(page)) {
1274 debug_check_no_locks_freed(page_address(page),
1275 PAGE_SIZE << order);
1276 debug_check_no_obj_freed(page_address(page),
1277 PAGE_SIZE << order);
1279 if (want_init_on_free())
1280 kernel_init_free_pages(page, 1 << order);
1282 kernel_poison_pages(page, 1 << order);
1285 * arch_free_page() can make the page's contents inaccessible. s390
1286 * does this. So nothing which can access the page's contents should
1287 * happen after this.
1289 arch_free_page(page, order);
1291 debug_pagealloc_unmap_pages(page, 1 << order);
1293 kasan_free_nondeferred_pages(page, order);
1298 #ifdef CONFIG_DEBUG_VM
1300 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1301 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1302 * moved from pcp lists to free lists.
1304 static bool free_pcp_prepare(struct page *page)
1306 return free_pages_prepare(page, 0, true);
1309 static bool bulkfree_pcp_prepare(struct page *page)
1311 if (debug_pagealloc_enabled_static())
1312 return check_free_page(page);
1318 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1319 * moving from pcp lists to free list in order to reduce overhead. With
1320 * debug_pagealloc enabled, they are checked also immediately when being freed
1323 static bool free_pcp_prepare(struct page *page)
1325 if (debug_pagealloc_enabled_static())
1326 return free_pages_prepare(page, 0, true);
1328 return free_pages_prepare(page, 0, false);
1331 static bool bulkfree_pcp_prepare(struct page *page)
1333 return check_free_page(page);
1335 #endif /* CONFIG_DEBUG_VM */
1337 static inline void prefetch_buddy(struct page *page)
1339 unsigned long pfn = page_to_pfn(page);
1340 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1341 struct page *buddy = page + (buddy_pfn - pfn);
1347 * Frees a number of pages from the PCP lists
1348 * Assumes all pages on list are in same zone, and of same order.
1349 * count is the number of pages to free.
1351 * If the zone was previously in an "all pages pinned" state then look to
1352 * see if this freeing clears that state.
1354 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1355 * pinned" detection logic.
1357 static void free_pcppages_bulk(struct zone *zone, int count,
1358 struct per_cpu_pages *pcp)
1360 int migratetype = 0;
1362 int prefetch_nr = READ_ONCE(pcp->batch);
1363 bool isolated_pageblocks;
1364 struct page *page, *tmp;
1368 * Ensure proper count is passed which otherwise would stuck in the
1369 * below while (list_empty(list)) loop.
1371 count = min(pcp->count, count);
1373 struct list_head *list;
1376 * Remove pages from lists in a round-robin fashion. A
1377 * batch_free count is maintained that is incremented when an
1378 * empty list is encountered. This is so more pages are freed
1379 * off fuller lists instead of spinning excessively around empty
1384 if (++migratetype == MIGRATE_PCPTYPES)
1386 list = &pcp->lists[migratetype];
1387 } while (list_empty(list));
1389 /* This is the only non-empty list. Free them all. */
1390 if (batch_free == MIGRATE_PCPTYPES)
1394 page = list_last_entry(list, struct page, lru);
1395 /* must delete to avoid corrupting pcp list */
1396 list_del(&page->lru);
1399 if (bulkfree_pcp_prepare(page))
1402 list_add_tail(&page->lru, &head);
1405 * We are going to put the page back to the global
1406 * pool, prefetch its buddy to speed up later access
1407 * under zone->lock. It is believed the overhead of
1408 * an additional test and calculating buddy_pfn here
1409 * can be offset by reduced memory latency later. To
1410 * avoid excessive prefetching due to large count, only
1411 * prefetch buddy for the first pcp->batch nr of pages.
1414 prefetch_buddy(page);
1417 } while (--count && --batch_free && !list_empty(list));
1420 spin_lock(&zone->lock);
1421 isolated_pageblocks = has_isolate_pageblock(zone);
1424 * Use safe version since after __free_one_page(),
1425 * page->lru.next will not point to original list.
1427 list_for_each_entry_safe(page, tmp, &head, lru) {
1428 int mt = get_pcppage_migratetype(page);
1429 /* MIGRATE_ISOLATE page should not go to pcplists */
1430 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1431 /* Pageblock could have been isolated meanwhile */
1432 if (unlikely(isolated_pageblocks))
1433 mt = get_pageblock_migratetype(page);
1435 __free_one_page(page, page_to_pfn(page), zone, 0, mt, FPI_NONE);
1436 trace_mm_page_pcpu_drain(page, 0, mt);
1438 spin_unlock(&zone->lock);
1441 static void free_one_page(struct zone *zone,
1442 struct page *page, unsigned long pfn,
1444 int migratetype, fpi_t fpi_flags)
1446 spin_lock(&zone->lock);
1447 if (unlikely(has_isolate_pageblock(zone) ||
1448 is_migrate_isolate(migratetype))) {
1449 migratetype = get_pfnblock_migratetype(page, pfn);
1451 __free_one_page(page, pfn, zone, order, migratetype, fpi_flags);
1452 spin_unlock(&zone->lock);
1455 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1456 unsigned long zone, int nid)
1458 mm_zero_struct_page(page);
1459 set_page_links(page, zone, nid, pfn);
1460 init_page_count(page);
1461 page_mapcount_reset(page);
1462 page_cpupid_reset_last(page);
1463 page_kasan_tag_reset(page);
1465 INIT_LIST_HEAD(&page->lru);
1466 #ifdef WANT_PAGE_VIRTUAL
1467 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1468 if (!is_highmem_idx(zone))
1469 set_page_address(page, __va(pfn << PAGE_SHIFT));
1473 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1474 static void __meminit init_reserved_page(unsigned long pfn)
1479 if (!early_page_uninitialised(pfn))
1482 nid = early_pfn_to_nid(pfn);
1483 pgdat = NODE_DATA(nid);
1485 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1486 struct zone *zone = &pgdat->node_zones[zid];
1488 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1491 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1494 static inline void init_reserved_page(unsigned long pfn)
1497 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1500 * Initialised pages do not have PageReserved set. This function is
1501 * called for each range allocated by the bootmem allocator and
1502 * marks the pages PageReserved. The remaining valid pages are later
1503 * sent to the buddy page allocator.
1505 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1507 unsigned long start_pfn = PFN_DOWN(start);
1508 unsigned long end_pfn = PFN_UP(end);
1510 for (; start_pfn < end_pfn; start_pfn++) {
1511 if (pfn_valid(start_pfn)) {
1512 struct page *page = pfn_to_page(start_pfn);
1514 init_reserved_page(start_pfn);
1516 /* Avoid false-positive PageTail() */
1517 INIT_LIST_HEAD(&page->lru);
1520 * no need for atomic set_bit because the struct
1521 * page is not visible yet so nobody should
1524 __SetPageReserved(page);
1529 static void __free_pages_ok(struct page *page, unsigned int order,
1532 unsigned long flags;
1534 unsigned long pfn = page_to_pfn(page);
1536 if (!free_pages_prepare(page, order, true))
1539 migratetype = get_pfnblock_migratetype(page, pfn);
1540 local_irq_save(flags);
1541 __count_vm_events(PGFREE, 1 << order);
1542 free_one_page(page_zone(page), page, pfn, order, migratetype,
1544 local_irq_restore(flags);
1547 void __free_pages_core(struct page *page, unsigned int order)
1549 unsigned int nr_pages = 1 << order;
1550 struct page *p = page;
1554 * When initializing the memmap, __init_single_page() sets the refcount
1555 * of all pages to 1 ("allocated"/"not free"). We have to set the
1556 * refcount of all involved pages to 0.
1559 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1561 __ClearPageReserved(p);
1562 set_page_count(p, 0);
1564 __ClearPageReserved(p);
1565 set_page_count(p, 0);
1567 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1570 * Bypass PCP and place fresh pages right to the tail, primarily
1571 * relevant for memory onlining.
1573 __free_pages_ok(page, order, FPI_TO_TAIL);
1576 #ifdef CONFIG_NEED_MULTIPLE_NODES
1579 * During memory init memblocks map pfns to nids. The search is expensive and
1580 * this caches recent lookups. The implementation of __early_pfn_to_nid
1581 * treats start/end as pfns.
1583 struct mminit_pfnnid_cache {
1584 unsigned long last_start;
1585 unsigned long last_end;
1589 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1592 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1594 static int __meminit __early_pfn_to_nid(unsigned long pfn,
1595 struct mminit_pfnnid_cache *state)
1597 unsigned long start_pfn, end_pfn;
1600 if (state->last_start <= pfn && pfn < state->last_end)
1601 return state->last_nid;
1603 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1604 if (nid != NUMA_NO_NODE) {
1605 state->last_start = start_pfn;
1606 state->last_end = end_pfn;
1607 state->last_nid = nid;
1613 int __meminit early_pfn_to_nid(unsigned long pfn)
1615 static DEFINE_SPINLOCK(early_pfn_lock);
1618 spin_lock(&early_pfn_lock);
1619 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1621 nid = first_online_node;
1622 spin_unlock(&early_pfn_lock);
1626 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1628 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1631 if (early_page_uninitialised(pfn))
1633 __free_pages_core(page, order);
1637 * Check that the whole (or subset of) a pageblock given by the interval of
1638 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1639 * with the migration of free compaction scanner. The scanners then need to
1640 * use only pfn_valid_within() check for arches that allow holes within
1643 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1645 * It's possible on some configurations to have a setup like node0 node1 node0
1646 * i.e. it's possible that all pages within a zones range of pages do not
1647 * belong to a single zone. We assume that a border between node0 and node1
1648 * can occur within a single pageblock, but not a node0 node1 node0
1649 * interleaving within a single pageblock. It is therefore sufficient to check
1650 * the first and last page of a pageblock and avoid checking each individual
1651 * page in a pageblock.
1653 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1654 unsigned long end_pfn, struct zone *zone)
1656 struct page *start_page;
1657 struct page *end_page;
1659 /* end_pfn is one past the range we are checking */
1662 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1665 start_page = pfn_to_online_page(start_pfn);
1669 if (page_zone(start_page) != zone)
1672 end_page = pfn_to_page(end_pfn);
1674 /* This gives a shorter code than deriving page_zone(end_page) */
1675 if (page_zone_id(start_page) != page_zone_id(end_page))
1681 void set_zone_contiguous(struct zone *zone)
1683 unsigned long block_start_pfn = zone->zone_start_pfn;
1684 unsigned long block_end_pfn;
1686 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1687 for (; block_start_pfn < zone_end_pfn(zone);
1688 block_start_pfn = block_end_pfn,
1689 block_end_pfn += pageblock_nr_pages) {
1691 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1693 if (!__pageblock_pfn_to_page(block_start_pfn,
1694 block_end_pfn, zone))
1699 /* We confirm that there is no hole */
1700 zone->contiguous = true;
1703 void clear_zone_contiguous(struct zone *zone)
1705 zone->contiguous = false;
1708 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1709 static void __init deferred_free_range(unsigned long pfn,
1710 unsigned long nr_pages)
1718 page = pfn_to_page(pfn);
1720 /* Free a large naturally-aligned chunk if possible */
1721 if (nr_pages == pageblock_nr_pages &&
1722 (pfn & (pageblock_nr_pages - 1)) == 0) {
1723 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1724 __free_pages_core(page, pageblock_order);
1728 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1729 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1730 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1731 __free_pages_core(page, 0);
1735 /* Completion tracking for deferred_init_memmap() threads */
1736 static atomic_t pgdat_init_n_undone __initdata;
1737 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1739 static inline void __init pgdat_init_report_one_done(void)
1741 if (atomic_dec_and_test(&pgdat_init_n_undone))
1742 complete(&pgdat_init_all_done_comp);
1746 * Returns true if page needs to be initialized or freed to buddy allocator.
1748 * First we check if pfn is valid on architectures where it is possible to have
1749 * holes within pageblock_nr_pages. On systems where it is not possible, this
1750 * function is optimized out.
1752 * Then, we check if a current large page is valid by only checking the validity
1755 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1757 if (!pfn_valid_within(pfn))
1759 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1765 * Free pages to buddy allocator. Try to free aligned pages in
1766 * pageblock_nr_pages sizes.
1768 static void __init deferred_free_pages(unsigned long pfn,
1769 unsigned long end_pfn)
1771 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1772 unsigned long nr_free = 0;
1774 for (; pfn < end_pfn; pfn++) {
1775 if (!deferred_pfn_valid(pfn)) {
1776 deferred_free_range(pfn - nr_free, nr_free);
1778 } else if (!(pfn & nr_pgmask)) {
1779 deferred_free_range(pfn - nr_free, nr_free);
1785 /* Free the last block of pages to allocator */
1786 deferred_free_range(pfn - nr_free, nr_free);
1790 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1791 * by performing it only once every pageblock_nr_pages.
1792 * Return number of pages initialized.
1794 static unsigned long __init deferred_init_pages(struct zone *zone,
1796 unsigned long end_pfn)
1798 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1799 int nid = zone_to_nid(zone);
1800 unsigned long nr_pages = 0;
1801 int zid = zone_idx(zone);
1802 struct page *page = NULL;
1804 for (; pfn < end_pfn; pfn++) {
1805 if (!deferred_pfn_valid(pfn)) {
1808 } else if (!page || !(pfn & nr_pgmask)) {
1809 page = pfn_to_page(pfn);
1813 __init_single_page(page, pfn, zid, nid);
1820 * This function is meant to pre-load the iterator for the zone init.
1821 * Specifically it walks through the ranges until we are caught up to the
1822 * first_init_pfn value and exits there. If we never encounter the value we
1823 * return false indicating there are no valid ranges left.
1826 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1827 unsigned long *spfn, unsigned long *epfn,
1828 unsigned long first_init_pfn)
1833 * Start out by walking through the ranges in this zone that have
1834 * already been initialized. We don't need to do anything with them
1835 * so we just need to flush them out of the system.
1837 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1838 if (*epfn <= first_init_pfn)
1840 if (*spfn < first_init_pfn)
1841 *spfn = first_init_pfn;
1850 * Initialize and free pages. We do it in two loops: first we initialize
1851 * struct page, then free to buddy allocator, because while we are
1852 * freeing pages we can access pages that are ahead (computing buddy
1853 * page in __free_one_page()).
1855 * In order to try and keep some memory in the cache we have the loop
1856 * broken along max page order boundaries. This way we will not cause
1857 * any issues with the buddy page computation.
1859 static unsigned long __init
1860 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1861 unsigned long *end_pfn)
1863 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1864 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1865 unsigned long nr_pages = 0;
1868 /* First we loop through and initialize the page values */
1869 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1872 if (mo_pfn <= *start_pfn)
1875 t = min(mo_pfn, *end_pfn);
1876 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1878 if (mo_pfn < *end_pfn) {
1879 *start_pfn = mo_pfn;
1884 /* Reset values and now loop through freeing pages as needed */
1887 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1893 t = min(mo_pfn, epfn);
1894 deferred_free_pages(spfn, t);
1904 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1907 unsigned long spfn, epfn;
1908 struct zone *zone = arg;
1911 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1914 * Initialize and free pages in MAX_ORDER sized increments so that we
1915 * can avoid introducing any issues with the buddy allocator.
1917 while (spfn < end_pfn) {
1918 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1923 /* An arch may override for more concurrency. */
1925 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1930 /* Initialise remaining memory on a node */
1931 static int __init deferred_init_memmap(void *data)
1933 pg_data_t *pgdat = data;
1934 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1935 unsigned long spfn = 0, epfn = 0;
1936 unsigned long first_init_pfn, flags;
1937 unsigned long start = jiffies;
1939 int zid, max_threads;
1942 /* Bind memory initialisation thread to a local node if possible */
1943 if (!cpumask_empty(cpumask))
1944 set_cpus_allowed_ptr(current, cpumask);
1946 pgdat_resize_lock(pgdat, &flags);
1947 first_init_pfn = pgdat->first_deferred_pfn;
1948 if (first_init_pfn == ULONG_MAX) {
1949 pgdat_resize_unlock(pgdat, &flags);
1950 pgdat_init_report_one_done();
1954 /* Sanity check boundaries */
1955 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1956 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1957 pgdat->first_deferred_pfn = ULONG_MAX;
1960 * Once we unlock here, the zone cannot be grown anymore, thus if an
1961 * interrupt thread must allocate this early in boot, zone must be
1962 * pre-grown prior to start of deferred page initialization.
1964 pgdat_resize_unlock(pgdat, &flags);
1966 /* Only the highest zone is deferred so find it */
1967 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1968 zone = pgdat->node_zones + zid;
1969 if (first_init_pfn < zone_end_pfn(zone))
1973 /* If the zone is empty somebody else may have cleared out the zone */
1974 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1978 max_threads = deferred_page_init_max_threads(cpumask);
1980 while (spfn < epfn) {
1981 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1982 struct padata_mt_job job = {
1983 .thread_fn = deferred_init_memmap_chunk,
1986 .size = epfn_align - spfn,
1987 .align = PAGES_PER_SECTION,
1988 .min_chunk = PAGES_PER_SECTION,
1989 .max_threads = max_threads,
1992 padata_do_multithreaded(&job);
1993 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1997 /* Sanity check that the next zone really is unpopulated */
1998 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
2000 pr_info("node %d deferred pages initialised in %ums\n",
2001 pgdat->node_id, jiffies_to_msecs(jiffies - start));
2003 pgdat_init_report_one_done();
2008 * If this zone has deferred pages, try to grow it by initializing enough
2009 * deferred pages to satisfy the allocation specified by order, rounded up to
2010 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
2011 * of SECTION_SIZE bytes by initializing struct pages in increments of
2012 * PAGES_PER_SECTION * sizeof(struct page) bytes.
2014 * Return true when zone was grown, otherwise return false. We return true even
2015 * when we grow less than requested, to let the caller decide if there are
2016 * enough pages to satisfy the allocation.
2018 * Note: We use noinline because this function is needed only during boot, and
2019 * it is called from a __ref function _deferred_grow_zone. This way we are
2020 * making sure that it is not inlined into permanent text section.
2022 static noinline bool __init
2023 deferred_grow_zone(struct zone *zone, unsigned int order)
2025 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
2026 pg_data_t *pgdat = zone->zone_pgdat;
2027 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
2028 unsigned long spfn, epfn, flags;
2029 unsigned long nr_pages = 0;
2032 /* Only the last zone may have deferred pages */
2033 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
2036 pgdat_resize_lock(pgdat, &flags);
2039 * If someone grew this zone while we were waiting for spinlock, return
2040 * true, as there might be enough pages already.
2042 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
2043 pgdat_resize_unlock(pgdat, &flags);
2047 /* If the zone is empty somebody else may have cleared out the zone */
2048 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
2049 first_deferred_pfn)) {
2050 pgdat->first_deferred_pfn = ULONG_MAX;
2051 pgdat_resize_unlock(pgdat, &flags);
2052 /* Retry only once. */
2053 return first_deferred_pfn != ULONG_MAX;
2057 * Initialize and free pages in MAX_ORDER sized increments so
2058 * that we can avoid introducing any issues with the buddy
2061 while (spfn < epfn) {
2062 /* update our first deferred PFN for this section */
2063 first_deferred_pfn = spfn;
2065 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
2066 touch_nmi_watchdog();
2068 /* We should only stop along section boundaries */
2069 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
2072 /* If our quota has been met we can stop here */
2073 if (nr_pages >= nr_pages_needed)
2077 pgdat->first_deferred_pfn = spfn;
2078 pgdat_resize_unlock(pgdat, &flags);
2080 return nr_pages > 0;
2084 * deferred_grow_zone() is __init, but it is called from
2085 * get_page_from_freelist() during early boot until deferred_pages permanently
2086 * disables this call. This is why we have refdata wrapper to avoid warning,
2087 * and to ensure that the function body gets unloaded.
2090 _deferred_grow_zone(struct zone *zone, unsigned int order)
2092 return deferred_grow_zone(zone, order);
2095 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2097 void __init page_alloc_init_late(void)
2102 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2104 /* There will be num_node_state(N_MEMORY) threads */
2105 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2106 for_each_node_state(nid, N_MEMORY) {
2107 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2110 /* Block until all are initialised */
2111 wait_for_completion(&pgdat_init_all_done_comp);
2114 * The number of managed pages has changed due to the initialisation
2115 * so the pcpu batch and high limits needs to be updated or the limits
2116 * will be artificially small.
2118 for_each_populated_zone(zone)
2119 zone_pcp_update(zone);
2122 * We initialized the rest of the deferred pages. Permanently disable
2123 * on-demand struct page initialization.
2125 static_branch_disable(&deferred_pages);
2127 /* Reinit limits that are based on free pages after the kernel is up */
2128 files_maxfiles_init();
2133 /* Discard memblock private memory */
2136 for_each_node_state(nid, N_MEMORY)
2137 shuffle_free_memory(NODE_DATA(nid));
2139 for_each_populated_zone(zone)
2140 set_zone_contiguous(zone);
2144 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2145 void __init init_cma_reserved_pageblock(struct page *page)
2147 unsigned i = pageblock_nr_pages;
2148 struct page *p = page;
2151 __ClearPageReserved(p);
2152 set_page_count(p, 0);
2155 set_pageblock_migratetype(page, MIGRATE_CMA);
2157 if (pageblock_order >= MAX_ORDER) {
2158 i = pageblock_nr_pages;
2161 set_page_refcounted(p);
2162 __free_pages(p, MAX_ORDER - 1);
2163 p += MAX_ORDER_NR_PAGES;
2164 } while (i -= MAX_ORDER_NR_PAGES);
2166 set_page_refcounted(page);
2167 __free_pages(page, pageblock_order);
2170 adjust_managed_page_count(page, pageblock_nr_pages);
2175 * The order of subdivision here is critical for the IO subsystem.
2176 * Please do not alter this order without good reasons and regression
2177 * testing. Specifically, as large blocks of memory are subdivided,
2178 * the order in which smaller blocks are delivered depends on the order
2179 * they're subdivided in this function. This is the primary factor
2180 * influencing the order in which pages are delivered to the IO
2181 * subsystem according to empirical testing, and this is also justified
2182 * by considering the behavior of a buddy system containing a single
2183 * large block of memory acted on by a series of small allocations.
2184 * This behavior is a critical factor in sglist merging's success.
2188 static inline void expand(struct zone *zone, struct page *page,
2189 int low, int high, int migratetype)
2191 unsigned long size = 1 << high;
2193 while (high > low) {
2196 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2199 * Mark as guard pages (or page), that will allow to
2200 * merge back to allocator when buddy will be freed.
2201 * Corresponding page table entries will not be touched,
2202 * pages will stay not present in virtual address space
2204 if (set_page_guard(zone, &page[size], high, migratetype))
2207 add_to_free_list(&page[size], zone, high, migratetype);
2208 set_buddy_order(&page[size], high);
2212 static void check_new_page_bad(struct page *page)
2214 if (unlikely(page->flags & __PG_HWPOISON)) {
2215 /* Don't complain about hwpoisoned pages */
2216 page_mapcount_reset(page); /* remove PageBuddy */
2221 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2225 * This page is about to be returned from the page allocator
2227 static inline int check_new_page(struct page *page)
2229 if (likely(page_expected_state(page,
2230 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2233 check_new_page_bad(page);
2237 #ifdef CONFIG_DEBUG_VM
2239 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2240 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2241 * also checked when pcp lists are refilled from the free lists.
2243 static inline bool check_pcp_refill(struct page *page)
2245 if (debug_pagealloc_enabled_static())
2246 return check_new_page(page);
2251 static inline bool check_new_pcp(struct page *page)
2253 return check_new_page(page);
2257 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2258 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2259 * enabled, they are also checked when being allocated from the pcp lists.
2261 static inline bool check_pcp_refill(struct page *page)
2263 return check_new_page(page);
2265 static inline bool check_new_pcp(struct page *page)
2267 if (debug_pagealloc_enabled_static())
2268 return check_new_page(page);
2272 #endif /* CONFIG_DEBUG_VM */
2274 static bool check_new_pages(struct page *page, unsigned int order)
2277 for (i = 0; i < (1 << order); i++) {
2278 struct page *p = page + i;
2280 if (unlikely(check_new_page(p)))
2287 inline void post_alloc_hook(struct page *page, unsigned int order,
2290 set_page_private(page, 0);
2291 set_page_refcounted(page);
2293 arch_alloc_page(page, order);
2294 debug_pagealloc_map_pages(page, 1 << order);
2295 kasan_alloc_pages(page, order);
2296 kernel_unpoison_pages(page, 1 << order);
2297 set_page_owner(page, order, gfp_flags);
2299 if (!want_init_on_free() && want_init_on_alloc(gfp_flags))
2300 kernel_init_free_pages(page, 1 << order);
2303 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2304 unsigned int alloc_flags)
2306 post_alloc_hook(page, order, gfp_flags);
2308 if (order && (gfp_flags & __GFP_COMP))
2309 prep_compound_page(page, order);
2312 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2313 * allocate the page. The expectation is that the caller is taking
2314 * steps that will free more memory. The caller should avoid the page
2315 * being used for !PFMEMALLOC purposes.
2317 if (alloc_flags & ALLOC_NO_WATERMARKS)
2318 set_page_pfmemalloc(page);
2320 clear_page_pfmemalloc(page);
2324 * Go through the free lists for the given migratetype and remove
2325 * the smallest available page from the freelists
2327 static __always_inline
2328 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2331 unsigned int current_order;
2332 struct free_area *area;
2335 /* Find a page of the appropriate size in the preferred list */
2336 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2337 area = &(zone->free_area[current_order]);
2338 page = get_page_from_free_area(area, migratetype);
2341 del_page_from_free_list(page, zone, current_order);
2342 expand(zone, page, order, current_order, migratetype);
2343 set_pcppage_migratetype(page, migratetype);
2352 * This array describes the order lists are fallen back to when
2353 * the free lists for the desirable migrate type are depleted
2355 static int fallbacks[MIGRATE_TYPES][3] = {
2356 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2357 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2358 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2360 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2362 #ifdef CONFIG_MEMORY_ISOLATION
2363 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2368 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2371 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2374 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2375 unsigned int order) { return NULL; }
2379 * Move the free pages in a range to the freelist tail of the requested type.
2380 * Note that start_page and end_pages are not aligned on a pageblock
2381 * boundary. If alignment is required, use move_freepages_block()
2383 static int move_freepages(struct zone *zone,
2384 struct page *start_page, struct page *end_page,
2385 int migratetype, int *num_movable)
2389 int pages_moved = 0;
2391 for (page = start_page; page <= end_page;) {
2392 if (!pfn_valid_within(page_to_pfn(page))) {
2397 if (!PageBuddy(page)) {
2399 * We assume that pages that could be isolated for
2400 * migration are movable. But we don't actually try
2401 * isolating, as that would be expensive.
2404 (PageLRU(page) || __PageMovable(page)))
2411 /* Make sure we are not inadvertently changing nodes */
2412 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2413 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2415 order = buddy_order(page);
2416 move_to_free_list(page, zone, order, migratetype);
2418 pages_moved += 1 << order;
2424 int move_freepages_block(struct zone *zone, struct page *page,
2425 int migratetype, int *num_movable)
2427 unsigned long start_pfn, end_pfn;
2428 struct page *start_page, *end_page;
2433 start_pfn = page_to_pfn(page);
2434 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2435 start_page = pfn_to_page(start_pfn);
2436 end_page = start_page + pageblock_nr_pages - 1;
2437 end_pfn = start_pfn + pageblock_nr_pages - 1;
2439 /* Do not cross zone boundaries */
2440 if (!zone_spans_pfn(zone, start_pfn))
2442 if (!zone_spans_pfn(zone, end_pfn))
2445 return move_freepages(zone, start_page, end_page, migratetype,
2449 static void change_pageblock_range(struct page *pageblock_page,
2450 int start_order, int migratetype)
2452 int nr_pageblocks = 1 << (start_order - pageblock_order);
2454 while (nr_pageblocks--) {
2455 set_pageblock_migratetype(pageblock_page, migratetype);
2456 pageblock_page += pageblock_nr_pages;
2461 * When we are falling back to another migratetype during allocation, try to
2462 * steal extra free pages from the same pageblocks to satisfy further
2463 * allocations, instead of polluting multiple pageblocks.
2465 * If we are stealing a relatively large buddy page, it is likely there will
2466 * be more free pages in the pageblock, so try to steal them all. For
2467 * reclaimable and unmovable allocations, we steal regardless of page size,
2468 * as fragmentation caused by those allocations polluting movable pageblocks
2469 * is worse than movable allocations stealing from unmovable and reclaimable
2472 static bool can_steal_fallback(unsigned int order, int start_mt)
2475 * Leaving this order check is intended, although there is
2476 * relaxed order check in next check. The reason is that
2477 * we can actually steal whole pageblock if this condition met,
2478 * but, below check doesn't guarantee it and that is just heuristic
2479 * so could be changed anytime.
2481 if (order >= pageblock_order)
2484 if (order >= pageblock_order / 2 ||
2485 start_mt == MIGRATE_RECLAIMABLE ||
2486 start_mt == MIGRATE_UNMOVABLE ||
2487 page_group_by_mobility_disabled)
2493 static inline bool boost_watermark(struct zone *zone)
2495 unsigned long max_boost;
2497 if (!watermark_boost_factor)
2500 * Don't bother in zones that are unlikely to produce results.
2501 * On small machines, including kdump capture kernels running
2502 * in a small area, boosting the watermark can cause an out of
2503 * memory situation immediately.
2505 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2508 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2509 watermark_boost_factor, 10000);
2512 * high watermark may be uninitialised if fragmentation occurs
2513 * very early in boot so do not boost. We do not fall
2514 * through and boost by pageblock_nr_pages as failing
2515 * allocations that early means that reclaim is not going
2516 * to help and it may even be impossible to reclaim the
2517 * boosted watermark resulting in a hang.
2522 max_boost = max(pageblock_nr_pages, max_boost);
2524 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2531 * This function implements actual steal behaviour. If order is large enough,
2532 * we can steal whole pageblock. If not, we first move freepages in this
2533 * pageblock to our migratetype and determine how many already-allocated pages
2534 * are there in the pageblock with a compatible migratetype. If at least half
2535 * of pages are free or compatible, we can change migratetype of the pageblock
2536 * itself, so pages freed in the future will be put on the correct free list.
2538 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2539 unsigned int alloc_flags, int start_type, bool whole_block)
2541 unsigned int current_order = buddy_order(page);
2542 int free_pages, movable_pages, alike_pages;
2545 old_block_type = get_pageblock_migratetype(page);
2548 * This can happen due to races and we want to prevent broken
2549 * highatomic accounting.
2551 if (is_migrate_highatomic(old_block_type))
2554 /* Take ownership for orders >= pageblock_order */
2555 if (current_order >= pageblock_order) {
2556 change_pageblock_range(page, current_order, start_type);
2561 * Boost watermarks to increase reclaim pressure to reduce the
2562 * likelihood of future fallbacks. Wake kswapd now as the node
2563 * may be balanced overall and kswapd will not wake naturally.
2565 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2566 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2568 /* We are not allowed to try stealing from the whole block */
2572 free_pages = move_freepages_block(zone, page, start_type,
2575 * Determine how many pages are compatible with our allocation.
2576 * For movable allocation, it's the number of movable pages which
2577 * we just obtained. For other types it's a bit more tricky.
2579 if (start_type == MIGRATE_MOVABLE) {
2580 alike_pages = movable_pages;
2583 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2584 * to MOVABLE pageblock, consider all non-movable pages as
2585 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2586 * vice versa, be conservative since we can't distinguish the
2587 * exact migratetype of non-movable pages.
2589 if (old_block_type == MIGRATE_MOVABLE)
2590 alike_pages = pageblock_nr_pages
2591 - (free_pages + movable_pages);
2596 /* moving whole block can fail due to zone boundary conditions */
2601 * If a sufficient number of pages in the block are either free or of
2602 * comparable migratability as our allocation, claim the whole block.
2604 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2605 page_group_by_mobility_disabled)
2606 set_pageblock_migratetype(page, start_type);
2611 move_to_free_list(page, zone, current_order, start_type);
2615 * Check whether there is a suitable fallback freepage with requested order.
2616 * If only_stealable is true, this function returns fallback_mt only if
2617 * we can steal other freepages all together. This would help to reduce
2618 * fragmentation due to mixed migratetype pages in one pageblock.
2620 int find_suitable_fallback(struct free_area *area, unsigned int order,
2621 int migratetype, bool only_stealable, bool *can_steal)
2626 if (area->nr_free == 0)
2631 fallback_mt = fallbacks[migratetype][i];
2632 if (fallback_mt == MIGRATE_TYPES)
2635 if (free_area_empty(area, fallback_mt))
2638 if (can_steal_fallback(order, migratetype))
2641 if (!only_stealable)
2652 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2653 * there are no empty page blocks that contain a page with a suitable order
2655 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2656 unsigned int alloc_order)
2659 unsigned long max_managed, flags;
2662 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2663 * Check is race-prone but harmless.
2665 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2666 if (zone->nr_reserved_highatomic >= max_managed)
2669 spin_lock_irqsave(&zone->lock, flags);
2671 /* Recheck the nr_reserved_highatomic limit under the lock */
2672 if (zone->nr_reserved_highatomic >= max_managed)
2676 mt = get_pageblock_migratetype(page);
2677 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2678 && !is_migrate_cma(mt)) {
2679 zone->nr_reserved_highatomic += pageblock_nr_pages;
2680 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2681 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2685 spin_unlock_irqrestore(&zone->lock, flags);
2689 * Used when an allocation is about to fail under memory pressure. This
2690 * potentially hurts the reliability of high-order allocations when under
2691 * intense memory pressure but failed atomic allocations should be easier
2692 * to recover from than an OOM.
2694 * If @force is true, try to unreserve a pageblock even though highatomic
2695 * pageblock is exhausted.
2697 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2700 struct zonelist *zonelist = ac->zonelist;
2701 unsigned long flags;
2708 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2711 * Preserve at least one pageblock unless memory pressure
2714 if (!force && zone->nr_reserved_highatomic <=
2718 spin_lock_irqsave(&zone->lock, flags);
2719 for (order = 0; order < MAX_ORDER; order++) {
2720 struct free_area *area = &(zone->free_area[order]);
2722 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2727 * In page freeing path, migratetype change is racy so
2728 * we can counter several free pages in a pageblock
2729 * in this loop althoug we changed the pageblock type
2730 * from highatomic to ac->migratetype. So we should
2731 * adjust the count once.
2733 if (is_migrate_highatomic_page(page)) {
2735 * It should never happen but changes to
2736 * locking could inadvertently allow a per-cpu
2737 * drain to add pages to MIGRATE_HIGHATOMIC
2738 * while unreserving so be safe and watch for
2741 zone->nr_reserved_highatomic -= min(
2743 zone->nr_reserved_highatomic);
2747 * Convert to ac->migratetype and avoid the normal
2748 * pageblock stealing heuristics. Minimally, the caller
2749 * is doing the work and needs the pages. More
2750 * importantly, if the block was always converted to
2751 * MIGRATE_UNMOVABLE or another type then the number
2752 * of pageblocks that cannot be completely freed
2755 set_pageblock_migratetype(page, ac->migratetype);
2756 ret = move_freepages_block(zone, page, ac->migratetype,
2759 spin_unlock_irqrestore(&zone->lock, flags);
2763 spin_unlock_irqrestore(&zone->lock, flags);
2770 * Try finding a free buddy page on the fallback list and put it on the free
2771 * list of requested migratetype, possibly along with other pages from the same
2772 * block, depending on fragmentation avoidance heuristics. Returns true if
2773 * fallback was found so that __rmqueue_smallest() can grab it.
2775 * The use of signed ints for order and current_order is a deliberate
2776 * deviation from the rest of this file, to make the for loop
2777 * condition simpler.
2779 static __always_inline bool
2780 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2781 unsigned int alloc_flags)
2783 struct free_area *area;
2785 int min_order = order;
2791 * Do not steal pages from freelists belonging to other pageblocks
2792 * i.e. orders < pageblock_order. If there are no local zones free,
2793 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2795 if (alloc_flags & ALLOC_NOFRAGMENT)
2796 min_order = pageblock_order;
2799 * Find the largest available free page in the other list. This roughly
2800 * approximates finding the pageblock with the most free pages, which
2801 * would be too costly to do exactly.
2803 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2805 area = &(zone->free_area[current_order]);
2806 fallback_mt = find_suitable_fallback(area, current_order,
2807 start_migratetype, false, &can_steal);
2808 if (fallback_mt == -1)
2812 * We cannot steal all free pages from the pageblock and the
2813 * requested migratetype is movable. In that case it's better to
2814 * steal and split the smallest available page instead of the
2815 * largest available page, because even if the next movable
2816 * allocation falls back into a different pageblock than this
2817 * one, it won't cause permanent fragmentation.
2819 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2820 && current_order > order)
2829 for (current_order = order; current_order < MAX_ORDER;
2831 area = &(zone->free_area[current_order]);
2832 fallback_mt = find_suitable_fallback(area, current_order,
2833 start_migratetype, false, &can_steal);
2834 if (fallback_mt != -1)
2839 * This should not happen - we already found a suitable fallback
2840 * when looking for the largest page.
2842 VM_BUG_ON(current_order == MAX_ORDER);
2845 page = get_page_from_free_area(area, fallback_mt);
2847 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2850 trace_mm_page_alloc_extfrag(page, order, current_order,
2851 start_migratetype, fallback_mt);
2858 * Do the hard work of removing an element from the buddy allocator.
2859 * Call me with the zone->lock already held.
2861 static __always_inline struct page *
2862 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2863 unsigned int alloc_flags)
2867 if (IS_ENABLED(CONFIG_CMA)) {
2869 * Balance movable allocations between regular and CMA areas by
2870 * allocating from CMA when over half of the zone's free memory
2871 * is in the CMA area.
2873 if (alloc_flags & ALLOC_CMA &&
2874 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2875 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2876 page = __rmqueue_cma_fallback(zone, order);
2882 page = __rmqueue_smallest(zone, order, migratetype);
2883 if (unlikely(!page)) {
2884 if (alloc_flags & ALLOC_CMA)
2885 page = __rmqueue_cma_fallback(zone, order);
2887 if (!page && __rmqueue_fallback(zone, order, migratetype,
2893 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2898 * Obtain a specified number of elements from the buddy allocator, all under
2899 * a single hold of the lock, for efficiency. Add them to the supplied list.
2900 * Returns the number of new pages which were placed at *list.
2902 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2903 unsigned long count, struct list_head *list,
2904 int migratetype, unsigned int alloc_flags)
2908 spin_lock(&zone->lock);
2909 for (i = 0; i < count; ++i) {
2910 struct page *page = __rmqueue(zone, order, migratetype,
2912 if (unlikely(page == NULL))
2915 if (unlikely(check_pcp_refill(page)))
2919 * Split buddy pages returned by expand() are received here in
2920 * physical page order. The page is added to the tail of
2921 * caller's list. From the callers perspective, the linked list
2922 * is ordered by page number under some conditions. This is
2923 * useful for IO devices that can forward direction from the
2924 * head, thus also in the physical page order. This is useful
2925 * for IO devices that can merge IO requests if the physical
2926 * pages are ordered properly.
2928 list_add_tail(&page->lru, list);
2930 if (is_migrate_cma(get_pcppage_migratetype(page)))
2931 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2936 * i pages were removed from the buddy list even if some leak due
2937 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2938 * on i. Do not confuse with 'alloced' which is the number of
2939 * pages added to the pcp list.
2941 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2942 spin_unlock(&zone->lock);
2948 * Called from the vmstat counter updater to drain pagesets of this
2949 * currently executing processor on remote nodes after they have
2952 * Note that this function must be called with the thread pinned to
2953 * a single processor.
2955 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2957 unsigned long flags;
2958 int to_drain, batch;
2960 local_irq_save(flags);
2961 batch = READ_ONCE(pcp->batch);
2962 to_drain = min(pcp->count, batch);
2964 free_pcppages_bulk(zone, to_drain, pcp);
2965 local_irq_restore(flags);
2970 * Drain pcplists of the indicated processor and zone.
2972 * The processor must either be the current processor and the
2973 * thread pinned to the current processor or a processor that
2976 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2978 unsigned long flags;
2979 struct per_cpu_pageset *pset;
2980 struct per_cpu_pages *pcp;
2982 local_irq_save(flags);
2983 pset = per_cpu_ptr(zone->pageset, cpu);
2987 free_pcppages_bulk(zone, pcp->count, pcp);
2988 local_irq_restore(flags);
2992 * Drain pcplists of all zones on the indicated processor.
2994 * The processor must either be the current processor and the
2995 * thread pinned to the current processor or a processor that
2998 static void drain_pages(unsigned int cpu)
3002 for_each_populated_zone(zone) {
3003 drain_pages_zone(cpu, zone);
3008 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
3010 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
3011 * the single zone's pages.
3013 void drain_local_pages(struct zone *zone)
3015 int cpu = smp_processor_id();
3018 drain_pages_zone(cpu, zone);
3023 static void drain_local_pages_wq(struct work_struct *work)
3025 struct pcpu_drain *drain;
3027 drain = container_of(work, struct pcpu_drain, work);
3030 * drain_all_pages doesn't use proper cpu hotplug protection so
3031 * we can race with cpu offline when the WQ can move this from
3032 * a cpu pinned worker to an unbound one. We can operate on a different
3033 * cpu which is allright but we also have to make sure to not move to
3037 drain_local_pages(drain->zone);
3042 * The implementation of drain_all_pages(), exposing an extra parameter to
3043 * drain on all cpus.
3045 * drain_all_pages() is optimized to only execute on cpus where pcplists are
3046 * not empty. The check for non-emptiness can however race with a free to
3047 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
3048 * that need the guarantee that every CPU has drained can disable the
3049 * optimizing racy check.
3051 static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
3056 * Allocate in the BSS so we wont require allocation in
3057 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
3059 static cpumask_t cpus_with_pcps;
3062 * Make sure nobody triggers this path before mm_percpu_wq is fully
3065 if (WARN_ON_ONCE(!mm_percpu_wq))
3069 * Do not drain if one is already in progress unless it's specific to
3070 * a zone. Such callers are primarily CMA and memory hotplug and need
3071 * the drain to be complete when the call returns.
3073 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
3076 mutex_lock(&pcpu_drain_mutex);
3080 * We don't care about racing with CPU hotplug event
3081 * as offline notification will cause the notified
3082 * cpu to drain that CPU pcps and on_each_cpu_mask
3083 * disables preemption as part of its processing
3085 for_each_online_cpu(cpu) {
3086 struct per_cpu_pageset *pcp;
3088 bool has_pcps = false;
3090 if (force_all_cpus) {
3092 * The pcp.count check is racy, some callers need a
3093 * guarantee that no cpu is missed.
3097 pcp = per_cpu_ptr(zone->pageset, cpu);
3101 for_each_populated_zone(z) {
3102 pcp = per_cpu_ptr(z->pageset, cpu);
3103 if (pcp->pcp.count) {
3111 cpumask_set_cpu(cpu, &cpus_with_pcps);
3113 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3116 for_each_cpu(cpu, &cpus_with_pcps) {
3117 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3120 INIT_WORK(&drain->work, drain_local_pages_wq);
3121 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3123 for_each_cpu(cpu, &cpus_with_pcps)
3124 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3126 mutex_unlock(&pcpu_drain_mutex);
3130 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
3132 * When zone parameter is non-NULL, spill just the single zone's pages.
3134 * Note that this can be extremely slow as the draining happens in a workqueue.
3136 void drain_all_pages(struct zone *zone)
3138 __drain_all_pages(zone, false);
3141 #ifdef CONFIG_HIBERNATION
3144 * Touch the watchdog for every WD_PAGE_COUNT pages.
3146 #define WD_PAGE_COUNT (128*1024)
3148 void mark_free_pages(struct zone *zone)
3150 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3151 unsigned long flags;
3152 unsigned int order, t;
3155 if (zone_is_empty(zone))
3158 spin_lock_irqsave(&zone->lock, flags);
3160 max_zone_pfn = zone_end_pfn(zone);
3161 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3162 if (pfn_valid(pfn)) {
3163 page = pfn_to_page(pfn);
3165 if (!--page_count) {
3166 touch_nmi_watchdog();
3167 page_count = WD_PAGE_COUNT;
3170 if (page_zone(page) != zone)
3173 if (!swsusp_page_is_forbidden(page))
3174 swsusp_unset_page_free(page);
3177 for_each_migratetype_order(order, t) {
3178 list_for_each_entry(page,
3179 &zone->free_area[order].free_list[t], lru) {
3182 pfn = page_to_pfn(page);
3183 for (i = 0; i < (1UL << order); i++) {
3184 if (!--page_count) {
3185 touch_nmi_watchdog();
3186 page_count = WD_PAGE_COUNT;
3188 swsusp_set_page_free(pfn_to_page(pfn + i));
3192 spin_unlock_irqrestore(&zone->lock, flags);
3194 #endif /* CONFIG_PM */
3196 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3200 if (!free_pcp_prepare(page))
3203 migratetype = get_pfnblock_migratetype(page, pfn);
3204 set_pcppage_migratetype(page, migratetype);
3208 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3210 struct zone *zone = page_zone(page);
3211 struct per_cpu_pages *pcp;
3214 migratetype = get_pcppage_migratetype(page);
3215 __count_vm_event(PGFREE);
3218 * We only track unmovable, reclaimable and movable on pcp lists.
3219 * Free ISOLATE pages back to the allocator because they are being
3220 * offlined but treat HIGHATOMIC as movable pages so we can get those
3221 * areas back if necessary. Otherwise, we may have to free
3222 * excessively into the page allocator
3224 if (migratetype >= MIGRATE_PCPTYPES) {
3225 if (unlikely(is_migrate_isolate(migratetype))) {
3226 free_one_page(zone, page, pfn, 0, migratetype,
3230 migratetype = MIGRATE_MOVABLE;
3233 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3234 list_add(&page->lru, &pcp->lists[migratetype]);
3236 if (pcp->count >= READ_ONCE(pcp->high))
3237 free_pcppages_bulk(zone, READ_ONCE(pcp->batch), pcp);
3241 * Free a 0-order page
3243 void free_unref_page(struct page *page)
3245 unsigned long flags;
3246 unsigned long pfn = page_to_pfn(page);
3248 if (!free_unref_page_prepare(page, pfn))
3251 local_irq_save(flags);
3252 free_unref_page_commit(page, pfn);
3253 local_irq_restore(flags);
3257 * Free a list of 0-order pages
3259 void free_unref_page_list(struct list_head *list)
3261 struct page *page, *next;
3262 unsigned long flags, pfn;
3263 int batch_count = 0;
3265 /* Prepare pages for freeing */
3266 list_for_each_entry_safe(page, next, list, lru) {
3267 pfn = page_to_pfn(page);
3268 if (!free_unref_page_prepare(page, pfn))
3269 list_del(&page->lru);
3270 set_page_private(page, pfn);
3273 local_irq_save(flags);
3274 list_for_each_entry_safe(page, next, list, lru) {
3275 unsigned long pfn = page_private(page);
3277 set_page_private(page, 0);
3278 trace_mm_page_free_batched(page);
3279 free_unref_page_commit(page, pfn);
3282 * Guard against excessive IRQ disabled times when we get
3283 * a large list of pages to free.
3285 if (++batch_count == SWAP_CLUSTER_MAX) {
3286 local_irq_restore(flags);
3288 local_irq_save(flags);
3291 local_irq_restore(flags);
3295 * split_page takes a non-compound higher-order page, and splits it into
3296 * n (1<<order) sub-pages: page[0..n]
3297 * Each sub-page must be freed individually.
3299 * Note: this is probably too low level an operation for use in drivers.
3300 * Please consult with lkml before using this in your driver.
3302 void split_page(struct page *page, unsigned int order)
3306 VM_BUG_ON_PAGE(PageCompound(page), page);
3307 VM_BUG_ON_PAGE(!page_count(page), page);
3309 for (i = 1; i < (1 << order); i++)
3310 set_page_refcounted(page + i);
3311 split_page_owner(page, 1 << order);
3313 EXPORT_SYMBOL_GPL(split_page);
3315 int __isolate_free_page(struct page *page, unsigned int order)
3317 unsigned long watermark;
3321 BUG_ON(!PageBuddy(page));
3323 zone = page_zone(page);
3324 mt = get_pageblock_migratetype(page);
3326 if (!is_migrate_isolate(mt)) {
3328 * Obey watermarks as if the page was being allocated. We can
3329 * emulate a high-order watermark check with a raised order-0
3330 * watermark, because we already know our high-order page
3333 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3334 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3337 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3340 /* Remove page from free list */
3342 del_page_from_free_list(page, zone, order);
3345 * Set the pageblock if the isolated page is at least half of a
3348 if (order >= pageblock_order - 1) {
3349 struct page *endpage = page + (1 << order) - 1;
3350 for (; page < endpage; page += pageblock_nr_pages) {
3351 int mt = get_pageblock_migratetype(page);
3352 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3353 && !is_migrate_highatomic(mt))
3354 set_pageblock_migratetype(page,
3360 return 1UL << order;
3364 * __putback_isolated_page - Return a now-isolated page back where we got it
3365 * @page: Page that was isolated
3366 * @order: Order of the isolated page
3367 * @mt: The page's pageblock's migratetype
3369 * This function is meant to return a page pulled from the free lists via
3370 * __isolate_free_page back to the free lists they were pulled from.
3372 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3374 struct zone *zone = page_zone(page);
3376 /* zone lock should be held when this function is called */
3377 lockdep_assert_held(&zone->lock);
3379 /* Return isolated page to tail of freelist. */
3380 __free_one_page(page, page_to_pfn(page), zone, order, mt,
3381 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
3385 * Update NUMA hit/miss statistics
3387 * Must be called with interrupts disabled.
3389 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3392 enum numa_stat_item local_stat = NUMA_LOCAL;
3394 /* skip numa counters update if numa stats is disabled */
3395 if (!static_branch_likely(&vm_numa_stat_key))
3398 if (zone_to_nid(z) != numa_node_id())
3399 local_stat = NUMA_OTHER;
3401 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3402 __inc_numa_state(z, NUMA_HIT);
3404 __inc_numa_state(z, NUMA_MISS);
3405 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3407 __inc_numa_state(z, local_stat);
3411 /* Remove page from the per-cpu list, caller must protect the list */
3412 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3413 unsigned int alloc_flags,
3414 struct per_cpu_pages *pcp,
3415 struct list_head *list)
3420 if (list_empty(list)) {
3421 pcp->count += rmqueue_bulk(zone, 0,
3422 READ_ONCE(pcp->batch), list,
3423 migratetype, alloc_flags);
3424 if (unlikely(list_empty(list)))
3428 page = list_first_entry(list, struct page, lru);
3429 list_del(&page->lru);
3431 } while (check_new_pcp(page));
3436 /* Lock and remove page from the per-cpu list */
3437 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3438 struct zone *zone, gfp_t gfp_flags,
3439 int migratetype, unsigned int alloc_flags)
3441 struct per_cpu_pages *pcp;
3442 struct list_head *list;
3444 unsigned long flags;
3446 local_irq_save(flags);
3447 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3448 list = &pcp->lists[migratetype];
3449 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3451 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3452 zone_statistics(preferred_zone, zone);
3454 local_irq_restore(flags);
3459 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3462 struct page *rmqueue(struct zone *preferred_zone,
3463 struct zone *zone, unsigned int order,
3464 gfp_t gfp_flags, unsigned int alloc_flags,
3467 unsigned long flags;
3470 if (likely(order == 0)) {
3472 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3473 * we need to skip it when CMA area isn't allowed.
3475 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3476 migratetype != MIGRATE_MOVABLE) {
3477 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3478 migratetype, alloc_flags);
3484 * We most definitely don't want callers attempting to
3485 * allocate greater than order-1 page units with __GFP_NOFAIL.
3487 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3488 spin_lock_irqsave(&zone->lock, flags);
3493 * order-0 request can reach here when the pcplist is skipped
3494 * due to non-CMA allocation context. HIGHATOMIC area is
3495 * reserved for high-order atomic allocation, so order-0
3496 * request should skip it.
3498 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3499 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3501 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3504 page = __rmqueue(zone, order, migratetype, alloc_flags);
3505 } while (page && check_new_pages(page, order));
3506 spin_unlock(&zone->lock);
3509 __mod_zone_freepage_state(zone, -(1 << order),
3510 get_pcppage_migratetype(page));
3512 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3513 zone_statistics(preferred_zone, zone);
3514 local_irq_restore(flags);
3517 /* Separate test+clear to avoid unnecessary atomics */
3518 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3519 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3520 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3523 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3527 local_irq_restore(flags);
3531 #ifdef CONFIG_FAIL_PAGE_ALLOC
3534 struct fault_attr attr;
3536 bool ignore_gfp_highmem;
3537 bool ignore_gfp_reclaim;
3539 } fail_page_alloc = {
3540 .attr = FAULT_ATTR_INITIALIZER,
3541 .ignore_gfp_reclaim = true,
3542 .ignore_gfp_highmem = true,
3546 static int __init setup_fail_page_alloc(char *str)
3548 return setup_fault_attr(&fail_page_alloc.attr, str);
3550 __setup("fail_page_alloc=", setup_fail_page_alloc);
3552 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3554 if (order < fail_page_alloc.min_order)
3556 if (gfp_mask & __GFP_NOFAIL)
3558 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3560 if (fail_page_alloc.ignore_gfp_reclaim &&
3561 (gfp_mask & __GFP_DIRECT_RECLAIM))
3564 return should_fail(&fail_page_alloc.attr, 1 << order);
3567 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3569 static int __init fail_page_alloc_debugfs(void)
3571 umode_t mode = S_IFREG | 0600;
3574 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3575 &fail_page_alloc.attr);
3577 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3578 &fail_page_alloc.ignore_gfp_reclaim);
3579 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3580 &fail_page_alloc.ignore_gfp_highmem);
3581 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3586 late_initcall(fail_page_alloc_debugfs);
3588 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3590 #else /* CONFIG_FAIL_PAGE_ALLOC */
3592 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3597 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3599 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3601 return __should_fail_alloc_page(gfp_mask, order);
3603 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3605 static inline long __zone_watermark_unusable_free(struct zone *z,
3606 unsigned int order, unsigned int alloc_flags)
3608 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3609 long unusable_free = (1 << order) - 1;
3612 * If the caller does not have rights to ALLOC_HARDER then subtract
3613 * the high-atomic reserves. This will over-estimate the size of the
3614 * atomic reserve but it avoids a search.
3616 if (likely(!alloc_harder))
3617 unusable_free += z->nr_reserved_highatomic;
3620 /* If allocation can't use CMA areas don't use free CMA pages */
3621 if (!(alloc_flags & ALLOC_CMA))
3622 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3625 return unusable_free;
3629 * Return true if free base pages are above 'mark'. For high-order checks it
3630 * will return true of the order-0 watermark is reached and there is at least
3631 * one free page of a suitable size. Checking now avoids taking the zone lock
3632 * to check in the allocation paths if no pages are free.
3634 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3635 int highest_zoneidx, unsigned int alloc_flags,
3640 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3642 /* free_pages may go negative - that's OK */
3643 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3645 if (alloc_flags & ALLOC_HIGH)
3648 if (unlikely(alloc_harder)) {
3650 * OOM victims can try even harder than normal ALLOC_HARDER
3651 * users on the grounds that it's definitely going to be in
3652 * the exit path shortly and free memory. Any allocation it
3653 * makes during the free path will be small and short-lived.
3655 if (alloc_flags & ALLOC_OOM)
3662 * Check watermarks for an order-0 allocation request. If these
3663 * are not met, then a high-order request also cannot go ahead
3664 * even if a suitable page happened to be free.
3666 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3669 /* If this is an order-0 request then the watermark is fine */
3673 /* For a high-order request, check at least one suitable page is free */
3674 for (o = order; o < MAX_ORDER; o++) {
3675 struct free_area *area = &z->free_area[o];
3681 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3682 if (!free_area_empty(area, mt))
3687 if ((alloc_flags & ALLOC_CMA) &&
3688 !free_area_empty(area, MIGRATE_CMA)) {
3692 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3698 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3699 int highest_zoneidx, unsigned int alloc_flags)
3701 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3702 zone_page_state(z, NR_FREE_PAGES));
3705 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3706 unsigned long mark, int highest_zoneidx,
3707 unsigned int alloc_flags, gfp_t gfp_mask)
3711 free_pages = zone_page_state(z, NR_FREE_PAGES);
3714 * Fast check for order-0 only. If this fails then the reserves
3715 * need to be calculated.
3720 fast_free = free_pages;
3721 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3722 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3726 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3730 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3731 * when checking the min watermark. The min watermark is the
3732 * point where boosting is ignored so that kswapd is woken up
3733 * when below the low watermark.
3735 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3736 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3737 mark = z->_watermark[WMARK_MIN];
3738 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3739 alloc_flags, free_pages);
3745 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3746 unsigned long mark, int highest_zoneidx)
3748 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3750 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3751 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3753 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3758 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3760 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3761 node_reclaim_distance;
3763 #else /* CONFIG_NUMA */
3764 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3768 #endif /* CONFIG_NUMA */
3771 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3772 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3773 * premature use of a lower zone may cause lowmem pressure problems that
3774 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3775 * probably too small. It only makes sense to spread allocations to avoid
3776 * fragmentation between the Normal and DMA32 zones.
3778 static inline unsigned int
3779 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3781 unsigned int alloc_flags;
3784 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3787 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3789 #ifdef CONFIG_ZONE_DMA32
3793 if (zone_idx(zone) != ZONE_NORMAL)
3797 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3798 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3799 * on UMA that if Normal is populated then so is DMA32.
3801 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3802 if (nr_online_nodes > 1 && !populated_zone(--zone))
3805 alloc_flags |= ALLOC_NOFRAGMENT;
3806 #endif /* CONFIG_ZONE_DMA32 */
3810 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3811 unsigned int alloc_flags)
3814 unsigned int pflags = current->flags;
3816 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3817 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3818 alloc_flags |= ALLOC_CMA;
3825 * get_page_from_freelist goes through the zonelist trying to allocate
3828 static struct page *
3829 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3830 const struct alloc_context *ac)
3834 struct pglist_data *last_pgdat_dirty_limit = NULL;
3839 * Scan zonelist, looking for a zone with enough free.
3840 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3842 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3843 z = ac->preferred_zoneref;
3844 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3849 if (cpusets_enabled() &&
3850 (alloc_flags & ALLOC_CPUSET) &&
3851 !__cpuset_zone_allowed(zone, gfp_mask))
3854 * When allocating a page cache page for writing, we
3855 * want to get it from a node that is within its dirty
3856 * limit, such that no single node holds more than its
3857 * proportional share of globally allowed dirty pages.
3858 * The dirty limits take into account the node's
3859 * lowmem reserves and high watermark so that kswapd
3860 * should be able to balance it without having to
3861 * write pages from its LRU list.
3863 * XXX: For now, allow allocations to potentially
3864 * exceed the per-node dirty limit in the slowpath
3865 * (spread_dirty_pages unset) before going into reclaim,
3866 * which is important when on a NUMA setup the allowed
3867 * nodes are together not big enough to reach the
3868 * global limit. The proper fix for these situations
3869 * will require awareness of nodes in the
3870 * dirty-throttling and the flusher threads.
3872 if (ac->spread_dirty_pages) {
3873 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3876 if (!node_dirty_ok(zone->zone_pgdat)) {
3877 last_pgdat_dirty_limit = zone->zone_pgdat;
3882 if (no_fallback && nr_online_nodes > 1 &&
3883 zone != ac->preferred_zoneref->zone) {
3887 * If moving to a remote node, retry but allow
3888 * fragmenting fallbacks. Locality is more important
3889 * than fragmentation avoidance.
3891 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3892 if (zone_to_nid(zone) != local_nid) {
3893 alloc_flags &= ~ALLOC_NOFRAGMENT;
3898 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3899 if (!zone_watermark_fast(zone, order, mark,
3900 ac->highest_zoneidx, alloc_flags,
3904 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3906 * Watermark failed for this zone, but see if we can
3907 * grow this zone if it contains deferred pages.
3909 if (static_branch_unlikely(&deferred_pages)) {
3910 if (_deferred_grow_zone(zone, order))
3914 /* Checked here to keep the fast path fast */
3915 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3916 if (alloc_flags & ALLOC_NO_WATERMARKS)
3919 if (node_reclaim_mode == 0 ||
3920 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3923 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3925 case NODE_RECLAIM_NOSCAN:
3928 case NODE_RECLAIM_FULL:
3929 /* scanned but unreclaimable */
3932 /* did we reclaim enough */
3933 if (zone_watermark_ok(zone, order, mark,
3934 ac->highest_zoneidx, alloc_flags))
3942 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3943 gfp_mask, alloc_flags, ac->migratetype);
3945 prep_new_page(page, order, gfp_mask, alloc_flags);
3948 * If this is a high-order atomic allocation then check
3949 * if the pageblock should be reserved for the future
3951 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3952 reserve_highatomic_pageblock(page, zone, order);
3956 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3957 /* Try again if zone has deferred pages */
3958 if (static_branch_unlikely(&deferred_pages)) {
3959 if (_deferred_grow_zone(zone, order))
3967 * It's possible on a UMA machine to get through all zones that are
3968 * fragmented. If avoiding fragmentation, reset and try again.
3971 alloc_flags &= ~ALLOC_NOFRAGMENT;
3978 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3980 unsigned int filter = SHOW_MEM_FILTER_NODES;
3983 * This documents exceptions given to allocations in certain
3984 * contexts that are allowed to allocate outside current's set
3987 if (!(gfp_mask & __GFP_NOMEMALLOC))
3988 if (tsk_is_oom_victim(current) ||
3989 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3990 filter &= ~SHOW_MEM_FILTER_NODES;
3991 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3992 filter &= ~SHOW_MEM_FILTER_NODES;
3994 show_mem(filter, nodemask);
3997 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3999 struct va_format vaf;
4001 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
4003 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
4006 va_start(args, fmt);
4009 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
4010 current->comm, &vaf, gfp_mask, &gfp_mask,
4011 nodemask_pr_args(nodemask));
4014 cpuset_print_current_mems_allowed();
4017 warn_alloc_show_mem(gfp_mask, nodemask);
4020 static inline struct page *
4021 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
4022 unsigned int alloc_flags,
4023 const struct alloc_context *ac)
4027 page = get_page_from_freelist(gfp_mask, order,
4028 alloc_flags|ALLOC_CPUSET, ac);
4030 * fallback to ignore cpuset restriction if our nodes
4034 page = get_page_from_freelist(gfp_mask, order,
4040 static inline struct page *
4041 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
4042 const struct alloc_context *ac, unsigned long *did_some_progress)
4044 struct oom_control oc = {
4045 .zonelist = ac->zonelist,
4046 .nodemask = ac->nodemask,
4048 .gfp_mask = gfp_mask,
4053 *did_some_progress = 0;
4056 * Acquire the oom lock. If that fails, somebody else is
4057 * making progress for us.
4059 if (!mutex_trylock(&oom_lock)) {
4060 *did_some_progress = 1;
4061 schedule_timeout_uninterruptible(1);
4066 * Go through the zonelist yet one more time, keep very high watermark
4067 * here, this is only to catch a parallel oom killing, we must fail if
4068 * we're still under heavy pressure. But make sure that this reclaim
4069 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
4070 * allocation which will never fail due to oom_lock already held.
4072 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
4073 ~__GFP_DIRECT_RECLAIM, order,
4074 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
4078 /* Coredumps can quickly deplete all memory reserves */
4079 if (current->flags & PF_DUMPCORE)
4081 /* The OOM killer will not help higher order allocs */
4082 if (order > PAGE_ALLOC_COSTLY_ORDER)
4085 * We have already exhausted all our reclaim opportunities without any
4086 * success so it is time to admit defeat. We will skip the OOM killer
4087 * because it is very likely that the caller has a more reasonable
4088 * fallback than shooting a random task.
4090 * The OOM killer may not free memory on a specific node.
4092 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
4094 /* The OOM killer does not needlessly kill tasks for lowmem */
4095 if (ac->highest_zoneidx < ZONE_NORMAL)
4097 if (pm_suspended_storage())
4100 * XXX: GFP_NOFS allocations should rather fail than rely on
4101 * other request to make a forward progress.
4102 * We are in an unfortunate situation where out_of_memory cannot
4103 * do much for this context but let's try it to at least get
4104 * access to memory reserved if the current task is killed (see
4105 * out_of_memory). Once filesystems are ready to handle allocation
4106 * failures more gracefully we should just bail out here.
4109 /* Exhausted what can be done so it's blame time */
4110 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4111 *did_some_progress = 1;
4114 * Help non-failing allocations by giving them access to memory
4117 if (gfp_mask & __GFP_NOFAIL)
4118 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4119 ALLOC_NO_WATERMARKS, ac);
4122 mutex_unlock(&oom_lock);
4127 * Maximum number of compaction retries wit a progress before OOM
4128 * killer is consider as the only way to move forward.
4130 #define MAX_COMPACT_RETRIES 16
4132 #ifdef CONFIG_COMPACTION
4133 /* Try memory compaction for high-order allocations before reclaim */
4134 static struct page *
4135 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4136 unsigned int alloc_flags, const struct alloc_context *ac,
4137 enum compact_priority prio, enum compact_result *compact_result)
4139 struct page *page = NULL;
4140 unsigned long pflags;
4141 unsigned int noreclaim_flag;
4146 psi_memstall_enter(&pflags);
4147 noreclaim_flag = memalloc_noreclaim_save();
4149 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4152 memalloc_noreclaim_restore(noreclaim_flag);
4153 psi_memstall_leave(&pflags);
4156 * At least in one zone compaction wasn't deferred or skipped, so let's
4157 * count a compaction stall
4159 count_vm_event(COMPACTSTALL);
4161 /* Prep a captured page if available */
4163 prep_new_page(page, order, gfp_mask, alloc_flags);
4165 /* Try get a page from the freelist if available */
4167 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4170 struct zone *zone = page_zone(page);
4172 zone->compact_blockskip_flush = false;
4173 compaction_defer_reset(zone, order, true);
4174 count_vm_event(COMPACTSUCCESS);
4179 * It's bad if compaction run occurs and fails. The most likely reason
4180 * is that pages exist, but not enough to satisfy watermarks.
4182 count_vm_event(COMPACTFAIL);
4190 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4191 enum compact_result compact_result,
4192 enum compact_priority *compact_priority,
4193 int *compaction_retries)
4195 int max_retries = MAX_COMPACT_RETRIES;
4198 int retries = *compaction_retries;
4199 enum compact_priority priority = *compact_priority;
4204 if (compaction_made_progress(compact_result))
4205 (*compaction_retries)++;
4208 * compaction considers all the zone as desperately out of memory
4209 * so it doesn't really make much sense to retry except when the
4210 * failure could be caused by insufficient priority
4212 if (compaction_failed(compact_result))
4213 goto check_priority;
4216 * compaction was skipped because there are not enough order-0 pages
4217 * to work with, so we retry only if it looks like reclaim can help.
4219 if (compaction_needs_reclaim(compact_result)) {
4220 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4225 * make sure the compaction wasn't deferred or didn't bail out early
4226 * due to locks contention before we declare that we should give up.
4227 * But the next retry should use a higher priority if allowed, so
4228 * we don't just keep bailing out endlessly.
4230 if (compaction_withdrawn(compact_result)) {
4231 goto check_priority;
4235 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4236 * costly ones because they are de facto nofail and invoke OOM
4237 * killer to move on while costly can fail and users are ready
4238 * to cope with that. 1/4 retries is rather arbitrary but we
4239 * would need much more detailed feedback from compaction to
4240 * make a better decision.
4242 if (order > PAGE_ALLOC_COSTLY_ORDER)
4244 if (*compaction_retries <= max_retries) {
4250 * Make sure there are attempts at the highest priority if we exhausted
4251 * all retries or failed at the lower priorities.
4254 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4255 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4257 if (*compact_priority > min_priority) {
4258 (*compact_priority)--;
4259 *compaction_retries = 0;
4263 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4267 static inline struct page *
4268 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4269 unsigned int alloc_flags, const struct alloc_context *ac,
4270 enum compact_priority prio, enum compact_result *compact_result)
4272 *compact_result = COMPACT_SKIPPED;
4277 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4278 enum compact_result compact_result,
4279 enum compact_priority *compact_priority,
4280 int *compaction_retries)
4285 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4289 * There are setups with compaction disabled which would prefer to loop
4290 * inside the allocator rather than hit the oom killer prematurely.
4291 * Let's give them a good hope and keep retrying while the order-0
4292 * watermarks are OK.
4294 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4295 ac->highest_zoneidx, ac->nodemask) {
4296 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4297 ac->highest_zoneidx, alloc_flags))
4302 #endif /* CONFIG_COMPACTION */
4304 #ifdef CONFIG_LOCKDEP
4305 static struct lockdep_map __fs_reclaim_map =
4306 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4308 static bool __need_reclaim(gfp_t gfp_mask)
4310 /* no reclaim without waiting on it */
4311 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4314 /* this guy won't enter reclaim */
4315 if (current->flags & PF_MEMALLOC)
4318 if (gfp_mask & __GFP_NOLOCKDEP)
4324 void __fs_reclaim_acquire(void)
4326 lock_map_acquire(&__fs_reclaim_map);
4329 void __fs_reclaim_release(void)
4331 lock_map_release(&__fs_reclaim_map);
4334 void fs_reclaim_acquire(gfp_t gfp_mask)
4336 gfp_mask = current_gfp_context(gfp_mask);
4338 if (__need_reclaim(gfp_mask)) {
4339 if (gfp_mask & __GFP_FS)
4340 __fs_reclaim_acquire();
4342 #ifdef CONFIG_MMU_NOTIFIER
4343 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
4344 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
4349 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4351 void fs_reclaim_release(gfp_t gfp_mask)
4353 gfp_mask = current_gfp_context(gfp_mask);
4355 if (__need_reclaim(gfp_mask)) {
4356 if (gfp_mask & __GFP_FS)
4357 __fs_reclaim_release();
4360 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4363 /* Perform direct synchronous page reclaim */
4364 static unsigned long
4365 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4366 const struct alloc_context *ac)
4368 unsigned int noreclaim_flag;
4369 unsigned long pflags, progress;
4373 /* We now go into synchronous reclaim */
4374 cpuset_memory_pressure_bump();
4375 psi_memstall_enter(&pflags);
4376 fs_reclaim_acquire(gfp_mask);
4377 noreclaim_flag = memalloc_noreclaim_save();
4379 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4382 memalloc_noreclaim_restore(noreclaim_flag);
4383 fs_reclaim_release(gfp_mask);
4384 psi_memstall_leave(&pflags);
4391 /* The really slow allocator path where we enter direct reclaim */
4392 static inline struct page *
4393 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4394 unsigned int alloc_flags, const struct alloc_context *ac,
4395 unsigned long *did_some_progress)
4397 struct page *page = NULL;
4398 bool drained = false;
4400 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4401 if (unlikely(!(*did_some_progress)))
4405 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4408 * If an allocation failed after direct reclaim, it could be because
4409 * pages are pinned on the per-cpu lists or in high alloc reserves.
4410 * Shrink them and try again
4412 if (!page && !drained) {
4413 unreserve_highatomic_pageblock(ac, false);
4414 drain_all_pages(NULL);
4422 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4423 const struct alloc_context *ac)
4427 pg_data_t *last_pgdat = NULL;
4428 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4430 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4432 if (last_pgdat != zone->zone_pgdat)
4433 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4434 last_pgdat = zone->zone_pgdat;
4438 static inline unsigned int
4439 gfp_to_alloc_flags(gfp_t gfp_mask)
4441 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4444 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4445 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4446 * to save two branches.
4448 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4449 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4452 * The caller may dip into page reserves a bit more if the caller
4453 * cannot run direct reclaim, or if the caller has realtime scheduling
4454 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4455 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4457 alloc_flags |= (__force int)
4458 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4460 if (gfp_mask & __GFP_ATOMIC) {
4462 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4463 * if it can't schedule.
4465 if (!(gfp_mask & __GFP_NOMEMALLOC))
4466 alloc_flags |= ALLOC_HARDER;
4468 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4469 * comment for __cpuset_node_allowed().
4471 alloc_flags &= ~ALLOC_CPUSET;
4472 } else if (unlikely(rt_task(current)) && !in_interrupt())
4473 alloc_flags |= ALLOC_HARDER;
4475 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4480 static bool oom_reserves_allowed(struct task_struct *tsk)
4482 if (!tsk_is_oom_victim(tsk))
4486 * !MMU doesn't have oom reaper so give access to memory reserves
4487 * only to the thread with TIF_MEMDIE set
4489 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4496 * Distinguish requests which really need access to full memory
4497 * reserves from oom victims which can live with a portion of it
4499 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4501 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4503 if (gfp_mask & __GFP_MEMALLOC)
4504 return ALLOC_NO_WATERMARKS;
4505 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4506 return ALLOC_NO_WATERMARKS;
4507 if (!in_interrupt()) {
4508 if (current->flags & PF_MEMALLOC)
4509 return ALLOC_NO_WATERMARKS;
4510 else if (oom_reserves_allowed(current))
4517 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4519 return !!__gfp_pfmemalloc_flags(gfp_mask);
4523 * Checks whether it makes sense to retry the reclaim to make a forward progress
4524 * for the given allocation request.
4526 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4527 * without success, or when we couldn't even meet the watermark if we
4528 * reclaimed all remaining pages on the LRU lists.
4530 * Returns true if a retry is viable or false to enter the oom path.
4533 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4534 struct alloc_context *ac, int alloc_flags,
4535 bool did_some_progress, int *no_progress_loops)
4542 * Costly allocations might have made a progress but this doesn't mean
4543 * their order will become available due to high fragmentation so
4544 * always increment the no progress counter for them
4546 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4547 *no_progress_loops = 0;
4549 (*no_progress_loops)++;
4552 * Make sure we converge to OOM if we cannot make any progress
4553 * several times in the row.
4555 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4556 /* Before OOM, exhaust highatomic_reserve */
4557 return unreserve_highatomic_pageblock(ac, true);
4561 * Keep reclaiming pages while there is a chance this will lead
4562 * somewhere. If none of the target zones can satisfy our allocation
4563 * request even if all reclaimable pages are considered then we are
4564 * screwed and have to go OOM.
4566 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4567 ac->highest_zoneidx, ac->nodemask) {
4568 unsigned long available;
4569 unsigned long reclaimable;
4570 unsigned long min_wmark = min_wmark_pages(zone);
4573 available = reclaimable = zone_reclaimable_pages(zone);
4574 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4577 * Would the allocation succeed if we reclaimed all
4578 * reclaimable pages?
4580 wmark = __zone_watermark_ok(zone, order, min_wmark,
4581 ac->highest_zoneidx, alloc_flags, available);
4582 trace_reclaim_retry_zone(z, order, reclaimable,
4583 available, min_wmark, *no_progress_loops, wmark);
4586 * If we didn't make any progress and have a lot of
4587 * dirty + writeback pages then we should wait for
4588 * an IO to complete to slow down the reclaim and
4589 * prevent from pre mature OOM
4591 if (!did_some_progress) {
4592 unsigned long write_pending;
4594 write_pending = zone_page_state_snapshot(zone,
4595 NR_ZONE_WRITE_PENDING);
4597 if (2 * write_pending > reclaimable) {
4598 congestion_wait(BLK_RW_ASYNC, HZ/10);
4610 * Memory allocation/reclaim might be called from a WQ context and the
4611 * current implementation of the WQ concurrency control doesn't
4612 * recognize that a particular WQ is congested if the worker thread is
4613 * looping without ever sleeping. Therefore we have to do a short sleep
4614 * here rather than calling cond_resched().
4616 if (current->flags & PF_WQ_WORKER)
4617 schedule_timeout_uninterruptible(1);
4624 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4627 * It's possible that cpuset's mems_allowed and the nodemask from
4628 * mempolicy don't intersect. This should be normally dealt with by
4629 * policy_nodemask(), but it's possible to race with cpuset update in
4630 * such a way the check therein was true, and then it became false
4631 * before we got our cpuset_mems_cookie here.
4632 * This assumes that for all allocations, ac->nodemask can come only
4633 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4634 * when it does not intersect with the cpuset restrictions) or the
4635 * caller can deal with a violated nodemask.
4637 if (cpusets_enabled() && ac->nodemask &&
4638 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4639 ac->nodemask = NULL;
4644 * When updating a task's mems_allowed or mempolicy nodemask, it is
4645 * possible to race with parallel threads in such a way that our
4646 * allocation can fail while the mask is being updated. If we are about
4647 * to fail, check if the cpuset changed during allocation and if so,
4650 if (read_mems_allowed_retry(cpuset_mems_cookie))
4656 static inline struct page *
4657 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4658 struct alloc_context *ac)
4660 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4661 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4662 struct page *page = NULL;
4663 unsigned int alloc_flags;
4664 unsigned long did_some_progress;
4665 enum compact_priority compact_priority;
4666 enum compact_result compact_result;
4667 int compaction_retries;
4668 int no_progress_loops;
4669 unsigned int cpuset_mems_cookie;
4673 * We also sanity check to catch abuse of atomic reserves being used by
4674 * callers that are not in atomic context.
4676 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4677 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4678 gfp_mask &= ~__GFP_ATOMIC;
4681 compaction_retries = 0;
4682 no_progress_loops = 0;
4683 compact_priority = DEF_COMPACT_PRIORITY;
4684 cpuset_mems_cookie = read_mems_allowed_begin();
4687 * The fast path uses conservative alloc_flags to succeed only until
4688 * kswapd needs to be woken up, and to avoid the cost of setting up
4689 * alloc_flags precisely. So we do that now.
4691 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4694 * We need to recalculate the starting point for the zonelist iterator
4695 * because we might have used different nodemask in the fast path, or
4696 * there was a cpuset modification and we are retrying - otherwise we
4697 * could end up iterating over non-eligible zones endlessly.
4699 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4700 ac->highest_zoneidx, ac->nodemask);
4701 if (!ac->preferred_zoneref->zone)
4704 if (alloc_flags & ALLOC_KSWAPD)
4705 wake_all_kswapds(order, gfp_mask, ac);
4708 * The adjusted alloc_flags might result in immediate success, so try
4711 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4716 * For costly allocations, try direct compaction first, as it's likely
4717 * that we have enough base pages and don't need to reclaim. For non-
4718 * movable high-order allocations, do that as well, as compaction will
4719 * try prevent permanent fragmentation by migrating from blocks of the
4721 * Don't try this for allocations that are allowed to ignore
4722 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4724 if (can_direct_reclaim &&
4726 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4727 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4728 page = __alloc_pages_direct_compact(gfp_mask, order,
4730 INIT_COMPACT_PRIORITY,
4736 * Checks for costly allocations with __GFP_NORETRY, which
4737 * includes some THP page fault allocations
4739 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4741 * If allocating entire pageblock(s) and compaction
4742 * failed because all zones are below low watermarks
4743 * or is prohibited because it recently failed at this
4744 * order, fail immediately unless the allocator has
4745 * requested compaction and reclaim retry.
4748 * - potentially very expensive because zones are far
4749 * below their low watermarks or this is part of very
4750 * bursty high order allocations,
4751 * - not guaranteed to help because isolate_freepages()
4752 * may not iterate over freed pages as part of its
4754 * - unlikely to make entire pageblocks free on its
4757 if (compact_result == COMPACT_SKIPPED ||
4758 compact_result == COMPACT_DEFERRED)
4762 * Looks like reclaim/compaction is worth trying, but
4763 * sync compaction could be very expensive, so keep
4764 * using async compaction.
4766 compact_priority = INIT_COMPACT_PRIORITY;
4771 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4772 if (alloc_flags & ALLOC_KSWAPD)
4773 wake_all_kswapds(order, gfp_mask, ac);
4775 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4777 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4780 * Reset the nodemask and zonelist iterators if memory policies can be
4781 * ignored. These allocations are high priority and system rather than
4784 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4785 ac->nodemask = NULL;
4786 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4787 ac->highest_zoneidx, ac->nodemask);
4790 /* Attempt with potentially adjusted zonelist and alloc_flags */
4791 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4795 /* Caller is not willing to reclaim, we can't balance anything */
4796 if (!can_direct_reclaim)
4799 /* Avoid recursion of direct reclaim */
4800 if (current->flags & PF_MEMALLOC)
4803 /* Try direct reclaim and then allocating */
4804 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4805 &did_some_progress);
4809 /* Try direct compaction and then allocating */
4810 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4811 compact_priority, &compact_result);
4815 /* Do not loop if specifically requested */
4816 if (gfp_mask & __GFP_NORETRY)
4820 * Do not retry costly high order allocations unless they are
4821 * __GFP_RETRY_MAYFAIL
4823 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4826 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4827 did_some_progress > 0, &no_progress_loops))
4831 * It doesn't make any sense to retry for the compaction if the order-0
4832 * reclaim is not able to make any progress because the current
4833 * implementation of the compaction depends on the sufficient amount
4834 * of free memory (see __compaction_suitable)
4836 if (did_some_progress > 0 &&
4837 should_compact_retry(ac, order, alloc_flags,
4838 compact_result, &compact_priority,
4839 &compaction_retries))
4843 /* Deal with possible cpuset update races before we start OOM killing */
4844 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4847 /* Reclaim has failed us, start killing things */
4848 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4852 /* Avoid allocations with no watermarks from looping endlessly */
4853 if (tsk_is_oom_victim(current) &&
4854 (alloc_flags & ALLOC_OOM ||
4855 (gfp_mask & __GFP_NOMEMALLOC)))
4858 /* Retry as long as the OOM killer is making progress */
4859 if (did_some_progress) {
4860 no_progress_loops = 0;
4865 /* Deal with possible cpuset update races before we fail */
4866 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4870 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4873 if (gfp_mask & __GFP_NOFAIL) {
4875 * All existing users of the __GFP_NOFAIL are blockable, so warn
4876 * of any new users that actually require GFP_NOWAIT
4878 if (WARN_ON_ONCE(!can_direct_reclaim))
4882 * PF_MEMALLOC request from this context is rather bizarre
4883 * because we cannot reclaim anything and only can loop waiting
4884 * for somebody to do a work for us
4886 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4889 * non failing costly orders are a hard requirement which we
4890 * are not prepared for much so let's warn about these users
4891 * so that we can identify them and convert them to something
4894 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4897 * Help non-failing allocations by giving them access to memory
4898 * reserves but do not use ALLOC_NO_WATERMARKS because this
4899 * could deplete whole memory reserves which would just make
4900 * the situation worse
4902 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4910 warn_alloc(gfp_mask, ac->nodemask,
4911 "page allocation failure: order:%u", order);
4916 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4917 int preferred_nid, nodemask_t *nodemask,
4918 struct alloc_context *ac, gfp_t *alloc_mask,
4919 unsigned int *alloc_flags)
4921 ac->highest_zoneidx = gfp_zone(gfp_mask);
4922 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4923 ac->nodemask = nodemask;
4924 ac->migratetype = gfp_migratetype(gfp_mask);
4926 if (cpusets_enabled()) {
4927 *alloc_mask |= __GFP_HARDWALL;
4929 * When we are in the interrupt context, it is irrelevant
4930 * to the current task context. It means that any node ok.
4932 if (!in_interrupt() && !ac->nodemask)
4933 ac->nodemask = &cpuset_current_mems_allowed;
4935 *alloc_flags |= ALLOC_CPUSET;
4938 fs_reclaim_acquire(gfp_mask);
4939 fs_reclaim_release(gfp_mask);
4941 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4943 if (should_fail_alloc_page(gfp_mask, order))
4946 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4948 /* Dirty zone balancing only done in the fast path */
4949 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4952 * The preferred zone is used for statistics but crucially it is
4953 * also used as the starting point for the zonelist iterator. It
4954 * may get reset for allocations that ignore memory policies.
4956 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4957 ac->highest_zoneidx, ac->nodemask);
4963 * This is the 'heart' of the zoned buddy allocator.
4966 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4967 nodemask_t *nodemask)
4970 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4971 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4972 struct alloc_context ac = { };
4975 * There are several places where we assume that the order value is sane
4976 * so bail out early if the request is out of bound.
4978 if (unlikely(order >= MAX_ORDER)) {
4979 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4983 gfp_mask &= gfp_allowed_mask;
4984 alloc_mask = gfp_mask;
4985 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4989 * Forbid the first pass from falling back to types that fragment
4990 * memory until all local zones are considered.
4992 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4994 /* First allocation attempt */
4995 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
5000 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
5001 * resp. GFP_NOIO which has to be inherited for all allocation requests
5002 * from a particular context which has been marked by
5003 * memalloc_no{fs,io}_{save,restore}.
5005 alloc_mask = current_gfp_context(gfp_mask);
5006 ac.spread_dirty_pages = false;
5009 * Restore the original nodemask if it was potentially replaced with
5010 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
5012 ac.nodemask = nodemask;
5014 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
5017 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
5018 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
5019 __free_pages(page, order);
5023 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
5027 EXPORT_SYMBOL(__alloc_pages_nodemask);
5030 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
5031 * address cannot represent highmem pages. Use alloc_pages and then kmap if
5032 * you need to access high mem.
5034 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
5038 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
5041 return (unsigned long) page_address(page);
5043 EXPORT_SYMBOL(__get_free_pages);
5045 unsigned long get_zeroed_page(gfp_t gfp_mask)
5047 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
5049 EXPORT_SYMBOL(get_zeroed_page);
5051 static inline void free_the_page(struct page *page, unsigned int order)
5053 if (order == 0) /* Via pcp? */
5054 free_unref_page(page);
5056 __free_pages_ok(page, order, FPI_NONE);
5060 * __free_pages - Free pages allocated with alloc_pages().
5061 * @page: The page pointer returned from alloc_pages().
5062 * @order: The order of the allocation.
5064 * This function can free multi-page allocations that are not compound
5065 * pages. It does not check that the @order passed in matches that of
5066 * the allocation, so it is easy to leak memory. Freeing more memory
5067 * than was allocated will probably emit a warning.
5069 * If the last reference to this page is speculative, it will be released
5070 * by put_page() which only frees the first page of a non-compound
5071 * allocation. To prevent the remaining pages from being leaked, we free
5072 * the subsequent pages here. If you want to use the page's reference
5073 * count to decide when to free the allocation, you should allocate a
5074 * compound page, and use put_page() instead of __free_pages().
5076 * Context: May be called in interrupt context or while holding a normal
5077 * spinlock, but not in NMI context or while holding a raw spinlock.
5079 void __free_pages(struct page *page, unsigned int order)
5081 if (put_page_testzero(page))
5082 free_the_page(page, order);
5083 else if (!PageHead(page))
5085 free_the_page(page + (1 << order), order);
5087 EXPORT_SYMBOL(__free_pages);
5089 void free_pages(unsigned long addr, unsigned int order)
5092 VM_BUG_ON(!virt_addr_valid((void *)addr));
5093 __free_pages(virt_to_page((void *)addr), order);
5097 EXPORT_SYMBOL(free_pages);
5101 * An arbitrary-length arbitrary-offset area of memory which resides
5102 * within a 0 or higher order page. Multiple fragments within that page
5103 * are individually refcounted, in the page's reference counter.
5105 * The page_frag functions below provide a simple allocation framework for
5106 * page fragments. This is used by the network stack and network device
5107 * drivers to provide a backing region of memory for use as either an
5108 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
5110 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
5113 struct page *page = NULL;
5114 gfp_t gfp = gfp_mask;
5116 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5117 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
5119 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
5120 PAGE_FRAG_CACHE_MAX_ORDER);
5121 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5123 if (unlikely(!page))
5124 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5126 nc->va = page ? page_address(page) : NULL;
5131 void __page_frag_cache_drain(struct page *page, unsigned int count)
5133 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5135 if (page_ref_sub_and_test(page, count))
5136 free_the_page(page, compound_order(page));
5138 EXPORT_SYMBOL(__page_frag_cache_drain);
5140 void *page_frag_alloc(struct page_frag_cache *nc,
5141 unsigned int fragsz, gfp_t gfp_mask)
5143 unsigned int size = PAGE_SIZE;
5147 if (unlikely(!nc->va)) {
5149 page = __page_frag_cache_refill(nc, gfp_mask);
5153 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5154 /* if size can vary use size else just use PAGE_SIZE */
5157 /* Even if we own the page, we do not use atomic_set().
5158 * This would break get_page_unless_zero() users.
5160 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5162 /* reset page count bias and offset to start of new frag */
5163 nc->pfmemalloc = page_is_pfmemalloc(page);
5164 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5168 offset = nc->offset - fragsz;
5169 if (unlikely(offset < 0)) {
5170 page = virt_to_page(nc->va);
5172 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5175 if (unlikely(nc->pfmemalloc)) {
5176 free_the_page(page, compound_order(page));
5180 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5181 /* if size can vary use size else just use PAGE_SIZE */
5184 /* OK, page count is 0, we can safely set it */
5185 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5187 /* reset page count bias and offset to start of new frag */
5188 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5189 offset = size - fragsz;
5193 nc->offset = offset;
5195 return nc->va + offset;
5197 EXPORT_SYMBOL(page_frag_alloc);
5200 * Frees a page fragment allocated out of either a compound or order 0 page.
5202 void page_frag_free(void *addr)
5204 struct page *page = virt_to_head_page(addr);
5206 if (unlikely(put_page_testzero(page)))
5207 free_the_page(page, compound_order(page));
5209 EXPORT_SYMBOL(page_frag_free);
5211 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5215 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5216 unsigned long used = addr + PAGE_ALIGN(size);
5218 split_page(virt_to_page((void *)addr), order);
5219 while (used < alloc_end) {
5224 return (void *)addr;
5228 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5229 * @size: the number of bytes to allocate
5230 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5232 * This function is similar to alloc_pages(), except that it allocates the
5233 * minimum number of pages to satisfy the request. alloc_pages() can only
5234 * allocate memory in power-of-two pages.
5236 * This function is also limited by MAX_ORDER.
5238 * Memory allocated by this function must be released by free_pages_exact().
5240 * Return: pointer to the allocated area or %NULL in case of error.
5242 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5244 unsigned int order = get_order(size);
5247 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5248 gfp_mask &= ~__GFP_COMP;
5250 addr = __get_free_pages(gfp_mask, order);
5251 return make_alloc_exact(addr, order, size);
5253 EXPORT_SYMBOL(alloc_pages_exact);
5256 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5258 * @nid: the preferred node ID where memory should be allocated
5259 * @size: the number of bytes to allocate
5260 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5262 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5265 * Return: pointer to the allocated area or %NULL in case of error.
5267 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5269 unsigned int order = get_order(size);
5272 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5273 gfp_mask &= ~__GFP_COMP;
5275 p = alloc_pages_node(nid, gfp_mask, order);
5278 return make_alloc_exact((unsigned long)page_address(p), order, size);
5282 * free_pages_exact - release memory allocated via alloc_pages_exact()
5283 * @virt: the value returned by alloc_pages_exact.
5284 * @size: size of allocation, same value as passed to alloc_pages_exact().
5286 * Release the memory allocated by a previous call to alloc_pages_exact.
5288 void free_pages_exact(void *virt, size_t size)
5290 unsigned long addr = (unsigned long)virt;
5291 unsigned long end = addr + PAGE_ALIGN(size);
5293 while (addr < end) {
5298 EXPORT_SYMBOL(free_pages_exact);
5301 * nr_free_zone_pages - count number of pages beyond high watermark
5302 * @offset: The zone index of the highest zone
5304 * nr_free_zone_pages() counts the number of pages which are beyond the
5305 * high watermark within all zones at or below a given zone index. For each
5306 * zone, the number of pages is calculated as:
5308 * nr_free_zone_pages = managed_pages - high_pages
5310 * Return: number of pages beyond high watermark.
5312 static unsigned long nr_free_zone_pages(int offset)
5317 /* Just pick one node, since fallback list is circular */
5318 unsigned long sum = 0;
5320 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5322 for_each_zone_zonelist(zone, z, zonelist, offset) {
5323 unsigned long size = zone_managed_pages(zone);
5324 unsigned long high = high_wmark_pages(zone);
5333 * nr_free_buffer_pages - count number of pages beyond high watermark
5335 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5336 * watermark within ZONE_DMA and ZONE_NORMAL.
5338 * Return: number of pages beyond high watermark within ZONE_DMA and
5341 unsigned long nr_free_buffer_pages(void)
5343 return nr_free_zone_pages(gfp_zone(GFP_USER));
5345 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5347 static inline void show_node(struct zone *zone)
5349 if (IS_ENABLED(CONFIG_NUMA))
5350 printk("Node %d ", zone_to_nid(zone));
5353 long si_mem_available(void)
5356 unsigned long pagecache;
5357 unsigned long wmark_low = 0;
5358 unsigned long pages[NR_LRU_LISTS];
5359 unsigned long reclaimable;
5363 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5364 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5367 wmark_low += low_wmark_pages(zone);
5370 * Estimate the amount of memory available for userspace allocations,
5371 * without causing swapping.
5373 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5376 * Not all the page cache can be freed, otherwise the system will
5377 * start swapping. Assume at least half of the page cache, or the
5378 * low watermark worth of cache, needs to stay.
5380 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5381 pagecache -= min(pagecache / 2, wmark_low);
5382 available += pagecache;
5385 * Part of the reclaimable slab and other kernel memory consists of
5386 * items that are in use, and cannot be freed. Cap this estimate at the
5389 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5390 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5391 available += reclaimable - min(reclaimable / 2, wmark_low);
5397 EXPORT_SYMBOL_GPL(si_mem_available);
5399 void si_meminfo(struct sysinfo *val)
5401 val->totalram = totalram_pages();
5402 val->sharedram = global_node_page_state(NR_SHMEM);
5403 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5404 val->bufferram = nr_blockdev_pages();
5405 val->totalhigh = totalhigh_pages();
5406 val->freehigh = nr_free_highpages();
5407 val->mem_unit = PAGE_SIZE;
5410 EXPORT_SYMBOL(si_meminfo);
5413 void si_meminfo_node(struct sysinfo *val, int nid)
5415 int zone_type; /* needs to be signed */
5416 unsigned long managed_pages = 0;
5417 unsigned long managed_highpages = 0;
5418 unsigned long free_highpages = 0;
5419 pg_data_t *pgdat = NODE_DATA(nid);
5421 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5422 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5423 val->totalram = managed_pages;
5424 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5425 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5426 #ifdef CONFIG_HIGHMEM
5427 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5428 struct zone *zone = &pgdat->node_zones[zone_type];
5430 if (is_highmem(zone)) {
5431 managed_highpages += zone_managed_pages(zone);
5432 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5435 val->totalhigh = managed_highpages;
5436 val->freehigh = free_highpages;
5438 val->totalhigh = managed_highpages;
5439 val->freehigh = free_highpages;
5441 val->mem_unit = PAGE_SIZE;
5446 * Determine whether the node should be displayed or not, depending on whether
5447 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5449 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5451 if (!(flags & SHOW_MEM_FILTER_NODES))
5455 * no node mask - aka implicit memory numa policy. Do not bother with
5456 * the synchronization - read_mems_allowed_begin - because we do not
5457 * have to be precise here.
5460 nodemask = &cpuset_current_mems_allowed;
5462 return !node_isset(nid, *nodemask);
5465 #define K(x) ((x) << (PAGE_SHIFT-10))
5467 static void show_migration_types(unsigned char type)
5469 static const char types[MIGRATE_TYPES] = {
5470 [MIGRATE_UNMOVABLE] = 'U',
5471 [MIGRATE_MOVABLE] = 'M',
5472 [MIGRATE_RECLAIMABLE] = 'E',
5473 [MIGRATE_HIGHATOMIC] = 'H',
5475 [MIGRATE_CMA] = 'C',
5477 #ifdef CONFIG_MEMORY_ISOLATION
5478 [MIGRATE_ISOLATE] = 'I',
5481 char tmp[MIGRATE_TYPES + 1];
5485 for (i = 0; i < MIGRATE_TYPES; i++) {
5486 if (type & (1 << i))
5491 printk(KERN_CONT "(%s) ", tmp);
5495 * Show free area list (used inside shift_scroll-lock stuff)
5496 * We also calculate the percentage fragmentation. We do this by counting the
5497 * memory on each free list with the exception of the first item on the list.
5500 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5503 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5505 unsigned long free_pcp = 0;
5510 for_each_populated_zone(zone) {
5511 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5514 for_each_online_cpu(cpu)
5515 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5518 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5519 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5520 " unevictable:%lu dirty:%lu writeback:%lu\n"
5521 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5522 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5523 " free:%lu free_pcp:%lu free_cma:%lu\n",
5524 global_node_page_state(NR_ACTIVE_ANON),
5525 global_node_page_state(NR_INACTIVE_ANON),
5526 global_node_page_state(NR_ISOLATED_ANON),
5527 global_node_page_state(NR_ACTIVE_FILE),
5528 global_node_page_state(NR_INACTIVE_FILE),
5529 global_node_page_state(NR_ISOLATED_FILE),
5530 global_node_page_state(NR_UNEVICTABLE),
5531 global_node_page_state(NR_FILE_DIRTY),
5532 global_node_page_state(NR_WRITEBACK),
5533 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5534 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5535 global_node_page_state(NR_FILE_MAPPED),
5536 global_node_page_state(NR_SHMEM),
5537 global_node_page_state(NR_PAGETABLE),
5538 global_zone_page_state(NR_BOUNCE),
5539 global_zone_page_state(NR_FREE_PAGES),
5541 global_zone_page_state(NR_FREE_CMA_PAGES));
5543 for_each_online_pgdat(pgdat) {
5544 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5548 " active_anon:%lukB"
5549 " inactive_anon:%lukB"
5550 " active_file:%lukB"
5551 " inactive_file:%lukB"
5552 " unevictable:%lukB"
5553 " isolated(anon):%lukB"
5554 " isolated(file):%lukB"
5559 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5561 " shmem_pmdmapped: %lukB"
5564 " writeback_tmp:%lukB"
5565 " kernel_stack:%lukB"
5566 #ifdef CONFIG_SHADOW_CALL_STACK
5567 " shadow_call_stack:%lukB"
5570 " all_unreclaimable? %s"
5573 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5574 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5575 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5576 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5577 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5578 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5579 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5580 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5581 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5582 K(node_page_state(pgdat, NR_WRITEBACK)),
5583 K(node_page_state(pgdat, NR_SHMEM)),
5584 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5585 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5586 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5588 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5590 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5591 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5592 #ifdef CONFIG_SHADOW_CALL_STACK
5593 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5595 K(node_page_state(pgdat, NR_PAGETABLE)),
5596 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5600 for_each_populated_zone(zone) {
5603 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5607 for_each_online_cpu(cpu)
5608 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5617 " reserved_highatomic:%luKB"
5618 " active_anon:%lukB"
5619 " inactive_anon:%lukB"
5620 " active_file:%lukB"
5621 " inactive_file:%lukB"
5622 " unevictable:%lukB"
5623 " writepending:%lukB"
5633 K(zone_page_state(zone, NR_FREE_PAGES)),
5634 K(min_wmark_pages(zone)),
5635 K(low_wmark_pages(zone)),
5636 K(high_wmark_pages(zone)),
5637 K(zone->nr_reserved_highatomic),
5638 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5639 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5640 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5641 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5642 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5643 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5644 K(zone->present_pages),
5645 K(zone_managed_pages(zone)),
5646 K(zone_page_state(zone, NR_MLOCK)),
5647 K(zone_page_state(zone, NR_BOUNCE)),
5649 K(this_cpu_read(zone->pageset->pcp.count)),
5650 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5651 printk("lowmem_reserve[]:");
5652 for (i = 0; i < MAX_NR_ZONES; i++)
5653 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5654 printk(KERN_CONT "\n");
5657 for_each_populated_zone(zone) {
5659 unsigned long nr[MAX_ORDER], flags, total = 0;
5660 unsigned char types[MAX_ORDER];
5662 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5665 printk(KERN_CONT "%s: ", zone->name);
5667 spin_lock_irqsave(&zone->lock, flags);
5668 for (order = 0; order < MAX_ORDER; order++) {
5669 struct free_area *area = &zone->free_area[order];
5672 nr[order] = area->nr_free;
5673 total += nr[order] << order;
5676 for (type = 0; type < MIGRATE_TYPES; type++) {
5677 if (!free_area_empty(area, type))
5678 types[order] |= 1 << type;
5681 spin_unlock_irqrestore(&zone->lock, flags);
5682 for (order = 0; order < MAX_ORDER; order++) {
5683 printk(KERN_CONT "%lu*%lukB ",
5684 nr[order], K(1UL) << order);
5686 show_migration_types(types[order]);
5688 printk(KERN_CONT "= %lukB\n", K(total));
5691 hugetlb_show_meminfo();
5693 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5695 show_swap_cache_info();
5698 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5700 zoneref->zone = zone;
5701 zoneref->zone_idx = zone_idx(zone);
5705 * Builds allocation fallback zone lists.
5707 * Add all populated zones of a node to the zonelist.
5709 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5712 enum zone_type zone_type = MAX_NR_ZONES;
5717 zone = pgdat->node_zones + zone_type;
5718 if (managed_zone(zone)) {
5719 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5720 check_highest_zone(zone_type);
5722 } while (zone_type);
5729 static int __parse_numa_zonelist_order(char *s)
5732 * We used to support different zonlists modes but they turned
5733 * out to be just not useful. Let's keep the warning in place
5734 * if somebody still use the cmd line parameter so that we do
5735 * not fail it silently
5737 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5738 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5744 char numa_zonelist_order[] = "Node";
5747 * sysctl handler for numa_zonelist_order
5749 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5750 void *buffer, size_t *length, loff_t *ppos)
5753 return __parse_numa_zonelist_order(buffer);
5754 return proc_dostring(table, write, buffer, length, ppos);
5758 #define MAX_NODE_LOAD (nr_online_nodes)
5759 static int node_load[MAX_NUMNODES];
5762 * find_next_best_node - find the next node that should appear in a given node's fallback list
5763 * @node: node whose fallback list we're appending
5764 * @used_node_mask: nodemask_t of already used nodes
5766 * We use a number of factors to determine which is the next node that should
5767 * appear on a given node's fallback list. The node should not have appeared
5768 * already in @node's fallback list, and it should be the next closest node
5769 * according to the distance array (which contains arbitrary distance values
5770 * from each node to each node in the system), and should also prefer nodes
5771 * with no CPUs, since presumably they'll have very little allocation pressure
5772 * on them otherwise.
5774 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5776 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5779 int min_val = INT_MAX;
5780 int best_node = NUMA_NO_NODE;
5782 /* Use the local node if we haven't already */
5783 if (!node_isset(node, *used_node_mask)) {
5784 node_set(node, *used_node_mask);
5788 for_each_node_state(n, N_MEMORY) {
5790 /* Don't want a node to appear more than once */
5791 if (node_isset(n, *used_node_mask))
5794 /* Use the distance array to find the distance */
5795 val = node_distance(node, n);
5797 /* Penalize nodes under us ("prefer the next node") */
5800 /* Give preference to headless and unused nodes */
5801 if (!cpumask_empty(cpumask_of_node(n)))
5802 val += PENALTY_FOR_NODE_WITH_CPUS;
5804 /* Slight preference for less loaded node */
5805 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5806 val += node_load[n];
5808 if (val < min_val) {
5815 node_set(best_node, *used_node_mask);
5822 * Build zonelists ordered by node and zones within node.
5823 * This results in maximum locality--normal zone overflows into local
5824 * DMA zone, if any--but risks exhausting DMA zone.
5826 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5829 struct zoneref *zonerefs;
5832 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5834 for (i = 0; i < nr_nodes; i++) {
5837 pg_data_t *node = NODE_DATA(node_order[i]);
5839 nr_zones = build_zonerefs_node(node, zonerefs);
5840 zonerefs += nr_zones;
5842 zonerefs->zone = NULL;
5843 zonerefs->zone_idx = 0;
5847 * Build gfp_thisnode zonelists
5849 static void build_thisnode_zonelists(pg_data_t *pgdat)
5851 struct zoneref *zonerefs;
5854 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5855 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5856 zonerefs += nr_zones;
5857 zonerefs->zone = NULL;
5858 zonerefs->zone_idx = 0;
5862 * Build zonelists ordered by zone and nodes within zones.
5863 * This results in conserving DMA zone[s] until all Normal memory is
5864 * exhausted, but results in overflowing to remote node while memory
5865 * may still exist in local DMA zone.
5868 static void build_zonelists(pg_data_t *pgdat)
5870 static int node_order[MAX_NUMNODES];
5871 int node, load, nr_nodes = 0;
5872 nodemask_t used_mask = NODE_MASK_NONE;
5873 int local_node, prev_node;
5875 /* NUMA-aware ordering of nodes */
5876 local_node = pgdat->node_id;
5877 load = nr_online_nodes;
5878 prev_node = local_node;
5880 memset(node_order, 0, sizeof(node_order));
5881 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5883 * We don't want to pressure a particular node.
5884 * So adding penalty to the first node in same
5885 * distance group to make it round-robin.
5887 if (node_distance(local_node, node) !=
5888 node_distance(local_node, prev_node))
5889 node_load[node] = load;
5891 node_order[nr_nodes++] = node;
5896 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5897 build_thisnode_zonelists(pgdat);
5900 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5902 * Return node id of node used for "local" allocations.
5903 * I.e., first node id of first zone in arg node's generic zonelist.
5904 * Used for initializing percpu 'numa_mem', which is used primarily
5905 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5907 int local_memory_node(int node)
5911 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5912 gfp_zone(GFP_KERNEL),
5914 return zone_to_nid(z->zone);
5918 static void setup_min_unmapped_ratio(void);
5919 static void setup_min_slab_ratio(void);
5920 #else /* CONFIG_NUMA */
5922 static void build_zonelists(pg_data_t *pgdat)
5924 int node, local_node;
5925 struct zoneref *zonerefs;
5928 local_node = pgdat->node_id;
5930 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5931 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5932 zonerefs += nr_zones;
5935 * Now we build the zonelist so that it contains the zones
5936 * of all the other nodes.
5937 * We don't want to pressure a particular node, so when
5938 * building the zones for node N, we make sure that the
5939 * zones coming right after the local ones are those from
5940 * node N+1 (modulo N)
5942 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5943 if (!node_online(node))
5945 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5946 zonerefs += nr_zones;
5948 for (node = 0; node < local_node; node++) {
5949 if (!node_online(node))
5951 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5952 zonerefs += nr_zones;
5955 zonerefs->zone = NULL;
5956 zonerefs->zone_idx = 0;
5959 #endif /* CONFIG_NUMA */
5962 * Boot pageset table. One per cpu which is going to be used for all
5963 * zones and all nodes. The parameters will be set in such a way
5964 * that an item put on a list will immediately be handed over to
5965 * the buddy list. This is safe since pageset manipulation is done
5966 * with interrupts disabled.
5968 * The boot_pagesets must be kept even after bootup is complete for
5969 * unused processors and/or zones. They do play a role for bootstrapping
5970 * hotplugged processors.
5972 * zoneinfo_show() and maybe other functions do
5973 * not check if the processor is online before following the pageset pointer.
5974 * Other parts of the kernel may not check if the zone is available.
5976 static void pageset_init(struct per_cpu_pageset *p);
5977 /* These effectively disable the pcplists in the boot pageset completely */
5978 #define BOOT_PAGESET_HIGH 0
5979 #define BOOT_PAGESET_BATCH 1
5980 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5981 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5983 static void __build_all_zonelists(void *data)
5986 int __maybe_unused cpu;
5987 pg_data_t *self = data;
5988 static DEFINE_SPINLOCK(lock);
5993 memset(node_load, 0, sizeof(node_load));
5997 * This node is hotadded and no memory is yet present. So just
5998 * building zonelists is fine - no need to touch other nodes.
6000 if (self && !node_online(self->node_id)) {
6001 build_zonelists(self);
6003 for_each_online_node(nid) {
6004 pg_data_t *pgdat = NODE_DATA(nid);
6006 build_zonelists(pgdat);
6009 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
6011 * We now know the "local memory node" for each node--
6012 * i.e., the node of the first zone in the generic zonelist.
6013 * Set up numa_mem percpu variable for on-line cpus. During
6014 * boot, only the boot cpu should be on-line; we'll init the
6015 * secondary cpus' numa_mem as they come on-line. During
6016 * node/memory hotplug, we'll fixup all on-line cpus.
6018 for_each_online_cpu(cpu)
6019 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
6026 static noinline void __init
6027 build_all_zonelists_init(void)
6031 __build_all_zonelists(NULL);
6034 * Initialize the boot_pagesets that are going to be used
6035 * for bootstrapping processors. The real pagesets for
6036 * each zone will be allocated later when the per cpu
6037 * allocator is available.
6039 * boot_pagesets are used also for bootstrapping offline
6040 * cpus if the system is already booted because the pagesets
6041 * are needed to initialize allocators on a specific cpu too.
6042 * F.e. the percpu allocator needs the page allocator which
6043 * needs the percpu allocator in order to allocate its pagesets
6044 * (a chicken-egg dilemma).
6046 for_each_possible_cpu(cpu)
6047 pageset_init(&per_cpu(boot_pageset, cpu));
6049 mminit_verify_zonelist();
6050 cpuset_init_current_mems_allowed();
6054 * unless system_state == SYSTEM_BOOTING.
6056 * __ref due to call of __init annotated helper build_all_zonelists_init
6057 * [protected by SYSTEM_BOOTING].
6059 void __ref build_all_zonelists(pg_data_t *pgdat)
6061 unsigned long vm_total_pages;
6063 if (system_state == SYSTEM_BOOTING) {
6064 build_all_zonelists_init();
6066 __build_all_zonelists(pgdat);
6067 /* cpuset refresh routine should be here */
6069 /* Get the number of free pages beyond high watermark in all zones. */
6070 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
6072 * Disable grouping by mobility if the number of pages in the
6073 * system is too low to allow the mechanism to work. It would be
6074 * more accurate, but expensive to check per-zone. This check is
6075 * made on memory-hotadd so a system can start with mobility
6076 * disabled and enable it later
6078 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
6079 page_group_by_mobility_disabled = 1;
6081 page_group_by_mobility_disabled = 0;
6083 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
6085 page_group_by_mobility_disabled ? "off" : "on",
6088 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
6092 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
6093 static bool __meminit
6094 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
6096 static struct memblock_region *r;
6098 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
6099 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
6100 for_each_mem_region(r) {
6101 if (*pfn < memblock_region_memory_end_pfn(r))
6105 if (*pfn >= memblock_region_memory_base_pfn(r) &&
6106 memblock_is_mirror(r)) {
6107 *pfn = memblock_region_memory_end_pfn(r);
6115 * Initially all pages are reserved - free ones are freed
6116 * up by memblock_free_all() once the early boot process is
6117 * done. Non-atomic initialization, single-pass.
6119 * All aligned pageblocks are initialized to the specified migratetype
6120 * (usually MIGRATE_MOVABLE). Besides setting the migratetype, no related
6121 * zone stats (e.g., nr_isolate_pageblock) are touched.
6123 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
6124 unsigned long start_pfn, unsigned long zone_end_pfn,
6125 enum meminit_context context,
6126 struct vmem_altmap *altmap, int migratetype)
6128 unsigned long pfn, end_pfn = start_pfn + size;
6131 if (highest_memmap_pfn < end_pfn - 1)
6132 highest_memmap_pfn = end_pfn - 1;
6134 #ifdef CONFIG_ZONE_DEVICE
6136 * Honor reservation requested by the driver for this ZONE_DEVICE
6137 * memory. We limit the total number of pages to initialize to just
6138 * those that might contain the memory mapping. We will defer the
6139 * ZONE_DEVICE page initialization until after we have released
6142 if (zone == ZONE_DEVICE) {
6146 if (start_pfn == altmap->base_pfn)
6147 start_pfn += altmap->reserve;
6148 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6152 for (pfn = start_pfn; pfn < end_pfn; ) {
6154 * There can be holes in boot-time mem_map[]s handed to this
6155 * function. They do not exist on hotplugged memory.
6157 if (context == MEMINIT_EARLY) {
6158 if (overlap_memmap_init(zone, &pfn))
6160 if (defer_init(nid, pfn, zone_end_pfn))
6164 page = pfn_to_page(pfn);
6165 __init_single_page(page, pfn, zone, nid);
6166 if (context == MEMINIT_HOTPLUG)
6167 __SetPageReserved(page);
6170 * Usually, we want to mark the pageblock MIGRATE_MOVABLE,
6171 * such that unmovable allocations won't be scattered all
6172 * over the place during system boot.
6174 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6175 set_pageblock_migratetype(page, migratetype);
6182 #ifdef CONFIG_ZONE_DEVICE
6183 void __ref memmap_init_zone_device(struct zone *zone,
6184 unsigned long start_pfn,
6185 unsigned long nr_pages,
6186 struct dev_pagemap *pgmap)
6188 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6189 struct pglist_data *pgdat = zone->zone_pgdat;
6190 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6191 unsigned long zone_idx = zone_idx(zone);
6192 unsigned long start = jiffies;
6193 int nid = pgdat->node_id;
6195 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6199 * The call to memmap_init_zone should have already taken care
6200 * of the pages reserved for the memmap, so we can just jump to
6201 * the end of that region and start processing the device pages.
6204 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6205 nr_pages = end_pfn - start_pfn;
6208 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6209 struct page *page = pfn_to_page(pfn);
6211 __init_single_page(page, pfn, zone_idx, nid);
6214 * Mark page reserved as it will need to wait for onlining
6215 * phase for it to be fully associated with a zone.
6217 * We can use the non-atomic __set_bit operation for setting
6218 * the flag as we are still initializing the pages.
6220 __SetPageReserved(page);
6223 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6224 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6225 * ever freed or placed on a driver-private list.
6227 page->pgmap = pgmap;
6228 page->zone_device_data = NULL;
6231 * Mark the block movable so that blocks are reserved for
6232 * movable at startup. This will force kernel allocations
6233 * to reserve their blocks rather than leaking throughout
6234 * the address space during boot when many long-lived
6235 * kernel allocations are made.
6237 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6238 * because this is done early in section_activate()
6240 if (IS_ALIGNED(pfn, pageblock_nr_pages)) {
6241 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6246 pr_info("%s initialised %lu pages in %ums\n", __func__,
6247 nr_pages, jiffies_to_msecs(jiffies - start));
6251 static void __meminit zone_init_free_lists(struct zone *zone)
6253 unsigned int order, t;
6254 for_each_migratetype_order(order, t) {
6255 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6256 zone->free_area[order].nr_free = 0;
6260 void __meminit __weak memmap_init(unsigned long size, int nid,
6262 unsigned long range_start_pfn)
6264 unsigned long start_pfn, end_pfn;
6265 unsigned long range_end_pfn = range_start_pfn + size;
6268 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6269 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6270 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6272 if (end_pfn > start_pfn) {
6273 size = end_pfn - start_pfn;
6274 memmap_init_zone(size, nid, zone, start_pfn, range_end_pfn,
6275 MEMINIT_EARLY, NULL, MIGRATE_MOVABLE);
6280 static int zone_batchsize(struct zone *zone)
6286 * The per-cpu-pages pools are set to around 1000th of the
6289 batch = zone_managed_pages(zone) / 1024;
6290 /* But no more than a meg. */
6291 if (batch * PAGE_SIZE > 1024 * 1024)
6292 batch = (1024 * 1024) / PAGE_SIZE;
6293 batch /= 4; /* We effectively *= 4 below */
6298 * Clamp the batch to a 2^n - 1 value. Having a power
6299 * of 2 value was found to be more likely to have
6300 * suboptimal cache aliasing properties in some cases.
6302 * For example if 2 tasks are alternately allocating
6303 * batches of pages, one task can end up with a lot
6304 * of pages of one half of the possible page colors
6305 * and the other with pages of the other colors.
6307 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6312 /* The deferral and batching of frees should be suppressed under NOMMU
6315 * The problem is that NOMMU needs to be able to allocate large chunks
6316 * of contiguous memory as there's no hardware page translation to
6317 * assemble apparent contiguous memory from discontiguous pages.
6319 * Queueing large contiguous runs of pages for batching, however,
6320 * causes the pages to actually be freed in smaller chunks. As there
6321 * can be a significant delay between the individual batches being
6322 * recycled, this leads to the once large chunks of space being
6323 * fragmented and becoming unavailable for high-order allocations.
6330 * pcp->high and pcp->batch values are related and generally batch is lower
6331 * than high. They are also related to pcp->count such that count is lower
6332 * than high, and as soon as it reaches high, the pcplist is flushed.
6334 * However, guaranteeing these relations at all times would require e.g. write
6335 * barriers here but also careful usage of read barriers at the read side, and
6336 * thus be prone to error and bad for performance. Thus the update only prevents
6337 * store tearing. Any new users of pcp->batch and pcp->high should ensure they
6338 * can cope with those fields changing asynchronously, and fully trust only the
6339 * pcp->count field on the local CPU with interrupts disabled.
6341 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6342 * outside of boot time (or some other assurance that no concurrent updaters
6345 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6346 unsigned long batch)
6348 WRITE_ONCE(pcp->batch, batch);
6349 WRITE_ONCE(pcp->high, high);
6352 static void pageset_init(struct per_cpu_pageset *p)
6354 struct per_cpu_pages *pcp;
6357 memset(p, 0, sizeof(*p));
6360 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6361 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6364 * Set batch and high values safe for a boot pageset. A true percpu
6365 * pageset's initialization will update them subsequently. Here we don't
6366 * need to be as careful as pageset_update() as nobody can access the
6369 pcp->high = BOOT_PAGESET_HIGH;
6370 pcp->batch = BOOT_PAGESET_BATCH;
6373 static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high,
6374 unsigned long batch)
6376 struct per_cpu_pageset *p;
6379 for_each_possible_cpu(cpu) {
6380 p = per_cpu_ptr(zone->pageset, cpu);
6381 pageset_update(&p->pcp, high, batch);
6386 * Calculate and set new high and batch values for all per-cpu pagesets of a
6387 * zone, based on the zone's size and the percpu_pagelist_fraction sysctl.
6389 static void zone_set_pageset_high_and_batch(struct zone *zone)
6391 unsigned long new_high, new_batch;
6393 if (percpu_pagelist_fraction) {
6394 new_high = zone_managed_pages(zone) / percpu_pagelist_fraction;
6395 new_batch = max(1UL, new_high / 4);
6396 if ((new_high / 4) > (PAGE_SHIFT * 8))
6397 new_batch = PAGE_SHIFT * 8;
6399 new_batch = zone_batchsize(zone);
6400 new_high = 6 * new_batch;
6401 new_batch = max(1UL, 1 * new_batch);
6404 if (zone->pageset_high == new_high &&
6405 zone->pageset_batch == new_batch)
6408 zone->pageset_high = new_high;
6409 zone->pageset_batch = new_batch;
6411 __zone_set_pageset_high_and_batch(zone, new_high, new_batch);
6414 void __meminit setup_zone_pageset(struct zone *zone)
6416 struct per_cpu_pageset *p;
6419 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6420 for_each_possible_cpu(cpu) {
6421 p = per_cpu_ptr(zone->pageset, cpu);
6425 zone_set_pageset_high_and_batch(zone);
6429 * Allocate per cpu pagesets and initialize them.
6430 * Before this call only boot pagesets were available.
6432 void __init setup_per_cpu_pageset(void)
6434 struct pglist_data *pgdat;
6436 int __maybe_unused cpu;
6438 for_each_populated_zone(zone)
6439 setup_zone_pageset(zone);
6443 * Unpopulated zones continue using the boot pagesets.
6444 * The numa stats for these pagesets need to be reset.
6445 * Otherwise, they will end up skewing the stats of
6446 * the nodes these zones are associated with.
6448 for_each_possible_cpu(cpu) {
6449 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6450 memset(pcp->vm_numa_stat_diff, 0,
6451 sizeof(pcp->vm_numa_stat_diff));
6455 for_each_online_pgdat(pgdat)
6456 pgdat->per_cpu_nodestats =
6457 alloc_percpu(struct per_cpu_nodestat);
6460 static __meminit void zone_pcp_init(struct zone *zone)
6463 * per cpu subsystem is not up at this point. The following code
6464 * relies on the ability of the linker to provide the
6465 * offset of a (static) per cpu variable into the per cpu area.
6467 zone->pageset = &boot_pageset;
6468 zone->pageset_high = BOOT_PAGESET_HIGH;
6469 zone->pageset_batch = BOOT_PAGESET_BATCH;
6471 if (populated_zone(zone))
6472 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6473 zone->name, zone->present_pages,
6474 zone_batchsize(zone));
6477 void __meminit init_currently_empty_zone(struct zone *zone,
6478 unsigned long zone_start_pfn,
6481 struct pglist_data *pgdat = zone->zone_pgdat;
6482 int zone_idx = zone_idx(zone) + 1;
6484 if (zone_idx > pgdat->nr_zones)
6485 pgdat->nr_zones = zone_idx;
6487 zone->zone_start_pfn = zone_start_pfn;
6489 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6490 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6492 (unsigned long)zone_idx(zone),
6493 zone_start_pfn, (zone_start_pfn + size));
6495 zone_init_free_lists(zone);
6496 zone->initialized = 1;
6500 * get_pfn_range_for_nid - Return the start and end page frames for a node
6501 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6502 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6503 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6505 * It returns the start and end page frame of a node based on information
6506 * provided by memblock_set_node(). If called for a node
6507 * with no available memory, a warning is printed and the start and end
6510 void __init get_pfn_range_for_nid(unsigned int nid,
6511 unsigned long *start_pfn, unsigned long *end_pfn)
6513 unsigned long this_start_pfn, this_end_pfn;
6519 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6520 *start_pfn = min(*start_pfn, this_start_pfn);
6521 *end_pfn = max(*end_pfn, this_end_pfn);
6524 if (*start_pfn == -1UL)
6529 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6530 * assumption is made that zones within a node are ordered in monotonic
6531 * increasing memory addresses so that the "highest" populated zone is used
6533 static void __init find_usable_zone_for_movable(void)
6536 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6537 if (zone_index == ZONE_MOVABLE)
6540 if (arch_zone_highest_possible_pfn[zone_index] >
6541 arch_zone_lowest_possible_pfn[zone_index])
6545 VM_BUG_ON(zone_index == -1);
6546 movable_zone = zone_index;
6550 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6551 * because it is sized independent of architecture. Unlike the other zones,
6552 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6553 * in each node depending on the size of each node and how evenly kernelcore
6554 * is distributed. This helper function adjusts the zone ranges
6555 * provided by the architecture for a given node by using the end of the
6556 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6557 * zones within a node are in order of monotonic increases memory addresses
6559 static void __init adjust_zone_range_for_zone_movable(int nid,
6560 unsigned long zone_type,
6561 unsigned long node_start_pfn,
6562 unsigned long node_end_pfn,
6563 unsigned long *zone_start_pfn,
6564 unsigned long *zone_end_pfn)
6566 /* Only adjust if ZONE_MOVABLE is on this node */
6567 if (zone_movable_pfn[nid]) {
6568 /* Size ZONE_MOVABLE */
6569 if (zone_type == ZONE_MOVABLE) {
6570 *zone_start_pfn = zone_movable_pfn[nid];
6571 *zone_end_pfn = min(node_end_pfn,
6572 arch_zone_highest_possible_pfn[movable_zone]);
6574 /* Adjust for ZONE_MOVABLE starting within this range */
6575 } else if (!mirrored_kernelcore &&
6576 *zone_start_pfn < zone_movable_pfn[nid] &&
6577 *zone_end_pfn > zone_movable_pfn[nid]) {
6578 *zone_end_pfn = zone_movable_pfn[nid];
6580 /* Check if this whole range is within ZONE_MOVABLE */
6581 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6582 *zone_start_pfn = *zone_end_pfn;
6587 * Return the number of pages a zone spans in a node, including holes
6588 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6590 static unsigned long __init zone_spanned_pages_in_node(int nid,
6591 unsigned long zone_type,
6592 unsigned long node_start_pfn,
6593 unsigned long node_end_pfn,
6594 unsigned long *zone_start_pfn,
6595 unsigned long *zone_end_pfn)
6597 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6598 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6599 /* When hotadd a new node from cpu_up(), the node should be empty */
6600 if (!node_start_pfn && !node_end_pfn)
6603 /* Get the start and end of the zone */
6604 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6605 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6606 adjust_zone_range_for_zone_movable(nid, zone_type,
6607 node_start_pfn, node_end_pfn,
6608 zone_start_pfn, zone_end_pfn);
6610 /* Check that this node has pages within the zone's required range */
6611 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6614 /* Move the zone boundaries inside the node if necessary */
6615 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6616 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6618 /* Return the spanned pages */
6619 return *zone_end_pfn - *zone_start_pfn;
6623 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6624 * then all holes in the requested range will be accounted for.
6626 unsigned long __init __absent_pages_in_range(int nid,
6627 unsigned long range_start_pfn,
6628 unsigned long range_end_pfn)
6630 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6631 unsigned long start_pfn, end_pfn;
6634 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6635 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6636 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6637 nr_absent -= end_pfn - start_pfn;
6643 * absent_pages_in_range - Return number of page frames in holes within a range
6644 * @start_pfn: The start PFN to start searching for holes
6645 * @end_pfn: The end PFN to stop searching for holes
6647 * Return: the number of pages frames in memory holes within a range.
6649 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6650 unsigned long end_pfn)
6652 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6655 /* Return the number of page frames in holes in a zone on a node */
6656 static unsigned long __init zone_absent_pages_in_node(int nid,
6657 unsigned long zone_type,
6658 unsigned long node_start_pfn,
6659 unsigned long node_end_pfn)
6661 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6662 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6663 unsigned long zone_start_pfn, zone_end_pfn;
6664 unsigned long nr_absent;
6666 /* When hotadd a new node from cpu_up(), the node should be empty */
6667 if (!node_start_pfn && !node_end_pfn)
6670 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6671 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6673 adjust_zone_range_for_zone_movable(nid, zone_type,
6674 node_start_pfn, node_end_pfn,
6675 &zone_start_pfn, &zone_end_pfn);
6676 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6679 * ZONE_MOVABLE handling.
6680 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6683 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6684 unsigned long start_pfn, end_pfn;
6685 struct memblock_region *r;
6687 for_each_mem_region(r) {
6688 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6689 zone_start_pfn, zone_end_pfn);
6690 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6691 zone_start_pfn, zone_end_pfn);
6693 if (zone_type == ZONE_MOVABLE &&
6694 memblock_is_mirror(r))
6695 nr_absent += end_pfn - start_pfn;
6697 if (zone_type == ZONE_NORMAL &&
6698 !memblock_is_mirror(r))
6699 nr_absent += end_pfn - start_pfn;
6706 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6707 unsigned long node_start_pfn,
6708 unsigned long node_end_pfn)
6710 unsigned long realtotalpages = 0, totalpages = 0;
6713 for (i = 0; i < MAX_NR_ZONES; i++) {
6714 struct zone *zone = pgdat->node_zones + i;
6715 unsigned long zone_start_pfn, zone_end_pfn;
6716 unsigned long spanned, absent;
6717 unsigned long size, real_size;
6719 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6724 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6729 real_size = size - absent;
6732 zone->zone_start_pfn = zone_start_pfn;
6734 zone->zone_start_pfn = 0;
6735 zone->spanned_pages = size;
6736 zone->present_pages = real_size;
6739 realtotalpages += real_size;
6742 pgdat->node_spanned_pages = totalpages;
6743 pgdat->node_present_pages = realtotalpages;
6744 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6748 #ifndef CONFIG_SPARSEMEM
6750 * Calculate the size of the zone->blockflags rounded to an unsigned long
6751 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6752 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6753 * round what is now in bits to nearest long in bits, then return it in
6756 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6758 unsigned long usemapsize;
6760 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6761 usemapsize = roundup(zonesize, pageblock_nr_pages);
6762 usemapsize = usemapsize >> pageblock_order;
6763 usemapsize *= NR_PAGEBLOCK_BITS;
6764 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6766 return usemapsize / 8;
6769 static void __ref setup_usemap(struct pglist_data *pgdat,
6771 unsigned long zone_start_pfn,
6772 unsigned long zonesize)
6774 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6775 zone->pageblock_flags = NULL;
6777 zone->pageblock_flags =
6778 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6780 if (!zone->pageblock_flags)
6781 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6782 usemapsize, zone->name, pgdat->node_id);
6786 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6787 unsigned long zone_start_pfn, unsigned long zonesize) {}
6788 #endif /* CONFIG_SPARSEMEM */
6790 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6792 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6793 void __init set_pageblock_order(void)
6797 /* Check that pageblock_nr_pages has not already been setup */
6798 if (pageblock_order)
6801 if (HPAGE_SHIFT > PAGE_SHIFT)
6802 order = HUGETLB_PAGE_ORDER;
6804 order = MAX_ORDER - 1;
6807 * Assume the largest contiguous order of interest is a huge page.
6808 * This value may be variable depending on boot parameters on IA64 and
6811 pageblock_order = order;
6813 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6816 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6817 * is unused as pageblock_order is set at compile-time. See
6818 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6821 void __init set_pageblock_order(void)
6825 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6827 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6828 unsigned long present_pages)
6830 unsigned long pages = spanned_pages;
6833 * Provide a more accurate estimation if there are holes within
6834 * the zone and SPARSEMEM is in use. If there are holes within the
6835 * zone, each populated memory region may cost us one or two extra
6836 * memmap pages due to alignment because memmap pages for each
6837 * populated regions may not be naturally aligned on page boundary.
6838 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6840 if (spanned_pages > present_pages + (present_pages >> 4) &&
6841 IS_ENABLED(CONFIG_SPARSEMEM))
6842 pages = present_pages;
6844 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6847 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6848 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6850 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6852 spin_lock_init(&ds_queue->split_queue_lock);
6853 INIT_LIST_HEAD(&ds_queue->split_queue);
6854 ds_queue->split_queue_len = 0;
6857 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6860 #ifdef CONFIG_COMPACTION
6861 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6863 init_waitqueue_head(&pgdat->kcompactd_wait);
6866 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6869 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6871 pgdat_resize_init(pgdat);
6873 pgdat_init_split_queue(pgdat);
6874 pgdat_init_kcompactd(pgdat);
6876 init_waitqueue_head(&pgdat->kswapd_wait);
6877 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6879 pgdat_page_ext_init(pgdat);
6880 lruvec_init(&pgdat->__lruvec);
6883 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6884 unsigned long remaining_pages)
6886 atomic_long_set(&zone->managed_pages, remaining_pages);
6887 zone_set_nid(zone, nid);
6888 zone->name = zone_names[idx];
6889 zone->zone_pgdat = NODE_DATA(nid);
6890 spin_lock_init(&zone->lock);
6891 zone_seqlock_init(zone);
6892 zone_pcp_init(zone);
6896 * Set up the zone data structures
6897 * - init pgdat internals
6898 * - init all zones belonging to this node
6900 * NOTE: this function is only called during memory hotplug
6902 #ifdef CONFIG_MEMORY_HOTPLUG
6903 void __ref free_area_init_core_hotplug(int nid)
6906 pg_data_t *pgdat = NODE_DATA(nid);
6908 pgdat_init_internals(pgdat);
6909 for (z = 0; z < MAX_NR_ZONES; z++)
6910 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6915 * Set up the zone data structures:
6916 * - mark all pages reserved
6917 * - mark all memory queues empty
6918 * - clear the memory bitmaps
6920 * NOTE: pgdat should get zeroed by caller.
6921 * NOTE: this function is only called during early init.
6923 static void __init free_area_init_core(struct pglist_data *pgdat)
6926 int nid = pgdat->node_id;
6928 pgdat_init_internals(pgdat);
6929 pgdat->per_cpu_nodestats = &boot_nodestats;
6931 for (j = 0; j < MAX_NR_ZONES; j++) {
6932 struct zone *zone = pgdat->node_zones + j;
6933 unsigned long size, freesize, memmap_pages;
6934 unsigned long zone_start_pfn = zone->zone_start_pfn;
6936 size = zone->spanned_pages;
6937 freesize = zone->present_pages;
6940 * Adjust freesize so that it accounts for how much memory
6941 * is used by this zone for memmap. This affects the watermark
6942 * and per-cpu initialisations
6944 memmap_pages = calc_memmap_size(size, freesize);
6945 if (!is_highmem_idx(j)) {
6946 if (freesize >= memmap_pages) {
6947 freesize -= memmap_pages;
6950 " %s zone: %lu pages used for memmap\n",
6951 zone_names[j], memmap_pages);
6953 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6954 zone_names[j], memmap_pages, freesize);
6957 /* Account for reserved pages */
6958 if (j == 0 && freesize > dma_reserve) {
6959 freesize -= dma_reserve;
6960 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6961 zone_names[0], dma_reserve);
6964 if (!is_highmem_idx(j))
6965 nr_kernel_pages += freesize;
6966 /* Charge for highmem memmap if there are enough kernel pages */
6967 else if (nr_kernel_pages > memmap_pages * 2)
6968 nr_kernel_pages -= memmap_pages;
6969 nr_all_pages += freesize;
6972 * Set an approximate value for lowmem here, it will be adjusted
6973 * when the bootmem allocator frees pages into the buddy system.
6974 * And all highmem pages will be managed by the buddy system.
6976 zone_init_internals(zone, j, nid, freesize);
6981 set_pageblock_order();
6982 setup_usemap(pgdat, zone, zone_start_pfn, size);
6983 init_currently_empty_zone(zone, zone_start_pfn, size);
6984 memmap_init(size, nid, j, zone_start_pfn);
6988 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6989 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6991 unsigned long __maybe_unused start = 0;
6992 unsigned long __maybe_unused offset = 0;
6994 /* Skip empty nodes */
6995 if (!pgdat->node_spanned_pages)
6998 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6999 offset = pgdat->node_start_pfn - start;
7000 /* ia64 gets its own node_mem_map, before this, without bootmem */
7001 if (!pgdat->node_mem_map) {
7002 unsigned long size, end;
7006 * The zone's endpoints aren't required to be MAX_ORDER
7007 * aligned but the node_mem_map endpoints must be in order
7008 * for the buddy allocator to function correctly.
7010 end = pgdat_end_pfn(pgdat);
7011 end = ALIGN(end, MAX_ORDER_NR_PAGES);
7012 size = (end - start) * sizeof(struct page);
7013 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
7016 panic("Failed to allocate %ld bytes for node %d memory map\n",
7017 size, pgdat->node_id);
7018 pgdat->node_mem_map = map + offset;
7020 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
7021 __func__, pgdat->node_id, (unsigned long)pgdat,
7022 (unsigned long)pgdat->node_mem_map);
7023 #ifndef CONFIG_NEED_MULTIPLE_NODES
7025 * With no DISCONTIG, the global mem_map is just set as node 0's
7027 if (pgdat == NODE_DATA(0)) {
7028 mem_map = NODE_DATA(0)->node_mem_map;
7029 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
7035 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
7036 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
7038 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
7039 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
7041 pgdat->first_deferred_pfn = ULONG_MAX;
7044 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
7047 static void __init free_area_init_node(int nid)
7049 pg_data_t *pgdat = NODE_DATA(nid);
7050 unsigned long start_pfn = 0;
7051 unsigned long end_pfn = 0;
7053 /* pg_data_t should be reset to zero when it's allocated */
7054 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
7056 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7058 pgdat->node_id = nid;
7059 pgdat->node_start_pfn = start_pfn;
7060 pgdat->per_cpu_nodestats = NULL;
7062 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
7063 (u64)start_pfn << PAGE_SHIFT,
7064 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
7065 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
7067 alloc_node_mem_map(pgdat);
7068 pgdat_set_deferred_range(pgdat);
7070 free_area_init_core(pgdat);
7073 void __init free_area_init_memoryless_node(int nid)
7075 free_area_init_node(nid);
7078 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
7080 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
7081 * PageReserved(). Return the number of struct pages that were initialized.
7083 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
7088 for (pfn = spfn; pfn < epfn; pfn++) {
7089 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
7090 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
7091 + pageblock_nr_pages - 1;
7095 * Use a fake node/zone (0) for now. Some of these pages
7096 * (in memblock.reserved but not in memblock.memory) will
7097 * get re-initialized via reserve_bootmem_region() later.
7099 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
7100 __SetPageReserved(pfn_to_page(pfn));
7108 * Only struct pages that are backed by physical memory are zeroed and
7109 * initialized by going through __init_single_page(). But, there are some
7110 * struct pages which are reserved in memblock allocator and their fields
7111 * may be accessed (for example page_to_pfn() on some configuration accesses
7112 * flags). We must explicitly initialize those struct pages.
7114 * This function also addresses a similar issue where struct pages are left
7115 * uninitialized because the physical address range is not covered by
7116 * memblock.memory or memblock.reserved. That could happen when memblock
7117 * layout is manually configured via memmap=, or when the highest physical
7118 * address (max_pfn) does not end on a section boundary.
7120 static void __init init_unavailable_mem(void)
7122 phys_addr_t start, end;
7124 phys_addr_t next = 0;
7127 * Loop through unavailable ranges not covered by memblock.memory.
7130 for_each_mem_range(i, &start, &end) {
7132 pgcnt += init_unavailable_range(PFN_DOWN(next),
7138 * Early sections always have a fully populated memmap for the whole
7139 * section - see pfn_valid(). If the last section has holes at the
7140 * end and that section is marked "online", the memmap will be
7141 * considered initialized. Make sure that memmap has a well defined
7144 pgcnt += init_unavailable_range(PFN_DOWN(next),
7145 round_up(max_pfn, PAGES_PER_SECTION));
7148 * Struct pages that do not have backing memory. This could be because
7149 * firmware is using some of this memory, or for some other reasons.
7152 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7155 static inline void __init init_unavailable_mem(void)
7158 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7160 #if MAX_NUMNODES > 1
7162 * Figure out the number of possible node ids.
7164 void __init setup_nr_node_ids(void)
7166 unsigned int highest;
7168 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7169 nr_node_ids = highest + 1;
7174 * node_map_pfn_alignment - determine the maximum internode alignment
7176 * This function should be called after node map is populated and sorted.
7177 * It calculates the maximum power of two alignment which can distinguish
7180 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7181 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7182 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7183 * shifted, 1GiB is enough and this function will indicate so.
7185 * This is used to test whether pfn -> nid mapping of the chosen memory
7186 * model has fine enough granularity to avoid incorrect mapping for the
7187 * populated node map.
7189 * Return: the determined alignment in pfn's. 0 if there is no alignment
7190 * requirement (single node).
7192 unsigned long __init node_map_pfn_alignment(void)
7194 unsigned long accl_mask = 0, last_end = 0;
7195 unsigned long start, end, mask;
7196 int last_nid = NUMA_NO_NODE;
7199 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7200 if (!start || last_nid < 0 || last_nid == nid) {
7207 * Start with a mask granular enough to pin-point to the
7208 * start pfn and tick off bits one-by-one until it becomes
7209 * too coarse to separate the current node from the last.
7211 mask = ~((1 << __ffs(start)) - 1);
7212 while (mask && last_end <= (start & (mask << 1)))
7215 /* accumulate all internode masks */
7219 /* convert mask to number of pages */
7220 return ~accl_mask + 1;
7224 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7226 * Return: the minimum PFN based on information provided via
7227 * memblock_set_node().
7229 unsigned long __init find_min_pfn_with_active_regions(void)
7231 return PHYS_PFN(memblock_start_of_DRAM());
7235 * early_calculate_totalpages()
7236 * Sum pages in active regions for movable zone.
7237 * Populate N_MEMORY for calculating usable_nodes.
7239 static unsigned long __init early_calculate_totalpages(void)
7241 unsigned long totalpages = 0;
7242 unsigned long start_pfn, end_pfn;
7245 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7246 unsigned long pages = end_pfn - start_pfn;
7248 totalpages += pages;
7250 node_set_state(nid, N_MEMORY);
7256 * Find the PFN the Movable zone begins in each node. Kernel memory
7257 * is spread evenly between nodes as long as the nodes have enough
7258 * memory. When they don't, some nodes will have more kernelcore than
7261 static void __init find_zone_movable_pfns_for_nodes(void)
7264 unsigned long usable_startpfn;
7265 unsigned long kernelcore_node, kernelcore_remaining;
7266 /* save the state before borrow the nodemask */
7267 nodemask_t saved_node_state = node_states[N_MEMORY];
7268 unsigned long totalpages = early_calculate_totalpages();
7269 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7270 struct memblock_region *r;
7272 /* Need to find movable_zone earlier when movable_node is specified. */
7273 find_usable_zone_for_movable();
7276 * If movable_node is specified, ignore kernelcore and movablecore
7279 if (movable_node_is_enabled()) {
7280 for_each_mem_region(r) {
7281 if (!memblock_is_hotpluggable(r))
7284 nid = memblock_get_region_node(r);
7286 usable_startpfn = PFN_DOWN(r->base);
7287 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7288 min(usable_startpfn, zone_movable_pfn[nid]) :
7296 * If kernelcore=mirror is specified, ignore movablecore option
7298 if (mirrored_kernelcore) {
7299 bool mem_below_4gb_not_mirrored = false;
7301 for_each_mem_region(r) {
7302 if (memblock_is_mirror(r))
7305 nid = memblock_get_region_node(r);
7307 usable_startpfn = memblock_region_memory_base_pfn(r);
7309 if (usable_startpfn < 0x100000) {
7310 mem_below_4gb_not_mirrored = true;
7314 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7315 min(usable_startpfn, zone_movable_pfn[nid]) :
7319 if (mem_below_4gb_not_mirrored)
7320 pr_warn("This configuration results in unmirrored kernel memory.\n");
7326 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7327 * amount of necessary memory.
7329 if (required_kernelcore_percent)
7330 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7332 if (required_movablecore_percent)
7333 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7337 * If movablecore= was specified, calculate what size of
7338 * kernelcore that corresponds so that memory usable for
7339 * any allocation type is evenly spread. If both kernelcore
7340 * and movablecore are specified, then the value of kernelcore
7341 * will be used for required_kernelcore if it's greater than
7342 * what movablecore would have allowed.
7344 if (required_movablecore) {
7345 unsigned long corepages;
7348 * Round-up so that ZONE_MOVABLE is at least as large as what
7349 * was requested by the user
7351 required_movablecore =
7352 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7353 required_movablecore = min(totalpages, required_movablecore);
7354 corepages = totalpages - required_movablecore;
7356 required_kernelcore = max(required_kernelcore, corepages);
7360 * If kernelcore was not specified or kernelcore size is larger
7361 * than totalpages, there is no ZONE_MOVABLE.
7363 if (!required_kernelcore || required_kernelcore >= totalpages)
7366 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7367 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7370 /* Spread kernelcore memory as evenly as possible throughout nodes */
7371 kernelcore_node = required_kernelcore / usable_nodes;
7372 for_each_node_state(nid, N_MEMORY) {
7373 unsigned long start_pfn, end_pfn;
7376 * Recalculate kernelcore_node if the division per node
7377 * now exceeds what is necessary to satisfy the requested
7378 * amount of memory for the kernel
7380 if (required_kernelcore < kernelcore_node)
7381 kernelcore_node = required_kernelcore / usable_nodes;
7384 * As the map is walked, we track how much memory is usable
7385 * by the kernel using kernelcore_remaining. When it is
7386 * 0, the rest of the node is usable by ZONE_MOVABLE
7388 kernelcore_remaining = kernelcore_node;
7390 /* Go through each range of PFNs within this node */
7391 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7392 unsigned long size_pages;
7394 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7395 if (start_pfn >= end_pfn)
7398 /* Account for what is only usable for kernelcore */
7399 if (start_pfn < usable_startpfn) {
7400 unsigned long kernel_pages;
7401 kernel_pages = min(end_pfn, usable_startpfn)
7404 kernelcore_remaining -= min(kernel_pages,
7405 kernelcore_remaining);
7406 required_kernelcore -= min(kernel_pages,
7407 required_kernelcore);
7409 /* Continue if range is now fully accounted */
7410 if (end_pfn <= usable_startpfn) {
7413 * Push zone_movable_pfn to the end so
7414 * that if we have to rebalance
7415 * kernelcore across nodes, we will
7416 * not double account here
7418 zone_movable_pfn[nid] = end_pfn;
7421 start_pfn = usable_startpfn;
7425 * The usable PFN range for ZONE_MOVABLE is from
7426 * start_pfn->end_pfn. Calculate size_pages as the
7427 * number of pages used as kernelcore
7429 size_pages = end_pfn - start_pfn;
7430 if (size_pages > kernelcore_remaining)
7431 size_pages = kernelcore_remaining;
7432 zone_movable_pfn[nid] = start_pfn + size_pages;
7435 * Some kernelcore has been met, update counts and
7436 * break if the kernelcore for this node has been
7439 required_kernelcore -= min(required_kernelcore,
7441 kernelcore_remaining -= size_pages;
7442 if (!kernelcore_remaining)
7448 * If there is still required_kernelcore, we do another pass with one
7449 * less node in the count. This will push zone_movable_pfn[nid] further
7450 * along on the nodes that still have memory until kernelcore is
7454 if (usable_nodes && required_kernelcore > usable_nodes)
7458 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7459 for (nid = 0; nid < MAX_NUMNODES; nid++)
7460 zone_movable_pfn[nid] =
7461 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7464 /* restore the node_state */
7465 node_states[N_MEMORY] = saved_node_state;
7468 /* Any regular or high memory on that node ? */
7469 static void check_for_memory(pg_data_t *pgdat, int nid)
7471 enum zone_type zone_type;
7473 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7474 struct zone *zone = &pgdat->node_zones[zone_type];
7475 if (populated_zone(zone)) {
7476 if (IS_ENABLED(CONFIG_HIGHMEM))
7477 node_set_state(nid, N_HIGH_MEMORY);
7478 if (zone_type <= ZONE_NORMAL)
7479 node_set_state(nid, N_NORMAL_MEMORY);
7486 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7487 * such cases we allow max_zone_pfn sorted in the descending order
7489 bool __weak arch_has_descending_max_zone_pfns(void)
7495 * free_area_init - Initialise all pg_data_t and zone data
7496 * @max_zone_pfn: an array of max PFNs for each zone
7498 * This will call free_area_init_node() for each active node in the system.
7499 * Using the page ranges provided by memblock_set_node(), the size of each
7500 * zone in each node and their holes is calculated. If the maximum PFN
7501 * between two adjacent zones match, it is assumed that the zone is empty.
7502 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7503 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7504 * starts where the previous one ended. For example, ZONE_DMA32 starts
7505 * at arch_max_dma_pfn.
7507 void __init free_area_init(unsigned long *max_zone_pfn)
7509 unsigned long start_pfn, end_pfn;
7513 /* Record where the zone boundaries are */
7514 memset(arch_zone_lowest_possible_pfn, 0,
7515 sizeof(arch_zone_lowest_possible_pfn));
7516 memset(arch_zone_highest_possible_pfn, 0,
7517 sizeof(arch_zone_highest_possible_pfn));
7519 start_pfn = find_min_pfn_with_active_regions();
7520 descending = arch_has_descending_max_zone_pfns();
7522 for (i = 0; i < MAX_NR_ZONES; i++) {
7524 zone = MAX_NR_ZONES - i - 1;
7528 if (zone == ZONE_MOVABLE)
7531 end_pfn = max(max_zone_pfn[zone], start_pfn);
7532 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7533 arch_zone_highest_possible_pfn[zone] = end_pfn;
7535 start_pfn = end_pfn;
7538 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7539 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7540 find_zone_movable_pfns_for_nodes();
7542 /* Print out the zone ranges */
7543 pr_info("Zone ranges:\n");
7544 for (i = 0; i < MAX_NR_ZONES; i++) {
7545 if (i == ZONE_MOVABLE)
7547 pr_info(" %-8s ", zone_names[i]);
7548 if (arch_zone_lowest_possible_pfn[i] ==
7549 arch_zone_highest_possible_pfn[i])
7552 pr_cont("[mem %#018Lx-%#018Lx]\n",
7553 (u64)arch_zone_lowest_possible_pfn[i]
7555 ((u64)arch_zone_highest_possible_pfn[i]
7556 << PAGE_SHIFT) - 1);
7559 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7560 pr_info("Movable zone start for each node\n");
7561 for (i = 0; i < MAX_NUMNODES; i++) {
7562 if (zone_movable_pfn[i])
7563 pr_info(" Node %d: %#018Lx\n", i,
7564 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7568 * Print out the early node map, and initialize the
7569 * subsection-map relative to active online memory ranges to
7570 * enable future "sub-section" extensions of the memory map.
7572 pr_info("Early memory node ranges\n");
7573 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7574 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7575 (u64)start_pfn << PAGE_SHIFT,
7576 ((u64)end_pfn << PAGE_SHIFT) - 1);
7577 subsection_map_init(start_pfn, end_pfn - start_pfn);
7580 /* Initialise every node */
7581 mminit_verify_pageflags_layout();
7582 setup_nr_node_ids();
7583 init_unavailable_mem();
7584 for_each_online_node(nid) {
7585 pg_data_t *pgdat = NODE_DATA(nid);
7586 free_area_init_node(nid);
7588 /* Any memory on that node */
7589 if (pgdat->node_present_pages)
7590 node_set_state(nid, N_MEMORY);
7591 check_for_memory(pgdat, nid);
7595 static int __init cmdline_parse_core(char *p, unsigned long *core,
7596 unsigned long *percent)
7598 unsigned long long coremem;
7604 /* Value may be a percentage of total memory, otherwise bytes */
7605 coremem = simple_strtoull(p, &endptr, 0);
7606 if (*endptr == '%') {
7607 /* Paranoid check for percent values greater than 100 */
7608 WARN_ON(coremem > 100);
7612 coremem = memparse(p, &p);
7613 /* Paranoid check that UL is enough for the coremem value */
7614 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7616 *core = coremem >> PAGE_SHIFT;
7623 * kernelcore=size sets the amount of memory for use for allocations that
7624 * cannot be reclaimed or migrated.
7626 static int __init cmdline_parse_kernelcore(char *p)
7628 /* parse kernelcore=mirror */
7629 if (parse_option_str(p, "mirror")) {
7630 mirrored_kernelcore = true;
7634 return cmdline_parse_core(p, &required_kernelcore,
7635 &required_kernelcore_percent);
7639 * movablecore=size sets the amount of memory for use for allocations that
7640 * can be reclaimed or migrated.
7642 static int __init cmdline_parse_movablecore(char *p)
7644 return cmdline_parse_core(p, &required_movablecore,
7645 &required_movablecore_percent);
7648 early_param("kernelcore", cmdline_parse_kernelcore);
7649 early_param("movablecore", cmdline_parse_movablecore);
7651 void adjust_managed_page_count(struct page *page, long count)
7653 atomic_long_add(count, &page_zone(page)->managed_pages);
7654 totalram_pages_add(count);
7655 #ifdef CONFIG_HIGHMEM
7656 if (PageHighMem(page))
7657 totalhigh_pages_add(count);
7660 EXPORT_SYMBOL(adjust_managed_page_count);
7662 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7665 unsigned long pages = 0;
7667 start = (void *)PAGE_ALIGN((unsigned long)start);
7668 end = (void *)((unsigned long)end & PAGE_MASK);
7669 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7670 struct page *page = virt_to_page(pos);
7671 void *direct_map_addr;
7674 * 'direct_map_addr' might be different from 'pos'
7675 * because some architectures' virt_to_page()
7676 * work with aliases. Getting the direct map
7677 * address ensures that we get a _writeable_
7678 * alias for the memset().
7680 direct_map_addr = page_address(page);
7682 * Perform a kasan-unchecked memset() since this memory
7683 * has not been initialized.
7685 direct_map_addr = kasan_reset_tag(direct_map_addr);
7686 if ((unsigned int)poison <= 0xFF)
7687 memset(direct_map_addr, poison, PAGE_SIZE);
7689 free_reserved_page(page);
7693 pr_info("Freeing %s memory: %ldK\n",
7694 s, pages << (PAGE_SHIFT - 10));
7699 #ifdef CONFIG_HIGHMEM
7700 void free_highmem_page(struct page *page)
7702 __free_reserved_page(page);
7703 totalram_pages_inc();
7704 atomic_long_inc(&page_zone(page)->managed_pages);
7705 totalhigh_pages_inc();
7710 void __init mem_init_print_info(const char *str)
7712 unsigned long physpages, codesize, datasize, rosize, bss_size;
7713 unsigned long init_code_size, init_data_size;
7715 physpages = get_num_physpages();
7716 codesize = _etext - _stext;
7717 datasize = _edata - _sdata;
7718 rosize = __end_rodata - __start_rodata;
7719 bss_size = __bss_stop - __bss_start;
7720 init_data_size = __init_end - __init_begin;
7721 init_code_size = _einittext - _sinittext;
7724 * Detect special cases and adjust section sizes accordingly:
7725 * 1) .init.* may be embedded into .data sections
7726 * 2) .init.text.* may be out of [__init_begin, __init_end],
7727 * please refer to arch/tile/kernel/vmlinux.lds.S.
7728 * 3) .rodata.* may be embedded into .text or .data sections.
7730 #define adj_init_size(start, end, size, pos, adj) \
7732 if (start <= pos && pos < end && size > adj) \
7736 adj_init_size(__init_begin, __init_end, init_data_size,
7737 _sinittext, init_code_size);
7738 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7739 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7740 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7741 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7743 #undef adj_init_size
7745 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7746 #ifdef CONFIG_HIGHMEM
7750 nr_free_pages() << (PAGE_SHIFT - 10),
7751 physpages << (PAGE_SHIFT - 10),
7752 codesize >> 10, datasize >> 10, rosize >> 10,
7753 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7754 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7755 totalcma_pages << (PAGE_SHIFT - 10),
7756 #ifdef CONFIG_HIGHMEM
7757 totalhigh_pages() << (PAGE_SHIFT - 10),
7759 str ? ", " : "", str ? str : "");
7763 * set_dma_reserve - set the specified number of pages reserved in the first zone
7764 * @new_dma_reserve: The number of pages to mark reserved
7766 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7767 * In the DMA zone, a significant percentage may be consumed by kernel image
7768 * and other unfreeable allocations which can skew the watermarks badly. This
7769 * function may optionally be used to account for unfreeable pages in the
7770 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7771 * smaller per-cpu batchsize.
7773 void __init set_dma_reserve(unsigned long new_dma_reserve)
7775 dma_reserve = new_dma_reserve;
7778 static int page_alloc_cpu_dead(unsigned int cpu)
7781 lru_add_drain_cpu(cpu);
7785 * Spill the event counters of the dead processor
7786 * into the current processors event counters.
7787 * This artificially elevates the count of the current
7790 vm_events_fold_cpu(cpu);
7793 * Zero the differential counters of the dead processor
7794 * so that the vm statistics are consistent.
7796 * This is only okay since the processor is dead and cannot
7797 * race with what we are doing.
7799 cpu_vm_stats_fold(cpu);
7804 int hashdist = HASHDIST_DEFAULT;
7806 static int __init set_hashdist(char *str)
7810 hashdist = simple_strtoul(str, &str, 0);
7813 __setup("hashdist=", set_hashdist);
7816 void __init page_alloc_init(void)
7821 if (num_node_state(N_MEMORY) == 1)
7825 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7826 "mm/page_alloc:dead", NULL,
7827 page_alloc_cpu_dead);
7832 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7833 * or min_free_kbytes changes.
7835 static void calculate_totalreserve_pages(void)
7837 struct pglist_data *pgdat;
7838 unsigned long reserve_pages = 0;
7839 enum zone_type i, j;
7841 for_each_online_pgdat(pgdat) {
7843 pgdat->totalreserve_pages = 0;
7845 for (i = 0; i < MAX_NR_ZONES; i++) {
7846 struct zone *zone = pgdat->node_zones + i;
7848 unsigned long managed_pages = zone_managed_pages(zone);
7850 /* Find valid and maximum lowmem_reserve in the zone */
7851 for (j = i; j < MAX_NR_ZONES; j++) {
7852 if (zone->lowmem_reserve[j] > max)
7853 max = zone->lowmem_reserve[j];
7856 /* we treat the high watermark as reserved pages. */
7857 max += high_wmark_pages(zone);
7859 if (max > managed_pages)
7860 max = managed_pages;
7862 pgdat->totalreserve_pages += max;
7864 reserve_pages += max;
7867 totalreserve_pages = reserve_pages;
7871 * setup_per_zone_lowmem_reserve - called whenever
7872 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7873 * has a correct pages reserved value, so an adequate number of
7874 * pages are left in the zone after a successful __alloc_pages().
7876 static void setup_per_zone_lowmem_reserve(void)
7878 struct pglist_data *pgdat;
7879 enum zone_type i, j;
7881 for_each_online_pgdat(pgdat) {
7882 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
7883 struct zone *zone = &pgdat->node_zones[i];
7884 int ratio = sysctl_lowmem_reserve_ratio[i];
7885 bool clear = !ratio || !zone_managed_pages(zone);
7886 unsigned long managed_pages = 0;
7888 for (j = i + 1; j < MAX_NR_ZONES; j++) {
7890 zone->lowmem_reserve[j] = 0;
7892 struct zone *upper_zone = &pgdat->node_zones[j];
7894 managed_pages += zone_managed_pages(upper_zone);
7895 zone->lowmem_reserve[j] = managed_pages / ratio;
7901 /* update totalreserve_pages */
7902 calculate_totalreserve_pages();
7905 static void __setup_per_zone_wmarks(void)
7907 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7908 unsigned long lowmem_pages = 0;
7910 unsigned long flags;
7912 /* Calculate total number of !ZONE_HIGHMEM pages */
7913 for_each_zone(zone) {
7914 if (!is_highmem(zone))
7915 lowmem_pages += zone_managed_pages(zone);
7918 for_each_zone(zone) {
7921 spin_lock_irqsave(&zone->lock, flags);
7922 tmp = (u64)pages_min * zone_managed_pages(zone);
7923 do_div(tmp, lowmem_pages);
7924 if (is_highmem(zone)) {
7926 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7927 * need highmem pages, so cap pages_min to a small
7930 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7931 * deltas control async page reclaim, and so should
7932 * not be capped for highmem.
7934 unsigned long min_pages;
7936 min_pages = zone_managed_pages(zone) / 1024;
7937 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7938 zone->_watermark[WMARK_MIN] = min_pages;
7941 * If it's a lowmem zone, reserve a number of pages
7942 * proportionate to the zone's size.
7944 zone->_watermark[WMARK_MIN] = tmp;
7948 * Set the kswapd watermarks distance according to the
7949 * scale factor in proportion to available memory, but
7950 * ensure a minimum size on small systems.
7952 tmp = max_t(u64, tmp >> 2,
7953 mult_frac(zone_managed_pages(zone),
7954 watermark_scale_factor, 10000));
7956 zone->watermark_boost = 0;
7957 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7958 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7960 spin_unlock_irqrestore(&zone->lock, flags);
7963 /* update totalreserve_pages */
7964 calculate_totalreserve_pages();
7968 * setup_per_zone_wmarks - called when min_free_kbytes changes
7969 * or when memory is hot-{added|removed}
7971 * Ensures that the watermark[min,low,high] values for each zone are set
7972 * correctly with respect to min_free_kbytes.
7974 void setup_per_zone_wmarks(void)
7976 static DEFINE_SPINLOCK(lock);
7979 __setup_per_zone_wmarks();
7984 * Initialise min_free_kbytes.
7986 * For small machines we want it small (128k min). For large machines
7987 * we want it large (256MB max). But it is not linear, because network
7988 * bandwidth does not increase linearly with machine size. We use
7990 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7991 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
8007 int __meminit init_per_zone_wmark_min(void)
8009 unsigned long lowmem_kbytes;
8010 int new_min_free_kbytes;
8012 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
8013 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
8015 if (new_min_free_kbytes > user_min_free_kbytes) {
8016 min_free_kbytes = new_min_free_kbytes;
8017 if (min_free_kbytes < 128)
8018 min_free_kbytes = 128;
8019 if (min_free_kbytes > 262144)
8020 min_free_kbytes = 262144;
8022 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
8023 new_min_free_kbytes, user_min_free_kbytes);
8025 setup_per_zone_wmarks();
8026 refresh_zone_stat_thresholds();
8027 setup_per_zone_lowmem_reserve();
8030 setup_min_unmapped_ratio();
8031 setup_min_slab_ratio();
8034 khugepaged_min_free_kbytes_update();
8038 postcore_initcall(init_per_zone_wmark_min)
8041 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
8042 * that we can call two helper functions whenever min_free_kbytes
8045 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
8046 void *buffer, size_t *length, loff_t *ppos)
8050 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8055 user_min_free_kbytes = min_free_kbytes;
8056 setup_per_zone_wmarks();
8061 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
8062 void *buffer, size_t *length, loff_t *ppos)
8066 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8071 setup_per_zone_wmarks();
8077 static void setup_min_unmapped_ratio(void)
8082 for_each_online_pgdat(pgdat)
8083 pgdat->min_unmapped_pages = 0;
8086 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
8087 sysctl_min_unmapped_ratio) / 100;
8091 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
8092 void *buffer, size_t *length, loff_t *ppos)
8096 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8100 setup_min_unmapped_ratio();
8105 static void setup_min_slab_ratio(void)
8110 for_each_online_pgdat(pgdat)
8111 pgdat->min_slab_pages = 0;
8114 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
8115 sysctl_min_slab_ratio) / 100;
8118 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8119 void *buffer, size_t *length, loff_t *ppos)
8123 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8127 setup_min_slab_ratio();
8134 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8135 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8136 * whenever sysctl_lowmem_reserve_ratio changes.
8138 * The reserve ratio obviously has absolutely no relation with the
8139 * minimum watermarks. The lowmem reserve ratio can only make sense
8140 * if in function of the boot time zone sizes.
8142 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8143 void *buffer, size_t *length, loff_t *ppos)
8147 proc_dointvec_minmax(table, write, buffer, length, ppos);
8149 for (i = 0; i < MAX_NR_ZONES; i++) {
8150 if (sysctl_lowmem_reserve_ratio[i] < 1)
8151 sysctl_lowmem_reserve_ratio[i] = 0;
8154 setup_per_zone_lowmem_reserve();
8159 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8160 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8161 * pagelist can have before it gets flushed back to buddy allocator.
8163 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8164 void *buffer, size_t *length, loff_t *ppos)
8167 int old_percpu_pagelist_fraction;
8170 mutex_lock(&pcp_batch_high_lock);
8171 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8173 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8174 if (!write || ret < 0)
8177 /* Sanity checking to avoid pcp imbalance */
8178 if (percpu_pagelist_fraction &&
8179 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8180 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8186 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8189 for_each_populated_zone(zone)
8190 zone_set_pageset_high_and_batch(zone);
8192 mutex_unlock(&pcp_batch_high_lock);
8196 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8198 * Returns the number of pages that arch has reserved but
8199 * is not known to alloc_large_system_hash().
8201 static unsigned long __init arch_reserved_kernel_pages(void)
8208 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8209 * machines. As memory size is increased the scale is also increased but at
8210 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8211 * quadruples the scale is increased by one, which means the size of hash table
8212 * only doubles, instead of quadrupling as well.
8213 * Because 32-bit systems cannot have large physical memory, where this scaling
8214 * makes sense, it is disabled on such platforms.
8216 #if __BITS_PER_LONG > 32
8217 #define ADAPT_SCALE_BASE (64ul << 30)
8218 #define ADAPT_SCALE_SHIFT 2
8219 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8223 * allocate a large system hash table from bootmem
8224 * - it is assumed that the hash table must contain an exact power-of-2
8225 * quantity of entries
8226 * - limit is the number of hash buckets, not the total allocation size
8228 void *__init alloc_large_system_hash(const char *tablename,
8229 unsigned long bucketsize,
8230 unsigned long numentries,
8233 unsigned int *_hash_shift,
8234 unsigned int *_hash_mask,
8235 unsigned long low_limit,
8236 unsigned long high_limit)
8238 unsigned long long max = high_limit;
8239 unsigned long log2qty, size;
8244 /* allow the kernel cmdline to have a say */
8246 /* round applicable memory size up to nearest megabyte */
8247 numentries = nr_kernel_pages;
8248 numentries -= arch_reserved_kernel_pages();
8250 /* It isn't necessary when PAGE_SIZE >= 1MB */
8251 if (PAGE_SHIFT < 20)
8252 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8254 #if __BITS_PER_LONG > 32
8256 unsigned long adapt;
8258 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8259 adapt <<= ADAPT_SCALE_SHIFT)
8264 /* limit to 1 bucket per 2^scale bytes of low memory */
8265 if (scale > PAGE_SHIFT)
8266 numentries >>= (scale - PAGE_SHIFT);
8268 numentries <<= (PAGE_SHIFT - scale);
8270 /* Make sure we've got at least a 0-order allocation.. */
8271 if (unlikely(flags & HASH_SMALL)) {
8272 /* Makes no sense without HASH_EARLY */
8273 WARN_ON(!(flags & HASH_EARLY));
8274 if (!(numentries >> *_hash_shift)) {
8275 numentries = 1UL << *_hash_shift;
8276 BUG_ON(!numentries);
8278 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8279 numentries = PAGE_SIZE / bucketsize;
8281 numentries = roundup_pow_of_two(numentries);
8283 /* limit allocation size to 1/16 total memory by default */
8285 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8286 do_div(max, bucketsize);
8288 max = min(max, 0x80000000ULL);
8290 if (numentries < low_limit)
8291 numentries = low_limit;
8292 if (numentries > max)
8295 log2qty = ilog2(numentries);
8297 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8300 size = bucketsize << log2qty;
8301 if (flags & HASH_EARLY) {
8302 if (flags & HASH_ZERO)
8303 table = memblock_alloc(size, SMP_CACHE_BYTES);
8305 table = memblock_alloc_raw(size,
8307 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8308 table = __vmalloc(size, gfp_flags);
8312 * If bucketsize is not a power-of-two, we may free
8313 * some pages at the end of hash table which
8314 * alloc_pages_exact() automatically does
8316 table = alloc_pages_exact(size, gfp_flags);
8317 kmemleak_alloc(table, size, 1, gfp_flags);
8319 } while (!table && size > PAGE_SIZE && --log2qty);
8322 panic("Failed to allocate %s hash table\n", tablename);
8324 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8325 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8326 virt ? "vmalloc" : "linear");
8329 *_hash_shift = log2qty;
8331 *_hash_mask = (1 << log2qty) - 1;
8337 * This function checks whether pageblock includes unmovable pages or not.
8339 * PageLRU check without isolation or lru_lock could race so that
8340 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8341 * check without lock_page also may miss some movable non-lru pages at
8342 * race condition. So you can't expect this function should be exact.
8344 * Returns a page without holding a reference. If the caller wants to
8345 * dereference that page (e.g., dumping), it has to make sure that it
8346 * cannot get removed (e.g., via memory unplug) concurrently.
8349 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8350 int migratetype, int flags)
8352 unsigned long iter = 0;
8353 unsigned long pfn = page_to_pfn(page);
8354 unsigned long offset = pfn % pageblock_nr_pages;
8356 if (is_migrate_cma_page(page)) {
8358 * CMA allocations (alloc_contig_range) really need to mark
8359 * isolate CMA pageblocks even when they are not movable in fact
8360 * so consider them movable here.
8362 if (is_migrate_cma(migratetype))
8368 for (; iter < pageblock_nr_pages - offset; iter++) {
8369 if (!pfn_valid_within(pfn + iter))
8372 page = pfn_to_page(pfn + iter);
8375 * Both, bootmem allocations and memory holes are marked
8376 * PG_reserved and are unmovable. We can even have unmovable
8377 * allocations inside ZONE_MOVABLE, for example when
8378 * specifying "movablecore".
8380 if (PageReserved(page))
8384 * If the zone is movable and we have ruled out all reserved
8385 * pages then it should be reasonably safe to assume the rest
8388 if (zone_idx(zone) == ZONE_MOVABLE)
8392 * Hugepages are not in LRU lists, but they're movable.
8393 * THPs are on the LRU, but need to be counted as #small pages.
8394 * We need not scan over tail pages because we don't
8395 * handle each tail page individually in migration.
8397 if (PageHuge(page) || PageTransCompound(page)) {
8398 struct page *head = compound_head(page);
8399 unsigned int skip_pages;
8401 if (PageHuge(page)) {
8402 if (!hugepage_migration_supported(page_hstate(head)))
8404 } else if (!PageLRU(head) && !__PageMovable(head)) {
8408 skip_pages = compound_nr(head) - (page - head);
8409 iter += skip_pages - 1;
8414 * We can't use page_count without pin a page
8415 * because another CPU can free compound page.
8416 * This check already skips compound tails of THP
8417 * because their page->_refcount is zero at all time.
8419 if (!page_ref_count(page)) {
8420 if (PageBuddy(page))
8421 iter += (1 << buddy_order(page)) - 1;
8426 * The HWPoisoned page may be not in buddy system, and
8427 * page_count() is not 0.
8429 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8433 * We treat all PageOffline() pages as movable when offlining
8434 * to give drivers a chance to decrement their reference count
8435 * in MEM_GOING_OFFLINE in order to indicate that these pages
8436 * can be offlined as there are no direct references anymore.
8437 * For actually unmovable PageOffline() where the driver does
8438 * not support this, we will fail later when trying to actually
8439 * move these pages that still have a reference count > 0.
8440 * (false negatives in this function only)
8442 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8445 if (__PageMovable(page) || PageLRU(page))
8449 * If there are RECLAIMABLE pages, we need to check
8450 * it. But now, memory offline itself doesn't call
8451 * shrink_node_slabs() and it still to be fixed.
8458 #ifdef CONFIG_CONTIG_ALLOC
8459 static unsigned long pfn_max_align_down(unsigned long pfn)
8461 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8462 pageblock_nr_pages) - 1);
8465 static unsigned long pfn_max_align_up(unsigned long pfn)
8467 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8468 pageblock_nr_pages));
8471 /* [start, end) must belong to a single zone. */
8472 static int __alloc_contig_migrate_range(struct compact_control *cc,
8473 unsigned long start, unsigned long end)
8475 /* This function is based on compact_zone() from compaction.c. */
8476 unsigned int nr_reclaimed;
8477 unsigned long pfn = start;
8478 unsigned int tries = 0;
8480 struct migration_target_control mtc = {
8481 .nid = zone_to_nid(cc->zone),
8482 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8487 while (pfn < end || !list_empty(&cc->migratepages)) {
8488 if (fatal_signal_pending(current)) {
8493 if (list_empty(&cc->migratepages)) {
8494 cc->nr_migratepages = 0;
8495 pfn = isolate_migratepages_range(cc, pfn, end);
8501 } else if (++tries == 5) {
8502 ret = ret < 0 ? ret : -EBUSY;
8506 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8508 cc->nr_migratepages -= nr_reclaimed;
8510 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8511 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8514 putback_movable_pages(&cc->migratepages);
8521 * alloc_contig_range() -- tries to allocate given range of pages
8522 * @start: start PFN to allocate
8523 * @end: one-past-the-last PFN to allocate
8524 * @migratetype: migratetype of the underlaying pageblocks (either
8525 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8526 * in range must have the same migratetype and it must
8527 * be either of the two.
8528 * @gfp_mask: GFP mask to use during compaction
8530 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8531 * aligned. The PFN range must belong to a single zone.
8533 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8534 * pageblocks in the range. Once isolated, the pageblocks should not
8535 * be modified by others.
8537 * Return: zero on success or negative error code. On success all
8538 * pages which PFN is in [start, end) are allocated for the caller and
8539 * need to be freed with free_contig_range().
8541 int alloc_contig_range(unsigned long start, unsigned long end,
8542 unsigned migratetype, gfp_t gfp_mask)
8544 unsigned long outer_start, outer_end;
8548 struct compact_control cc = {
8549 .nr_migratepages = 0,
8551 .zone = page_zone(pfn_to_page(start)),
8552 .mode = MIGRATE_SYNC,
8553 .ignore_skip_hint = true,
8554 .no_set_skip_hint = true,
8555 .gfp_mask = current_gfp_context(gfp_mask),
8556 .alloc_contig = true,
8558 INIT_LIST_HEAD(&cc.migratepages);
8561 * What we do here is we mark all pageblocks in range as
8562 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8563 * have different sizes, and due to the way page allocator
8564 * work, we align the range to biggest of the two pages so
8565 * that page allocator won't try to merge buddies from
8566 * different pageblocks and change MIGRATE_ISOLATE to some
8567 * other migration type.
8569 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8570 * migrate the pages from an unaligned range (ie. pages that
8571 * we are interested in). This will put all the pages in
8572 * range back to page allocator as MIGRATE_ISOLATE.
8574 * When this is done, we take the pages in range from page
8575 * allocator removing them from the buddy system. This way
8576 * page allocator will never consider using them.
8578 * This lets us mark the pageblocks back as
8579 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8580 * aligned range but not in the unaligned, original range are
8581 * put back to page allocator so that buddy can use them.
8584 ret = start_isolate_page_range(pfn_max_align_down(start),
8585 pfn_max_align_up(end), migratetype, 0);
8589 drain_all_pages(cc.zone);
8592 * In case of -EBUSY, we'd like to know which page causes problem.
8593 * So, just fall through. test_pages_isolated() has a tracepoint
8594 * which will report the busy page.
8596 * It is possible that busy pages could become available before
8597 * the call to test_pages_isolated, and the range will actually be
8598 * allocated. So, if we fall through be sure to clear ret so that
8599 * -EBUSY is not accidentally used or returned to caller.
8601 ret = __alloc_contig_migrate_range(&cc, start, end);
8602 if (ret && ret != -EBUSY)
8607 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8608 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8609 * more, all pages in [start, end) are free in page allocator.
8610 * What we are going to do is to allocate all pages from
8611 * [start, end) (that is remove them from page allocator).
8613 * The only problem is that pages at the beginning and at the
8614 * end of interesting range may be not aligned with pages that
8615 * page allocator holds, ie. they can be part of higher order
8616 * pages. Because of this, we reserve the bigger range and
8617 * once this is done free the pages we are not interested in.
8619 * We don't have to hold zone->lock here because the pages are
8620 * isolated thus they won't get removed from buddy.
8623 lru_add_drain_all();
8626 outer_start = start;
8627 while (!PageBuddy(pfn_to_page(outer_start))) {
8628 if (++order >= MAX_ORDER) {
8629 outer_start = start;
8632 outer_start &= ~0UL << order;
8635 if (outer_start != start) {
8636 order = buddy_order(pfn_to_page(outer_start));
8639 * outer_start page could be small order buddy page and
8640 * it doesn't include start page. Adjust outer_start
8641 * in this case to report failed page properly
8642 * on tracepoint in test_pages_isolated()
8644 if (outer_start + (1UL << order) <= start)
8645 outer_start = start;
8648 /* Make sure the range is really isolated. */
8649 if (test_pages_isolated(outer_start, end, 0)) {
8650 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8651 __func__, outer_start, end);
8656 /* Grab isolated pages from freelists. */
8657 outer_end = isolate_freepages_range(&cc, outer_start, end);
8663 /* Free head and tail (if any) */
8664 if (start != outer_start)
8665 free_contig_range(outer_start, start - outer_start);
8666 if (end != outer_end)
8667 free_contig_range(end, outer_end - end);
8670 undo_isolate_page_range(pfn_max_align_down(start),
8671 pfn_max_align_up(end), migratetype);
8674 EXPORT_SYMBOL(alloc_contig_range);
8676 static int __alloc_contig_pages(unsigned long start_pfn,
8677 unsigned long nr_pages, gfp_t gfp_mask)
8679 unsigned long end_pfn = start_pfn + nr_pages;
8681 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8685 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8686 unsigned long nr_pages)
8688 unsigned long i, end_pfn = start_pfn + nr_pages;
8691 for (i = start_pfn; i < end_pfn; i++) {
8692 page = pfn_to_online_page(i);
8696 if (page_zone(page) != z)
8699 if (PageReserved(page))
8702 if (page_count(page) > 0)
8711 static bool zone_spans_last_pfn(const struct zone *zone,
8712 unsigned long start_pfn, unsigned long nr_pages)
8714 unsigned long last_pfn = start_pfn + nr_pages - 1;
8716 return zone_spans_pfn(zone, last_pfn);
8720 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8721 * @nr_pages: Number of contiguous pages to allocate
8722 * @gfp_mask: GFP mask to limit search and used during compaction
8724 * @nodemask: Mask for other possible nodes
8726 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8727 * on an applicable zonelist to find a contiguous pfn range which can then be
8728 * tried for allocation with alloc_contig_range(). This routine is intended
8729 * for allocation requests which can not be fulfilled with the buddy allocator.
8731 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8732 * power of two then the alignment is guaranteed to be to the given nr_pages
8733 * (e.g. 1GB request would be aligned to 1GB).
8735 * Allocated pages can be freed with free_contig_range() or by manually calling
8736 * __free_page() on each allocated page.
8738 * Return: pointer to contiguous pages on success, or NULL if not successful.
8740 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8741 int nid, nodemask_t *nodemask)
8743 unsigned long ret, pfn, flags;
8744 struct zonelist *zonelist;
8748 zonelist = node_zonelist(nid, gfp_mask);
8749 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8750 gfp_zone(gfp_mask), nodemask) {
8751 spin_lock_irqsave(&zone->lock, flags);
8753 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8754 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8755 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8757 * We release the zone lock here because
8758 * alloc_contig_range() will also lock the zone
8759 * at some point. If there's an allocation
8760 * spinning on this lock, it may win the race
8761 * and cause alloc_contig_range() to fail...
8763 spin_unlock_irqrestore(&zone->lock, flags);
8764 ret = __alloc_contig_pages(pfn, nr_pages,
8767 return pfn_to_page(pfn);
8768 spin_lock_irqsave(&zone->lock, flags);
8772 spin_unlock_irqrestore(&zone->lock, flags);
8776 #endif /* CONFIG_CONTIG_ALLOC */
8778 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8780 unsigned int count = 0;
8782 for (; nr_pages--; pfn++) {
8783 struct page *page = pfn_to_page(pfn);
8785 count += page_count(page) != 1;
8788 WARN(count != 0, "%d pages are still in use!\n", count);
8790 EXPORT_SYMBOL(free_contig_range);
8793 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8794 * page high values need to be recalulated.
8796 void __meminit zone_pcp_update(struct zone *zone)
8798 mutex_lock(&pcp_batch_high_lock);
8799 zone_set_pageset_high_and_batch(zone);
8800 mutex_unlock(&pcp_batch_high_lock);
8804 * Effectively disable pcplists for the zone by setting the high limit to 0
8805 * and draining all cpus. A concurrent page freeing on another CPU that's about
8806 * to put the page on pcplist will either finish before the drain and the page
8807 * will be drained, or observe the new high limit and skip the pcplist.
8809 * Must be paired with a call to zone_pcp_enable().
8811 void zone_pcp_disable(struct zone *zone)
8813 mutex_lock(&pcp_batch_high_lock);
8814 __zone_set_pageset_high_and_batch(zone, 0, 1);
8815 __drain_all_pages(zone, true);
8818 void zone_pcp_enable(struct zone *zone)
8820 __zone_set_pageset_high_and_batch(zone, zone->pageset_high, zone->pageset_batch);
8821 mutex_unlock(&pcp_batch_high_lock);
8824 void zone_pcp_reset(struct zone *zone)
8826 unsigned long flags;
8828 struct per_cpu_pageset *pset;
8830 /* avoid races with drain_pages() */
8831 local_irq_save(flags);
8832 if (zone->pageset != &boot_pageset) {
8833 for_each_online_cpu(cpu) {
8834 pset = per_cpu_ptr(zone->pageset, cpu);
8835 drain_zonestat(zone, pset);
8837 free_percpu(zone->pageset);
8838 zone->pageset = &boot_pageset;
8840 local_irq_restore(flags);
8843 #ifdef CONFIG_MEMORY_HOTREMOVE
8845 * All pages in the range must be in a single zone, must not contain holes,
8846 * must span full sections, and must be isolated before calling this function.
8848 void __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8850 unsigned long pfn = start_pfn;
8854 unsigned long flags;
8856 offline_mem_sections(pfn, end_pfn);
8857 zone = page_zone(pfn_to_page(pfn));
8858 spin_lock_irqsave(&zone->lock, flags);
8859 while (pfn < end_pfn) {
8860 page = pfn_to_page(pfn);
8862 * The HWPoisoned page may be not in buddy system, and
8863 * page_count() is not 0.
8865 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8870 * At this point all remaining PageOffline() pages have a
8871 * reference count of 0 and can simply be skipped.
8873 if (PageOffline(page)) {
8874 BUG_ON(page_count(page));
8875 BUG_ON(PageBuddy(page));
8880 BUG_ON(page_count(page));
8881 BUG_ON(!PageBuddy(page));
8882 order = buddy_order(page);
8883 del_page_from_free_list(page, zone, order);
8884 pfn += (1 << order);
8886 spin_unlock_irqrestore(&zone->lock, flags);
8890 bool is_free_buddy_page(struct page *page)
8892 struct zone *zone = page_zone(page);
8893 unsigned long pfn = page_to_pfn(page);
8894 unsigned long flags;
8897 spin_lock_irqsave(&zone->lock, flags);
8898 for (order = 0; order < MAX_ORDER; order++) {
8899 struct page *page_head = page - (pfn & ((1 << order) - 1));
8901 if (PageBuddy(page_head) && buddy_order(page_head) >= order)
8904 spin_unlock_irqrestore(&zone->lock, flags);
8906 return order < MAX_ORDER;
8909 #ifdef CONFIG_MEMORY_FAILURE
8911 * Break down a higher-order page in sub-pages, and keep our target out of
8914 static void break_down_buddy_pages(struct zone *zone, struct page *page,
8915 struct page *target, int low, int high,
8918 unsigned long size = 1 << high;
8919 struct page *current_buddy, *next_page;
8921 while (high > low) {
8925 if (target >= &page[size]) {
8926 next_page = page + size;
8927 current_buddy = page;
8930 current_buddy = page + size;
8933 if (set_page_guard(zone, current_buddy, high, migratetype))
8936 if (current_buddy != target) {
8937 add_to_free_list(current_buddy, zone, high, migratetype);
8938 set_buddy_order(current_buddy, high);
8945 * Take a page that will be marked as poisoned off the buddy allocator.
8947 bool take_page_off_buddy(struct page *page)
8949 struct zone *zone = page_zone(page);
8950 unsigned long pfn = page_to_pfn(page);
8951 unsigned long flags;
8955 spin_lock_irqsave(&zone->lock, flags);
8956 for (order = 0; order < MAX_ORDER; order++) {
8957 struct page *page_head = page - (pfn & ((1 << order) - 1));
8958 int page_order = buddy_order(page_head);
8960 if (PageBuddy(page_head) && page_order >= order) {
8961 unsigned long pfn_head = page_to_pfn(page_head);
8962 int migratetype = get_pfnblock_migratetype(page_head,
8965 del_page_from_free_list(page_head, zone, page_order);
8966 break_down_buddy_pages(zone, page_head, page, 0,
8967 page_order, migratetype);
8971 if (page_count(page_head) > 0)
8974 spin_unlock_irqrestore(&zone->lock, flags);