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/khugepaged.h>
73 #include <asm/sections.h>
74 #include <asm/tlbflush.h>
75 #include <asm/div64.h>
79 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
80 static DEFINE_MUTEX(pcp_batch_high_lock);
81 #define MIN_PERCPU_PAGELIST_FRACTION (8)
83 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
84 DEFINE_PER_CPU(int, numa_node);
85 EXPORT_PER_CPU_SYMBOL(numa_node);
88 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
90 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
92 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
93 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
94 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
95 * defined in <linux/topology.h>.
97 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
98 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
99 int _node_numa_mem_[MAX_NUMNODES];
102 /* work_structs for global per-cpu drains */
105 struct work_struct work;
107 DEFINE_MUTEX(pcpu_drain_mutex);
108 DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
110 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
111 volatile unsigned long latent_entropy __latent_entropy;
112 EXPORT_SYMBOL(latent_entropy);
116 * Array of node states.
118 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
119 [N_POSSIBLE] = NODE_MASK_ALL,
120 [N_ONLINE] = { { [0] = 1UL } },
122 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
123 #ifdef CONFIG_HIGHMEM
124 [N_HIGH_MEMORY] = { { [0] = 1UL } },
126 [N_MEMORY] = { { [0] = 1UL } },
127 [N_CPU] = { { [0] = 1UL } },
130 EXPORT_SYMBOL(node_states);
132 atomic_long_t _totalram_pages __read_mostly;
133 EXPORT_SYMBOL(_totalram_pages);
134 unsigned long totalreserve_pages __read_mostly;
135 unsigned long totalcma_pages __read_mostly;
137 int percpu_pagelist_fraction;
138 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
139 #ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
140 DEFINE_STATIC_KEY_TRUE(init_on_alloc);
142 DEFINE_STATIC_KEY_FALSE(init_on_alloc);
144 EXPORT_SYMBOL(init_on_alloc);
146 #ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
147 DEFINE_STATIC_KEY_TRUE(init_on_free);
149 DEFINE_STATIC_KEY_FALSE(init_on_free);
151 EXPORT_SYMBOL(init_on_free);
153 static int __init early_init_on_alloc(char *buf)
160 ret = kstrtobool(buf, &bool_result);
161 if (bool_result && page_poisoning_enabled())
162 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
164 static_branch_enable(&init_on_alloc);
166 static_branch_disable(&init_on_alloc);
169 early_param("init_on_alloc", early_init_on_alloc);
171 static int __init early_init_on_free(char *buf)
178 ret = kstrtobool(buf, &bool_result);
179 if (bool_result && page_poisoning_enabled())
180 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
182 static_branch_enable(&init_on_free);
184 static_branch_disable(&init_on_free);
187 early_param("init_on_free", early_init_on_free);
190 * A cached value of the page's pageblock's migratetype, used when the page is
191 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
192 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
193 * Also the migratetype set in the page does not necessarily match the pcplist
194 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
195 * other index - this ensures that it will be put on the correct CMA freelist.
197 static inline int get_pcppage_migratetype(struct page *page)
202 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
204 page->index = migratetype;
207 #ifdef CONFIG_PM_SLEEP
209 * The following functions are used by the suspend/hibernate code to temporarily
210 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
211 * while devices are suspended. To avoid races with the suspend/hibernate code,
212 * they should always be called with system_transition_mutex held
213 * (gfp_allowed_mask also should only be modified with system_transition_mutex
214 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
215 * with that modification).
218 static gfp_t saved_gfp_mask;
220 void pm_restore_gfp_mask(void)
222 WARN_ON(!mutex_is_locked(&system_transition_mutex));
223 if (saved_gfp_mask) {
224 gfp_allowed_mask = saved_gfp_mask;
229 void pm_restrict_gfp_mask(void)
231 WARN_ON(!mutex_is_locked(&system_transition_mutex));
232 WARN_ON(saved_gfp_mask);
233 saved_gfp_mask = gfp_allowed_mask;
234 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
237 bool pm_suspended_storage(void)
239 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
243 #endif /* CONFIG_PM_SLEEP */
245 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
246 unsigned int pageblock_order __read_mostly;
249 static void __free_pages_ok(struct page *page, unsigned int order);
252 * results with 256, 32 in the lowmem_reserve sysctl:
253 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
254 * 1G machine -> (16M dma, 784M normal, 224M high)
255 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
256 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
257 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
259 * TBD: should special case ZONE_DMA32 machines here - in those we normally
260 * don't need any ZONE_NORMAL reservation
262 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
263 #ifdef CONFIG_ZONE_DMA
266 #ifdef CONFIG_ZONE_DMA32
270 #ifdef CONFIG_HIGHMEM
276 static char * const zone_names[MAX_NR_ZONES] = {
277 #ifdef CONFIG_ZONE_DMA
280 #ifdef CONFIG_ZONE_DMA32
284 #ifdef CONFIG_HIGHMEM
288 #ifdef CONFIG_ZONE_DEVICE
293 const char * const migratetype_names[MIGRATE_TYPES] = {
301 #ifdef CONFIG_MEMORY_ISOLATION
306 compound_page_dtor * const compound_page_dtors[] = {
309 #ifdef CONFIG_HUGETLB_PAGE
312 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
317 int min_free_kbytes = 1024;
318 int user_min_free_kbytes = -1;
319 #ifdef CONFIG_DISCONTIGMEM
321 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
322 * are not on separate NUMA nodes. Functionally this works but with
323 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
324 * quite small. By default, do not boost watermarks on discontigmem as in
325 * many cases very high-order allocations like THP are likely to be
326 * unsupported and the premature reclaim offsets the advantage of long-term
327 * fragmentation avoidance.
329 int watermark_boost_factor __read_mostly;
331 int watermark_boost_factor __read_mostly = 15000;
333 int watermark_scale_factor = 10;
335 static unsigned long nr_kernel_pages __initdata;
336 static unsigned long nr_all_pages __initdata;
337 static unsigned long dma_reserve __initdata;
339 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
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);
352 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
355 unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
356 unsigned int nr_online_nodes __read_mostly = 1;
357 EXPORT_SYMBOL(nr_node_ids);
358 EXPORT_SYMBOL(nr_online_nodes);
361 int page_group_by_mobility_disabled __read_mostly;
363 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
365 * During boot we initialize deferred pages on-demand, as needed, but once
366 * page_alloc_init_late() has finished, the deferred pages are all initialized,
367 * and we can permanently disable that path.
369 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
372 * Calling kasan_free_pages() only after deferred memory initialization
373 * has completed. Poisoning pages during deferred memory init will greatly
374 * lengthen the process and cause problem in large memory systems as the
375 * deferred pages initialization is done with interrupt disabled.
377 * Assuming that there will be no reference to those newly initialized
378 * pages before they are ever allocated, this should have no effect on
379 * KASAN memory tracking as the poison will be properly inserted at page
380 * allocation time. The only corner case is when pages are allocated by
381 * on-demand allocation and then freed again before the deferred pages
382 * initialization is done, but this is not likely to happen.
384 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
386 if (!static_branch_unlikely(&deferred_pages))
387 kasan_free_pages(page, order);
390 /* Returns true if the struct page for the pfn is uninitialised */
391 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
393 int nid = early_pfn_to_nid(pfn);
395 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
402 * Returns true when the remaining initialisation should be deferred until
403 * later in the boot cycle when it can be parallelised.
405 static bool __meminit
406 defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
408 static unsigned long prev_end_pfn, nr_initialised;
411 * prev_end_pfn static that contains the end of previous zone
412 * No need to protect because called very early in boot before smp_init.
414 if (prev_end_pfn != end_pfn) {
415 prev_end_pfn = end_pfn;
419 /* Always populate low zones for address-constrained allocations */
420 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
424 * We start only with one section of pages, more pages are added as
425 * needed until the rest of deferred pages are initialized.
428 if ((nr_initialised > PAGES_PER_SECTION) &&
429 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
430 NODE_DATA(nid)->first_deferred_pfn = pfn;
436 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
438 static inline bool early_page_uninitialised(unsigned long pfn)
443 static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
449 /* Return a pointer to the bitmap storing bits affecting a block of pages */
450 static inline unsigned long *get_pageblock_bitmap(struct page *page,
453 #ifdef CONFIG_SPARSEMEM
454 return section_to_usemap(__pfn_to_section(pfn));
456 return page_zone(page)->pageblock_flags;
457 #endif /* CONFIG_SPARSEMEM */
460 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
462 #ifdef CONFIG_SPARSEMEM
463 pfn &= (PAGES_PER_SECTION-1);
464 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
466 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
467 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
468 #endif /* CONFIG_SPARSEMEM */
472 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
473 * @page: The page within the block of interest
474 * @pfn: The target page frame number
475 * @end_bitidx: The last bit of interest to retrieve
476 * @mask: mask of bits that the caller is interested in
478 * Return: pageblock_bits flags
480 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
482 unsigned long end_bitidx,
485 unsigned long *bitmap;
486 unsigned long bitidx, word_bitidx;
489 bitmap = get_pageblock_bitmap(page, pfn);
490 bitidx = pfn_to_bitidx(page, pfn);
491 word_bitidx = bitidx / BITS_PER_LONG;
492 bitidx &= (BITS_PER_LONG-1);
494 word = bitmap[word_bitidx];
495 bitidx += end_bitidx;
496 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
499 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
500 unsigned long end_bitidx,
503 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
506 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
508 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
512 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
513 * @page: The page within the block of interest
514 * @flags: The flags to set
515 * @pfn: The target page frame number
516 * @end_bitidx: The last bit of interest
517 * @mask: mask of bits that the caller is interested in
519 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
521 unsigned long end_bitidx,
524 unsigned long *bitmap;
525 unsigned long bitidx, word_bitidx;
526 unsigned long old_word, word;
528 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
529 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
531 bitmap = get_pageblock_bitmap(page, pfn);
532 bitidx = pfn_to_bitidx(page, pfn);
533 word_bitidx = bitidx / BITS_PER_LONG;
534 bitidx &= (BITS_PER_LONG-1);
536 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
538 bitidx += end_bitidx;
539 mask <<= (BITS_PER_LONG - bitidx - 1);
540 flags <<= (BITS_PER_LONG - bitidx - 1);
542 word = READ_ONCE(bitmap[word_bitidx]);
544 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
545 if (word == old_word)
551 void set_pageblock_migratetype(struct page *page, int migratetype)
553 if (unlikely(page_group_by_mobility_disabled &&
554 migratetype < MIGRATE_PCPTYPES))
555 migratetype = MIGRATE_UNMOVABLE;
557 set_pageblock_flags_group(page, (unsigned long)migratetype,
558 PB_migrate, PB_migrate_end);
561 #ifdef CONFIG_DEBUG_VM
562 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
566 unsigned long pfn = page_to_pfn(page);
567 unsigned long sp, start_pfn;
570 seq = zone_span_seqbegin(zone);
571 start_pfn = zone->zone_start_pfn;
572 sp = zone->spanned_pages;
573 if (!zone_spans_pfn(zone, pfn))
575 } while (zone_span_seqretry(zone, seq));
578 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
579 pfn, zone_to_nid(zone), zone->name,
580 start_pfn, start_pfn + sp);
585 static int page_is_consistent(struct zone *zone, struct page *page)
587 if (!pfn_valid_within(page_to_pfn(page)))
589 if (zone != page_zone(page))
595 * Temporary debugging check for pages not lying within a given zone.
597 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
599 if (page_outside_zone_boundaries(zone, page))
601 if (!page_is_consistent(zone, page))
607 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
613 static void bad_page(struct page *page, const char *reason,
614 unsigned long bad_flags)
616 static unsigned long resume;
617 static unsigned long nr_shown;
618 static unsigned long nr_unshown;
621 * Allow a burst of 60 reports, then keep quiet for that minute;
622 * or allow a steady drip of one report per second.
624 if (nr_shown == 60) {
625 if (time_before(jiffies, resume)) {
631 "BUG: Bad page state: %lu messages suppressed\n",
638 resume = jiffies + 60 * HZ;
640 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
641 current->comm, page_to_pfn(page));
642 __dump_page(page, reason);
643 bad_flags &= page->flags;
645 pr_alert("bad because of flags: %#lx(%pGp)\n",
646 bad_flags, &bad_flags);
647 dump_page_owner(page);
652 /* Leave bad fields for debug, except PageBuddy could make trouble */
653 page_mapcount_reset(page); /* remove PageBuddy */
654 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
658 * Higher-order pages are called "compound pages". They are structured thusly:
660 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
662 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
663 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
665 * The first tail page's ->compound_dtor holds the offset in array of compound
666 * page destructors. See compound_page_dtors.
668 * The first tail page's ->compound_order holds the order of allocation.
669 * This usage means that zero-order pages may not be compound.
672 void free_compound_page(struct page *page)
674 mem_cgroup_uncharge(page);
675 __free_pages_ok(page, compound_order(page));
678 void prep_compound_page(struct page *page, unsigned int order)
681 int nr_pages = 1 << order;
683 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
684 set_compound_order(page, order);
686 for (i = 1; i < nr_pages; i++) {
687 struct page *p = page + i;
688 set_page_count(p, 0);
689 p->mapping = TAIL_MAPPING;
690 set_compound_head(p, page);
692 atomic_set(compound_mapcount_ptr(page), -1);
695 #ifdef CONFIG_DEBUG_PAGEALLOC
696 unsigned int _debug_guardpage_minorder;
698 bool _debug_pagealloc_enabled_early __read_mostly
699 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
700 EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
701 DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
702 EXPORT_SYMBOL(_debug_pagealloc_enabled);
704 DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
706 static int __init early_debug_pagealloc(char *buf)
708 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
710 early_param("debug_pagealloc", early_debug_pagealloc);
712 void init_debug_pagealloc(void)
714 if (!debug_pagealloc_enabled())
717 static_branch_enable(&_debug_pagealloc_enabled);
719 if (!debug_guardpage_minorder())
722 static_branch_enable(&_debug_guardpage_enabled);
725 static int __init debug_guardpage_minorder_setup(char *buf)
729 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
730 pr_err("Bad debug_guardpage_minorder value\n");
733 _debug_guardpage_minorder = res;
734 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
737 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
739 static inline bool set_page_guard(struct zone *zone, struct page *page,
740 unsigned int order, int migratetype)
742 if (!debug_guardpage_enabled())
745 if (order >= debug_guardpage_minorder())
748 __SetPageGuard(page);
749 INIT_LIST_HEAD(&page->lru);
750 set_page_private(page, order);
751 /* Guard pages are not available for any usage */
752 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
757 static inline void clear_page_guard(struct zone *zone, struct page *page,
758 unsigned int order, int migratetype)
760 if (!debug_guardpage_enabled())
763 __ClearPageGuard(page);
765 set_page_private(page, 0);
766 if (!is_migrate_isolate(migratetype))
767 __mod_zone_freepage_state(zone, (1 << order), migratetype);
770 static inline bool set_page_guard(struct zone *zone, struct page *page,
771 unsigned int order, int migratetype) { return false; }
772 static inline void clear_page_guard(struct zone *zone, struct page *page,
773 unsigned int order, int migratetype) {}
776 static inline void set_page_order(struct page *page, unsigned int order)
778 set_page_private(page, order);
779 __SetPageBuddy(page);
783 * This function checks whether a page is free && is the buddy
784 * we can coalesce a page and its buddy if
785 * (a) the buddy is not in a hole (check before calling!) &&
786 * (b) the buddy is in the buddy system &&
787 * (c) a page and its buddy have the same order &&
788 * (d) a page and its buddy are in the same zone.
790 * For recording whether a page is in the buddy system, we set PageBuddy.
791 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
793 * For recording page's order, we use page_private(page).
795 static inline int page_is_buddy(struct page *page, struct page *buddy,
798 if (page_is_guard(buddy) && page_order(buddy) == order) {
799 if (page_zone_id(page) != page_zone_id(buddy))
802 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
807 if (PageBuddy(buddy) && page_order(buddy) == order) {
809 * zone check is done late to avoid uselessly
810 * calculating zone/node ids for pages that could
813 if (page_zone_id(page) != page_zone_id(buddy))
816 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
823 #ifdef CONFIG_COMPACTION
824 static inline struct capture_control *task_capc(struct zone *zone)
826 struct capture_control *capc = current->capture_control;
829 !(current->flags & PF_KTHREAD) &&
831 capc->cc->zone == zone &&
832 capc->cc->direct_compaction ? capc : NULL;
836 compaction_capture(struct capture_control *capc, struct page *page,
837 int order, int migratetype)
839 if (!capc || order != capc->cc->order)
842 /* Do not accidentally pollute CMA or isolated regions*/
843 if (is_migrate_cma(migratetype) ||
844 is_migrate_isolate(migratetype))
848 * Do not let lower order allocations polluate a movable pageblock.
849 * This might let an unmovable request use a reclaimable pageblock
850 * and vice-versa but no more than normal fallback logic which can
851 * have trouble finding a high-order free page.
853 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
861 static inline struct capture_control *task_capc(struct zone *zone)
867 compaction_capture(struct capture_control *capc, struct page *page,
868 int order, int migratetype)
872 #endif /* CONFIG_COMPACTION */
875 * Freeing function for a buddy system allocator.
877 * The concept of a buddy system is to maintain direct-mapped table
878 * (containing bit values) for memory blocks of various "orders".
879 * The bottom level table contains the map for the smallest allocatable
880 * units of memory (here, pages), and each level above it describes
881 * pairs of units from the levels below, hence, "buddies".
882 * At a high level, all that happens here is marking the table entry
883 * at the bottom level available, and propagating the changes upward
884 * as necessary, plus some accounting needed to play nicely with other
885 * parts of the VM system.
886 * At each level, we keep a list of pages, which are heads of continuous
887 * free pages of length of (1 << order) and marked with PageBuddy.
888 * Page's order is recorded in page_private(page) field.
889 * So when we are allocating or freeing one, we can derive the state of the
890 * other. That is, if we allocate a small block, and both were
891 * free, the remainder of the region must be split into blocks.
892 * If a block is freed, and its buddy is also free, then this
893 * triggers coalescing into a block of larger size.
898 static inline void __free_one_page(struct page *page,
900 struct zone *zone, unsigned int order,
903 unsigned long combined_pfn;
904 unsigned long uninitialized_var(buddy_pfn);
906 unsigned int max_order;
907 struct capture_control *capc = task_capc(zone);
909 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
911 VM_BUG_ON(!zone_is_initialized(zone));
912 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
914 VM_BUG_ON(migratetype == -1);
915 if (likely(!is_migrate_isolate(migratetype)))
916 __mod_zone_freepage_state(zone, 1 << order, migratetype);
918 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
919 VM_BUG_ON_PAGE(bad_range(zone, page), page);
922 while (order < max_order) {
923 if (compaction_capture(capc, page, order, migratetype)) {
924 __mod_zone_freepage_state(zone, -(1 << order),
928 buddy_pfn = __find_buddy_pfn(pfn, order);
929 buddy = page + (buddy_pfn - pfn);
931 if (!pfn_valid_within(buddy_pfn))
933 if (!page_is_buddy(page, buddy, order))
936 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
937 * merge with it and move up one order.
939 if (page_is_guard(buddy))
940 clear_page_guard(zone, buddy, order, migratetype);
942 del_page_from_free_area(buddy, &zone->free_area[order]);
943 combined_pfn = buddy_pfn & pfn;
944 page = page + (combined_pfn - pfn);
948 if (order < MAX_ORDER - 1) {
949 /* If we are here, it means order is >= pageblock_order.
950 * We want to prevent merge between freepages on isolate
951 * pageblock and normal pageblock. Without this, pageblock
952 * isolation could cause incorrect freepage or CMA accounting.
954 * We don't want to hit this code for the more frequent
957 if (unlikely(has_isolate_pageblock(zone))) {
960 buddy_pfn = __find_buddy_pfn(pfn, order);
961 buddy = page + (buddy_pfn - pfn);
962 buddy_mt = get_pageblock_migratetype(buddy);
964 if (migratetype != buddy_mt
965 && (is_migrate_isolate(migratetype) ||
966 is_migrate_isolate(buddy_mt)))
969 max_order = order + 1;
970 goto continue_merging;
974 set_page_order(page, order);
977 * If this is not the largest possible page, check if the buddy
978 * of the next-highest order is free. If it is, it's possible
979 * that pages are being freed that will coalesce soon. In case,
980 * that is happening, add the free page to the tail of the list
981 * so it's less likely to be used soon and more likely to be merged
982 * as a higher order page
984 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)
985 && !is_shuffle_order(order)) {
986 struct page *higher_page, *higher_buddy;
987 combined_pfn = buddy_pfn & pfn;
988 higher_page = page + (combined_pfn - pfn);
989 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
990 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
991 if (pfn_valid_within(buddy_pfn) &&
992 page_is_buddy(higher_page, higher_buddy, order + 1)) {
993 add_to_free_area_tail(page, &zone->free_area[order],
999 if (is_shuffle_order(order))
1000 add_to_free_area_random(page, &zone->free_area[order],
1003 add_to_free_area(page, &zone->free_area[order], migratetype);
1008 * A bad page could be due to a number of fields. Instead of multiple branches,
1009 * try and check multiple fields with one check. The caller must do a detailed
1010 * check if necessary.
1012 static inline bool page_expected_state(struct page *page,
1013 unsigned long check_flags)
1015 if (unlikely(atomic_read(&page->_mapcount) != -1))
1018 if (unlikely((unsigned long)page->mapping |
1019 page_ref_count(page) |
1021 (unsigned long)page->mem_cgroup |
1023 (page->flags & check_flags)))
1029 static void free_pages_check_bad(struct page *page)
1031 const char *bad_reason;
1032 unsigned long bad_flags;
1037 if (unlikely(atomic_read(&page->_mapcount) != -1))
1038 bad_reason = "nonzero mapcount";
1039 if (unlikely(page->mapping != NULL))
1040 bad_reason = "non-NULL mapping";
1041 if (unlikely(page_ref_count(page) != 0))
1042 bad_reason = "nonzero _refcount";
1043 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1044 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1045 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1048 if (unlikely(page->mem_cgroup))
1049 bad_reason = "page still charged to cgroup";
1051 bad_page(page, bad_reason, bad_flags);
1054 static inline int free_pages_check(struct page *page)
1056 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1059 /* Something has gone sideways, find it */
1060 free_pages_check_bad(page);
1064 static int free_tail_pages_check(struct page *head_page, struct page *page)
1069 * We rely page->lru.next never has bit 0 set, unless the page
1070 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1072 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1074 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1078 switch (page - head_page) {
1080 /* the first tail page: ->mapping may be compound_mapcount() */
1081 if (unlikely(compound_mapcount(page))) {
1082 bad_page(page, "nonzero compound_mapcount", 0);
1088 * the second tail page: ->mapping is
1089 * deferred_list.next -- ignore value.
1093 if (page->mapping != TAIL_MAPPING) {
1094 bad_page(page, "corrupted mapping in tail page", 0);
1099 if (unlikely(!PageTail(page))) {
1100 bad_page(page, "PageTail not set", 0);
1103 if (unlikely(compound_head(page) != head_page)) {
1104 bad_page(page, "compound_head not consistent", 0);
1109 page->mapping = NULL;
1110 clear_compound_head(page);
1114 static void kernel_init_free_pages(struct page *page, int numpages)
1118 for (i = 0; i < numpages; i++)
1119 clear_highpage(page + i);
1122 static __always_inline bool free_pages_prepare(struct page *page,
1123 unsigned int order, bool check_free)
1127 VM_BUG_ON_PAGE(PageTail(page), page);
1129 trace_mm_page_free(page, order);
1132 * Check tail pages before head page information is cleared to
1133 * avoid checking PageCompound for order-0 pages.
1135 if (unlikely(order)) {
1136 bool compound = PageCompound(page);
1139 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1142 ClearPageDoubleMap(page);
1143 for (i = 1; i < (1 << order); i++) {
1145 bad += free_tail_pages_check(page, page + i);
1146 if (unlikely(free_pages_check(page + i))) {
1150 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1153 if (PageMappingFlags(page))
1154 page->mapping = NULL;
1155 if (memcg_kmem_enabled() && PageKmemcg(page))
1156 __memcg_kmem_uncharge(page, order);
1158 bad += free_pages_check(page);
1162 page_cpupid_reset_last(page);
1163 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1164 reset_page_owner(page, order);
1166 if (!PageHighMem(page)) {
1167 debug_check_no_locks_freed(page_address(page),
1168 PAGE_SIZE << order);
1169 debug_check_no_obj_freed(page_address(page),
1170 PAGE_SIZE << order);
1172 if (want_init_on_free())
1173 kernel_init_free_pages(page, 1 << order);
1175 kernel_poison_pages(page, 1 << order, 0);
1177 * arch_free_page() can make the page's contents inaccessible. s390
1178 * does this. So nothing which can access the page's contents should
1179 * happen after this.
1181 arch_free_page(page, order);
1183 if (debug_pagealloc_enabled_static())
1184 kernel_map_pages(page, 1 << order, 0);
1186 kasan_free_nondeferred_pages(page, order);
1191 #ifdef CONFIG_DEBUG_VM
1193 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1194 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1195 * moved from pcp lists to free lists.
1197 static bool free_pcp_prepare(struct page *page)
1199 return free_pages_prepare(page, 0, true);
1202 static bool bulkfree_pcp_prepare(struct page *page)
1204 if (debug_pagealloc_enabled_static())
1205 return free_pages_check(page);
1211 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1212 * moving from pcp lists to free list in order to reduce overhead. With
1213 * debug_pagealloc enabled, they are checked also immediately when being freed
1216 static bool free_pcp_prepare(struct page *page)
1218 if (debug_pagealloc_enabled_static())
1219 return free_pages_prepare(page, 0, true);
1221 return free_pages_prepare(page, 0, false);
1224 static bool bulkfree_pcp_prepare(struct page *page)
1226 return free_pages_check(page);
1228 #endif /* CONFIG_DEBUG_VM */
1230 static inline void prefetch_buddy(struct page *page)
1232 unsigned long pfn = page_to_pfn(page);
1233 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1234 struct page *buddy = page + (buddy_pfn - pfn);
1240 * Frees a number of pages from the PCP lists
1241 * Assumes all pages on list are in same zone, and of same order.
1242 * count is the number of pages to free.
1244 * If the zone was previously in an "all pages pinned" state then look to
1245 * see if this freeing clears that state.
1247 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1248 * pinned" detection logic.
1250 static void free_pcppages_bulk(struct zone *zone, int count,
1251 struct per_cpu_pages *pcp)
1253 int migratetype = 0;
1255 int prefetch_nr = 0;
1256 bool isolated_pageblocks;
1257 struct page *page, *tmp;
1261 * Ensure proper count is passed which otherwise would stuck in the
1262 * below while (list_empty(list)) loop.
1264 count = min(pcp->count, count);
1266 struct list_head *list;
1269 * Remove pages from lists in a round-robin fashion. A
1270 * batch_free count is maintained that is incremented when an
1271 * empty list is encountered. This is so more pages are freed
1272 * off fuller lists instead of spinning excessively around empty
1277 if (++migratetype == MIGRATE_PCPTYPES)
1279 list = &pcp->lists[migratetype];
1280 } while (list_empty(list));
1282 /* This is the only non-empty list. Free them all. */
1283 if (batch_free == MIGRATE_PCPTYPES)
1287 page = list_last_entry(list, struct page, lru);
1288 /* must delete to avoid corrupting pcp list */
1289 list_del(&page->lru);
1292 if (bulkfree_pcp_prepare(page))
1295 list_add_tail(&page->lru, &head);
1298 * We are going to put the page back to the global
1299 * pool, prefetch its buddy to speed up later access
1300 * under zone->lock. It is believed the overhead of
1301 * an additional test and calculating buddy_pfn here
1302 * can be offset by reduced memory latency later. To
1303 * avoid excessive prefetching due to large count, only
1304 * prefetch buddy for the first pcp->batch nr of pages.
1306 if (prefetch_nr++ < pcp->batch)
1307 prefetch_buddy(page);
1308 } while (--count && --batch_free && !list_empty(list));
1311 spin_lock(&zone->lock);
1312 isolated_pageblocks = has_isolate_pageblock(zone);
1315 * Use safe version since after __free_one_page(),
1316 * page->lru.next will not point to original list.
1318 list_for_each_entry_safe(page, tmp, &head, lru) {
1319 int mt = get_pcppage_migratetype(page);
1320 /* MIGRATE_ISOLATE page should not go to pcplists */
1321 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1322 /* Pageblock could have been isolated meanwhile */
1323 if (unlikely(isolated_pageblocks))
1324 mt = get_pageblock_migratetype(page);
1326 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1327 trace_mm_page_pcpu_drain(page, 0, mt);
1329 spin_unlock(&zone->lock);
1332 static void free_one_page(struct zone *zone,
1333 struct page *page, unsigned long pfn,
1337 spin_lock(&zone->lock);
1338 if (unlikely(has_isolate_pageblock(zone) ||
1339 is_migrate_isolate(migratetype))) {
1340 migratetype = get_pfnblock_migratetype(page, pfn);
1342 __free_one_page(page, pfn, zone, order, migratetype);
1343 spin_unlock(&zone->lock);
1346 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1347 unsigned long zone, int nid)
1349 mm_zero_struct_page(page);
1350 set_page_links(page, zone, nid, pfn);
1351 init_page_count(page);
1352 page_mapcount_reset(page);
1353 page_cpupid_reset_last(page);
1354 page_kasan_tag_reset(page);
1356 INIT_LIST_HEAD(&page->lru);
1357 #ifdef WANT_PAGE_VIRTUAL
1358 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1359 if (!is_highmem_idx(zone))
1360 set_page_address(page, __va(pfn << PAGE_SHIFT));
1364 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1365 static void __meminit init_reserved_page(unsigned long pfn)
1370 if (!early_page_uninitialised(pfn))
1373 nid = early_pfn_to_nid(pfn);
1374 pgdat = NODE_DATA(nid);
1376 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1377 struct zone *zone = &pgdat->node_zones[zid];
1379 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1382 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1385 static inline void init_reserved_page(unsigned long pfn)
1388 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1391 * Initialised pages do not have PageReserved set. This function is
1392 * called for each range allocated by the bootmem allocator and
1393 * marks the pages PageReserved. The remaining valid pages are later
1394 * sent to the buddy page allocator.
1396 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1398 unsigned long start_pfn = PFN_DOWN(start);
1399 unsigned long end_pfn = PFN_UP(end);
1401 for (; start_pfn < end_pfn; start_pfn++) {
1402 if (pfn_valid(start_pfn)) {
1403 struct page *page = pfn_to_page(start_pfn);
1405 init_reserved_page(start_pfn);
1407 /* Avoid false-positive PageTail() */
1408 INIT_LIST_HEAD(&page->lru);
1411 * no need for atomic set_bit because the struct
1412 * page is not visible yet so nobody should
1415 __SetPageReserved(page);
1420 static void __free_pages_ok(struct page *page, unsigned int order)
1422 unsigned long flags;
1424 unsigned long pfn = page_to_pfn(page);
1426 if (!free_pages_prepare(page, order, true))
1429 migratetype = get_pfnblock_migratetype(page, pfn);
1430 local_irq_save(flags);
1431 __count_vm_events(PGFREE, 1 << order);
1432 free_one_page(page_zone(page), page, pfn, order, migratetype);
1433 local_irq_restore(flags);
1436 void __free_pages_core(struct page *page, unsigned int order)
1438 unsigned int nr_pages = 1 << order;
1439 struct page *p = page;
1443 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1445 __ClearPageReserved(p);
1446 set_page_count(p, 0);
1448 __ClearPageReserved(p);
1449 set_page_count(p, 0);
1451 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1452 set_page_refcounted(page);
1453 __free_pages(page, order);
1456 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1457 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1459 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1461 int __meminit early_pfn_to_nid(unsigned long pfn)
1463 static DEFINE_SPINLOCK(early_pfn_lock);
1466 spin_lock(&early_pfn_lock);
1467 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1469 nid = first_online_node;
1470 spin_unlock(&early_pfn_lock);
1476 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1477 /* Only safe to use early in boot when initialisation is single-threaded */
1478 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1482 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1483 if (nid >= 0 && nid != node)
1489 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1496 void __init memblock_free_pages(struct page *page, unsigned long pfn,
1499 if (early_page_uninitialised(pfn))
1501 __free_pages_core(page, order);
1505 * Check that the whole (or subset of) a pageblock given by the interval of
1506 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1507 * with the migration of free compaction scanner. The scanners then need to
1508 * use only pfn_valid_within() check for arches that allow holes within
1511 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1513 * It's possible on some configurations to have a setup like node0 node1 node0
1514 * i.e. it's possible that all pages within a zones range of pages do not
1515 * belong to a single zone. We assume that a border between node0 and node1
1516 * can occur within a single pageblock, but not a node0 node1 node0
1517 * interleaving within a single pageblock. It is therefore sufficient to check
1518 * the first and last page of a pageblock and avoid checking each individual
1519 * page in a pageblock.
1521 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1522 unsigned long end_pfn, struct zone *zone)
1524 struct page *start_page;
1525 struct page *end_page;
1527 /* end_pfn is one past the range we are checking */
1530 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1533 start_page = pfn_to_online_page(start_pfn);
1537 if (page_zone(start_page) != zone)
1540 end_page = pfn_to_page(end_pfn);
1542 /* This gives a shorter code than deriving page_zone(end_page) */
1543 if (page_zone_id(start_page) != page_zone_id(end_page))
1549 void set_zone_contiguous(struct zone *zone)
1551 unsigned long block_start_pfn = zone->zone_start_pfn;
1552 unsigned long block_end_pfn;
1554 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1555 for (; block_start_pfn < zone_end_pfn(zone);
1556 block_start_pfn = block_end_pfn,
1557 block_end_pfn += pageblock_nr_pages) {
1559 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1561 if (!__pageblock_pfn_to_page(block_start_pfn,
1562 block_end_pfn, zone))
1567 /* We confirm that there is no hole */
1568 zone->contiguous = true;
1571 void clear_zone_contiguous(struct zone *zone)
1573 zone->contiguous = false;
1576 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1577 static void __init deferred_free_range(unsigned long pfn,
1578 unsigned long nr_pages)
1586 page = pfn_to_page(pfn);
1588 /* Free a large naturally-aligned chunk if possible */
1589 if (nr_pages == pageblock_nr_pages &&
1590 (pfn & (pageblock_nr_pages - 1)) == 0) {
1591 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1592 __free_pages_core(page, pageblock_order);
1596 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1597 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1598 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1599 __free_pages_core(page, 0);
1603 /* Completion tracking for deferred_init_memmap() threads */
1604 static atomic_t pgdat_init_n_undone __initdata;
1605 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1607 static inline void __init pgdat_init_report_one_done(void)
1609 if (atomic_dec_and_test(&pgdat_init_n_undone))
1610 complete(&pgdat_init_all_done_comp);
1614 * Returns true if page needs to be initialized or freed to buddy allocator.
1616 * First we check if pfn is valid on architectures where it is possible to have
1617 * holes within pageblock_nr_pages. On systems where it is not possible, this
1618 * function is optimized out.
1620 * Then, we check if a current large page is valid by only checking the validity
1623 static inline bool __init deferred_pfn_valid(unsigned long pfn)
1625 if (!pfn_valid_within(pfn))
1627 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1633 * Free pages to buddy allocator. Try to free aligned pages in
1634 * pageblock_nr_pages sizes.
1636 static void __init deferred_free_pages(unsigned long pfn,
1637 unsigned long end_pfn)
1639 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1640 unsigned long nr_free = 0;
1642 for (; pfn < end_pfn; pfn++) {
1643 if (!deferred_pfn_valid(pfn)) {
1644 deferred_free_range(pfn - nr_free, nr_free);
1646 } else if (!(pfn & nr_pgmask)) {
1647 deferred_free_range(pfn - nr_free, nr_free);
1653 /* Free the last block of pages to allocator */
1654 deferred_free_range(pfn - nr_free, nr_free);
1658 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1659 * by performing it only once every pageblock_nr_pages.
1660 * Return number of pages initialized.
1662 static unsigned long __init deferred_init_pages(struct zone *zone,
1664 unsigned long end_pfn)
1666 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1667 int nid = zone_to_nid(zone);
1668 unsigned long nr_pages = 0;
1669 int zid = zone_idx(zone);
1670 struct page *page = NULL;
1672 for (; pfn < end_pfn; pfn++) {
1673 if (!deferred_pfn_valid(pfn)) {
1676 } else if (!page || !(pfn & nr_pgmask)) {
1677 page = pfn_to_page(pfn);
1681 __init_single_page(page, pfn, zid, nid);
1688 * This function is meant to pre-load the iterator for the zone init.
1689 * Specifically it walks through the ranges until we are caught up to the
1690 * first_init_pfn value and exits there. If we never encounter the value we
1691 * return false indicating there are no valid ranges left.
1694 deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1695 unsigned long *spfn, unsigned long *epfn,
1696 unsigned long first_init_pfn)
1701 * Start out by walking through the ranges in this zone that have
1702 * already been initialized. We don't need to do anything with them
1703 * so we just need to flush them out of the system.
1705 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1706 if (*epfn <= first_init_pfn)
1708 if (*spfn < first_init_pfn)
1709 *spfn = first_init_pfn;
1718 * Initialize and free pages. We do it in two loops: first we initialize
1719 * struct page, then free to buddy allocator, because while we are
1720 * freeing pages we can access pages that are ahead (computing buddy
1721 * page in __free_one_page()).
1723 * In order to try and keep some memory in the cache we have the loop
1724 * broken along max page order boundaries. This way we will not cause
1725 * any issues with the buddy page computation.
1727 static unsigned long __init
1728 deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1729 unsigned long *end_pfn)
1731 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1732 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1733 unsigned long nr_pages = 0;
1736 /* First we loop through and initialize the page values */
1737 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1740 if (mo_pfn <= *start_pfn)
1743 t = min(mo_pfn, *end_pfn);
1744 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1746 if (mo_pfn < *end_pfn) {
1747 *start_pfn = mo_pfn;
1752 /* Reset values and now loop through freeing pages as needed */
1755 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1761 t = min(mo_pfn, epfn);
1762 deferred_free_pages(spfn, t);
1771 /* Initialise remaining memory on a node */
1772 static int __init deferred_init_memmap(void *data)
1774 pg_data_t *pgdat = data;
1775 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1776 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1777 unsigned long first_init_pfn, flags;
1778 unsigned long start = jiffies;
1783 /* Bind memory initialisation thread to a local node if possible */
1784 if (!cpumask_empty(cpumask))
1785 set_cpus_allowed_ptr(current, cpumask);
1787 pgdat_resize_lock(pgdat, &flags);
1788 first_init_pfn = pgdat->first_deferred_pfn;
1789 if (first_init_pfn == ULONG_MAX) {
1790 pgdat_resize_unlock(pgdat, &flags);
1791 pgdat_init_report_one_done();
1795 /* Sanity check boundaries */
1796 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1797 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1798 pgdat->first_deferred_pfn = ULONG_MAX;
1801 * Once we unlock here, the zone cannot be grown anymore, thus if an
1802 * interrupt thread must allocate this early in boot, zone must be
1803 * pre-grown prior to start of deferred page initialization.
1805 pgdat_resize_unlock(pgdat, &flags);
1807 /* Only the highest zone is deferred so find it */
1808 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1809 zone = pgdat->node_zones + zid;
1810 if (first_init_pfn < zone_end_pfn(zone))
1814 /* If the zone is empty somebody else may have cleared out the zone */
1815 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1820 * Initialize and free pages in MAX_ORDER sized increments so
1821 * that we can avoid introducing any issues with the buddy
1824 while (spfn < epfn) {
1825 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1829 /* Sanity check that the next zone really is unpopulated */
1830 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1832 pr_info("node %d initialised, %lu pages in %ums\n",
1833 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1835 pgdat_init_report_one_done();
1840 * If this zone has deferred pages, try to grow it by initializing enough
1841 * deferred pages to satisfy the allocation specified by order, rounded up to
1842 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1843 * of SECTION_SIZE bytes by initializing struct pages in increments of
1844 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1846 * Return true when zone was grown, otherwise return false. We return true even
1847 * when we grow less than requested, to let the caller decide if there are
1848 * enough pages to satisfy the allocation.
1850 * Note: We use noinline because this function is needed only during boot, and
1851 * it is called from a __ref function _deferred_grow_zone. This way we are
1852 * making sure that it is not inlined into permanent text section.
1854 static noinline bool __init
1855 deferred_grow_zone(struct zone *zone, unsigned int order)
1857 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1858 pg_data_t *pgdat = zone->zone_pgdat;
1859 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1860 unsigned long spfn, epfn, flags;
1861 unsigned long nr_pages = 0;
1864 /* Only the last zone may have deferred pages */
1865 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1868 pgdat_resize_lock(pgdat, &flags);
1871 * If someone grew this zone while we were waiting for spinlock, return
1872 * true, as there might be enough pages already.
1874 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1875 pgdat_resize_unlock(pgdat, &flags);
1879 /* If the zone is empty somebody else may have cleared out the zone */
1880 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1881 first_deferred_pfn)) {
1882 pgdat->first_deferred_pfn = ULONG_MAX;
1883 pgdat_resize_unlock(pgdat, &flags);
1884 /* Retry only once. */
1885 return first_deferred_pfn != ULONG_MAX;
1889 * Initialize and free pages in MAX_ORDER sized increments so
1890 * that we can avoid introducing any issues with the buddy
1893 while (spfn < epfn) {
1894 /* update our first deferred PFN for this section */
1895 first_deferred_pfn = spfn;
1897 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1898 touch_nmi_watchdog();
1900 /* We should only stop along section boundaries */
1901 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1904 /* If our quota has been met we can stop here */
1905 if (nr_pages >= nr_pages_needed)
1909 pgdat->first_deferred_pfn = spfn;
1910 pgdat_resize_unlock(pgdat, &flags);
1912 return nr_pages > 0;
1916 * deferred_grow_zone() is __init, but it is called from
1917 * get_page_from_freelist() during early boot until deferred_pages permanently
1918 * disables this call. This is why we have refdata wrapper to avoid warning,
1919 * and to ensure that the function body gets unloaded.
1922 _deferred_grow_zone(struct zone *zone, unsigned int order)
1924 return deferred_grow_zone(zone, order);
1927 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1929 void __init page_alloc_init_late(void)
1934 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1936 /* There will be num_node_state(N_MEMORY) threads */
1937 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1938 for_each_node_state(nid, N_MEMORY) {
1939 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1942 /* Block until all are initialised */
1943 wait_for_completion(&pgdat_init_all_done_comp);
1946 * The number of managed pages has changed due to the initialisation
1947 * so the pcpu batch and high limits needs to be updated or the limits
1948 * will be artificially small.
1950 for_each_populated_zone(zone)
1951 zone_pcp_update(zone);
1954 * We initialized the rest of the deferred pages. Permanently disable
1955 * on-demand struct page initialization.
1957 static_branch_disable(&deferred_pages);
1959 /* Reinit limits that are based on free pages after the kernel is up */
1960 files_maxfiles_init();
1963 /* Discard memblock private memory */
1966 for_each_node_state(nid, N_MEMORY)
1967 shuffle_free_memory(NODE_DATA(nid));
1969 for_each_populated_zone(zone)
1970 set_zone_contiguous(zone);
1974 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1975 void __init init_cma_reserved_pageblock(struct page *page)
1977 unsigned i = pageblock_nr_pages;
1978 struct page *p = page;
1981 __ClearPageReserved(p);
1982 set_page_count(p, 0);
1985 set_pageblock_migratetype(page, MIGRATE_CMA);
1987 if (pageblock_order >= MAX_ORDER) {
1988 i = pageblock_nr_pages;
1991 set_page_refcounted(p);
1992 __free_pages(p, MAX_ORDER - 1);
1993 p += MAX_ORDER_NR_PAGES;
1994 } while (i -= MAX_ORDER_NR_PAGES);
1996 set_page_refcounted(page);
1997 __free_pages(page, pageblock_order);
2000 adjust_managed_page_count(page, pageblock_nr_pages);
2005 * The order of subdivision here is critical for the IO subsystem.
2006 * Please do not alter this order without good reasons and regression
2007 * testing. Specifically, as large blocks of memory are subdivided,
2008 * the order in which smaller blocks are delivered depends on the order
2009 * they're subdivided in this function. This is the primary factor
2010 * influencing the order in which pages are delivered to the IO
2011 * subsystem according to empirical testing, and this is also justified
2012 * by considering the behavior of a buddy system containing a single
2013 * large block of memory acted on by a series of small allocations.
2014 * This behavior is a critical factor in sglist merging's success.
2018 static inline void expand(struct zone *zone, struct page *page,
2019 int low, int high, struct free_area *area,
2022 unsigned long size = 1 << high;
2024 while (high > low) {
2028 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2031 * Mark as guard pages (or page), that will allow to
2032 * merge back to allocator when buddy will be freed.
2033 * Corresponding page table entries will not be touched,
2034 * pages will stay not present in virtual address space
2036 if (set_page_guard(zone, &page[size], high, migratetype))
2039 add_to_free_area(&page[size], area, migratetype);
2040 set_page_order(&page[size], high);
2044 static void check_new_page_bad(struct page *page)
2046 const char *bad_reason = NULL;
2047 unsigned long bad_flags = 0;
2049 if (unlikely(atomic_read(&page->_mapcount) != -1))
2050 bad_reason = "nonzero mapcount";
2051 if (unlikely(page->mapping != NULL))
2052 bad_reason = "non-NULL mapping";
2053 if (unlikely(page_ref_count(page) != 0))
2054 bad_reason = "nonzero _refcount";
2055 if (unlikely(page->flags & __PG_HWPOISON)) {
2056 bad_reason = "HWPoisoned (hardware-corrupted)";
2057 bad_flags = __PG_HWPOISON;
2058 /* Don't complain about hwpoisoned pages */
2059 page_mapcount_reset(page); /* remove PageBuddy */
2062 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2063 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2064 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2067 if (unlikely(page->mem_cgroup))
2068 bad_reason = "page still charged to cgroup";
2070 bad_page(page, bad_reason, bad_flags);
2074 * This page is about to be returned from the page allocator
2076 static inline int check_new_page(struct page *page)
2078 if (likely(page_expected_state(page,
2079 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2082 check_new_page_bad(page);
2086 static inline bool free_pages_prezeroed(void)
2088 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2089 page_poisoning_enabled()) || want_init_on_free();
2092 #ifdef CONFIG_DEBUG_VM
2094 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2095 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2096 * also checked when pcp lists are refilled from the free lists.
2098 static inline bool check_pcp_refill(struct page *page)
2100 if (debug_pagealloc_enabled_static())
2101 return check_new_page(page);
2106 static inline bool check_new_pcp(struct page *page)
2108 return check_new_page(page);
2112 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2113 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2114 * enabled, they are also checked when being allocated from the pcp lists.
2116 static inline bool check_pcp_refill(struct page *page)
2118 return check_new_page(page);
2120 static inline bool check_new_pcp(struct page *page)
2122 if (debug_pagealloc_enabled_static())
2123 return check_new_page(page);
2127 #endif /* CONFIG_DEBUG_VM */
2129 static bool check_new_pages(struct page *page, unsigned int order)
2132 for (i = 0; i < (1 << order); i++) {
2133 struct page *p = page + i;
2135 if (unlikely(check_new_page(p)))
2142 inline void post_alloc_hook(struct page *page, unsigned int order,
2145 set_page_private(page, 0);
2146 set_page_refcounted(page);
2148 arch_alloc_page(page, order);
2149 if (debug_pagealloc_enabled_static())
2150 kernel_map_pages(page, 1 << order, 1);
2151 kasan_alloc_pages(page, order);
2152 kernel_poison_pages(page, 1 << order, 1);
2153 set_page_owner(page, order, gfp_flags);
2156 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2157 unsigned int alloc_flags)
2159 post_alloc_hook(page, order, gfp_flags);
2161 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2162 kernel_init_free_pages(page, 1 << order);
2164 if (order && (gfp_flags & __GFP_COMP))
2165 prep_compound_page(page, order);
2168 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2169 * allocate the page. The expectation is that the caller is taking
2170 * steps that will free more memory. The caller should avoid the page
2171 * being used for !PFMEMALLOC purposes.
2173 if (alloc_flags & ALLOC_NO_WATERMARKS)
2174 set_page_pfmemalloc(page);
2176 clear_page_pfmemalloc(page);
2180 * Go through the free lists for the given migratetype and remove
2181 * the smallest available page from the freelists
2183 static __always_inline
2184 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2187 unsigned int current_order;
2188 struct free_area *area;
2191 /* Find a page of the appropriate size in the preferred list */
2192 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2193 area = &(zone->free_area[current_order]);
2194 page = get_page_from_free_area(area, migratetype);
2197 del_page_from_free_area(page, area);
2198 expand(zone, page, order, current_order, area, migratetype);
2199 set_pcppage_migratetype(page, migratetype);
2208 * This array describes the order lists are fallen back to when
2209 * the free lists for the desirable migrate type are depleted
2211 static int fallbacks[MIGRATE_TYPES][4] = {
2212 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2213 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2214 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2216 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2218 #ifdef CONFIG_MEMORY_ISOLATION
2219 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2224 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2227 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2230 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2231 unsigned int order) { return NULL; }
2235 * Move the free pages in a range to the free lists of the requested type.
2236 * Note that start_page and end_pages are not aligned on a pageblock
2237 * boundary. If alignment is required, use move_freepages_block()
2239 static int move_freepages(struct zone *zone,
2240 struct page *start_page, struct page *end_page,
2241 int migratetype, int *num_movable)
2245 int pages_moved = 0;
2247 for (page = start_page; page <= end_page;) {
2248 if (!pfn_valid_within(page_to_pfn(page))) {
2253 if (!PageBuddy(page)) {
2255 * We assume that pages that could be isolated for
2256 * migration are movable. But we don't actually try
2257 * isolating, as that would be expensive.
2260 (PageLRU(page) || __PageMovable(page)))
2267 /* Make sure we are not inadvertently changing nodes */
2268 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2269 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2271 order = page_order(page);
2272 move_to_free_area(page, &zone->free_area[order], migratetype);
2274 pages_moved += 1 << order;
2280 int move_freepages_block(struct zone *zone, struct page *page,
2281 int migratetype, int *num_movable)
2283 unsigned long start_pfn, end_pfn;
2284 struct page *start_page, *end_page;
2289 start_pfn = page_to_pfn(page);
2290 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2291 start_page = pfn_to_page(start_pfn);
2292 end_page = start_page + pageblock_nr_pages - 1;
2293 end_pfn = start_pfn + pageblock_nr_pages - 1;
2295 /* Do not cross zone boundaries */
2296 if (!zone_spans_pfn(zone, start_pfn))
2298 if (!zone_spans_pfn(zone, end_pfn))
2301 return move_freepages(zone, start_page, end_page, migratetype,
2305 static void change_pageblock_range(struct page *pageblock_page,
2306 int start_order, int migratetype)
2308 int nr_pageblocks = 1 << (start_order - pageblock_order);
2310 while (nr_pageblocks--) {
2311 set_pageblock_migratetype(pageblock_page, migratetype);
2312 pageblock_page += pageblock_nr_pages;
2317 * When we are falling back to another migratetype during allocation, try to
2318 * steal extra free pages from the same pageblocks to satisfy further
2319 * allocations, instead of polluting multiple pageblocks.
2321 * If we are stealing a relatively large buddy page, it is likely there will
2322 * be more free pages in the pageblock, so try to steal them all. For
2323 * reclaimable and unmovable allocations, we steal regardless of page size,
2324 * as fragmentation caused by those allocations polluting movable pageblocks
2325 * is worse than movable allocations stealing from unmovable and reclaimable
2328 static bool can_steal_fallback(unsigned int order, int start_mt)
2331 * Leaving this order check is intended, although there is
2332 * relaxed order check in next check. The reason is that
2333 * we can actually steal whole pageblock if this condition met,
2334 * but, below check doesn't guarantee it and that is just heuristic
2335 * so could be changed anytime.
2337 if (order >= pageblock_order)
2340 if (order >= pageblock_order / 2 ||
2341 start_mt == MIGRATE_RECLAIMABLE ||
2342 start_mt == MIGRATE_UNMOVABLE ||
2343 page_group_by_mobility_disabled)
2349 static inline bool boost_watermark(struct zone *zone)
2351 unsigned long max_boost;
2353 if (!watermark_boost_factor)
2356 * Don't bother in zones that are unlikely to produce results.
2357 * On small machines, including kdump capture kernels running
2358 * in a small area, boosting the watermark can cause an out of
2359 * memory situation immediately.
2361 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2364 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2365 watermark_boost_factor, 10000);
2368 * high watermark may be uninitialised if fragmentation occurs
2369 * very early in boot so do not boost. We do not fall
2370 * through and boost by pageblock_nr_pages as failing
2371 * allocations that early means that reclaim is not going
2372 * to help and it may even be impossible to reclaim the
2373 * boosted watermark resulting in a hang.
2378 max_boost = max(pageblock_nr_pages, max_boost);
2380 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2387 * This function implements actual steal behaviour. If order is large enough,
2388 * we can steal whole pageblock. If not, we first move freepages in this
2389 * pageblock to our migratetype and determine how many already-allocated pages
2390 * are there in the pageblock with a compatible migratetype. If at least half
2391 * of pages are free or compatible, we can change migratetype of the pageblock
2392 * itself, so pages freed in the future will be put on the correct free list.
2394 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2395 unsigned int alloc_flags, int start_type, bool whole_block)
2397 unsigned int current_order = page_order(page);
2398 struct free_area *area;
2399 int free_pages, movable_pages, alike_pages;
2402 old_block_type = get_pageblock_migratetype(page);
2405 * This can happen due to races and we want to prevent broken
2406 * highatomic accounting.
2408 if (is_migrate_highatomic(old_block_type))
2411 /* Take ownership for orders >= pageblock_order */
2412 if (current_order >= pageblock_order) {
2413 change_pageblock_range(page, current_order, start_type);
2418 * Boost watermarks to increase reclaim pressure to reduce the
2419 * likelihood of future fallbacks. Wake kswapd now as the node
2420 * may be balanced overall and kswapd will not wake naturally.
2422 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
2423 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2425 /* We are not allowed to try stealing from the whole block */
2429 free_pages = move_freepages_block(zone, page, start_type,
2432 * Determine how many pages are compatible with our allocation.
2433 * For movable allocation, it's the number of movable pages which
2434 * we just obtained. For other types it's a bit more tricky.
2436 if (start_type == MIGRATE_MOVABLE) {
2437 alike_pages = movable_pages;
2440 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2441 * to MOVABLE pageblock, consider all non-movable pages as
2442 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2443 * vice versa, be conservative since we can't distinguish the
2444 * exact migratetype of non-movable pages.
2446 if (old_block_type == MIGRATE_MOVABLE)
2447 alike_pages = pageblock_nr_pages
2448 - (free_pages + movable_pages);
2453 /* moving whole block can fail due to zone boundary conditions */
2458 * If a sufficient number of pages in the block are either free or of
2459 * comparable migratability as our allocation, claim the whole block.
2461 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2462 page_group_by_mobility_disabled)
2463 set_pageblock_migratetype(page, start_type);
2468 area = &zone->free_area[current_order];
2469 move_to_free_area(page, area, start_type);
2473 * Check whether there is a suitable fallback freepage with requested order.
2474 * If only_stealable is true, this function returns fallback_mt only if
2475 * we can steal other freepages all together. This would help to reduce
2476 * fragmentation due to mixed migratetype pages in one pageblock.
2478 int find_suitable_fallback(struct free_area *area, unsigned int order,
2479 int migratetype, bool only_stealable, bool *can_steal)
2484 if (area->nr_free == 0)
2489 fallback_mt = fallbacks[migratetype][i];
2490 if (fallback_mt == MIGRATE_TYPES)
2493 if (free_area_empty(area, fallback_mt))
2496 if (can_steal_fallback(order, migratetype))
2499 if (!only_stealable)
2510 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2511 * there are no empty page blocks that contain a page with a suitable order
2513 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2514 unsigned int alloc_order)
2517 unsigned long max_managed, flags;
2520 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2521 * Check is race-prone but harmless.
2523 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2524 if (zone->nr_reserved_highatomic >= max_managed)
2527 spin_lock_irqsave(&zone->lock, flags);
2529 /* Recheck the nr_reserved_highatomic limit under the lock */
2530 if (zone->nr_reserved_highatomic >= max_managed)
2534 mt = get_pageblock_migratetype(page);
2535 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2536 && !is_migrate_cma(mt)) {
2537 zone->nr_reserved_highatomic += pageblock_nr_pages;
2538 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2539 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2543 spin_unlock_irqrestore(&zone->lock, flags);
2547 * Used when an allocation is about to fail under memory pressure. This
2548 * potentially hurts the reliability of high-order allocations when under
2549 * intense memory pressure but failed atomic allocations should be easier
2550 * to recover from than an OOM.
2552 * If @force is true, try to unreserve a pageblock even though highatomic
2553 * pageblock is exhausted.
2555 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2558 struct zonelist *zonelist = ac->zonelist;
2559 unsigned long flags;
2566 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2569 * Preserve at least one pageblock unless memory pressure
2572 if (!force && zone->nr_reserved_highatomic <=
2576 spin_lock_irqsave(&zone->lock, flags);
2577 for (order = 0; order < MAX_ORDER; order++) {
2578 struct free_area *area = &(zone->free_area[order]);
2580 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2585 * In page freeing path, migratetype change is racy so
2586 * we can counter several free pages in a pageblock
2587 * in this loop althoug we changed the pageblock type
2588 * from highatomic to ac->migratetype. So we should
2589 * adjust the count once.
2591 if (is_migrate_highatomic_page(page)) {
2593 * It should never happen but changes to
2594 * locking could inadvertently allow a per-cpu
2595 * drain to add pages to MIGRATE_HIGHATOMIC
2596 * while unreserving so be safe and watch for
2599 zone->nr_reserved_highatomic -= min(
2601 zone->nr_reserved_highatomic);
2605 * Convert to ac->migratetype and avoid the normal
2606 * pageblock stealing heuristics. Minimally, the caller
2607 * is doing the work and needs the pages. More
2608 * importantly, if the block was always converted to
2609 * MIGRATE_UNMOVABLE or another type then the number
2610 * of pageblocks that cannot be completely freed
2613 set_pageblock_migratetype(page, ac->migratetype);
2614 ret = move_freepages_block(zone, page, ac->migratetype,
2617 spin_unlock_irqrestore(&zone->lock, flags);
2621 spin_unlock_irqrestore(&zone->lock, flags);
2628 * Try finding a free buddy page on the fallback list and put it on the free
2629 * list of requested migratetype, possibly along with other pages from the same
2630 * block, depending on fragmentation avoidance heuristics. Returns true if
2631 * fallback was found so that __rmqueue_smallest() can grab it.
2633 * The use of signed ints for order and current_order is a deliberate
2634 * deviation from the rest of this file, to make the for loop
2635 * condition simpler.
2637 static __always_inline bool
2638 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2639 unsigned int alloc_flags)
2641 struct free_area *area;
2643 int min_order = order;
2649 * Do not steal pages from freelists belonging to other pageblocks
2650 * i.e. orders < pageblock_order. If there are no local zones free,
2651 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2653 if (alloc_flags & ALLOC_NOFRAGMENT)
2654 min_order = pageblock_order;
2657 * Find the largest available free page in the other list. This roughly
2658 * approximates finding the pageblock with the most free pages, which
2659 * would be too costly to do exactly.
2661 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2663 area = &(zone->free_area[current_order]);
2664 fallback_mt = find_suitable_fallback(area, current_order,
2665 start_migratetype, false, &can_steal);
2666 if (fallback_mt == -1)
2670 * We cannot steal all free pages from the pageblock and the
2671 * requested migratetype is movable. In that case it's better to
2672 * steal and split the smallest available page instead of the
2673 * largest available page, because even if the next movable
2674 * allocation falls back into a different pageblock than this
2675 * one, it won't cause permanent fragmentation.
2677 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2678 && current_order > order)
2687 for (current_order = order; current_order < MAX_ORDER;
2689 area = &(zone->free_area[current_order]);
2690 fallback_mt = find_suitable_fallback(area, current_order,
2691 start_migratetype, false, &can_steal);
2692 if (fallback_mt != -1)
2697 * This should not happen - we already found a suitable fallback
2698 * when looking for the largest page.
2700 VM_BUG_ON(current_order == MAX_ORDER);
2703 page = get_page_from_free_area(area, fallback_mt);
2705 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2708 trace_mm_page_alloc_extfrag(page, order, current_order,
2709 start_migratetype, fallback_mt);
2716 * Do the hard work of removing an element from the buddy allocator.
2717 * Call me with the zone->lock already held.
2719 static __always_inline struct page *
2720 __rmqueue(struct zone *zone, unsigned int order, int migratetype,
2721 unsigned int alloc_flags)
2726 page = __rmqueue_smallest(zone, order, migratetype);
2727 if (unlikely(!page)) {
2728 if (migratetype == MIGRATE_MOVABLE)
2729 page = __rmqueue_cma_fallback(zone, order);
2731 if (!page && __rmqueue_fallback(zone, order, migratetype,
2736 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2741 * Obtain a specified number of elements from the buddy allocator, all under
2742 * a single hold of the lock, for efficiency. Add them to the supplied list.
2743 * Returns the number of new pages which were placed at *list.
2745 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2746 unsigned long count, struct list_head *list,
2747 int migratetype, unsigned int alloc_flags)
2751 spin_lock(&zone->lock);
2752 for (i = 0; i < count; ++i) {
2753 struct page *page = __rmqueue(zone, order, migratetype,
2755 if (unlikely(page == NULL))
2758 if (unlikely(check_pcp_refill(page)))
2762 * Split buddy pages returned by expand() are received here in
2763 * physical page order. The page is added to the tail of
2764 * caller's list. From the callers perspective, the linked list
2765 * is ordered by page number under some conditions. This is
2766 * useful for IO devices that can forward direction from the
2767 * head, thus also in the physical page order. This is useful
2768 * for IO devices that can merge IO requests if the physical
2769 * pages are ordered properly.
2771 list_add_tail(&page->lru, list);
2773 if (is_migrate_cma(get_pcppage_migratetype(page)))
2774 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2779 * i pages were removed from the buddy list even if some leak due
2780 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2781 * on i. Do not confuse with 'alloced' which is the number of
2782 * pages added to the pcp list.
2784 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2785 spin_unlock(&zone->lock);
2791 * Called from the vmstat counter updater to drain pagesets of this
2792 * currently executing processor on remote nodes after they have
2795 * Note that this function must be called with the thread pinned to
2796 * a single processor.
2798 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2800 unsigned long flags;
2801 int to_drain, batch;
2803 local_irq_save(flags);
2804 batch = READ_ONCE(pcp->batch);
2805 to_drain = min(pcp->count, batch);
2807 free_pcppages_bulk(zone, to_drain, pcp);
2808 local_irq_restore(flags);
2813 * Drain pcplists of the indicated processor and zone.
2815 * The processor must either be the current processor and the
2816 * thread pinned to the current processor or a processor that
2819 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2821 unsigned long flags;
2822 struct per_cpu_pageset *pset;
2823 struct per_cpu_pages *pcp;
2825 local_irq_save(flags);
2826 pset = per_cpu_ptr(zone->pageset, cpu);
2830 free_pcppages_bulk(zone, pcp->count, pcp);
2831 local_irq_restore(flags);
2835 * Drain pcplists of all zones on the indicated processor.
2837 * The processor must either be the current processor and the
2838 * thread pinned to the current processor or a processor that
2841 static void drain_pages(unsigned int cpu)
2845 for_each_populated_zone(zone) {
2846 drain_pages_zone(cpu, zone);
2851 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2853 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2854 * the single zone's pages.
2856 void drain_local_pages(struct zone *zone)
2858 int cpu = smp_processor_id();
2861 drain_pages_zone(cpu, zone);
2866 static void drain_local_pages_wq(struct work_struct *work)
2868 struct pcpu_drain *drain;
2870 drain = container_of(work, struct pcpu_drain, work);
2873 * drain_all_pages doesn't use proper cpu hotplug protection so
2874 * we can race with cpu offline when the WQ can move this from
2875 * a cpu pinned worker to an unbound one. We can operate on a different
2876 * cpu which is allright but we also have to make sure to not move to
2880 drain_local_pages(drain->zone);
2885 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2887 * When zone parameter is non-NULL, spill just the single zone's pages.
2889 * Note that this can be extremely slow as the draining happens in a workqueue.
2891 void drain_all_pages(struct zone *zone)
2896 * Allocate in the BSS so we wont require allocation in
2897 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2899 static cpumask_t cpus_with_pcps;
2902 * Make sure nobody triggers this path before mm_percpu_wq is fully
2905 if (WARN_ON_ONCE(!mm_percpu_wq))
2909 * Do not drain if one is already in progress unless it's specific to
2910 * a zone. Such callers are primarily CMA and memory hotplug and need
2911 * the drain to be complete when the call returns.
2913 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2916 mutex_lock(&pcpu_drain_mutex);
2920 * We don't care about racing with CPU hotplug event
2921 * as offline notification will cause the notified
2922 * cpu to drain that CPU pcps and on_each_cpu_mask
2923 * disables preemption as part of its processing
2925 for_each_online_cpu(cpu) {
2926 struct per_cpu_pageset *pcp;
2928 bool has_pcps = false;
2931 pcp = per_cpu_ptr(zone->pageset, cpu);
2935 for_each_populated_zone(z) {
2936 pcp = per_cpu_ptr(z->pageset, cpu);
2937 if (pcp->pcp.count) {
2945 cpumask_set_cpu(cpu, &cpus_with_pcps);
2947 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2950 for_each_cpu(cpu, &cpus_with_pcps) {
2951 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2954 INIT_WORK(&drain->work, drain_local_pages_wq);
2955 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2957 for_each_cpu(cpu, &cpus_with_pcps)
2958 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2960 mutex_unlock(&pcpu_drain_mutex);
2963 #ifdef CONFIG_HIBERNATION
2966 * Touch the watchdog for every WD_PAGE_COUNT pages.
2968 #define WD_PAGE_COUNT (128*1024)
2970 void mark_free_pages(struct zone *zone)
2972 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2973 unsigned long flags;
2974 unsigned int order, t;
2977 if (zone_is_empty(zone))
2980 spin_lock_irqsave(&zone->lock, flags);
2982 max_zone_pfn = zone_end_pfn(zone);
2983 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2984 if (pfn_valid(pfn)) {
2985 page = pfn_to_page(pfn);
2987 if (!--page_count) {
2988 touch_nmi_watchdog();
2989 page_count = WD_PAGE_COUNT;
2992 if (page_zone(page) != zone)
2995 if (!swsusp_page_is_forbidden(page))
2996 swsusp_unset_page_free(page);
2999 for_each_migratetype_order(order, t) {
3000 list_for_each_entry(page,
3001 &zone->free_area[order].free_list[t], lru) {
3004 pfn = page_to_pfn(page);
3005 for (i = 0; i < (1UL << order); i++) {
3006 if (!--page_count) {
3007 touch_nmi_watchdog();
3008 page_count = WD_PAGE_COUNT;
3010 swsusp_set_page_free(pfn_to_page(pfn + i));
3014 spin_unlock_irqrestore(&zone->lock, flags);
3016 #endif /* CONFIG_PM */
3018 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3022 if (!free_pcp_prepare(page))
3025 migratetype = get_pfnblock_migratetype(page, pfn);
3026 set_pcppage_migratetype(page, migratetype);
3030 static void free_unref_page_commit(struct page *page, unsigned long pfn)
3032 struct zone *zone = page_zone(page);
3033 struct per_cpu_pages *pcp;
3036 migratetype = get_pcppage_migratetype(page);
3037 __count_vm_event(PGFREE);
3040 * We only track unmovable, reclaimable and movable on pcp lists.
3041 * Free ISOLATE pages back to the allocator because they are being
3042 * offlined but treat HIGHATOMIC as movable pages so we can get those
3043 * areas back if necessary. Otherwise, we may have to free
3044 * excessively into the page allocator
3046 if (migratetype >= MIGRATE_PCPTYPES) {
3047 if (unlikely(is_migrate_isolate(migratetype))) {
3048 free_one_page(zone, page, pfn, 0, migratetype);
3051 migratetype = MIGRATE_MOVABLE;
3054 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3055 list_add(&page->lru, &pcp->lists[migratetype]);
3057 if (pcp->count >= pcp->high) {
3058 unsigned long batch = READ_ONCE(pcp->batch);
3059 free_pcppages_bulk(zone, batch, pcp);
3064 * Free a 0-order page
3066 void free_unref_page(struct page *page)
3068 unsigned long flags;
3069 unsigned long pfn = page_to_pfn(page);
3071 if (!free_unref_page_prepare(page, pfn))
3074 local_irq_save(flags);
3075 free_unref_page_commit(page, pfn);
3076 local_irq_restore(flags);
3080 * Free a list of 0-order pages
3082 void free_unref_page_list(struct list_head *list)
3084 struct page *page, *next;
3085 unsigned long flags, pfn;
3086 int batch_count = 0;
3088 /* Prepare pages for freeing */
3089 list_for_each_entry_safe(page, next, list, lru) {
3090 pfn = page_to_pfn(page);
3091 if (!free_unref_page_prepare(page, pfn))
3092 list_del(&page->lru);
3093 set_page_private(page, pfn);
3096 local_irq_save(flags);
3097 list_for_each_entry_safe(page, next, list, lru) {
3098 unsigned long pfn = page_private(page);
3100 set_page_private(page, 0);
3101 trace_mm_page_free_batched(page);
3102 free_unref_page_commit(page, pfn);
3105 * Guard against excessive IRQ disabled times when we get
3106 * a large list of pages to free.
3108 if (++batch_count == SWAP_CLUSTER_MAX) {
3109 local_irq_restore(flags);
3111 local_irq_save(flags);
3114 local_irq_restore(flags);
3118 * split_page takes a non-compound higher-order page, and splits it into
3119 * n (1<<order) sub-pages: page[0..n]
3120 * Each sub-page must be freed individually.
3122 * Note: this is probably too low level an operation for use in drivers.
3123 * Please consult with lkml before using this in your driver.
3125 void split_page(struct page *page, unsigned int order)
3129 VM_BUG_ON_PAGE(PageCompound(page), page);
3130 VM_BUG_ON_PAGE(!page_count(page), page);
3132 for (i = 1; i < (1 << order); i++)
3133 set_page_refcounted(page + i);
3134 split_page_owner(page, 1 << order);
3136 EXPORT_SYMBOL_GPL(split_page);
3138 int __isolate_free_page(struct page *page, unsigned int order)
3140 struct free_area *area = &page_zone(page)->free_area[order];
3141 unsigned long watermark;
3145 BUG_ON(!PageBuddy(page));
3147 zone = page_zone(page);
3148 mt = get_pageblock_migratetype(page);
3150 if (!is_migrate_isolate(mt)) {
3152 * Obey watermarks as if the page was being allocated. We can
3153 * emulate a high-order watermark check with a raised order-0
3154 * watermark, because we already know our high-order page
3157 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3158 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3161 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3164 /* Remove page from free list */
3166 del_page_from_free_area(page, area);
3169 * Set the pageblock if the isolated page is at least half of a
3172 if (order >= pageblock_order - 1) {
3173 struct page *endpage = page + (1 << order) - 1;
3174 for (; page < endpage; page += pageblock_nr_pages) {
3175 int mt = get_pageblock_migratetype(page);
3176 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3177 && !is_migrate_highatomic(mt))
3178 set_pageblock_migratetype(page,
3184 return 1UL << order;
3188 * Update NUMA hit/miss statistics
3190 * Must be called with interrupts disabled.
3192 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3195 enum numa_stat_item local_stat = NUMA_LOCAL;
3197 /* skip numa counters update if numa stats is disabled */
3198 if (!static_branch_likely(&vm_numa_stat_key))
3201 if (zone_to_nid(z) != numa_node_id())
3202 local_stat = NUMA_OTHER;
3204 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3205 __inc_numa_state(z, NUMA_HIT);
3207 __inc_numa_state(z, NUMA_MISS);
3208 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3210 __inc_numa_state(z, local_stat);
3214 /* Remove page from the per-cpu list, caller must protect the list */
3215 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3216 unsigned int alloc_flags,
3217 struct per_cpu_pages *pcp,
3218 struct list_head *list)
3223 if (list_empty(list)) {
3224 pcp->count += rmqueue_bulk(zone, 0,
3226 migratetype, alloc_flags);
3227 if (unlikely(list_empty(list)))
3231 page = list_first_entry(list, struct page, lru);
3232 list_del(&page->lru);
3234 } while (check_new_pcp(page));
3239 /* Lock and remove page from the per-cpu list */
3240 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3241 struct zone *zone, gfp_t gfp_flags,
3242 int migratetype, unsigned int alloc_flags)
3244 struct per_cpu_pages *pcp;
3245 struct list_head *list;
3247 unsigned long flags;
3249 local_irq_save(flags);
3250 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3251 list = &pcp->lists[migratetype];
3252 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3254 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3255 zone_statistics(preferred_zone, zone);
3257 local_irq_restore(flags);
3262 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3265 struct page *rmqueue(struct zone *preferred_zone,
3266 struct zone *zone, unsigned int order,
3267 gfp_t gfp_flags, unsigned int alloc_flags,
3270 unsigned long flags;
3273 if (likely(order == 0)) {
3274 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3275 migratetype, alloc_flags);
3280 * We most definitely don't want callers attempting to
3281 * allocate greater than order-1 page units with __GFP_NOFAIL.
3283 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3284 spin_lock_irqsave(&zone->lock, flags);
3288 if (alloc_flags & ALLOC_HARDER) {
3289 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3291 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3294 page = __rmqueue(zone, order, migratetype, alloc_flags);
3295 } while (page && check_new_pages(page, order));
3296 spin_unlock(&zone->lock);
3299 __mod_zone_freepage_state(zone, -(1 << order),
3300 get_pcppage_migratetype(page));
3302 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3303 zone_statistics(preferred_zone, zone);
3304 local_irq_restore(flags);
3307 /* Separate test+clear to avoid unnecessary atomics */
3308 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3309 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3310 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3313 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3317 local_irq_restore(flags);
3321 #ifdef CONFIG_FAIL_PAGE_ALLOC
3324 struct fault_attr attr;
3326 bool ignore_gfp_highmem;
3327 bool ignore_gfp_reclaim;
3329 } fail_page_alloc = {
3330 .attr = FAULT_ATTR_INITIALIZER,
3331 .ignore_gfp_reclaim = true,
3332 .ignore_gfp_highmem = true,
3336 static int __init setup_fail_page_alloc(char *str)
3338 return setup_fault_attr(&fail_page_alloc.attr, str);
3340 __setup("fail_page_alloc=", setup_fail_page_alloc);
3342 static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3344 if (order < fail_page_alloc.min_order)
3346 if (gfp_mask & __GFP_NOFAIL)
3348 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3350 if (fail_page_alloc.ignore_gfp_reclaim &&
3351 (gfp_mask & __GFP_DIRECT_RECLAIM))
3354 return should_fail(&fail_page_alloc.attr, 1 << order);
3357 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3359 static int __init fail_page_alloc_debugfs(void)
3361 umode_t mode = S_IFREG | 0600;
3364 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3365 &fail_page_alloc.attr);
3367 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3368 &fail_page_alloc.ignore_gfp_reclaim);
3369 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3370 &fail_page_alloc.ignore_gfp_highmem);
3371 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3376 late_initcall(fail_page_alloc_debugfs);
3378 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3380 #else /* CONFIG_FAIL_PAGE_ALLOC */
3382 static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3387 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3389 noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3391 return __should_fail_alloc_page(gfp_mask, order);
3393 ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3396 * Return true if free base pages are above 'mark'. For high-order checks it
3397 * will return true of the order-0 watermark is reached and there is at least
3398 * one free page of a suitable size. Checking now avoids taking the zone lock
3399 * to check in the allocation paths if no pages are free.
3401 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3402 int classzone_idx, unsigned int alloc_flags,
3407 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3409 /* free_pages may go negative - that's OK */
3410 free_pages -= (1 << order) - 1;
3412 if (alloc_flags & ALLOC_HIGH)
3416 * If the caller does not have rights to ALLOC_HARDER then subtract
3417 * the high-atomic reserves. This will over-estimate the size of the
3418 * atomic reserve but it avoids a search.
3420 if (likely(!alloc_harder)) {
3421 free_pages -= z->nr_reserved_highatomic;
3424 * OOM victims can try even harder than normal ALLOC_HARDER
3425 * users on the grounds that it's definitely going to be in
3426 * the exit path shortly and free memory. Any allocation it
3427 * makes during the free path will be small and short-lived.
3429 if (alloc_flags & ALLOC_OOM)
3437 /* If allocation can't use CMA areas don't use free CMA pages */
3438 if (!(alloc_flags & ALLOC_CMA))
3439 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3443 * Check watermarks for an order-0 allocation request. If these
3444 * are not met, then a high-order request also cannot go ahead
3445 * even if a suitable page happened to be free.
3447 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3450 /* If this is an order-0 request then the watermark is fine */
3454 /* For a high-order request, check at least one suitable page is free */
3455 for (o = order; o < MAX_ORDER; o++) {
3456 struct free_area *area = &z->free_area[o];
3462 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3463 if (!free_area_empty(area, mt))
3468 if ((alloc_flags & ALLOC_CMA) &&
3469 !free_area_empty(area, MIGRATE_CMA)) {
3474 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3480 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3481 int classzone_idx, unsigned int alloc_flags)
3483 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3484 zone_page_state(z, NR_FREE_PAGES));
3487 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3488 unsigned long mark, int classzone_idx,
3489 unsigned int alloc_flags, gfp_t gfp_mask)
3491 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3495 /* If allocation can't use CMA areas don't use free CMA pages */
3496 if (!(alloc_flags & ALLOC_CMA))
3497 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3501 * Fast check for order-0 only. If this fails then the reserves
3502 * need to be calculated. There is a corner case where the check
3503 * passes but only the high-order atomic reserve are free. If
3504 * the caller is !atomic then it'll uselessly search the free
3505 * list. That corner case is then slower but it is harmless.
3507 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3510 if (__zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3514 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3515 * when checking the min watermark. The min watermark is the
3516 * point where boosting is ignored so that kswapd is woken up
3517 * when below the low watermark.
3519 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3520 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3521 mark = z->_watermark[WMARK_MIN];
3522 return __zone_watermark_ok(z, order, mark, classzone_idx,
3523 alloc_flags, free_pages);
3529 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3530 unsigned long mark, int classzone_idx)
3532 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3534 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3535 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3537 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3542 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3544 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3545 node_reclaim_distance;
3547 #else /* CONFIG_NUMA */
3548 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3552 #endif /* CONFIG_NUMA */
3555 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3556 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3557 * premature use of a lower zone may cause lowmem pressure problems that
3558 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3559 * probably too small. It only makes sense to spread allocations to avoid
3560 * fragmentation between the Normal and DMA32 zones.
3562 static inline unsigned int
3563 alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3565 unsigned int alloc_flags = 0;
3567 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3568 alloc_flags |= ALLOC_KSWAPD;
3570 #ifdef CONFIG_ZONE_DMA32
3574 if (zone_idx(zone) != ZONE_NORMAL)
3578 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3579 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3580 * on UMA that if Normal is populated then so is DMA32.
3582 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3583 if (nr_online_nodes > 1 && !populated_zone(--zone))
3586 alloc_flags |= ALLOC_NOFRAGMENT;
3587 #endif /* CONFIG_ZONE_DMA32 */
3592 * get_page_from_freelist goes through the zonelist trying to allocate
3595 static struct page *
3596 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3597 const struct alloc_context *ac)
3601 struct pglist_data *last_pgdat_dirty_limit = NULL;
3606 * Scan zonelist, looking for a zone with enough free.
3607 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3609 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3610 z = ac->preferred_zoneref;
3611 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3616 if (cpusets_enabled() &&
3617 (alloc_flags & ALLOC_CPUSET) &&
3618 !__cpuset_zone_allowed(zone, gfp_mask))
3621 * When allocating a page cache page for writing, we
3622 * want to get it from a node that is within its dirty
3623 * limit, such that no single node holds more than its
3624 * proportional share of globally allowed dirty pages.
3625 * The dirty limits take into account the node's
3626 * lowmem reserves and high watermark so that kswapd
3627 * should be able to balance it without having to
3628 * write pages from its LRU list.
3630 * XXX: For now, allow allocations to potentially
3631 * exceed the per-node dirty limit in the slowpath
3632 * (spread_dirty_pages unset) before going into reclaim,
3633 * which is important when on a NUMA setup the allowed
3634 * nodes are together not big enough to reach the
3635 * global limit. The proper fix for these situations
3636 * will require awareness of nodes in the
3637 * dirty-throttling and the flusher threads.
3639 if (ac->spread_dirty_pages) {
3640 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3643 if (!node_dirty_ok(zone->zone_pgdat)) {
3644 last_pgdat_dirty_limit = zone->zone_pgdat;
3649 if (no_fallback && nr_online_nodes > 1 &&
3650 zone != ac->preferred_zoneref->zone) {
3654 * If moving to a remote node, retry but allow
3655 * fragmenting fallbacks. Locality is more important
3656 * than fragmentation avoidance.
3658 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3659 if (zone_to_nid(zone) != local_nid) {
3660 alloc_flags &= ~ALLOC_NOFRAGMENT;
3665 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3666 if (!zone_watermark_fast(zone, order, mark,
3667 ac_classzone_idx(ac), alloc_flags,
3671 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3673 * Watermark failed for this zone, but see if we can
3674 * grow this zone if it contains deferred pages.
3676 if (static_branch_unlikely(&deferred_pages)) {
3677 if (_deferred_grow_zone(zone, order))
3681 /* Checked here to keep the fast path fast */
3682 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3683 if (alloc_flags & ALLOC_NO_WATERMARKS)
3686 if (node_reclaim_mode == 0 ||
3687 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3690 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3692 case NODE_RECLAIM_NOSCAN:
3695 case NODE_RECLAIM_FULL:
3696 /* scanned but unreclaimable */
3699 /* did we reclaim enough */
3700 if (zone_watermark_ok(zone, order, mark,
3701 ac_classzone_idx(ac), alloc_flags))
3709 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3710 gfp_mask, alloc_flags, ac->migratetype);
3712 prep_new_page(page, order, gfp_mask, alloc_flags);
3715 * If this is a high-order atomic allocation then check
3716 * if the pageblock should be reserved for the future
3718 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3719 reserve_highatomic_pageblock(page, zone, order);
3723 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3724 /* Try again if zone has deferred pages */
3725 if (static_branch_unlikely(&deferred_pages)) {
3726 if (_deferred_grow_zone(zone, order))
3734 * It's possible on a UMA machine to get through all zones that are
3735 * fragmented. If avoiding fragmentation, reset and try again.
3738 alloc_flags &= ~ALLOC_NOFRAGMENT;
3745 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3747 unsigned int filter = SHOW_MEM_FILTER_NODES;
3750 * This documents exceptions given to allocations in certain
3751 * contexts that are allowed to allocate outside current's set
3754 if (!(gfp_mask & __GFP_NOMEMALLOC))
3755 if (tsk_is_oom_victim(current) ||
3756 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3757 filter &= ~SHOW_MEM_FILTER_NODES;
3758 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3759 filter &= ~SHOW_MEM_FILTER_NODES;
3761 show_mem(filter, nodemask);
3764 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3766 struct va_format vaf;
3768 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3770 if ((gfp_mask & __GFP_NOWARN) ||
3771 !__ratelimit(&nopage_rs) ||
3772 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3775 va_start(args, fmt);
3778 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3779 current->comm, &vaf, gfp_mask, &gfp_mask,
3780 nodemask_pr_args(nodemask));
3783 cpuset_print_current_mems_allowed();
3786 warn_alloc_show_mem(gfp_mask, nodemask);
3789 static inline struct page *
3790 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3791 unsigned int alloc_flags,
3792 const struct alloc_context *ac)
3796 page = get_page_from_freelist(gfp_mask, order,
3797 alloc_flags|ALLOC_CPUSET, ac);
3799 * fallback to ignore cpuset restriction if our nodes
3803 page = get_page_from_freelist(gfp_mask, order,
3809 static inline struct page *
3810 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3811 const struct alloc_context *ac, unsigned long *did_some_progress)
3813 struct oom_control oc = {
3814 .zonelist = ac->zonelist,
3815 .nodemask = ac->nodemask,
3817 .gfp_mask = gfp_mask,
3822 *did_some_progress = 0;
3825 * Acquire the oom lock. If that fails, somebody else is
3826 * making progress for us.
3828 if (!mutex_trylock(&oom_lock)) {
3829 *did_some_progress = 1;
3830 schedule_timeout_uninterruptible(1);
3835 * Go through the zonelist yet one more time, keep very high watermark
3836 * here, this is only to catch a parallel oom killing, we must fail if
3837 * we're still under heavy pressure. But make sure that this reclaim
3838 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3839 * allocation which will never fail due to oom_lock already held.
3841 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3842 ~__GFP_DIRECT_RECLAIM, order,
3843 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3847 /* Coredumps can quickly deplete all memory reserves */
3848 if (current->flags & PF_DUMPCORE)
3850 /* The OOM killer will not help higher order allocs */
3851 if (order > PAGE_ALLOC_COSTLY_ORDER)
3854 * We have already exhausted all our reclaim opportunities without any
3855 * success so it is time to admit defeat. We will skip the OOM killer
3856 * because it is very likely that the caller has a more reasonable
3857 * fallback than shooting a random task.
3859 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3861 /* The OOM killer does not needlessly kill tasks for lowmem */
3862 if (ac->high_zoneidx < ZONE_NORMAL)
3864 if (pm_suspended_storage())
3867 * XXX: GFP_NOFS allocations should rather fail than rely on
3868 * other request to make a forward progress.
3869 * We are in an unfortunate situation where out_of_memory cannot
3870 * do much for this context but let's try it to at least get
3871 * access to memory reserved if the current task is killed (see
3872 * out_of_memory). Once filesystems are ready to handle allocation
3873 * failures more gracefully we should just bail out here.
3876 /* The OOM killer may not free memory on a specific node */
3877 if (gfp_mask & __GFP_THISNODE)
3880 /* Exhausted what can be done so it's blame time */
3881 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3882 *did_some_progress = 1;
3885 * Help non-failing allocations by giving them access to memory
3888 if (gfp_mask & __GFP_NOFAIL)
3889 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3890 ALLOC_NO_WATERMARKS, ac);
3893 mutex_unlock(&oom_lock);
3898 * Maximum number of compaction retries wit a progress before OOM
3899 * killer is consider as the only way to move forward.
3901 #define MAX_COMPACT_RETRIES 16
3903 #ifdef CONFIG_COMPACTION
3904 /* Try memory compaction for high-order allocations before reclaim */
3905 static struct page *
3906 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3907 unsigned int alloc_flags, const struct alloc_context *ac,
3908 enum compact_priority prio, enum compact_result *compact_result)
3910 struct page *page = NULL;
3911 unsigned long pflags;
3912 unsigned int noreclaim_flag;
3917 psi_memstall_enter(&pflags);
3918 noreclaim_flag = memalloc_noreclaim_save();
3920 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3923 memalloc_noreclaim_restore(noreclaim_flag);
3924 psi_memstall_leave(&pflags);
3927 * At least in one zone compaction wasn't deferred or skipped, so let's
3928 * count a compaction stall
3930 count_vm_event(COMPACTSTALL);
3932 /* Prep a captured page if available */
3934 prep_new_page(page, order, gfp_mask, alloc_flags);
3936 /* Try get a page from the freelist if available */
3938 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3941 struct zone *zone = page_zone(page);
3943 zone->compact_blockskip_flush = false;
3944 compaction_defer_reset(zone, order, true);
3945 count_vm_event(COMPACTSUCCESS);
3950 * It's bad if compaction run occurs and fails. The most likely reason
3951 * is that pages exist, but not enough to satisfy watermarks.
3953 count_vm_event(COMPACTFAIL);
3961 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3962 enum compact_result compact_result,
3963 enum compact_priority *compact_priority,
3964 int *compaction_retries)
3966 int max_retries = MAX_COMPACT_RETRIES;
3969 int retries = *compaction_retries;
3970 enum compact_priority priority = *compact_priority;
3975 if (compaction_made_progress(compact_result))
3976 (*compaction_retries)++;
3979 * compaction considers all the zone as desperately out of memory
3980 * so it doesn't really make much sense to retry except when the
3981 * failure could be caused by insufficient priority
3983 if (compaction_failed(compact_result))
3984 goto check_priority;
3987 * compaction was skipped because there are not enough order-0 pages
3988 * to work with, so we retry only if it looks like reclaim can help.
3990 if (compaction_needs_reclaim(compact_result)) {
3991 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3996 * make sure the compaction wasn't deferred or didn't bail out early
3997 * due to locks contention before we declare that we should give up.
3998 * But the next retry should use a higher priority if allowed, so
3999 * we don't just keep bailing out endlessly.
4001 if (compaction_withdrawn(compact_result)) {
4002 goto check_priority;
4006 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4007 * costly ones because they are de facto nofail and invoke OOM
4008 * killer to move on while costly can fail and users are ready
4009 * to cope with that. 1/4 retries is rather arbitrary but we
4010 * would need much more detailed feedback from compaction to
4011 * make a better decision.
4013 if (order > PAGE_ALLOC_COSTLY_ORDER)
4015 if (*compaction_retries <= max_retries) {
4021 * Make sure there are attempts at the highest priority if we exhausted
4022 * all retries or failed at the lower priorities.
4025 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4026 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4028 if (*compact_priority > min_priority) {
4029 (*compact_priority)--;
4030 *compaction_retries = 0;
4034 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4038 static inline struct page *
4039 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4040 unsigned int alloc_flags, const struct alloc_context *ac,
4041 enum compact_priority prio, enum compact_result *compact_result)
4043 *compact_result = COMPACT_SKIPPED;
4048 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4049 enum compact_result compact_result,
4050 enum compact_priority *compact_priority,
4051 int *compaction_retries)
4056 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4060 * There are setups with compaction disabled which would prefer to loop
4061 * inside the allocator rather than hit the oom killer prematurely.
4062 * Let's give them a good hope and keep retrying while the order-0
4063 * watermarks are OK.
4065 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4067 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4068 ac_classzone_idx(ac), alloc_flags))
4073 #endif /* CONFIG_COMPACTION */
4075 #ifdef CONFIG_LOCKDEP
4076 static struct lockdep_map __fs_reclaim_map =
4077 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4079 static bool __need_fs_reclaim(gfp_t gfp_mask)
4081 gfp_mask = current_gfp_context(gfp_mask);
4083 /* no reclaim without waiting on it */
4084 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4087 /* this guy won't enter reclaim */
4088 if (current->flags & PF_MEMALLOC)
4091 /* We're only interested __GFP_FS allocations for now */
4092 if (!(gfp_mask & __GFP_FS))
4095 if (gfp_mask & __GFP_NOLOCKDEP)
4101 void __fs_reclaim_acquire(void)
4103 lock_map_acquire(&__fs_reclaim_map);
4106 void __fs_reclaim_release(void)
4108 lock_map_release(&__fs_reclaim_map);
4111 void fs_reclaim_acquire(gfp_t gfp_mask)
4113 if (__need_fs_reclaim(gfp_mask))
4114 __fs_reclaim_acquire();
4116 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4118 void fs_reclaim_release(gfp_t gfp_mask)
4120 if (__need_fs_reclaim(gfp_mask))
4121 __fs_reclaim_release();
4123 EXPORT_SYMBOL_GPL(fs_reclaim_release);
4126 /* Perform direct synchronous page reclaim */
4128 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
4129 const struct alloc_context *ac)
4132 unsigned int noreclaim_flag;
4133 unsigned long pflags;
4137 /* We now go into synchronous reclaim */
4138 cpuset_memory_pressure_bump();
4139 psi_memstall_enter(&pflags);
4140 fs_reclaim_acquire(gfp_mask);
4141 noreclaim_flag = memalloc_noreclaim_save();
4143 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4146 memalloc_noreclaim_restore(noreclaim_flag);
4147 fs_reclaim_release(gfp_mask);
4148 psi_memstall_leave(&pflags);
4155 /* The really slow allocator path where we enter direct reclaim */
4156 static inline struct page *
4157 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4158 unsigned int alloc_flags, const struct alloc_context *ac,
4159 unsigned long *did_some_progress)
4161 struct page *page = NULL;
4162 bool drained = false;
4164 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4165 if (unlikely(!(*did_some_progress)))
4169 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4172 * If an allocation failed after direct reclaim, it could be because
4173 * pages are pinned on the per-cpu lists or in high alloc reserves.
4174 * Shrink them them and try again
4176 if (!page && !drained) {
4177 unreserve_highatomic_pageblock(ac, false);
4178 drain_all_pages(NULL);
4186 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4187 const struct alloc_context *ac)
4191 pg_data_t *last_pgdat = NULL;
4192 enum zone_type high_zoneidx = ac->high_zoneidx;
4194 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4196 if (last_pgdat != zone->zone_pgdat)
4197 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4198 last_pgdat = zone->zone_pgdat;
4202 static inline unsigned int
4203 gfp_to_alloc_flags(gfp_t gfp_mask)
4205 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4207 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4208 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4211 * The caller may dip into page reserves a bit more if the caller
4212 * cannot run direct reclaim, or if the caller has realtime scheduling
4213 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4214 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4216 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4218 if (gfp_mask & __GFP_ATOMIC) {
4220 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4221 * if it can't schedule.
4223 if (!(gfp_mask & __GFP_NOMEMALLOC))
4224 alloc_flags |= ALLOC_HARDER;
4226 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4227 * comment for __cpuset_node_allowed().
4229 alloc_flags &= ~ALLOC_CPUSET;
4230 } else if (unlikely(rt_task(current)) && !in_interrupt())
4231 alloc_flags |= ALLOC_HARDER;
4233 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4234 alloc_flags |= ALLOC_KSWAPD;
4237 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4238 alloc_flags |= ALLOC_CMA;
4243 static bool oom_reserves_allowed(struct task_struct *tsk)
4245 if (!tsk_is_oom_victim(tsk))
4249 * !MMU doesn't have oom reaper so give access to memory reserves
4250 * only to the thread with TIF_MEMDIE set
4252 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4259 * Distinguish requests which really need access to full memory
4260 * reserves from oom victims which can live with a portion of it
4262 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4264 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4266 if (gfp_mask & __GFP_MEMALLOC)
4267 return ALLOC_NO_WATERMARKS;
4268 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4269 return ALLOC_NO_WATERMARKS;
4270 if (!in_interrupt()) {
4271 if (current->flags & PF_MEMALLOC)
4272 return ALLOC_NO_WATERMARKS;
4273 else if (oom_reserves_allowed(current))
4280 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4282 return !!__gfp_pfmemalloc_flags(gfp_mask);
4286 * Checks whether it makes sense to retry the reclaim to make a forward progress
4287 * for the given allocation request.
4289 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4290 * without success, or when we couldn't even meet the watermark if we
4291 * reclaimed all remaining pages on the LRU lists.
4293 * Returns true if a retry is viable or false to enter the oom path.
4296 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4297 struct alloc_context *ac, int alloc_flags,
4298 bool did_some_progress, int *no_progress_loops)
4305 * Costly allocations might have made a progress but this doesn't mean
4306 * their order will become available due to high fragmentation so
4307 * always increment the no progress counter for them
4309 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4310 *no_progress_loops = 0;
4312 (*no_progress_loops)++;
4315 * Make sure we converge to OOM if we cannot make any progress
4316 * several times in the row.
4318 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4319 /* Before OOM, exhaust highatomic_reserve */
4320 return unreserve_highatomic_pageblock(ac, true);
4324 * Keep reclaiming pages while there is a chance this will lead
4325 * somewhere. If none of the target zones can satisfy our allocation
4326 * request even if all reclaimable pages are considered then we are
4327 * screwed and have to go OOM.
4329 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4331 unsigned long available;
4332 unsigned long reclaimable;
4333 unsigned long min_wmark = min_wmark_pages(zone);
4336 available = reclaimable = zone_reclaimable_pages(zone);
4337 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4340 * Would the allocation succeed if we reclaimed all
4341 * reclaimable pages?
4343 wmark = __zone_watermark_ok(zone, order, min_wmark,
4344 ac_classzone_idx(ac), alloc_flags, available);
4345 trace_reclaim_retry_zone(z, order, reclaimable,
4346 available, min_wmark, *no_progress_loops, wmark);
4349 * If we didn't make any progress and have a lot of
4350 * dirty + writeback pages then we should wait for
4351 * an IO to complete to slow down the reclaim and
4352 * prevent from pre mature OOM
4354 if (!did_some_progress) {
4355 unsigned long write_pending;
4357 write_pending = zone_page_state_snapshot(zone,
4358 NR_ZONE_WRITE_PENDING);
4360 if (2 * write_pending > reclaimable) {
4361 congestion_wait(BLK_RW_ASYNC, HZ/10);
4373 * Memory allocation/reclaim might be called from a WQ context and the
4374 * current implementation of the WQ concurrency control doesn't
4375 * recognize that a particular WQ is congested if the worker thread is
4376 * looping without ever sleeping. Therefore we have to do a short sleep
4377 * here rather than calling cond_resched().
4379 if (current->flags & PF_WQ_WORKER)
4380 schedule_timeout_uninterruptible(1);
4387 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4390 * It's possible that cpuset's mems_allowed and the nodemask from
4391 * mempolicy don't intersect. This should be normally dealt with by
4392 * policy_nodemask(), but it's possible to race with cpuset update in
4393 * such a way the check therein was true, and then it became false
4394 * before we got our cpuset_mems_cookie here.
4395 * This assumes that for all allocations, ac->nodemask can come only
4396 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4397 * when it does not intersect with the cpuset restrictions) or the
4398 * caller can deal with a violated nodemask.
4400 if (cpusets_enabled() && ac->nodemask &&
4401 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4402 ac->nodemask = NULL;
4407 * When updating a task's mems_allowed or mempolicy nodemask, it is
4408 * possible to race with parallel threads in such a way that our
4409 * allocation can fail while the mask is being updated. If we are about
4410 * to fail, check if the cpuset changed during allocation and if so,
4413 if (read_mems_allowed_retry(cpuset_mems_cookie))
4419 static inline struct page *
4420 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4421 struct alloc_context *ac)
4423 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4424 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4425 struct page *page = NULL;
4426 unsigned int alloc_flags;
4427 unsigned long did_some_progress;
4428 enum compact_priority compact_priority;
4429 enum compact_result compact_result;
4430 int compaction_retries;
4431 int no_progress_loops;
4432 unsigned int cpuset_mems_cookie;
4436 * We also sanity check to catch abuse of atomic reserves being used by
4437 * callers that are not in atomic context.
4439 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4440 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4441 gfp_mask &= ~__GFP_ATOMIC;
4444 compaction_retries = 0;
4445 no_progress_loops = 0;
4446 compact_priority = DEF_COMPACT_PRIORITY;
4447 cpuset_mems_cookie = read_mems_allowed_begin();
4450 * The fast path uses conservative alloc_flags to succeed only until
4451 * kswapd needs to be woken up, and to avoid the cost of setting up
4452 * alloc_flags precisely. So we do that now.
4454 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4457 * We need to recalculate the starting point for the zonelist iterator
4458 * because we might have used different nodemask in the fast path, or
4459 * there was a cpuset modification and we are retrying - otherwise we
4460 * could end up iterating over non-eligible zones endlessly.
4462 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4463 ac->high_zoneidx, ac->nodemask);
4464 if (!ac->preferred_zoneref->zone)
4467 if (alloc_flags & ALLOC_KSWAPD)
4468 wake_all_kswapds(order, gfp_mask, ac);
4471 * The adjusted alloc_flags might result in immediate success, so try
4474 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4479 * For costly allocations, try direct compaction first, as it's likely
4480 * that we have enough base pages and don't need to reclaim. For non-
4481 * movable high-order allocations, do that as well, as compaction will
4482 * try prevent permanent fragmentation by migrating from blocks of the
4484 * Don't try this for allocations that are allowed to ignore
4485 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4487 if (can_direct_reclaim &&
4489 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4490 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4491 page = __alloc_pages_direct_compact(gfp_mask, order,
4493 INIT_COMPACT_PRIORITY,
4498 if (order >= pageblock_order && (gfp_mask & __GFP_IO) &&
4499 !(gfp_mask & __GFP_RETRY_MAYFAIL)) {
4501 * If allocating entire pageblock(s) and compaction
4502 * failed because all zones are below low watermarks
4503 * or is prohibited because it recently failed at this
4504 * order, fail immediately unless the allocator has
4505 * requested compaction and reclaim retry.
4508 * - potentially very expensive because zones are far
4509 * below their low watermarks or this is part of very
4510 * bursty high order allocations,
4511 * - not guaranteed to help because isolate_freepages()
4512 * may not iterate over freed pages as part of its
4514 * - unlikely to make entire pageblocks free on its
4517 if (compact_result == COMPACT_SKIPPED ||
4518 compact_result == COMPACT_DEFERRED)
4523 * Checks for costly allocations with __GFP_NORETRY, which
4524 * includes THP page fault allocations
4526 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4528 * If compaction is deferred for high-order allocations,
4529 * it is because sync compaction recently failed. If
4530 * this is the case and the caller requested a THP
4531 * allocation, we do not want to heavily disrupt the
4532 * system, so we fail the allocation instead of entering
4535 if (compact_result == COMPACT_DEFERRED)
4539 * Looks like reclaim/compaction is worth trying, but
4540 * sync compaction could be very expensive, so keep
4541 * using async compaction.
4543 compact_priority = INIT_COMPACT_PRIORITY;
4548 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4549 if (alloc_flags & ALLOC_KSWAPD)
4550 wake_all_kswapds(order, gfp_mask, ac);
4552 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4554 alloc_flags = reserve_flags;
4557 * Reset the nodemask and zonelist iterators if memory policies can be
4558 * ignored. These allocations are high priority and system rather than
4561 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4562 ac->nodemask = NULL;
4563 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4564 ac->high_zoneidx, ac->nodemask);
4567 /* Attempt with potentially adjusted zonelist and alloc_flags */
4568 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4572 /* Caller is not willing to reclaim, we can't balance anything */
4573 if (!can_direct_reclaim)
4576 /* Avoid recursion of direct reclaim */
4577 if (current->flags & PF_MEMALLOC)
4580 /* Try direct reclaim and then allocating */
4581 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4582 &did_some_progress);
4586 /* Try direct compaction and then allocating */
4587 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4588 compact_priority, &compact_result);
4592 /* Do not loop if specifically requested */
4593 if (gfp_mask & __GFP_NORETRY)
4597 * Do not retry costly high order allocations unless they are
4598 * __GFP_RETRY_MAYFAIL
4600 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4603 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4604 did_some_progress > 0, &no_progress_loops))
4608 * It doesn't make any sense to retry for the compaction if the order-0
4609 * reclaim is not able to make any progress because the current
4610 * implementation of the compaction depends on the sufficient amount
4611 * of free memory (see __compaction_suitable)
4613 if (did_some_progress > 0 &&
4614 should_compact_retry(ac, order, alloc_flags,
4615 compact_result, &compact_priority,
4616 &compaction_retries))
4620 /* Deal with possible cpuset update races before we start OOM killing */
4621 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4624 /* Reclaim has failed us, start killing things */
4625 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4629 /* Avoid allocations with no watermarks from looping endlessly */
4630 if (tsk_is_oom_victim(current) &&
4631 (alloc_flags == ALLOC_OOM ||
4632 (gfp_mask & __GFP_NOMEMALLOC)))
4635 /* Retry as long as the OOM killer is making progress */
4636 if (did_some_progress) {
4637 no_progress_loops = 0;
4642 /* Deal with possible cpuset update races before we fail */
4643 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4647 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4650 if (gfp_mask & __GFP_NOFAIL) {
4652 * All existing users of the __GFP_NOFAIL are blockable, so warn
4653 * of any new users that actually require GFP_NOWAIT
4655 if (WARN_ON_ONCE(!can_direct_reclaim))
4659 * PF_MEMALLOC request from this context is rather bizarre
4660 * because we cannot reclaim anything and only can loop waiting
4661 * for somebody to do a work for us
4663 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4666 * non failing costly orders are a hard requirement which we
4667 * are not prepared for much so let's warn about these users
4668 * so that we can identify them and convert them to something
4671 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4674 * Help non-failing allocations by giving them access to memory
4675 * reserves but do not use ALLOC_NO_WATERMARKS because this
4676 * could deplete whole memory reserves which would just make
4677 * the situation worse
4679 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4687 warn_alloc(gfp_mask, ac->nodemask,
4688 "page allocation failure: order:%u", order);
4693 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4694 int preferred_nid, nodemask_t *nodemask,
4695 struct alloc_context *ac, gfp_t *alloc_mask,
4696 unsigned int *alloc_flags)
4698 ac->high_zoneidx = gfp_zone(gfp_mask);
4699 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4700 ac->nodemask = nodemask;
4701 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4703 if (cpusets_enabled()) {
4704 *alloc_mask |= __GFP_HARDWALL;
4706 ac->nodemask = &cpuset_current_mems_allowed;
4708 *alloc_flags |= ALLOC_CPUSET;
4711 fs_reclaim_acquire(gfp_mask);
4712 fs_reclaim_release(gfp_mask);
4714 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4716 if (should_fail_alloc_page(gfp_mask, order))
4719 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4720 *alloc_flags |= ALLOC_CMA;
4725 /* Determine whether to spread dirty pages and what the first usable zone */
4726 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4728 /* Dirty zone balancing only done in the fast path */
4729 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4732 * The preferred zone is used for statistics but crucially it is
4733 * also used as the starting point for the zonelist iterator. It
4734 * may get reset for allocations that ignore memory policies.
4736 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4737 ac->high_zoneidx, ac->nodemask);
4741 * This is the 'heart' of the zoned buddy allocator.
4744 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4745 nodemask_t *nodemask)
4748 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4749 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4750 struct alloc_context ac = { };
4753 * There are several places where we assume that the order value is sane
4754 * so bail out early if the request is out of bound.
4756 if (unlikely(order >= MAX_ORDER)) {
4757 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4761 gfp_mask &= gfp_allowed_mask;
4762 alloc_mask = gfp_mask;
4763 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4766 finalise_ac(gfp_mask, &ac);
4769 * Forbid the first pass from falling back to types that fragment
4770 * memory until all local zones are considered.
4772 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4774 /* First allocation attempt */
4775 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4780 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4781 * resp. GFP_NOIO which has to be inherited for all allocation requests
4782 * from a particular context which has been marked by
4783 * memalloc_no{fs,io}_{save,restore}.
4785 alloc_mask = current_gfp_context(gfp_mask);
4786 ac.spread_dirty_pages = false;
4789 * Restore the original nodemask if it was potentially replaced with
4790 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4792 if (unlikely(ac.nodemask != nodemask))
4793 ac.nodemask = nodemask;
4795 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4798 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4799 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4800 __free_pages(page, order);
4804 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4808 EXPORT_SYMBOL(__alloc_pages_nodemask);
4811 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4812 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4813 * you need to access high mem.
4815 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4819 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4822 return (unsigned long) page_address(page);
4824 EXPORT_SYMBOL(__get_free_pages);
4826 unsigned long get_zeroed_page(gfp_t gfp_mask)
4828 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4830 EXPORT_SYMBOL(get_zeroed_page);
4832 static inline void free_the_page(struct page *page, unsigned int order)
4834 if (order == 0) /* Via pcp? */
4835 free_unref_page(page);
4837 __free_pages_ok(page, order);
4840 void __free_pages(struct page *page, unsigned int order)
4842 if (put_page_testzero(page))
4843 free_the_page(page, order);
4845 EXPORT_SYMBOL(__free_pages);
4847 void free_pages(unsigned long addr, unsigned int order)
4850 VM_BUG_ON(!virt_addr_valid((void *)addr));
4851 __free_pages(virt_to_page((void *)addr), order);
4855 EXPORT_SYMBOL(free_pages);
4859 * An arbitrary-length arbitrary-offset area of memory which resides
4860 * within a 0 or higher order page. Multiple fragments within that page
4861 * are individually refcounted, in the page's reference counter.
4863 * The page_frag functions below provide a simple allocation framework for
4864 * page fragments. This is used by the network stack and network device
4865 * drivers to provide a backing region of memory for use as either an
4866 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4868 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4871 struct page *page = NULL;
4872 gfp_t gfp = gfp_mask;
4874 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4875 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4877 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4878 PAGE_FRAG_CACHE_MAX_ORDER);
4879 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4881 if (unlikely(!page))
4882 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4884 nc->va = page ? page_address(page) : NULL;
4889 void __page_frag_cache_drain(struct page *page, unsigned int count)
4891 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4893 if (page_ref_sub_and_test(page, count))
4894 free_the_page(page, compound_order(page));
4896 EXPORT_SYMBOL(__page_frag_cache_drain);
4898 void *page_frag_alloc(struct page_frag_cache *nc,
4899 unsigned int fragsz, gfp_t gfp_mask)
4901 unsigned int size = PAGE_SIZE;
4905 if (unlikely(!nc->va)) {
4907 page = __page_frag_cache_refill(nc, gfp_mask);
4911 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4912 /* if size can vary use size else just use PAGE_SIZE */
4915 /* Even if we own the page, we do not use atomic_set().
4916 * This would break get_page_unless_zero() users.
4918 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4920 /* reset page count bias and offset to start of new frag */
4921 nc->pfmemalloc = page_is_pfmemalloc(page);
4922 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4926 offset = nc->offset - fragsz;
4927 if (unlikely(offset < 0)) {
4928 page = virt_to_page(nc->va);
4930 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4933 if (unlikely(nc->pfmemalloc)) {
4934 free_the_page(page, compound_order(page));
4938 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4939 /* if size can vary use size else just use PAGE_SIZE */
4942 /* OK, page count is 0, we can safely set it */
4943 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4945 /* reset page count bias and offset to start of new frag */
4946 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4947 offset = size - fragsz;
4951 nc->offset = offset;
4953 return nc->va + offset;
4955 EXPORT_SYMBOL(page_frag_alloc);
4958 * Frees a page fragment allocated out of either a compound or order 0 page.
4960 void page_frag_free(void *addr)
4962 struct page *page = virt_to_head_page(addr);
4964 if (unlikely(put_page_testzero(page)))
4965 free_the_page(page, compound_order(page));
4967 EXPORT_SYMBOL(page_frag_free);
4969 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4973 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4974 unsigned long used = addr + PAGE_ALIGN(size);
4976 split_page(virt_to_page((void *)addr), order);
4977 while (used < alloc_end) {
4982 return (void *)addr;
4986 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4987 * @size: the number of bytes to allocate
4988 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4990 * This function is similar to alloc_pages(), except that it allocates the
4991 * minimum number of pages to satisfy the request. alloc_pages() can only
4992 * allocate memory in power-of-two pages.
4994 * This function is also limited by MAX_ORDER.
4996 * Memory allocated by this function must be released by free_pages_exact().
4998 * Return: pointer to the allocated area or %NULL in case of error.
5000 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5002 unsigned int order = get_order(size);
5005 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5006 gfp_mask &= ~__GFP_COMP;
5008 addr = __get_free_pages(gfp_mask, order);
5009 return make_alloc_exact(addr, order, size);
5011 EXPORT_SYMBOL(alloc_pages_exact);
5014 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5016 * @nid: the preferred node ID where memory should be allocated
5017 * @size: the number of bytes to allocate
5018 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5020 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5023 * Return: pointer to the allocated area or %NULL in case of error.
5025 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5027 unsigned int order = get_order(size);
5030 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5031 gfp_mask &= ~__GFP_COMP;
5033 p = alloc_pages_node(nid, gfp_mask, order);
5036 return make_alloc_exact((unsigned long)page_address(p), order, size);
5040 * free_pages_exact - release memory allocated via alloc_pages_exact()
5041 * @virt: the value returned by alloc_pages_exact.
5042 * @size: size of allocation, same value as passed to alloc_pages_exact().
5044 * Release the memory allocated by a previous call to alloc_pages_exact.
5046 void free_pages_exact(void *virt, size_t size)
5048 unsigned long addr = (unsigned long)virt;
5049 unsigned long end = addr + PAGE_ALIGN(size);
5051 while (addr < end) {
5056 EXPORT_SYMBOL(free_pages_exact);
5059 * nr_free_zone_pages - count number of pages beyond high watermark
5060 * @offset: The zone index of the highest zone
5062 * nr_free_zone_pages() counts the number of pages which are beyond the
5063 * high watermark within all zones at or below a given zone index. For each
5064 * zone, the number of pages is calculated as:
5066 * nr_free_zone_pages = managed_pages - high_pages
5068 * Return: number of pages beyond high watermark.
5070 static unsigned long nr_free_zone_pages(int offset)
5075 /* Just pick one node, since fallback list is circular */
5076 unsigned long sum = 0;
5078 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5080 for_each_zone_zonelist(zone, z, zonelist, offset) {
5081 unsigned long size = zone_managed_pages(zone);
5082 unsigned long high = high_wmark_pages(zone);
5091 * nr_free_buffer_pages - count number of pages beyond high watermark
5093 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5094 * watermark within ZONE_DMA and ZONE_NORMAL.
5096 * Return: number of pages beyond high watermark within ZONE_DMA and
5099 unsigned long nr_free_buffer_pages(void)
5101 return nr_free_zone_pages(gfp_zone(GFP_USER));
5103 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5106 * nr_free_pagecache_pages - count number of pages beyond high watermark
5108 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5109 * high watermark within all zones.
5111 * Return: number of pages beyond high watermark within all zones.
5113 unsigned long nr_free_pagecache_pages(void)
5115 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5118 static inline void show_node(struct zone *zone)
5120 if (IS_ENABLED(CONFIG_NUMA))
5121 printk("Node %d ", zone_to_nid(zone));
5124 long si_mem_available(void)
5127 unsigned long pagecache;
5128 unsigned long wmark_low = 0;
5129 unsigned long pages[NR_LRU_LISTS];
5130 unsigned long reclaimable;
5134 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5135 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5138 wmark_low += low_wmark_pages(zone);
5141 * Estimate the amount of memory available for userspace allocations,
5142 * without causing swapping.
5144 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5147 * Not all the page cache can be freed, otherwise the system will
5148 * start swapping. Assume at least half of the page cache, or the
5149 * low watermark worth of cache, needs to stay.
5151 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5152 pagecache -= min(pagecache / 2, wmark_low);
5153 available += pagecache;
5156 * Part of the reclaimable slab and other kernel memory consists of
5157 * items that are in use, and cannot be freed. Cap this estimate at the
5160 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5161 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5162 available += reclaimable - min(reclaimable / 2, wmark_low);
5168 EXPORT_SYMBOL_GPL(si_mem_available);
5170 void si_meminfo(struct sysinfo *val)
5172 val->totalram = totalram_pages();
5173 val->sharedram = global_node_page_state(NR_SHMEM);
5174 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5175 val->bufferram = nr_blockdev_pages();
5176 val->totalhigh = totalhigh_pages();
5177 val->freehigh = nr_free_highpages();
5178 val->mem_unit = PAGE_SIZE;
5181 EXPORT_SYMBOL(si_meminfo);
5184 void si_meminfo_node(struct sysinfo *val, int nid)
5186 int zone_type; /* needs to be signed */
5187 unsigned long managed_pages = 0;
5188 unsigned long managed_highpages = 0;
5189 unsigned long free_highpages = 0;
5190 pg_data_t *pgdat = NODE_DATA(nid);
5192 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5193 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5194 val->totalram = managed_pages;
5195 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5196 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5197 #ifdef CONFIG_HIGHMEM
5198 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5199 struct zone *zone = &pgdat->node_zones[zone_type];
5201 if (is_highmem(zone)) {
5202 managed_highpages += zone_managed_pages(zone);
5203 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5206 val->totalhigh = managed_highpages;
5207 val->freehigh = free_highpages;
5209 val->totalhigh = managed_highpages;
5210 val->freehigh = free_highpages;
5212 val->mem_unit = PAGE_SIZE;
5217 * Determine whether the node should be displayed or not, depending on whether
5218 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5220 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5222 if (!(flags & SHOW_MEM_FILTER_NODES))
5226 * no node mask - aka implicit memory numa policy. Do not bother with
5227 * the synchronization - read_mems_allowed_begin - because we do not
5228 * have to be precise here.
5231 nodemask = &cpuset_current_mems_allowed;
5233 return !node_isset(nid, *nodemask);
5236 #define K(x) ((x) << (PAGE_SHIFT-10))
5238 static void show_migration_types(unsigned char type)
5240 static const char types[MIGRATE_TYPES] = {
5241 [MIGRATE_UNMOVABLE] = 'U',
5242 [MIGRATE_MOVABLE] = 'M',
5243 [MIGRATE_RECLAIMABLE] = 'E',
5244 [MIGRATE_HIGHATOMIC] = 'H',
5246 [MIGRATE_CMA] = 'C',
5248 #ifdef CONFIG_MEMORY_ISOLATION
5249 [MIGRATE_ISOLATE] = 'I',
5252 char tmp[MIGRATE_TYPES + 1];
5256 for (i = 0; i < MIGRATE_TYPES; i++) {
5257 if (type & (1 << i))
5262 printk(KERN_CONT "(%s) ", tmp);
5266 * Show free area list (used inside shift_scroll-lock stuff)
5267 * We also calculate the percentage fragmentation. We do this by counting the
5268 * memory on each free list with the exception of the first item on the list.
5271 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5274 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5276 unsigned long free_pcp = 0;
5281 for_each_populated_zone(zone) {
5282 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5285 for_each_online_cpu(cpu)
5286 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5289 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5290 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5291 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5292 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5293 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5294 " free:%lu free_pcp:%lu free_cma:%lu\n",
5295 global_node_page_state(NR_ACTIVE_ANON),
5296 global_node_page_state(NR_INACTIVE_ANON),
5297 global_node_page_state(NR_ISOLATED_ANON),
5298 global_node_page_state(NR_ACTIVE_FILE),
5299 global_node_page_state(NR_INACTIVE_FILE),
5300 global_node_page_state(NR_ISOLATED_FILE),
5301 global_node_page_state(NR_UNEVICTABLE),
5302 global_node_page_state(NR_FILE_DIRTY),
5303 global_node_page_state(NR_WRITEBACK),
5304 global_node_page_state(NR_UNSTABLE_NFS),
5305 global_node_page_state(NR_SLAB_RECLAIMABLE),
5306 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5307 global_node_page_state(NR_FILE_MAPPED),
5308 global_node_page_state(NR_SHMEM),
5309 global_zone_page_state(NR_PAGETABLE),
5310 global_zone_page_state(NR_BOUNCE),
5311 global_zone_page_state(NR_FREE_PAGES),
5313 global_zone_page_state(NR_FREE_CMA_PAGES));
5315 for_each_online_pgdat(pgdat) {
5316 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5320 " active_anon:%lukB"
5321 " inactive_anon:%lukB"
5322 " active_file:%lukB"
5323 " inactive_file:%lukB"
5324 " unevictable:%lukB"
5325 " isolated(anon):%lukB"
5326 " isolated(file):%lukB"
5331 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5333 " shmem_pmdmapped: %lukB"
5336 " writeback_tmp:%lukB"
5338 " all_unreclaimable? %s"
5341 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5342 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5343 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5344 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5345 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5346 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5347 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5348 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5349 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5350 K(node_page_state(pgdat, NR_WRITEBACK)),
5351 K(node_page_state(pgdat, NR_SHMEM)),
5352 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5353 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5354 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5356 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5358 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5359 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5360 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5364 for_each_populated_zone(zone) {
5367 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5371 for_each_online_cpu(cpu)
5372 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5381 " active_anon:%lukB"
5382 " inactive_anon:%lukB"
5383 " active_file:%lukB"
5384 " inactive_file:%lukB"
5385 " unevictable:%lukB"
5386 " writepending:%lukB"
5390 " kernel_stack:%lukB"
5398 K(zone_page_state(zone, NR_FREE_PAGES)),
5399 K(min_wmark_pages(zone)),
5400 K(low_wmark_pages(zone)),
5401 K(high_wmark_pages(zone)),
5402 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5403 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5404 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5405 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5406 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5407 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5408 K(zone->present_pages),
5409 K(zone_managed_pages(zone)),
5410 K(zone_page_state(zone, NR_MLOCK)),
5411 zone_page_state(zone, NR_KERNEL_STACK_KB),
5412 K(zone_page_state(zone, NR_PAGETABLE)),
5413 K(zone_page_state(zone, NR_BOUNCE)),
5415 K(this_cpu_read(zone->pageset->pcp.count)),
5416 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5417 printk("lowmem_reserve[]:");
5418 for (i = 0; i < MAX_NR_ZONES; i++)
5419 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5420 printk(KERN_CONT "\n");
5423 for_each_populated_zone(zone) {
5425 unsigned long nr[MAX_ORDER], flags, total = 0;
5426 unsigned char types[MAX_ORDER];
5428 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5431 printk(KERN_CONT "%s: ", zone->name);
5433 spin_lock_irqsave(&zone->lock, flags);
5434 for (order = 0; order < MAX_ORDER; order++) {
5435 struct free_area *area = &zone->free_area[order];
5438 nr[order] = area->nr_free;
5439 total += nr[order] << order;
5442 for (type = 0; type < MIGRATE_TYPES; type++) {
5443 if (!free_area_empty(area, type))
5444 types[order] |= 1 << type;
5447 spin_unlock_irqrestore(&zone->lock, flags);
5448 for (order = 0; order < MAX_ORDER; order++) {
5449 printk(KERN_CONT "%lu*%lukB ",
5450 nr[order], K(1UL) << order);
5452 show_migration_types(types[order]);
5454 printk(KERN_CONT "= %lukB\n", K(total));
5457 hugetlb_show_meminfo();
5459 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5461 show_swap_cache_info();
5464 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5466 zoneref->zone = zone;
5467 zoneref->zone_idx = zone_idx(zone);
5471 * Builds allocation fallback zone lists.
5473 * Add all populated zones of a node to the zonelist.
5475 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5478 enum zone_type zone_type = MAX_NR_ZONES;
5483 zone = pgdat->node_zones + zone_type;
5484 if (populated_zone(zone)) {
5485 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5486 check_highest_zone(zone_type);
5488 } while (zone_type);
5495 static int __parse_numa_zonelist_order(char *s)
5498 * We used to support different zonlists modes but they turned
5499 * out to be just not useful. Let's keep the warning in place
5500 * if somebody still use the cmd line parameter so that we do
5501 * not fail it silently
5503 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5504 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5510 static __init int setup_numa_zonelist_order(char *s)
5515 return __parse_numa_zonelist_order(s);
5517 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5519 char numa_zonelist_order[] = "Node";
5522 * sysctl handler for numa_zonelist_order
5524 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5525 void __user *buffer, size_t *length,
5532 return proc_dostring(table, write, buffer, length, ppos);
5533 str = memdup_user_nul(buffer, 16);
5535 return PTR_ERR(str);
5537 ret = __parse_numa_zonelist_order(str);
5543 #define MAX_NODE_LOAD (nr_online_nodes)
5544 static int node_load[MAX_NUMNODES];
5547 * find_next_best_node - find the next node that should appear in a given node's fallback list
5548 * @node: node whose fallback list we're appending
5549 * @used_node_mask: nodemask_t of already used nodes
5551 * We use a number of factors to determine which is the next node that should
5552 * appear on a given node's fallback list. The node should not have appeared
5553 * already in @node's fallback list, and it should be the next closest node
5554 * according to the distance array (which contains arbitrary distance values
5555 * from each node to each node in the system), and should also prefer nodes
5556 * with no CPUs, since presumably they'll have very little allocation pressure
5557 * on them otherwise.
5559 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5561 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5564 int min_val = INT_MAX;
5565 int best_node = NUMA_NO_NODE;
5566 const struct cpumask *tmp = cpumask_of_node(0);
5568 /* Use the local node if we haven't already */
5569 if (!node_isset(node, *used_node_mask)) {
5570 node_set(node, *used_node_mask);
5574 for_each_node_state(n, N_MEMORY) {
5576 /* Don't want a node to appear more than once */
5577 if (node_isset(n, *used_node_mask))
5580 /* Use the distance array to find the distance */
5581 val = node_distance(node, n);
5583 /* Penalize nodes under us ("prefer the next node") */
5586 /* Give preference to headless and unused nodes */
5587 tmp = cpumask_of_node(n);
5588 if (!cpumask_empty(tmp))
5589 val += PENALTY_FOR_NODE_WITH_CPUS;
5591 /* Slight preference for less loaded node */
5592 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5593 val += node_load[n];
5595 if (val < min_val) {
5602 node_set(best_node, *used_node_mask);
5609 * Build zonelists ordered by node and zones within node.
5610 * This results in maximum locality--normal zone overflows into local
5611 * DMA zone, if any--but risks exhausting DMA zone.
5613 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5616 struct zoneref *zonerefs;
5619 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5621 for (i = 0; i < nr_nodes; i++) {
5624 pg_data_t *node = NODE_DATA(node_order[i]);
5626 nr_zones = build_zonerefs_node(node, zonerefs);
5627 zonerefs += nr_zones;
5629 zonerefs->zone = NULL;
5630 zonerefs->zone_idx = 0;
5634 * Build gfp_thisnode zonelists
5636 static void build_thisnode_zonelists(pg_data_t *pgdat)
5638 struct zoneref *zonerefs;
5641 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5642 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5643 zonerefs += nr_zones;
5644 zonerefs->zone = NULL;
5645 zonerefs->zone_idx = 0;
5649 * Build zonelists ordered by zone and nodes within zones.
5650 * This results in conserving DMA zone[s] until all Normal memory is
5651 * exhausted, but results in overflowing to remote node while memory
5652 * may still exist in local DMA zone.
5655 static void build_zonelists(pg_data_t *pgdat)
5657 static int node_order[MAX_NUMNODES];
5658 int node, load, nr_nodes = 0;
5659 nodemask_t used_mask;
5660 int local_node, prev_node;
5662 /* NUMA-aware ordering of nodes */
5663 local_node = pgdat->node_id;
5664 load = nr_online_nodes;
5665 prev_node = local_node;
5666 nodes_clear(used_mask);
5668 memset(node_order, 0, sizeof(node_order));
5669 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5671 * We don't want to pressure a particular node.
5672 * So adding penalty to the first node in same
5673 * distance group to make it round-robin.
5675 if (node_distance(local_node, node) !=
5676 node_distance(local_node, prev_node))
5677 node_load[node] = load;
5679 node_order[nr_nodes++] = node;
5684 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5685 build_thisnode_zonelists(pgdat);
5688 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5690 * Return node id of node used for "local" allocations.
5691 * I.e., first node id of first zone in arg node's generic zonelist.
5692 * Used for initializing percpu 'numa_mem', which is used primarily
5693 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5695 int local_memory_node(int node)
5699 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5700 gfp_zone(GFP_KERNEL),
5702 return zone_to_nid(z->zone);
5706 static void setup_min_unmapped_ratio(void);
5707 static void setup_min_slab_ratio(void);
5708 #else /* CONFIG_NUMA */
5710 static void build_zonelists(pg_data_t *pgdat)
5712 int node, local_node;
5713 struct zoneref *zonerefs;
5716 local_node = pgdat->node_id;
5718 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5719 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5720 zonerefs += nr_zones;
5723 * Now we build the zonelist so that it contains the zones
5724 * of all the other nodes.
5725 * We don't want to pressure a particular node, so when
5726 * building the zones for node N, we make sure that the
5727 * zones coming right after the local ones are those from
5728 * node N+1 (modulo N)
5730 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5731 if (!node_online(node))
5733 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5734 zonerefs += nr_zones;
5736 for (node = 0; node < local_node; node++) {
5737 if (!node_online(node))
5739 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5740 zonerefs += nr_zones;
5743 zonerefs->zone = NULL;
5744 zonerefs->zone_idx = 0;
5747 #endif /* CONFIG_NUMA */
5750 * Boot pageset table. One per cpu which is going to be used for all
5751 * zones and all nodes. The parameters will be set in such a way
5752 * that an item put on a list will immediately be handed over to
5753 * the buddy list. This is safe since pageset manipulation is done
5754 * with interrupts disabled.
5756 * The boot_pagesets must be kept even after bootup is complete for
5757 * unused processors and/or zones. They do play a role for bootstrapping
5758 * hotplugged processors.
5760 * zoneinfo_show() and maybe other functions do
5761 * not check if the processor is online before following the pageset pointer.
5762 * Other parts of the kernel may not check if the zone is available.
5764 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5765 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5766 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5768 static void __build_all_zonelists(void *data)
5771 int __maybe_unused cpu;
5772 pg_data_t *self = data;
5773 static DEFINE_SPINLOCK(lock);
5778 memset(node_load, 0, sizeof(node_load));
5782 * This node is hotadded and no memory is yet present. So just
5783 * building zonelists is fine - no need to touch other nodes.
5785 if (self && !node_online(self->node_id)) {
5786 build_zonelists(self);
5788 for_each_online_node(nid) {
5789 pg_data_t *pgdat = NODE_DATA(nid);
5791 build_zonelists(pgdat);
5794 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5796 * We now know the "local memory node" for each node--
5797 * i.e., the node of the first zone in the generic zonelist.
5798 * Set up numa_mem percpu variable for on-line cpus. During
5799 * boot, only the boot cpu should be on-line; we'll init the
5800 * secondary cpus' numa_mem as they come on-line. During
5801 * node/memory hotplug, we'll fixup all on-line cpus.
5803 for_each_online_cpu(cpu)
5804 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5811 static noinline void __init
5812 build_all_zonelists_init(void)
5816 __build_all_zonelists(NULL);
5819 * Initialize the boot_pagesets that are going to be used
5820 * for bootstrapping processors. The real pagesets for
5821 * each zone will be allocated later when the per cpu
5822 * allocator is available.
5824 * boot_pagesets are used also for bootstrapping offline
5825 * cpus if the system is already booted because the pagesets
5826 * are needed to initialize allocators on a specific cpu too.
5827 * F.e. the percpu allocator needs the page allocator which
5828 * needs the percpu allocator in order to allocate its pagesets
5829 * (a chicken-egg dilemma).
5831 for_each_possible_cpu(cpu)
5832 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5834 mminit_verify_zonelist();
5835 cpuset_init_current_mems_allowed();
5839 * unless system_state == SYSTEM_BOOTING.
5841 * __ref due to call of __init annotated helper build_all_zonelists_init
5842 * [protected by SYSTEM_BOOTING].
5844 void __ref build_all_zonelists(pg_data_t *pgdat)
5846 if (system_state == SYSTEM_BOOTING) {
5847 build_all_zonelists_init();
5849 __build_all_zonelists(pgdat);
5850 /* cpuset refresh routine should be here */
5852 vm_total_pages = nr_free_pagecache_pages();
5854 * Disable grouping by mobility if the number of pages in the
5855 * system is too low to allow the mechanism to work. It would be
5856 * more accurate, but expensive to check per-zone. This check is
5857 * made on memory-hotadd so a system can start with mobility
5858 * disabled and enable it later
5860 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5861 page_group_by_mobility_disabled = 1;
5863 page_group_by_mobility_disabled = 0;
5865 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5867 page_group_by_mobility_disabled ? "off" : "on",
5870 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5874 /* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5875 static bool __meminit
5876 overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5878 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5879 static struct memblock_region *r;
5881 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5882 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5883 for_each_memblock(memory, r) {
5884 if (*pfn < memblock_region_memory_end_pfn(r))
5888 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5889 memblock_is_mirror(r)) {
5890 *pfn = memblock_region_memory_end_pfn(r);
5899 * Initially all pages are reserved - free ones are freed
5900 * up by memblock_free_all() once the early boot process is
5901 * done. Non-atomic initialization, single-pass.
5903 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5904 unsigned long start_pfn, enum meminit_context context,
5905 struct vmem_altmap *altmap)
5907 unsigned long pfn, end_pfn = start_pfn + size;
5910 if (highest_memmap_pfn < end_pfn - 1)
5911 highest_memmap_pfn = end_pfn - 1;
5913 #ifdef CONFIG_ZONE_DEVICE
5915 * Honor reservation requested by the driver for this ZONE_DEVICE
5916 * memory. We limit the total number of pages to initialize to just
5917 * those that might contain the memory mapping. We will defer the
5918 * ZONE_DEVICE page initialization until after we have released
5921 if (zone == ZONE_DEVICE) {
5925 if (start_pfn == altmap->base_pfn)
5926 start_pfn += altmap->reserve;
5927 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5931 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5933 * There can be holes in boot-time mem_map[]s handed to this
5934 * function. They do not exist on hotplugged memory.
5936 if (context == MEMINIT_EARLY) {
5937 if (!early_pfn_valid(pfn))
5939 if (!early_pfn_in_nid(pfn, nid))
5941 if (overlap_memmap_init(zone, &pfn))
5943 if (defer_init(nid, pfn, end_pfn))
5947 page = pfn_to_page(pfn);
5948 __init_single_page(page, pfn, zone, nid);
5949 if (context == MEMINIT_HOTPLUG)
5950 __SetPageReserved(page);
5953 * Mark the block movable so that blocks are reserved for
5954 * movable at startup. This will force kernel allocations
5955 * to reserve their blocks rather than leaking throughout
5956 * the address space during boot when many long-lived
5957 * kernel allocations are made.
5959 * bitmap is created for zone's valid pfn range. but memmap
5960 * can be created for invalid pages (for alignment)
5961 * check here not to call set_pageblock_migratetype() against
5964 if (!(pfn & (pageblock_nr_pages - 1))) {
5965 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5971 #ifdef CONFIG_ZONE_DEVICE
5972 void __ref memmap_init_zone_device(struct zone *zone,
5973 unsigned long start_pfn,
5975 struct dev_pagemap *pgmap)
5977 unsigned long pfn, end_pfn = start_pfn + size;
5978 struct pglist_data *pgdat = zone->zone_pgdat;
5979 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
5980 unsigned long zone_idx = zone_idx(zone);
5981 unsigned long start = jiffies;
5982 int nid = pgdat->node_id;
5984 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
5988 * The call to memmap_init_zone should have already taken care
5989 * of the pages reserved for the memmap, so we can just jump to
5990 * the end of that region and start processing the device pages.
5993 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5994 size = end_pfn - start_pfn;
5997 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5998 struct page *page = pfn_to_page(pfn);
6000 __init_single_page(page, pfn, zone_idx, nid);
6003 * Mark page reserved as it will need to wait for onlining
6004 * phase for it to be fully associated with a zone.
6006 * We can use the non-atomic __set_bit operation for setting
6007 * the flag as we are still initializing the pages.
6009 __SetPageReserved(page);
6012 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6013 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6014 * ever freed or placed on a driver-private list.
6016 page->pgmap = pgmap;
6017 page->zone_device_data = NULL;
6020 * Mark the block movable so that blocks are reserved for
6021 * movable at startup. This will force kernel allocations
6022 * to reserve their blocks rather than leaking throughout
6023 * the address space during boot when many long-lived
6024 * kernel allocations are made.
6026 * bitmap is created for zone's valid pfn range. but memmap
6027 * can be created for invalid pages (for alignment)
6028 * check here not to call set_pageblock_migratetype() against
6031 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6032 * because this is done early in section_activate()
6034 if (!(pfn & (pageblock_nr_pages - 1))) {
6035 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6040 pr_info("%s initialised %lu pages in %ums\n", __func__,
6041 size, jiffies_to_msecs(jiffies - start));
6045 static void __meminit zone_init_free_lists(struct zone *zone)
6047 unsigned int order, t;
6048 for_each_migratetype_order(order, t) {
6049 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6050 zone->free_area[order].nr_free = 0;
6054 void __meminit __weak memmap_init(unsigned long size, int nid,
6055 unsigned long zone, unsigned long start_pfn)
6057 memmap_init_zone(size, nid, zone, start_pfn, MEMINIT_EARLY, NULL);
6060 static int zone_batchsize(struct zone *zone)
6066 * The per-cpu-pages pools are set to around 1000th of the
6069 batch = zone_managed_pages(zone) / 1024;
6070 /* But no more than a meg. */
6071 if (batch * PAGE_SIZE > 1024 * 1024)
6072 batch = (1024 * 1024) / PAGE_SIZE;
6073 batch /= 4; /* We effectively *= 4 below */
6078 * Clamp the batch to a 2^n - 1 value. Having a power
6079 * of 2 value was found to be more likely to have
6080 * suboptimal cache aliasing properties in some cases.
6082 * For example if 2 tasks are alternately allocating
6083 * batches of pages, one task can end up with a lot
6084 * of pages of one half of the possible page colors
6085 * and the other with pages of the other colors.
6087 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6092 /* The deferral and batching of frees should be suppressed under NOMMU
6095 * The problem is that NOMMU needs to be able to allocate large chunks
6096 * of contiguous memory as there's no hardware page translation to
6097 * assemble apparent contiguous memory from discontiguous pages.
6099 * Queueing large contiguous runs of pages for batching, however,
6100 * causes the pages to actually be freed in smaller chunks. As there
6101 * can be a significant delay between the individual batches being
6102 * recycled, this leads to the once large chunks of space being
6103 * fragmented and becoming unavailable for high-order allocations.
6110 * pcp->high and pcp->batch values are related and dependent on one another:
6111 * ->batch must never be higher then ->high.
6112 * The following function updates them in a safe manner without read side
6115 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6116 * those fields changing asynchronously (acording the the above rule).
6118 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6119 * outside of boot time (or some other assurance that no concurrent updaters
6122 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6123 unsigned long batch)
6125 /* start with a fail safe value for batch */
6129 /* Update high, then batch, in order */
6136 /* a companion to pageset_set_high() */
6137 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6139 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6142 static void pageset_init(struct per_cpu_pageset *p)
6144 struct per_cpu_pages *pcp;
6147 memset(p, 0, sizeof(*p));
6150 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6151 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6154 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6157 pageset_set_batch(p, batch);
6161 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6162 * to the value high for the pageset p.
6164 static void pageset_set_high(struct per_cpu_pageset *p,
6167 unsigned long batch = max(1UL, high / 4);
6168 if ((high / 4) > (PAGE_SHIFT * 8))
6169 batch = PAGE_SHIFT * 8;
6171 pageset_update(&p->pcp, high, batch);
6174 static void pageset_set_high_and_batch(struct zone *zone,
6175 struct per_cpu_pageset *pcp)
6177 if (percpu_pagelist_fraction)
6178 pageset_set_high(pcp,
6179 (zone_managed_pages(zone) /
6180 percpu_pagelist_fraction));
6182 pageset_set_batch(pcp, zone_batchsize(zone));
6185 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6187 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6190 pageset_set_high_and_batch(zone, pcp);
6193 void __meminit setup_zone_pageset(struct zone *zone)
6196 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6197 for_each_possible_cpu(cpu)
6198 zone_pageset_init(zone, cpu);
6202 * Allocate per cpu pagesets and initialize them.
6203 * Before this call only boot pagesets were available.
6205 void __init setup_per_cpu_pageset(void)
6207 struct pglist_data *pgdat;
6210 for_each_populated_zone(zone)
6211 setup_zone_pageset(zone);
6213 for_each_online_pgdat(pgdat)
6214 pgdat->per_cpu_nodestats =
6215 alloc_percpu(struct per_cpu_nodestat);
6218 static __meminit void zone_pcp_init(struct zone *zone)
6221 * per cpu subsystem is not up at this point. The following code
6222 * relies on the ability of the linker to provide the
6223 * offset of a (static) per cpu variable into the per cpu area.
6225 zone->pageset = &boot_pageset;
6227 if (populated_zone(zone))
6228 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6229 zone->name, zone->present_pages,
6230 zone_batchsize(zone));
6233 void __meminit init_currently_empty_zone(struct zone *zone,
6234 unsigned long zone_start_pfn,
6237 struct pglist_data *pgdat = zone->zone_pgdat;
6238 int zone_idx = zone_idx(zone) + 1;
6240 if (zone_idx > pgdat->nr_zones)
6241 pgdat->nr_zones = zone_idx;
6243 zone->zone_start_pfn = zone_start_pfn;
6245 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6246 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6248 (unsigned long)zone_idx(zone),
6249 zone_start_pfn, (zone_start_pfn + size));
6251 zone_init_free_lists(zone);
6252 zone->initialized = 1;
6255 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6256 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6259 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6261 int __meminit __early_pfn_to_nid(unsigned long pfn,
6262 struct mminit_pfnnid_cache *state)
6264 unsigned long start_pfn, end_pfn;
6267 if (state->last_start <= pfn && pfn < state->last_end)
6268 return state->last_nid;
6270 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6271 if (nid != NUMA_NO_NODE) {
6272 state->last_start = start_pfn;
6273 state->last_end = end_pfn;
6274 state->last_nid = nid;
6279 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6282 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6283 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6284 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6286 * If an architecture guarantees that all ranges registered contain no holes
6287 * and may be freed, this this function may be used instead of calling
6288 * memblock_free_early_nid() manually.
6290 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6292 unsigned long start_pfn, end_pfn;
6295 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6296 start_pfn = min(start_pfn, max_low_pfn);
6297 end_pfn = min(end_pfn, max_low_pfn);
6299 if (start_pfn < end_pfn)
6300 memblock_free_early_nid(PFN_PHYS(start_pfn),
6301 (end_pfn - start_pfn) << PAGE_SHIFT,
6307 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6308 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6310 * If an architecture guarantees that all ranges registered contain no holes and may
6311 * be freed, this function may be used instead of calling memory_present() manually.
6313 void __init sparse_memory_present_with_active_regions(int nid)
6315 unsigned long start_pfn, end_pfn;
6318 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6319 memory_present(this_nid, start_pfn, end_pfn);
6323 * get_pfn_range_for_nid - Return the start and end page frames for a node
6324 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6325 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6326 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6328 * It returns the start and end page frame of a node based on information
6329 * provided by memblock_set_node(). If called for a node
6330 * with no available memory, a warning is printed and the start and end
6333 void __init get_pfn_range_for_nid(unsigned int nid,
6334 unsigned long *start_pfn, unsigned long *end_pfn)
6336 unsigned long this_start_pfn, this_end_pfn;
6342 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6343 *start_pfn = min(*start_pfn, this_start_pfn);
6344 *end_pfn = max(*end_pfn, this_end_pfn);
6347 if (*start_pfn == -1UL)
6352 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6353 * assumption is made that zones within a node are ordered in monotonic
6354 * increasing memory addresses so that the "highest" populated zone is used
6356 static void __init find_usable_zone_for_movable(void)
6359 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6360 if (zone_index == ZONE_MOVABLE)
6363 if (arch_zone_highest_possible_pfn[zone_index] >
6364 arch_zone_lowest_possible_pfn[zone_index])
6368 VM_BUG_ON(zone_index == -1);
6369 movable_zone = zone_index;
6373 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6374 * because it is sized independent of architecture. Unlike the other zones,
6375 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6376 * in each node depending on the size of each node and how evenly kernelcore
6377 * is distributed. This helper function adjusts the zone ranges
6378 * provided by the architecture for a given node by using the end of the
6379 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6380 * zones within a node are in order of monotonic increases memory addresses
6382 static void __init adjust_zone_range_for_zone_movable(int nid,
6383 unsigned long zone_type,
6384 unsigned long node_start_pfn,
6385 unsigned long node_end_pfn,
6386 unsigned long *zone_start_pfn,
6387 unsigned long *zone_end_pfn)
6389 /* Only adjust if ZONE_MOVABLE is on this node */
6390 if (zone_movable_pfn[nid]) {
6391 /* Size ZONE_MOVABLE */
6392 if (zone_type == ZONE_MOVABLE) {
6393 *zone_start_pfn = zone_movable_pfn[nid];
6394 *zone_end_pfn = min(node_end_pfn,
6395 arch_zone_highest_possible_pfn[movable_zone]);
6397 /* Adjust for ZONE_MOVABLE starting within this range */
6398 } else if (!mirrored_kernelcore &&
6399 *zone_start_pfn < zone_movable_pfn[nid] &&
6400 *zone_end_pfn > zone_movable_pfn[nid]) {
6401 *zone_end_pfn = zone_movable_pfn[nid];
6403 /* Check if this whole range is within ZONE_MOVABLE */
6404 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6405 *zone_start_pfn = *zone_end_pfn;
6410 * Return the number of pages a zone spans in a node, including holes
6411 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6413 static unsigned long __init zone_spanned_pages_in_node(int nid,
6414 unsigned long zone_type,
6415 unsigned long node_start_pfn,
6416 unsigned long node_end_pfn,
6417 unsigned long *zone_start_pfn,
6418 unsigned long *zone_end_pfn,
6419 unsigned long *ignored)
6421 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6422 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6423 /* When hotadd a new node from cpu_up(), the node should be empty */
6424 if (!node_start_pfn && !node_end_pfn)
6427 /* Get the start and end of the zone */
6428 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6429 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6430 adjust_zone_range_for_zone_movable(nid, zone_type,
6431 node_start_pfn, node_end_pfn,
6432 zone_start_pfn, zone_end_pfn);
6434 /* Check that this node has pages within the zone's required range */
6435 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6438 /* Move the zone boundaries inside the node if necessary */
6439 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6440 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6442 /* Return the spanned pages */
6443 return *zone_end_pfn - *zone_start_pfn;
6447 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6448 * then all holes in the requested range will be accounted for.
6450 unsigned long __init __absent_pages_in_range(int nid,
6451 unsigned long range_start_pfn,
6452 unsigned long range_end_pfn)
6454 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6455 unsigned long start_pfn, end_pfn;
6458 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6459 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6460 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6461 nr_absent -= end_pfn - start_pfn;
6467 * absent_pages_in_range - Return number of page frames in holes within a range
6468 * @start_pfn: The start PFN to start searching for holes
6469 * @end_pfn: The end PFN to stop searching for holes
6471 * Return: the number of pages frames in memory holes within a range.
6473 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6474 unsigned long end_pfn)
6476 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6479 /* Return the number of page frames in holes in a zone on a node */
6480 static unsigned long __init zone_absent_pages_in_node(int nid,
6481 unsigned long zone_type,
6482 unsigned long node_start_pfn,
6483 unsigned long node_end_pfn,
6484 unsigned long *ignored)
6486 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6487 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6488 unsigned long zone_start_pfn, zone_end_pfn;
6489 unsigned long nr_absent;
6491 /* When hotadd a new node from cpu_up(), the node should be empty */
6492 if (!node_start_pfn && !node_end_pfn)
6495 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6496 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6498 adjust_zone_range_for_zone_movable(nid, zone_type,
6499 node_start_pfn, node_end_pfn,
6500 &zone_start_pfn, &zone_end_pfn);
6501 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6504 * ZONE_MOVABLE handling.
6505 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6508 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6509 unsigned long start_pfn, end_pfn;
6510 struct memblock_region *r;
6512 for_each_memblock(memory, r) {
6513 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6514 zone_start_pfn, zone_end_pfn);
6515 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6516 zone_start_pfn, zone_end_pfn);
6518 if (zone_type == ZONE_MOVABLE &&
6519 memblock_is_mirror(r))
6520 nr_absent += end_pfn - start_pfn;
6522 if (zone_type == ZONE_NORMAL &&
6523 !memblock_is_mirror(r))
6524 nr_absent += end_pfn - start_pfn;
6531 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6532 static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6533 unsigned long zone_type,
6534 unsigned long node_start_pfn,
6535 unsigned long node_end_pfn,
6536 unsigned long *zone_start_pfn,
6537 unsigned long *zone_end_pfn,
6538 unsigned long *zones_size)
6542 *zone_start_pfn = node_start_pfn;
6543 for (zone = 0; zone < zone_type; zone++)
6544 *zone_start_pfn += zones_size[zone];
6546 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6548 return zones_size[zone_type];
6551 static inline unsigned long __init zone_absent_pages_in_node(int nid,
6552 unsigned long zone_type,
6553 unsigned long node_start_pfn,
6554 unsigned long node_end_pfn,
6555 unsigned long *zholes_size)
6560 return zholes_size[zone_type];
6563 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6565 static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6566 unsigned long node_start_pfn,
6567 unsigned long node_end_pfn,
6568 unsigned long *zones_size,
6569 unsigned long *zholes_size)
6571 unsigned long realtotalpages = 0, totalpages = 0;
6574 for (i = 0; i < MAX_NR_ZONES; i++) {
6575 struct zone *zone = pgdat->node_zones + i;
6576 unsigned long zone_start_pfn, zone_end_pfn;
6577 unsigned long size, real_size;
6579 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6585 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6586 node_start_pfn, node_end_pfn,
6589 zone->zone_start_pfn = zone_start_pfn;
6591 zone->zone_start_pfn = 0;
6592 zone->spanned_pages = size;
6593 zone->present_pages = real_size;
6596 realtotalpages += real_size;
6599 pgdat->node_spanned_pages = totalpages;
6600 pgdat->node_present_pages = realtotalpages;
6601 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6605 #ifndef CONFIG_SPARSEMEM
6607 * Calculate the size of the zone->blockflags rounded to an unsigned long
6608 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6609 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6610 * round what is now in bits to nearest long in bits, then return it in
6613 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6615 unsigned long usemapsize;
6617 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6618 usemapsize = roundup(zonesize, pageblock_nr_pages);
6619 usemapsize = usemapsize >> pageblock_order;
6620 usemapsize *= NR_PAGEBLOCK_BITS;
6621 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6623 return usemapsize / 8;
6626 static void __ref setup_usemap(struct pglist_data *pgdat,
6628 unsigned long zone_start_pfn,
6629 unsigned long zonesize)
6631 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6632 zone->pageblock_flags = NULL;
6634 zone->pageblock_flags =
6635 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6637 if (!zone->pageblock_flags)
6638 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6639 usemapsize, zone->name, pgdat->node_id);
6643 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6644 unsigned long zone_start_pfn, unsigned long zonesize) {}
6645 #endif /* CONFIG_SPARSEMEM */
6647 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6649 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6650 void __init set_pageblock_order(void)
6654 /* Check that pageblock_nr_pages has not already been setup */
6655 if (pageblock_order)
6658 if (HPAGE_SHIFT > PAGE_SHIFT)
6659 order = HUGETLB_PAGE_ORDER;
6661 order = MAX_ORDER - 1;
6664 * Assume the largest contiguous order of interest is a huge page.
6665 * This value may be variable depending on boot parameters on IA64 and
6668 pageblock_order = order;
6670 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6673 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6674 * is unused as pageblock_order is set at compile-time. See
6675 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6678 void __init set_pageblock_order(void)
6682 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6684 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6685 unsigned long present_pages)
6687 unsigned long pages = spanned_pages;
6690 * Provide a more accurate estimation if there are holes within
6691 * the zone and SPARSEMEM is in use. If there are holes within the
6692 * zone, each populated memory region may cost us one or two extra
6693 * memmap pages due to alignment because memmap pages for each
6694 * populated regions may not be naturally aligned on page boundary.
6695 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6697 if (spanned_pages > present_pages + (present_pages >> 4) &&
6698 IS_ENABLED(CONFIG_SPARSEMEM))
6699 pages = present_pages;
6701 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6704 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6705 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6707 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6709 spin_lock_init(&ds_queue->split_queue_lock);
6710 INIT_LIST_HEAD(&ds_queue->split_queue);
6711 ds_queue->split_queue_len = 0;
6714 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6717 #ifdef CONFIG_COMPACTION
6718 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6720 init_waitqueue_head(&pgdat->kcompactd_wait);
6723 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6726 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6728 pgdat_resize_init(pgdat);
6730 pgdat_init_split_queue(pgdat);
6731 pgdat_init_kcompactd(pgdat);
6733 init_waitqueue_head(&pgdat->kswapd_wait);
6734 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6736 pgdat_page_ext_init(pgdat);
6737 spin_lock_init(&pgdat->lru_lock);
6738 lruvec_init(node_lruvec(pgdat));
6741 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6742 unsigned long remaining_pages)
6744 atomic_long_set(&zone->managed_pages, remaining_pages);
6745 zone_set_nid(zone, nid);
6746 zone->name = zone_names[idx];
6747 zone->zone_pgdat = NODE_DATA(nid);
6748 spin_lock_init(&zone->lock);
6749 zone_seqlock_init(zone);
6750 zone_pcp_init(zone);
6754 * Set up the zone data structures
6755 * - init pgdat internals
6756 * - init all zones belonging to this node
6758 * NOTE: this function is only called during memory hotplug
6760 #ifdef CONFIG_MEMORY_HOTPLUG
6761 void __ref free_area_init_core_hotplug(int nid)
6764 pg_data_t *pgdat = NODE_DATA(nid);
6766 pgdat_init_internals(pgdat);
6767 for (z = 0; z < MAX_NR_ZONES; z++)
6768 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6773 * Set up the zone data structures:
6774 * - mark all pages reserved
6775 * - mark all memory queues empty
6776 * - clear the memory bitmaps
6778 * NOTE: pgdat should get zeroed by caller.
6779 * NOTE: this function is only called during early init.
6781 static void __init free_area_init_core(struct pglist_data *pgdat)
6784 int nid = pgdat->node_id;
6786 pgdat_init_internals(pgdat);
6787 pgdat->per_cpu_nodestats = &boot_nodestats;
6789 for (j = 0; j < MAX_NR_ZONES; j++) {
6790 struct zone *zone = pgdat->node_zones + j;
6791 unsigned long size, freesize, memmap_pages;
6792 unsigned long zone_start_pfn = zone->zone_start_pfn;
6794 size = zone->spanned_pages;
6795 freesize = zone->present_pages;
6798 * Adjust freesize so that it accounts for how much memory
6799 * is used by this zone for memmap. This affects the watermark
6800 * and per-cpu initialisations
6802 memmap_pages = calc_memmap_size(size, freesize);
6803 if (!is_highmem_idx(j)) {
6804 if (freesize >= memmap_pages) {
6805 freesize -= memmap_pages;
6808 " %s zone: %lu pages used for memmap\n",
6809 zone_names[j], memmap_pages);
6811 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6812 zone_names[j], memmap_pages, freesize);
6815 /* Account for reserved pages */
6816 if (j == 0 && freesize > dma_reserve) {
6817 freesize -= dma_reserve;
6818 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6819 zone_names[0], dma_reserve);
6822 if (!is_highmem_idx(j))
6823 nr_kernel_pages += freesize;
6824 /* Charge for highmem memmap if there are enough kernel pages */
6825 else if (nr_kernel_pages > memmap_pages * 2)
6826 nr_kernel_pages -= memmap_pages;
6827 nr_all_pages += freesize;
6830 * Set an approximate value for lowmem here, it will be adjusted
6831 * when the bootmem allocator frees pages into the buddy system.
6832 * And all highmem pages will be managed by the buddy system.
6834 zone_init_internals(zone, j, nid, freesize);
6839 set_pageblock_order();
6840 setup_usemap(pgdat, zone, zone_start_pfn, size);
6841 init_currently_empty_zone(zone, zone_start_pfn, size);
6842 memmap_init(size, nid, j, zone_start_pfn);
6846 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6847 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6849 unsigned long __maybe_unused start = 0;
6850 unsigned long __maybe_unused offset = 0;
6852 /* Skip empty nodes */
6853 if (!pgdat->node_spanned_pages)
6856 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6857 offset = pgdat->node_start_pfn - start;
6858 /* ia64 gets its own node_mem_map, before this, without bootmem */
6859 if (!pgdat->node_mem_map) {
6860 unsigned long size, end;
6864 * The zone's endpoints aren't required to be MAX_ORDER
6865 * aligned but the node_mem_map endpoints must be in order
6866 * for the buddy allocator to function correctly.
6868 end = pgdat_end_pfn(pgdat);
6869 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6870 size = (end - start) * sizeof(struct page);
6871 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6874 panic("Failed to allocate %ld bytes for node %d memory map\n",
6875 size, pgdat->node_id);
6876 pgdat->node_mem_map = map + offset;
6878 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6879 __func__, pgdat->node_id, (unsigned long)pgdat,
6880 (unsigned long)pgdat->node_mem_map);
6881 #ifndef CONFIG_NEED_MULTIPLE_NODES
6883 * With no DISCONTIG, the global mem_map is just set as node 0's
6885 if (pgdat == NODE_DATA(0)) {
6886 mem_map = NODE_DATA(0)->node_mem_map;
6887 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6888 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6890 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6895 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6896 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6898 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6899 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6901 pgdat->first_deferred_pfn = ULONG_MAX;
6904 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6907 void __init free_area_init_node(int nid, unsigned long *zones_size,
6908 unsigned long node_start_pfn,
6909 unsigned long *zholes_size)
6911 pg_data_t *pgdat = NODE_DATA(nid);
6912 unsigned long start_pfn = 0;
6913 unsigned long end_pfn = 0;
6915 /* pg_data_t should be reset to zero when it's allocated */
6916 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6918 pgdat->node_id = nid;
6919 pgdat->node_start_pfn = node_start_pfn;
6920 pgdat->per_cpu_nodestats = NULL;
6921 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6922 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6923 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6924 (u64)start_pfn << PAGE_SHIFT,
6925 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6927 start_pfn = node_start_pfn;
6929 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6930 zones_size, zholes_size);
6932 alloc_node_mem_map(pgdat);
6933 pgdat_set_deferred_range(pgdat);
6935 free_area_init_core(pgdat);
6938 #if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6940 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6943 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6948 for (pfn = spfn; pfn < epfn; pfn++) {
6949 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6950 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6951 + pageblock_nr_pages - 1;
6954 mm_zero_struct_page(pfn_to_page(pfn));
6962 * Only struct pages that are backed by physical memory are zeroed and
6963 * initialized by going through __init_single_page(). But, there are some
6964 * struct pages which are reserved in memblock allocator and their fields
6965 * may be accessed (for example page_to_pfn() on some configuration accesses
6966 * flags). We must explicitly zero those struct pages.
6968 * This function also addresses a similar issue where struct pages are left
6969 * uninitialized because the physical address range is not covered by
6970 * memblock.memory or memblock.reserved. That could happen when memblock
6971 * layout is manually configured via memmap=, or when the highest physical
6972 * address (max_pfn) does not end on a section boundary.
6974 void __init zero_resv_unavail(void)
6976 phys_addr_t start, end;
6978 phys_addr_t next = 0;
6981 * Loop through unavailable ranges not covered by memblock.memory.
6984 for_each_mem_range(i, &memblock.memory, NULL,
6985 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6987 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6992 * Early sections always have a fully populated memmap for the whole
6993 * section - see pfn_valid(). If the last section has holes at the
6994 * end and that section is marked "online", the memmap will be
6995 * considered initialized. Make sure that memmap has a well defined
6998 pgcnt += zero_pfn_range(PFN_DOWN(next),
6999 round_up(max_pfn, PAGES_PER_SECTION));
7002 * Struct pages that do not have backing memory. This could be because
7003 * firmware is using some of this memory, or for some other reasons.
7006 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7008 #endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7010 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
7012 #if MAX_NUMNODES > 1
7014 * Figure out the number of possible node ids.
7016 void __init setup_nr_node_ids(void)
7018 unsigned int highest;
7020 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7021 nr_node_ids = highest + 1;
7026 * node_map_pfn_alignment - determine the maximum internode alignment
7028 * This function should be called after node map is populated and sorted.
7029 * It calculates the maximum power of two alignment which can distinguish
7032 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7033 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7034 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7035 * shifted, 1GiB is enough and this function will indicate so.
7037 * This is used to test whether pfn -> nid mapping of the chosen memory
7038 * model has fine enough granularity to avoid incorrect mapping for the
7039 * populated node map.
7041 * Return: the determined alignment in pfn's. 0 if there is no alignment
7042 * requirement (single node).
7044 unsigned long __init node_map_pfn_alignment(void)
7046 unsigned long accl_mask = 0, last_end = 0;
7047 unsigned long start, end, mask;
7048 int last_nid = NUMA_NO_NODE;
7051 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7052 if (!start || last_nid < 0 || last_nid == nid) {
7059 * Start with a mask granular enough to pin-point to the
7060 * start pfn and tick off bits one-by-one until it becomes
7061 * too coarse to separate the current node from the last.
7063 mask = ~((1 << __ffs(start)) - 1);
7064 while (mask && last_end <= (start & (mask << 1)))
7067 /* accumulate all internode masks */
7071 /* convert mask to number of pages */
7072 return ~accl_mask + 1;
7075 /* Find the lowest pfn for a node */
7076 static unsigned long __init find_min_pfn_for_node(int nid)
7078 unsigned long min_pfn = ULONG_MAX;
7079 unsigned long start_pfn;
7082 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7083 min_pfn = min(min_pfn, start_pfn);
7085 if (min_pfn == ULONG_MAX) {
7086 pr_warn("Could not find start_pfn for node %d\n", nid);
7094 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7096 * Return: the minimum PFN based on information provided via
7097 * memblock_set_node().
7099 unsigned long __init find_min_pfn_with_active_regions(void)
7101 return find_min_pfn_for_node(MAX_NUMNODES);
7105 * early_calculate_totalpages()
7106 * Sum pages in active regions for movable zone.
7107 * Populate N_MEMORY for calculating usable_nodes.
7109 static unsigned long __init early_calculate_totalpages(void)
7111 unsigned long totalpages = 0;
7112 unsigned long start_pfn, end_pfn;
7115 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7116 unsigned long pages = end_pfn - start_pfn;
7118 totalpages += pages;
7120 node_set_state(nid, N_MEMORY);
7126 * Find the PFN the Movable zone begins in each node. Kernel memory
7127 * is spread evenly between nodes as long as the nodes have enough
7128 * memory. When they don't, some nodes will have more kernelcore than
7131 static void __init find_zone_movable_pfns_for_nodes(void)
7134 unsigned long usable_startpfn;
7135 unsigned long kernelcore_node, kernelcore_remaining;
7136 /* save the state before borrow the nodemask */
7137 nodemask_t saved_node_state = node_states[N_MEMORY];
7138 unsigned long totalpages = early_calculate_totalpages();
7139 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7140 struct memblock_region *r;
7142 /* Need to find movable_zone earlier when movable_node is specified. */
7143 find_usable_zone_for_movable();
7146 * If movable_node is specified, ignore kernelcore and movablecore
7149 if (movable_node_is_enabled()) {
7150 for_each_memblock(memory, r) {
7151 if (!memblock_is_hotpluggable(r))
7156 usable_startpfn = PFN_DOWN(r->base);
7157 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7158 min(usable_startpfn, zone_movable_pfn[nid]) :
7166 * If kernelcore=mirror is specified, ignore movablecore option
7168 if (mirrored_kernelcore) {
7169 bool mem_below_4gb_not_mirrored = false;
7171 for_each_memblock(memory, r) {
7172 if (memblock_is_mirror(r))
7177 usable_startpfn = memblock_region_memory_base_pfn(r);
7179 if (usable_startpfn < 0x100000) {
7180 mem_below_4gb_not_mirrored = true;
7184 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7185 min(usable_startpfn, zone_movable_pfn[nid]) :
7189 if (mem_below_4gb_not_mirrored)
7190 pr_warn("This configuration results in unmirrored kernel memory.");
7196 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7197 * amount of necessary memory.
7199 if (required_kernelcore_percent)
7200 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7202 if (required_movablecore_percent)
7203 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7207 * If movablecore= was specified, calculate what size of
7208 * kernelcore that corresponds so that memory usable for
7209 * any allocation type is evenly spread. If both kernelcore
7210 * and movablecore are specified, then the value of kernelcore
7211 * will be used for required_kernelcore if it's greater than
7212 * what movablecore would have allowed.
7214 if (required_movablecore) {
7215 unsigned long corepages;
7218 * Round-up so that ZONE_MOVABLE is at least as large as what
7219 * was requested by the user
7221 required_movablecore =
7222 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7223 required_movablecore = min(totalpages, required_movablecore);
7224 corepages = totalpages - required_movablecore;
7226 required_kernelcore = max(required_kernelcore, corepages);
7230 * If kernelcore was not specified or kernelcore size is larger
7231 * than totalpages, there is no ZONE_MOVABLE.
7233 if (!required_kernelcore || required_kernelcore >= totalpages)
7236 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7237 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7240 /* Spread kernelcore memory as evenly as possible throughout nodes */
7241 kernelcore_node = required_kernelcore / usable_nodes;
7242 for_each_node_state(nid, N_MEMORY) {
7243 unsigned long start_pfn, end_pfn;
7246 * Recalculate kernelcore_node if the division per node
7247 * now exceeds what is necessary to satisfy the requested
7248 * amount of memory for the kernel
7250 if (required_kernelcore < kernelcore_node)
7251 kernelcore_node = required_kernelcore / usable_nodes;
7254 * As the map is walked, we track how much memory is usable
7255 * by the kernel using kernelcore_remaining. When it is
7256 * 0, the rest of the node is usable by ZONE_MOVABLE
7258 kernelcore_remaining = kernelcore_node;
7260 /* Go through each range of PFNs within this node */
7261 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7262 unsigned long size_pages;
7264 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7265 if (start_pfn >= end_pfn)
7268 /* Account for what is only usable for kernelcore */
7269 if (start_pfn < usable_startpfn) {
7270 unsigned long kernel_pages;
7271 kernel_pages = min(end_pfn, usable_startpfn)
7274 kernelcore_remaining -= min(kernel_pages,
7275 kernelcore_remaining);
7276 required_kernelcore -= min(kernel_pages,
7277 required_kernelcore);
7279 /* Continue if range is now fully accounted */
7280 if (end_pfn <= usable_startpfn) {
7283 * Push zone_movable_pfn to the end so
7284 * that if we have to rebalance
7285 * kernelcore across nodes, we will
7286 * not double account here
7288 zone_movable_pfn[nid] = end_pfn;
7291 start_pfn = usable_startpfn;
7295 * The usable PFN range for ZONE_MOVABLE is from
7296 * start_pfn->end_pfn. Calculate size_pages as the
7297 * number of pages used as kernelcore
7299 size_pages = end_pfn - start_pfn;
7300 if (size_pages > kernelcore_remaining)
7301 size_pages = kernelcore_remaining;
7302 zone_movable_pfn[nid] = start_pfn + size_pages;
7305 * Some kernelcore has been met, update counts and
7306 * break if the kernelcore for this node has been
7309 required_kernelcore -= min(required_kernelcore,
7311 kernelcore_remaining -= size_pages;
7312 if (!kernelcore_remaining)
7318 * If there is still required_kernelcore, we do another pass with one
7319 * less node in the count. This will push zone_movable_pfn[nid] further
7320 * along on the nodes that still have memory until kernelcore is
7324 if (usable_nodes && required_kernelcore > usable_nodes)
7328 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7329 for (nid = 0; nid < MAX_NUMNODES; nid++) {
7330 unsigned long start_pfn, end_pfn;
7332 zone_movable_pfn[nid] =
7333 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7335 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
7336 if (zone_movable_pfn[nid] >= end_pfn)
7337 zone_movable_pfn[nid] = 0;
7341 /* restore the node_state */
7342 node_states[N_MEMORY] = saved_node_state;
7345 /* Any regular or high memory on that node ? */
7346 static void check_for_memory(pg_data_t *pgdat, int nid)
7348 enum zone_type zone_type;
7350 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7351 struct zone *zone = &pgdat->node_zones[zone_type];
7352 if (populated_zone(zone)) {
7353 if (IS_ENABLED(CONFIG_HIGHMEM))
7354 node_set_state(nid, N_HIGH_MEMORY);
7355 if (zone_type <= ZONE_NORMAL)
7356 node_set_state(nid, N_NORMAL_MEMORY);
7363 * free_area_init_nodes - Initialise all pg_data_t and zone data
7364 * @max_zone_pfn: an array of max PFNs for each zone
7366 * This will call free_area_init_node() for each active node in the system.
7367 * Using the page ranges provided by memblock_set_node(), the size of each
7368 * zone in each node and their holes is calculated. If the maximum PFN
7369 * between two adjacent zones match, it is assumed that the zone is empty.
7370 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7371 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7372 * starts where the previous one ended. For example, ZONE_DMA32 starts
7373 * at arch_max_dma_pfn.
7375 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7377 unsigned long start_pfn, end_pfn;
7380 /* Record where the zone boundaries are */
7381 memset(arch_zone_lowest_possible_pfn, 0,
7382 sizeof(arch_zone_lowest_possible_pfn));
7383 memset(arch_zone_highest_possible_pfn, 0,
7384 sizeof(arch_zone_highest_possible_pfn));
7386 start_pfn = find_min_pfn_with_active_regions();
7388 for (i = 0; i < MAX_NR_ZONES; i++) {
7389 if (i == ZONE_MOVABLE)
7392 end_pfn = max(max_zone_pfn[i], start_pfn);
7393 arch_zone_lowest_possible_pfn[i] = start_pfn;
7394 arch_zone_highest_possible_pfn[i] = end_pfn;
7396 start_pfn = end_pfn;
7399 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7400 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7401 find_zone_movable_pfns_for_nodes();
7403 /* Print out the zone ranges */
7404 pr_info("Zone ranges:\n");
7405 for (i = 0; i < MAX_NR_ZONES; i++) {
7406 if (i == ZONE_MOVABLE)
7408 pr_info(" %-8s ", zone_names[i]);
7409 if (arch_zone_lowest_possible_pfn[i] ==
7410 arch_zone_highest_possible_pfn[i])
7413 pr_cont("[mem %#018Lx-%#018Lx]\n",
7414 (u64)arch_zone_lowest_possible_pfn[i]
7416 ((u64)arch_zone_highest_possible_pfn[i]
7417 << PAGE_SHIFT) - 1);
7420 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7421 pr_info("Movable zone start for each node\n");
7422 for (i = 0; i < MAX_NUMNODES; i++) {
7423 if (zone_movable_pfn[i])
7424 pr_info(" Node %d: %#018Lx\n", i,
7425 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7429 * Print out the early node map, and initialize the
7430 * subsection-map relative to active online memory ranges to
7431 * enable future "sub-section" extensions of the memory map.
7433 pr_info("Early memory node ranges\n");
7434 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7435 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7436 (u64)start_pfn << PAGE_SHIFT,
7437 ((u64)end_pfn << PAGE_SHIFT) - 1);
7438 subsection_map_init(start_pfn, end_pfn - start_pfn);
7441 /* Initialise every node */
7442 mminit_verify_pageflags_layout();
7443 setup_nr_node_ids();
7444 zero_resv_unavail();
7445 for_each_online_node(nid) {
7446 pg_data_t *pgdat = NODE_DATA(nid);
7447 free_area_init_node(nid, NULL,
7448 find_min_pfn_for_node(nid), NULL);
7450 /* Any memory on that node */
7451 if (pgdat->node_present_pages)
7452 node_set_state(nid, N_MEMORY);
7453 check_for_memory(pgdat, nid);
7457 static int __init cmdline_parse_core(char *p, unsigned long *core,
7458 unsigned long *percent)
7460 unsigned long long coremem;
7466 /* Value may be a percentage of total memory, otherwise bytes */
7467 coremem = simple_strtoull(p, &endptr, 0);
7468 if (*endptr == '%') {
7469 /* Paranoid check for percent values greater than 100 */
7470 WARN_ON(coremem > 100);
7474 coremem = memparse(p, &p);
7475 /* Paranoid check that UL is enough for the coremem value */
7476 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7478 *core = coremem >> PAGE_SHIFT;
7485 * kernelcore=size sets the amount of memory for use for allocations that
7486 * cannot be reclaimed or migrated.
7488 static int __init cmdline_parse_kernelcore(char *p)
7490 /* parse kernelcore=mirror */
7491 if (parse_option_str(p, "mirror")) {
7492 mirrored_kernelcore = true;
7496 return cmdline_parse_core(p, &required_kernelcore,
7497 &required_kernelcore_percent);
7501 * movablecore=size sets the amount of memory for use for allocations that
7502 * can be reclaimed or migrated.
7504 static int __init cmdline_parse_movablecore(char *p)
7506 return cmdline_parse_core(p, &required_movablecore,
7507 &required_movablecore_percent);
7510 early_param("kernelcore", cmdline_parse_kernelcore);
7511 early_param("movablecore", cmdline_parse_movablecore);
7513 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7515 void adjust_managed_page_count(struct page *page, long count)
7517 atomic_long_add(count, &page_zone(page)->managed_pages);
7518 totalram_pages_add(count);
7519 #ifdef CONFIG_HIGHMEM
7520 if (PageHighMem(page))
7521 totalhigh_pages_add(count);
7524 EXPORT_SYMBOL(adjust_managed_page_count);
7526 unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7529 unsigned long pages = 0;
7531 start = (void *)PAGE_ALIGN((unsigned long)start);
7532 end = (void *)((unsigned long)end & PAGE_MASK);
7533 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7534 struct page *page = virt_to_page(pos);
7535 void *direct_map_addr;
7538 * 'direct_map_addr' might be different from 'pos'
7539 * because some architectures' virt_to_page()
7540 * work with aliases. Getting the direct map
7541 * address ensures that we get a _writeable_
7542 * alias for the memset().
7544 direct_map_addr = page_address(page);
7545 if ((unsigned int)poison <= 0xFF)
7546 memset(direct_map_addr, poison, PAGE_SIZE);
7548 free_reserved_page(page);
7552 pr_info("Freeing %s memory: %ldK\n",
7553 s, pages << (PAGE_SHIFT - 10));
7558 #ifdef CONFIG_HIGHMEM
7559 void free_highmem_page(struct page *page)
7561 __free_reserved_page(page);
7562 totalram_pages_inc();
7563 atomic_long_inc(&page_zone(page)->managed_pages);
7564 totalhigh_pages_inc();
7569 void __init mem_init_print_info(const char *str)
7571 unsigned long physpages, codesize, datasize, rosize, bss_size;
7572 unsigned long init_code_size, init_data_size;
7574 physpages = get_num_physpages();
7575 codesize = _etext - _stext;
7576 datasize = _edata - _sdata;
7577 rosize = __end_rodata - __start_rodata;
7578 bss_size = __bss_stop - __bss_start;
7579 init_data_size = __init_end - __init_begin;
7580 init_code_size = _einittext - _sinittext;
7583 * Detect special cases and adjust section sizes accordingly:
7584 * 1) .init.* may be embedded into .data sections
7585 * 2) .init.text.* may be out of [__init_begin, __init_end],
7586 * please refer to arch/tile/kernel/vmlinux.lds.S.
7587 * 3) .rodata.* may be embedded into .text or .data sections.
7589 #define adj_init_size(start, end, size, pos, adj) \
7591 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
7595 adj_init_size(__init_begin, __init_end, init_data_size,
7596 _sinittext, init_code_size);
7597 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7598 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7599 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7600 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7602 #undef adj_init_size
7604 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7605 #ifdef CONFIG_HIGHMEM
7609 nr_free_pages() << (PAGE_SHIFT - 10),
7610 physpages << (PAGE_SHIFT - 10),
7611 codesize >> 10, datasize >> 10, rosize >> 10,
7612 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7613 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7614 totalcma_pages << (PAGE_SHIFT - 10),
7615 #ifdef CONFIG_HIGHMEM
7616 totalhigh_pages() << (PAGE_SHIFT - 10),
7618 str ? ", " : "", str ? str : "");
7622 * set_dma_reserve - set the specified number of pages reserved in the first zone
7623 * @new_dma_reserve: The number of pages to mark reserved
7625 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7626 * In the DMA zone, a significant percentage may be consumed by kernel image
7627 * and other unfreeable allocations which can skew the watermarks badly. This
7628 * function may optionally be used to account for unfreeable pages in the
7629 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7630 * smaller per-cpu batchsize.
7632 void __init set_dma_reserve(unsigned long new_dma_reserve)
7634 dma_reserve = new_dma_reserve;
7637 void __init free_area_init(unsigned long *zones_size)
7639 zero_resv_unavail();
7640 free_area_init_node(0, zones_size,
7641 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7644 static int page_alloc_cpu_dead(unsigned int cpu)
7647 lru_add_drain_cpu(cpu);
7651 * Spill the event counters of the dead processor
7652 * into the current processors event counters.
7653 * This artificially elevates the count of the current
7656 vm_events_fold_cpu(cpu);
7659 * Zero the differential counters of the dead processor
7660 * so that the vm statistics are consistent.
7662 * This is only okay since the processor is dead and cannot
7663 * race with what we are doing.
7665 cpu_vm_stats_fold(cpu);
7670 int hashdist = HASHDIST_DEFAULT;
7672 static int __init set_hashdist(char *str)
7676 hashdist = simple_strtoul(str, &str, 0);
7679 __setup("hashdist=", set_hashdist);
7682 void __init page_alloc_init(void)
7687 if (num_node_state(N_MEMORY) == 1)
7691 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7692 "mm/page_alloc:dead", NULL,
7693 page_alloc_cpu_dead);
7698 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7699 * or min_free_kbytes changes.
7701 static void calculate_totalreserve_pages(void)
7703 struct pglist_data *pgdat;
7704 unsigned long reserve_pages = 0;
7705 enum zone_type i, j;
7707 for_each_online_pgdat(pgdat) {
7709 pgdat->totalreserve_pages = 0;
7711 for (i = 0; i < MAX_NR_ZONES; i++) {
7712 struct zone *zone = pgdat->node_zones + i;
7714 unsigned long managed_pages = zone_managed_pages(zone);
7716 /* Find valid and maximum lowmem_reserve in the zone */
7717 for (j = i; j < MAX_NR_ZONES; j++) {
7718 if (zone->lowmem_reserve[j] > max)
7719 max = zone->lowmem_reserve[j];
7722 /* we treat the high watermark as reserved pages. */
7723 max += high_wmark_pages(zone);
7725 if (max > managed_pages)
7726 max = managed_pages;
7728 pgdat->totalreserve_pages += max;
7730 reserve_pages += max;
7733 totalreserve_pages = reserve_pages;
7737 * setup_per_zone_lowmem_reserve - called whenever
7738 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7739 * has a correct pages reserved value, so an adequate number of
7740 * pages are left in the zone after a successful __alloc_pages().
7742 static void setup_per_zone_lowmem_reserve(void)
7744 struct pglist_data *pgdat;
7745 enum zone_type j, idx;
7747 for_each_online_pgdat(pgdat) {
7748 for (j = 0; j < MAX_NR_ZONES; j++) {
7749 struct zone *zone = pgdat->node_zones + j;
7750 unsigned long managed_pages = zone_managed_pages(zone);
7752 zone->lowmem_reserve[j] = 0;
7756 struct zone *lower_zone;
7759 lower_zone = pgdat->node_zones + idx;
7761 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7762 sysctl_lowmem_reserve_ratio[idx] = 0;
7763 lower_zone->lowmem_reserve[j] = 0;
7765 lower_zone->lowmem_reserve[j] =
7766 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7768 managed_pages += zone_managed_pages(lower_zone);
7773 /* update totalreserve_pages */
7774 calculate_totalreserve_pages();
7777 static void __setup_per_zone_wmarks(void)
7779 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7780 unsigned long lowmem_pages = 0;
7782 unsigned long flags;
7784 /* Calculate total number of !ZONE_HIGHMEM pages */
7785 for_each_zone(zone) {
7786 if (!is_highmem(zone))
7787 lowmem_pages += zone_managed_pages(zone);
7790 for_each_zone(zone) {
7793 spin_lock_irqsave(&zone->lock, flags);
7794 tmp = (u64)pages_min * zone_managed_pages(zone);
7795 do_div(tmp, lowmem_pages);
7796 if (is_highmem(zone)) {
7798 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7799 * need highmem pages, so cap pages_min to a small
7802 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7803 * deltas control async page reclaim, and so should
7804 * not be capped for highmem.
7806 unsigned long min_pages;
7808 min_pages = zone_managed_pages(zone) / 1024;
7809 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7810 zone->_watermark[WMARK_MIN] = min_pages;
7813 * If it's a lowmem zone, reserve a number of pages
7814 * proportionate to the zone's size.
7816 zone->_watermark[WMARK_MIN] = tmp;
7820 * Set the kswapd watermarks distance according to the
7821 * scale factor in proportion to available memory, but
7822 * ensure a minimum size on small systems.
7824 tmp = max_t(u64, tmp >> 2,
7825 mult_frac(zone_managed_pages(zone),
7826 watermark_scale_factor, 10000));
7828 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7829 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7830 zone->watermark_boost = 0;
7832 spin_unlock_irqrestore(&zone->lock, flags);
7835 /* update totalreserve_pages */
7836 calculate_totalreserve_pages();
7840 * setup_per_zone_wmarks - called when min_free_kbytes changes
7841 * or when memory is hot-{added|removed}
7843 * Ensures that the watermark[min,low,high] values for each zone are set
7844 * correctly with respect to min_free_kbytes.
7846 void setup_per_zone_wmarks(void)
7848 static DEFINE_SPINLOCK(lock);
7851 __setup_per_zone_wmarks();
7856 * Initialise min_free_kbytes.
7858 * For small machines we want it small (128k min). For large machines
7859 * we want it large (64MB max). But it is not linear, because network
7860 * bandwidth does not increase linearly with machine size. We use
7862 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7863 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7879 int __meminit init_per_zone_wmark_min(void)
7881 unsigned long lowmem_kbytes;
7882 int new_min_free_kbytes;
7884 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7885 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7887 if (new_min_free_kbytes > user_min_free_kbytes) {
7888 min_free_kbytes = new_min_free_kbytes;
7889 if (min_free_kbytes < 128)
7890 min_free_kbytes = 128;
7891 if (min_free_kbytes > 65536)
7892 min_free_kbytes = 65536;
7894 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7895 new_min_free_kbytes, user_min_free_kbytes);
7897 setup_per_zone_wmarks();
7898 refresh_zone_stat_thresholds();
7899 setup_per_zone_lowmem_reserve();
7902 setup_min_unmapped_ratio();
7903 setup_min_slab_ratio();
7906 khugepaged_min_free_kbytes_update();
7910 postcore_initcall(init_per_zone_wmark_min)
7913 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7914 * that we can call two helper functions whenever min_free_kbytes
7917 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7918 void __user *buffer, size_t *length, loff_t *ppos)
7922 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7927 user_min_free_kbytes = min_free_kbytes;
7928 setup_per_zone_wmarks();
7933 int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7934 void __user *buffer, size_t *length, loff_t *ppos)
7938 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7945 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7946 void __user *buffer, size_t *length, loff_t *ppos)
7950 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7955 setup_per_zone_wmarks();
7961 static void setup_min_unmapped_ratio(void)
7966 for_each_online_pgdat(pgdat)
7967 pgdat->min_unmapped_pages = 0;
7970 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7971 sysctl_min_unmapped_ratio) / 100;
7975 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7976 void __user *buffer, size_t *length, loff_t *ppos)
7980 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7984 setup_min_unmapped_ratio();
7989 static void setup_min_slab_ratio(void)
7994 for_each_online_pgdat(pgdat)
7995 pgdat->min_slab_pages = 0;
7998 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7999 sysctl_min_slab_ratio) / 100;
8002 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
8003 void __user *buffer, size_t *length, loff_t *ppos)
8007 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
8011 setup_min_slab_ratio();
8018 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8019 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8020 * whenever sysctl_lowmem_reserve_ratio changes.
8022 * The reserve ratio obviously has absolutely no relation with the
8023 * minimum watermarks. The lowmem reserve ratio can only make sense
8024 * if in function of the boot time zone sizes.
8026 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8027 void __user *buffer, size_t *length, loff_t *ppos)
8029 proc_dointvec_minmax(table, write, buffer, length, ppos);
8030 setup_per_zone_lowmem_reserve();
8035 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8036 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8037 * pagelist can have before it gets flushed back to buddy allocator.
8039 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8040 void __user *buffer, size_t *length, loff_t *ppos)
8043 int old_percpu_pagelist_fraction;
8046 mutex_lock(&pcp_batch_high_lock);
8047 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8049 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8050 if (!write || ret < 0)
8053 /* Sanity checking to avoid pcp imbalance */
8054 if (percpu_pagelist_fraction &&
8055 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8056 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8062 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8065 for_each_populated_zone(zone) {
8068 for_each_possible_cpu(cpu)
8069 pageset_set_high_and_batch(zone,
8070 per_cpu_ptr(zone->pageset, cpu));
8073 mutex_unlock(&pcp_batch_high_lock);
8077 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8079 * Returns the number of pages that arch has reserved but
8080 * is not known to alloc_large_system_hash().
8082 static unsigned long __init arch_reserved_kernel_pages(void)
8089 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8090 * machines. As memory size is increased the scale is also increased but at
8091 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8092 * quadruples the scale is increased by one, which means the size of hash table
8093 * only doubles, instead of quadrupling as well.
8094 * Because 32-bit systems cannot have large physical memory, where this scaling
8095 * makes sense, it is disabled on such platforms.
8097 #if __BITS_PER_LONG > 32
8098 #define ADAPT_SCALE_BASE (64ul << 30)
8099 #define ADAPT_SCALE_SHIFT 2
8100 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8104 * allocate a large system hash table from bootmem
8105 * - it is assumed that the hash table must contain an exact power-of-2
8106 * quantity of entries
8107 * - limit is the number of hash buckets, not the total allocation size
8109 void *__init alloc_large_system_hash(const char *tablename,
8110 unsigned long bucketsize,
8111 unsigned long numentries,
8114 unsigned int *_hash_shift,
8115 unsigned int *_hash_mask,
8116 unsigned long low_limit,
8117 unsigned long high_limit)
8119 unsigned long long max = high_limit;
8120 unsigned long log2qty, size;
8125 /* allow the kernel cmdline to have a say */
8127 /* round applicable memory size up to nearest megabyte */
8128 numentries = nr_kernel_pages;
8129 numentries -= arch_reserved_kernel_pages();
8131 /* It isn't necessary when PAGE_SIZE >= 1MB */
8132 if (PAGE_SHIFT < 20)
8133 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8135 #if __BITS_PER_LONG > 32
8137 unsigned long adapt;
8139 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8140 adapt <<= ADAPT_SCALE_SHIFT)
8145 /* limit to 1 bucket per 2^scale bytes of low memory */
8146 if (scale > PAGE_SHIFT)
8147 numentries >>= (scale - PAGE_SHIFT);
8149 numentries <<= (PAGE_SHIFT - scale);
8151 /* Make sure we've got at least a 0-order allocation.. */
8152 if (unlikely(flags & HASH_SMALL)) {
8153 /* Makes no sense without HASH_EARLY */
8154 WARN_ON(!(flags & HASH_EARLY));
8155 if (!(numentries >> *_hash_shift)) {
8156 numentries = 1UL << *_hash_shift;
8157 BUG_ON(!numentries);
8159 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8160 numentries = PAGE_SIZE / bucketsize;
8162 numentries = roundup_pow_of_two(numentries);
8164 /* limit allocation size to 1/16 total memory by default */
8166 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8167 do_div(max, bucketsize);
8169 max = min(max, 0x80000000ULL);
8171 if (numentries < low_limit)
8172 numentries = low_limit;
8173 if (numentries > max)
8176 log2qty = ilog2(numentries);
8178 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8181 size = bucketsize << log2qty;
8182 if (flags & HASH_EARLY) {
8183 if (flags & HASH_ZERO)
8184 table = memblock_alloc(size, SMP_CACHE_BYTES);
8186 table = memblock_alloc_raw(size,
8188 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8189 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8193 * If bucketsize is not a power-of-two, we may free
8194 * some pages at the end of hash table which
8195 * alloc_pages_exact() automatically does
8197 table = alloc_pages_exact(size, gfp_flags);
8198 kmemleak_alloc(table, size, 1, gfp_flags);
8200 } while (!table && size > PAGE_SIZE && --log2qty);
8203 panic("Failed to allocate %s hash table\n", tablename);
8205 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8206 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8207 virt ? "vmalloc" : "linear");
8210 *_hash_shift = log2qty;
8212 *_hash_mask = (1 << log2qty) - 1;
8218 * This function checks whether pageblock includes unmovable pages or not.
8219 * If @count is not zero, it is okay to include less @count unmovable pages
8221 * PageLRU check without isolation or lru_lock could race so that
8222 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8223 * check without lock_page also may miss some movable non-lru pages at
8224 * race condition. So you can't expect this function should be exact.
8226 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
8227 int migratetype, int flags)
8229 unsigned long found;
8230 unsigned long iter = 0;
8231 unsigned long pfn = page_to_pfn(page);
8232 const char *reason = "unmovable page";
8235 * TODO we could make this much more efficient by not checking every
8236 * page in the range if we know all of them are in MOVABLE_ZONE and
8237 * that the movable zone guarantees that pages are migratable but
8238 * the later is not the case right now unfortunatelly. E.g. movablecore
8239 * can still lead to having bootmem allocations in zone_movable.
8242 if (is_migrate_cma_page(page)) {
8244 * CMA allocations (alloc_contig_range) really need to mark
8245 * isolate CMA pageblocks even when they are not movable in fact
8246 * so consider them movable here.
8248 if (is_migrate_cma(migratetype))
8251 reason = "CMA page";
8255 for (found = 0; iter < pageblock_nr_pages; iter++) {
8256 unsigned long check = pfn + iter;
8258 if (!pfn_valid_within(check))
8261 page = pfn_to_page(check);
8263 if (PageReserved(page))
8267 * If the zone is movable and we have ruled out all reserved
8268 * pages then it should be reasonably safe to assume the rest
8271 if (zone_idx(zone) == ZONE_MOVABLE)
8275 * Hugepages are not in LRU lists, but they're movable.
8276 * We need not scan over tail pages because we don't
8277 * handle each tail page individually in migration.
8279 if (PageHuge(page)) {
8280 struct page *head = compound_head(page);
8281 unsigned int skip_pages;
8283 if (!hugepage_migration_supported(page_hstate(head)))
8286 skip_pages = compound_nr(head) - (page - head);
8287 iter += skip_pages - 1;
8292 * We can't use page_count without pin a page
8293 * because another CPU can free compound page.
8294 * This check already skips compound tails of THP
8295 * because their page->_refcount is zero at all time.
8297 if (!page_ref_count(page)) {
8298 if (PageBuddy(page))
8299 iter += (1 << page_order(page)) - 1;
8304 * The HWPoisoned page may be not in buddy system, and
8305 * page_count() is not 0.
8307 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
8310 if (__PageMovable(page))
8316 * If there are RECLAIMABLE pages, we need to check
8317 * it. But now, memory offline itself doesn't call
8318 * shrink_node_slabs() and it still to be fixed.
8321 * If the page is not RAM, page_count()should be 0.
8322 * we don't need more check. This is an _used_ not-movable page.
8324 * The problematic thing here is PG_reserved pages. PG_reserved
8325 * is set to both of a memory hole page and a _used_ kernel
8333 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8334 if (flags & REPORT_FAILURE)
8335 dump_page(pfn_to_page(pfn + iter), reason);
8339 #ifdef CONFIG_CONTIG_ALLOC
8340 static unsigned long pfn_max_align_down(unsigned long pfn)
8342 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8343 pageblock_nr_pages) - 1);
8346 static unsigned long pfn_max_align_up(unsigned long pfn)
8348 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8349 pageblock_nr_pages));
8352 /* [start, end) must belong to a single zone. */
8353 static int __alloc_contig_migrate_range(struct compact_control *cc,
8354 unsigned long start, unsigned long end)
8356 /* This function is based on compact_zone() from compaction.c. */
8357 unsigned long nr_reclaimed;
8358 unsigned long pfn = start;
8359 unsigned int tries = 0;
8364 while (pfn < end || !list_empty(&cc->migratepages)) {
8365 if (fatal_signal_pending(current)) {
8370 if (list_empty(&cc->migratepages)) {
8371 cc->nr_migratepages = 0;
8372 pfn = isolate_migratepages_range(cc, pfn, end);
8378 } else if (++tries == 5) {
8379 ret = ret < 0 ? ret : -EBUSY;
8383 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8385 cc->nr_migratepages -= nr_reclaimed;
8387 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8388 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8391 putback_movable_pages(&cc->migratepages);
8398 * alloc_contig_range() -- tries to allocate given range of pages
8399 * @start: start PFN to allocate
8400 * @end: one-past-the-last PFN to allocate
8401 * @migratetype: migratetype of the underlaying pageblocks (either
8402 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8403 * in range must have the same migratetype and it must
8404 * be either of the two.
8405 * @gfp_mask: GFP mask to use during compaction
8407 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8408 * aligned. The PFN range must belong to a single zone.
8410 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8411 * pageblocks in the range. Once isolated, the pageblocks should not
8412 * be modified by others.
8414 * Return: zero on success or negative error code. On success all
8415 * pages which PFN is in [start, end) are allocated for the caller and
8416 * need to be freed with free_contig_range().
8418 int alloc_contig_range(unsigned long start, unsigned long end,
8419 unsigned migratetype, gfp_t gfp_mask)
8421 unsigned long outer_start, outer_end;
8425 struct compact_control cc = {
8426 .nr_migratepages = 0,
8428 .zone = page_zone(pfn_to_page(start)),
8429 .mode = MIGRATE_SYNC,
8430 .ignore_skip_hint = true,
8431 .no_set_skip_hint = true,
8432 .gfp_mask = current_gfp_context(gfp_mask),
8434 INIT_LIST_HEAD(&cc.migratepages);
8437 * What we do here is we mark all pageblocks in range as
8438 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8439 * have different sizes, and due to the way page allocator
8440 * work, we align the range to biggest of the two pages so
8441 * that page allocator won't try to merge buddies from
8442 * different pageblocks and change MIGRATE_ISOLATE to some
8443 * other migration type.
8445 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8446 * migrate the pages from an unaligned range (ie. pages that
8447 * we are interested in). This will put all the pages in
8448 * range back to page allocator as MIGRATE_ISOLATE.
8450 * When this is done, we take the pages in range from page
8451 * allocator removing them from the buddy system. This way
8452 * page allocator will never consider using them.
8454 * This lets us mark the pageblocks back as
8455 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8456 * aligned range but not in the unaligned, original range are
8457 * put back to page allocator so that buddy can use them.
8460 ret = start_isolate_page_range(pfn_max_align_down(start),
8461 pfn_max_align_up(end), migratetype, 0);
8466 * In case of -EBUSY, we'd like to know which page causes problem.
8467 * So, just fall through. test_pages_isolated() has a tracepoint
8468 * which will report the busy page.
8470 * It is possible that busy pages could become available before
8471 * the call to test_pages_isolated, and the range will actually be
8472 * allocated. So, if we fall through be sure to clear ret so that
8473 * -EBUSY is not accidentally used or returned to caller.
8475 ret = __alloc_contig_migrate_range(&cc, start, end);
8476 if (ret && ret != -EBUSY)
8481 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8482 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8483 * more, all pages in [start, end) are free in page allocator.
8484 * What we are going to do is to allocate all pages from
8485 * [start, end) (that is remove them from page allocator).
8487 * The only problem is that pages at the beginning and at the
8488 * end of interesting range may be not aligned with pages that
8489 * page allocator holds, ie. they can be part of higher order
8490 * pages. Because of this, we reserve the bigger range and
8491 * once this is done free the pages we are not interested in.
8493 * We don't have to hold zone->lock here because the pages are
8494 * isolated thus they won't get removed from buddy.
8497 lru_add_drain_all();
8500 outer_start = start;
8501 while (!PageBuddy(pfn_to_page(outer_start))) {
8502 if (++order >= MAX_ORDER) {
8503 outer_start = start;
8506 outer_start &= ~0UL << order;
8509 if (outer_start != start) {
8510 order = page_order(pfn_to_page(outer_start));
8513 * outer_start page could be small order buddy page and
8514 * it doesn't include start page. Adjust outer_start
8515 * in this case to report failed page properly
8516 * on tracepoint in test_pages_isolated()
8518 if (outer_start + (1UL << order) <= start)
8519 outer_start = start;
8522 /* Make sure the range is really isolated. */
8523 if (test_pages_isolated(outer_start, end, false)) {
8524 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8525 __func__, outer_start, end);
8530 /* Grab isolated pages from freelists. */
8531 outer_end = isolate_freepages_range(&cc, outer_start, end);
8537 /* Free head and tail (if any) */
8538 if (start != outer_start)
8539 free_contig_range(outer_start, start - outer_start);
8540 if (end != outer_end)
8541 free_contig_range(end, outer_end - end);
8544 undo_isolate_page_range(pfn_max_align_down(start),
8545 pfn_max_align_up(end), migratetype);
8548 #endif /* CONFIG_CONTIG_ALLOC */
8550 void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8552 unsigned int count = 0;
8554 for (; nr_pages--; pfn++) {
8555 struct page *page = pfn_to_page(pfn);
8557 count += page_count(page) != 1;
8560 WARN(count != 0, "%d pages are still in use!\n", count);
8564 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8565 * page high values need to be recalulated.
8567 void __meminit zone_pcp_update(struct zone *zone)
8570 mutex_lock(&pcp_batch_high_lock);
8571 for_each_possible_cpu(cpu)
8572 pageset_set_high_and_batch(zone,
8573 per_cpu_ptr(zone->pageset, cpu));
8574 mutex_unlock(&pcp_batch_high_lock);
8577 void zone_pcp_reset(struct zone *zone)
8579 unsigned long flags;
8581 struct per_cpu_pageset *pset;
8583 /* avoid races with drain_pages() */
8584 local_irq_save(flags);
8585 if (zone->pageset != &boot_pageset) {
8586 for_each_online_cpu(cpu) {
8587 pset = per_cpu_ptr(zone->pageset, cpu);
8588 drain_zonestat(zone, pset);
8590 free_percpu(zone->pageset);
8591 zone->pageset = &boot_pageset;
8593 local_irq_restore(flags);
8596 #ifdef CONFIG_MEMORY_HOTREMOVE
8598 * All pages in the range must be in a single zone and isolated
8599 * before calling this.
8602 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8606 unsigned int order, i;
8608 unsigned long flags;
8609 unsigned long offlined_pages = 0;
8611 /* find the first valid pfn */
8612 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8616 return offlined_pages;
8618 offline_mem_sections(pfn, end_pfn);
8619 zone = page_zone(pfn_to_page(pfn));
8620 spin_lock_irqsave(&zone->lock, flags);
8622 while (pfn < end_pfn) {
8623 if (!pfn_valid(pfn)) {
8627 page = pfn_to_page(pfn);
8629 * The HWPoisoned page may be not in buddy system, and
8630 * page_count() is not 0.
8632 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8634 SetPageReserved(page);
8639 BUG_ON(page_count(page));
8640 BUG_ON(!PageBuddy(page));
8641 order = page_order(page);
8642 offlined_pages += 1 << order;
8643 #ifdef CONFIG_DEBUG_VM
8644 pr_info("remove from free list %lx %d %lx\n",
8645 pfn, 1 << order, end_pfn);
8647 del_page_from_free_area(page, &zone->free_area[order]);
8648 for (i = 0; i < (1 << order); i++)
8649 SetPageReserved((page+i));
8650 pfn += (1 << order);
8652 spin_unlock_irqrestore(&zone->lock, flags);
8654 return offlined_pages;
8658 bool is_free_buddy_page(struct page *page)
8660 struct zone *zone = page_zone(page);
8661 unsigned long pfn = page_to_pfn(page);
8662 unsigned long flags;
8665 spin_lock_irqsave(&zone->lock, flags);
8666 for (order = 0; order < MAX_ORDER; order++) {
8667 struct page *page_head = page - (pfn & ((1 << order) - 1));
8669 if (PageBuddy(page_head) && page_order(page_head) >= order)
8672 spin_unlock_irqrestore(&zone->lock, flags);
8674 return order < MAX_ORDER;
8677 #ifdef CONFIG_MEMORY_FAILURE
8679 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8680 * test is performed under the zone lock to prevent a race against page
8683 bool set_hwpoison_free_buddy_page(struct page *page)
8685 struct zone *zone = page_zone(page);
8686 unsigned long pfn = page_to_pfn(page);
8687 unsigned long flags;
8689 bool hwpoisoned = false;
8691 spin_lock_irqsave(&zone->lock, flags);
8692 for (order = 0; order < MAX_ORDER; order++) {
8693 struct page *page_head = page - (pfn & ((1 << order) - 1));
8695 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8696 if (!TestSetPageHWPoison(page))
8701 spin_unlock_irqrestore(&zone->lock, flags);
8707 #ifdef CONFIG_ZONE_DMA
8708 bool has_managed_dma(void)
8710 struct pglist_data *pgdat;
8712 for_each_online_pgdat(pgdat) {
8713 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
8715 if (managed_zone(zone))
8720 #endif /* CONFIG_ZONE_DMA */