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/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/psi.h>
71 #include <linux/padata.h>
72 #include <linux/khugepaged.h>
74 #include <asm/sections.h>
75 #include <asm/tlbflush.h>
76 #include <asm/div64.h>
79 #include "page_reporting.h"
81 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
82 static DEFINE_MUTEX(pcp_batch_high_lock);
83 #define MIN_PERCPU_PAGELIST_FRACTION (8)
85 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
86 DEFINE_PER_CPU(int, numa_node);
87 EXPORT_PER_CPU_SYMBOL(numa_node);
90 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
92 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
94 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
95 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
96 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
97 * defined in <linux/topology.h>.
99 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
100 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
103 /* work_structs for global per-cpu drains */
106 struct work_struct work;
108 static DEFINE_MUTEX(pcpu_drain_mutex);
109 static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
111 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
112 volatile unsigned long latent_entropy __latent_entropy;
113 EXPORT_SYMBOL(latent_entropy);
117 * Array of node states.
119 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
120 [N_POSSIBLE] = NODE_MASK_ALL,
121 [N_ONLINE] = { { [0] = 1UL } },
123 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
124 #ifdef CONFIG_HIGHMEM
125 [N_HIGH_MEMORY] = { { [0] = 1UL } },
127 [N_MEMORY] = { { [0] = 1UL } },
128 [N_CPU] = { { [0] = 1UL } },
131 EXPORT_SYMBOL(node_states);
133 atomic_long_t _totalram_pages __read_mostly;
134 EXPORT_SYMBOL(_totalram_pages);
135 unsigned long totalreserve_pages __read_mostly;
136 unsigned long totalcma_pages __read_mostly;
138 int percpu_pagelist_fraction;
139 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
140 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
141 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
143 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
145 EXPORT_SYMBOL(init_on_alloc);
147 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
148 DEFINE_STATIC_KEY_TRUE(init_on_free);
150 DEFINE_STATIC_KEY_FALSE(init_on_free);
152 EXPORT_SYMBOL(init_on_free);
154 static int __init early_init_on_alloc(char *buf)
161 ret = kstrtobool(buf, &bool_result);
162 if (bool_result && page_poisoning_enabled())
163 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
165 static_branch_enable(&init_on_alloc);
167 static_branch_disable(&init_on_alloc);
170 early_param("init_on_alloc", early_init_on_alloc);
172 static int __init early_init_on_free(char *buf)
179 ret = kstrtobool(buf, &bool_result);
180 if (bool_result && page_poisoning_enabled())
181 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
183 static_branch_enable(&init_on_free);
185 static_branch_disable(&init_on_free);
188 early_param("init_on_free", early_init_on_free);
191 * A cached value of the page's pageblock's migratetype, used when the page is
192 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
193 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
194 * Also the migratetype set in the page does not necessarily match the pcplist
195 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
196 * other index - this ensures that it will be put on the correct CMA freelist.
198 static inline int get_pcppage_migratetype(struct page *page)
203 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
205 page->index = migratetype;
208 #ifdef CONFIG_PM_SLEEP
210 * The following functions are used by the suspend/hibernate code to temporarily
211 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
212 * while devices are suspended. To avoid races with the suspend/hibernate code,
213 * they should always be called with system_transition_mutex held
214 * (gfp_allowed_mask also should only be modified with system_transition_mutex
215 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
216 * with that modification).
219 static gfp_t saved_gfp_mask;
221 void pm_restore_gfp_mask(void)
223 WARN_ON(!mutex_is_locked(&system_transition_mutex));
224 if (saved_gfp_mask) {
225 gfp_allowed_mask = saved_gfp_mask;
230 void pm_restrict_gfp_mask(void)
232 WARN_ON(!mutex_is_locked(&system_transition_mutex));
233 WARN_ON(saved_gfp_mask);
234 saved_gfp_mask = gfp_allowed_mask;
235 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
238 bool pm_suspended_storage(void)
240 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
244 #endif /* CONFIG_PM_SLEEP */
246 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
247 unsigned int pageblock_order __read_mostly;
250 static void __free_pages_ok(struct page *page, unsigned int order);
253 * results with 256, 32 in the lowmem_reserve sysctl:
254 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
255 * 1G machine -> (16M dma, 784M normal, 224M high)
256 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
257 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
258 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
260 * TBD: should special case ZONE_DMA32 machines here - in those we normally
261 * don't need any ZONE_NORMAL reservation
263 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
264 #ifdef CONFIG_ZONE_DMA
267 #ifdef CONFIG_ZONE_DMA32
271 #ifdef CONFIG_HIGHMEM
277 static char * const zone_names[MAX_NR_ZONES] = {
278 #ifdef CONFIG_ZONE_DMA
281 #ifdef CONFIG_ZONE_DMA32
285 #ifdef CONFIG_HIGHMEM
289 #ifdef CONFIG_ZONE_DEVICE
294 const char * const migratetype_names[MIGRATE_TYPES] = {
302 #ifdef CONFIG_MEMORY_ISOLATION
307 compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
308 [NULL_COMPOUND_DTOR] = NULL,
309 [COMPOUND_PAGE_DTOR] = free_compound_page,
310 #ifdef CONFIG_HUGETLB_PAGE
311 [HUGETLB_PAGE_DTOR] = free_huge_page,
313 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
314 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
318 int min_free_kbytes = 1024;
319 int user_min_free_kbytes = -1;
320 #ifdef CONFIG_DISCONTIGMEM
322 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
323 * are not on separate NUMA nodes. Functionally this works but with
324 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
325 * quite small. By default, do not boost watermarks on discontigmem as in
326 * many cases very high-order allocations like THP are likely to be
327 * unsupported and the premature reclaim offsets the advantage of long-term
328 * fragmentation avoidance.
330 int watermark_boost_factor __read_mostly;
332 int watermark_boost_factor __read_mostly = 15000;
334 int watermark_scale_factor = 10;
336 static unsigned long nr_kernel_pages __initdata;
337 static unsigned long nr_all_pages __initdata;
338 static unsigned long dma_reserve __initdata;
340 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
341 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
342 static unsigned long required_kernelcore __initdata;
343 static unsigned long required_kernelcore_percent __initdata;
344 static unsigned long required_movablecore __initdata;
345 static unsigned long required_movablecore_percent __initdata;
346 static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
347 static bool mirrored_kernelcore __meminitdata;
349 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
351 EXPORT_SYMBOL(movable_zone);
354 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
355 unsigned int nr_online_nodes __read_mostly = 1;
356 EXPORT_SYMBOL(nr_node_ids);
357 EXPORT_SYMBOL(nr_online_nodes);
360 int page_group_by_mobility_disabled __read_mostly;
362 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
364 * During boot we initialize deferred pages on-demand, as needed, but once
365 * page_alloc_init_late() has finished, the deferred pages are all initialized,
366 * and we can permanently disable that path.
368 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
371 * Calling kasan_free_pages() only after deferred memory initialization
372 * has completed. Poisoning pages during deferred memory init will greatly
373 * lengthen the process and cause problem in large memory systems as the
374 * deferred pages initialization is done with interrupt disabled.
376 * Assuming that there will be no reference to those newly initialized
377 * pages before they are ever allocated, this should have no effect on
378 * KASAN memory tracking as the poison will be properly inserted at page
379 * allocation time. The only corner case is when pages are allocated by
380 * on-demand allocation and then freed again before the deferred pages
381 * initialization is done, but this is not likely to happen.
383 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
385 if (!static_branch_unlikely(&deferred_pages))
386 kasan_free_pages(page, order);
389 /* Returns true if the struct page for the pfn is uninitialised */
390 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
392 int nid = early_pfn_to_nid(pfn);
394 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
401 * Returns true when the remaining initialisation should be deferred until
402 * later in the boot cycle when it can be parallelised.
404 static bool __meminit
405 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
407 static unsigned long prev_end_pfn, nr_initialised;
410 * prev_end_pfn static that contains the end of previous zone
411 * No need to protect because called very early in boot before smp_init.
413 if (prev_end_pfn != end_pfn) {
414 prev_end_pfn = end_pfn;
418 /* Always populate low zones for address-constrained allocations */
419 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
423 * We start only with one section of pages, more pages are added as
424 * needed until the rest of deferred pages are initialized.
427 if ((nr_initialised > PAGES_PER_SECTION) &&
428 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
429 NODE_DATA(nid)->first_deferred_pfn = pfn;
435 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
437 static inline bool early_page_uninitialised(unsigned long pfn)
442 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
448 /* Return a pointer to the bitmap storing bits affecting a block of pages */
449 static inline unsigned long *get_pageblock_bitmap(struct page *page,
452 #ifdef CONFIG_SPARSEMEM
453 return section_to_usemap(__pfn_to_section(pfn));
455 return page_zone(page)->pageblock_flags;
456 #endif /* CONFIG_SPARSEMEM */
459 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
461 #ifdef CONFIG_SPARSEMEM
462 pfn &= (PAGES_PER_SECTION-1);
464 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
465 #endif /* CONFIG_SPARSEMEM */
466 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
470 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
471 * @page: The page within the block of interest
472 * @pfn: The target page frame number
473 * @mask: mask of bits that the caller is interested in
475 * Return: pageblock_bits flags
477 static __always_inline
478 unsigned long __get_pfnblock_flags_mask(struct page *page,
482 unsigned long *bitmap;
483 unsigned long bitidx, word_bitidx;
486 bitmap = get_pageblock_bitmap(page, pfn);
487 bitidx = pfn_to_bitidx(page, pfn);
488 word_bitidx = bitidx / BITS_PER_LONG;
489 bitidx &= (BITS_PER_LONG-1);
491 word = bitmap[word_bitidx];
492 return (word >> bitidx) & mask;
495 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
498 return __get_pfnblock_flags_mask(page, pfn, mask);
501 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
503 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
507 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
508 * @page: The page within the block of interest
509 * @flags: The flags to set
510 * @pfn: The target page frame number
511 * @mask: mask of bits that the caller is interested in
513 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
517 unsigned long *bitmap;
518 unsigned long bitidx, word_bitidx;
519 unsigned long old_word, word;
521 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
522 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
524 bitmap = get_pageblock_bitmap(page, pfn);
525 bitidx = pfn_to_bitidx(page, pfn);
526 word_bitidx = bitidx / BITS_PER_LONG;
527 bitidx &= (BITS_PER_LONG-1);
529 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
534 word = READ_ONCE(bitmap[word_bitidx]);
536 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
537 if (word == old_word)
543 void set_pageblock_migratetype(struct page *page, int migratetype)
545 if (unlikely(page_group_by_mobility_disabled &&
546 migratetype < MIGRATE_PCPTYPES))
547 migratetype = MIGRATE_UNMOVABLE;
549 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
550 page_to_pfn(page), MIGRATETYPE_MASK);
553 #ifdef CONFIG_DEBUG_VM
554 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
558 unsigned long pfn = page_to_pfn(page);
559 unsigned long sp, start_pfn;
562 seq = zone_span_seqbegin(zone);
563 start_pfn = zone->zone_start_pfn;
564 sp = zone->spanned_pages;
565 if (!zone_spans_pfn(zone, pfn))
567 } while (zone_span_seqretry(zone, seq));
570 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
571 pfn, zone_to_nid(zone), zone->name,
572 start_pfn, start_pfn + sp);
577 static int page_is_consistent(struct zone *zone, struct page *page)
579 if (!pfn_valid_within(page_to_pfn(page)))
581 if (zone != page_zone(page))
587 * Temporary debugging check for pages not lying within a given zone.
589 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
591 if (page_outside_zone_boundaries(zone, page))
593 if (!page_is_consistent(zone, page))
599 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
605 static void bad_page(struct page *page, const char *reason)
607 static unsigned long resume;
608 static unsigned long nr_shown;
609 static unsigned long nr_unshown;
612 * Allow a burst of 60 reports, then keep quiet for that minute;
613 * or allow a steady drip of one report per second.
615 if (nr_shown == 60) {
616 if (time_before(jiffies, resume)) {
622 "BUG: Bad page state: %lu messages suppressed\n",
629 resume = jiffies + 60 * HZ;
631 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
632 current->comm, page_to_pfn(page));
633 __dump_page(page, reason);
634 dump_page_owner(page);
639 /* Leave bad fields for debug, except PageBuddy could make trouble */
640 page_mapcount_reset(page); /* remove PageBuddy */
641 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
645 * Higher-order pages are called "compound pages". They are structured thusly:
647 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
649 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
650 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
652 * The first tail page's ->compound_dtor holds the offset in array of compound
653 * page destructors. See compound_page_dtors.
655 * The first tail page's ->compound_order holds the order of allocation.
656 * This usage means that zero-order pages may not be compound.
659 void free_compound_page(struct page *page)
661 mem_cgroup_uncharge(page);
662 __free_pages_ok(page, compound_order(page));
665 void prep_compound_page(struct page *page, unsigned int order)
668 int nr_pages = 1 << order;
671 for (i = 1; i < nr_pages; i++) {
672 struct page *p = page + i;
673 set_page_count(p, 0);
674 p->mapping = TAIL_MAPPING;
675 set_compound_head(p, page);
678 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
679 set_compound_order(page, order);
680 atomic_set(compound_mapcount_ptr(page), -1);
681 if (hpage_pincount_available(page))
682 atomic_set(compound_pincount_ptr(page), 0);
685 #ifdef CONFIG_DEBUG_PAGEALLOC
686 unsigned int _debug_guardpage_minorder;
688 bool _debug_pagealloc_enabled_early __read_mostly
689 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
690 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
691 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
692 EXPORT_SYMBOL(_debug_pagealloc_enabled);
694 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
696 static int __init early_debug_pagealloc(char *buf)
698 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
700 early_param("debug_pagealloc", early_debug_pagealloc);
702 void init_debug_pagealloc(void)
704 if (!debug_pagealloc_enabled())
707 static_branch_enable(&_debug_pagealloc_enabled);
709 if (!debug_guardpage_minorder())
712 static_branch_enable(&_debug_guardpage_enabled);
715 static int __init debug_guardpage_minorder_setup(char *buf)
719 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
720 pr_err("Bad debug_guardpage_minorder value\n");
723 _debug_guardpage_minorder = res;
724 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
727 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
729 static inline bool set_page_guard(struct zone *zone, struct page *page,
730 unsigned int order, int migratetype)
732 if (!debug_guardpage_enabled())
735 if (order >= debug_guardpage_minorder())
738 __SetPageGuard(page);
739 INIT_LIST_HEAD(&page->lru);
740 set_page_private(page, order);
741 /* Guard pages are not available for any usage */
742 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
747 static inline void clear_page_guard(struct zone *zone, struct page *page,
748 unsigned int order, int migratetype)
750 if (!debug_guardpage_enabled())
753 __ClearPageGuard(page);
755 set_page_private(page, 0);
756 if (!is_migrate_isolate(migratetype))
757 __mod_zone_freepage_state(zone, (1 << order), migratetype);
760 static inline bool set_page_guard(struct zone *zone, struct page *page,
761 unsigned int order, int migratetype) { return false; }
762 static inline void clear_page_guard(struct zone *zone, struct page *page,
763 unsigned int order, int migratetype) {}
766 static inline void set_page_order(struct page *page, unsigned int order)
768 set_page_private(page, order);
769 __SetPageBuddy(page);
773 * This function checks whether a page is free && is the buddy
774 * we can coalesce a page and its buddy if
775 * (a) the buddy is not in a hole (check before calling!) &&
776 * (b) the buddy is in the buddy system &&
777 * (c) a page and its buddy have the same order &&
778 * (d) a page and its buddy are in the same zone.
780 * For recording whether a page is in the buddy system, we set PageBuddy.
781 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
783 * For recording page's order, we use page_private(page).
785 static inline bool page_is_buddy(struct page *page, struct page *buddy,
788 if (!page_is_guard(buddy) && !PageBuddy(buddy))
791 if (page_order(buddy) != order)
795 * zone check is done late to avoid uselessly calculating
796 * zone/node ids for pages that could never merge.
798 if (page_zone_id(page) != page_zone_id(buddy))
801 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
806 #ifdef CONFIG_COMPACTION
807 static inline struct capture_control *task_capc(struct zone *zone)
809 struct capture_control *capc = current->capture_control;
811 return unlikely(capc) &&
812 !(current->flags & PF_KTHREAD) &&
814 capc->cc->zone == zone ? capc : NULL;
818 compaction_capture(struct capture_control *capc, struct page *page,
819 int order, int migratetype)
821 if (!capc || order != capc->cc->order)
824 /* Do not accidentally pollute CMA or isolated regions*/
825 if (is_migrate_cma(migratetype) ||
826 is_migrate_isolate(migratetype))
830 * Do not let lower order allocations polluate a movable pageblock.
831 * This might let an unmovable request use a reclaimable pageblock
832 * and vice-versa but no more than normal fallback logic which can
833 * have trouble finding a high-order free page.
835 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
843 static inline struct capture_control *task_capc(struct zone *zone)
849 compaction_capture(struct capture_control *capc, struct page *page,
850 int order, int migratetype)
854 #endif /* CONFIG_COMPACTION */
856 /* Used for pages not on another list */
857 static inline void add_to_free_list(struct page *page, struct zone *zone,
858 unsigned int order, int migratetype)
860 struct free_area *area = &zone->free_area[order];
862 list_add(&page->lru, &area->free_list[migratetype]);
866 /* Used for pages not on another list */
867 static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
868 unsigned int order, int migratetype)
870 struct free_area *area = &zone->free_area[order];
872 list_add_tail(&page->lru, &area->free_list[migratetype]);
876 /* Used for pages which are on another list */
877 static inline void move_to_free_list(struct page *page, struct zone *zone,
878 unsigned int order, int migratetype)
880 struct free_area *area = &zone->free_area[order];
882 list_move(&page->lru, &area->free_list[migratetype]);
885 static inline void del_page_from_free_list(struct page *page, struct zone *zone,
888 /* clear reported state and update reported page count */
889 if (page_reported(page))
890 __ClearPageReported(page);
892 list_del(&page->lru);
893 __ClearPageBuddy(page);
894 set_page_private(page, 0);
895 zone->free_area[order].nr_free--;
899 * If this is not the largest possible page, check if the buddy
900 * of the next-highest order is free. If it is, it's possible
901 * that pages are being freed that will coalesce soon. In case,
902 * that is happening, add the free page to the tail of the list
903 * so it's less likely to be used soon and more likely to be merged
904 * as a higher order page
907 buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
908 struct page *page, unsigned int order)
910 struct page *higher_page, *higher_buddy;
911 unsigned long combined_pfn;
913 if (order >= MAX_ORDER - 2)
916 if (!pfn_valid_within(buddy_pfn))
919 combined_pfn = buddy_pfn & pfn;
920 higher_page = page + (combined_pfn - pfn);
921 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
922 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
924 return pfn_valid_within(buddy_pfn) &&
925 page_is_buddy(higher_page, higher_buddy, order + 1);
929 * Freeing function for a buddy system allocator.
931 * The concept of a buddy system is to maintain direct-mapped table
932 * (containing bit values) for memory blocks of various "orders".
933 * The bottom level table contains the map for the smallest allocatable
934 * units of memory (here, pages), and each level above it describes
935 * pairs of units from the levels below, hence, "buddies".
936 * At a high level, all that happens here is marking the table entry
937 * at the bottom level available, and propagating the changes upward
938 * as necessary, plus some accounting needed to play nicely with other
939 * parts of the VM system.
940 * At each level, we keep a list of pages, which are heads of continuous
941 * free pages of length of (1 << order) and marked with PageBuddy.
942 * Page's order is recorded in page_private(page) field.
943 * So when we are allocating or freeing one, we can derive the state of the
944 * other. That is, if we allocate a small block, and both were
945 * free, the remainder of the region must be split into blocks.
946 * If a block is freed, and its buddy is also free, then this
947 * triggers coalescing into a block of larger size.
952 static inline void __free_one_page(struct page *page,
954 struct zone *zone, unsigned int order,
955 int migratetype, bool report)
957 struct capture_control *capc = task_capc(zone);
958 unsigned long buddy_pfn;
959 unsigned long combined_pfn;
960 unsigned int max_order;
964 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
966 VM_BUG_ON(!zone_is_initialized(zone));
967 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
969 VM_BUG_ON(migratetype == -1);
970 if (likely(!is_migrate_isolate(migratetype)))
971 __mod_zone_freepage_state(zone, 1 << order, migratetype);
973 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
974 VM_BUG_ON_PAGE(bad_range(zone, page), page);
977 while (order < max_order - 1) {
978 if (compaction_capture(capc, page, order, migratetype)) {
979 __mod_zone_freepage_state(zone, -(1 << order),
983 buddy_pfn = __find_buddy_pfn(pfn, order);
984 buddy = page + (buddy_pfn - pfn);
986 if (!pfn_valid_within(buddy_pfn))
988 if (!page_is_buddy(page, buddy, order))
991 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
992 * merge with it and move up one order.
994 if (page_is_guard(buddy))
995 clear_page_guard(zone, buddy, order, migratetype);
997 del_page_from_free_list(buddy, zone, order);
998 combined_pfn = buddy_pfn & pfn;
999 page = page + (combined_pfn - pfn);
1003 if (max_order < MAX_ORDER) {
1004 /* If we are here, it means order is >= pageblock_order.
1005 * We want to prevent merge between freepages on isolate
1006 * pageblock and normal pageblock. Without this, pageblock
1007 * isolation could cause incorrect freepage or CMA accounting.
1009 * We don't want to hit this code for the more frequent
1010 * low-order merging.
1012 if (unlikely(has_isolate_pageblock(zone))) {
1015 buddy_pfn = __find_buddy_pfn(pfn, order);
1016 buddy = page + (buddy_pfn - pfn);
1017 buddy_mt = get_pageblock_migratetype(buddy);
1019 if (migratetype != buddy_mt
1020 && (is_migrate_isolate(migratetype) ||
1021 is_migrate_isolate(buddy_mt)))
1025 goto continue_merging;
1029 set_page_order(page, order);
1031 if (is_shuffle_order(order))
1032 to_tail = shuffle_pick_tail();
1034 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1037 add_to_free_list_tail(page, zone, order, migratetype);
1039 add_to_free_list(page, zone, order, migratetype);
1041 /* Notify page reporting subsystem of freed page */
1043 page_reporting_notify_free(order);
1047 * A bad page could be due to a number of fields. Instead of multiple branches,
1048 * try and check multiple fields with one check. The caller must do a detailed
1049 * check if necessary.
1051 static inline bool page_expected_state(struct page *page,
1052 unsigned long check_flags)
1054 if (unlikely(atomic_read(&page->_mapcount) != -1))
1057 if (unlikely((unsigned long)page->mapping |
1058 page_ref_count(page) |
1060 (unsigned long)page->mem_cgroup |
1062 (page->flags & check_flags)))
1068 static const char *page_bad_reason(struct page *page, unsigned long flags)
1070 const char *bad_reason = NULL;
1072 if (unlikely(atomic_read(&page->_mapcount) != -1))
1073 bad_reason = "nonzero mapcount";
1074 if (unlikely(page->mapping != NULL))
1075 bad_reason = "non-NULL mapping";
1076 if (unlikely(page_ref_count(page) != 0))
1077 bad_reason = "nonzero _refcount";
1078 if (unlikely(page->flags & flags)) {
1079 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1080 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1082 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1085 if (unlikely(page->mem_cgroup))
1086 bad_reason = "page still charged to cgroup";
1091 static void check_free_page_bad(struct page *page)
1094 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1097 static inline int check_free_page(struct page *page)
1099 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1102 /* Something has gone sideways, find it */
1103 check_free_page_bad(page);
1107 static int free_tail_pages_check(struct page *head_page, struct page *page)
1112 * We rely page->lru.next never has bit 0 set, unless the page
1113 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1115 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1117 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1121 switch (page - head_page) {
1123 /* the first tail page: ->mapping may be compound_mapcount() */
1124 if (unlikely(compound_mapcount(page))) {
1125 bad_page(page, "nonzero compound_mapcount");
1131 * the second tail page: ->mapping is
1132 * deferred_list.next -- ignore value.
1136 if (page->mapping != TAIL_MAPPING) {
1137 bad_page(page, "corrupted mapping in tail page");
1142 if (unlikely(!PageTail(page))) {
1143 bad_page(page, "PageTail not set");
1146 if (unlikely(compound_head(page) != head_page)) {
1147 bad_page(page, "compound_head not consistent");
1152 page->mapping = NULL;
1153 clear_compound_head(page);
1157 static void kernel_init_free_pages(struct page *page, int numpages)
1161 /* s390's use of memset() could override KASAN redzones. */
1162 kasan_disable_current();
1163 for (i = 0; i < numpages; i++)
1164 clear_highpage(page + i);
1165 kasan_enable_current();
1168 static __always_inline bool free_pages_prepare(struct page *page,
1169 unsigned int order, bool check_free)
1173 VM_BUG_ON_PAGE(PageTail(page), page);
1175 trace_mm_page_free(page, order);
1178 * Check tail pages before head page information is cleared to
1179 * avoid checking PageCompound for order-0 pages.
1181 if (unlikely(order)) {
1182 bool compound = PageCompound(page);
1185 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1188 ClearPageDoubleMap(page);
1189 for (i = 1; i < (1 << order); i++) {
1191 bad += free_tail_pages_check(page, page + i);
1192 if (unlikely(check_free_page(page + i))) {
1196 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1199 if (PageMappingFlags(page))
1200 page->mapping = NULL;
1201 if (memcg_kmem_enabled() && PageKmemcg(page))
1202 __memcg_kmem_uncharge_page(page, order);
1204 bad += check_free_page(page);
1208 page_cpupid_reset_last(page);
1209 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1210 reset_page_owner(page, order);
1212 if (!PageHighMem(page)) {
1213 debug_check_no_locks_freed(page_address(page),
1214 PAGE_SIZE << order);
1215 debug_check_no_obj_freed(page_address(page),
1216 PAGE_SIZE << order);
1218 if (want_init_on_free())
1219 kernel_init_free_pages(page, 1 << order);
1221 kernel_poison_pages(page, 1 << order, 0);
1223 * arch_free_page() can make the page's contents inaccessible. s390
1224 * does this. So nothing which can access the page's contents should
1225 * happen after this.
1227 arch_free_page(page, order);
1229 if (debug_pagealloc_enabled_static())
1230 kernel_map_pages(page, 1 << order, 0);
1232 kasan_free_nondeferred_pages(page, order);
1237 #ifdef CONFIG_DEBUG_VM
1239 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1240 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1241 * moved from pcp lists to free lists.
1243 static bool free_pcp_prepare(struct page *page)
1245 return free_pages_prepare(page, 0, true);
1248 static bool bulkfree_pcp_prepare(struct page *page)
1250 if (debug_pagealloc_enabled_static())
1251 return check_free_page(page);
1257 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1258 * moving from pcp lists to free list in order to reduce overhead. With
1259 * debug_pagealloc enabled, they are checked also immediately when being freed
1262 static bool free_pcp_prepare(struct page *page)
1264 if (debug_pagealloc_enabled_static())
1265 return free_pages_prepare(page, 0, true);
1267 return free_pages_prepare(page, 0, false);
1270 static bool bulkfree_pcp_prepare(struct page *page)
1272 return check_free_page(page);
1274 #endif /* CONFIG_DEBUG_VM */
1276 static inline void prefetch_buddy(struct page *page)
1278 unsigned long pfn = page_to_pfn(page);
1279 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1280 struct page *buddy = page + (buddy_pfn - pfn);
1286 * Frees a number of pages from the PCP lists
1287 * Assumes all pages on list are in same zone, and of same order.
1288 * count is the number of pages to free.
1290 * If the zone was previously in an "all pages pinned" state then look to
1291 * see if this freeing clears that state.
1293 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1294 * pinned" detection logic.
1296 static void free_pcppages_bulk(struct zone *zone, int count,
1297 struct per_cpu_pages *pcp)
1299 int migratetype = 0;
1301 int prefetch_nr = 0;
1302 bool isolated_pageblocks;
1303 struct page *page, *tmp;
1307 * Ensure proper count is passed which otherwise would stuck in the
1308 * below while (list_empty(list)) loop.
1310 count = min(pcp->count, count);
1312 struct list_head *list;
1315 * Remove pages from lists in a round-robin fashion. A
1316 * batch_free count is maintained that is incremented when an
1317 * empty list is encountered. This is so more pages are freed
1318 * off fuller lists instead of spinning excessively around empty
1323 if (++migratetype == MIGRATE_PCPTYPES)
1325 list = &pcp->lists[migratetype];
1326 } while (list_empty(list));
1328 /* This is the only non-empty list. Free them all. */
1329 if (batch_free == MIGRATE_PCPTYPES)
1333 page = list_last_entry(list, struct page, lru);
1334 /* must delete to avoid corrupting pcp list */
1335 list_del(&page->lru);
1338 if (bulkfree_pcp_prepare(page))
1341 list_add_tail(&page->lru, &head);
1344 * We are going to put the page back to the global
1345 * pool, prefetch its buddy to speed up later access
1346 * under zone->lock. It is believed the overhead of
1347 * an additional test and calculating buddy_pfn here
1348 * can be offset by reduced memory latency later. To
1349 * avoid excessive prefetching due to large count, only
1350 * prefetch buddy for the first pcp->batch nr of pages.
1352 if (prefetch_nr++ < pcp->batch)
1353 prefetch_buddy(page);
1354 } while (--count && --batch_free && !list_empty(list));
1357 spin_lock(&zone->lock);
1358 isolated_pageblocks = has_isolate_pageblock(zone);
1361 * Use safe version since after __free_one_page(),
1362 * page->lru.next will not point to original list.
1364 list_for_each_entry_safe(page, tmp, &head, lru) {
1365 int mt = get_pcppage_migratetype(page);
1366 /* MIGRATE_ISOLATE page should not go to pcplists */
1367 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1368 /* Pageblock could have been isolated meanwhile */
1369 if (unlikely(isolated_pageblocks))
1370 mt = get_pageblock_migratetype(page);
1372 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1373 trace_mm_page_pcpu_drain(page, 0, mt);
1375 spin_unlock(&zone->lock);
1378 static void free_one_page(struct zone *zone,
1379 struct page *page, unsigned long pfn,
1383 spin_lock(&zone->lock);
1384 if (unlikely(has_isolate_pageblock(zone) ||
1385 is_migrate_isolate(migratetype))) {
1386 migratetype = get_pfnblock_migratetype(page, pfn);
1388 __free_one_page(page, pfn, zone, order, migratetype, true);
1389 spin_unlock(&zone->lock);
1392 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1393 unsigned long zone, int nid)
1395 mm_zero_struct_page(page);
1396 set_page_links(page, zone, nid, pfn);
1397 init_page_count(page);
1398 page_mapcount_reset(page);
1399 page_cpupid_reset_last(page);
1400 page_kasan_tag_reset(page);
1402 INIT_LIST_HEAD(&page->lru);
1403 #ifdef WANT_PAGE_VIRTUAL
1404 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1405 if (!is_highmem_idx(zone))
1406 set_page_address(page, __va(pfn << PAGE_SHIFT));
1410 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1411 static void __meminit init_reserved_page(unsigned long pfn)
1416 if (!early_page_uninitialised(pfn))
1419 nid = early_pfn_to_nid(pfn);
1420 pgdat = NODE_DATA(nid);
1422 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1423 struct zone *zone = &pgdat->node_zones[zid];
1425 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1428 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1431 static inline void init_reserved_page(unsigned long pfn)
1434 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1437 * Initialised pages do not have PageReserved set. This function is
1438 * called for each range allocated by the bootmem allocator and
1439 * marks the pages PageReserved. The remaining valid pages are later
1440 * sent to the buddy page allocator.
1442 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1444 unsigned long start_pfn = PFN_DOWN(start);
1445 unsigned long end_pfn = PFN_UP(end);
1447 for (; start_pfn < end_pfn; start_pfn++) {
1448 if (pfn_valid(start_pfn)) {
1449 struct page *page = pfn_to_page(start_pfn);
1451 init_reserved_page(start_pfn);
1453 /* Avoid false-positive PageTail() */
1454 INIT_LIST_HEAD(&page->lru);
1457 * no need for atomic set_bit because the struct
1458 * page is not visible yet so nobody should
1461 __SetPageReserved(page);
1466 static void __free_pages_ok(struct page *page, unsigned int order)
1468 unsigned long flags;
1470 unsigned long pfn = page_to_pfn(page);
1472 if (!free_pages_prepare(page, order, true))
1475 migratetype = get_pfnblock_migratetype(page, pfn);
1476 local_irq_save(flags);
1477 __count_vm_events(PGFREE, 1 << order);
1478 free_one_page(page_zone(page), page, pfn, order, migratetype);
1479 local_irq_restore(flags);
1482 void __free_pages_core(struct page *page, unsigned int order)
1484 unsigned int nr_pages = 1 << order;
1485 struct page *p = page;
1489 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1491 __ClearPageReserved(p);
1492 set_page_count(p, 0);
1494 __ClearPageReserved(p);
1495 set_page_count(p, 0);
1497 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1498 set_page_refcounted(page);
1499 __free_pages(page, order);
1502 #ifdef CONFIG_NEED_MULTIPLE_NODES
1504 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1506 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1509 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1511 int __meminit __early_pfn_to_nid(unsigned long pfn,
1512 struct mminit_pfnnid_cache *state)
1514 unsigned long start_pfn, end_pfn;
1517 if (state->last_start <= pfn && pfn < state->last_end)
1518 return state->last_nid;
1520 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1521 if (nid != NUMA_NO_NODE) {
1522 state->last_start = start_pfn;
1523 state->last_end = end_pfn;
1524 state->last_nid = nid;
1529 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1531 int __meminit early_pfn_to_nid(unsigned long pfn)
1533 static DEFINE_SPINLOCK(early_pfn_lock);
1536 spin_lock(&early_pfn_lock);
1537 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1539 nid = first_online_node;
1540 spin_unlock(&early_pfn_lock);
1544 #endif /* CONFIG_NEED_MULTIPLE_NODES */
1546 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1549 if (early_page_uninitialised(pfn))
1551 __free_pages_core(page, order);
1555 * Check that the whole (or subset of) a pageblock given by the interval of
1556 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1557 * with the migration of free compaction scanner. The scanners then need to
1558 * use only pfn_valid_within() check for arches that allow holes within
1561 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1563 * It's possible on some configurations to have a setup like node0 node1 node0
1564 * i.e. it's possible that all pages within a zones range of pages do not
1565 * belong to a single zone. We assume that a border between node0 and node1
1566 * can occur within a single pageblock, but not a node0 node1 node0
1567 * interleaving within a single pageblock. It is therefore sufficient to check
1568 * the first and last page of a pageblock and avoid checking each individual
1569 * page in a pageblock.
1571 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1572 unsigned long end_pfn, struct zone *zone)
1574 struct page *start_page;
1575 struct page *end_page;
1577 /* end_pfn is one past the range we are checking */
1580 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1583 start_page = pfn_to_online_page(start_pfn);
1587 if (page_zone(start_page) != zone)
1590 end_page = pfn_to_page(end_pfn);
1592 /* This gives a shorter code than deriving page_zone(end_page) */
1593 if (page_zone_id(start_page) != page_zone_id(end_page))
1599 void set_zone_contiguous(struct zone *zone)
1601 unsigned long block_start_pfn = zone->zone_start_pfn;
1602 unsigned long block_end_pfn;
1604 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1605 for (; block_start_pfn < zone_end_pfn(zone);
1606 block_start_pfn = block_end_pfn,
1607 block_end_pfn += pageblock_nr_pages) {
1609 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1611 if (!__pageblock_pfn_to_page(block_start_pfn,
1612 block_end_pfn, zone))
1617 /* We confirm that there is no hole */
1618 zone->contiguous = true;
1621 void clear_zone_contiguous(struct zone *zone)
1623 zone->contiguous = false;
1626 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1627 static void __init deferred_free_range(unsigned long pfn,
1628 unsigned long nr_pages)
1636 page = pfn_to_page(pfn);
1638 /* Free a large naturally-aligned chunk if possible */
1639 if (nr_pages == pageblock_nr_pages &&
1640 (pfn & (pageblock_nr_pages - 1)) == 0) {
1641 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1642 __free_pages_core(page, pageblock_order);
1646 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1647 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1648 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1649 __free_pages_core(page, 0);
1653 /* Completion tracking for deferred_init_memmap() threads */
1654 static atomic_t pgdat_init_n_undone __initdata;
1655 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1657 static inline void __init pgdat_init_report_one_done(void)
1659 if (atomic_dec_and_test(&pgdat_init_n_undone))
1660 complete(&pgdat_init_all_done_comp);
1664 * Returns true if page needs to be initialized or freed to buddy allocator.
1666 * First we check if pfn is valid on architectures where it is possible to have
1667 * holes within pageblock_nr_pages. On systems where it is not possible, this
1668 * function is optimized out.
1670 * Then, we check if a current large page is valid by only checking the validity
1673 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1675 if (!pfn_valid_within(pfn))
1677 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1683 * Free pages to buddy allocator. Try to free aligned pages in
1684 * pageblock_nr_pages sizes.
1686 static void __init deferred_free_pages(unsigned long pfn,
1687 unsigned long end_pfn)
1689 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1690 unsigned long nr_free = 0;
1692 for (; pfn < end_pfn; pfn++) {
1693 if (!deferred_pfn_valid(pfn)) {
1694 deferred_free_range(pfn - nr_free, nr_free);
1696 } else if (!(pfn & nr_pgmask)) {
1697 deferred_free_range(pfn - nr_free, nr_free);
1703 /* Free the last block of pages to allocator */
1704 deferred_free_range(pfn - nr_free, nr_free);
1708 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1709 * by performing it only once every pageblock_nr_pages.
1710 * Return number of pages initialized.
1712 static unsigned long __init deferred_init_pages(struct zone *zone,
1714 unsigned long end_pfn)
1716 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1717 int nid = zone_to_nid(zone);
1718 unsigned long nr_pages = 0;
1719 int zid = zone_idx(zone);
1720 struct page *page = NULL;
1722 for (; pfn < end_pfn; pfn++) {
1723 if (!deferred_pfn_valid(pfn)) {
1726 } else if (!page || !(pfn & nr_pgmask)) {
1727 page = pfn_to_page(pfn);
1731 __init_single_page(page, pfn, zid, nid);
1738 * This function is meant to pre-load the iterator for the zone init.
1739 * Specifically it walks through the ranges until we are caught up to the
1740 * first_init_pfn value and exits there. If we never encounter the value we
1741 * return false indicating there are no valid ranges left.
1744 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1745 unsigned long *spfn, unsigned long *epfn,
1746 unsigned long first_init_pfn)
1751 * Start out by walking through the ranges in this zone that have
1752 * already been initialized. We don't need to do anything with them
1753 * so we just need to flush them out of the system.
1755 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1756 if (*epfn <= first_init_pfn)
1758 if (*spfn < first_init_pfn)
1759 *spfn = first_init_pfn;
1768 * Initialize and free pages. We do it in two loops: first we initialize
1769 * struct page, then free to buddy allocator, because while we are
1770 * freeing pages we can access pages that are ahead (computing buddy
1771 * page in __free_one_page()).
1773 * In order to try and keep some memory in the cache we have the loop
1774 * broken along max page order boundaries. This way we will not cause
1775 * any issues with the buddy page computation.
1777 static unsigned long __init
1778 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1779 unsigned long *end_pfn)
1781 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1782 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1783 unsigned long nr_pages = 0;
1786 /* First we loop through and initialize the page values */
1787 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1790 if (mo_pfn <= *start_pfn)
1793 t = min(mo_pfn, *end_pfn);
1794 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1796 if (mo_pfn < *end_pfn) {
1797 *start_pfn = mo_pfn;
1802 /* Reset values and now loop through freeing pages as needed */
1805 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1811 t = min(mo_pfn, epfn);
1812 deferred_free_pages(spfn, t);
1822 deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1825 unsigned long spfn, epfn;
1826 struct zone *zone = arg;
1829 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1832 * Initialize and free pages in MAX_ORDER sized increments so that we
1833 * can avoid introducing any issues with the buddy allocator.
1835 while (spfn < end_pfn) {
1836 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1841 /* An arch may override for more concurrency. */
1843 deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1848 /* Initialise remaining memory on a node */
1849 static int __init deferred_init_memmap(void *data)
1851 pg_data_t *pgdat = data;
1852 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1853 unsigned long spfn = 0, epfn = 0;
1854 unsigned long first_init_pfn, flags;
1855 unsigned long start = jiffies;
1857 int zid, max_threads;
1860 /* Bind memory initialisation thread to a local node if possible */
1861 if (!cpumask_empty(cpumask))
1862 set_cpus_allowed_ptr(current, cpumask);
1864 pgdat_resize_lock(pgdat, &flags);
1865 first_init_pfn = pgdat->first_deferred_pfn;
1866 if (first_init_pfn == ULONG_MAX) {
1867 pgdat_resize_unlock(pgdat, &flags);
1868 pgdat_init_report_one_done();
1872 /* Sanity check boundaries */
1873 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1874 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1875 pgdat->first_deferred_pfn = ULONG_MAX;
1878 * Once we unlock here, the zone cannot be grown anymore, thus if an
1879 * interrupt thread must allocate this early in boot, zone must be
1880 * pre-grown prior to start of deferred page initialization.
1882 pgdat_resize_unlock(pgdat, &flags);
1884 /* Only the highest zone is deferred so find it */
1885 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1886 zone = pgdat->node_zones + zid;
1887 if (first_init_pfn < zone_end_pfn(zone))
1891 /* If the zone is empty somebody else may have cleared out the zone */
1892 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1896 max_threads = deferred_page_init_max_threads(cpumask);
1898 while (spfn < epfn) {
1899 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1900 struct padata_mt_job job = {
1901 .thread_fn = deferred_init_memmap_chunk,
1904 .size = epfn_align - spfn,
1905 .align = PAGES_PER_SECTION,
1906 .min_chunk = PAGES_PER_SECTION,
1907 .max_threads = max_threads,
1910 padata_do_multithreaded(&job);
1911 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1915 /* Sanity check that the next zone really is unpopulated */
1916 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1918 pr_info("node %d deferred pages initialised in %ums\n",
1919 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1921 pgdat_init_report_one_done();
1926 * If this zone has deferred pages, try to grow it by initializing enough
1927 * deferred pages to satisfy the allocation specified by order, rounded up to
1928 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1929 * of SECTION_SIZE bytes by initializing struct pages in increments of
1930 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1932 * Return true when zone was grown, otherwise return false. We return true even
1933 * when we grow less than requested, to let the caller decide if there are
1934 * enough pages to satisfy the allocation.
1936 * Note: We use noinline because this function is needed only during boot, and
1937 * it is called from a __ref function _deferred_grow_zone. This way we are
1938 * making sure that it is not inlined into permanent text section.
1940 static noinline bool __init
1941 deferred_grow_zone(struct zone *zone, unsigned int order)
1943 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1944 pg_data_t *pgdat = zone->zone_pgdat;
1945 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1946 unsigned long spfn, epfn, flags;
1947 unsigned long nr_pages = 0;
1950 /* Only the last zone may have deferred pages */
1951 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1954 pgdat_resize_lock(pgdat, &flags);
1957 * If someone grew this zone while we were waiting for spinlock, return
1958 * true, as there might be enough pages already.
1960 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1961 pgdat_resize_unlock(pgdat, &flags);
1965 /* If the zone is empty somebody else may have cleared out the zone */
1966 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1967 first_deferred_pfn)) {
1968 pgdat->first_deferred_pfn = ULONG_MAX;
1969 pgdat_resize_unlock(pgdat, &flags);
1970 /* Retry only once. */
1971 return first_deferred_pfn != ULONG_MAX;
1975 * Initialize and free pages in MAX_ORDER sized increments so
1976 * that we can avoid introducing any issues with the buddy
1979 while (spfn < epfn) {
1980 /* update our first deferred PFN for this section */
1981 first_deferred_pfn = spfn;
1983 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1984 touch_nmi_watchdog();
1986 /* We should only stop along section boundaries */
1987 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1990 /* If our quota has been met we can stop here */
1991 if (nr_pages >= nr_pages_needed)
1995 pgdat->first_deferred_pfn = spfn;
1996 pgdat_resize_unlock(pgdat, &flags);
1998 return nr_pages > 0;
2002 * deferred_grow_zone() is __init, but it is called from
2003 * get_page_from_freelist() during early boot until deferred_pages permanently
2004 * disables this call. This is why we have refdata wrapper to avoid warning,
2005 * and to ensure that the function body gets unloaded.
2008 _deferred_grow_zone(struct zone *zone, unsigned int order)
2010 return deferred_grow_zone(zone, order);
2013 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2015 void __init page_alloc_init_late(void)
2020 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2022 /* There will be num_node_state(N_MEMORY) threads */
2023 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2024 for_each_node_state(nid, N_MEMORY) {
2025 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2028 /* Block until all are initialised */
2029 wait_for_completion(&pgdat_init_all_done_comp);
2032 * The number of managed pages has changed due to the initialisation
2033 * so the pcpu batch and high limits needs to be updated or the limits
2034 * will be artificially small.
2036 for_each_populated_zone(zone)
2037 zone_pcp_update(zone);
2040 * We initialized the rest of the deferred pages. Permanently disable
2041 * on-demand struct page initialization.
2043 static_branch_disable(&deferred_pages);
2045 /* Reinit limits that are based on free pages after the kernel is up */
2046 files_maxfiles_init();
2049 /* Discard memblock private memory */
2052 for_each_node_state(nid, N_MEMORY)
2053 shuffle_free_memory(NODE_DATA(nid));
2055 for_each_populated_zone(zone)
2056 set_zone_contiguous(zone);
2060 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2061 void __init init_cma_reserved_pageblock(struct page *page)
2063 unsigned i = pageblock_nr_pages;
2064 struct page *p = page;
2067 __ClearPageReserved(p);
2068 set_page_count(p, 0);
2071 set_pageblock_migratetype(page, MIGRATE_CMA);
2073 if (pageblock_order >= MAX_ORDER) {
2074 i = pageblock_nr_pages;
2077 set_page_refcounted(p);
2078 __free_pages(p, MAX_ORDER - 1);
2079 p += MAX_ORDER_NR_PAGES;
2080 } while (i -= MAX_ORDER_NR_PAGES);
2082 set_page_refcounted(page);
2083 __free_pages(page, pageblock_order);
2086 adjust_managed_page_count(page, pageblock_nr_pages);
2091 * The order of subdivision here is critical for the IO subsystem.
2092 * Please do not alter this order without good reasons and regression
2093 * testing. Specifically, as large blocks of memory are subdivided,
2094 * the order in which smaller blocks are delivered depends on the order
2095 * they're subdivided in this function. This is the primary factor
2096 * influencing the order in which pages are delivered to the IO
2097 * subsystem according to empirical testing, and this is also justified
2098 * by considering the behavior of a buddy system containing a single
2099 * large block of memory acted on by a series of small allocations.
2100 * This behavior is a critical factor in sglist merging's success.
2104 static inline void expand(struct zone *zone, struct page *page,
2105 int low, int high, int migratetype)
2107 unsigned long size = 1 << high;
2109 while (high > low) {
2112 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2115 * Mark as guard pages (or page), that will allow to
2116 * merge back to allocator when buddy will be freed.
2117 * Corresponding page table entries will not be touched,
2118 * pages will stay not present in virtual address space
2120 if (set_page_guard(zone, &page[size], high, migratetype))
2123 add_to_free_list(&page[size], zone, high, migratetype);
2124 set_page_order(&page[size], high);
2128 static void check_new_page_bad(struct page *page)
2130 if (unlikely(page->flags & __PG_HWPOISON)) {
2131 /* Don't complain about hwpoisoned pages */
2132 page_mapcount_reset(page); /* remove PageBuddy */
2137 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2141 * This page is about to be returned from the page allocator
2143 static inline int check_new_page(struct page *page)
2145 if (likely(page_expected_state(page,
2146 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2149 check_new_page_bad(page);
2153 static inline bool free_pages_prezeroed(void)
2155 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2156 page_poisoning_enabled()) || want_init_on_free();
2159 #ifdef CONFIG_DEBUG_VM
2161 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2162 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2163 * also checked when pcp lists are refilled from the free lists.
2165 static inline bool check_pcp_refill(struct page *page)
2167 if (debug_pagealloc_enabled_static())
2168 return check_new_page(page);
2173 static inline bool check_new_pcp(struct page *page)
2175 return check_new_page(page);
2179 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2180 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2181 * enabled, they are also checked when being allocated from the pcp lists.
2183 static inline bool check_pcp_refill(struct page *page)
2185 return check_new_page(page);
2187 static inline bool check_new_pcp(struct page *page)
2189 if (debug_pagealloc_enabled_static())
2190 return check_new_page(page);
2194 #endif /* CONFIG_DEBUG_VM */
2196 static bool check_new_pages(struct page *page, unsigned int order)
2199 for (i = 0; i < (1 << order); i++) {
2200 struct page *p = page + i;
2202 if (unlikely(check_new_page(p)))
2209 inline void post_alloc_hook(struct page *page, unsigned int order,
2212 set_page_private(page, 0);
2213 set_page_refcounted(page);
2215 arch_alloc_page(page, order);
2216 if (debug_pagealloc_enabled_static())
2217 kernel_map_pages(page, 1 << order, 1);
2218 kasan_alloc_pages(page, order);
2219 kernel_poison_pages(page, 1 << order, 1);
2220 set_page_owner(page, order, gfp_flags);
2223 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2224 unsigned int alloc_flags)
2226 post_alloc_hook(page, order, gfp_flags);
2228 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2229 kernel_init_free_pages(page, 1 << order);
2231 if (order && (gfp_flags & __GFP_COMP))
2232 prep_compound_page(page, order);
2235 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2236 * allocate the page. The expectation is that the caller is taking
2237 * steps that will free more memory. The caller should avoid the page
2238 * being used for !PFMEMALLOC purposes.
2240 if (alloc_flags & ALLOC_NO_WATERMARKS)
2241 set_page_pfmemalloc(page);
2243 clear_page_pfmemalloc(page);
2247 * Go through the free lists for the given migratetype and remove
2248 * the smallest available page from the freelists
2250 static __always_inline
2251 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2254 unsigned int current_order;
2255 struct free_area *area;
2258 /* Find a page of the appropriate size in the preferred list */
2259 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2260 area = &(zone->free_area[current_order]);
2261 page = get_page_from_free_area(area, migratetype);
2264 del_page_from_free_list(page, zone, current_order);
2265 expand(zone, page, order, current_order, migratetype);
2266 set_pcppage_migratetype(page, migratetype);
2275 * This array describes the order lists are fallen back to when
2276 * the free lists for the desirable migrate type are depleted
2278 static int fallbacks[MIGRATE_TYPES][3] = {
2279 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2280 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2281 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2283 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2285 #ifdef CONFIG_MEMORY_ISOLATION
2286 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2291 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2294 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2297 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2298 unsigned int order) { return NULL; }
2302 * Move the free pages in a range to the free lists of the requested type.
2303 * Note that start_page and end_pages are not aligned on a pageblock
2304 * boundary. If alignment is required, use move_freepages_block()
2306 static int move_freepages(struct zone *zone,
2307 struct page *start_page, struct page *end_page,
2308 int migratetype, int *num_movable)
2312 int pages_moved = 0;
2314 for (page = start_page; page <= end_page;) {
2315 if (!pfn_valid_within(page_to_pfn(page))) {
2320 if (!PageBuddy(page)) {
2322 * We assume that pages that could be isolated for
2323 * migration are movable. But we don't actually try
2324 * isolating, as that would be expensive.
2327 (PageLRU(page) || __PageMovable(page)))
2334 /* Make sure we are not inadvertently changing nodes */
2335 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2336 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2338 order = page_order(page);
2339 move_to_free_list(page, zone, order, migratetype);
2341 pages_moved += 1 << order;
2347 int move_freepages_block(struct zone *zone, struct page *page,
2348 int migratetype, int *num_movable)
2350 unsigned long start_pfn, end_pfn;
2351 struct page *start_page, *end_page;
2356 start_pfn = page_to_pfn(page);
2357 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2358 start_page = pfn_to_page(start_pfn);
2359 end_page = start_page + pageblock_nr_pages - 1;
2360 end_pfn = start_pfn + pageblock_nr_pages - 1;
2362 /* Do not cross zone boundaries */
2363 if (!zone_spans_pfn(zone, start_pfn))
2365 if (!zone_spans_pfn(zone, end_pfn))
2368 return move_freepages(zone, start_page, end_page, migratetype,
2372 static void change_pageblock_range(struct page *pageblock_page,
2373 int start_order, int migratetype)
2375 int nr_pageblocks = 1 << (start_order - pageblock_order);
2377 while (nr_pageblocks--) {
2378 set_pageblock_migratetype(pageblock_page, migratetype);
2379 pageblock_page += pageblock_nr_pages;
2384 * When we are falling back to another migratetype during allocation, try to
2385 * steal extra free pages from the same pageblocks to satisfy further
2386 * allocations, instead of polluting multiple pageblocks.
2388 * If we are stealing a relatively large buddy page, it is likely there will
2389 * be more free pages in the pageblock, so try to steal them all. For
2390 * reclaimable and unmovable allocations, we steal regardless of page size,
2391 * as fragmentation caused by those allocations polluting movable pageblocks
2392 * is worse than movable allocations stealing from unmovable and reclaimable
2395 static bool can_steal_fallback(unsigned int order, int start_mt)
2398 * Leaving this order check is intended, although there is
2399 * relaxed order check in next check. The reason is that
2400 * we can actually steal whole pageblock if this condition met,
2401 * but, below check doesn't guarantee it and that is just heuristic
2402 * so could be changed anytime.
2404 if (order >= pageblock_order)
2407 if (order >= pageblock_order / 2 ||
2408 start_mt == MIGRATE_RECLAIMABLE ||
2409 start_mt == MIGRATE_UNMOVABLE ||
2410 page_group_by_mobility_disabled)
2416 static inline void boost_watermark(struct zone *zone)
2418 unsigned long max_boost;
2420 if (!watermark_boost_factor)
2423 * Don't bother in zones that are unlikely to produce results.
2424 * On small machines, including kdump capture kernels running
2425 * in a small area, boosting the watermark can cause an out of
2426 * memory situation immediately.
2428 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2431 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2432 watermark_boost_factor, 10000);
2435 * high watermark may be uninitialised if fragmentation occurs
2436 * very early in boot so do not boost. We do not fall
2437 * through and boost by pageblock_nr_pages as failing
2438 * allocations that early means that reclaim is not going
2439 * to help and it may even be impossible to reclaim the
2440 * boosted watermark resulting in a hang.
2445 max_boost = max(pageblock_nr_pages, max_boost);
2447 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2452 * This function implements actual steal behaviour. If order is large enough,
2453 * we can steal whole pageblock. If not, we first move freepages in this
2454 * pageblock to our migratetype and determine how many already-allocated pages
2455 * are there in the pageblock with a compatible migratetype. If at least half
2456 * of pages are free or compatible, we can change migratetype of the pageblock
2457 * itself, so pages freed in the future will be put on the correct free list.
2459 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2460 unsigned int alloc_flags, int start_type, bool whole_block)
2462 unsigned int current_order = page_order(page);
2463 int free_pages, movable_pages, alike_pages;
2466 old_block_type = get_pageblock_migratetype(page);
2469 * This can happen due to races and we want to prevent broken
2470 * highatomic accounting.
2472 if (is_migrate_highatomic(old_block_type))
2475 /* Take ownership for orders >= pageblock_order */
2476 if (current_order >= pageblock_order) {
2477 change_pageblock_range(page, current_order, start_type);
2482 * Boost watermarks to increase reclaim pressure to reduce the
2483 * likelihood of future fallbacks. Wake kswapd now as the node
2484 * may be balanced overall and kswapd will not wake naturally.
2486 boost_watermark(zone);
2487 if (alloc_flags & ALLOC_KSWAPD)
2488 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2490 /* We are not allowed to try stealing from the whole block */
2494 free_pages = move_freepages_block(zone, page, start_type,
2497 * Determine how many pages are compatible with our allocation.
2498 * For movable allocation, it's the number of movable pages which
2499 * we just obtained. For other types it's a bit more tricky.
2501 if (start_type == MIGRATE_MOVABLE) {
2502 alike_pages = movable_pages;
2505 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2506 * to MOVABLE pageblock, consider all non-movable pages as
2507 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2508 * vice versa, be conservative since we can't distinguish the
2509 * exact migratetype of non-movable pages.
2511 if (old_block_type == MIGRATE_MOVABLE)
2512 alike_pages = pageblock_nr_pages
2513 - (free_pages + movable_pages);
2518 /* moving whole block can fail due to zone boundary conditions */
2523 * If a sufficient number of pages in the block are either free or of
2524 * comparable migratability as our allocation, claim the whole block.
2526 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2527 page_group_by_mobility_disabled)
2528 set_pageblock_migratetype(page, start_type);
2533 move_to_free_list(page, zone, current_order, start_type);
2537 * Check whether there is a suitable fallback freepage with requested order.
2538 * If only_stealable is true, this function returns fallback_mt only if
2539 * we can steal other freepages all together. This would help to reduce
2540 * fragmentation due to mixed migratetype pages in one pageblock.
2542 int find_suitable_fallback(struct free_area *area, unsigned int order,
2543 int migratetype, bool only_stealable, bool *can_steal)
2548 if (area->nr_free == 0)
2553 fallback_mt = fallbacks[migratetype][i];
2554 if (fallback_mt == MIGRATE_TYPES)
2557 if (free_area_empty(area, fallback_mt))
2560 if (can_steal_fallback(order, migratetype))
2563 if (!only_stealable)
2574 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2575 * there are no empty page blocks that contain a page with a suitable order
2577 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2578 unsigned int alloc_order)
2581 unsigned long max_managed, flags;
2584 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2585 * Check is race-prone but harmless.
2587 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2588 if (zone->nr_reserved_highatomic >= max_managed)
2591 spin_lock_irqsave(&zone->lock, flags);
2593 /* Recheck the nr_reserved_highatomic limit under the lock */
2594 if (zone->nr_reserved_highatomic >= max_managed)
2598 mt = get_pageblock_migratetype(page);
2599 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2600 && !is_migrate_cma(mt)) {
2601 zone->nr_reserved_highatomic += pageblock_nr_pages;
2602 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2603 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2607 spin_unlock_irqrestore(&zone->lock, flags);
2611 * Used when an allocation is about to fail under memory pressure. This
2612 * potentially hurts the reliability of high-order allocations when under
2613 * intense memory pressure but failed atomic allocations should be easier
2614 * to recover from than an OOM.
2616 * If @force is true, try to unreserve a pageblock even though highatomic
2617 * pageblock is exhausted.
2619 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2622 struct zonelist *zonelist = ac->zonelist;
2623 unsigned long flags;
2630 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2633 * Preserve at least one pageblock unless memory pressure
2636 if (!force && zone->nr_reserved_highatomic <=
2640 spin_lock_irqsave(&zone->lock, flags);
2641 for (order = 0; order < MAX_ORDER; order++) {
2642 struct free_area *area = &(zone->free_area[order]);
2644 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2649 * In page freeing path, migratetype change is racy so
2650 * we can counter several free pages in a pageblock
2651 * in this loop althoug we changed the pageblock type
2652 * from highatomic to ac->migratetype. So we should
2653 * adjust the count once.
2655 if (is_migrate_highatomic_page(page)) {
2657 * It should never happen but changes to
2658 * locking could inadvertently allow a per-cpu
2659 * drain to add pages to MIGRATE_HIGHATOMIC
2660 * while unreserving so be safe and watch for
2663 zone->nr_reserved_highatomic -= min(
2665 zone->nr_reserved_highatomic);
2669 * Convert to ac->migratetype and avoid the normal
2670 * pageblock stealing heuristics. Minimally, the caller
2671 * is doing the work and needs the pages. More
2672 * importantly, if the block was always converted to
2673 * MIGRATE_UNMOVABLE or another type then the number
2674 * of pageblocks that cannot be completely freed
2677 set_pageblock_migratetype(page, ac->migratetype);
2678 ret = move_freepages_block(zone, page, ac->migratetype,
2681 spin_unlock_irqrestore(&zone->lock, flags);
2685 spin_unlock_irqrestore(&zone->lock, flags);
2692 * Try finding a free buddy page on the fallback list and put it on the free
2693 * list of requested migratetype, possibly along with other pages from the same
2694 * block, depending on fragmentation avoidance heuristics. Returns true if
2695 * fallback was found so that __rmqueue_smallest() can grab it.
2697 * The use of signed ints for order and current_order is a deliberate
2698 * deviation from the rest of this file, to make the for loop
2699 * condition simpler.
2701 static __always_inline bool
2702 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2703 unsigned int alloc_flags)
2705 struct free_area *area;
2707 int min_order = order;
2713 * Do not steal pages from freelists belonging to other pageblocks
2714 * i.e. orders < pageblock_order. If there are no local zones free,
2715 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2717 if (alloc_flags & ALLOC_NOFRAGMENT)
2718 min_order = pageblock_order;
2721 * Find the largest available free page in the other list. This roughly
2722 * approximates finding the pageblock with the most free pages, which
2723 * would be too costly to do exactly.
2725 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2727 area = &(zone->free_area[current_order]);
2728 fallback_mt = find_suitable_fallback(area, current_order,
2729 start_migratetype, false, &can_steal);
2730 if (fallback_mt == -1)
2734 * We cannot steal all free pages from the pageblock and the
2735 * requested migratetype is movable. In that case it's better to
2736 * steal and split the smallest available page instead of the
2737 * largest available page, because even if the next movable
2738 * allocation falls back into a different pageblock than this
2739 * one, it won't cause permanent fragmentation.
2741 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2742 && current_order > order)
2751 for (current_order = order; current_order < MAX_ORDER;
2753 area = &(zone->free_area[current_order]);
2754 fallback_mt = find_suitable_fallback(area, current_order,
2755 start_migratetype, false, &can_steal);
2756 if (fallback_mt != -1)
2761 * This should not happen - we already found a suitable fallback
2762 * when looking for the largest page.
2764 VM_BUG_ON(current_order == MAX_ORDER);
2767 page = get_page_from_free_area(area, fallback_mt);
2769 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2772 trace_mm_page_alloc_extfrag(page, order, current_order,
2773 start_migratetype, fallback_mt);
2780 * Do the hard work of removing an element from the buddy allocator.
2781 * Call me with the zone->lock already held.
2783 static __always_inline struct page *
2784 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2785 unsigned int alloc_flags)
2791 * Balance movable allocations between regular and CMA areas by
2792 * allocating from CMA when over half of the zone's free memory
2793 * is in the CMA area.
2795 if (alloc_flags & ALLOC_CMA &&
2796 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2797 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2798 page = __rmqueue_cma_fallback(zone, order);
2804 page = __rmqueue_smallest(zone, order, migratetype);
2805 if (unlikely(!page)) {
2806 if (alloc_flags & ALLOC_CMA)
2807 page = __rmqueue_cma_fallback(zone, order);
2809 if (!page && __rmqueue_fallback(zone, order, migratetype,
2814 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2819 * Obtain a specified number of elements from the buddy allocator, all under
2820 * a single hold of the lock, for efficiency. Add them to the supplied list.
2821 * Returns the number of new pages which were placed at *list.
2823 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2824 unsigned long count, struct list_head *list,
2825 int migratetype, unsigned int alloc_flags)
2829 spin_lock(&zone->lock);
2830 for (i = 0; i < count; ++i) {
2831 struct page *page = __rmqueue(zone, order, migratetype,
2833 if (unlikely(page == NULL))
2836 if (unlikely(check_pcp_refill(page)))
2840 * Split buddy pages returned by expand() are received here in
2841 * physical page order. The page is added to the tail of
2842 * caller's list. From the callers perspective, the linked list
2843 * is ordered by page number under some conditions. This is
2844 * useful for IO devices that can forward direction from the
2845 * head, thus also in the physical page order. This is useful
2846 * for IO devices that can merge IO requests if the physical
2847 * pages are ordered properly.
2849 list_add_tail(&page->lru, list);
2851 if (is_migrate_cma(get_pcppage_migratetype(page)))
2852 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2857 * i pages were removed from the buddy list even if some leak due
2858 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2859 * on i. Do not confuse with 'alloced' which is the number of
2860 * pages added to the pcp list.
2862 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2863 spin_unlock(&zone->lock);
2869 * Called from the vmstat counter updater to drain pagesets of this
2870 * currently executing processor on remote nodes after they have
2873 * Note that this function must be called with the thread pinned to
2874 * a single processor.
2876 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2878 unsigned long flags;
2879 int to_drain, batch;
2881 local_irq_save(flags);
2882 batch = READ_ONCE(pcp->batch);
2883 to_drain = min(pcp->count, batch);
2885 free_pcppages_bulk(zone, to_drain, pcp);
2886 local_irq_restore(flags);
2891 * Drain pcplists of the indicated processor and zone.
2893 * The processor must either be the current processor and the
2894 * thread pinned to the current processor or a processor that
2897 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2899 unsigned long flags;
2900 struct per_cpu_pageset *pset;
2901 struct per_cpu_pages *pcp;
2903 local_irq_save(flags);
2904 pset = per_cpu_ptr(zone->pageset, cpu);
2908 free_pcppages_bulk(zone, pcp->count, pcp);
2909 local_irq_restore(flags);
2913 * Drain pcplists of all zones on the indicated processor.
2915 * The processor must either be the current processor and the
2916 * thread pinned to the current processor or a processor that
2919 static void drain_pages(unsigned int cpu)
2923 for_each_populated_zone(zone) {
2924 drain_pages_zone(cpu, zone);
2929 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2931 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2932 * the single zone's pages.
2934 void drain_local_pages(struct zone *zone)
2936 int cpu = smp_processor_id();
2939 drain_pages_zone(cpu, zone);
2944 static void drain_local_pages_wq(struct work_struct *work)
2946 struct pcpu_drain *drain;
2948 drain = container_of(work, struct pcpu_drain, work);
2951 * drain_all_pages doesn't use proper cpu hotplug protection so
2952 * we can race with cpu offline when the WQ can move this from
2953 * a cpu pinned worker to an unbound one. We can operate on a different
2954 * cpu which is allright but we also have to make sure to not move to
2958 drain_local_pages(drain->zone);
2963 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2965 * When zone parameter is non-NULL, spill just the single zone's pages.
2967 * Note that this can be extremely slow as the draining happens in a workqueue.
2969 void drain_all_pages(struct zone *zone)
2974 * Allocate in the BSS so we wont require allocation in
2975 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2977 static cpumask_t cpus_with_pcps;
2980 * Make sure nobody triggers this path before mm_percpu_wq is fully
2983 if (WARN_ON_ONCE(!mm_percpu_wq))
2987 * Do not drain if one is already in progress unless it's specific to
2988 * a zone. Such callers are primarily CMA and memory hotplug and need
2989 * the drain to be complete when the call returns.
2991 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2994 mutex_lock(&pcpu_drain_mutex);
2998 * We don't care about racing with CPU hotplug event
2999 * as offline notification will cause the notified
3000 * cpu to drain that CPU pcps and on_each_cpu_mask
3001 * disables preemption as part of its processing
3003 for_each_online_cpu(cpu) {
3004 struct per_cpu_pageset *pcp;
3006 bool has_pcps = false;
3009 pcp = per_cpu_ptr(zone->pageset, cpu);
3013 for_each_populated_zone(z) {
3014 pcp = per_cpu_ptr(z->pageset, cpu);
3015 if (pcp->pcp.count) {
3023 cpumask_set_cpu(cpu, &cpus_with_pcps);
3025 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3028 for_each_cpu(cpu, &cpus_with_pcps) {
3029 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3032 INIT_WORK(&drain->work, drain_local_pages_wq);
3033 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3035 for_each_cpu(cpu, &cpus_with_pcps)
3036 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3038 mutex_unlock(&pcpu_drain_mutex);
3041 #ifdef CONFIG_HIBERNATION
3044 * Touch the watchdog for every WD_PAGE_COUNT pages.
3046 #define WD_PAGE_COUNT (128*1024)
3048 void mark_free_pages(struct zone *zone)
3050 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3051 unsigned long flags;
3052 unsigned int order, t;
3055 if (zone_is_empty(zone))
3058 spin_lock_irqsave(&zone->lock, flags);
3060 max_zone_pfn = zone_end_pfn(zone);
3061 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3062 if (pfn_valid(pfn)) {
3063 page = pfn_to_page(pfn);
3065 if (!--page_count) {
3066 touch_nmi_watchdog();
3067 page_count = WD_PAGE_COUNT;
3070 if (page_zone(page) != zone)
3073 if (!swsusp_page_is_forbidden(page))
3074 swsusp_unset_page_free(page);
3077 for_each_migratetype_order(order, t) {
3078 list_for_each_entry(page,
3079 &zone->free_area[order].free_list[t], lru) {
3082 pfn = page_to_pfn(page);
3083 for (i = 0; i < (1UL << order); i++) {
3084 if (!--page_count) {
3085 touch_nmi_watchdog();
3086 page_count = WD_PAGE_COUNT;
3088 swsusp_set_page_free(pfn_to_page(pfn + i));
3092 spin_unlock_irqrestore(&zone->lock, flags);
3094 #endif /* CONFIG_PM */
3096 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3100 if (!free_pcp_prepare(page))
3103 migratetype = get_pfnblock_migratetype(page, pfn);
3104 set_pcppage_migratetype(page, migratetype);
3108 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3110 struct zone *zone = page_zone(page);
3111 struct per_cpu_pages *pcp;
3114 migratetype = get_pcppage_migratetype(page);
3115 __count_vm_event(PGFREE);
3118 * We only track unmovable, reclaimable and movable on pcp lists.
3119 * Free ISOLATE pages back to the allocator because they are being
3120 * offlined but treat HIGHATOMIC as movable pages so we can get those
3121 * areas back if necessary. Otherwise, we may have to free
3122 * excessively into the page allocator
3124 if (migratetype >= MIGRATE_PCPTYPES) {
3125 if (unlikely(is_migrate_isolate(migratetype))) {
3126 free_one_page(zone, page, pfn, 0, migratetype);
3129 migratetype = MIGRATE_MOVABLE;
3132 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3133 list_add(&page->lru, &pcp->lists[migratetype]);
3135 if (pcp->count >= pcp->high) {
3136 unsigned long batch = READ_ONCE(pcp->batch);
3137 free_pcppages_bulk(zone, batch, pcp);
3142 * Free a 0-order page
3144 void free_unref_page(struct page *page)
3146 unsigned long flags;
3147 unsigned long pfn = page_to_pfn(page);
3149 if (!free_unref_page_prepare(page, pfn))
3152 local_irq_save(flags);
3153 free_unref_page_commit(page, pfn);
3154 local_irq_restore(flags);
3158 * Free a list of 0-order pages
3160 void free_unref_page_list(struct list_head *list)
3162 struct page *page, *next;
3163 unsigned long flags, pfn;
3164 int batch_count = 0;
3166 /* Prepare pages for freeing */
3167 list_for_each_entry_safe(page, next, list, lru) {
3168 pfn = page_to_pfn(page);
3169 if (!free_unref_page_prepare(page, pfn))
3170 list_del(&page->lru);
3171 set_page_private(page, pfn);
3174 local_irq_save(flags);
3175 list_for_each_entry_safe(page, next, list, lru) {
3176 unsigned long pfn = page_private(page);
3178 set_page_private(page, 0);
3179 trace_mm_page_free_batched(page);
3180 free_unref_page_commit(page, pfn);
3183 * Guard against excessive IRQ disabled times when we get
3184 * a large list of pages to free.
3186 if (++batch_count == SWAP_CLUSTER_MAX) {
3187 local_irq_restore(flags);
3189 local_irq_save(flags);
3192 local_irq_restore(flags);
3196 * split_page takes a non-compound higher-order page, and splits it into
3197 * n (1<<order) sub-pages: page[0..n]
3198 * Each sub-page must be freed individually.
3200 * Note: this is probably too low level an operation for use in drivers.
3201 * Please consult with lkml before using this in your driver.
3203 void split_page(struct page *page, unsigned int order)
3207 VM_BUG_ON_PAGE(PageCompound(page), page);
3208 VM_BUG_ON_PAGE(!page_count(page), page);
3210 for (i = 1; i < (1 << order); i++)
3211 set_page_refcounted(page + i);
3212 split_page_owner(page, order);
3214 EXPORT_SYMBOL_GPL(split_page);
3216 int __isolate_free_page(struct page *page, unsigned int order)
3218 unsigned long watermark;
3222 BUG_ON(!PageBuddy(page));
3224 zone = page_zone(page);
3225 mt = get_pageblock_migratetype(page);
3227 if (!is_migrate_isolate(mt)) {
3229 * Obey watermarks as if the page was being allocated. We can
3230 * emulate a high-order watermark check with a raised order-0
3231 * watermark, because we already know our high-order page
3234 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3235 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3238 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3241 /* Remove page from free list */
3243 del_page_from_free_list(page, zone, order);
3246 * Set the pageblock if the isolated page is at least half of a
3249 if (order >= pageblock_order - 1) {
3250 struct page *endpage = page + (1 << order) - 1;
3251 for (; page < endpage; page += pageblock_nr_pages) {
3252 int mt = get_pageblock_migratetype(page);
3253 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3254 && !is_migrate_highatomic(mt))
3255 set_pageblock_migratetype(page,
3261 return 1UL << order;
3265 * __putback_isolated_page - Return a now-isolated page back where we got it
3266 * @page: Page that was isolated
3267 * @order: Order of the isolated page
3268 * @mt: The page's pageblock's migratetype
3270 * This function is meant to return a page pulled from the free lists via
3271 * __isolate_free_page back to the free lists they were pulled from.
3273 void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3275 struct zone *zone = page_zone(page);
3277 /* zone lock should be held when this function is called */
3278 lockdep_assert_held(&zone->lock);
3280 /* Return isolated page to tail of freelist. */
3281 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3285 * Update NUMA hit/miss statistics
3287 * Must be called with interrupts disabled.
3289 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3292 enum numa_stat_item local_stat = NUMA_LOCAL;
3294 /* skip numa counters update if numa stats is disabled */
3295 if (!static_branch_likely(&vm_numa_stat_key))
3298 if (zone_to_nid(z) != numa_node_id())
3299 local_stat = NUMA_OTHER;
3301 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3302 __inc_numa_state(z, NUMA_HIT);
3304 __inc_numa_state(z, NUMA_MISS);
3305 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3307 __inc_numa_state(z, local_stat);
3311 /* Remove page from the per-cpu list, caller must protect the list */
3312 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3313 unsigned int alloc_flags,
3314 struct per_cpu_pages *pcp,
3315 struct list_head *list)
3320 if (list_empty(list)) {
3321 pcp->count += rmqueue_bulk(zone, 0,
3323 migratetype, alloc_flags);
3324 if (unlikely(list_empty(list)))
3328 page = list_first_entry(list, struct page, lru);
3329 list_del(&page->lru);
3331 } while (check_new_pcp(page));
3336 /* Lock and remove page from the per-cpu list */
3337 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3338 struct zone *zone, gfp_t gfp_flags,
3339 int migratetype, unsigned int alloc_flags)
3341 struct per_cpu_pages *pcp;
3342 struct list_head *list;
3344 unsigned long flags;
3346 local_irq_save(flags);
3347 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3348 list = &pcp->lists[migratetype];
3349 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3351 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3352 zone_statistics(preferred_zone, zone);
3354 local_irq_restore(flags);
3359 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3362 struct page *rmqueue(struct zone *preferred_zone,
3363 struct zone *zone, unsigned int order,
3364 gfp_t gfp_flags, unsigned int alloc_flags,
3367 unsigned long flags;
3370 if (likely(order == 0)) {
3372 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3373 * we need to skip it when CMA area isn't allowed.
3375 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3376 migratetype != MIGRATE_MOVABLE) {
3377 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3378 migratetype, alloc_flags);
3384 * We most definitely don't want callers attempting to
3385 * allocate greater than order-1 page units with __GFP_NOFAIL.
3387 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3388 spin_lock_irqsave(&zone->lock, flags);
3393 * order-0 request can reach here when the pcplist is skipped
3394 * due to non-CMA allocation context. HIGHATOMIC area is
3395 * reserved for high-order atomic allocation, so order-0
3396 * request should skip it.
3398 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3399 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3401 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3404 page = __rmqueue(zone, order, migratetype, alloc_flags);
3405 } while (page && check_new_pages(page, order));
3406 spin_unlock(&zone->lock);
3409 __mod_zone_freepage_state(zone, -(1 << order),
3410 get_pcppage_migratetype(page));
3412 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3413 zone_statistics(preferred_zone, zone);
3414 local_irq_restore(flags);
3417 /* Separate test+clear to avoid unnecessary atomics */
3418 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3419 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3420 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3423 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3427 local_irq_restore(flags);
3431 #ifdef CONFIG_FAIL_PAGE_ALLOC
3434 struct fault_attr attr;
3436 bool ignore_gfp_highmem;
3437 bool ignore_gfp_reclaim;
3439 } fail_page_alloc = {
3440 .attr = FAULT_ATTR_INITIALIZER,
3441 .ignore_gfp_reclaim = true,
3442 .ignore_gfp_highmem = true,
3446 static int __init setup_fail_page_alloc(char *str)
3448 return setup_fault_attr(&fail_page_alloc.attr, str);
3450 __setup("fail_page_alloc=", setup_fail_page_alloc);
3452 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3454 if (order < fail_page_alloc.min_order)
3456 if (gfp_mask & __GFP_NOFAIL)
3458 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3460 if (fail_page_alloc.ignore_gfp_reclaim &&
3461 (gfp_mask & __GFP_DIRECT_RECLAIM))
3464 return should_fail(&fail_page_alloc.attr, 1 << order);
3467 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3469 static int __init fail_page_alloc_debugfs(void)
3471 umode_t mode = S_IFREG | 0600;
3474 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3475 &fail_page_alloc.attr);
3477 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3478 &fail_page_alloc.ignore_gfp_reclaim);
3479 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3480 &fail_page_alloc.ignore_gfp_highmem);
3481 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3486 late_initcall(fail_page_alloc_debugfs);
3488 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3490 #else /* CONFIG_FAIL_PAGE_ALLOC */
3492 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3497 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3499 static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3501 return __should_fail_alloc_page(gfp_mask, order);
3503 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3505 static inline long __zone_watermark_unusable_free(struct zone *z,
3506 unsigned int order, unsigned int alloc_flags)
3508 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3509 long unusable_free = (1 << order) - 1;
3512 * If the caller does not have rights to ALLOC_HARDER then subtract
3513 * the high-atomic reserves. This will over-estimate the size of the
3514 * atomic reserve but it avoids a search.
3516 if (likely(!alloc_harder))
3517 unusable_free += z->nr_reserved_highatomic;
3520 /* If allocation can't use CMA areas don't use free CMA pages */
3521 if (!(alloc_flags & ALLOC_CMA))
3522 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3525 return unusable_free;
3529 * Return true if free base pages are above 'mark'. For high-order checks it
3530 * will return true of the order-0 watermark is reached and there is at least
3531 * one free page of a suitable size. Checking now avoids taking the zone lock
3532 * to check in the allocation paths if no pages are free.
3534 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3535 int highest_zoneidx, unsigned int alloc_flags,
3540 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3542 /* free_pages may go negative - that's OK */
3543 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3545 if (alloc_flags & ALLOC_HIGH)
3548 if (unlikely(alloc_harder)) {
3550 * OOM victims can try even harder than normal ALLOC_HARDER
3551 * users on the grounds that it's definitely going to be in
3552 * the exit path shortly and free memory. Any allocation it
3553 * makes during the free path will be small and short-lived.
3555 if (alloc_flags & ALLOC_OOM)
3562 * Check watermarks for an order-0 allocation request. If these
3563 * are not met, then a high-order request also cannot go ahead
3564 * even if a suitable page happened to be free.
3566 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3569 /* If this is an order-0 request then the watermark is fine */
3573 /* For a high-order request, check at least one suitable page is free */
3574 for (o = order; o < MAX_ORDER; o++) {
3575 struct free_area *area = &z->free_area[o];
3581 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3582 if (!free_area_empty(area, mt))
3587 if ((alloc_flags & ALLOC_CMA) &&
3588 !free_area_empty(area, MIGRATE_CMA)) {
3592 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3598 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3599 int highest_zoneidx, unsigned int alloc_flags)
3601 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3602 zone_page_state(z, NR_FREE_PAGES));
3605 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3606 unsigned long mark, int highest_zoneidx,
3607 unsigned int alloc_flags, gfp_t gfp_mask)
3611 free_pages = zone_page_state(z, NR_FREE_PAGES);
3614 * Fast check for order-0 only. If this fails then the reserves
3615 * need to be calculated.
3620 fast_free = free_pages;
3621 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3622 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3626 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3630 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3631 * when checking the min watermark. The min watermark is the
3632 * point where boosting is ignored so that kswapd is woken up
3633 * when below the low watermark.
3635 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3636 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3637 mark = z->_watermark[WMARK_MIN];
3638 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3639 alloc_flags, free_pages);
3645 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3646 unsigned long mark, int highest_zoneidx)
3648 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3650 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3651 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3653 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3658 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3660 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3661 node_reclaim_distance;
3663 #else /* CONFIG_NUMA */
3664 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3668 #endif /* CONFIG_NUMA */
3671 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3672 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3673 * premature use of a lower zone may cause lowmem pressure problems that
3674 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3675 * probably too small. It only makes sense to spread allocations to avoid
3676 * fragmentation between the Normal and DMA32 zones.
3678 static inline unsigned int
3679 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3681 unsigned int alloc_flags;
3684 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3687 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3689 #ifdef CONFIG_ZONE_DMA32
3693 if (zone_idx(zone) != ZONE_NORMAL)
3697 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3698 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3699 * on UMA that if Normal is populated then so is DMA32.
3701 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3702 if (nr_online_nodes > 1 && !populated_zone(--zone))
3705 alloc_flags |= ALLOC_NOFRAGMENT;
3706 #endif /* CONFIG_ZONE_DMA32 */
3710 static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3711 unsigned int alloc_flags)
3714 unsigned int pflags = current->flags;
3716 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3717 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3718 alloc_flags |= ALLOC_CMA;
3725 * get_page_from_freelist goes through the zonelist trying to allocate
3728 static struct page *
3729 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3730 const struct alloc_context *ac)
3734 struct pglist_data *last_pgdat_dirty_limit = NULL;
3739 * Scan zonelist, looking for a zone with enough free.
3740 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3742 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3743 z = ac->preferred_zoneref;
3744 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist,
3745 ac->highest_zoneidx, ac->nodemask) {
3749 if (cpusets_enabled() &&
3750 (alloc_flags & ALLOC_CPUSET) &&
3751 !__cpuset_zone_allowed(zone, gfp_mask))
3754 * When allocating a page cache page for writing, we
3755 * want to get it from a node that is within its dirty
3756 * limit, such that no single node holds more than its
3757 * proportional share of globally allowed dirty pages.
3758 * The dirty limits take into account the node's
3759 * lowmem reserves and high watermark so that kswapd
3760 * should be able to balance it without having to
3761 * write pages from its LRU list.
3763 * XXX: For now, allow allocations to potentially
3764 * exceed the per-node dirty limit in the slowpath
3765 * (spread_dirty_pages unset) before going into reclaim,
3766 * which is important when on a NUMA setup the allowed
3767 * nodes are together not big enough to reach the
3768 * global limit. The proper fix for these situations
3769 * will require awareness of nodes in the
3770 * dirty-throttling and the flusher threads.
3772 if (ac->spread_dirty_pages) {
3773 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3776 if (!node_dirty_ok(zone->zone_pgdat)) {
3777 last_pgdat_dirty_limit = zone->zone_pgdat;
3782 if (no_fallback && nr_online_nodes > 1 &&
3783 zone != ac->preferred_zoneref->zone) {
3787 * If moving to a remote node, retry but allow
3788 * fragmenting fallbacks. Locality is more important
3789 * than fragmentation avoidance.
3791 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3792 if (zone_to_nid(zone) != local_nid) {
3793 alloc_flags &= ~ALLOC_NOFRAGMENT;
3798 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3799 if (!zone_watermark_fast(zone, order, mark,
3800 ac->highest_zoneidx, alloc_flags,
3804 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3806 * Watermark failed for this zone, but see if we can
3807 * grow this zone if it contains deferred pages.
3809 if (static_branch_unlikely(&deferred_pages)) {
3810 if (_deferred_grow_zone(zone, order))
3814 /* Checked here to keep the fast path fast */
3815 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3816 if (alloc_flags & ALLOC_NO_WATERMARKS)
3819 if (node_reclaim_mode == 0 ||
3820 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3823 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3825 case NODE_RECLAIM_NOSCAN:
3828 case NODE_RECLAIM_FULL:
3829 /* scanned but unreclaimable */
3832 /* did we reclaim enough */
3833 if (zone_watermark_ok(zone, order, mark,
3834 ac->highest_zoneidx, alloc_flags))
3842 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3843 gfp_mask, alloc_flags, ac->migratetype);
3845 prep_new_page(page, order, gfp_mask, alloc_flags);
3848 * If this is a high-order atomic allocation then check
3849 * if the pageblock should be reserved for the future
3851 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3852 reserve_highatomic_pageblock(page, zone, order);
3856 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3857 /* Try again if zone has deferred pages */
3858 if (static_branch_unlikely(&deferred_pages)) {
3859 if (_deferred_grow_zone(zone, order))
3867 * It's possible on a UMA machine to get through all zones that are
3868 * fragmented. If avoiding fragmentation, reset and try again.
3871 alloc_flags &= ~ALLOC_NOFRAGMENT;
3878 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3880 unsigned int filter = SHOW_MEM_FILTER_NODES;
3883 * This documents exceptions given to allocations in certain
3884 * contexts that are allowed to allocate outside current's set
3887 if (!(gfp_mask & __GFP_NOMEMALLOC))
3888 if (tsk_is_oom_victim(current) ||
3889 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3890 filter &= ~SHOW_MEM_FILTER_NODES;
3891 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3892 filter &= ~SHOW_MEM_FILTER_NODES;
3894 show_mem(filter, nodemask);
3897 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3899 struct va_format vaf;
3901 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3903 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3906 va_start(args, fmt);
3909 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3910 current->comm, &vaf, gfp_mask, &gfp_mask,
3911 nodemask_pr_args(nodemask));
3914 cpuset_print_current_mems_allowed();
3917 warn_alloc_show_mem(gfp_mask, nodemask);
3920 static inline struct page *
3921 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3922 unsigned int alloc_flags,
3923 const struct alloc_context *ac)
3927 page = get_page_from_freelist(gfp_mask, order,
3928 alloc_flags|ALLOC_CPUSET, ac);
3930 * fallback to ignore cpuset restriction if our nodes
3934 page = get_page_from_freelist(gfp_mask, order,
3940 static inline struct page *
3941 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3942 const struct alloc_context *ac, unsigned long *did_some_progress)
3944 struct oom_control oc = {
3945 .zonelist = ac->zonelist,
3946 .nodemask = ac->nodemask,
3948 .gfp_mask = gfp_mask,
3953 *did_some_progress = 0;
3956 * Acquire the oom lock. If that fails, somebody else is
3957 * making progress for us.
3959 if (!mutex_trylock(&oom_lock)) {
3960 *did_some_progress = 1;
3961 schedule_timeout_uninterruptible(1);
3966 * Go through the zonelist yet one more time, keep very high watermark
3967 * here, this is only to catch a parallel oom killing, we must fail if
3968 * we're still under heavy pressure. But make sure that this reclaim
3969 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3970 * allocation which will never fail due to oom_lock already held.
3972 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3973 ~__GFP_DIRECT_RECLAIM, order,
3974 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3978 /* Coredumps can quickly deplete all memory reserves */
3979 if (current->flags & PF_DUMPCORE)
3981 /* The OOM killer will not help higher order allocs */
3982 if (order > PAGE_ALLOC_COSTLY_ORDER)
3985 * We have already exhausted all our reclaim opportunities without any
3986 * success so it is time to admit defeat. We will skip the OOM killer
3987 * because it is very likely that the caller has a more reasonable
3988 * fallback than shooting a random task.
3990 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3992 /* The OOM killer does not needlessly kill tasks for lowmem */
3993 if (ac->highest_zoneidx < ZONE_NORMAL)
3995 if (pm_suspended_storage())
3998 * XXX: GFP_NOFS allocations should rather fail than rely on
3999 * other request to make a forward progress.
4000 * We are in an unfortunate situation where out_of_memory cannot
4001 * do much for this context but let's try it to at least get
4002 * access to memory reserved if the current task is killed (see
4003 * out_of_memory). Once filesystems are ready to handle allocation
4004 * failures more gracefully we should just bail out here.
4007 /* The OOM killer may not free memory on a specific node */
4008 if (gfp_mask & __GFP_THISNODE)
4011 /* Exhausted what can be done so it's blame time */
4012 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4013 *did_some_progress = 1;
4016 * Help non-failing allocations by giving them access to memory
4019 if (gfp_mask & __GFP_NOFAIL)
4020 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4021 ALLOC_NO_WATERMARKS, ac);
4024 mutex_unlock(&oom_lock);
4029 * Maximum number of compaction retries wit a progress before OOM
4030 * killer is consider as the only way to move forward.
4032 #define MAX_COMPACT_RETRIES 16
4034 #ifdef CONFIG_COMPACTION
4035 /* Try memory compaction for high-order allocations before reclaim */
4036 static struct page *
4037 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4038 unsigned int alloc_flags, const struct alloc_context *ac,
4039 enum compact_priority prio, enum compact_result *compact_result)
4041 struct page *page = NULL;
4042 unsigned long pflags;
4043 unsigned int noreclaim_flag;
4048 psi_memstall_enter(&pflags);
4049 noreclaim_flag = memalloc_noreclaim_save();
4051 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4054 memalloc_noreclaim_restore(noreclaim_flag);
4055 psi_memstall_leave(&pflags);
4058 * At least in one zone compaction wasn't deferred or skipped, so let's
4059 * count a compaction stall
4061 count_vm_event(COMPACTSTALL);
4063 /* Prep a captured page if available */
4065 prep_new_page(page, order, gfp_mask, alloc_flags);
4067 /* Try get a page from the freelist if available */
4069 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4072 struct zone *zone = page_zone(page);
4074 zone->compact_blockskip_flush = false;
4075 compaction_defer_reset(zone, order, true);
4076 count_vm_event(COMPACTSUCCESS);
4081 * It's bad if compaction run occurs and fails. The most likely reason
4082 * is that pages exist, but not enough to satisfy watermarks.
4084 count_vm_event(COMPACTFAIL);
4092 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4093 enum compact_result compact_result,
4094 enum compact_priority *compact_priority,
4095 int *compaction_retries)
4097 int max_retries = MAX_COMPACT_RETRIES;
4100 int retries = *compaction_retries;
4101 enum compact_priority priority = *compact_priority;
4106 if (compaction_made_progress(compact_result))
4107 (*compaction_retries)++;
4110 * compaction considers all the zone as desperately out of memory
4111 * so it doesn't really make much sense to retry except when the
4112 * failure could be caused by insufficient priority
4114 if (compaction_failed(compact_result))
4115 goto check_priority;
4118 * compaction was skipped because there are not enough order-0 pages
4119 * to work with, so we retry only if it looks like reclaim can help.
4121 if (compaction_needs_reclaim(compact_result)) {
4122 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4127 * make sure the compaction wasn't deferred or didn't bail out early
4128 * due to locks contention before we declare that we should give up.
4129 * But the next retry should use a higher priority if allowed, so
4130 * we don't just keep bailing out endlessly.
4132 if (compaction_withdrawn(compact_result)) {
4133 goto check_priority;
4137 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4138 * costly ones because they are de facto nofail and invoke OOM
4139 * killer to move on while costly can fail and users are ready
4140 * to cope with that. 1/4 retries is rather arbitrary but we
4141 * would need much more detailed feedback from compaction to
4142 * make a better decision.
4144 if (order > PAGE_ALLOC_COSTLY_ORDER)
4146 if (*compaction_retries <= max_retries) {
4152 * Make sure there are attempts at the highest priority if we exhausted
4153 * all retries or failed at the lower priorities.
4156 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4157 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4159 if (*compact_priority > min_priority) {
4160 (*compact_priority)--;
4161 *compaction_retries = 0;
4165 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4169 static inline struct page *
4170 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4171 unsigned int alloc_flags, const struct alloc_context *ac,
4172 enum compact_priority prio, enum compact_result *compact_result)
4174 *compact_result = COMPACT_SKIPPED;
4179 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4180 enum compact_result compact_result,
4181 enum compact_priority *compact_priority,
4182 int *compaction_retries)
4187 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4191 * There are setups with compaction disabled which would prefer to loop
4192 * inside the allocator rather than hit the oom killer prematurely.
4193 * Let's give them a good hope and keep retrying while the order-0
4194 * watermarks are OK.
4196 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4197 ac->highest_zoneidx, ac->nodemask) {
4198 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4199 ac->highest_zoneidx, alloc_flags))
4204 #endif /* CONFIG_COMPACTION */
4206 #ifdef CONFIG_LOCKDEP
4207 static struct lockdep_map __fs_reclaim_map =
4208 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4210 static bool __need_fs_reclaim(gfp_t gfp_mask)
4212 gfp_mask = current_gfp_context(gfp_mask);
4214 /* no reclaim without waiting on it */
4215 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4218 /* this guy won't enter reclaim */
4219 if (current->flags & PF_MEMALLOC)
4222 /* We're only interested __GFP_FS allocations for now */
4223 if (!(gfp_mask & __GFP_FS))
4226 if (gfp_mask & __GFP_NOLOCKDEP)
4232 void __fs_reclaim_acquire(void)
4234 lock_map_acquire(&__fs_reclaim_map);
4237 void __fs_reclaim_release(void)
4239 lock_map_release(&__fs_reclaim_map);
4242 void fs_reclaim_acquire(gfp_t gfp_mask)
4244 if (__need_fs_reclaim(gfp_mask))
4245 __fs_reclaim_acquire();
4247 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4249 void fs_reclaim_release(gfp_t gfp_mask)
4251 if (__need_fs_reclaim(gfp_mask))
4252 __fs_reclaim_release();
4254 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4257 /* Perform direct synchronous page reclaim */
4259 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4260 const struct alloc_context *ac)
4263 unsigned int noreclaim_flag;
4264 unsigned long pflags;
4268 /* We now go into synchronous reclaim */
4269 cpuset_memory_pressure_bump();
4270 psi_memstall_enter(&pflags);
4271 fs_reclaim_acquire(gfp_mask);
4272 noreclaim_flag = memalloc_noreclaim_save();
4274 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4277 memalloc_noreclaim_restore(noreclaim_flag);
4278 fs_reclaim_release(gfp_mask);
4279 psi_memstall_leave(&pflags);
4286 /* The really slow allocator path where we enter direct reclaim */
4287 static inline struct page *
4288 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4289 unsigned int alloc_flags, const struct alloc_context *ac,
4290 unsigned long *did_some_progress)
4292 struct page *page = NULL;
4293 bool drained = false;
4295 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4296 if (unlikely(!(*did_some_progress)))
4300 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4303 * If an allocation failed after direct reclaim, it could be because
4304 * pages are pinned on the per-cpu lists or in high alloc reserves.
4305 * Shrink them and try again
4307 if (!page && !drained) {
4308 unreserve_highatomic_pageblock(ac, false);
4309 drain_all_pages(NULL);
4317 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4318 const struct alloc_context *ac)
4322 pg_data_t *last_pgdat = NULL;
4323 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4325 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4327 if (last_pgdat != zone->zone_pgdat)
4328 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4329 last_pgdat = zone->zone_pgdat;
4333 static inline unsigned int
4334 gfp_to_alloc_flags(gfp_t gfp_mask)
4336 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4339 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4340 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4341 * to save two branches.
4343 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4344 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4347 * The caller may dip into page reserves a bit more if the caller
4348 * cannot run direct reclaim, or if the caller has realtime scheduling
4349 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4350 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4352 alloc_flags |= (__force int)
4353 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4355 if (gfp_mask & __GFP_ATOMIC) {
4357 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4358 * if it can't schedule.
4360 if (!(gfp_mask & __GFP_NOMEMALLOC))
4361 alloc_flags |= ALLOC_HARDER;
4363 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4364 * comment for __cpuset_node_allowed().
4366 alloc_flags &= ~ALLOC_CPUSET;
4367 } else if (unlikely(rt_task(current)) && !in_interrupt())
4368 alloc_flags |= ALLOC_HARDER;
4370 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4375 static bool oom_reserves_allowed(struct task_struct *tsk)
4377 if (!tsk_is_oom_victim(tsk))
4381 * !MMU doesn't have oom reaper so give access to memory reserves
4382 * only to the thread with TIF_MEMDIE set
4384 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4391 * Distinguish requests which really need access to full memory
4392 * reserves from oom victims which can live with a portion of it
4394 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4396 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4398 if (gfp_mask & __GFP_MEMALLOC)
4399 return ALLOC_NO_WATERMARKS;
4400 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4401 return ALLOC_NO_WATERMARKS;
4402 if (!in_interrupt()) {
4403 if (current->flags & PF_MEMALLOC)
4404 return ALLOC_NO_WATERMARKS;
4405 else if (oom_reserves_allowed(current))
4412 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4414 return !!__gfp_pfmemalloc_flags(gfp_mask);
4418 * Checks whether it makes sense to retry the reclaim to make a forward progress
4419 * for the given allocation request.
4421 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4422 * without success, or when we couldn't even meet the watermark if we
4423 * reclaimed all remaining pages on the LRU lists.
4425 * Returns true if a retry is viable or false to enter the oom path.
4428 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4429 struct alloc_context *ac, int alloc_flags,
4430 bool did_some_progress, int *no_progress_loops)
4437 * Costly allocations might have made a progress but this doesn't mean
4438 * their order will become available due to high fragmentation so
4439 * always increment the no progress counter for them
4441 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4442 *no_progress_loops = 0;
4444 (*no_progress_loops)++;
4447 * Make sure we converge to OOM if we cannot make any progress
4448 * several times in the row.
4450 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4451 /* Before OOM, exhaust highatomic_reserve */
4452 return unreserve_highatomic_pageblock(ac, true);
4456 * Keep reclaiming pages while there is a chance this will lead
4457 * somewhere. If none of the target zones can satisfy our allocation
4458 * request even if all reclaimable pages are considered then we are
4459 * screwed and have to go OOM.
4461 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4462 ac->highest_zoneidx, ac->nodemask) {
4463 unsigned long available;
4464 unsigned long reclaimable;
4465 unsigned long min_wmark = min_wmark_pages(zone);
4468 available = reclaimable = zone_reclaimable_pages(zone);
4469 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4472 * Would the allocation succeed if we reclaimed all
4473 * reclaimable pages?
4475 wmark = __zone_watermark_ok(zone, order, min_wmark,
4476 ac->highest_zoneidx, alloc_flags, available);
4477 trace_reclaim_retry_zone(z, order, reclaimable,
4478 available, min_wmark, *no_progress_loops, wmark);
4481 * If we didn't make any progress and have a lot of
4482 * dirty + writeback pages then we should wait for
4483 * an IO to complete to slow down the reclaim and
4484 * prevent from pre mature OOM
4486 if (!did_some_progress) {
4487 unsigned long write_pending;
4489 write_pending = zone_page_state_snapshot(zone,
4490 NR_ZONE_WRITE_PENDING);
4492 if (2 * write_pending > reclaimable) {
4493 congestion_wait(BLK_RW_ASYNC, HZ/10);
4505 * Memory allocation/reclaim might be called from a WQ context and the
4506 * current implementation of the WQ concurrency control doesn't
4507 * recognize that a particular WQ is congested if the worker thread is
4508 * looping without ever sleeping. Therefore we have to do a short sleep
4509 * here rather than calling cond_resched().
4511 if (current->flags & PF_WQ_WORKER)
4512 schedule_timeout_uninterruptible(1);
4519 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4522 * It's possible that cpuset's mems_allowed and the nodemask from
4523 * mempolicy don't intersect. This should be normally dealt with by
4524 * policy_nodemask(), but it's possible to race with cpuset update in
4525 * such a way the check therein was true, and then it became false
4526 * before we got our cpuset_mems_cookie here.
4527 * This assumes that for all allocations, ac->nodemask can come only
4528 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4529 * when it does not intersect with the cpuset restrictions) or the
4530 * caller can deal with a violated nodemask.
4532 if (cpusets_enabled() && ac->nodemask &&
4533 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4534 ac->nodemask = NULL;
4539 * When updating a task's mems_allowed or mempolicy nodemask, it is
4540 * possible to race with parallel threads in such a way that our
4541 * allocation can fail while the mask is being updated. If we are about
4542 * to fail, check if the cpuset changed during allocation and if so,
4545 if (read_mems_allowed_retry(cpuset_mems_cookie))
4551 static inline struct page *
4552 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4553 struct alloc_context *ac)
4555 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4556 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4557 struct page *page = NULL;
4558 unsigned int alloc_flags;
4559 unsigned long did_some_progress;
4560 enum compact_priority compact_priority;
4561 enum compact_result compact_result;
4562 int compaction_retries;
4563 int no_progress_loops;
4564 unsigned int cpuset_mems_cookie;
4568 * We also sanity check to catch abuse of atomic reserves being used by
4569 * callers that are not in atomic context.
4571 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4572 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4573 gfp_mask &= ~__GFP_ATOMIC;
4576 compaction_retries = 0;
4577 no_progress_loops = 0;
4578 compact_priority = DEF_COMPACT_PRIORITY;
4579 cpuset_mems_cookie = read_mems_allowed_begin();
4582 * The fast path uses conservative alloc_flags to succeed only until
4583 * kswapd needs to be woken up, and to avoid the cost of setting up
4584 * alloc_flags precisely. So we do that now.
4586 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4589 * We need to recalculate the starting point for the zonelist iterator
4590 * because we might have used different nodemask in the fast path, or
4591 * there was a cpuset modification and we are retrying - otherwise we
4592 * could end up iterating over non-eligible zones endlessly.
4594 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4595 ac->highest_zoneidx, ac->nodemask);
4596 if (!ac->preferred_zoneref->zone)
4599 if (alloc_flags & ALLOC_KSWAPD)
4600 wake_all_kswapds(order, gfp_mask, ac);
4603 * The adjusted alloc_flags might result in immediate success, so try
4606 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4611 * For costly allocations, try direct compaction first, as it's likely
4612 * that we have enough base pages and don't need to reclaim. For non-
4613 * movable high-order allocations, do that as well, as compaction will
4614 * try prevent permanent fragmentation by migrating from blocks of the
4616 * Don't try this for allocations that are allowed to ignore
4617 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4619 if (can_direct_reclaim &&
4621 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4622 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4623 page = __alloc_pages_direct_compact(gfp_mask, order,
4625 INIT_COMPACT_PRIORITY,
4631 * Checks for costly allocations with __GFP_NORETRY, which
4632 * includes some THP page fault allocations
4634 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4636 * If allocating entire pageblock(s) and compaction
4637 * failed because all zones are below low watermarks
4638 * or is prohibited because it recently failed at this
4639 * order, fail immediately unless the allocator has
4640 * requested compaction and reclaim retry.
4643 * - potentially very expensive because zones are far
4644 * below their low watermarks or this is part of very
4645 * bursty high order allocations,
4646 * - not guaranteed to help because isolate_freepages()
4647 * may not iterate over freed pages as part of its
4649 * - unlikely to make entire pageblocks free on its
4652 if (compact_result == COMPACT_SKIPPED ||
4653 compact_result == COMPACT_DEFERRED)
4657 * Looks like reclaim/compaction is worth trying, but
4658 * sync compaction could be very expensive, so keep
4659 * using async compaction.
4661 compact_priority = INIT_COMPACT_PRIORITY;
4666 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4667 if (alloc_flags & ALLOC_KSWAPD)
4668 wake_all_kswapds(order, gfp_mask, ac);
4670 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4672 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4675 * Reset the nodemask and zonelist iterators if memory policies can be
4676 * ignored. These allocations are high priority and system rather than
4679 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4680 ac->nodemask = NULL;
4681 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4682 ac->highest_zoneidx, ac->nodemask);
4685 /* Attempt with potentially adjusted zonelist and alloc_flags */
4686 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4690 /* Caller is not willing to reclaim, we can't balance anything */
4691 if (!can_direct_reclaim)
4694 /* Avoid recursion of direct reclaim */
4695 if (current->flags & PF_MEMALLOC)
4698 /* Try direct reclaim and then allocating */
4699 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4700 &did_some_progress);
4704 /* Try direct compaction and then allocating */
4705 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4706 compact_priority, &compact_result);
4710 /* Do not loop if specifically requested */
4711 if (gfp_mask & __GFP_NORETRY)
4715 * Do not retry costly high order allocations unless they are
4716 * __GFP_RETRY_MAYFAIL
4718 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4721 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4722 did_some_progress > 0, &no_progress_loops))
4726 * It doesn't make any sense to retry for the compaction if the order-0
4727 * reclaim is not able to make any progress because the current
4728 * implementation of the compaction depends on the sufficient amount
4729 * of free memory (see __compaction_suitable)
4731 if (did_some_progress > 0 &&
4732 should_compact_retry(ac, order, alloc_flags,
4733 compact_result, &compact_priority,
4734 &compaction_retries))
4738 /* Deal with possible cpuset update races before we start OOM killing */
4739 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4742 /* Reclaim has failed us, start killing things */
4743 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4747 /* Avoid allocations with no watermarks from looping endlessly */
4748 if (tsk_is_oom_victim(current) &&
4749 (alloc_flags & ALLOC_OOM ||
4750 (gfp_mask & __GFP_NOMEMALLOC)))
4753 /* Retry as long as the OOM killer is making progress */
4754 if (did_some_progress) {
4755 no_progress_loops = 0;
4760 /* Deal with possible cpuset update races before we fail */
4761 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4765 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4768 if (gfp_mask & __GFP_NOFAIL) {
4770 * All existing users of the __GFP_NOFAIL are blockable, so warn
4771 * of any new users that actually require GFP_NOWAIT
4773 if (WARN_ON_ONCE(!can_direct_reclaim))
4777 * PF_MEMALLOC request from this context is rather bizarre
4778 * because we cannot reclaim anything and only can loop waiting
4779 * for somebody to do a work for us
4781 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4784 * non failing costly orders are a hard requirement which we
4785 * are not prepared for much so let's warn about these users
4786 * so that we can identify them and convert them to something
4789 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4792 * Help non-failing allocations by giving them access to memory
4793 * reserves but do not use ALLOC_NO_WATERMARKS because this
4794 * could deplete whole memory reserves which would just make
4795 * the situation worse
4797 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4805 warn_alloc(gfp_mask, ac->nodemask,
4806 "page allocation failure: order:%u", order);
4811 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4812 int preferred_nid, nodemask_t *nodemask,
4813 struct alloc_context *ac, gfp_t *alloc_mask,
4814 unsigned int *alloc_flags)
4816 ac->highest_zoneidx = gfp_zone(gfp_mask);
4817 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4818 ac->nodemask = nodemask;
4819 ac->migratetype = gfp_migratetype(gfp_mask);
4821 if (cpusets_enabled()) {
4822 *alloc_mask |= __GFP_HARDWALL;
4824 * When we are in the interrupt context, it is irrelevant
4825 * to the current task context. It means that any node ok.
4827 if (!in_interrupt() && !ac->nodemask)
4828 ac->nodemask = &cpuset_current_mems_allowed;
4830 *alloc_flags |= ALLOC_CPUSET;
4833 fs_reclaim_acquire(gfp_mask);
4834 fs_reclaim_release(gfp_mask);
4836 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4838 if (should_fail_alloc_page(gfp_mask, order))
4841 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4846 /* Determine whether to spread dirty pages and what the first usable zone */
4847 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4849 /* Dirty zone balancing only done in the fast path */
4850 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4853 * The preferred zone is used for statistics but crucially it is
4854 * also used as the starting point for the zonelist iterator. It
4855 * may get reset for allocations that ignore memory policies.
4857 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4858 ac->highest_zoneidx, ac->nodemask);
4862 * This is the 'heart' of the zoned buddy allocator.
4865 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4866 nodemask_t *nodemask)
4869 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4870 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4871 struct alloc_context ac = { };
4874 * There are several places where we assume that the order value is sane
4875 * so bail out early if the request is out of bound.
4877 if (unlikely(order >= MAX_ORDER)) {
4878 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4882 gfp_mask &= gfp_allowed_mask;
4883 alloc_mask = gfp_mask;
4884 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4887 finalise_ac(gfp_mask, &ac);
4890 * Forbid the first pass from falling back to types that fragment
4891 * memory until all local zones are considered.
4893 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4895 /* First allocation attempt */
4896 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4901 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4902 * resp. GFP_NOIO which has to be inherited for all allocation requests
4903 * from a particular context which has been marked by
4904 * memalloc_no{fs,io}_{save,restore}.
4906 alloc_mask = current_gfp_context(gfp_mask);
4907 ac.spread_dirty_pages = false;
4910 * Restore the original nodemask if it was potentially replaced with
4911 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4913 ac.nodemask = nodemask;
4915 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4918 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4919 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4920 __free_pages(page, order);
4924 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4928 EXPORT_SYMBOL(__alloc_pages_nodemask);
4931 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4932 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4933 * you need to access high mem.
4935 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4939 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4942 return (unsigned long) page_address(page);
4944 EXPORT_SYMBOL(__get_free_pages);
4946 unsigned long get_zeroed_page(gfp_t gfp_mask)
4948 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4950 EXPORT_SYMBOL(get_zeroed_page);
4952 static inline void free_the_page(struct page *page, unsigned int order)
4954 if (order == 0) /* Via pcp? */
4955 free_unref_page(page);
4957 __free_pages_ok(page, order);
4960 void __free_pages(struct page *page, unsigned int order)
4962 if (put_page_testzero(page))
4963 free_the_page(page, order);
4965 EXPORT_SYMBOL(__free_pages);
4967 void free_pages(unsigned long addr, unsigned int order)
4970 VM_BUG_ON(!virt_addr_valid((void *)addr));
4971 __free_pages(virt_to_page((void *)addr), order);
4975 EXPORT_SYMBOL(free_pages);
4979 * An arbitrary-length arbitrary-offset area of memory which resides
4980 * within a 0 or higher order page. Multiple fragments within that page
4981 * are individually refcounted, in the page's reference counter.
4983 * The page_frag functions below provide a simple allocation framework for
4984 * page fragments. This is used by the network stack and network device
4985 * drivers to provide a backing region of memory for use as either an
4986 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4988 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4991 struct page *page = NULL;
4992 gfp_t gfp = gfp_mask;
4994 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4995 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4997 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4998 PAGE_FRAG_CACHE_MAX_ORDER);
4999 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5001 if (unlikely(!page))
5002 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5004 nc->va = page ? page_address(page) : NULL;
5009 void __page_frag_cache_drain(struct page *page, unsigned int count)
5011 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5013 if (page_ref_sub_and_test(page, count))
5014 free_the_page(page, compound_order(page));
5016 EXPORT_SYMBOL(__page_frag_cache_drain);
5018 void *page_frag_alloc(struct page_frag_cache *nc,
5019 unsigned int fragsz, gfp_t gfp_mask)
5021 unsigned int size = PAGE_SIZE;
5025 if (unlikely(!nc->va)) {
5027 page = __page_frag_cache_refill(nc, gfp_mask);
5031 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5032 /* if size can vary use size else just use PAGE_SIZE */
5035 /* Even if we own the page, we do not use atomic_set().
5036 * This would break get_page_unless_zero() users.
5038 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5040 /* reset page count bias and offset to start of new frag */
5041 nc->pfmemalloc = page_is_pfmemalloc(page);
5042 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5046 offset = nc->offset - fragsz;
5047 if (unlikely(offset < 0)) {
5048 page = virt_to_page(nc->va);
5050 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5053 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5054 /* if size can vary use size else just use PAGE_SIZE */
5057 /* OK, page count is 0, we can safely set it */
5058 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5060 /* reset page count bias and offset to start of new frag */
5061 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5062 offset = size - fragsz;
5066 nc->offset = offset;
5068 return nc->va + offset;
5070 EXPORT_SYMBOL(page_frag_alloc);
5073 * Frees a page fragment allocated out of either a compound or order 0 page.
5075 void page_frag_free(void *addr)
5077 struct page *page = virt_to_head_page(addr);
5079 if (unlikely(put_page_testzero(page)))
5080 free_the_page(page, compound_order(page));
5082 EXPORT_SYMBOL(page_frag_free);
5084 static void *make_alloc_exact(unsigned long addr, unsigned int order,
5088 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5089 unsigned long used = addr + PAGE_ALIGN(size);
5091 split_page(virt_to_page((void *)addr), order);
5092 while (used < alloc_end) {
5097 return (void *)addr;
5101 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5102 * @size: the number of bytes to allocate
5103 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5105 * This function is similar to alloc_pages(), except that it allocates the
5106 * minimum number of pages to satisfy the request. alloc_pages() can only
5107 * allocate memory in power-of-two pages.
5109 * This function is also limited by MAX_ORDER.
5111 * Memory allocated by this function must be released by free_pages_exact().
5113 * Return: pointer to the allocated area or %NULL in case of error.
5115 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5117 unsigned int order = get_order(size);
5120 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5121 gfp_mask &= ~__GFP_COMP;
5123 addr = __get_free_pages(gfp_mask, order);
5124 return make_alloc_exact(addr, order, size);
5126 EXPORT_SYMBOL(alloc_pages_exact);
5129 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5131 * @nid: the preferred node ID where memory should be allocated
5132 * @size: the number of bytes to allocate
5133 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5135 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5138 * Return: pointer to the allocated area or %NULL in case of error.
5140 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5142 unsigned int order = get_order(size);
5145 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5146 gfp_mask &= ~__GFP_COMP;
5148 p = alloc_pages_node(nid, gfp_mask, order);
5151 return make_alloc_exact((unsigned long)page_address(p), order, size);
5155 * free_pages_exact - release memory allocated via alloc_pages_exact()
5156 * @virt: the value returned by alloc_pages_exact.
5157 * @size: size of allocation, same value as passed to alloc_pages_exact().
5159 * Release the memory allocated by a previous call to alloc_pages_exact.
5161 void free_pages_exact(void *virt, size_t size)
5163 unsigned long addr = (unsigned long)virt;
5164 unsigned long end = addr + PAGE_ALIGN(size);
5166 while (addr < end) {
5171 EXPORT_SYMBOL(free_pages_exact);
5174 * nr_free_zone_pages - count number of pages beyond high watermark
5175 * @offset: The zone index of the highest zone
5177 * nr_free_zone_pages() counts the number of pages which are beyond the
5178 * high watermark within all zones at or below a given zone index. For each
5179 * zone, the number of pages is calculated as:
5181 * nr_free_zone_pages = managed_pages - high_pages
5183 * Return: number of pages beyond high watermark.
5185 static unsigned long nr_free_zone_pages(int offset)
5190 /* Just pick one node, since fallback list is circular */
5191 unsigned long sum = 0;
5193 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5195 for_each_zone_zonelist(zone, z, zonelist, offset) {
5196 unsigned long size = zone_managed_pages(zone);
5197 unsigned long high = high_wmark_pages(zone);
5206 * nr_free_buffer_pages - count number of pages beyond high watermark
5208 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5209 * watermark within ZONE_DMA and ZONE_NORMAL.
5211 * Return: number of pages beyond high watermark within ZONE_DMA and
5214 unsigned long nr_free_buffer_pages(void)
5216 return nr_free_zone_pages(gfp_zone(GFP_USER));
5218 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5220 static inline void show_node(struct zone *zone)
5222 if (IS_ENABLED(CONFIG_NUMA))
5223 printk("Node %d ", zone_to_nid(zone));
5226 long si_mem_available(void)
5229 unsigned long pagecache;
5230 unsigned long wmark_low = 0;
5231 unsigned long pages[NR_LRU_LISTS];
5232 unsigned long reclaimable;
5236 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5237 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5240 wmark_low += low_wmark_pages(zone);
5243 * Estimate the amount of memory available for userspace allocations,
5244 * without causing swapping.
5246 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5249 * Not all the page cache can be freed, otherwise the system will
5250 * start swapping. Assume at least half of the page cache, or the
5251 * low watermark worth of cache, needs to stay.
5253 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5254 pagecache -= min(pagecache / 2, wmark_low);
5255 available += pagecache;
5258 * Part of the reclaimable slab and other kernel memory consists of
5259 * items that are in use, and cannot be freed. Cap this estimate at the
5262 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5263 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5264 available += reclaimable - min(reclaimable / 2, wmark_low);
5270 EXPORT_SYMBOL_GPL(si_mem_available);
5272 void si_meminfo(struct sysinfo *val)
5274 val->totalram = totalram_pages();
5275 val->sharedram = global_node_page_state(NR_SHMEM);
5276 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5277 val->bufferram = nr_blockdev_pages();
5278 val->totalhigh = totalhigh_pages();
5279 val->freehigh = nr_free_highpages();
5280 val->mem_unit = PAGE_SIZE;
5283 EXPORT_SYMBOL(si_meminfo);
5286 void si_meminfo_node(struct sysinfo *val, int nid)
5288 int zone_type; /* needs to be signed */
5289 unsigned long managed_pages = 0;
5290 unsigned long managed_highpages = 0;
5291 unsigned long free_highpages = 0;
5292 pg_data_t *pgdat = NODE_DATA(nid);
5294 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5295 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5296 val->totalram = managed_pages;
5297 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5298 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5299 #ifdef CONFIG_HIGHMEM
5300 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5301 struct zone *zone = &pgdat->node_zones[zone_type];
5303 if (is_highmem(zone)) {
5304 managed_highpages += zone_managed_pages(zone);
5305 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5308 val->totalhigh = managed_highpages;
5309 val->freehigh = free_highpages;
5311 val->totalhigh = managed_highpages;
5312 val->freehigh = free_highpages;
5314 val->mem_unit = PAGE_SIZE;
5319 * Determine whether the node should be displayed or not, depending on whether
5320 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5322 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5324 if (!(flags & SHOW_MEM_FILTER_NODES))
5328 * no node mask - aka implicit memory numa policy. Do not bother with
5329 * the synchronization - read_mems_allowed_begin - because we do not
5330 * have to be precise here.
5333 nodemask = &cpuset_current_mems_allowed;
5335 return !node_isset(nid, *nodemask);
5338 #define K(x) ((x) << (PAGE_SHIFT-10))
5340 static void show_migration_types(unsigned char type)
5342 static const char types[MIGRATE_TYPES] = {
5343 [MIGRATE_UNMOVABLE] = 'U',
5344 [MIGRATE_MOVABLE] = 'M',
5345 [MIGRATE_RECLAIMABLE] = 'E',
5346 [MIGRATE_HIGHATOMIC] = 'H',
5348 [MIGRATE_CMA] = 'C',
5350 #ifdef CONFIG_MEMORY_ISOLATION
5351 [MIGRATE_ISOLATE] = 'I',
5354 char tmp[MIGRATE_TYPES + 1];
5358 for (i = 0; i < MIGRATE_TYPES; i++) {
5359 if (type & (1 << i))
5364 printk(KERN_CONT "(%s) ", tmp);
5368 * Show free area list (used inside shift_scroll-lock stuff)
5369 * We also calculate the percentage fragmentation. We do this by counting the
5370 * memory on each free list with the exception of the first item on the list.
5373 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5376 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5378 unsigned long free_pcp = 0;
5383 for_each_populated_zone(zone) {
5384 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5387 for_each_online_cpu(cpu)
5388 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5391 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5392 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5393 " unevictable:%lu dirty:%lu writeback:%lu\n"
5394 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5395 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5396 " free:%lu free_pcp:%lu free_cma:%lu\n",
5397 global_node_page_state(NR_ACTIVE_ANON),
5398 global_node_page_state(NR_INACTIVE_ANON),
5399 global_node_page_state(NR_ISOLATED_ANON),
5400 global_node_page_state(NR_ACTIVE_FILE),
5401 global_node_page_state(NR_INACTIVE_FILE),
5402 global_node_page_state(NR_ISOLATED_FILE),
5403 global_node_page_state(NR_UNEVICTABLE),
5404 global_node_page_state(NR_FILE_DIRTY),
5405 global_node_page_state(NR_WRITEBACK),
5406 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5407 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5408 global_node_page_state(NR_FILE_MAPPED),
5409 global_node_page_state(NR_SHMEM),
5410 global_zone_page_state(NR_PAGETABLE),
5411 global_zone_page_state(NR_BOUNCE),
5412 global_zone_page_state(NR_FREE_PAGES),
5414 global_zone_page_state(NR_FREE_CMA_PAGES));
5416 for_each_online_pgdat(pgdat) {
5417 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5421 " active_anon:%lukB"
5422 " inactive_anon:%lukB"
5423 " active_file:%lukB"
5424 " inactive_file:%lukB"
5425 " unevictable:%lukB"
5426 " isolated(anon):%lukB"
5427 " isolated(file):%lukB"
5432 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5434 " shmem_pmdmapped: %lukB"
5437 " writeback_tmp:%lukB"
5438 " kernel_stack:%lukB"
5439 #ifdef CONFIG_SHADOW_CALL_STACK
5440 " shadow_call_stack:%lukB"
5442 " all_unreclaimable? %s"
5445 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5446 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5447 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5448 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5449 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5450 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5451 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5452 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5453 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5454 K(node_page_state(pgdat, NR_WRITEBACK)),
5455 K(node_page_state(pgdat, NR_SHMEM)),
5456 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5457 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5458 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5460 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5462 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5463 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5464 #ifdef CONFIG_SHADOW_CALL_STACK
5465 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5467 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5471 for_each_populated_zone(zone) {
5474 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5478 for_each_online_cpu(cpu)
5479 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5488 " reserved_highatomic:%luKB"
5489 " active_anon:%lukB"
5490 " inactive_anon:%lukB"
5491 " active_file:%lukB"
5492 " inactive_file:%lukB"
5493 " unevictable:%lukB"
5494 " writepending:%lukB"
5505 K(zone_page_state(zone, NR_FREE_PAGES)),
5506 K(min_wmark_pages(zone)),
5507 K(low_wmark_pages(zone)),
5508 K(high_wmark_pages(zone)),
5509 K(zone->nr_reserved_highatomic),
5510 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5511 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5512 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5513 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5514 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5515 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5516 K(zone->present_pages),
5517 K(zone_managed_pages(zone)),
5518 K(zone_page_state(zone, NR_MLOCK)),
5519 K(zone_page_state(zone, NR_PAGETABLE)),
5520 K(zone_page_state(zone, NR_BOUNCE)),
5522 K(this_cpu_read(zone->pageset->pcp.count)),
5523 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5524 printk("lowmem_reserve[]:");
5525 for (i = 0; i < MAX_NR_ZONES; i++)
5526 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5527 printk(KERN_CONT "\n");
5530 for_each_populated_zone(zone) {
5532 unsigned long nr[MAX_ORDER], flags, total = 0;
5533 unsigned char types[MAX_ORDER];
5535 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5538 printk(KERN_CONT "%s: ", zone->name);
5540 spin_lock_irqsave(&zone->lock, flags);
5541 for (order = 0; order < MAX_ORDER; order++) {
5542 struct free_area *area = &zone->free_area[order];
5545 nr[order] = area->nr_free;
5546 total += nr[order] << order;
5549 for (type = 0; type < MIGRATE_TYPES; type++) {
5550 if (!free_area_empty(area, type))
5551 types[order] |= 1 << type;
5554 spin_unlock_irqrestore(&zone->lock, flags);
5555 for (order = 0; order < MAX_ORDER; order++) {
5556 printk(KERN_CONT "%lu*%lukB ",
5557 nr[order], K(1UL) << order);
5559 show_migration_types(types[order]);
5561 printk(KERN_CONT "= %lukB\n", K(total));
5564 hugetlb_show_meminfo();
5566 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5568 show_swap_cache_info();
5571 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5573 zoneref->zone = zone;
5574 zoneref->zone_idx = zone_idx(zone);
5578 * Builds allocation fallback zone lists.
5580 * Add all populated zones of a node to the zonelist.
5582 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5585 enum zone_type zone_type = MAX_NR_ZONES;
5590 zone = pgdat->node_zones + zone_type;
5591 if (managed_zone(zone)) {
5592 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5593 check_highest_zone(zone_type);
5595 } while (zone_type);
5602 static int __parse_numa_zonelist_order(char *s)
5605 * We used to support different zonlists modes but they turned
5606 * out to be just not useful. Let's keep the warning in place
5607 * if somebody still use the cmd line parameter so that we do
5608 * not fail it silently
5610 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5611 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5617 char numa_zonelist_order[] = "Node";
5620 * sysctl handler for numa_zonelist_order
5622 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5623 void *buffer, size_t *length, loff_t *ppos)
5626 return __parse_numa_zonelist_order(buffer);
5627 return proc_dostring(table, write, buffer, length, ppos);
5631 #define MAX_NODE_LOAD (nr_online_nodes)
5632 static int node_load[MAX_NUMNODES];
5635 * find_next_best_node - find the next node that should appear in a given node's fallback list
5636 * @node: node whose fallback list we're appending
5637 * @used_node_mask: nodemask_t of already used nodes
5639 * We use a number of factors to determine which is the next node that should
5640 * appear on a given node's fallback list. The node should not have appeared
5641 * already in @node's fallback list, and it should be the next closest node
5642 * according to the distance array (which contains arbitrary distance values
5643 * from each node to each node in the system), and should also prefer nodes
5644 * with no CPUs, since presumably they'll have very little allocation pressure
5645 * on them otherwise.
5647 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5649 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5652 int min_val = INT_MAX;
5653 int best_node = NUMA_NO_NODE;
5654 const struct cpumask *tmp = cpumask_of_node(0);
5656 /* Use the local node if we haven't already */
5657 if (!node_isset(node, *used_node_mask)) {
5658 node_set(node, *used_node_mask);
5662 for_each_node_state(n, N_MEMORY) {
5664 /* Don't want a node to appear more than once */
5665 if (node_isset(n, *used_node_mask))
5668 /* Use the distance array to find the distance */
5669 val = node_distance(node, n);
5671 /* Penalize nodes under us ("prefer the next node") */
5674 /* Give preference to headless and unused nodes */
5675 tmp = cpumask_of_node(n);
5676 if (!cpumask_empty(tmp))
5677 val += PENALTY_FOR_NODE_WITH_CPUS;
5679 /* Slight preference for less loaded node */
5680 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5681 val += node_load[n];
5683 if (val < min_val) {
5690 node_set(best_node, *used_node_mask);
5697 * Build zonelists ordered by node and zones within node.
5698 * This results in maximum locality--normal zone overflows into local
5699 * DMA zone, if any--but risks exhausting DMA zone.
5701 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5704 struct zoneref *zonerefs;
5707 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5709 for (i = 0; i < nr_nodes; i++) {
5712 pg_data_t *node = NODE_DATA(node_order[i]);
5714 nr_zones = build_zonerefs_node(node, zonerefs);
5715 zonerefs += nr_zones;
5717 zonerefs->zone = NULL;
5718 zonerefs->zone_idx = 0;
5722 * Build gfp_thisnode zonelists
5724 static void build_thisnode_zonelists(pg_data_t *pgdat)
5726 struct zoneref *zonerefs;
5729 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5730 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5731 zonerefs += nr_zones;
5732 zonerefs->zone = NULL;
5733 zonerefs->zone_idx = 0;
5737 * Build zonelists ordered by zone and nodes within zones.
5738 * This results in conserving DMA zone[s] until all Normal memory is
5739 * exhausted, but results in overflowing to remote node while memory
5740 * may still exist in local DMA zone.
5743 static void build_zonelists(pg_data_t *pgdat)
5745 static int node_order[MAX_NUMNODES];
5746 int node, load, nr_nodes = 0;
5747 nodemask_t used_mask = NODE_MASK_NONE;
5748 int local_node, prev_node;
5750 /* NUMA-aware ordering of nodes */
5751 local_node = pgdat->node_id;
5752 load = nr_online_nodes;
5753 prev_node = local_node;
5755 memset(node_order, 0, sizeof(node_order));
5756 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5758 * We don't want to pressure a particular node.
5759 * So adding penalty to the first node in same
5760 * distance group to make it round-robin.
5762 if (node_distance(local_node, node) !=
5763 node_distance(local_node, prev_node))
5764 node_load[node] = load;
5766 node_order[nr_nodes++] = node;
5771 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5772 build_thisnode_zonelists(pgdat);
5775 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5777 * Return node id of node used for "local" allocations.
5778 * I.e., first node id of first zone in arg node's generic zonelist.
5779 * Used for initializing percpu 'numa_mem', which is used primarily
5780 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5782 int local_memory_node(int node)
5786 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5787 gfp_zone(GFP_KERNEL),
5789 return zone_to_nid(z->zone);
5793 static void setup_min_unmapped_ratio(void);
5794 static void setup_min_slab_ratio(void);
5795 #else /* CONFIG_NUMA */
5797 static void build_zonelists(pg_data_t *pgdat)
5799 int node, local_node;
5800 struct zoneref *zonerefs;
5803 local_node = pgdat->node_id;
5805 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5806 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5807 zonerefs += nr_zones;
5810 * Now we build the zonelist so that it contains the zones
5811 * of all the other nodes.
5812 * We don't want to pressure a particular node, so when
5813 * building the zones for node N, we make sure that the
5814 * zones coming right after the local ones are those from
5815 * node N+1 (modulo N)
5817 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5818 if (!node_online(node))
5820 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5821 zonerefs += nr_zones;
5823 for (node = 0; node < local_node; node++) {
5824 if (!node_online(node))
5826 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5827 zonerefs += nr_zones;
5830 zonerefs->zone = NULL;
5831 zonerefs->zone_idx = 0;
5834 #endif /* CONFIG_NUMA */
5837 * Boot pageset table. One per cpu which is going to be used for all
5838 * zones and all nodes. The parameters will be set in such a way
5839 * that an item put on a list will immediately be handed over to
5840 * the buddy list. This is safe since pageset manipulation is done
5841 * with interrupts disabled.
5843 * The boot_pagesets must be kept even after bootup is complete for
5844 * unused processors and/or zones. They do play a role for bootstrapping
5845 * hotplugged processors.
5847 * zoneinfo_show() and maybe other functions do
5848 * not check if the processor is online before following the pageset pointer.
5849 * Other parts of the kernel may not check if the zone is available.
5851 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5852 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5853 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5855 static void __build_all_zonelists(void *data)
5858 int __maybe_unused cpu;
5859 pg_data_t *self = data;
5860 static DEFINE_SPINLOCK(lock);
5865 memset(node_load, 0, sizeof(node_load));
5869 * This node is hotadded and no memory is yet present. So just
5870 * building zonelists is fine - no need to touch other nodes.
5872 if (self && !node_online(self->node_id)) {
5873 build_zonelists(self);
5875 for_each_online_node(nid) {
5876 pg_data_t *pgdat = NODE_DATA(nid);
5878 build_zonelists(pgdat);
5881 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5883 * We now know the "local memory node" for each node--
5884 * i.e., the node of the first zone in the generic zonelist.
5885 * Set up numa_mem percpu variable for on-line cpus. During
5886 * boot, only the boot cpu should be on-line; we'll init the
5887 * secondary cpus' numa_mem as they come on-line. During
5888 * node/memory hotplug, we'll fixup all on-line cpus.
5890 for_each_online_cpu(cpu)
5891 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5898 static noinline void __init
5899 build_all_zonelists_init(void)
5903 __build_all_zonelists(NULL);
5906 * Initialize the boot_pagesets that are going to be used
5907 * for bootstrapping processors. The real pagesets for
5908 * each zone will be allocated later when the per cpu
5909 * allocator is available.
5911 * boot_pagesets are used also for bootstrapping offline
5912 * cpus if the system is already booted because the pagesets
5913 * are needed to initialize allocators on a specific cpu too.
5914 * F.e. the percpu allocator needs the page allocator which
5915 * needs the percpu allocator in order to allocate its pagesets
5916 * (a chicken-egg dilemma).
5918 for_each_possible_cpu(cpu)
5919 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5921 mminit_verify_zonelist();
5922 cpuset_init_current_mems_allowed();
5926 * unless system_state == SYSTEM_BOOTING.
5928 * __ref due to call of __init annotated helper build_all_zonelists_init
5929 * [protected by SYSTEM_BOOTING].
5931 void __ref build_all_zonelists(pg_data_t *pgdat)
5933 unsigned long vm_total_pages;
5935 if (system_state == SYSTEM_BOOTING) {
5936 build_all_zonelists_init();
5938 __build_all_zonelists(pgdat);
5939 /* cpuset refresh routine should be here */
5941 /* Get the number of free pages beyond high watermark in all zones. */
5942 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5944 * Disable grouping by mobility if the number of pages in the
5945 * system is too low to allow the mechanism to work. It would be
5946 * more accurate, but expensive to check per-zone. This check is
5947 * made on memory-hotadd so a system can start with mobility
5948 * disabled and enable it later
5950 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5951 page_group_by_mobility_disabled = 1;
5953 page_group_by_mobility_disabled = 0;
5955 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5957 page_group_by_mobility_disabled ? "off" : "on",
5960 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5964 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5965 static bool __meminit
5966 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5968 static struct memblock_region *r;
5970 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5971 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5972 for_each_memblock(memory, r) {
5973 if (*pfn < memblock_region_memory_end_pfn(r))
5977 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5978 memblock_is_mirror(r)) {
5979 *pfn = memblock_region_memory_end_pfn(r);
5987 * Initially all pages are reserved - free ones are freed
5988 * up by memblock_free_all() once the early boot process is
5989 * done. Non-atomic initialization, single-pass.
5991 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5992 unsigned long start_pfn, enum meminit_context context,
5993 struct vmem_altmap *altmap)
5995 unsigned long pfn, end_pfn = start_pfn + size;
5998 if (highest_memmap_pfn < end_pfn - 1)
5999 highest_memmap_pfn = end_pfn - 1;
6001 #ifdef CONFIG_ZONE_DEVICE
6003 * Honor reservation requested by the driver for this ZONE_DEVICE
6004 * memory. We limit the total number of pages to initialize to just
6005 * those that might contain the memory mapping. We will defer the
6006 * ZONE_DEVICE page initialization until after we have released
6009 if (zone == ZONE_DEVICE) {
6013 if (start_pfn == altmap->base_pfn)
6014 start_pfn += altmap->reserve;
6015 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6019 for (pfn = start_pfn; pfn < end_pfn; ) {
6021 * There can be holes in boot-time mem_map[]s handed to this
6022 * function. They do not exist on hotplugged memory.
6024 if (context == MEMINIT_EARLY) {
6025 if (overlap_memmap_init(zone, &pfn))
6027 if (defer_init(nid, pfn, end_pfn))
6031 page = pfn_to_page(pfn);
6032 __init_single_page(page, pfn, zone, nid);
6033 if (context == MEMINIT_HOTPLUG)
6034 __SetPageReserved(page);
6037 * Mark the block movable so that blocks are reserved for
6038 * movable at startup. This will force kernel allocations
6039 * to reserve their blocks rather than leaking throughout
6040 * the address space during boot when many long-lived
6041 * kernel allocations are made.
6043 * bitmap is created for zone's valid pfn range. but memmap
6044 * can be created for invalid pages (for alignment)
6045 * check here not to call set_pageblock_migratetype() against
6048 if (!(pfn & (pageblock_nr_pages - 1))) {
6049 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6056 #ifdef CONFIG_ZONE_DEVICE
6057 void __ref memmap_init_zone_device(struct zone *zone,
6058 unsigned long start_pfn,
6059 unsigned long nr_pages,
6060 struct dev_pagemap *pgmap)
6062 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6063 struct pglist_data *pgdat = zone->zone_pgdat;
6064 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6065 unsigned long zone_idx = zone_idx(zone);
6066 unsigned long start = jiffies;
6067 int nid = pgdat->node_id;
6069 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6073 * The call to memmap_init_zone should have already taken care
6074 * of the pages reserved for the memmap, so we can just jump to
6075 * the end of that region and start processing the device pages.
6078 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6079 nr_pages = end_pfn - start_pfn;
6082 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6083 struct page *page = pfn_to_page(pfn);
6085 __init_single_page(page, pfn, zone_idx, nid);
6088 * Mark page reserved as it will need to wait for onlining
6089 * phase for it to be fully associated with a zone.
6091 * We can use the non-atomic __set_bit operation for setting
6092 * the flag as we are still initializing the pages.
6094 __SetPageReserved(page);
6097 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6098 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6099 * ever freed or placed on a driver-private list.
6101 page->pgmap = pgmap;
6102 page->zone_device_data = NULL;
6105 * Mark the block movable so that blocks are reserved for
6106 * movable at startup. This will force kernel allocations
6107 * to reserve their blocks rather than leaking throughout
6108 * the address space during boot when many long-lived
6109 * kernel allocations are made.
6111 * bitmap is created for zone's valid pfn range. but memmap
6112 * can be created for invalid pages (for alignment)
6113 * check here not to call set_pageblock_migratetype() against
6116 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6117 * because this is done early in section_activate()
6119 if (!(pfn & (pageblock_nr_pages - 1))) {
6120 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6125 pr_info("%s initialised %lu pages in %ums\n", __func__,
6126 nr_pages, jiffies_to_msecs(jiffies - start));
6130 static void __meminit zone_init_free_lists(struct zone *zone)
6132 unsigned int order, t;
6133 for_each_migratetype_order(order, t) {
6134 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6135 zone->free_area[order].nr_free = 0;
6139 void __meminit __weak memmap_init(unsigned long size, int nid,
6141 unsigned long range_start_pfn)
6143 unsigned long start_pfn, end_pfn;
6144 unsigned long range_end_pfn = range_start_pfn + size;
6147 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6148 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6149 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6151 if (end_pfn > start_pfn) {
6152 size = end_pfn - start_pfn;
6153 memmap_init_zone(size, nid, zone, start_pfn,
6154 MEMINIT_EARLY, NULL);
6159 static int zone_batchsize(struct zone *zone)
6165 * The per-cpu-pages pools are set to around 1000th of the
6168 batch = zone_managed_pages(zone) / 1024;
6169 /* But no more than a meg. */
6170 if (batch * PAGE_SIZE > 1024 * 1024)
6171 batch = (1024 * 1024) / PAGE_SIZE;
6172 batch /= 4; /* We effectively *= 4 below */
6177 * Clamp the batch to a 2^n - 1 value. Having a power
6178 * of 2 value was found to be more likely to have
6179 * suboptimal cache aliasing properties in some cases.
6181 * For example if 2 tasks are alternately allocating
6182 * batches of pages, one task can end up with a lot
6183 * of pages of one half of the possible page colors
6184 * and the other with pages of the other colors.
6186 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6191 /* The deferral and batching of frees should be suppressed under NOMMU
6194 * The problem is that NOMMU needs to be able to allocate large chunks
6195 * of contiguous memory as there's no hardware page translation to
6196 * assemble apparent contiguous memory from discontiguous pages.
6198 * Queueing large contiguous runs of pages for batching, however,
6199 * causes the pages to actually be freed in smaller chunks. As there
6200 * can be a significant delay between the individual batches being
6201 * recycled, this leads to the once large chunks of space being
6202 * fragmented and becoming unavailable for high-order allocations.
6209 * pcp->high and pcp->batch values are related and dependent on one another:
6210 * ->batch must never be higher then ->high.
6211 * The following function updates them in a safe manner without read side
6214 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6215 * those fields changing asynchronously (acording to the above rule).
6217 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6218 * outside of boot time (or some other assurance that no concurrent updaters
6221 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6222 unsigned long batch)
6224 /* start with a fail safe value for batch */
6228 /* Update high, then batch, in order */
6235 /* a companion to pageset_set_high() */
6236 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6238 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6241 static void pageset_init(struct per_cpu_pageset *p)
6243 struct per_cpu_pages *pcp;
6246 memset(p, 0, sizeof(*p));
6249 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6250 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6253 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6256 pageset_set_batch(p, batch);
6260 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6261 * to the value high for the pageset p.
6263 static void pageset_set_high(struct per_cpu_pageset *p,
6266 unsigned long batch = max(1UL, high / 4);
6267 if ((high / 4) > (PAGE_SHIFT * 8))
6268 batch = PAGE_SHIFT * 8;
6270 pageset_update(&p->pcp, high, batch);
6273 static void pageset_set_high_and_batch(struct zone *zone,
6274 struct per_cpu_pageset *pcp)
6276 if (percpu_pagelist_fraction)
6277 pageset_set_high(pcp,
6278 (zone_managed_pages(zone) /
6279 percpu_pagelist_fraction));
6281 pageset_set_batch(pcp, zone_batchsize(zone));
6284 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6286 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6289 pageset_set_high_and_batch(zone, pcp);
6292 void __meminit setup_zone_pageset(struct zone *zone)
6295 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6296 for_each_possible_cpu(cpu)
6297 zone_pageset_init(zone, cpu);
6301 * Allocate per cpu pagesets and initialize them.
6302 * Before this call only boot pagesets were available.
6304 void __init setup_per_cpu_pageset(void)
6306 struct pglist_data *pgdat;
6308 int __maybe_unused cpu;
6310 for_each_populated_zone(zone)
6311 setup_zone_pageset(zone);
6315 * Unpopulated zones continue using the boot pagesets.
6316 * The numa stats for these pagesets need to be reset.
6317 * Otherwise, they will end up skewing the stats of
6318 * the nodes these zones are associated with.
6320 for_each_possible_cpu(cpu) {
6321 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6322 memset(pcp->vm_numa_stat_diff, 0,
6323 sizeof(pcp->vm_numa_stat_diff));
6327 for_each_online_pgdat(pgdat)
6328 pgdat->per_cpu_nodestats =
6329 alloc_percpu(struct per_cpu_nodestat);
6332 static __meminit void zone_pcp_init(struct zone *zone)
6335 * per cpu subsystem is not up at this point. The following code
6336 * relies on the ability of the linker to provide the
6337 * offset of a (static) per cpu variable into the per cpu area.
6339 zone->pageset = &boot_pageset;
6341 if (populated_zone(zone))
6342 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6343 zone->name, zone->present_pages,
6344 zone_batchsize(zone));
6347 void __meminit init_currently_empty_zone(struct zone *zone,
6348 unsigned long zone_start_pfn,
6351 struct pglist_data *pgdat = zone->zone_pgdat;
6352 int zone_idx = zone_idx(zone) + 1;
6354 if (zone_idx > pgdat->nr_zones)
6355 pgdat->nr_zones = zone_idx;
6357 zone->zone_start_pfn = zone_start_pfn;
6359 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6360 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6362 (unsigned long)zone_idx(zone),
6363 zone_start_pfn, (zone_start_pfn + size));
6365 zone_init_free_lists(zone);
6366 zone->initialized = 1;
6370 * get_pfn_range_for_nid - Return the start and end page frames for a node
6371 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6372 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6373 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6375 * It returns the start and end page frame of a node based on information
6376 * provided by memblock_set_node(). If called for a node
6377 * with no available memory, a warning is printed and the start and end
6380 void __init get_pfn_range_for_nid(unsigned int nid,
6381 unsigned long *start_pfn, unsigned long *end_pfn)
6383 unsigned long this_start_pfn, this_end_pfn;
6389 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6390 *start_pfn = min(*start_pfn, this_start_pfn);
6391 *end_pfn = max(*end_pfn, this_end_pfn);
6394 if (*start_pfn == -1UL)
6399 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6400 * assumption is made that zones within a node are ordered in monotonic
6401 * increasing memory addresses so that the "highest" populated zone is used
6403 static void __init find_usable_zone_for_movable(void)
6406 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6407 if (zone_index == ZONE_MOVABLE)
6410 if (arch_zone_highest_possible_pfn[zone_index] >
6411 arch_zone_lowest_possible_pfn[zone_index])
6415 VM_BUG_ON(zone_index == -1);
6416 movable_zone = zone_index;
6420 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6421 * because it is sized independent of architecture. Unlike the other zones,
6422 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6423 * in each node depending on the size of each node and how evenly kernelcore
6424 * is distributed. This helper function adjusts the zone ranges
6425 * provided by the architecture for a given node by using the end of the
6426 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6427 * zones within a node are in order of monotonic increases memory addresses
6429 static void __init adjust_zone_range_for_zone_movable(int nid,
6430 unsigned long zone_type,
6431 unsigned long node_start_pfn,
6432 unsigned long node_end_pfn,
6433 unsigned long *zone_start_pfn,
6434 unsigned long *zone_end_pfn)
6436 /* Only adjust if ZONE_MOVABLE is on this node */
6437 if (zone_movable_pfn[nid]) {
6438 /* Size ZONE_MOVABLE */
6439 if (zone_type == ZONE_MOVABLE) {
6440 *zone_start_pfn = zone_movable_pfn[nid];
6441 *zone_end_pfn = min(node_end_pfn,
6442 arch_zone_highest_possible_pfn[movable_zone]);
6444 /* Adjust for ZONE_MOVABLE starting within this range */
6445 } else if (!mirrored_kernelcore &&
6446 *zone_start_pfn < zone_movable_pfn[nid] &&
6447 *zone_end_pfn > zone_movable_pfn[nid]) {
6448 *zone_end_pfn = zone_movable_pfn[nid];
6450 /* Check if this whole range is within ZONE_MOVABLE */
6451 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6452 *zone_start_pfn = *zone_end_pfn;
6457 * Return the number of pages a zone spans in a node, including holes
6458 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6460 static unsigned long __init zone_spanned_pages_in_node(int nid,
6461 unsigned long zone_type,
6462 unsigned long node_start_pfn,
6463 unsigned long node_end_pfn,
6464 unsigned long *zone_start_pfn,
6465 unsigned long *zone_end_pfn)
6467 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6468 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6469 /* When hotadd a new node from cpu_up(), the node should be empty */
6470 if (!node_start_pfn && !node_end_pfn)
6473 /* Get the start and end of the zone */
6474 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6475 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6476 adjust_zone_range_for_zone_movable(nid, zone_type,
6477 node_start_pfn, node_end_pfn,
6478 zone_start_pfn, zone_end_pfn);
6480 /* Check that this node has pages within the zone's required range */
6481 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6484 /* Move the zone boundaries inside the node if necessary */
6485 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6486 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6488 /* Return the spanned pages */
6489 return *zone_end_pfn - *zone_start_pfn;
6493 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6494 * then all holes in the requested range will be accounted for.
6496 unsigned long __init __absent_pages_in_range(int nid,
6497 unsigned long range_start_pfn,
6498 unsigned long range_end_pfn)
6500 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6501 unsigned long start_pfn, end_pfn;
6504 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6505 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6506 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6507 nr_absent -= end_pfn - start_pfn;
6513 * absent_pages_in_range - Return number of page frames in holes within a range
6514 * @start_pfn: The start PFN to start searching for holes
6515 * @end_pfn: The end PFN to stop searching for holes
6517 * Return: the number of pages frames in memory holes within a range.
6519 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6520 unsigned long end_pfn)
6522 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6525 /* Return the number of page frames in holes in a zone on a node */
6526 static unsigned long __init zone_absent_pages_in_node(int nid,
6527 unsigned long zone_type,
6528 unsigned long node_start_pfn,
6529 unsigned long node_end_pfn)
6531 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6532 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6533 unsigned long zone_start_pfn, zone_end_pfn;
6534 unsigned long nr_absent;
6536 /* When hotadd a new node from cpu_up(), the node should be empty */
6537 if (!node_start_pfn && !node_end_pfn)
6540 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6541 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6543 adjust_zone_range_for_zone_movable(nid, zone_type,
6544 node_start_pfn, node_end_pfn,
6545 &zone_start_pfn, &zone_end_pfn);
6546 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6549 * ZONE_MOVABLE handling.
6550 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6553 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6554 unsigned long start_pfn, end_pfn;
6555 struct memblock_region *r;
6557 for_each_memblock(memory, r) {
6558 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6559 zone_start_pfn, zone_end_pfn);
6560 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6561 zone_start_pfn, zone_end_pfn);
6563 if (zone_type == ZONE_MOVABLE &&
6564 memblock_is_mirror(r))
6565 nr_absent += end_pfn - start_pfn;
6567 if (zone_type == ZONE_NORMAL &&
6568 !memblock_is_mirror(r))
6569 nr_absent += end_pfn - start_pfn;
6576 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6577 unsigned long node_start_pfn,
6578 unsigned long node_end_pfn)
6580 unsigned long realtotalpages = 0, totalpages = 0;
6583 for (i = 0; i < MAX_NR_ZONES; i++) {
6584 struct zone *zone = pgdat->node_zones + i;
6585 unsigned long zone_start_pfn, zone_end_pfn;
6586 unsigned long spanned, absent;
6587 unsigned long size, real_size;
6589 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6594 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6599 real_size = size - absent;
6602 zone->zone_start_pfn = zone_start_pfn;
6604 zone->zone_start_pfn = 0;
6605 zone->spanned_pages = size;
6606 zone->present_pages = real_size;
6609 realtotalpages += real_size;
6612 pgdat->node_spanned_pages = totalpages;
6613 pgdat->node_present_pages = realtotalpages;
6614 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6618 #ifndef CONFIG_SPARSEMEM
6620 * Calculate the size of the zone->blockflags rounded to an unsigned long
6621 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6622 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6623 * round what is now in bits to nearest long in bits, then return it in
6626 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6628 unsigned long usemapsize;
6630 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6631 usemapsize = roundup(zonesize, pageblock_nr_pages);
6632 usemapsize = usemapsize >> pageblock_order;
6633 usemapsize *= NR_PAGEBLOCK_BITS;
6634 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6636 return usemapsize / 8;
6639 static void __ref setup_usemap(struct pglist_data *pgdat,
6641 unsigned long zone_start_pfn,
6642 unsigned long zonesize)
6644 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6645 zone->pageblock_flags = NULL;
6647 zone->pageblock_flags =
6648 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6650 if (!zone->pageblock_flags)
6651 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6652 usemapsize, zone->name, pgdat->node_id);
6656 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6657 unsigned long zone_start_pfn, unsigned long zonesize) {}
6658 #endif /* CONFIG_SPARSEMEM */
6660 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6662 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6663 void __init set_pageblock_order(void)
6667 /* Check that pageblock_nr_pages has not already been setup */
6668 if (pageblock_order)
6671 if (HPAGE_SHIFT > PAGE_SHIFT)
6672 order = HUGETLB_PAGE_ORDER;
6674 order = MAX_ORDER - 1;
6677 * Assume the largest contiguous order of interest is a huge page.
6678 * This value may be variable depending on boot parameters on IA64 and
6681 pageblock_order = order;
6683 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6686 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6687 * is unused as pageblock_order is set at compile-time. See
6688 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6691 void __init set_pageblock_order(void)
6695 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6697 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6698 unsigned long present_pages)
6700 unsigned long pages = spanned_pages;
6703 * Provide a more accurate estimation if there are holes within
6704 * the zone and SPARSEMEM is in use. If there are holes within the
6705 * zone, each populated memory region may cost us one or two extra
6706 * memmap pages due to alignment because memmap pages for each
6707 * populated regions may not be naturally aligned on page boundary.
6708 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6710 if (spanned_pages > present_pages + (present_pages >> 4) &&
6711 IS_ENABLED(CONFIG_SPARSEMEM))
6712 pages = present_pages;
6714 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6717 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6718 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6720 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6722 spin_lock_init(&ds_queue->split_queue_lock);
6723 INIT_LIST_HEAD(&ds_queue->split_queue);
6724 ds_queue->split_queue_len = 0;
6727 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6730 #ifdef CONFIG_COMPACTION
6731 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6733 init_waitqueue_head(&pgdat->kcompactd_wait);
6736 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6739 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6741 pgdat_resize_init(pgdat);
6743 pgdat_init_split_queue(pgdat);
6744 pgdat_init_kcompactd(pgdat);
6746 init_waitqueue_head(&pgdat->kswapd_wait);
6747 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6749 pgdat_page_ext_init(pgdat);
6750 spin_lock_init(&pgdat->lru_lock);
6751 lruvec_init(&pgdat->__lruvec);
6754 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6755 unsigned long remaining_pages)
6757 atomic_long_set(&zone->managed_pages, remaining_pages);
6758 zone_set_nid(zone, nid);
6759 zone->name = zone_names[idx];
6760 zone->zone_pgdat = NODE_DATA(nid);
6761 spin_lock_init(&zone->lock);
6762 zone_seqlock_init(zone);
6763 zone_pcp_init(zone);
6767 * Set up the zone data structures
6768 * - init pgdat internals
6769 * - init all zones belonging to this node
6771 * NOTE: this function is only called during memory hotplug
6773 #ifdef CONFIG_MEMORY_HOTPLUG
6774 void __ref free_area_init_core_hotplug(int nid)
6777 pg_data_t *pgdat = NODE_DATA(nid);
6779 pgdat_init_internals(pgdat);
6780 for (z = 0; z < MAX_NR_ZONES; z++)
6781 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6786 * Set up the zone data structures:
6787 * - mark all pages reserved
6788 * - mark all memory queues empty
6789 * - clear the memory bitmaps
6791 * NOTE: pgdat should get zeroed by caller.
6792 * NOTE: this function is only called during early init.
6794 static void __init free_area_init_core(struct pglist_data *pgdat)
6797 int nid = pgdat->node_id;
6799 pgdat_init_internals(pgdat);
6800 pgdat->per_cpu_nodestats = &boot_nodestats;
6802 for (j = 0; j < MAX_NR_ZONES; j++) {
6803 struct zone *zone = pgdat->node_zones + j;
6804 unsigned long size, freesize, memmap_pages;
6805 unsigned long zone_start_pfn = zone->zone_start_pfn;
6807 size = zone->spanned_pages;
6808 freesize = zone->present_pages;
6811 * Adjust freesize so that it accounts for how much memory
6812 * is used by this zone for memmap. This affects the watermark
6813 * and per-cpu initialisations
6815 memmap_pages = calc_memmap_size(size, freesize);
6816 if (!is_highmem_idx(j)) {
6817 if (freesize >= memmap_pages) {
6818 freesize -= memmap_pages;
6821 " %s zone: %lu pages used for memmap\n",
6822 zone_names[j], memmap_pages);
6824 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6825 zone_names[j], memmap_pages, freesize);
6828 /* Account for reserved pages */
6829 if (j == 0 && freesize > dma_reserve) {
6830 freesize -= dma_reserve;
6831 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6832 zone_names[0], dma_reserve);
6835 if (!is_highmem_idx(j))
6836 nr_kernel_pages += freesize;
6837 /* Charge for highmem memmap if there are enough kernel pages */
6838 else if (nr_kernel_pages > memmap_pages * 2)
6839 nr_kernel_pages -= memmap_pages;
6840 nr_all_pages += freesize;
6843 * Set an approximate value for lowmem here, it will be adjusted
6844 * when the bootmem allocator frees pages into the buddy system.
6845 * And all highmem pages will be managed by the buddy system.
6847 zone_init_internals(zone, j, nid, freesize);
6852 set_pageblock_order();
6853 setup_usemap(pgdat, zone, zone_start_pfn, size);
6854 init_currently_empty_zone(zone, zone_start_pfn, size);
6855 memmap_init(size, nid, j, zone_start_pfn);
6859 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6860 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6862 unsigned long __maybe_unused start = 0;
6863 unsigned long __maybe_unused offset = 0;
6865 /* Skip empty nodes */
6866 if (!pgdat->node_spanned_pages)
6869 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6870 offset = pgdat->node_start_pfn - start;
6871 /* ia64 gets its own node_mem_map, before this, without bootmem */
6872 if (!pgdat->node_mem_map) {
6873 unsigned long size, end;
6877 * The zone's endpoints aren't required to be MAX_ORDER
6878 * aligned but the node_mem_map endpoints must be in order
6879 * for the buddy allocator to function correctly.
6881 end = pgdat_end_pfn(pgdat);
6882 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6883 size = (end - start) * sizeof(struct page);
6884 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6887 panic("Failed to allocate %ld bytes for node %d memory map\n",
6888 size, pgdat->node_id);
6889 pgdat->node_mem_map = map + offset;
6891 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6892 __func__, pgdat->node_id, (unsigned long)pgdat,
6893 (unsigned long)pgdat->node_mem_map);
6894 #ifndef CONFIG_NEED_MULTIPLE_NODES
6896 * With no DISCONTIG, the global mem_map is just set as node 0's
6898 if (pgdat == NODE_DATA(0)) {
6899 mem_map = NODE_DATA(0)->node_mem_map;
6900 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6906 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6907 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6909 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6910 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6912 pgdat->first_deferred_pfn = ULONG_MAX;
6915 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6918 static void __init free_area_init_node(int nid)
6920 pg_data_t *pgdat = NODE_DATA(nid);
6921 unsigned long start_pfn = 0;
6922 unsigned long end_pfn = 0;
6924 /* pg_data_t should be reset to zero when it's allocated */
6925 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6927 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6929 pgdat->node_id = nid;
6930 pgdat->node_start_pfn = start_pfn;
6931 pgdat->per_cpu_nodestats = NULL;
6933 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6934 (u64)start_pfn << PAGE_SHIFT,
6935 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6936 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6938 alloc_node_mem_map(pgdat);
6939 pgdat_set_deferred_range(pgdat);
6941 free_area_init_core(pgdat);
6944 void __init free_area_init_memoryless_node(int nid)
6946 free_area_init_node(nid);
6949 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6951 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6952 * PageReserved(). Return the number of struct pages that were initialized.
6954 static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6959 for (pfn = spfn; pfn < epfn; pfn++) {
6960 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6961 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6962 + pageblock_nr_pages - 1;
6966 * Use a fake node/zone (0) for now. Some of these pages
6967 * (in memblock.reserved but not in memblock.memory) will
6968 * get re-initialized via reserve_bootmem_region() later.
6970 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6971 __SetPageReserved(pfn_to_page(pfn));
6979 * Only struct pages that are backed by physical memory are zeroed and
6980 * initialized by going through __init_single_page(). But, there are some
6981 * struct pages which are reserved in memblock allocator and their fields
6982 * may be accessed (for example page_to_pfn() on some configuration accesses
6983 * flags). We must explicitly initialize those struct pages.
6985 * This function also addresses a similar issue where struct pages are left
6986 * uninitialized because the physical address range is not covered by
6987 * memblock.memory or memblock.reserved. That could happen when memblock
6988 * layout is manually configured via memmap=, or when the highest physical
6989 * address (max_pfn) does not end on a section boundary.
6991 static void __init init_unavailable_mem(void)
6993 phys_addr_t start, end;
6995 phys_addr_t next = 0;
6998 * Loop through unavailable ranges not covered by memblock.memory.
7001 for_each_mem_range(i, &memblock.memory, NULL,
7002 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
7004 pgcnt += init_unavailable_range(PFN_DOWN(next),
7010 * Early sections always have a fully populated memmap for the whole
7011 * section - see pfn_valid(). If the last section has holes at the
7012 * end and that section is marked "online", the memmap will be
7013 * considered initialized. Make sure that memmap has a well defined
7016 pgcnt += init_unavailable_range(PFN_DOWN(next),
7017 round_up(max_pfn, PAGES_PER_SECTION));
7020 * Struct pages that do not have backing memory. This could be because
7021 * firmware is using some of this memory, or for some other reasons.
7024 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7027 static inline void __init init_unavailable_mem(void)
7030 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7032 #if MAX_NUMNODES > 1
7034 * Figure out the number of possible node ids.
7036 void __init setup_nr_node_ids(void)
7038 unsigned int highest;
7040 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7041 nr_node_ids = highest + 1;
7046 * node_map_pfn_alignment - determine the maximum internode alignment
7048 * This function should be called after node map is populated and sorted.
7049 * It calculates the maximum power of two alignment which can distinguish
7052 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7053 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7054 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7055 * shifted, 1GiB is enough and this function will indicate so.
7057 * This is used to test whether pfn -> nid mapping of the chosen memory
7058 * model has fine enough granularity to avoid incorrect mapping for the
7059 * populated node map.
7061 * Return: the determined alignment in pfn's. 0 if there is no alignment
7062 * requirement (single node).
7064 unsigned long __init node_map_pfn_alignment(void)
7066 unsigned long accl_mask = 0, last_end = 0;
7067 unsigned long start, end, mask;
7068 int last_nid = NUMA_NO_NODE;
7071 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7072 if (!start || last_nid < 0 || last_nid == nid) {
7079 * Start with a mask granular enough to pin-point to the
7080 * start pfn and tick off bits one-by-one until it becomes
7081 * too coarse to separate the current node from the last.
7083 mask = ~((1 << __ffs(start)) - 1);
7084 while (mask && last_end <= (start & (mask << 1)))
7087 /* accumulate all internode masks */
7091 /* convert mask to number of pages */
7092 return ~accl_mask + 1;
7096 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7098 * Return: the minimum PFN based on information provided via
7099 * memblock_set_node().
7101 unsigned long __init find_min_pfn_with_active_regions(void)
7103 return PHYS_PFN(memblock_start_of_DRAM());
7107 * early_calculate_totalpages()
7108 * Sum pages in active regions for movable zone.
7109 * Populate N_MEMORY for calculating usable_nodes.
7111 static unsigned long __init early_calculate_totalpages(void)
7113 unsigned long totalpages = 0;
7114 unsigned long start_pfn, end_pfn;
7117 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7118 unsigned long pages = end_pfn - start_pfn;
7120 totalpages += pages;
7122 node_set_state(nid, N_MEMORY);
7128 * Find the PFN the Movable zone begins in each node. Kernel memory
7129 * is spread evenly between nodes as long as the nodes have enough
7130 * memory. When they don't, some nodes will have more kernelcore than
7133 static void __init find_zone_movable_pfns_for_nodes(void)
7136 unsigned long usable_startpfn;
7137 unsigned long kernelcore_node, kernelcore_remaining;
7138 /* save the state before borrow the nodemask */
7139 nodemask_t saved_node_state = node_states[N_MEMORY];
7140 unsigned long totalpages = early_calculate_totalpages();
7141 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7142 struct memblock_region *r;
7144 /* Need to find movable_zone earlier when movable_node is specified. */
7145 find_usable_zone_for_movable();
7148 * If movable_node is specified, ignore kernelcore and movablecore
7151 if (movable_node_is_enabled()) {
7152 for_each_memblock(memory, r) {
7153 if (!memblock_is_hotpluggable(r))
7156 nid = memblock_get_region_node(r);
7158 usable_startpfn = PFN_DOWN(r->base);
7159 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7160 min(usable_startpfn, zone_movable_pfn[nid]) :
7168 * If kernelcore=mirror is specified, ignore movablecore option
7170 if (mirrored_kernelcore) {
7171 bool mem_below_4gb_not_mirrored = false;
7173 for_each_memblock(memory, r) {
7174 if (memblock_is_mirror(r))
7177 nid = memblock_get_region_node(r);
7179 usable_startpfn = memblock_region_memory_base_pfn(r);
7181 if (usable_startpfn < 0x100000) {
7182 mem_below_4gb_not_mirrored = true;
7186 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7187 min(usable_startpfn, zone_movable_pfn[nid]) :
7191 if (mem_below_4gb_not_mirrored)
7192 pr_warn("This configuration results in unmirrored kernel memory.\n");
7198 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7199 * amount of necessary memory.
7201 if (required_kernelcore_percent)
7202 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7204 if (required_movablecore_percent)
7205 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7209 * If movablecore= was specified, calculate what size of
7210 * kernelcore that corresponds so that memory usable for
7211 * any allocation type is evenly spread. If both kernelcore
7212 * and movablecore are specified, then the value of kernelcore
7213 * will be used for required_kernelcore if it's greater than
7214 * what movablecore would have allowed.
7216 if (required_movablecore) {
7217 unsigned long corepages;
7220 * Round-up so that ZONE_MOVABLE is at least as large as what
7221 * was requested by the user
7223 required_movablecore =
7224 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7225 required_movablecore = min(totalpages, required_movablecore);
7226 corepages = totalpages - required_movablecore;
7228 required_kernelcore = max(required_kernelcore, corepages);
7232 * If kernelcore was not specified or kernelcore size is larger
7233 * than totalpages, there is no ZONE_MOVABLE.
7235 if (!required_kernelcore || required_kernelcore >= totalpages)
7238 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7239 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7242 /* Spread kernelcore memory as evenly as possible throughout nodes */
7243 kernelcore_node = required_kernelcore / usable_nodes;
7244 for_each_node_state(nid, N_MEMORY) {
7245 unsigned long start_pfn, end_pfn;
7248 * Recalculate kernelcore_node if the division per node
7249 * now exceeds what is necessary to satisfy the requested
7250 * amount of memory for the kernel
7252 if (required_kernelcore < kernelcore_node)
7253 kernelcore_node = required_kernelcore / usable_nodes;
7256 * As the map is walked, we track how much memory is usable
7257 * by the kernel using kernelcore_remaining. When it is
7258 * 0, the rest of the node is usable by ZONE_MOVABLE
7260 kernelcore_remaining = kernelcore_node;
7262 /* Go through each range of PFNs within this node */
7263 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7264 unsigned long size_pages;
7266 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7267 if (start_pfn >= end_pfn)
7270 /* Account for what is only usable for kernelcore */
7271 if (start_pfn < usable_startpfn) {
7272 unsigned long kernel_pages;
7273 kernel_pages = min(end_pfn, usable_startpfn)
7276 kernelcore_remaining -= min(kernel_pages,
7277 kernelcore_remaining);
7278 required_kernelcore -= min(kernel_pages,
7279 required_kernelcore);
7281 /* Continue if range is now fully accounted */
7282 if (end_pfn <= usable_startpfn) {
7285 * Push zone_movable_pfn to the end so
7286 * that if we have to rebalance
7287 * kernelcore across nodes, we will
7288 * not double account here
7290 zone_movable_pfn[nid] = end_pfn;
7293 start_pfn = usable_startpfn;
7297 * The usable PFN range for ZONE_MOVABLE is from
7298 * start_pfn->end_pfn. Calculate size_pages as the
7299 * number of pages used as kernelcore
7301 size_pages = end_pfn - start_pfn;
7302 if (size_pages > kernelcore_remaining)
7303 size_pages = kernelcore_remaining;
7304 zone_movable_pfn[nid] = start_pfn + size_pages;
7307 * Some kernelcore has been met, update counts and
7308 * break if the kernelcore for this node has been
7311 required_kernelcore -= min(required_kernelcore,
7313 kernelcore_remaining -= size_pages;
7314 if (!kernelcore_remaining)
7320 * If there is still required_kernelcore, we do another pass with one
7321 * less node in the count. This will push zone_movable_pfn[nid] further
7322 * along on the nodes that still have memory until kernelcore is
7326 if (usable_nodes && required_kernelcore > usable_nodes)
7330 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7331 for (nid = 0; nid < MAX_NUMNODES; nid++)
7332 zone_movable_pfn[nid] =
7333 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7336 /* restore the node_state */
7337 node_states[N_MEMORY] = saved_node_state;
7340 /* Any regular or high memory on that node ? */
7341 static void check_for_memory(pg_data_t *pgdat, int nid)
7343 enum zone_type zone_type;
7345 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7346 struct zone *zone = &pgdat->node_zones[zone_type];
7347 if (populated_zone(zone)) {
7348 if (IS_ENABLED(CONFIG_HIGHMEM))
7349 node_set_state(nid, N_HIGH_MEMORY);
7350 if (zone_type <= ZONE_NORMAL)
7351 node_set_state(nid, N_NORMAL_MEMORY);
7358 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7359 * such cases we allow max_zone_pfn sorted in the descending order
7361 bool __weak arch_has_descending_max_zone_pfns(void)
7367 * free_area_init - Initialise all pg_data_t and zone data
7368 * @max_zone_pfn: an array of max PFNs for each zone
7370 * This will call free_area_init_node() for each active node in the system.
7371 * Using the page ranges provided by memblock_set_node(), the size of each
7372 * zone in each node and their holes is calculated. If the maximum PFN
7373 * between two adjacent zones match, it is assumed that the zone is empty.
7374 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7375 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7376 * starts where the previous one ended. For example, ZONE_DMA32 starts
7377 * at arch_max_dma_pfn.
7379 void __init free_area_init(unsigned long *max_zone_pfn)
7381 unsigned long start_pfn, end_pfn;
7385 /* Record where the zone boundaries are */
7386 memset(arch_zone_lowest_possible_pfn, 0,
7387 sizeof(arch_zone_lowest_possible_pfn));
7388 memset(arch_zone_highest_possible_pfn, 0,
7389 sizeof(arch_zone_highest_possible_pfn));
7391 start_pfn = find_min_pfn_with_active_regions();
7392 descending = arch_has_descending_max_zone_pfns();
7394 for (i = 0; i < MAX_NR_ZONES; i++) {
7396 zone = MAX_NR_ZONES - i - 1;
7400 if (zone == ZONE_MOVABLE)
7403 end_pfn = max(max_zone_pfn[zone], start_pfn);
7404 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7405 arch_zone_highest_possible_pfn[zone] = end_pfn;
7407 start_pfn = end_pfn;
7410 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7411 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7412 find_zone_movable_pfns_for_nodes();
7414 /* Print out the zone ranges */
7415 pr_info("Zone ranges:\n");
7416 for (i = 0; i < MAX_NR_ZONES; i++) {
7417 if (i == ZONE_MOVABLE)
7419 pr_info(" %-8s ", zone_names[i]);
7420 if (arch_zone_lowest_possible_pfn[i] ==
7421 arch_zone_highest_possible_pfn[i])
7424 pr_cont("[mem %#018Lx-%#018Lx]\n",
7425 (u64)arch_zone_lowest_possible_pfn[i]
7427 ((u64)arch_zone_highest_possible_pfn[i]
7428 << PAGE_SHIFT) - 1);
7431 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7432 pr_info("Movable zone start for each node\n");
7433 for (i = 0; i < MAX_NUMNODES; i++) {
7434 if (zone_movable_pfn[i])
7435 pr_info(" Node %d: %#018Lx\n", i,
7436 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7440 * Print out the early node map, and initialize the
7441 * subsection-map relative to active online memory ranges to
7442 * enable future "sub-section" extensions of the memory map.
7444 pr_info("Early memory node ranges\n");
7445 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7446 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7447 (u64)start_pfn << PAGE_SHIFT,
7448 ((u64)end_pfn << PAGE_SHIFT) - 1);
7449 subsection_map_init(start_pfn, end_pfn - start_pfn);
7452 /* Initialise every node */
7453 mminit_verify_pageflags_layout();
7454 setup_nr_node_ids();
7455 init_unavailable_mem();
7456 for_each_online_node(nid) {
7457 pg_data_t *pgdat = NODE_DATA(nid);
7458 free_area_init_node(nid);
7460 /* Any memory on that node */
7461 if (pgdat->node_present_pages)
7462 node_set_state(nid, N_MEMORY);
7463 check_for_memory(pgdat, nid);
7467 static int __init cmdline_parse_core(char *p, unsigned long *core,
7468 unsigned long *percent)
7470 unsigned long long coremem;
7476 /* Value may be a percentage of total memory, otherwise bytes */
7477 coremem = simple_strtoull(p, &endptr, 0);
7478 if (*endptr == '%') {
7479 /* Paranoid check for percent values greater than 100 */
7480 WARN_ON(coremem > 100);
7484 coremem = memparse(p, &p);
7485 /* Paranoid check that UL is enough for the coremem value */
7486 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7488 *core = coremem >> PAGE_SHIFT;
7495 * kernelcore=size sets the amount of memory for use for allocations that
7496 * cannot be reclaimed or migrated.
7498 static int __init cmdline_parse_kernelcore(char *p)
7500 /* parse kernelcore=mirror */
7501 if (parse_option_str(p, "mirror")) {
7502 mirrored_kernelcore = true;
7506 return cmdline_parse_core(p, &required_kernelcore,
7507 &required_kernelcore_percent);
7511 * movablecore=size sets the amount of memory for use for allocations that
7512 * can be reclaimed or migrated.
7514 static int __init cmdline_parse_movablecore(char *p)
7516 return cmdline_parse_core(p, &required_movablecore,
7517 &required_movablecore_percent);
7520 early_param("kernelcore", cmdline_parse_kernelcore);
7521 early_param("movablecore", cmdline_parse_movablecore);
7523 void adjust_managed_page_count(struct page *page, long count)
7525 atomic_long_add(count, &page_zone(page)->managed_pages);
7526 totalram_pages_add(count);
7527 #ifdef CONFIG_HIGHMEM
7528 if (PageHighMem(page))
7529 totalhigh_pages_add(count);
7532 EXPORT_SYMBOL(adjust_managed_page_count);
7534 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7537 unsigned long pages = 0;
7539 start = (void *)PAGE_ALIGN((unsigned long)start);
7540 end = (void *)((unsigned long)end & PAGE_MASK);
7541 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7542 struct page *page = virt_to_page(pos);
7543 void *direct_map_addr;
7546 * 'direct_map_addr' might be different from 'pos'
7547 * because some architectures' virt_to_page()
7548 * work with aliases. Getting the direct map
7549 * address ensures that we get a _writeable_
7550 * alias for the memset().
7552 direct_map_addr = page_address(page);
7553 if ((unsigned int)poison <= 0xFF)
7554 memset(direct_map_addr, poison, PAGE_SIZE);
7556 free_reserved_page(page);
7560 pr_info("Freeing %s memory: %ldK\n",
7561 s, pages << (PAGE_SHIFT - 10));
7566 #ifdef CONFIG_HIGHMEM
7567 void free_highmem_page(struct page *page)
7569 __free_reserved_page(page);
7570 totalram_pages_inc();
7571 atomic_long_inc(&page_zone(page)->managed_pages);
7572 totalhigh_pages_inc();
7577 void __init mem_init_print_info(const char *str)
7579 unsigned long physpages, codesize, datasize, rosize, bss_size;
7580 unsigned long init_code_size, init_data_size;
7582 physpages = get_num_physpages();
7583 codesize = _etext - _stext;
7584 datasize = _edata - _sdata;
7585 rosize = __end_rodata - __start_rodata;
7586 bss_size = __bss_stop - __bss_start;
7587 init_data_size = __init_end - __init_begin;
7588 init_code_size = _einittext - _sinittext;
7591 * Detect special cases and adjust section sizes accordingly:
7592 * 1) .init.* may be embedded into .data sections
7593 * 2) .init.text.* may be out of [__init_begin, __init_end],
7594 * please refer to arch/tile/kernel/vmlinux.lds.S.
7595 * 3) .rodata.* may be embedded into .text or .data sections.
7597 #define adj_init_size(start, end, size, pos, adj) \
7599 if (start <= pos && pos < end && size > adj) \
7603 adj_init_size(__init_begin, __init_end, init_data_size,
7604 _sinittext, init_code_size);
7605 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7606 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7607 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7608 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7610 #undef adj_init_size
7612 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7613 #ifdef CONFIG_HIGHMEM
7617 nr_free_pages() << (PAGE_SHIFT - 10),
7618 physpages << (PAGE_SHIFT - 10),
7619 codesize >> 10, datasize >> 10, rosize >> 10,
7620 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7621 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7622 totalcma_pages << (PAGE_SHIFT - 10),
7623 #ifdef CONFIG_HIGHMEM
7624 totalhigh_pages() << (PAGE_SHIFT - 10),
7626 str ? ", " : "", str ? str : "");
7630 * set_dma_reserve - set the specified number of pages reserved in the first zone
7631 * @new_dma_reserve: The number of pages to mark reserved
7633 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7634 * In the DMA zone, a significant percentage may be consumed by kernel image
7635 * and other unfreeable allocations which can skew the watermarks badly. This
7636 * function may optionally be used to account for unfreeable pages in the
7637 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7638 * smaller per-cpu batchsize.
7640 void __init set_dma_reserve(unsigned long new_dma_reserve)
7642 dma_reserve = new_dma_reserve;
7645 static int page_alloc_cpu_dead(unsigned int cpu)
7648 lru_add_drain_cpu(cpu);
7652 * Spill the event counters of the dead processor
7653 * into the current processors event counters.
7654 * This artificially elevates the count of the current
7657 vm_events_fold_cpu(cpu);
7660 * Zero the differential counters of the dead processor
7661 * so that the vm statistics are consistent.
7663 * This is only okay since the processor is dead and cannot
7664 * race with what we are doing.
7666 cpu_vm_stats_fold(cpu);
7671 int hashdist = HASHDIST_DEFAULT;
7673 static int __init set_hashdist(char *str)
7677 hashdist = simple_strtoul(str, &str, 0);
7680 __setup("hashdist=", set_hashdist);
7683 void __init page_alloc_init(void)
7688 if (num_node_state(N_MEMORY) == 1)
7692 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7693 "mm/page_alloc:dead", NULL,
7694 page_alloc_cpu_dead);
7699 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7700 * or min_free_kbytes changes.
7702 static void calculate_totalreserve_pages(void)
7704 struct pglist_data *pgdat;
7705 unsigned long reserve_pages = 0;
7706 enum zone_type i, j;
7708 for_each_online_pgdat(pgdat) {
7710 pgdat->totalreserve_pages = 0;
7712 for (i = 0; i < MAX_NR_ZONES; i++) {
7713 struct zone *zone = pgdat->node_zones + i;
7715 unsigned long managed_pages = zone_managed_pages(zone);
7717 /* Find valid and maximum lowmem_reserve in the zone */
7718 for (j = i; j < MAX_NR_ZONES; j++) {
7719 if (zone->lowmem_reserve[j] > max)
7720 max = zone->lowmem_reserve[j];
7723 /* we treat the high watermark as reserved pages. */
7724 max += high_wmark_pages(zone);
7726 if (max > managed_pages)
7727 max = managed_pages;
7729 pgdat->totalreserve_pages += max;
7731 reserve_pages += max;
7734 totalreserve_pages = reserve_pages;
7738 * setup_per_zone_lowmem_reserve - called whenever
7739 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7740 * has a correct pages reserved value, so an adequate number of
7741 * pages are left in the zone after a successful __alloc_pages().
7743 static void setup_per_zone_lowmem_reserve(void)
7745 struct pglist_data *pgdat;
7746 enum zone_type j, idx;
7748 for_each_online_pgdat(pgdat) {
7749 for (j = 0; j < MAX_NR_ZONES; j++) {
7750 struct zone *zone = pgdat->node_zones + j;
7751 unsigned long managed_pages = zone_managed_pages(zone);
7753 zone->lowmem_reserve[j] = 0;
7757 struct zone *lower_zone;
7760 lower_zone = pgdat->node_zones + idx;
7762 if (!sysctl_lowmem_reserve_ratio[idx] ||
7763 !zone_managed_pages(lower_zone)) {
7764 lower_zone->lowmem_reserve[j] = 0;
7767 lower_zone->lowmem_reserve[j] =
7768 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7770 managed_pages += zone_managed_pages(lower_zone);
7775 /* update totalreserve_pages */
7776 calculate_totalreserve_pages();
7779 static void __setup_per_zone_wmarks(void)
7781 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7782 unsigned long lowmem_pages = 0;
7784 unsigned long flags;
7786 /* Calculate total number of !ZONE_HIGHMEM pages */
7787 for_each_zone(zone) {
7788 if (!is_highmem(zone))
7789 lowmem_pages += zone_managed_pages(zone);
7792 for_each_zone(zone) {
7795 spin_lock_irqsave(&zone->lock, flags);
7796 tmp = (u64)pages_min * zone_managed_pages(zone);
7797 do_div(tmp, lowmem_pages);
7798 if (is_highmem(zone)) {
7800 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7801 * need highmem pages, so cap pages_min to a small
7804 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7805 * deltas control async page reclaim, and so should
7806 * not be capped for highmem.
7808 unsigned long min_pages;
7810 min_pages = zone_managed_pages(zone) / 1024;
7811 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7812 zone->_watermark[WMARK_MIN] = min_pages;
7815 * If it's a lowmem zone, reserve a number of pages
7816 * proportionate to the zone's size.
7818 zone->_watermark[WMARK_MIN] = tmp;
7822 * Set the kswapd watermarks distance according to the
7823 * scale factor in proportion to available memory, but
7824 * ensure a minimum size on small systems.
7826 tmp = max_t(u64, tmp >> 2,
7827 mult_frac(zone_managed_pages(zone),
7828 watermark_scale_factor, 10000));
7830 zone->watermark_boost = 0;
7831 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7832 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7834 spin_unlock_irqrestore(&zone->lock, flags);
7837 /* update totalreserve_pages */
7838 calculate_totalreserve_pages();
7842 * setup_per_zone_wmarks - called when min_free_kbytes changes
7843 * or when memory is hot-{added|removed}
7845 * Ensures that the watermark[min,low,high] values for each zone are set
7846 * correctly with respect to min_free_kbytes.
7848 void setup_per_zone_wmarks(void)
7850 static DEFINE_SPINLOCK(lock);
7853 __setup_per_zone_wmarks();
7858 * Initialise min_free_kbytes.
7860 * For small machines we want it small (128k min). For large machines
7861 * we want it large (256MB max). But it is not linear, because network
7862 * bandwidth does not increase linearly with machine size. We use
7864 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7865 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7881 int __meminit init_per_zone_wmark_min(void)
7883 unsigned long lowmem_kbytes;
7884 int new_min_free_kbytes;
7886 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7887 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7889 if (new_min_free_kbytes > user_min_free_kbytes) {
7890 min_free_kbytes = new_min_free_kbytes;
7891 if (min_free_kbytes < 128)
7892 min_free_kbytes = 128;
7893 if (min_free_kbytes > 262144)
7894 min_free_kbytes = 262144;
7896 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7897 new_min_free_kbytes, user_min_free_kbytes);
7899 setup_per_zone_wmarks();
7900 refresh_zone_stat_thresholds();
7901 setup_per_zone_lowmem_reserve();
7904 setup_min_unmapped_ratio();
7905 setup_min_slab_ratio();
7908 khugepaged_min_free_kbytes_update();
7912 postcore_initcall(init_per_zone_wmark_min)
7915 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7916 * that we can call two helper functions whenever min_free_kbytes
7919 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7920 void *buffer, size_t *length, loff_t *ppos)
7924 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7929 user_min_free_kbytes = min_free_kbytes;
7930 setup_per_zone_wmarks();
7935 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7936 void *buffer, size_t *length, loff_t *ppos)
7940 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7945 setup_per_zone_wmarks();
7951 static void setup_min_unmapped_ratio(void)
7956 for_each_online_pgdat(pgdat)
7957 pgdat->min_unmapped_pages = 0;
7960 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7961 sysctl_min_unmapped_ratio) / 100;
7965 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7966 void *buffer, size_t *length, loff_t *ppos)
7970 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7974 setup_min_unmapped_ratio();
7979 static void setup_min_slab_ratio(void)
7984 for_each_online_pgdat(pgdat)
7985 pgdat->min_slab_pages = 0;
7988 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7989 sysctl_min_slab_ratio) / 100;
7992 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7993 void *buffer, size_t *length, loff_t *ppos)
7997 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8001 setup_min_slab_ratio();
8008 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8009 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8010 * whenever sysctl_lowmem_reserve_ratio changes.
8012 * The reserve ratio obviously has absolutely no relation with the
8013 * minimum watermarks. The lowmem reserve ratio can only make sense
8014 * if in function of the boot time zone sizes.
8016 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8017 void *buffer, size_t *length, loff_t *ppos)
8021 proc_dointvec_minmax(table, write, buffer, length, ppos);
8023 for (i = 0; i < MAX_NR_ZONES; i++) {
8024 if (sysctl_lowmem_reserve_ratio[i] < 1)
8025 sysctl_lowmem_reserve_ratio[i] = 0;
8028 setup_per_zone_lowmem_reserve();
8032 static void __zone_pcp_update(struct zone *zone)
8036 for_each_possible_cpu(cpu)
8037 pageset_set_high_and_batch(zone,
8038 per_cpu_ptr(zone->pageset, cpu));
8042 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8043 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8044 * pagelist can have before it gets flushed back to buddy allocator.
8046 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8047 void *buffer, size_t *length, loff_t *ppos)
8050 int old_percpu_pagelist_fraction;
8053 mutex_lock(&pcp_batch_high_lock);
8054 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8056 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8057 if (!write || ret < 0)
8060 /* Sanity checking to avoid pcp imbalance */
8061 if (percpu_pagelist_fraction &&
8062 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8063 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8069 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8072 for_each_populated_zone(zone)
8073 __zone_pcp_update(zone);
8075 mutex_unlock(&pcp_batch_high_lock);
8079 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8081 * Returns the number of pages that arch has reserved but
8082 * is not known to alloc_large_system_hash().
8084 static unsigned long __init arch_reserved_kernel_pages(void)
8091 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8092 * machines. As memory size is increased the scale is also increased but at
8093 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8094 * quadruples the scale is increased by one, which means the size of hash table
8095 * only doubles, instead of quadrupling as well.
8096 * Because 32-bit systems cannot have large physical memory, where this scaling
8097 * makes sense, it is disabled on such platforms.
8099 #if __BITS_PER_LONG > 32
8100 #define ADAPT_SCALE_BASE (64ul << 30)
8101 #define ADAPT_SCALE_SHIFT 2
8102 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8106 * allocate a large system hash table from bootmem
8107 * - it is assumed that the hash table must contain an exact power-of-2
8108 * quantity of entries
8109 * - limit is the number of hash buckets, not the total allocation size
8111 void *__init alloc_large_system_hash(const char *tablename,
8112 unsigned long bucketsize,
8113 unsigned long numentries,
8116 unsigned int *_hash_shift,
8117 unsigned int *_hash_mask,
8118 unsigned long low_limit,
8119 unsigned long high_limit)
8121 unsigned long long max = high_limit;
8122 unsigned long log2qty, size;
8127 /* allow the kernel cmdline to have a say */
8129 /* round applicable memory size up to nearest megabyte */
8130 numentries = nr_kernel_pages;
8131 numentries -= arch_reserved_kernel_pages();
8133 /* It isn't necessary when PAGE_SIZE >= 1MB */
8134 if (PAGE_SHIFT < 20)
8135 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8137 #if __BITS_PER_LONG > 32
8139 unsigned long adapt;
8141 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8142 adapt <<= ADAPT_SCALE_SHIFT)
8147 /* limit to 1 bucket per 2^scale bytes of low memory */
8148 if (scale > PAGE_SHIFT)
8149 numentries >>= (scale - PAGE_SHIFT);
8151 numentries <<= (PAGE_SHIFT - scale);
8153 /* Make sure we've got at least a 0-order allocation.. */
8154 if (unlikely(flags & HASH_SMALL)) {
8155 /* Makes no sense without HASH_EARLY */
8156 WARN_ON(!(flags & HASH_EARLY));
8157 if (!(numentries >> *_hash_shift)) {
8158 numentries = 1UL << *_hash_shift;
8159 BUG_ON(!numentries);
8161 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8162 numentries = PAGE_SIZE / bucketsize;
8164 numentries = roundup_pow_of_two(numentries);
8166 /* limit allocation size to 1/16 total memory by default */
8168 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8169 do_div(max, bucketsize);
8171 max = min(max, 0x80000000ULL);
8173 if (numentries < low_limit)
8174 numentries = low_limit;
8175 if (numentries > max)
8178 log2qty = ilog2(numentries);
8180 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8183 size = bucketsize << log2qty;
8184 if (flags & HASH_EARLY) {
8185 if (flags & HASH_ZERO)
8186 table = memblock_alloc(size, SMP_CACHE_BYTES);
8188 table = memblock_alloc_raw(size,
8190 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8191 table = __vmalloc(size, gfp_flags);
8195 * If bucketsize is not a power-of-two, we may free
8196 * some pages at the end of hash table which
8197 * alloc_pages_exact() automatically does
8199 table = alloc_pages_exact(size, gfp_flags);
8200 kmemleak_alloc(table, size, 1, gfp_flags);
8202 } while (!table && size > PAGE_SIZE && --log2qty);
8205 panic("Failed to allocate %s hash table\n", tablename);
8207 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8208 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8209 virt ? "vmalloc" : "linear");
8212 *_hash_shift = log2qty;
8214 *_hash_mask = (1 << log2qty) - 1;
8220 * This function checks whether pageblock includes unmovable pages or not.
8222 * PageLRU check without isolation or lru_lock could race so that
8223 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8224 * check without lock_page also may miss some movable non-lru pages at
8225 * race condition. So you can't expect this function should be exact.
8227 * Returns a page without holding a reference. If the caller wants to
8228 * dereference that page (e.g., dumping), it has to make sure that it
8229 * cannot get removed (e.g., via memory unplug) concurrently.
8232 struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8233 int migratetype, int flags)
8235 unsigned long iter = 0;
8236 unsigned long pfn = page_to_pfn(page);
8239 * TODO we could make this much more efficient by not checking every
8240 * page in the range if we know all of them are in MOVABLE_ZONE and
8241 * that the movable zone guarantees that pages are migratable but
8242 * the later is not the case right now unfortunatelly. E.g. movablecore
8243 * can still lead to having bootmem allocations in zone_movable.
8246 if (is_migrate_cma_page(page)) {
8248 * CMA allocations (alloc_contig_range) really need to mark
8249 * isolate CMA pageblocks even when they are not movable in fact
8250 * so consider them movable here.
8252 if (is_migrate_cma(migratetype))
8258 for (; iter < pageblock_nr_pages; iter++) {
8259 if (!pfn_valid_within(pfn + iter))
8262 page = pfn_to_page(pfn + iter);
8264 if (PageReserved(page))
8268 * If the zone is movable and we have ruled out all reserved
8269 * pages then it should be reasonably safe to assume the rest
8272 if (zone_idx(zone) == ZONE_MOVABLE)
8276 * Hugepages are not in LRU lists, but they're movable.
8277 * THPs are on the LRU, but need to be counted as #small pages.
8278 * We need not scan over tail pages because we don't
8279 * handle each tail page individually in migration.
8281 if (PageHuge(page) || PageTransCompound(page)) {
8282 struct page *head = compound_head(page);
8283 unsigned int skip_pages;
8285 if (PageHuge(page)) {
8286 if (!hugepage_migration_supported(page_hstate(head)))
8288 } else if (!PageLRU(head) && !__PageMovable(head)) {
8292 skip_pages = compound_nr(head) - (page - head);
8293 iter += skip_pages - 1;
8298 * We can't use page_count without pin a page
8299 * because another CPU can free compound page.
8300 * This check already skips compound tails of THP
8301 * because their page->_refcount is zero at all time.
8303 if (!page_ref_count(page)) {
8304 if (PageBuddy(page))
8305 iter += (1 << page_order(page)) - 1;
8310 * The HWPoisoned page may be not in buddy system, and
8311 * page_count() is not 0.
8313 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8317 * We treat all PageOffline() pages as movable when offlining
8318 * to give drivers a chance to decrement their reference count
8319 * in MEM_GOING_OFFLINE in order to indicate that these pages
8320 * can be offlined as there are no direct references anymore.
8321 * For actually unmovable PageOffline() where the driver does
8322 * not support this, we will fail later when trying to actually
8323 * move these pages that still have a reference count > 0.
8324 * (false negatives in this function only)
8326 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8329 if (__PageMovable(page) || PageLRU(page))
8333 * If there are RECLAIMABLE pages, we need to check
8334 * it. But now, memory offline itself doesn't call
8335 * shrink_node_slabs() and it still to be fixed.
8338 * If the page is not RAM, page_count()should be 0.
8339 * we don't need more check. This is an _used_ not-movable page.
8341 * The problematic thing here is PG_reserved pages. PG_reserved
8342 * is set to both of a memory hole page and a _used_ kernel
8350 #ifdef CONFIG_CONTIG_ALLOC
8351 static unsigned long pfn_max_align_down(unsigned long pfn)
8353 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8354 pageblock_nr_pages) - 1);
8357 static unsigned long pfn_max_align_up(unsigned long pfn)
8359 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8360 pageblock_nr_pages));
8363 /* [start, end) must belong to a single zone. */
8364 static int __alloc_contig_migrate_range(struct compact_control *cc,
8365 unsigned long start, unsigned long end)
8367 /* This function is based on compact_zone() from compaction.c. */
8368 unsigned int nr_reclaimed;
8369 unsigned long pfn = start;
8370 unsigned int tries = 0;
8372 struct migration_target_control mtc = {
8373 .nid = zone_to_nid(cc->zone),
8374 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8379 while (pfn < end || !list_empty(&cc->migratepages)) {
8380 if (fatal_signal_pending(current)) {
8385 if (list_empty(&cc->migratepages)) {
8386 cc->nr_migratepages = 0;
8387 pfn = isolate_migratepages_range(cc, pfn, end);
8393 } else if (++tries == 5) {
8394 ret = ret < 0 ? ret : -EBUSY;
8398 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8400 cc->nr_migratepages -= nr_reclaimed;
8402 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8403 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8406 putback_movable_pages(&cc->migratepages);
8413 * alloc_contig_range() -- tries to allocate given range of pages
8414 * @start: start PFN to allocate
8415 * @end: one-past-the-last PFN to allocate
8416 * @migratetype: migratetype of the underlaying pageblocks (either
8417 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8418 * in range must have the same migratetype and it must
8419 * be either of the two.
8420 * @gfp_mask: GFP mask to use during compaction
8422 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8423 * aligned. The PFN range must belong to a single zone.
8425 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8426 * pageblocks in the range. Once isolated, the pageblocks should not
8427 * be modified by others.
8429 * Return: zero on success or negative error code. On success all
8430 * pages which PFN is in [start, end) are allocated for the caller and
8431 * need to be freed with free_contig_range().
8433 int alloc_contig_range(unsigned long start, unsigned long end,
8434 unsigned migratetype, gfp_t gfp_mask)
8436 unsigned long outer_start, outer_end;
8440 struct compact_control cc = {
8441 .nr_migratepages = 0,
8443 .zone = page_zone(pfn_to_page(start)),
8444 .mode = MIGRATE_SYNC,
8445 .ignore_skip_hint = true,
8446 .no_set_skip_hint = true,
8447 .gfp_mask = current_gfp_context(gfp_mask),
8448 .alloc_contig = true,
8450 INIT_LIST_HEAD(&cc.migratepages);
8453 * What we do here is we mark all pageblocks in range as
8454 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8455 * have different sizes, and due to the way page allocator
8456 * work, we align the range to biggest of the two pages so
8457 * that page allocator won't try to merge buddies from
8458 * different pageblocks and change MIGRATE_ISOLATE to some
8459 * other migration type.
8461 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8462 * migrate the pages from an unaligned range (ie. pages that
8463 * we are interested in). This will put all the pages in
8464 * range back to page allocator as MIGRATE_ISOLATE.
8466 * When this is done, we take the pages in range from page
8467 * allocator removing them from the buddy system. This way
8468 * page allocator will never consider using them.
8470 * This lets us mark the pageblocks back as
8471 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8472 * aligned range but not in the unaligned, original range are
8473 * put back to page allocator so that buddy can use them.
8476 ret = start_isolate_page_range(pfn_max_align_down(start),
8477 pfn_max_align_up(end), migratetype, 0);
8482 * In case of -EBUSY, we'd like to know which page causes problem.
8483 * So, just fall through. test_pages_isolated() has a tracepoint
8484 * which will report the busy page.
8486 * It is possible that busy pages could become available before
8487 * the call to test_pages_isolated, and the range will actually be
8488 * allocated. So, if we fall through be sure to clear ret so that
8489 * -EBUSY is not accidentally used or returned to caller.
8491 ret = __alloc_contig_migrate_range(&cc, start, end);
8492 if (ret && ret != -EBUSY)
8497 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8498 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8499 * more, all pages in [start, end) are free in page allocator.
8500 * What we are going to do is to allocate all pages from
8501 * [start, end) (that is remove them from page allocator).
8503 * The only problem is that pages at the beginning and at the
8504 * end of interesting range may be not aligned with pages that
8505 * page allocator holds, ie. they can be part of higher order
8506 * pages. Because of this, we reserve the bigger range and
8507 * once this is done free the pages we are not interested in.
8509 * We don't have to hold zone->lock here because the pages are
8510 * isolated thus they won't get removed from buddy.
8513 lru_add_drain_all();
8516 outer_start = start;
8517 while (!PageBuddy(pfn_to_page(outer_start))) {
8518 if (++order >= MAX_ORDER) {
8519 outer_start = start;
8522 outer_start &= ~0UL << order;
8525 if (outer_start != start) {
8526 order = page_order(pfn_to_page(outer_start));
8529 * outer_start page could be small order buddy page and
8530 * it doesn't include start page. Adjust outer_start
8531 * in this case to report failed page properly
8532 * on tracepoint in test_pages_isolated()
8534 if (outer_start + (1UL << order) <= start)
8535 outer_start = start;
8538 /* Make sure the range is really isolated. */
8539 if (test_pages_isolated(outer_start, end, 0)) {
8540 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8541 __func__, outer_start, end);
8546 /* Grab isolated pages from freelists. */
8547 outer_end = isolate_freepages_range(&cc, outer_start, end);
8553 /* Free head and tail (if any) */
8554 if (start != outer_start)
8555 free_contig_range(outer_start, start - outer_start);
8556 if (end != outer_end)
8557 free_contig_range(end, outer_end - end);
8560 undo_isolate_page_range(pfn_max_align_down(start),
8561 pfn_max_align_up(end), migratetype);
8564 EXPORT_SYMBOL(alloc_contig_range);
8566 static int __alloc_contig_pages(unsigned long start_pfn,
8567 unsigned long nr_pages, gfp_t gfp_mask)
8569 unsigned long end_pfn = start_pfn + nr_pages;
8571 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8575 static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8576 unsigned long nr_pages)
8578 unsigned long i, end_pfn = start_pfn + nr_pages;
8581 for (i = start_pfn; i < end_pfn; i++) {
8582 page = pfn_to_online_page(i);
8586 if (page_zone(page) != z)
8589 if (PageReserved(page))
8592 if (page_count(page) > 0)
8601 static bool zone_spans_last_pfn(const struct zone *zone,
8602 unsigned long start_pfn, unsigned long nr_pages)
8604 unsigned long last_pfn = start_pfn + nr_pages - 1;
8606 return zone_spans_pfn(zone, last_pfn);
8610 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8611 * @nr_pages: Number of contiguous pages to allocate
8612 * @gfp_mask: GFP mask to limit search and used during compaction
8614 * @nodemask: Mask for other possible nodes
8616 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8617 * on an applicable zonelist to find a contiguous pfn range which can then be
8618 * tried for allocation with alloc_contig_range(). This routine is intended
8619 * for allocation requests which can not be fulfilled with the buddy allocator.
8621 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8622 * power of two then the alignment is guaranteed to be to the given nr_pages
8623 * (e.g. 1GB request would be aligned to 1GB).
8625 * Allocated pages can be freed with free_contig_range() or by manually calling
8626 * __free_page() on each allocated page.
8628 * Return: pointer to contiguous pages on success, or NULL if not successful.
8630 struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8631 int nid, nodemask_t *nodemask)
8633 unsigned long ret, pfn, flags;
8634 struct zonelist *zonelist;
8638 zonelist = node_zonelist(nid, gfp_mask);
8639 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8640 gfp_zone(gfp_mask), nodemask) {
8641 spin_lock_irqsave(&zone->lock, flags);
8643 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8644 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8645 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8647 * We release the zone lock here because
8648 * alloc_contig_range() will also lock the zone
8649 * at some point. If there's an allocation
8650 * spinning on this lock, it may win the race
8651 * and cause alloc_contig_range() to fail...
8653 spin_unlock_irqrestore(&zone->lock, flags);
8654 ret = __alloc_contig_pages(pfn, nr_pages,
8657 return pfn_to_page(pfn);
8658 spin_lock_irqsave(&zone->lock, flags);
8662 spin_unlock_irqrestore(&zone->lock, flags);
8666 #endif /* CONFIG_CONTIG_ALLOC */
8668 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8670 unsigned int count = 0;
8672 for (; nr_pages--; pfn++) {
8673 struct page *page = pfn_to_page(pfn);
8675 count += page_count(page) != 1;
8678 WARN(count != 0, "%d pages are still in use!\n", count);
8680 EXPORT_SYMBOL(free_contig_range);
8683 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8684 * page high values need to be recalulated.
8686 void __meminit zone_pcp_update(struct zone *zone)
8688 mutex_lock(&pcp_batch_high_lock);
8689 __zone_pcp_update(zone);
8690 mutex_unlock(&pcp_batch_high_lock);
8693 void zone_pcp_reset(struct zone *zone)
8695 unsigned long flags;
8697 struct per_cpu_pageset *pset;
8699 /* avoid races with drain_pages() */
8700 local_irq_save(flags);
8701 if (zone->pageset != &boot_pageset) {
8702 for_each_online_cpu(cpu) {
8703 pset = per_cpu_ptr(zone->pageset, cpu);
8704 drain_zonestat(zone, pset);
8706 free_percpu(zone->pageset);
8707 zone->pageset = &boot_pageset;
8709 local_irq_restore(flags);
8712 #ifdef CONFIG_MEMORY_HOTREMOVE
8714 * All pages in the range must be in a single zone and isolated
8715 * before calling this.
8718 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8724 unsigned long flags;
8725 unsigned long offlined_pages = 0;
8727 /* find the first valid pfn */
8728 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8732 return offlined_pages;
8734 offline_mem_sections(pfn, end_pfn);
8735 zone = page_zone(pfn_to_page(pfn));
8736 spin_lock_irqsave(&zone->lock, flags);
8738 while (pfn < end_pfn) {
8739 if (!pfn_valid(pfn)) {
8743 page = pfn_to_page(pfn);
8745 * The HWPoisoned page may be not in buddy system, and
8746 * page_count() is not 0.
8748 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8754 * At this point all remaining PageOffline() pages have a
8755 * reference count of 0 and can simply be skipped.
8757 if (PageOffline(page)) {
8758 BUG_ON(page_count(page));
8759 BUG_ON(PageBuddy(page));
8765 BUG_ON(page_count(page));
8766 BUG_ON(!PageBuddy(page));
8767 order = page_order(page);
8768 offlined_pages += 1 << order;
8769 del_page_from_free_list(page, zone, order);
8770 pfn += (1 << order);
8772 spin_unlock_irqrestore(&zone->lock, flags);
8774 return offlined_pages;
8778 bool is_free_buddy_page(struct page *page)
8780 struct zone *zone = page_zone(page);
8781 unsigned long pfn = page_to_pfn(page);
8782 unsigned long flags;
8785 spin_lock_irqsave(&zone->lock, flags);
8786 for (order = 0; order < MAX_ORDER; order++) {
8787 struct page *page_head = page - (pfn & ((1 << order) - 1));
8789 if (PageBuddy(page_head) && page_order(page_head) >= order)
8792 spin_unlock_irqrestore(&zone->lock, flags);
8794 return order < MAX_ORDER;
8797 #ifdef CONFIG_MEMORY_FAILURE
8799 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8800 * test is performed under the zone lock to prevent a race against page
8803 bool set_hwpoison_free_buddy_page(struct page *page)
8805 struct zone *zone = page_zone(page);
8806 unsigned long pfn = page_to_pfn(page);
8807 unsigned long flags;
8809 bool hwpoisoned = false;
8811 spin_lock_irqsave(&zone->lock, flags);
8812 for (order = 0; order < MAX_ORDER; order++) {
8813 struct page *page_head = page - (pfn & ((1 << order) - 1));
8815 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8816 if (!TestSetPageHWPoison(page))
8821 spin_unlock_irqrestore(&zone->lock, flags);