2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/topology.h>
36 #include <linux/sysctl.h>
37 #include <linux/cpu.h>
38 #include <linux/cpuset.h>
39 #include <linux/memory_hotplug.h>
40 #include <linux/nodemask.h>
41 #include <linux/vmalloc.h>
42 #include <linux/vmstat.h>
43 #include <linux/mempolicy.h>
44 #include <linux/memremap.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_ext.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <trace/events/oom.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
62 #include <linux/sched/mm.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
65 #include <linux/memcontrol.h>
66 #include <linux/ftrace.h>
67 #include <linux/lockdep.h>
68 #include <linux/nmi.h>
69 #include <linux/khugepaged.h>
71 #include <asm/sections.h>
72 #include <asm/tlbflush.h>
73 #include <asm/div64.h>
76 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77 static DEFINE_MUTEX(pcp_batch_high_lock);
78 #define MIN_PERCPU_PAGELIST_FRACTION (8)
80 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81 DEFINE_PER_CPU(int, numa_node);
82 EXPORT_PER_CPU_SYMBOL(numa_node);
85 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
87 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
94 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96 int _node_numa_mem_[MAX_NUMNODES];
99 /* work_structs for global per-cpu drains */
100 DEFINE_MUTEX(pcpu_drain_mutex);
101 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
103 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104 volatile unsigned long latent_entropy __latent_entropy;
105 EXPORT_SYMBOL(latent_entropy);
109 * Array of node states.
111 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
112 [N_POSSIBLE] = NODE_MASK_ALL,
113 [N_ONLINE] = { { [0] = 1UL } },
115 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
116 #ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY] = { { [0] = 1UL } },
119 [N_MEMORY] = { { [0] = 1UL } },
120 [N_CPU] = { { [0] = 1UL } },
123 EXPORT_SYMBOL(node_states);
125 /* Protect totalram_pages and zone->managed_pages */
126 static DEFINE_SPINLOCK(managed_page_count_lock);
128 unsigned long totalram_pages __read_mostly;
129 unsigned long totalreserve_pages __read_mostly;
130 unsigned long totalcma_pages __read_mostly;
132 int percpu_pagelist_fraction;
133 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
136 * A cached value of the page's pageblock's migratetype, used when the page is
137 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
138 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
139 * Also the migratetype set in the page does not necessarily match the pcplist
140 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
141 * other index - this ensures that it will be put on the correct CMA freelist.
143 static inline int get_pcppage_migratetype(struct page *page)
148 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
150 page->index = migratetype;
153 #ifdef CONFIG_PM_SLEEP
155 * The following functions are used by the suspend/hibernate code to temporarily
156 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
157 * while devices are suspended. To avoid races with the suspend/hibernate code,
158 * they should always be called with system_transition_mutex held
159 * (gfp_allowed_mask also should only be modified with system_transition_mutex
160 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
161 * with that modification).
164 static gfp_t saved_gfp_mask;
166 void pm_restore_gfp_mask(void)
168 WARN_ON(!mutex_is_locked(&system_transition_mutex));
169 if (saved_gfp_mask) {
170 gfp_allowed_mask = saved_gfp_mask;
175 void pm_restrict_gfp_mask(void)
177 WARN_ON(!mutex_is_locked(&system_transition_mutex));
178 WARN_ON(saved_gfp_mask);
179 saved_gfp_mask = gfp_allowed_mask;
180 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
183 bool pm_suspended_storage(void)
185 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
189 #endif /* CONFIG_PM_SLEEP */
191 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
192 unsigned int pageblock_order __read_mostly;
195 static void __free_pages_ok(struct page *page, unsigned int order);
198 * results with 256, 32 in the lowmem_reserve sysctl:
199 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
200 * 1G machine -> (16M dma, 784M normal, 224M high)
201 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
202 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
203 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
205 * TBD: should special case ZONE_DMA32 machines here - in those we normally
206 * don't need any ZONE_NORMAL reservation
208 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
209 #ifdef CONFIG_ZONE_DMA
212 #ifdef CONFIG_ZONE_DMA32
216 #ifdef CONFIG_HIGHMEM
222 EXPORT_SYMBOL(totalram_pages);
224 static char * const zone_names[MAX_NR_ZONES] = {
225 #ifdef CONFIG_ZONE_DMA
228 #ifdef CONFIG_ZONE_DMA32
232 #ifdef CONFIG_HIGHMEM
236 #ifdef CONFIG_ZONE_DEVICE
241 char * const migratetype_names[MIGRATE_TYPES] = {
249 #ifdef CONFIG_MEMORY_ISOLATION
254 compound_page_dtor * const compound_page_dtors[] = {
257 #ifdef CONFIG_HUGETLB_PAGE
260 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
265 int min_free_kbytes = 1024;
266 int user_min_free_kbytes = -1;
267 int watermark_scale_factor = 10;
269 static unsigned long nr_kernel_pages __meminitdata;
270 static unsigned long nr_all_pages __meminitdata;
271 static unsigned long dma_reserve __meminitdata;
273 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
274 static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __meminitdata;
275 static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __meminitdata;
276 static unsigned long required_kernelcore __initdata;
277 static unsigned long required_kernelcore_percent __initdata;
278 static unsigned long required_movablecore __initdata;
279 static unsigned long required_movablecore_percent __initdata;
280 static unsigned long zone_movable_pfn[MAX_NUMNODES] __meminitdata;
281 static bool mirrored_kernelcore __meminitdata;
283 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
285 EXPORT_SYMBOL(movable_zone);
286 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
289 int nr_node_ids __read_mostly = MAX_NUMNODES;
290 int nr_online_nodes __read_mostly = 1;
291 EXPORT_SYMBOL(nr_node_ids);
292 EXPORT_SYMBOL(nr_online_nodes);
295 int page_group_by_mobility_disabled __read_mostly;
297 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
299 * During boot we initialize deferred pages on-demand, as needed, but once
300 * page_alloc_init_late() has finished, the deferred pages are all initialized,
301 * and we can permanently disable that path.
303 static DEFINE_STATIC_KEY_TRUE(deferred_pages);
306 * Calling kasan_free_pages() only after deferred memory initialization
307 * has completed. Poisoning pages during deferred memory init will greatly
308 * lengthen the process and cause problem in large memory systems as the
309 * deferred pages initialization is done with interrupt disabled.
311 * Assuming that there will be no reference to those newly initialized
312 * pages before they are ever allocated, this should have no effect on
313 * KASAN memory tracking as the poison will be properly inserted at page
314 * allocation time. The only corner case is when pages are allocated by
315 * on-demand allocation and then freed again before the deferred pages
316 * initialization is done, but this is not likely to happen.
318 static inline void kasan_free_nondeferred_pages(struct page *page, int order)
320 if (!static_branch_unlikely(&deferred_pages))
321 kasan_free_pages(page, order);
324 /* Returns true if the struct page for the pfn is uninitialised */
325 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
327 int nid = early_pfn_to_nid(pfn);
329 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
336 * Returns false when the remaining initialisation should be deferred until
337 * later in the boot cycle when it can be parallelised.
339 static inline bool update_defer_init(pg_data_t *pgdat,
340 unsigned long pfn, unsigned long zone_end,
341 unsigned long *nr_initialised)
343 /* Always populate low zones for address-constrained allocations */
344 if (zone_end < pgdat_end_pfn(pgdat))
347 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
348 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
349 pgdat->first_deferred_pfn = pfn;
356 #define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
358 static inline bool early_page_uninitialised(unsigned long pfn)
363 static inline bool update_defer_init(pg_data_t *pgdat,
364 unsigned long pfn, unsigned long zone_end,
365 unsigned long *nr_initialised)
371 /* Return a pointer to the bitmap storing bits affecting a block of pages */
372 static inline unsigned long *get_pageblock_bitmap(struct page *page,
375 #ifdef CONFIG_SPARSEMEM
376 return __pfn_to_section(pfn)->pageblock_flags;
378 return page_zone(page)->pageblock_flags;
379 #endif /* CONFIG_SPARSEMEM */
382 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
384 #ifdef CONFIG_SPARSEMEM
385 pfn &= (PAGES_PER_SECTION-1);
386 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
388 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
389 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
390 #endif /* CONFIG_SPARSEMEM */
394 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
395 * @page: The page within the block of interest
396 * @pfn: The target page frame number
397 * @end_bitidx: The last bit of interest to retrieve
398 * @mask: mask of bits that the caller is interested in
400 * Return: pageblock_bits flags
402 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
404 unsigned long end_bitidx,
407 unsigned long *bitmap;
408 unsigned long bitidx, word_bitidx;
411 bitmap = get_pageblock_bitmap(page, pfn);
412 bitidx = pfn_to_bitidx(page, pfn);
413 word_bitidx = bitidx / BITS_PER_LONG;
414 bitidx &= (BITS_PER_LONG-1);
416 word = bitmap[word_bitidx];
417 bitidx += end_bitidx;
418 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
421 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
422 unsigned long end_bitidx,
425 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
428 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
430 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
434 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
435 * @page: The page within the block of interest
436 * @flags: The flags to set
437 * @pfn: The target page frame number
438 * @end_bitidx: The last bit of interest
439 * @mask: mask of bits that the caller is interested in
441 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
443 unsigned long end_bitidx,
446 unsigned long *bitmap;
447 unsigned long bitidx, word_bitidx;
448 unsigned long old_word, word;
450 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
452 bitmap = get_pageblock_bitmap(page, pfn);
453 bitidx = pfn_to_bitidx(page, pfn);
454 word_bitidx = bitidx / BITS_PER_LONG;
455 bitidx &= (BITS_PER_LONG-1);
457 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
459 bitidx += end_bitidx;
460 mask <<= (BITS_PER_LONG - bitidx - 1);
461 flags <<= (BITS_PER_LONG - bitidx - 1);
463 word = READ_ONCE(bitmap[word_bitidx]);
465 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
466 if (word == old_word)
472 void set_pageblock_migratetype(struct page *page, int migratetype)
474 if (unlikely(page_group_by_mobility_disabled &&
475 migratetype < MIGRATE_PCPTYPES))
476 migratetype = MIGRATE_UNMOVABLE;
478 set_pageblock_flags_group(page, (unsigned long)migratetype,
479 PB_migrate, PB_migrate_end);
482 #ifdef CONFIG_DEBUG_VM
483 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
487 unsigned long pfn = page_to_pfn(page);
488 unsigned long sp, start_pfn;
491 seq = zone_span_seqbegin(zone);
492 start_pfn = zone->zone_start_pfn;
493 sp = zone->spanned_pages;
494 if (!zone_spans_pfn(zone, pfn))
496 } while (zone_span_seqretry(zone, seq));
499 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
500 pfn, zone_to_nid(zone), zone->name,
501 start_pfn, start_pfn + sp);
506 static int page_is_consistent(struct zone *zone, struct page *page)
508 if (!pfn_valid_within(page_to_pfn(page)))
510 if (zone != page_zone(page))
516 * Temporary debugging check for pages not lying within a given zone.
518 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
520 if (page_outside_zone_boundaries(zone, page))
522 if (!page_is_consistent(zone, page))
528 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
534 static void bad_page(struct page *page, const char *reason,
535 unsigned long bad_flags)
537 static unsigned long resume;
538 static unsigned long nr_shown;
539 static unsigned long nr_unshown;
542 * Allow a burst of 60 reports, then keep quiet for that minute;
543 * or allow a steady drip of one report per second.
545 if (nr_shown == 60) {
546 if (time_before(jiffies, resume)) {
552 "BUG: Bad page state: %lu messages suppressed\n",
559 resume = jiffies + 60 * HZ;
561 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
562 current->comm, page_to_pfn(page));
563 __dump_page(page, reason);
564 bad_flags &= page->flags;
566 pr_alert("bad because of flags: %#lx(%pGp)\n",
567 bad_flags, &bad_flags);
568 dump_page_owner(page);
573 /* Leave bad fields for debug, except PageBuddy could make trouble */
574 page_mapcount_reset(page); /* remove PageBuddy */
575 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
579 * Higher-order pages are called "compound pages". They are structured thusly:
581 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
583 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
584 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
586 * The first tail page's ->compound_dtor holds the offset in array of compound
587 * page destructors. See compound_page_dtors.
589 * The first tail page's ->compound_order holds the order of allocation.
590 * This usage means that zero-order pages may not be compound.
593 void free_compound_page(struct page *page)
595 __free_pages_ok(page, compound_order(page));
598 void prep_compound_page(struct page *page, unsigned int order)
601 int nr_pages = 1 << order;
603 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
604 set_compound_order(page, order);
606 for (i = 1; i < nr_pages; i++) {
607 struct page *p = page + i;
608 set_page_count(p, 0);
609 p->mapping = TAIL_MAPPING;
610 set_compound_head(p, page);
612 atomic_set(compound_mapcount_ptr(page), -1);
615 #ifdef CONFIG_DEBUG_PAGEALLOC
616 unsigned int _debug_guardpage_minorder;
617 bool _debug_pagealloc_enabled __read_mostly
618 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
619 EXPORT_SYMBOL(_debug_pagealloc_enabled);
620 bool _debug_guardpage_enabled __read_mostly;
622 static int __init early_debug_pagealloc(char *buf)
626 return kstrtobool(buf, &_debug_pagealloc_enabled);
628 early_param("debug_pagealloc", early_debug_pagealloc);
630 static bool need_debug_guardpage(void)
632 /* If we don't use debug_pagealloc, we don't need guard page */
633 if (!debug_pagealloc_enabled())
636 if (!debug_guardpage_minorder())
642 static void init_debug_guardpage(void)
644 if (!debug_pagealloc_enabled())
647 if (!debug_guardpage_minorder())
650 _debug_guardpage_enabled = true;
653 struct page_ext_operations debug_guardpage_ops = {
654 .need = need_debug_guardpage,
655 .init = init_debug_guardpage,
658 static int __init debug_guardpage_minorder_setup(char *buf)
662 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
663 pr_err("Bad debug_guardpage_minorder value\n");
666 _debug_guardpage_minorder = res;
667 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
670 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
672 static inline bool set_page_guard(struct zone *zone, struct page *page,
673 unsigned int order, int migratetype)
675 struct page_ext *page_ext;
677 if (!debug_guardpage_enabled())
680 if (order >= debug_guardpage_minorder())
683 page_ext = lookup_page_ext(page);
684 if (unlikely(!page_ext))
687 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
689 INIT_LIST_HEAD(&page->lru);
690 set_page_private(page, order);
691 /* Guard pages are not available for any usage */
692 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
697 static inline void clear_page_guard(struct zone *zone, struct page *page,
698 unsigned int order, int migratetype)
700 struct page_ext *page_ext;
702 if (!debug_guardpage_enabled())
705 page_ext = lookup_page_ext(page);
706 if (unlikely(!page_ext))
709 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
711 set_page_private(page, 0);
712 if (!is_migrate_isolate(migratetype))
713 __mod_zone_freepage_state(zone, (1 << order), migratetype);
716 struct page_ext_operations debug_guardpage_ops;
717 static inline bool set_page_guard(struct zone *zone, struct page *page,
718 unsigned int order, int migratetype) { return false; }
719 static inline void clear_page_guard(struct zone *zone, struct page *page,
720 unsigned int order, int migratetype) {}
723 static inline void set_page_order(struct page *page, unsigned int order)
725 set_page_private(page, order);
726 __SetPageBuddy(page);
729 static inline void rmv_page_order(struct page *page)
731 __ClearPageBuddy(page);
732 set_page_private(page, 0);
736 * This function checks whether a page is free && is the buddy
737 * we can coalesce a page and its buddy if
738 * (a) the buddy is not in a hole (check before calling!) &&
739 * (b) the buddy is in the buddy system &&
740 * (c) a page and its buddy have the same order &&
741 * (d) a page and its buddy are in the same zone.
743 * For recording whether a page is in the buddy system, we set PageBuddy.
744 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
746 * For recording page's order, we use page_private(page).
748 static inline int page_is_buddy(struct page *page, struct page *buddy,
751 if (page_is_guard(buddy) && page_order(buddy) == order) {
752 if (page_zone_id(page) != page_zone_id(buddy))
755 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
760 if (PageBuddy(buddy) && page_order(buddy) == order) {
762 * zone check is done late to avoid uselessly
763 * calculating zone/node ids for pages that could
766 if (page_zone_id(page) != page_zone_id(buddy))
769 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
777 * Freeing function for a buddy system allocator.
779 * The concept of a buddy system is to maintain direct-mapped table
780 * (containing bit values) for memory blocks of various "orders".
781 * The bottom level table contains the map for the smallest allocatable
782 * units of memory (here, pages), and each level above it describes
783 * pairs of units from the levels below, hence, "buddies".
784 * At a high level, all that happens here is marking the table entry
785 * at the bottom level available, and propagating the changes upward
786 * as necessary, plus some accounting needed to play nicely with other
787 * parts of the VM system.
788 * At each level, we keep a list of pages, which are heads of continuous
789 * free pages of length of (1 << order) and marked with PageBuddy.
790 * Page's order is recorded in page_private(page) field.
791 * So when we are allocating or freeing one, we can derive the state of the
792 * other. That is, if we allocate a small block, and both were
793 * free, the remainder of the region must be split into blocks.
794 * If a block is freed, and its buddy is also free, then this
795 * triggers coalescing into a block of larger size.
800 static inline void __free_one_page(struct page *page,
802 struct zone *zone, unsigned int order,
805 unsigned long combined_pfn;
806 unsigned long uninitialized_var(buddy_pfn);
808 unsigned int max_order;
810 max_order = min_t(unsigned int, MAX_ORDER - 1, pageblock_order);
812 VM_BUG_ON(!zone_is_initialized(zone));
813 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
815 VM_BUG_ON(migratetype == -1);
816 if (likely(!is_migrate_isolate(migratetype)))
817 __mod_zone_freepage_state(zone, 1 << order, migratetype);
819 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
820 VM_BUG_ON_PAGE(bad_range(zone, page), page);
823 while (order < max_order) {
824 buddy_pfn = __find_buddy_pfn(pfn, order);
825 buddy = page + (buddy_pfn - pfn);
827 if (!pfn_valid_within(buddy_pfn))
829 if (!page_is_buddy(page, buddy, order))
832 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
833 * merge with it and move up one order.
835 if (page_is_guard(buddy)) {
836 clear_page_guard(zone, buddy, order, migratetype);
838 list_del(&buddy->lru);
839 zone->free_area[order].nr_free--;
840 rmv_page_order(buddy);
842 combined_pfn = buddy_pfn & pfn;
843 page = page + (combined_pfn - pfn);
847 if (order < MAX_ORDER - 1) {
848 /* If we are here, it means order is >= pageblock_order.
849 * We want to prevent merge between freepages on isolate
850 * pageblock and normal pageblock. Without this, pageblock
851 * isolation could cause incorrect freepage or CMA accounting.
853 * We don't want to hit this code for the more frequent
856 if (unlikely(has_isolate_pageblock(zone))) {
859 buddy_pfn = __find_buddy_pfn(pfn, order);
860 buddy = page + (buddy_pfn - pfn);
861 buddy_mt = get_pageblock_migratetype(buddy);
863 if (migratetype != buddy_mt
864 && (is_migrate_isolate(migratetype) ||
865 is_migrate_isolate(buddy_mt)))
868 max_order = order + 1;
869 goto continue_merging;
873 set_page_order(page, order);
876 * If this is not the largest possible page, check if the buddy
877 * of the next-highest order is free. If it is, it's possible
878 * that pages are being freed that will coalesce soon. In case,
879 * that is happening, add the free page to the tail of the list
880 * so it's less likely to be used soon and more likely to be merged
881 * as a higher order page
883 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
884 struct page *higher_page, *higher_buddy;
885 combined_pfn = buddy_pfn & pfn;
886 higher_page = page + (combined_pfn - pfn);
887 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
888 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
889 if (pfn_valid_within(buddy_pfn) &&
890 page_is_buddy(higher_page, higher_buddy, order + 1)) {
891 list_add_tail(&page->lru,
892 &zone->free_area[order].free_list[migratetype]);
897 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
899 zone->free_area[order].nr_free++;
903 * A bad page could be due to a number of fields. Instead of multiple branches,
904 * try and check multiple fields with one check. The caller must do a detailed
905 * check if necessary.
907 static inline bool page_expected_state(struct page *page,
908 unsigned long check_flags)
910 if (unlikely(atomic_read(&page->_mapcount) != -1))
913 if (unlikely((unsigned long)page->mapping |
914 page_ref_count(page) |
916 (unsigned long)page->mem_cgroup |
918 (page->flags & check_flags)))
924 static void free_pages_check_bad(struct page *page)
926 const char *bad_reason;
927 unsigned long bad_flags;
932 if (unlikely(atomic_read(&page->_mapcount) != -1))
933 bad_reason = "nonzero mapcount";
934 if (unlikely(page->mapping != NULL))
935 bad_reason = "non-NULL mapping";
936 if (unlikely(page_ref_count(page) != 0))
937 bad_reason = "nonzero _refcount";
938 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
939 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
940 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
943 if (unlikely(page->mem_cgroup))
944 bad_reason = "page still charged to cgroup";
946 bad_page(page, bad_reason, bad_flags);
949 static inline int free_pages_check(struct page *page)
951 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
954 /* Something has gone sideways, find it */
955 free_pages_check_bad(page);
959 static int free_tail_pages_check(struct page *head_page, struct page *page)
964 * We rely page->lru.next never has bit 0 set, unless the page
965 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
967 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
969 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
973 switch (page - head_page) {
975 /* the first tail page: ->mapping may be compound_mapcount() */
976 if (unlikely(compound_mapcount(page))) {
977 bad_page(page, "nonzero compound_mapcount", 0);
983 * the second tail page: ->mapping is
984 * deferred_list.next -- ignore value.
988 if (page->mapping != TAIL_MAPPING) {
989 bad_page(page, "corrupted mapping in tail page", 0);
994 if (unlikely(!PageTail(page))) {
995 bad_page(page, "PageTail not set", 0);
998 if (unlikely(compound_head(page) != head_page)) {
999 bad_page(page, "compound_head not consistent", 0);
1004 page->mapping = NULL;
1005 clear_compound_head(page);
1009 static __always_inline bool free_pages_prepare(struct page *page,
1010 unsigned int order, bool check_free)
1014 VM_BUG_ON_PAGE(PageTail(page), page);
1016 trace_mm_page_free(page, order);
1019 * Check tail pages before head page information is cleared to
1020 * avoid checking PageCompound for order-0 pages.
1022 if (unlikely(order)) {
1023 bool compound = PageCompound(page);
1026 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1029 ClearPageDoubleMap(page);
1030 for (i = 1; i < (1 << order); i++) {
1032 bad += free_tail_pages_check(page, page + i);
1033 if (unlikely(free_pages_check(page + i))) {
1037 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1040 if (PageMappingFlags(page))
1041 page->mapping = NULL;
1042 if (memcg_kmem_enabled() && PageKmemcg(page))
1043 memcg_kmem_uncharge(page, order);
1045 bad += free_pages_check(page);
1049 page_cpupid_reset_last(page);
1050 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1051 reset_page_owner(page, order);
1053 if (!PageHighMem(page)) {
1054 debug_check_no_locks_freed(page_address(page),
1055 PAGE_SIZE << order);
1056 debug_check_no_obj_freed(page_address(page),
1057 PAGE_SIZE << order);
1059 arch_free_page(page, order);
1060 kernel_poison_pages(page, 1 << order, 0);
1061 kernel_map_pages(page, 1 << order, 0);
1062 kasan_free_nondeferred_pages(page, order);
1067 #ifdef CONFIG_DEBUG_VM
1068 static inline bool free_pcp_prepare(struct page *page)
1070 return free_pages_prepare(page, 0, true);
1073 static inline bool bulkfree_pcp_prepare(struct page *page)
1078 static bool free_pcp_prepare(struct page *page)
1080 return free_pages_prepare(page, 0, false);
1083 static bool bulkfree_pcp_prepare(struct page *page)
1085 return free_pages_check(page);
1087 #endif /* CONFIG_DEBUG_VM */
1089 static inline void prefetch_buddy(struct page *page)
1091 unsigned long pfn = page_to_pfn(page);
1092 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1093 struct page *buddy = page + (buddy_pfn - pfn);
1099 * Frees a number of pages from the PCP lists
1100 * Assumes all pages on list are in same zone, and of same order.
1101 * count is the number of pages to free.
1103 * If the zone was previously in an "all pages pinned" state then look to
1104 * see if this freeing clears that state.
1106 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1107 * pinned" detection logic.
1109 static void free_pcppages_bulk(struct zone *zone, int count,
1110 struct per_cpu_pages *pcp)
1112 int migratetype = 0;
1114 int prefetch_nr = 0;
1115 bool isolated_pageblocks;
1116 struct page *page, *tmp;
1120 * Ensure proper count is passed which otherwise would stuck in the
1121 * below while (list_empty(list)) loop.
1123 count = min(pcp->count, count);
1125 struct list_head *list;
1128 * Remove pages from lists in a round-robin fashion. A
1129 * batch_free count is maintained that is incremented when an
1130 * empty list is encountered. This is so more pages are freed
1131 * off fuller lists instead of spinning excessively around empty
1136 if (++migratetype == MIGRATE_PCPTYPES)
1138 list = &pcp->lists[migratetype];
1139 } while (list_empty(list));
1141 /* This is the only non-empty list. Free them all. */
1142 if (batch_free == MIGRATE_PCPTYPES)
1146 page = list_last_entry(list, struct page, lru);
1147 /* must delete to avoid corrupting pcp list */
1148 list_del(&page->lru);
1151 if (bulkfree_pcp_prepare(page))
1154 list_add_tail(&page->lru, &head);
1157 * We are going to put the page back to the global
1158 * pool, prefetch its buddy to speed up later access
1159 * under zone->lock. It is believed the overhead of
1160 * an additional test and calculating buddy_pfn here
1161 * can be offset by reduced memory latency later. To
1162 * avoid excessive prefetching due to large count, only
1163 * prefetch buddy for the first pcp->batch nr of pages.
1165 if (prefetch_nr++ < pcp->batch)
1166 prefetch_buddy(page);
1167 } while (--count && --batch_free && !list_empty(list));
1170 spin_lock(&zone->lock);
1171 isolated_pageblocks = has_isolate_pageblock(zone);
1174 * Use safe version since after __free_one_page(),
1175 * page->lru.next will not point to original list.
1177 list_for_each_entry_safe(page, tmp, &head, lru) {
1178 int mt = get_pcppage_migratetype(page);
1179 /* MIGRATE_ISOLATE page should not go to pcplists */
1180 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1181 /* Pageblock could have been isolated meanwhile */
1182 if (unlikely(isolated_pageblocks))
1183 mt = get_pageblock_migratetype(page);
1185 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1186 trace_mm_page_pcpu_drain(page, 0, mt);
1188 spin_unlock(&zone->lock);
1191 static void free_one_page(struct zone *zone,
1192 struct page *page, unsigned long pfn,
1196 spin_lock(&zone->lock);
1197 if (unlikely(has_isolate_pageblock(zone) ||
1198 is_migrate_isolate(migratetype))) {
1199 migratetype = get_pfnblock_migratetype(page, pfn);
1201 __free_one_page(page, pfn, zone, order, migratetype);
1202 spin_unlock(&zone->lock);
1205 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1206 unsigned long zone, int nid)
1208 mm_zero_struct_page(page);
1209 set_page_links(page, zone, nid, pfn);
1210 init_page_count(page);
1211 page_mapcount_reset(page);
1212 page_cpupid_reset_last(page);
1214 INIT_LIST_HEAD(&page->lru);
1215 #ifdef WANT_PAGE_VIRTUAL
1216 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1217 if (!is_highmem_idx(zone))
1218 set_page_address(page, __va(pfn << PAGE_SHIFT));
1222 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1223 static void __meminit init_reserved_page(unsigned long pfn)
1228 if (!early_page_uninitialised(pfn))
1231 nid = early_pfn_to_nid(pfn);
1232 pgdat = NODE_DATA(nid);
1234 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1235 struct zone *zone = &pgdat->node_zones[zid];
1237 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1240 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1243 static inline void init_reserved_page(unsigned long pfn)
1246 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1249 * Initialised pages do not have PageReserved set. This function is
1250 * called for each range allocated by the bootmem allocator and
1251 * marks the pages PageReserved. The remaining valid pages are later
1252 * sent to the buddy page allocator.
1254 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1256 unsigned long start_pfn = PFN_DOWN(start);
1257 unsigned long end_pfn = PFN_UP(end);
1259 for (; start_pfn < end_pfn; start_pfn++) {
1260 if (pfn_valid(start_pfn)) {
1261 struct page *page = pfn_to_page(start_pfn);
1263 init_reserved_page(start_pfn);
1265 /* Avoid false-positive PageTail() */
1266 INIT_LIST_HEAD(&page->lru);
1268 SetPageReserved(page);
1273 static void __free_pages_ok(struct page *page, unsigned int order)
1275 unsigned long flags;
1277 unsigned long pfn = page_to_pfn(page);
1279 if (!free_pages_prepare(page, order, true))
1282 migratetype = get_pfnblock_migratetype(page, pfn);
1283 local_irq_save(flags);
1284 __count_vm_events(PGFREE, 1 << order);
1285 free_one_page(page_zone(page), page, pfn, order, migratetype);
1286 local_irq_restore(flags);
1289 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1291 unsigned int nr_pages = 1 << order;
1292 struct page *p = page;
1296 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1298 __ClearPageReserved(p);
1299 set_page_count(p, 0);
1301 __ClearPageReserved(p);
1302 set_page_count(p, 0);
1304 page_zone(page)->managed_pages += nr_pages;
1305 set_page_refcounted(page);
1306 __free_pages(page, order);
1309 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1310 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1312 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1314 int __meminit early_pfn_to_nid(unsigned long pfn)
1316 static DEFINE_SPINLOCK(early_pfn_lock);
1319 spin_lock(&early_pfn_lock);
1320 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1322 nid = first_online_node;
1323 spin_unlock(&early_pfn_lock);
1329 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1330 static inline bool __meminit __maybe_unused
1331 meminit_pfn_in_nid(unsigned long pfn, int node,
1332 struct mminit_pfnnid_cache *state)
1336 nid = __early_pfn_to_nid(pfn, state);
1337 if (nid >= 0 && nid != node)
1342 /* Only safe to use early in boot when initialisation is single-threaded */
1343 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1345 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1350 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1354 static inline bool __meminit __maybe_unused
1355 meminit_pfn_in_nid(unsigned long pfn, int node,
1356 struct mminit_pfnnid_cache *state)
1363 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1366 if (early_page_uninitialised(pfn))
1368 return __free_pages_boot_core(page, order);
1372 * Check that the whole (or subset of) a pageblock given by the interval of
1373 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1374 * with the migration of free compaction scanner. The scanners then need to
1375 * use only pfn_valid_within() check for arches that allow holes within
1378 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1380 * It's possible on some configurations to have a setup like node0 node1 node0
1381 * i.e. it's possible that all pages within a zones range of pages do not
1382 * belong to a single zone. We assume that a border between node0 and node1
1383 * can occur within a single pageblock, but not a node0 node1 node0
1384 * interleaving within a single pageblock. It is therefore sufficient to check
1385 * the first and last page of a pageblock and avoid checking each individual
1386 * page in a pageblock.
1388 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1389 unsigned long end_pfn, struct zone *zone)
1391 struct page *start_page;
1392 struct page *end_page;
1394 /* end_pfn is one past the range we are checking */
1397 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1400 start_page = pfn_to_online_page(start_pfn);
1404 if (page_zone(start_page) != zone)
1407 end_page = pfn_to_page(end_pfn);
1409 /* This gives a shorter code than deriving page_zone(end_page) */
1410 if (page_zone_id(start_page) != page_zone_id(end_page))
1416 void set_zone_contiguous(struct zone *zone)
1418 unsigned long block_start_pfn = zone->zone_start_pfn;
1419 unsigned long block_end_pfn;
1421 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1422 for (; block_start_pfn < zone_end_pfn(zone);
1423 block_start_pfn = block_end_pfn,
1424 block_end_pfn += pageblock_nr_pages) {
1426 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1428 if (!__pageblock_pfn_to_page(block_start_pfn,
1429 block_end_pfn, zone))
1434 /* We confirm that there is no hole */
1435 zone->contiguous = true;
1438 void clear_zone_contiguous(struct zone *zone)
1440 zone->contiguous = false;
1443 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1444 static void __init deferred_free_range(unsigned long pfn,
1445 unsigned long nr_pages)
1453 page = pfn_to_page(pfn);
1455 /* Free a large naturally-aligned chunk if possible */
1456 if (nr_pages == pageblock_nr_pages &&
1457 (pfn & (pageblock_nr_pages - 1)) == 0) {
1458 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1459 __free_pages_boot_core(page, pageblock_order);
1463 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1464 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1465 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1466 __free_pages_boot_core(page, 0);
1470 /* Completion tracking for deferred_init_memmap() threads */
1471 static atomic_t pgdat_init_n_undone __initdata;
1472 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1474 static inline void __init pgdat_init_report_one_done(void)
1476 if (atomic_dec_and_test(&pgdat_init_n_undone))
1477 complete(&pgdat_init_all_done_comp);
1481 * Returns true if page needs to be initialized or freed to buddy allocator.
1483 * First we check if pfn is valid on architectures where it is possible to have
1484 * holes within pageblock_nr_pages. On systems where it is not possible, this
1485 * function is optimized out.
1487 * Then, we check if a current large page is valid by only checking the validity
1490 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1491 * within a node: a pfn is between start and end of a node, but does not belong
1492 * to this memory node.
1494 static inline bool __init
1495 deferred_pfn_valid(int nid, unsigned long pfn,
1496 struct mminit_pfnnid_cache *nid_init_state)
1498 if (!pfn_valid_within(pfn))
1500 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1502 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1508 * Free pages to buddy allocator. Try to free aligned pages in
1509 * pageblock_nr_pages sizes.
1511 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1512 unsigned long end_pfn)
1514 struct mminit_pfnnid_cache nid_init_state = { };
1515 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1516 unsigned long nr_free = 0;
1518 for (; pfn < end_pfn; pfn++) {
1519 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1520 deferred_free_range(pfn - nr_free, nr_free);
1522 } else if (!(pfn & nr_pgmask)) {
1523 deferred_free_range(pfn - nr_free, nr_free);
1525 touch_nmi_watchdog();
1530 /* Free the last block of pages to allocator */
1531 deferred_free_range(pfn - nr_free, nr_free);
1535 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1536 * by performing it only once every pageblock_nr_pages.
1537 * Return number of pages initialized.
1539 static unsigned long __init deferred_init_pages(int nid, int zid,
1541 unsigned long end_pfn)
1543 struct mminit_pfnnid_cache nid_init_state = { };
1544 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1545 unsigned long nr_pages = 0;
1546 struct page *page = NULL;
1548 for (; pfn < end_pfn; pfn++) {
1549 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1552 } else if (!page || !(pfn & nr_pgmask)) {
1553 page = pfn_to_page(pfn);
1554 touch_nmi_watchdog();
1558 __init_single_page(page, pfn, zid, nid);
1564 /* Initialise remaining memory on a node */
1565 static int __init deferred_init_memmap(void *data)
1567 pg_data_t *pgdat = data;
1568 int nid = pgdat->node_id;
1569 unsigned long start = jiffies;
1570 unsigned long nr_pages = 0;
1571 unsigned long spfn, epfn, first_init_pfn, flags;
1572 phys_addr_t spa, epa;
1575 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1578 /* Bind memory initialisation thread to a local node if possible */
1579 if (!cpumask_empty(cpumask))
1580 set_cpus_allowed_ptr(current, cpumask);
1582 pgdat_resize_lock(pgdat, &flags);
1583 first_init_pfn = pgdat->first_deferred_pfn;
1584 if (first_init_pfn == ULONG_MAX) {
1585 pgdat_resize_unlock(pgdat, &flags);
1586 pgdat_init_report_one_done();
1590 /* Sanity check boundaries */
1591 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1592 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1593 pgdat->first_deferred_pfn = ULONG_MAX;
1596 * Once we unlock here, the zone cannot be grown anymore, thus if an
1597 * interrupt thread must allocate this early in boot, zone must be
1598 * pre-grown prior to start of deferred page initialization.
1600 pgdat_resize_unlock(pgdat, &flags);
1602 /* Only the highest zone is deferred so find it */
1603 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1604 zone = pgdat->node_zones + zid;
1605 if (first_init_pfn < zone_end_pfn(zone))
1608 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1611 * Initialize and free pages. We do it in two loops: first we initialize
1612 * struct page, than free to buddy allocator, because while we are
1613 * freeing pages we can access pages that are ahead (computing buddy
1614 * page in __free_one_page()).
1616 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1617 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1618 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1619 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1621 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1622 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1623 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1624 deferred_free_pages(nid, zid, spfn, epfn);
1627 /* Sanity check that the next zone really is unpopulated */
1628 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1630 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1631 jiffies_to_msecs(jiffies - start));
1633 pgdat_init_report_one_done();
1638 * If this zone has deferred pages, try to grow it by initializing enough
1639 * deferred pages to satisfy the allocation specified by order, rounded up to
1640 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1641 * of SECTION_SIZE bytes by initializing struct pages in increments of
1642 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1644 * Return true when zone was grown, otherwise return false. We return true even
1645 * when we grow less than requested, to let the caller decide if there are
1646 * enough pages to satisfy the allocation.
1648 * Note: We use noinline because this function is needed only during boot, and
1649 * it is called from a __ref function _deferred_grow_zone. This way we are
1650 * making sure that it is not inlined into permanent text section.
1652 static noinline bool __init
1653 deferred_grow_zone(struct zone *zone, unsigned int order)
1655 int zid = zone_idx(zone);
1656 int nid = zone_to_nid(zone);
1657 pg_data_t *pgdat = NODE_DATA(nid);
1658 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1659 unsigned long nr_pages = 0;
1660 unsigned long first_init_pfn, spfn, epfn, t, flags;
1661 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1662 phys_addr_t spa, epa;
1665 /* Only the last zone may have deferred pages */
1666 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1669 pgdat_resize_lock(pgdat, &flags);
1672 * If someone grew this zone while we were waiting for spinlock, return
1673 * true, as there might be enough pages already.
1675 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1676 pgdat_resize_unlock(pgdat, &flags);
1680 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1682 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1683 pgdat_resize_unlock(pgdat, &flags);
1687 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1688 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1689 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1691 while (spfn < epfn && nr_pages < nr_pages_needed) {
1692 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1693 first_deferred_pfn = min(t, epfn);
1694 nr_pages += deferred_init_pages(nid, zid, spfn,
1695 first_deferred_pfn);
1696 spfn = first_deferred_pfn;
1699 if (nr_pages >= nr_pages_needed)
1703 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1704 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1705 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1706 deferred_free_pages(nid, zid, spfn, epfn);
1708 if (first_deferred_pfn == epfn)
1711 pgdat->first_deferred_pfn = first_deferred_pfn;
1712 pgdat_resize_unlock(pgdat, &flags);
1714 return nr_pages > 0;
1718 * deferred_grow_zone() is __init, but it is called from
1719 * get_page_from_freelist() during early boot until deferred_pages permanently
1720 * disables this call. This is why we have refdata wrapper to avoid warning,
1721 * and to ensure that the function body gets unloaded.
1724 _deferred_grow_zone(struct zone *zone, unsigned int order)
1726 return deferred_grow_zone(zone, order);
1729 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1731 void __init page_alloc_init_late(void)
1735 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1738 /* There will be num_node_state(N_MEMORY) threads */
1739 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1740 for_each_node_state(nid, N_MEMORY) {
1741 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1744 /* Block until all are initialised */
1745 wait_for_completion(&pgdat_init_all_done_comp);
1748 * The number of managed pages has changed due to the initialisation
1749 * so the pcpu batch and high limits needs to be updated or the limits
1750 * will be artificially small.
1752 for_each_populated_zone(zone)
1753 zone_pcp_update(zone);
1756 * We initialized the rest of the deferred pages. Permanently disable
1757 * on-demand struct page initialization.
1759 static_branch_disable(&deferred_pages);
1761 /* Reinit limits that are based on free pages after the kernel is up */
1762 files_maxfiles_init();
1764 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1765 /* Discard memblock private memory */
1769 for_each_populated_zone(zone)
1770 set_zone_contiguous(zone);
1774 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1775 void __init init_cma_reserved_pageblock(struct page *page)
1777 unsigned i = pageblock_nr_pages;
1778 struct page *p = page;
1781 __ClearPageReserved(p);
1782 set_page_count(p, 0);
1785 set_pageblock_migratetype(page, MIGRATE_CMA);
1787 if (pageblock_order >= MAX_ORDER) {
1788 i = pageblock_nr_pages;
1791 set_page_refcounted(p);
1792 __free_pages(p, MAX_ORDER - 1);
1793 p += MAX_ORDER_NR_PAGES;
1794 } while (i -= MAX_ORDER_NR_PAGES);
1796 set_page_refcounted(page);
1797 __free_pages(page, pageblock_order);
1800 adjust_managed_page_count(page, pageblock_nr_pages);
1805 * The order of subdivision here is critical for the IO subsystem.
1806 * Please do not alter this order without good reasons and regression
1807 * testing. Specifically, as large blocks of memory are subdivided,
1808 * the order in which smaller blocks are delivered depends on the order
1809 * they're subdivided in this function. This is the primary factor
1810 * influencing the order in which pages are delivered to the IO
1811 * subsystem according to empirical testing, and this is also justified
1812 * by considering the behavior of a buddy system containing a single
1813 * large block of memory acted on by a series of small allocations.
1814 * This behavior is a critical factor in sglist merging's success.
1818 static inline void expand(struct zone *zone, struct page *page,
1819 int low, int high, struct free_area *area,
1822 unsigned long size = 1 << high;
1824 while (high > low) {
1828 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1831 * Mark as guard pages (or page), that will allow to
1832 * merge back to allocator when buddy will be freed.
1833 * Corresponding page table entries will not be touched,
1834 * pages will stay not present in virtual address space
1836 if (set_page_guard(zone, &page[size], high, migratetype))
1839 list_add(&page[size].lru, &area->free_list[migratetype]);
1841 set_page_order(&page[size], high);
1845 static void check_new_page_bad(struct page *page)
1847 const char *bad_reason = NULL;
1848 unsigned long bad_flags = 0;
1850 if (unlikely(atomic_read(&page->_mapcount) != -1))
1851 bad_reason = "nonzero mapcount";
1852 if (unlikely(page->mapping != NULL))
1853 bad_reason = "non-NULL mapping";
1854 if (unlikely(page_ref_count(page) != 0))
1855 bad_reason = "nonzero _count";
1856 if (unlikely(page->flags & __PG_HWPOISON)) {
1857 bad_reason = "HWPoisoned (hardware-corrupted)";
1858 bad_flags = __PG_HWPOISON;
1859 /* Don't complain about hwpoisoned pages */
1860 page_mapcount_reset(page); /* remove PageBuddy */
1863 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1864 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1865 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1868 if (unlikely(page->mem_cgroup))
1869 bad_reason = "page still charged to cgroup";
1871 bad_page(page, bad_reason, bad_flags);
1875 * This page is about to be returned from the page allocator
1877 static inline int check_new_page(struct page *page)
1879 if (likely(page_expected_state(page,
1880 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1883 check_new_page_bad(page);
1887 static inline bool free_pages_prezeroed(void)
1889 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1890 page_poisoning_enabled();
1893 #ifdef CONFIG_DEBUG_VM
1894 static bool check_pcp_refill(struct page *page)
1899 static bool check_new_pcp(struct page *page)
1901 return check_new_page(page);
1904 static bool check_pcp_refill(struct page *page)
1906 return check_new_page(page);
1908 static bool check_new_pcp(struct page *page)
1912 #endif /* CONFIG_DEBUG_VM */
1914 static bool check_new_pages(struct page *page, unsigned int order)
1917 for (i = 0; i < (1 << order); i++) {
1918 struct page *p = page + i;
1920 if (unlikely(check_new_page(p)))
1927 inline void post_alloc_hook(struct page *page, unsigned int order,
1930 set_page_private(page, 0);
1931 set_page_refcounted(page);
1933 arch_alloc_page(page, order);
1934 kernel_map_pages(page, 1 << order, 1);
1935 kasan_alloc_pages(page, order);
1936 kernel_poison_pages(page, 1 << order, 1);
1937 set_page_owner(page, order, gfp_flags);
1940 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1941 unsigned int alloc_flags)
1945 post_alloc_hook(page, order, gfp_flags);
1947 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1948 for (i = 0; i < (1 << order); i++)
1949 clear_highpage(page + i);
1951 if (order && (gfp_flags & __GFP_COMP))
1952 prep_compound_page(page, order);
1955 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1956 * allocate the page. The expectation is that the caller is taking
1957 * steps that will free more memory. The caller should avoid the page
1958 * being used for !PFMEMALLOC purposes.
1960 if (alloc_flags & ALLOC_NO_WATERMARKS)
1961 set_page_pfmemalloc(page);
1963 clear_page_pfmemalloc(page);
1967 * Go through the free lists for the given migratetype and remove
1968 * the smallest available page from the freelists
1970 static __always_inline
1971 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1974 unsigned int current_order;
1975 struct free_area *area;
1978 /* Find a page of the appropriate size in the preferred list */
1979 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1980 area = &(zone->free_area[current_order]);
1981 page = list_first_entry_or_null(&area->free_list[migratetype],
1985 list_del(&page->lru);
1986 rmv_page_order(page);
1988 expand(zone, page, order, current_order, area, migratetype);
1989 set_pcppage_migratetype(page, migratetype);
1998 * This array describes the order lists are fallen back to when
1999 * the free lists for the desirable migrate type are depleted
2001 static int fallbacks[MIGRATE_TYPES][4] = {
2002 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2003 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2004 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2006 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2008 #ifdef CONFIG_MEMORY_ISOLATION
2009 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2014 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2017 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2020 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2021 unsigned int order) { return NULL; }
2025 * Move the free pages in a range to the free lists of the requested type.
2026 * Note that start_page and end_pages are not aligned on a pageblock
2027 * boundary. If alignment is required, use move_freepages_block()
2029 static int move_freepages(struct zone *zone,
2030 struct page *start_page, struct page *end_page,
2031 int migratetype, int *num_movable)
2035 int pages_moved = 0;
2037 #ifndef CONFIG_HOLES_IN_ZONE
2039 * page_zone is not safe to call in this context when
2040 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2041 * anyway as we check zone boundaries in move_freepages_block().
2042 * Remove at a later date when no bug reports exist related to
2043 * grouping pages by mobility
2045 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2046 pfn_valid(page_to_pfn(end_page)) &&
2047 page_zone(start_page) != page_zone(end_page));
2053 for (page = start_page; page <= end_page;) {
2054 if (!pfn_valid_within(page_to_pfn(page))) {
2059 /* Make sure we are not inadvertently changing nodes */
2060 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2062 if (!PageBuddy(page)) {
2064 * We assume that pages that could be isolated for
2065 * migration are movable. But we don't actually try
2066 * isolating, as that would be expensive.
2069 (PageLRU(page) || __PageMovable(page)))
2076 order = page_order(page);
2077 list_move(&page->lru,
2078 &zone->free_area[order].free_list[migratetype]);
2080 pages_moved += 1 << order;
2086 int move_freepages_block(struct zone *zone, struct page *page,
2087 int migratetype, int *num_movable)
2089 unsigned long start_pfn, end_pfn;
2090 struct page *start_page, *end_page;
2092 start_pfn = page_to_pfn(page);
2093 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2094 start_page = pfn_to_page(start_pfn);
2095 end_page = start_page + pageblock_nr_pages - 1;
2096 end_pfn = start_pfn + pageblock_nr_pages - 1;
2098 /* Do not cross zone boundaries */
2099 if (!zone_spans_pfn(zone, start_pfn))
2101 if (!zone_spans_pfn(zone, end_pfn))
2104 return move_freepages(zone, start_page, end_page, migratetype,
2108 static void change_pageblock_range(struct page *pageblock_page,
2109 int start_order, int migratetype)
2111 int nr_pageblocks = 1 << (start_order - pageblock_order);
2113 while (nr_pageblocks--) {
2114 set_pageblock_migratetype(pageblock_page, migratetype);
2115 pageblock_page += pageblock_nr_pages;
2120 * When we are falling back to another migratetype during allocation, try to
2121 * steal extra free pages from the same pageblocks to satisfy further
2122 * allocations, instead of polluting multiple pageblocks.
2124 * If we are stealing a relatively large buddy page, it is likely there will
2125 * be more free pages in the pageblock, so try to steal them all. For
2126 * reclaimable and unmovable allocations, we steal regardless of page size,
2127 * as fragmentation caused by those allocations polluting movable pageblocks
2128 * is worse than movable allocations stealing from unmovable and reclaimable
2131 static bool can_steal_fallback(unsigned int order, int start_mt)
2134 * Leaving this order check is intended, although there is
2135 * relaxed order check in next check. The reason is that
2136 * we can actually steal whole pageblock if this condition met,
2137 * but, below check doesn't guarantee it and that is just heuristic
2138 * so could be changed anytime.
2140 if (order >= pageblock_order)
2143 if (order >= pageblock_order / 2 ||
2144 start_mt == MIGRATE_RECLAIMABLE ||
2145 start_mt == MIGRATE_UNMOVABLE ||
2146 page_group_by_mobility_disabled)
2153 * This function implements actual steal behaviour. If order is large enough,
2154 * we can steal whole pageblock. If not, we first move freepages in this
2155 * pageblock to our migratetype and determine how many already-allocated pages
2156 * are there in the pageblock with a compatible migratetype. If at least half
2157 * of pages are free or compatible, we can change migratetype of the pageblock
2158 * itself, so pages freed in the future will be put on the correct free list.
2160 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2161 int start_type, bool whole_block)
2163 unsigned int current_order = page_order(page);
2164 struct free_area *area;
2165 int free_pages, movable_pages, alike_pages;
2168 old_block_type = get_pageblock_migratetype(page);
2171 * This can happen due to races and we want to prevent broken
2172 * highatomic accounting.
2174 if (is_migrate_highatomic(old_block_type))
2177 /* Take ownership for orders >= pageblock_order */
2178 if (current_order >= pageblock_order) {
2179 change_pageblock_range(page, current_order, start_type);
2183 /* We are not allowed to try stealing from the whole block */
2187 free_pages = move_freepages_block(zone, page, start_type,
2190 * Determine how many pages are compatible with our allocation.
2191 * For movable allocation, it's the number of movable pages which
2192 * we just obtained. For other types it's a bit more tricky.
2194 if (start_type == MIGRATE_MOVABLE) {
2195 alike_pages = movable_pages;
2198 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2199 * to MOVABLE pageblock, consider all non-movable pages as
2200 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2201 * vice versa, be conservative since we can't distinguish the
2202 * exact migratetype of non-movable pages.
2204 if (old_block_type == MIGRATE_MOVABLE)
2205 alike_pages = pageblock_nr_pages
2206 - (free_pages + movable_pages);
2211 /* moving whole block can fail due to zone boundary conditions */
2216 * If a sufficient number of pages in the block are either free or of
2217 * comparable migratability as our allocation, claim the whole block.
2219 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2220 page_group_by_mobility_disabled)
2221 set_pageblock_migratetype(page, start_type);
2226 area = &zone->free_area[current_order];
2227 list_move(&page->lru, &area->free_list[start_type]);
2231 * Check whether there is a suitable fallback freepage with requested order.
2232 * If only_stealable is true, this function returns fallback_mt only if
2233 * we can steal other freepages all together. This would help to reduce
2234 * fragmentation due to mixed migratetype pages in one pageblock.
2236 int find_suitable_fallback(struct free_area *area, unsigned int order,
2237 int migratetype, bool only_stealable, bool *can_steal)
2242 if (area->nr_free == 0)
2247 fallback_mt = fallbacks[migratetype][i];
2248 if (fallback_mt == MIGRATE_TYPES)
2251 if (list_empty(&area->free_list[fallback_mt]))
2254 if (can_steal_fallback(order, migratetype))
2257 if (!only_stealable)
2268 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2269 * there are no empty page blocks that contain a page with a suitable order
2271 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2272 unsigned int alloc_order)
2275 unsigned long max_managed, flags;
2278 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2279 * Check is race-prone but harmless.
2281 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2282 if (zone->nr_reserved_highatomic >= max_managed)
2285 spin_lock_irqsave(&zone->lock, flags);
2287 /* Recheck the nr_reserved_highatomic limit under the lock */
2288 if (zone->nr_reserved_highatomic >= max_managed)
2292 mt = get_pageblock_migratetype(page);
2293 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2294 && !is_migrate_cma(mt)) {
2295 zone->nr_reserved_highatomic += pageblock_nr_pages;
2296 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2297 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2301 spin_unlock_irqrestore(&zone->lock, flags);
2305 * Used when an allocation is about to fail under memory pressure. This
2306 * potentially hurts the reliability of high-order allocations when under
2307 * intense memory pressure but failed atomic allocations should be easier
2308 * to recover from than an OOM.
2310 * If @force is true, try to unreserve a pageblock even though highatomic
2311 * pageblock is exhausted.
2313 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2316 struct zonelist *zonelist = ac->zonelist;
2317 unsigned long flags;
2324 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2327 * Preserve at least one pageblock unless memory pressure
2330 if (!force && zone->nr_reserved_highatomic <=
2334 spin_lock_irqsave(&zone->lock, flags);
2335 for (order = 0; order < MAX_ORDER; order++) {
2336 struct free_area *area = &(zone->free_area[order]);
2338 page = list_first_entry_or_null(
2339 &area->free_list[MIGRATE_HIGHATOMIC],
2345 * In page freeing path, migratetype change is racy so
2346 * we can counter several free pages in a pageblock
2347 * in this loop althoug we changed the pageblock type
2348 * from highatomic to ac->migratetype. So we should
2349 * adjust the count once.
2351 if (is_migrate_highatomic_page(page)) {
2353 * It should never happen but changes to
2354 * locking could inadvertently allow a per-cpu
2355 * drain to add pages to MIGRATE_HIGHATOMIC
2356 * while unreserving so be safe and watch for
2359 zone->nr_reserved_highatomic -= min(
2361 zone->nr_reserved_highatomic);
2365 * Convert to ac->migratetype and avoid the normal
2366 * pageblock stealing heuristics. Minimally, the caller
2367 * is doing the work and needs the pages. More
2368 * importantly, if the block was always converted to
2369 * MIGRATE_UNMOVABLE or another type then the number
2370 * of pageblocks that cannot be completely freed
2373 set_pageblock_migratetype(page, ac->migratetype);
2374 ret = move_freepages_block(zone, page, ac->migratetype,
2377 spin_unlock_irqrestore(&zone->lock, flags);
2381 spin_unlock_irqrestore(&zone->lock, flags);
2388 * Try finding a free buddy page on the fallback list and put it on the free
2389 * list of requested migratetype, possibly along with other pages from the same
2390 * block, depending on fragmentation avoidance heuristics. Returns true if
2391 * fallback was found so that __rmqueue_smallest() can grab it.
2393 * The use of signed ints for order and current_order is a deliberate
2394 * deviation from the rest of this file, to make the for loop
2395 * condition simpler.
2397 static __always_inline bool
2398 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2400 struct free_area *area;
2407 * Find the largest available free page in the other list. This roughly
2408 * approximates finding the pageblock with the most free pages, which
2409 * would be too costly to do exactly.
2411 for (current_order = MAX_ORDER - 1; current_order >= order;
2413 area = &(zone->free_area[current_order]);
2414 fallback_mt = find_suitable_fallback(area, current_order,
2415 start_migratetype, false, &can_steal);
2416 if (fallback_mt == -1)
2420 * We cannot steal all free pages from the pageblock and the
2421 * requested migratetype is movable. In that case it's better to
2422 * steal and split the smallest available page instead of the
2423 * largest available page, because even if the next movable
2424 * allocation falls back into a different pageblock than this
2425 * one, it won't cause permanent fragmentation.
2427 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2428 && current_order > order)
2437 for (current_order = order; current_order < MAX_ORDER;
2439 area = &(zone->free_area[current_order]);
2440 fallback_mt = find_suitable_fallback(area, current_order,
2441 start_migratetype, false, &can_steal);
2442 if (fallback_mt != -1)
2447 * This should not happen - we already found a suitable fallback
2448 * when looking for the largest page.
2450 VM_BUG_ON(current_order == MAX_ORDER);
2453 page = list_first_entry(&area->free_list[fallback_mt],
2456 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2458 trace_mm_page_alloc_extfrag(page, order, current_order,
2459 start_migratetype, fallback_mt);
2466 * Do the hard work of removing an element from the buddy allocator.
2467 * Call me with the zone->lock already held.
2469 static __always_inline struct page *
2470 __rmqueue(struct zone *zone, unsigned int order, int migratetype)
2475 page = __rmqueue_smallest(zone, order, migratetype);
2476 if (unlikely(!page)) {
2477 if (migratetype == MIGRATE_MOVABLE)
2478 page = __rmqueue_cma_fallback(zone, order);
2480 if (!page && __rmqueue_fallback(zone, order, migratetype))
2484 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2489 * Obtain a specified number of elements from the buddy allocator, all under
2490 * a single hold of the lock, for efficiency. Add them to the supplied list.
2491 * Returns the number of new pages which were placed at *list.
2493 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2494 unsigned long count, struct list_head *list,
2499 spin_lock(&zone->lock);
2500 for (i = 0; i < count; ++i) {
2501 struct page *page = __rmqueue(zone, order, migratetype);
2502 if (unlikely(page == NULL))
2505 if (unlikely(check_pcp_refill(page)))
2509 * Split buddy pages returned by expand() are received here in
2510 * physical page order. The page is added to the tail of
2511 * caller's list. From the callers perspective, the linked list
2512 * is ordered by page number under some conditions. This is
2513 * useful for IO devices that can forward direction from the
2514 * head, thus also in the physical page order. This is useful
2515 * for IO devices that can merge IO requests if the physical
2516 * pages are ordered properly.
2518 list_add_tail(&page->lru, list);
2520 if (is_migrate_cma(get_pcppage_migratetype(page)))
2521 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2526 * i pages were removed from the buddy list even if some leak due
2527 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2528 * on i. Do not confuse with 'alloced' which is the number of
2529 * pages added to the pcp list.
2531 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2532 spin_unlock(&zone->lock);
2538 * Called from the vmstat counter updater to drain pagesets of this
2539 * currently executing processor on remote nodes after they have
2542 * Note that this function must be called with the thread pinned to
2543 * a single processor.
2545 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2547 unsigned long flags;
2548 int to_drain, batch;
2550 local_irq_save(flags);
2551 batch = READ_ONCE(pcp->batch);
2552 to_drain = min(pcp->count, batch);
2554 free_pcppages_bulk(zone, to_drain, pcp);
2555 local_irq_restore(flags);
2560 * Drain pcplists of the indicated processor and zone.
2562 * The processor must either be the current processor and the
2563 * thread pinned to the current processor or a processor that
2566 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2568 unsigned long flags;
2569 struct per_cpu_pageset *pset;
2570 struct per_cpu_pages *pcp;
2572 local_irq_save(flags);
2573 pset = per_cpu_ptr(zone->pageset, cpu);
2577 free_pcppages_bulk(zone, pcp->count, pcp);
2578 local_irq_restore(flags);
2582 * Drain pcplists of all zones on the indicated processor.
2584 * The processor must either be the current processor and the
2585 * thread pinned to the current processor or a processor that
2588 static void drain_pages(unsigned int cpu)
2592 for_each_populated_zone(zone) {
2593 drain_pages_zone(cpu, zone);
2598 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2600 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2601 * the single zone's pages.
2603 void drain_local_pages(struct zone *zone)
2605 int cpu = smp_processor_id();
2608 drain_pages_zone(cpu, zone);
2613 static void drain_local_pages_wq(struct work_struct *work)
2616 * drain_all_pages doesn't use proper cpu hotplug protection so
2617 * we can race with cpu offline when the WQ can move this from
2618 * a cpu pinned worker to an unbound one. We can operate on a different
2619 * cpu which is allright but we also have to make sure to not move to
2623 drain_local_pages(NULL);
2628 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2630 * When zone parameter is non-NULL, spill just the single zone's pages.
2632 * Note that this can be extremely slow as the draining happens in a workqueue.
2634 void drain_all_pages(struct zone *zone)
2639 * Allocate in the BSS so we wont require allocation in
2640 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2642 static cpumask_t cpus_with_pcps;
2645 * Make sure nobody triggers this path before mm_percpu_wq is fully
2648 if (WARN_ON_ONCE(!mm_percpu_wq))
2652 * Do not drain if one is already in progress unless it's specific to
2653 * a zone. Such callers are primarily CMA and memory hotplug and need
2654 * the drain to be complete when the call returns.
2656 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2659 mutex_lock(&pcpu_drain_mutex);
2663 * We don't care about racing with CPU hotplug event
2664 * as offline notification will cause the notified
2665 * cpu to drain that CPU pcps and on_each_cpu_mask
2666 * disables preemption as part of its processing
2668 for_each_online_cpu(cpu) {
2669 struct per_cpu_pageset *pcp;
2671 bool has_pcps = false;
2674 pcp = per_cpu_ptr(zone->pageset, cpu);
2678 for_each_populated_zone(z) {
2679 pcp = per_cpu_ptr(z->pageset, cpu);
2680 if (pcp->pcp.count) {
2688 cpumask_set_cpu(cpu, &cpus_with_pcps);
2690 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2693 for_each_cpu(cpu, &cpus_with_pcps) {
2694 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2695 INIT_WORK(work, drain_local_pages_wq);
2696 queue_work_on(cpu, mm_percpu_wq, work);
2698 for_each_cpu(cpu, &cpus_with_pcps)
2699 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2701 mutex_unlock(&pcpu_drain_mutex);
2704 #ifdef CONFIG_HIBERNATION
2707 * Touch the watchdog for every WD_PAGE_COUNT pages.
2709 #define WD_PAGE_COUNT (128*1024)
2711 void mark_free_pages(struct zone *zone)
2713 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2714 unsigned long flags;
2715 unsigned int order, t;
2718 if (zone_is_empty(zone))
2721 spin_lock_irqsave(&zone->lock, flags);
2723 max_zone_pfn = zone_end_pfn(zone);
2724 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2725 if (pfn_valid(pfn)) {
2726 page = pfn_to_page(pfn);
2728 if (!--page_count) {
2729 touch_nmi_watchdog();
2730 page_count = WD_PAGE_COUNT;
2733 if (page_zone(page) != zone)
2736 if (!swsusp_page_is_forbidden(page))
2737 swsusp_unset_page_free(page);
2740 for_each_migratetype_order(order, t) {
2741 list_for_each_entry(page,
2742 &zone->free_area[order].free_list[t], lru) {
2745 pfn = page_to_pfn(page);
2746 for (i = 0; i < (1UL << order); i++) {
2747 if (!--page_count) {
2748 touch_nmi_watchdog();
2749 page_count = WD_PAGE_COUNT;
2751 swsusp_set_page_free(pfn_to_page(pfn + i));
2755 spin_unlock_irqrestore(&zone->lock, flags);
2757 #endif /* CONFIG_PM */
2759 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2763 if (!free_pcp_prepare(page))
2766 migratetype = get_pfnblock_migratetype(page, pfn);
2767 set_pcppage_migratetype(page, migratetype);
2771 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2773 struct zone *zone = page_zone(page);
2774 struct per_cpu_pages *pcp;
2777 migratetype = get_pcppage_migratetype(page);
2778 __count_vm_event(PGFREE);
2781 * We only track unmovable, reclaimable and movable on pcp lists.
2782 * Free ISOLATE pages back to the allocator because they are being
2783 * offlined but treat HIGHATOMIC as movable pages so we can get those
2784 * areas back if necessary. Otherwise, we may have to free
2785 * excessively into the page allocator
2787 if (migratetype >= MIGRATE_PCPTYPES) {
2788 if (unlikely(is_migrate_isolate(migratetype))) {
2789 free_one_page(zone, page, pfn, 0, migratetype);
2792 migratetype = MIGRATE_MOVABLE;
2795 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2796 list_add(&page->lru, &pcp->lists[migratetype]);
2798 if (pcp->count >= pcp->high) {
2799 unsigned long batch = READ_ONCE(pcp->batch);
2800 free_pcppages_bulk(zone, batch, pcp);
2805 * Free a 0-order page
2807 void free_unref_page(struct page *page)
2809 unsigned long flags;
2810 unsigned long pfn = page_to_pfn(page);
2812 if (!free_unref_page_prepare(page, pfn))
2815 local_irq_save(flags);
2816 free_unref_page_commit(page, pfn);
2817 local_irq_restore(flags);
2821 * Free a list of 0-order pages
2823 void free_unref_page_list(struct list_head *list)
2825 struct page *page, *next;
2826 unsigned long flags, pfn;
2827 int batch_count = 0;
2829 /* Prepare pages for freeing */
2830 list_for_each_entry_safe(page, next, list, lru) {
2831 pfn = page_to_pfn(page);
2832 if (!free_unref_page_prepare(page, pfn))
2833 list_del(&page->lru);
2834 set_page_private(page, pfn);
2837 local_irq_save(flags);
2838 list_for_each_entry_safe(page, next, list, lru) {
2839 unsigned long pfn = page_private(page);
2841 set_page_private(page, 0);
2842 trace_mm_page_free_batched(page);
2843 free_unref_page_commit(page, pfn);
2846 * Guard against excessive IRQ disabled times when we get
2847 * a large list of pages to free.
2849 if (++batch_count == SWAP_CLUSTER_MAX) {
2850 local_irq_restore(flags);
2852 local_irq_save(flags);
2855 local_irq_restore(flags);
2859 * split_page takes a non-compound higher-order page, and splits it into
2860 * n (1<<order) sub-pages: page[0..n]
2861 * Each sub-page must be freed individually.
2863 * Note: this is probably too low level an operation for use in drivers.
2864 * Please consult with lkml before using this in your driver.
2866 void split_page(struct page *page, unsigned int order)
2870 VM_BUG_ON_PAGE(PageCompound(page), page);
2871 VM_BUG_ON_PAGE(!page_count(page), page);
2873 for (i = 1; i < (1 << order); i++)
2874 set_page_refcounted(page + i);
2875 split_page_owner(page, order);
2877 EXPORT_SYMBOL_GPL(split_page);
2879 int __isolate_free_page(struct page *page, unsigned int order)
2881 unsigned long watermark;
2885 BUG_ON(!PageBuddy(page));
2887 zone = page_zone(page);
2888 mt = get_pageblock_migratetype(page);
2890 if (!is_migrate_isolate(mt)) {
2892 * Obey watermarks as if the page was being allocated. We can
2893 * emulate a high-order watermark check with a raised order-0
2894 * watermark, because we already know our high-order page
2897 watermark = min_wmark_pages(zone) + (1UL << order);
2898 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2901 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2904 /* Remove page from free list */
2905 list_del(&page->lru);
2906 zone->free_area[order].nr_free--;
2907 rmv_page_order(page);
2910 * Set the pageblock if the isolated page is at least half of a
2913 if (order >= pageblock_order - 1) {
2914 struct page *endpage = page + (1 << order) - 1;
2915 for (; page < endpage; page += pageblock_nr_pages) {
2916 int mt = get_pageblock_migratetype(page);
2917 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2918 && !is_migrate_highatomic(mt))
2919 set_pageblock_migratetype(page,
2925 return 1UL << order;
2929 * Update NUMA hit/miss statistics
2931 * Must be called with interrupts disabled.
2933 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2936 enum numa_stat_item local_stat = NUMA_LOCAL;
2938 /* skip numa counters update if numa stats is disabled */
2939 if (!static_branch_likely(&vm_numa_stat_key))
2942 if (zone_to_nid(z) != numa_node_id())
2943 local_stat = NUMA_OTHER;
2945 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2946 __inc_numa_state(z, NUMA_HIT);
2948 __inc_numa_state(z, NUMA_MISS);
2949 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2951 __inc_numa_state(z, local_stat);
2955 /* Remove page from the per-cpu list, caller must protect the list */
2956 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2957 struct per_cpu_pages *pcp,
2958 struct list_head *list)
2963 if (list_empty(list)) {
2964 pcp->count += rmqueue_bulk(zone, 0,
2967 if (unlikely(list_empty(list)))
2971 page = list_first_entry(list, struct page, lru);
2972 list_del(&page->lru);
2974 } while (check_new_pcp(page));
2979 /* Lock and remove page from the per-cpu list */
2980 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2981 struct zone *zone, unsigned int order,
2982 gfp_t gfp_flags, int migratetype)
2984 struct per_cpu_pages *pcp;
2985 struct list_head *list;
2987 unsigned long flags;
2989 local_irq_save(flags);
2990 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2991 list = &pcp->lists[migratetype];
2992 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2994 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2995 zone_statistics(preferred_zone, zone);
2997 local_irq_restore(flags);
3002 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3005 struct page *rmqueue(struct zone *preferred_zone,
3006 struct zone *zone, unsigned int order,
3007 gfp_t gfp_flags, unsigned int alloc_flags,
3010 unsigned long flags;
3013 if (likely(order == 0)) {
3014 page = rmqueue_pcplist(preferred_zone, zone, order,
3015 gfp_flags, migratetype);
3020 * We most definitely don't want callers attempting to
3021 * allocate greater than order-1 page units with __GFP_NOFAIL.
3023 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3024 spin_lock_irqsave(&zone->lock, flags);
3028 if (alloc_flags & ALLOC_HARDER) {
3029 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3031 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3034 page = __rmqueue(zone, order, migratetype);
3035 } while (page && check_new_pages(page, order));
3036 spin_unlock(&zone->lock);
3039 __mod_zone_freepage_state(zone, -(1 << order),
3040 get_pcppage_migratetype(page));
3042 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3043 zone_statistics(preferred_zone, zone);
3044 local_irq_restore(flags);
3047 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3051 local_irq_restore(flags);
3055 #ifdef CONFIG_FAIL_PAGE_ALLOC
3058 struct fault_attr attr;
3060 bool ignore_gfp_highmem;
3061 bool ignore_gfp_reclaim;
3063 } fail_page_alloc = {
3064 .attr = FAULT_ATTR_INITIALIZER,
3065 .ignore_gfp_reclaim = true,
3066 .ignore_gfp_highmem = true,
3070 static int __init setup_fail_page_alloc(char *str)
3072 return setup_fault_attr(&fail_page_alloc.attr, str);
3074 __setup("fail_page_alloc=", setup_fail_page_alloc);
3076 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3078 if (order < fail_page_alloc.min_order)
3080 if (gfp_mask & __GFP_NOFAIL)
3082 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3084 if (fail_page_alloc.ignore_gfp_reclaim &&
3085 (gfp_mask & __GFP_DIRECT_RECLAIM))
3088 return should_fail(&fail_page_alloc.attr, 1 << order);
3091 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3093 static int __init fail_page_alloc_debugfs(void)
3095 umode_t mode = S_IFREG | 0600;
3098 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3099 &fail_page_alloc.attr);
3101 return PTR_ERR(dir);
3103 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3104 &fail_page_alloc.ignore_gfp_reclaim))
3106 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3107 &fail_page_alloc.ignore_gfp_highmem))
3109 if (!debugfs_create_u32("min-order", mode, dir,
3110 &fail_page_alloc.min_order))
3115 debugfs_remove_recursive(dir);
3120 late_initcall(fail_page_alloc_debugfs);
3122 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3124 #else /* CONFIG_FAIL_PAGE_ALLOC */
3126 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3131 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3134 * Return true if free base pages are above 'mark'. For high-order checks it
3135 * will return true of the order-0 watermark is reached and there is at least
3136 * one free page of a suitable size. Checking now avoids taking the zone lock
3137 * to check in the allocation paths if no pages are free.
3139 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3140 int classzone_idx, unsigned int alloc_flags,
3145 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3147 /* free_pages may go negative - that's OK */
3148 free_pages -= (1 << order) - 1;
3150 if (alloc_flags & ALLOC_HIGH)
3154 * If the caller does not have rights to ALLOC_HARDER then subtract
3155 * the high-atomic reserves. This will over-estimate the size of the
3156 * atomic reserve but it avoids a search.
3158 if (likely(!alloc_harder)) {
3159 free_pages -= z->nr_reserved_highatomic;
3162 * OOM victims can try even harder than normal ALLOC_HARDER
3163 * users on the grounds that it's definitely going to be in
3164 * the exit path shortly and free memory. Any allocation it
3165 * makes during the free path will be small and short-lived.
3167 if (alloc_flags & ALLOC_OOM)
3175 /* If allocation can't use CMA areas don't use free CMA pages */
3176 if (!(alloc_flags & ALLOC_CMA))
3177 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3181 * Check watermarks for an order-0 allocation request. If these
3182 * are not met, then a high-order request also cannot go ahead
3183 * even if a suitable page happened to be free.
3185 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3188 /* If this is an order-0 request then the watermark is fine */
3192 /* For a high-order request, check at least one suitable page is free */
3193 for (o = order; o < MAX_ORDER; o++) {
3194 struct free_area *area = &z->free_area[o];
3200 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3201 if (!list_empty(&area->free_list[mt]))
3206 if ((alloc_flags & ALLOC_CMA) &&
3207 !list_empty(&area->free_list[MIGRATE_CMA])) {
3212 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3218 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3219 int classzone_idx, unsigned int alloc_flags)
3221 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3222 zone_page_state(z, NR_FREE_PAGES));
3225 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3226 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3228 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3232 /* If allocation can't use CMA areas don't use free CMA pages */
3233 if (!(alloc_flags & ALLOC_CMA))
3234 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3238 * Fast check for order-0 only. If this fails then the reserves
3239 * need to be calculated. There is a corner case where the check
3240 * passes but only the high-order atomic reserve are free. If
3241 * the caller is !atomic then it'll uselessly search the free
3242 * list. That corner case is then slower but it is harmless.
3244 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3247 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3251 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3252 unsigned long mark, int classzone_idx)
3254 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3256 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3257 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3259 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3264 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3266 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3269 #else /* CONFIG_NUMA */
3270 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3274 #endif /* CONFIG_NUMA */
3277 * get_page_from_freelist goes through the zonelist trying to allocate
3280 static struct page *
3281 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3282 const struct alloc_context *ac)
3284 struct zoneref *z = ac->preferred_zoneref;
3286 struct pglist_data *last_pgdat_dirty_limit = NULL;
3289 * Scan zonelist, looking for a zone with enough free.
3290 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3292 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3297 if (cpusets_enabled() &&
3298 (alloc_flags & ALLOC_CPUSET) &&
3299 !__cpuset_zone_allowed(zone, gfp_mask))
3302 * When allocating a page cache page for writing, we
3303 * want to get it from a node that is within its dirty
3304 * limit, such that no single node holds more than its
3305 * proportional share of globally allowed dirty pages.
3306 * The dirty limits take into account the node's
3307 * lowmem reserves and high watermark so that kswapd
3308 * should be able to balance it without having to
3309 * write pages from its LRU list.
3311 * XXX: For now, allow allocations to potentially
3312 * exceed the per-node dirty limit in the slowpath
3313 * (spread_dirty_pages unset) before going into reclaim,
3314 * which is important when on a NUMA setup the allowed
3315 * nodes are together not big enough to reach the
3316 * global limit. The proper fix for these situations
3317 * will require awareness of nodes in the
3318 * dirty-throttling and the flusher threads.
3320 if (ac->spread_dirty_pages) {
3321 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3324 if (!node_dirty_ok(zone->zone_pgdat)) {
3325 last_pgdat_dirty_limit = zone->zone_pgdat;
3330 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3331 if (!zone_watermark_fast(zone, order, mark,
3332 ac_classzone_idx(ac), alloc_flags)) {
3335 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3337 * Watermark failed for this zone, but see if we can
3338 * grow this zone if it contains deferred pages.
3340 if (static_branch_unlikely(&deferred_pages)) {
3341 if (_deferred_grow_zone(zone, order))
3345 /* Checked here to keep the fast path fast */
3346 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3347 if (alloc_flags & ALLOC_NO_WATERMARKS)
3350 if (node_reclaim_mode == 0 ||
3351 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3354 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3356 case NODE_RECLAIM_NOSCAN:
3359 case NODE_RECLAIM_FULL:
3360 /* scanned but unreclaimable */
3363 /* did we reclaim enough */
3364 if (zone_watermark_ok(zone, order, mark,
3365 ac_classzone_idx(ac), alloc_flags))
3373 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3374 gfp_mask, alloc_flags, ac->migratetype);
3376 prep_new_page(page, order, gfp_mask, alloc_flags);
3379 * If this is a high-order atomic allocation then check
3380 * if the pageblock should be reserved for the future
3382 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3383 reserve_highatomic_pageblock(page, zone, order);
3387 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3388 /* Try again if zone has deferred pages */
3389 if (static_branch_unlikely(&deferred_pages)) {
3390 if (_deferred_grow_zone(zone, order))
3401 * Large machines with many possible nodes should not always dump per-node
3402 * meminfo in irq context.
3404 static inline bool should_suppress_show_mem(void)
3409 ret = in_interrupt();
3414 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3416 unsigned int filter = SHOW_MEM_FILTER_NODES;
3417 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3419 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3423 * This documents exceptions given to allocations in certain
3424 * contexts that are allowed to allocate outside current's set
3427 if (!(gfp_mask & __GFP_NOMEMALLOC))
3428 if (tsk_is_oom_victim(current) ||
3429 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3430 filter &= ~SHOW_MEM_FILTER_NODES;
3431 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3432 filter &= ~SHOW_MEM_FILTER_NODES;
3434 show_mem(filter, nodemask);
3437 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3439 struct va_format vaf;
3441 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3442 DEFAULT_RATELIMIT_BURST);
3444 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3447 va_start(args, fmt);
3450 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3451 current->comm, &vaf, gfp_mask, &gfp_mask,
3452 nodemask_pr_args(nodemask));
3455 cpuset_print_current_mems_allowed();
3458 warn_alloc_show_mem(gfp_mask, nodemask);
3461 static inline struct page *
3462 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3463 unsigned int alloc_flags,
3464 const struct alloc_context *ac)
3468 page = get_page_from_freelist(gfp_mask, order,
3469 alloc_flags|ALLOC_CPUSET, ac);
3471 * fallback to ignore cpuset restriction if our nodes
3475 page = get_page_from_freelist(gfp_mask, order,
3481 static inline struct page *
3482 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3483 const struct alloc_context *ac, unsigned long *did_some_progress)
3485 struct oom_control oc = {
3486 .zonelist = ac->zonelist,
3487 .nodemask = ac->nodemask,
3489 .gfp_mask = gfp_mask,
3494 *did_some_progress = 0;
3497 * Acquire the oom lock. If that fails, somebody else is
3498 * making progress for us.
3500 if (!mutex_trylock(&oom_lock)) {
3501 *did_some_progress = 1;
3502 schedule_timeout_uninterruptible(1);
3507 * Go through the zonelist yet one more time, keep very high watermark
3508 * here, this is only to catch a parallel oom killing, we must fail if
3509 * we're still under heavy pressure. But make sure that this reclaim
3510 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3511 * allocation which will never fail due to oom_lock already held.
3513 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3514 ~__GFP_DIRECT_RECLAIM, order,
3515 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3519 /* Coredumps can quickly deplete all memory reserves */
3520 if (current->flags & PF_DUMPCORE)
3522 /* The OOM killer will not help higher order allocs */
3523 if (order > PAGE_ALLOC_COSTLY_ORDER)
3526 * We have already exhausted all our reclaim opportunities without any
3527 * success so it is time to admit defeat. We will skip the OOM killer
3528 * because it is very likely that the caller has a more reasonable
3529 * fallback than shooting a random task.
3531 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3533 /* The OOM killer does not needlessly kill tasks for lowmem */
3534 if (ac->high_zoneidx < ZONE_NORMAL)
3536 if (pm_suspended_storage())
3539 * XXX: GFP_NOFS allocations should rather fail than rely on
3540 * other request to make a forward progress.
3541 * We are in an unfortunate situation where out_of_memory cannot
3542 * do much for this context but let's try it to at least get
3543 * access to memory reserved if the current task is killed (see
3544 * out_of_memory). Once filesystems are ready to handle allocation
3545 * failures more gracefully we should just bail out here.
3548 /* The OOM killer may not free memory on a specific node */
3549 if (gfp_mask & __GFP_THISNODE)
3552 /* Exhausted what can be done so it's blame time */
3553 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3554 *did_some_progress = 1;
3557 * Help non-failing allocations by giving them access to memory
3560 if (gfp_mask & __GFP_NOFAIL)
3561 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3562 ALLOC_NO_WATERMARKS, ac);
3565 mutex_unlock(&oom_lock);
3570 * Maximum number of compaction retries wit a progress before OOM
3571 * killer is consider as the only way to move forward.
3573 #define MAX_COMPACT_RETRIES 16
3575 #ifdef CONFIG_COMPACTION
3576 /* Try memory compaction for high-order allocations before reclaim */
3577 static struct page *
3578 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3579 unsigned int alloc_flags, const struct alloc_context *ac,
3580 enum compact_priority prio, enum compact_result *compact_result)
3583 unsigned int noreclaim_flag;
3588 noreclaim_flag = memalloc_noreclaim_save();
3589 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3591 memalloc_noreclaim_restore(noreclaim_flag);
3593 if (*compact_result <= COMPACT_INACTIVE)
3597 * At least in one zone compaction wasn't deferred or skipped, so let's
3598 * count a compaction stall
3600 count_vm_event(COMPACTSTALL);
3602 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3605 struct zone *zone = page_zone(page);
3607 zone->compact_blockskip_flush = false;
3608 compaction_defer_reset(zone, order, true);
3609 count_vm_event(COMPACTSUCCESS);
3614 * It's bad if compaction run occurs and fails. The most likely reason
3615 * is that pages exist, but not enough to satisfy watermarks.
3617 count_vm_event(COMPACTFAIL);
3625 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3626 enum compact_result compact_result,
3627 enum compact_priority *compact_priority,
3628 int *compaction_retries)
3630 int max_retries = MAX_COMPACT_RETRIES;
3633 int retries = *compaction_retries;
3634 enum compact_priority priority = *compact_priority;
3639 if (compaction_made_progress(compact_result))
3640 (*compaction_retries)++;
3643 * compaction considers all the zone as desperately out of memory
3644 * so it doesn't really make much sense to retry except when the
3645 * failure could be caused by insufficient priority
3647 if (compaction_failed(compact_result))
3648 goto check_priority;
3651 * make sure the compaction wasn't deferred or didn't bail out early
3652 * due to locks contention before we declare that we should give up.
3653 * But do not retry if the given zonelist is not suitable for
3656 if (compaction_withdrawn(compact_result)) {
3657 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3662 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3663 * costly ones because they are de facto nofail and invoke OOM
3664 * killer to move on while costly can fail and users are ready
3665 * to cope with that. 1/4 retries is rather arbitrary but we
3666 * would need much more detailed feedback from compaction to
3667 * make a better decision.
3669 if (order > PAGE_ALLOC_COSTLY_ORDER)
3671 if (*compaction_retries <= max_retries) {
3677 * Make sure there are attempts at the highest priority if we exhausted
3678 * all retries or failed at the lower priorities.
3681 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3682 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3684 if (*compact_priority > min_priority) {
3685 (*compact_priority)--;
3686 *compaction_retries = 0;
3690 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3694 static inline struct page *
3695 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3696 unsigned int alloc_flags, const struct alloc_context *ac,
3697 enum compact_priority prio, enum compact_result *compact_result)
3699 *compact_result = COMPACT_SKIPPED;
3704 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3705 enum compact_result compact_result,
3706 enum compact_priority *compact_priority,
3707 int *compaction_retries)
3712 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3716 * There are setups with compaction disabled which would prefer to loop
3717 * inside the allocator rather than hit the oom killer prematurely.
3718 * Let's give them a good hope and keep retrying while the order-0
3719 * watermarks are OK.
3721 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3723 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3724 ac_classzone_idx(ac), alloc_flags))
3729 #endif /* CONFIG_COMPACTION */
3731 #ifdef CONFIG_LOCKDEP
3732 static struct lockdep_map __fs_reclaim_map =
3733 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3735 static bool __need_fs_reclaim(gfp_t gfp_mask)
3737 gfp_mask = current_gfp_context(gfp_mask);
3739 /* no reclaim without waiting on it */
3740 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3743 /* this guy won't enter reclaim */
3744 if (current->flags & PF_MEMALLOC)
3747 /* We're only interested __GFP_FS allocations for now */
3748 if (!(gfp_mask & __GFP_FS))
3751 if (gfp_mask & __GFP_NOLOCKDEP)
3757 void __fs_reclaim_acquire(void)
3759 lock_map_acquire(&__fs_reclaim_map);
3762 void __fs_reclaim_release(void)
3764 lock_map_release(&__fs_reclaim_map);
3767 void fs_reclaim_acquire(gfp_t gfp_mask)
3769 if (__need_fs_reclaim(gfp_mask))
3770 __fs_reclaim_acquire();
3772 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3774 void fs_reclaim_release(gfp_t gfp_mask)
3776 if (__need_fs_reclaim(gfp_mask))
3777 __fs_reclaim_release();
3779 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3783 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3784 * have been rebuilt so allocation retries. Reader side does not lock and
3785 * retries the allocation if zonelist changes. Writer side is protected by the
3786 * embedded spin_lock.
3788 static DEFINE_SEQLOCK(zonelist_update_seq);
3790 static unsigned int zonelist_iter_begin(void)
3792 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3793 return read_seqbegin(&zonelist_update_seq);
3798 static unsigned int check_retry_zonelist(unsigned int seq)
3800 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3801 return read_seqretry(&zonelist_update_seq, seq);
3806 /* Perform direct synchronous page reclaim */
3808 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3809 const struct alloc_context *ac)
3811 struct reclaim_state reclaim_state;
3813 unsigned int noreclaim_flag;
3817 /* We now go into synchronous reclaim */
3818 cpuset_memory_pressure_bump();
3819 fs_reclaim_acquire(gfp_mask);
3820 noreclaim_flag = memalloc_noreclaim_save();
3821 reclaim_state.reclaimed_slab = 0;
3822 current->reclaim_state = &reclaim_state;
3824 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3827 current->reclaim_state = NULL;
3828 memalloc_noreclaim_restore(noreclaim_flag);
3829 fs_reclaim_release(gfp_mask);
3836 /* The really slow allocator path where we enter direct reclaim */
3837 static inline struct page *
3838 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3839 unsigned int alloc_flags, const struct alloc_context *ac,
3840 unsigned long *did_some_progress)
3842 struct page *page = NULL;
3843 bool drained = false;
3845 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3846 if (unlikely(!(*did_some_progress)))
3850 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3853 * If an allocation failed after direct reclaim, it could be because
3854 * pages are pinned on the per-cpu lists or in high alloc reserves.
3855 * Shrink them them and try again
3857 if (!page && !drained) {
3858 unreserve_highatomic_pageblock(ac, false);
3859 drain_all_pages(NULL);
3867 static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3868 const struct alloc_context *ac)
3872 pg_data_t *last_pgdat = NULL;
3873 enum zone_type high_zoneidx = ac->high_zoneidx;
3875 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3877 if (last_pgdat != zone->zone_pgdat)
3878 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3879 last_pgdat = zone->zone_pgdat;
3883 static inline unsigned int
3884 gfp_to_alloc_flags(gfp_t gfp_mask)
3886 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3888 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3889 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3892 * The caller may dip into page reserves a bit more if the caller
3893 * cannot run direct reclaim, or if the caller has realtime scheduling
3894 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3895 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3897 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3899 if (gfp_mask & __GFP_ATOMIC) {
3901 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3902 * if it can't schedule.
3904 if (!(gfp_mask & __GFP_NOMEMALLOC))
3905 alloc_flags |= ALLOC_HARDER;
3907 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3908 * comment for __cpuset_node_allowed().
3910 alloc_flags &= ~ALLOC_CPUSET;
3911 } else if (unlikely(rt_task(current)) && !in_interrupt())
3912 alloc_flags |= ALLOC_HARDER;
3915 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3916 alloc_flags |= ALLOC_CMA;
3921 static bool oom_reserves_allowed(struct task_struct *tsk)
3923 if (!tsk_is_oom_victim(tsk))
3927 * !MMU doesn't have oom reaper so give access to memory reserves
3928 * only to the thread with TIF_MEMDIE set
3930 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3937 * Distinguish requests which really need access to full memory
3938 * reserves from oom victims which can live with a portion of it
3940 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3942 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3944 if (gfp_mask & __GFP_MEMALLOC)
3945 return ALLOC_NO_WATERMARKS;
3946 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3947 return ALLOC_NO_WATERMARKS;
3948 if (!in_interrupt()) {
3949 if (current->flags & PF_MEMALLOC)
3950 return ALLOC_NO_WATERMARKS;
3951 else if (oom_reserves_allowed(current))
3958 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3960 return !!__gfp_pfmemalloc_flags(gfp_mask);
3964 * Checks whether it makes sense to retry the reclaim to make a forward progress
3965 * for the given allocation request.
3967 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3968 * without success, or when we couldn't even meet the watermark if we
3969 * reclaimed all remaining pages on the LRU lists.
3971 * Returns true if a retry is viable or false to enter the oom path.
3974 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3975 struct alloc_context *ac, int alloc_flags,
3976 bool did_some_progress, int *no_progress_loops)
3982 * Costly allocations might have made a progress but this doesn't mean
3983 * their order will become available due to high fragmentation so
3984 * always increment the no progress counter for them
3986 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3987 *no_progress_loops = 0;
3989 (*no_progress_loops)++;
3992 * Make sure we converge to OOM if we cannot make any progress
3993 * several times in the row.
3995 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3996 /* Before OOM, exhaust highatomic_reserve */
3997 return unreserve_highatomic_pageblock(ac, true);
4001 * Keep reclaiming pages while there is a chance this will lead
4002 * somewhere. If none of the target zones can satisfy our allocation
4003 * request even if all reclaimable pages are considered then we are
4004 * screwed and have to go OOM.
4006 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4008 unsigned long available;
4009 unsigned long reclaimable;
4010 unsigned long min_wmark = min_wmark_pages(zone);
4013 available = reclaimable = zone_reclaimable_pages(zone);
4014 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4017 * Would the allocation succeed if we reclaimed all
4018 * reclaimable pages?
4020 wmark = __zone_watermark_ok(zone, order, min_wmark,
4021 ac_classzone_idx(ac), alloc_flags, available);
4022 trace_reclaim_retry_zone(z, order, reclaimable,
4023 available, min_wmark, *no_progress_loops, wmark);
4026 * If we didn't make any progress and have a lot of
4027 * dirty + writeback pages then we should wait for
4028 * an IO to complete to slow down the reclaim and
4029 * prevent from pre mature OOM
4031 if (!did_some_progress) {
4032 unsigned long write_pending;
4034 write_pending = zone_page_state_snapshot(zone,
4035 NR_ZONE_WRITE_PENDING);
4037 if (2 * write_pending > reclaimable) {
4038 congestion_wait(BLK_RW_ASYNC, HZ/10);
4044 * Memory allocation/reclaim might be called from a WQ
4045 * context and the current implementation of the WQ
4046 * concurrency control doesn't recognize that
4047 * a particular WQ is congested if the worker thread is
4048 * looping without ever sleeping. Therefore we have to
4049 * do a short sleep here rather than calling
4052 if (current->flags & PF_WQ_WORKER)
4053 schedule_timeout_uninterruptible(1);
4065 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4068 * It's possible that cpuset's mems_allowed and the nodemask from
4069 * mempolicy don't intersect. This should be normally dealt with by
4070 * policy_nodemask(), but it's possible to race with cpuset update in
4071 * such a way the check therein was true, and then it became false
4072 * before we got our cpuset_mems_cookie here.
4073 * This assumes that for all allocations, ac->nodemask can come only
4074 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4075 * when it does not intersect with the cpuset restrictions) or the
4076 * caller can deal with a violated nodemask.
4078 if (cpusets_enabled() && ac->nodemask &&
4079 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4080 ac->nodemask = NULL;
4085 * When updating a task's mems_allowed or mempolicy nodemask, it is
4086 * possible to race with parallel threads in such a way that our
4087 * allocation can fail while the mask is being updated. If we are about
4088 * to fail, check if the cpuset changed during allocation and if so,
4091 if (read_mems_allowed_retry(cpuset_mems_cookie))
4097 static inline struct page *
4098 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4099 struct alloc_context *ac)
4101 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4102 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4103 struct page *page = NULL;
4104 unsigned int alloc_flags;
4105 unsigned long did_some_progress;
4106 enum compact_priority compact_priority;
4107 enum compact_result compact_result;
4108 int compaction_retries;
4109 int no_progress_loops;
4110 unsigned int cpuset_mems_cookie;
4111 unsigned int zonelist_iter_cookie;
4115 * We also sanity check to catch abuse of atomic reserves being used by
4116 * callers that are not in atomic context.
4118 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4119 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4120 gfp_mask &= ~__GFP_ATOMIC;
4123 compaction_retries = 0;
4124 no_progress_loops = 0;
4125 compact_priority = DEF_COMPACT_PRIORITY;
4126 cpuset_mems_cookie = read_mems_allowed_begin();
4127 zonelist_iter_cookie = zonelist_iter_begin();
4130 * The fast path uses conservative alloc_flags to succeed only until
4131 * kswapd needs to be woken up, and to avoid the cost of setting up
4132 * alloc_flags precisely. So we do that now.
4134 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4137 * We need to recalculate the starting point for the zonelist iterator
4138 * because we might have used different nodemask in the fast path, or
4139 * there was a cpuset modification and we are retrying - otherwise we
4140 * could end up iterating over non-eligible zones endlessly.
4142 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4143 ac->high_zoneidx, ac->nodemask);
4144 if (!ac->preferred_zoneref->zone)
4147 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4148 wake_all_kswapds(order, gfp_mask, ac);
4151 * The adjusted alloc_flags might result in immediate success, so try
4154 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4159 * For costly allocations, try direct compaction first, as it's likely
4160 * that we have enough base pages and don't need to reclaim. For non-
4161 * movable high-order allocations, do that as well, as compaction will
4162 * try prevent permanent fragmentation by migrating from blocks of the
4164 * Don't try this for allocations that are allowed to ignore
4165 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4167 if (can_direct_reclaim &&
4169 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4170 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4171 page = __alloc_pages_direct_compact(gfp_mask, order,
4173 INIT_COMPACT_PRIORITY,
4179 * Checks for costly allocations with __GFP_NORETRY, which
4180 * includes THP page fault allocations
4182 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4184 * If compaction is deferred for high-order allocations,
4185 * it is because sync compaction recently failed. If
4186 * this is the case and the caller requested a THP
4187 * allocation, we do not want to heavily disrupt the
4188 * system, so we fail the allocation instead of entering
4191 if (compact_result == COMPACT_DEFERRED)
4195 * Looks like reclaim/compaction is worth trying, but
4196 * sync compaction could be very expensive, so keep
4197 * using async compaction.
4199 compact_priority = INIT_COMPACT_PRIORITY;
4204 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4205 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4206 wake_all_kswapds(order, gfp_mask, ac);
4208 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4210 alloc_flags = reserve_flags;
4213 * Reset the nodemask and zonelist iterators if memory policies can be
4214 * ignored. These allocations are high priority and system rather than
4217 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4218 ac->nodemask = NULL;
4219 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4220 ac->high_zoneidx, ac->nodemask);
4223 /* Attempt with potentially adjusted zonelist and alloc_flags */
4224 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4228 /* Caller is not willing to reclaim, we can't balance anything */
4229 if (!can_direct_reclaim)
4232 /* Avoid recursion of direct reclaim */
4233 if (current->flags & PF_MEMALLOC)
4236 /* Try direct reclaim and then allocating */
4237 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4238 &did_some_progress);
4242 /* Try direct compaction and then allocating */
4243 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4244 compact_priority, &compact_result);
4248 /* Do not loop if specifically requested */
4249 if (gfp_mask & __GFP_NORETRY)
4253 * Do not retry costly high order allocations unless they are
4254 * __GFP_RETRY_MAYFAIL
4256 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4259 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4260 did_some_progress > 0, &no_progress_loops))
4264 * It doesn't make any sense to retry for the compaction if the order-0
4265 * reclaim is not able to make any progress because the current
4266 * implementation of the compaction depends on the sufficient amount
4267 * of free memory (see __compaction_suitable)
4269 if (did_some_progress > 0 &&
4270 should_compact_retry(ac, order, alloc_flags,
4271 compact_result, &compact_priority,
4272 &compaction_retries))
4277 * Deal with possible cpuset update races or zonelist updates to avoid
4278 * a unnecessary OOM kill.
4280 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4281 check_retry_zonelist(zonelist_iter_cookie))
4284 /* Reclaim has failed us, start killing things */
4285 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4289 /* Avoid allocations with no watermarks from looping endlessly */
4290 if (tsk_is_oom_victim(current) &&
4291 (alloc_flags == ALLOC_OOM ||
4292 (gfp_mask & __GFP_NOMEMALLOC)))
4295 /* Retry as long as the OOM killer is making progress */
4296 if (did_some_progress) {
4297 no_progress_loops = 0;
4303 * Deal with possible cpuset update races or zonelist updates to avoid
4304 * a unnecessary OOM kill.
4306 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4307 check_retry_zonelist(zonelist_iter_cookie))
4311 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4314 if (gfp_mask & __GFP_NOFAIL) {
4316 * All existing users of the __GFP_NOFAIL are blockable, so warn
4317 * of any new users that actually require GFP_NOWAIT
4319 if (WARN_ON_ONCE(!can_direct_reclaim))
4323 * PF_MEMALLOC request from this context is rather bizarre
4324 * because we cannot reclaim anything and only can loop waiting
4325 * for somebody to do a work for us
4327 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4330 * non failing costly orders are a hard requirement which we
4331 * are not prepared for much so let's warn about these users
4332 * so that we can identify them and convert them to something
4335 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4338 * Help non-failing allocations by giving them access to memory
4339 * reserves but do not use ALLOC_NO_WATERMARKS because this
4340 * could deplete whole memory reserves which would just make
4341 * the situation worse
4343 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4351 warn_alloc(gfp_mask, ac->nodemask,
4352 "page allocation failure: order:%u", order);
4357 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4358 int preferred_nid, nodemask_t *nodemask,
4359 struct alloc_context *ac, gfp_t *alloc_mask,
4360 unsigned int *alloc_flags)
4362 ac->high_zoneidx = gfp_zone(gfp_mask);
4363 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4364 ac->nodemask = nodemask;
4365 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4367 if (cpusets_enabled()) {
4368 *alloc_mask |= __GFP_HARDWALL;
4370 ac->nodemask = &cpuset_current_mems_allowed;
4372 *alloc_flags |= ALLOC_CPUSET;
4375 fs_reclaim_acquire(gfp_mask);
4376 fs_reclaim_release(gfp_mask);
4378 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4380 if (should_fail_alloc_page(gfp_mask, order))
4383 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4384 *alloc_flags |= ALLOC_CMA;
4389 /* Determine whether to spread dirty pages and what the first usable zone */
4390 static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4392 /* Dirty zone balancing only done in the fast path */
4393 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4396 * The preferred zone is used for statistics but crucially it is
4397 * also used as the starting point for the zonelist iterator. It
4398 * may get reset for allocations that ignore memory policies.
4400 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4401 ac->high_zoneidx, ac->nodemask);
4405 * This is the 'heart' of the zoned buddy allocator.
4408 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4409 nodemask_t *nodemask)
4412 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4413 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4414 struct alloc_context ac = { };
4417 * There are several places where we assume that the order value is sane
4418 * so bail out early if the request is out of bound.
4420 if (unlikely(order >= MAX_ORDER)) {
4421 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4425 gfp_mask &= gfp_allowed_mask;
4426 alloc_mask = gfp_mask;
4427 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4430 finalise_ac(gfp_mask, &ac);
4432 /* First allocation attempt */
4433 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4438 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4439 * resp. GFP_NOIO which has to be inherited for all allocation requests
4440 * from a particular context which has been marked by
4441 * memalloc_no{fs,io}_{save,restore}.
4443 alloc_mask = current_gfp_context(gfp_mask);
4444 ac.spread_dirty_pages = false;
4447 * Restore the original nodemask if it was potentially replaced with
4448 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4450 if (unlikely(ac.nodemask != nodemask))
4451 ac.nodemask = nodemask;
4453 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4456 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4457 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4458 __free_pages(page, order);
4462 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4466 EXPORT_SYMBOL(__alloc_pages_nodemask);
4469 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4470 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4471 * you need to access high mem.
4473 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4477 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4480 return (unsigned long) page_address(page);
4482 EXPORT_SYMBOL(__get_free_pages);
4484 unsigned long get_zeroed_page(gfp_t gfp_mask)
4486 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4488 EXPORT_SYMBOL(get_zeroed_page);
4490 static inline void free_the_page(struct page *page, unsigned int order)
4492 if (order == 0) /* Via pcp? */
4493 free_unref_page(page);
4495 __free_pages_ok(page, order);
4498 void __free_pages(struct page *page, unsigned int order)
4500 if (put_page_testzero(page))
4501 free_the_page(page, order);
4503 EXPORT_SYMBOL(__free_pages);
4505 void free_pages(unsigned long addr, unsigned int order)
4508 VM_BUG_ON(!virt_addr_valid((void *)addr));
4509 __free_pages(virt_to_page((void *)addr), order);
4513 EXPORT_SYMBOL(free_pages);
4517 * An arbitrary-length arbitrary-offset area of memory which resides
4518 * within a 0 or higher order page. Multiple fragments within that page
4519 * are individually refcounted, in the page's reference counter.
4521 * The page_frag functions below provide a simple allocation framework for
4522 * page fragments. This is used by the network stack and network device
4523 * drivers to provide a backing region of memory for use as either an
4524 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4526 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4529 struct page *page = NULL;
4530 gfp_t gfp = gfp_mask;
4532 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4533 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4535 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4536 PAGE_FRAG_CACHE_MAX_ORDER);
4537 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4539 if (unlikely(!page))
4540 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4542 nc->va = page ? page_address(page) : NULL;
4547 void __page_frag_cache_drain(struct page *page, unsigned int count)
4549 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4551 if (page_ref_sub_and_test(page, count))
4552 free_the_page(page, compound_order(page));
4554 EXPORT_SYMBOL(__page_frag_cache_drain);
4556 void *page_frag_alloc(struct page_frag_cache *nc,
4557 unsigned int fragsz, gfp_t gfp_mask)
4559 unsigned int size = PAGE_SIZE;
4563 if (unlikely(!nc->va)) {
4565 page = __page_frag_cache_refill(nc, gfp_mask);
4569 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4570 /* if size can vary use size else just use PAGE_SIZE */
4573 /* Even if we own the page, we do not use atomic_set().
4574 * This would break get_page_unless_zero() users.
4576 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4578 /* reset page count bias and offset to start of new frag */
4579 nc->pfmemalloc = page_is_pfmemalloc(page);
4580 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4584 offset = nc->offset - fragsz;
4585 if (unlikely(offset < 0)) {
4586 page = virt_to_page(nc->va);
4588 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4591 if (unlikely(nc->pfmemalloc)) {
4592 free_the_page(page, compound_order(page));
4596 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4597 /* if size can vary use size else just use PAGE_SIZE */
4600 /* OK, page count is 0, we can safely set it */
4601 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4603 /* reset page count bias and offset to start of new frag */
4604 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4605 offset = size - fragsz;
4606 if (unlikely(offset < 0)) {
4608 * The caller is trying to allocate a fragment
4609 * with fragsz > PAGE_SIZE but the cache isn't big
4610 * enough to satisfy the request, this may
4611 * happen in low memory conditions.
4612 * We don't release the cache page because
4613 * it could make memory pressure worse
4614 * so we simply return NULL here.
4621 nc->offset = offset;
4623 return nc->va + offset;
4625 EXPORT_SYMBOL(page_frag_alloc);
4628 * Frees a page fragment allocated out of either a compound or order 0 page.
4630 void page_frag_free(void *addr)
4632 struct page *page = virt_to_head_page(addr);
4634 if (unlikely(put_page_testzero(page)))
4635 free_the_page(page, compound_order(page));
4637 EXPORT_SYMBOL(page_frag_free);
4639 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4643 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4644 unsigned long used = addr + PAGE_ALIGN(size);
4646 split_page(virt_to_page((void *)addr), order);
4647 while (used < alloc_end) {
4652 return (void *)addr;
4656 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4657 * @size: the number of bytes to allocate
4658 * @gfp_mask: GFP flags for the allocation
4660 * This function is similar to alloc_pages(), except that it allocates the
4661 * minimum number of pages to satisfy the request. alloc_pages() can only
4662 * allocate memory in power-of-two pages.
4664 * This function is also limited by MAX_ORDER.
4666 * Memory allocated by this function must be released by free_pages_exact().
4668 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4670 unsigned int order = get_order(size);
4673 addr = __get_free_pages(gfp_mask, order);
4674 return make_alloc_exact(addr, order, size);
4676 EXPORT_SYMBOL(alloc_pages_exact);
4679 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4681 * @nid: the preferred node ID where memory should be allocated
4682 * @size: the number of bytes to allocate
4683 * @gfp_mask: GFP flags for the allocation
4685 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4688 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4690 unsigned int order = get_order(size);
4691 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4694 return make_alloc_exact((unsigned long)page_address(p), order, size);
4698 * free_pages_exact - release memory allocated via alloc_pages_exact()
4699 * @virt: the value returned by alloc_pages_exact.
4700 * @size: size of allocation, same value as passed to alloc_pages_exact().
4702 * Release the memory allocated by a previous call to alloc_pages_exact.
4704 void free_pages_exact(void *virt, size_t size)
4706 unsigned long addr = (unsigned long)virt;
4707 unsigned long end = addr + PAGE_ALIGN(size);
4709 while (addr < end) {
4714 EXPORT_SYMBOL(free_pages_exact);
4717 * nr_free_zone_pages - count number of pages beyond high watermark
4718 * @offset: The zone index of the highest zone
4720 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4721 * high watermark within all zones at or below a given zone index. For each
4722 * zone, the number of pages is calculated as:
4724 * nr_free_zone_pages = managed_pages - high_pages
4726 static unsigned long nr_free_zone_pages(int offset)
4731 /* Just pick one node, since fallback list is circular */
4732 unsigned long sum = 0;
4734 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4736 for_each_zone_zonelist(zone, z, zonelist, offset) {
4737 unsigned long size = zone->managed_pages;
4738 unsigned long high = high_wmark_pages(zone);
4747 * nr_free_buffer_pages - count number of pages beyond high watermark
4749 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4750 * watermark within ZONE_DMA and ZONE_NORMAL.
4752 unsigned long nr_free_buffer_pages(void)
4754 return nr_free_zone_pages(gfp_zone(GFP_USER));
4756 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4759 * nr_free_pagecache_pages - count number of pages beyond high watermark
4761 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4762 * high watermark within all zones.
4764 unsigned long nr_free_pagecache_pages(void)
4766 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4769 static inline void show_node(struct zone *zone)
4771 if (IS_ENABLED(CONFIG_NUMA))
4772 printk("Node %d ", zone_to_nid(zone));
4775 long si_mem_available(void)
4778 unsigned long pagecache;
4779 unsigned long wmark_low = 0;
4780 unsigned long pages[NR_LRU_LISTS];
4784 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4785 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4788 wmark_low += zone->watermark[WMARK_LOW];
4791 * Estimate the amount of memory available for userspace allocations,
4792 * without causing swapping.
4794 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4797 * Not all the page cache can be freed, otherwise the system will
4798 * start swapping. Assume at least half of the page cache, or the
4799 * low watermark worth of cache, needs to stay.
4801 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4802 pagecache -= min(pagecache / 2, wmark_low);
4803 available += pagecache;
4806 * Part of the reclaimable slab consists of items that are in use,
4807 * and cannot be freed. Cap this estimate at the low watermark.
4809 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4810 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4814 * Part of the kernel memory, which can be released under memory
4817 available += global_node_page_state(NR_INDIRECTLY_RECLAIMABLE_BYTES) >>
4824 EXPORT_SYMBOL_GPL(si_mem_available);
4826 void si_meminfo(struct sysinfo *val)
4828 val->totalram = totalram_pages;
4829 val->sharedram = global_node_page_state(NR_SHMEM);
4830 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4831 val->bufferram = nr_blockdev_pages();
4832 val->totalhigh = totalhigh_pages;
4833 val->freehigh = nr_free_highpages();
4834 val->mem_unit = PAGE_SIZE;
4837 EXPORT_SYMBOL(si_meminfo);
4840 void si_meminfo_node(struct sysinfo *val, int nid)
4842 int zone_type; /* needs to be signed */
4843 unsigned long managed_pages = 0;
4844 unsigned long managed_highpages = 0;
4845 unsigned long free_highpages = 0;
4846 pg_data_t *pgdat = NODE_DATA(nid);
4848 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4849 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4850 val->totalram = managed_pages;
4851 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4852 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4853 #ifdef CONFIG_HIGHMEM
4854 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4855 struct zone *zone = &pgdat->node_zones[zone_type];
4857 if (is_highmem(zone)) {
4858 managed_highpages += zone->managed_pages;
4859 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4862 val->totalhigh = managed_highpages;
4863 val->freehigh = free_highpages;
4865 val->totalhigh = managed_highpages;
4866 val->freehigh = free_highpages;
4868 val->mem_unit = PAGE_SIZE;
4873 * Determine whether the node should be displayed or not, depending on whether
4874 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4876 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4878 if (!(flags & SHOW_MEM_FILTER_NODES))
4882 * no node mask - aka implicit memory numa policy. Do not bother with
4883 * the synchronization - read_mems_allowed_begin - because we do not
4884 * have to be precise here.
4887 nodemask = &cpuset_current_mems_allowed;
4889 return !node_isset(nid, *nodemask);
4892 #define K(x) ((x) << (PAGE_SHIFT-10))
4894 static void show_migration_types(unsigned char type)
4896 static const char types[MIGRATE_TYPES] = {
4897 [MIGRATE_UNMOVABLE] = 'U',
4898 [MIGRATE_MOVABLE] = 'M',
4899 [MIGRATE_RECLAIMABLE] = 'E',
4900 [MIGRATE_HIGHATOMIC] = 'H',
4902 [MIGRATE_CMA] = 'C',
4904 #ifdef CONFIG_MEMORY_ISOLATION
4905 [MIGRATE_ISOLATE] = 'I',
4908 char tmp[MIGRATE_TYPES + 1];
4912 for (i = 0; i < MIGRATE_TYPES; i++) {
4913 if (type & (1 << i))
4918 printk(KERN_CONT "(%s) ", tmp);
4922 * Show free area list (used inside shift_scroll-lock stuff)
4923 * We also calculate the percentage fragmentation. We do this by counting the
4924 * memory on each free list with the exception of the first item on the list.
4927 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4930 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4932 unsigned long free_pcp = 0;
4937 for_each_populated_zone(zone) {
4938 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4941 for_each_online_cpu(cpu)
4942 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4945 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4946 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4947 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4948 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4949 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4950 " free:%lu free_pcp:%lu free_cma:%lu\n",
4951 global_node_page_state(NR_ACTIVE_ANON),
4952 global_node_page_state(NR_INACTIVE_ANON),
4953 global_node_page_state(NR_ISOLATED_ANON),
4954 global_node_page_state(NR_ACTIVE_FILE),
4955 global_node_page_state(NR_INACTIVE_FILE),
4956 global_node_page_state(NR_ISOLATED_FILE),
4957 global_node_page_state(NR_UNEVICTABLE),
4958 global_node_page_state(NR_FILE_DIRTY),
4959 global_node_page_state(NR_WRITEBACK),
4960 global_node_page_state(NR_UNSTABLE_NFS),
4961 global_node_page_state(NR_SLAB_RECLAIMABLE),
4962 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4963 global_node_page_state(NR_FILE_MAPPED),
4964 global_node_page_state(NR_SHMEM),
4965 global_zone_page_state(NR_PAGETABLE),
4966 global_zone_page_state(NR_BOUNCE),
4967 global_zone_page_state(NR_FREE_PAGES),
4969 global_zone_page_state(NR_FREE_CMA_PAGES));
4971 for_each_online_pgdat(pgdat) {
4972 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4976 " active_anon:%lukB"
4977 " inactive_anon:%lukB"
4978 " active_file:%lukB"
4979 " inactive_file:%lukB"
4980 " unevictable:%lukB"
4981 " isolated(anon):%lukB"
4982 " isolated(file):%lukB"
4987 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4989 " shmem_pmdmapped: %lukB"
4992 " writeback_tmp:%lukB"
4994 " all_unreclaimable? %s"
4997 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4998 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4999 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5000 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5001 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5002 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5003 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5004 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5005 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5006 K(node_page_state(pgdat, NR_WRITEBACK)),
5007 K(node_page_state(pgdat, NR_SHMEM)),
5008 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5009 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5010 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5012 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5014 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5015 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5016 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5020 for_each_populated_zone(zone) {
5023 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5027 for_each_online_cpu(cpu)
5028 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5037 " active_anon:%lukB"
5038 " inactive_anon:%lukB"
5039 " active_file:%lukB"
5040 " inactive_file:%lukB"
5041 " unevictable:%lukB"
5042 " writepending:%lukB"
5046 " kernel_stack:%lukB"
5054 K(zone_page_state(zone, NR_FREE_PAGES)),
5055 K(min_wmark_pages(zone)),
5056 K(low_wmark_pages(zone)),
5057 K(high_wmark_pages(zone)),
5058 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5059 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5060 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5061 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5062 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5063 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5064 K(zone->present_pages),
5065 K(zone->managed_pages),
5066 K(zone_page_state(zone, NR_MLOCK)),
5067 zone_page_state(zone, NR_KERNEL_STACK_KB),
5068 K(zone_page_state(zone, NR_PAGETABLE)),
5069 K(zone_page_state(zone, NR_BOUNCE)),
5071 K(this_cpu_read(zone->pageset->pcp.count)),
5072 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5073 printk("lowmem_reserve[]:");
5074 for (i = 0; i < MAX_NR_ZONES; i++)
5075 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5076 printk(KERN_CONT "\n");
5079 for_each_populated_zone(zone) {
5081 unsigned long nr[MAX_ORDER], flags, total = 0;
5082 unsigned char types[MAX_ORDER];
5084 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5087 printk(KERN_CONT "%s: ", zone->name);
5089 spin_lock_irqsave(&zone->lock, flags);
5090 for (order = 0; order < MAX_ORDER; order++) {
5091 struct free_area *area = &zone->free_area[order];
5094 nr[order] = area->nr_free;
5095 total += nr[order] << order;
5098 for (type = 0; type < MIGRATE_TYPES; type++) {
5099 if (!list_empty(&area->free_list[type]))
5100 types[order] |= 1 << type;
5103 spin_unlock_irqrestore(&zone->lock, flags);
5104 for (order = 0; order < MAX_ORDER; order++) {
5105 printk(KERN_CONT "%lu*%lukB ",
5106 nr[order], K(1UL) << order);
5108 show_migration_types(types[order]);
5110 printk(KERN_CONT "= %lukB\n", K(total));
5113 hugetlb_show_meminfo();
5115 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5117 show_swap_cache_info();
5120 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5122 zoneref->zone = zone;
5123 zoneref->zone_idx = zone_idx(zone);
5127 * Builds allocation fallback zone lists.
5129 * Add all populated zones of a node to the zonelist.
5131 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5134 enum zone_type zone_type = MAX_NR_ZONES;
5139 zone = pgdat->node_zones + zone_type;
5140 if (populated_zone(zone)) {
5141 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5142 check_highest_zone(zone_type);
5144 } while (zone_type);
5151 static int __parse_numa_zonelist_order(char *s)
5154 * We used to support different zonlists modes but they turned
5155 * out to be just not useful. Let's keep the warning in place
5156 * if somebody still use the cmd line parameter so that we do
5157 * not fail it silently
5159 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5160 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5166 static __init int setup_numa_zonelist_order(char *s)
5171 return __parse_numa_zonelist_order(s);
5173 early_param("numa_zonelist_order", setup_numa_zonelist_order);
5175 char numa_zonelist_order[] = "Node";
5178 * sysctl handler for numa_zonelist_order
5180 int numa_zonelist_order_handler(struct ctl_table *table, int write,
5181 void __user *buffer, size_t *length,
5188 return proc_dostring(table, write, buffer, length, ppos);
5189 str = memdup_user_nul(buffer, 16);
5191 return PTR_ERR(str);
5193 ret = __parse_numa_zonelist_order(str);
5199 #define MAX_NODE_LOAD (nr_online_nodes)
5200 static int node_load[MAX_NUMNODES];
5203 * find_next_best_node - find the next node that should appear in a given node's fallback list
5204 * @node: node whose fallback list we're appending
5205 * @used_node_mask: nodemask_t of already used nodes
5207 * We use a number of factors to determine which is the next node that should
5208 * appear on a given node's fallback list. The node should not have appeared
5209 * already in @node's fallback list, and it should be the next closest node
5210 * according to the distance array (which contains arbitrary distance values
5211 * from each node to each node in the system), and should also prefer nodes
5212 * with no CPUs, since presumably they'll have very little allocation pressure
5213 * on them otherwise.
5214 * It returns -1 if no node is found.
5216 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5219 int min_val = INT_MAX;
5220 int best_node = NUMA_NO_NODE;
5221 const struct cpumask *tmp = cpumask_of_node(0);
5223 /* Use the local node if we haven't already */
5224 if (!node_isset(node, *used_node_mask)) {
5225 node_set(node, *used_node_mask);
5229 for_each_node_state(n, N_MEMORY) {
5231 /* Don't want a node to appear more than once */
5232 if (node_isset(n, *used_node_mask))
5235 /* Use the distance array to find the distance */
5236 val = node_distance(node, n);
5238 /* Penalize nodes under us ("prefer the next node") */
5241 /* Give preference to headless and unused nodes */
5242 tmp = cpumask_of_node(n);
5243 if (!cpumask_empty(tmp))
5244 val += PENALTY_FOR_NODE_WITH_CPUS;
5246 /* Slight preference for less loaded node */
5247 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5248 val += node_load[n];
5250 if (val < min_val) {
5257 node_set(best_node, *used_node_mask);
5264 * Build zonelists ordered by node and zones within node.
5265 * This results in maximum locality--normal zone overflows into local
5266 * DMA zone, if any--but risks exhausting DMA zone.
5268 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5271 struct zoneref *zonerefs;
5274 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5276 for (i = 0; i < nr_nodes; i++) {
5279 pg_data_t *node = NODE_DATA(node_order[i]);
5281 nr_zones = build_zonerefs_node(node, zonerefs);
5282 zonerefs += nr_zones;
5284 zonerefs->zone = NULL;
5285 zonerefs->zone_idx = 0;
5289 * Build gfp_thisnode zonelists
5291 static void build_thisnode_zonelists(pg_data_t *pgdat)
5293 struct zoneref *zonerefs;
5296 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5297 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5298 zonerefs += nr_zones;
5299 zonerefs->zone = NULL;
5300 zonerefs->zone_idx = 0;
5304 * Build zonelists ordered by zone and nodes within zones.
5305 * This results in conserving DMA zone[s] until all Normal memory is
5306 * exhausted, but results in overflowing to remote node while memory
5307 * may still exist in local DMA zone.
5310 static void build_zonelists(pg_data_t *pgdat)
5312 static int node_order[MAX_NUMNODES];
5313 int node, load, nr_nodes = 0;
5314 nodemask_t used_mask;
5315 int local_node, prev_node;
5317 /* NUMA-aware ordering of nodes */
5318 local_node = pgdat->node_id;
5319 load = nr_online_nodes;
5320 prev_node = local_node;
5321 nodes_clear(used_mask);
5323 memset(node_order, 0, sizeof(node_order));
5324 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5326 * We don't want to pressure a particular node.
5327 * So adding penalty to the first node in same
5328 * distance group to make it round-robin.
5330 if (node_distance(local_node, node) !=
5331 node_distance(local_node, prev_node))
5332 node_load[node] = load;
5334 node_order[nr_nodes++] = node;
5339 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5340 build_thisnode_zonelists(pgdat);
5343 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5345 * Return node id of node used for "local" allocations.
5346 * I.e., first node id of first zone in arg node's generic zonelist.
5347 * Used for initializing percpu 'numa_mem', which is used primarily
5348 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5350 int local_memory_node(int node)
5354 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5355 gfp_zone(GFP_KERNEL),
5357 return zone_to_nid(z->zone);
5361 static void setup_min_unmapped_ratio(void);
5362 static void setup_min_slab_ratio(void);
5363 #else /* CONFIG_NUMA */
5365 static void build_zonelists(pg_data_t *pgdat)
5367 int node, local_node;
5368 struct zoneref *zonerefs;
5371 local_node = pgdat->node_id;
5373 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5374 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5375 zonerefs += nr_zones;
5378 * Now we build the zonelist so that it contains the zones
5379 * of all the other nodes.
5380 * We don't want to pressure a particular node, so when
5381 * building the zones for node N, we make sure that the
5382 * zones coming right after the local ones are those from
5383 * node N+1 (modulo N)
5385 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5386 if (!node_online(node))
5388 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5389 zonerefs += nr_zones;
5391 for (node = 0; node < local_node; node++) {
5392 if (!node_online(node))
5394 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5395 zonerefs += nr_zones;
5398 zonerefs->zone = NULL;
5399 zonerefs->zone_idx = 0;
5402 #endif /* CONFIG_NUMA */
5405 * Boot pageset table. One per cpu which is going to be used for all
5406 * zones and all nodes. The parameters will be set in such a way
5407 * that an item put on a list will immediately be handed over to
5408 * the buddy list. This is safe since pageset manipulation is done
5409 * with interrupts disabled.
5411 * The boot_pagesets must be kept even after bootup is complete for
5412 * unused processors and/or zones. They do play a role for bootstrapping
5413 * hotplugged processors.
5415 * zoneinfo_show() and maybe other functions do
5416 * not check if the processor is online before following the pageset pointer.
5417 * Other parts of the kernel may not check if the zone is available.
5419 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5420 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5421 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5423 static void __build_all_zonelists(void *data)
5426 int __maybe_unused cpu;
5427 pg_data_t *self = data;
5429 write_seqlock(&zonelist_update_seq);
5432 memset(node_load, 0, sizeof(node_load));
5436 * This node is hotadded and no memory is yet present. So just
5437 * building zonelists is fine - no need to touch other nodes.
5439 if (self && !node_online(self->node_id)) {
5440 build_zonelists(self);
5442 for_each_online_node(nid) {
5443 pg_data_t *pgdat = NODE_DATA(nid);
5445 build_zonelists(pgdat);
5448 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5450 * We now know the "local memory node" for each node--
5451 * i.e., the node of the first zone in the generic zonelist.
5452 * Set up numa_mem percpu variable for on-line cpus. During
5453 * boot, only the boot cpu should be on-line; we'll init the
5454 * secondary cpus' numa_mem as they come on-line. During
5455 * node/memory hotplug, we'll fixup all on-line cpus.
5457 for_each_online_cpu(cpu)
5458 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5462 write_sequnlock(&zonelist_update_seq);
5465 static noinline void __init
5466 build_all_zonelists_init(void)
5470 __build_all_zonelists(NULL);
5473 * Initialize the boot_pagesets that are going to be used
5474 * for bootstrapping processors. The real pagesets for
5475 * each zone will be allocated later when the per cpu
5476 * allocator is available.
5478 * boot_pagesets are used also for bootstrapping offline
5479 * cpus if the system is already booted because the pagesets
5480 * are needed to initialize allocators on a specific cpu too.
5481 * F.e. the percpu allocator needs the page allocator which
5482 * needs the percpu allocator in order to allocate its pagesets
5483 * (a chicken-egg dilemma).
5485 for_each_possible_cpu(cpu)
5486 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5488 mminit_verify_zonelist();
5489 cpuset_init_current_mems_allowed();
5493 * unless system_state == SYSTEM_BOOTING.
5495 * __ref due to call of __init annotated helper build_all_zonelists_init
5496 * [protected by SYSTEM_BOOTING].
5498 void __ref build_all_zonelists(pg_data_t *pgdat)
5500 if (system_state == SYSTEM_BOOTING) {
5501 build_all_zonelists_init();
5503 __build_all_zonelists(pgdat);
5504 /* cpuset refresh routine should be here */
5506 vm_total_pages = nr_free_pagecache_pages();
5508 * Disable grouping by mobility if the number of pages in the
5509 * system is too low to allow the mechanism to work. It would be
5510 * more accurate, but expensive to check per-zone. This check is
5511 * made on memory-hotadd so a system can start with mobility
5512 * disabled and enable it later
5514 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5515 page_group_by_mobility_disabled = 1;
5517 page_group_by_mobility_disabled = 0;
5519 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5521 page_group_by_mobility_disabled ? "off" : "on",
5524 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5529 * Initially all pages are reserved - free ones are freed
5530 * up by free_all_bootmem() once the early boot process is
5531 * done. Non-atomic initialization, single-pass.
5533 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5534 unsigned long start_pfn, enum meminit_context context,
5535 struct vmem_altmap *altmap)
5537 unsigned long end_pfn = start_pfn + size;
5538 pg_data_t *pgdat = NODE_DATA(nid);
5540 unsigned long nr_initialised = 0;
5542 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5543 struct memblock_region *r = NULL, *tmp;
5546 if (highest_memmap_pfn < end_pfn - 1)
5547 highest_memmap_pfn = end_pfn - 1;
5550 * Honor reservation requested by the driver for this ZONE_DEVICE
5553 if (altmap && start_pfn == altmap->base_pfn)
5554 start_pfn += altmap->reserve;
5556 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5558 * There can be holes in boot-time mem_map[]s handed to this
5559 * function. They do not exist on hotplugged memory.
5561 if (context != MEMINIT_EARLY)
5564 if (!early_pfn_valid(pfn))
5566 if (!early_pfn_in_nid(pfn, nid))
5568 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5571 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5573 * Check given memblock attribute by firmware which can affect
5574 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5575 * mirrored, it's an overlapped memmap init. skip it.
5577 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5578 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5579 for_each_memblock(memory, tmp)
5580 if (pfn < memblock_region_memory_end_pfn(tmp))
5584 if (pfn >= memblock_region_memory_base_pfn(r) &&
5585 memblock_is_mirror(r)) {
5586 /* already initialized as NORMAL */
5587 pfn = memblock_region_memory_end_pfn(r);
5594 page = pfn_to_page(pfn);
5595 __init_single_page(page, pfn, zone, nid);
5596 if (context == MEMINIT_HOTPLUG)
5597 SetPageReserved(page);
5600 * Mark the block movable so that blocks are reserved for
5601 * movable at startup. This will force kernel allocations
5602 * to reserve their blocks rather than leaking throughout
5603 * the address space during boot when many long-lived
5604 * kernel allocations are made.
5606 * bitmap is created for zone's valid pfn range. but memmap
5607 * can be created for invalid pages (for alignment)
5608 * check here not to call set_pageblock_migratetype() against
5611 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
5612 * because this is done early in sparse_add_one_section
5614 if (!(pfn & (pageblock_nr_pages - 1))) {
5615 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5621 static void __meminit zone_init_free_lists(struct zone *zone)
5623 unsigned int order, t;
5624 for_each_migratetype_order(order, t) {
5625 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5626 zone->free_area[order].nr_free = 0;
5630 #ifndef __HAVE_ARCH_MEMMAP_INIT
5631 #define memmap_init(size, nid, zone, start_pfn) \
5632 memmap_init_zone((size), (nid), (zone), (start_pfn), \
5633 MEMINIT_EARLY, NULL)
5636 static int zone_batchsize(struct zone *zone)
5642 * The per-cpu-pages pools are set to around 1000th of the
5645 batch = zone->managed_pages / 1024;
5646 /* But no more than a meg. */
5647 if (batch * PAGE_SIZE > 1024 * 1024)
5648 batch = (1024 * 1024) / PAGE_SIZE;
5649 batch /= 4; /* We effectively *= 4 below */
5654 * Clamp the batch to a 2^n - 1 value. Having a power
5655 * of 2 value was found to be more likely to have
5656 * suboptimal cache aliasing properties in some cases.
5658 * For example if 2 tasks are alternately allocating
5659 * batches of pages, one task can end up with a lot
5660 * of pages of one half of the possible page colors
5661 * and the other with pages of the other colors.
5663 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5668 /* The deferral and batching of frees should be suppressed under NOMMU
5671 * The problem is that NOMMU needs to be able to allocate large chunks
5672 * of contiguous memory as there's no hardware page translation to
5673 * assemble apparent contiguous memory from discontiguous pages.
5675 * Queueing large contiguous runs of pages for batching, however,
5676 * causes the pages to actually be freed in smaller chunks. As there
5677 * can be a significant delay between the individual batches being
5678 * recycled, this leads to the once large chunks of space being
5679 * fragmented and becoming unavailable for high-order allocations.
5686 * pcp->high and pcp->batch values are related and dependent on one another:
5687 * ->batch must never be higher then ->high.
5688 * The following function updates them in a safe manner without read side
5691 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5692 * those fields changing asynchronously (acording the the above rule).
5694 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5695 * outside of boot time (or some other assurance that no concurrent updaters
5698 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5699 unsigned long batch)
5701 /* start with a fail safe value for batch */
5705 /* Update high, then batch, in order */
5712 /* a companion to pageset_set_high() */
5713 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5715 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5718 static void pageset_init(struct per_cpu_pageset *p)
5720 struct per_cpu_pages *pcp;
5723 memset(p, 0, sizeof(*p));
5727 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5728 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5731 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5734 pageset_set_batch(p, batch);
5738 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5739 * to the value high for the pageset p.
5741 static void pageset_set_high(struct per_cpu_pageset *p,
5744 unsigned long batch = max(1UL, high / 4);
5745 if ((high / 4) > (PAGE_SHIFT * 8))
5746 batch = PAGE_SHIFT * 8;
5748 pageset_update(&p->pcp, high, batch);
5751 static void pageset_set_high_and_batch(struct zone *zone,
5752 struct per_cpu_pageset *pcp)
5754 if (percpu_pagelist_fraction)
5755 pageset_set_high(pcp,
5756 (zone->managed_pages /
5757 percpu_pagelist_fraction));
5759 pageset_set_batch(pcp, zone_batchsize(zone));
5762 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5764 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5767 pageset_set_high_and_batch(zone, pcp);
5770 void __meminit setup_zone_pageset(struct zone *zone)
5773 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5774 for_each_possible_cpu(cpu)
5775 zone_pageset_init(zone, cpu);
5779 * Allocate per cpu pagesets and initialize them.
5780 * Before this call only boot pagesets were available.
5782 void __init setup_per_cpu_pageset(void)
5784 struct pglist_data *pgdat;
5787 for_each_populated_zone(zone)
5788 setup_zone_pageset(zone);
5790 for_each_online_pgdat(pgdat)
5791 pgdat->per_cpu_nodestats =
5792 alloc_percpu(struct per_cpu_nodestat);
5795 static __meminit void zone_pcp_init(struct zone *zone)
5798 * per cpu subsystem is not up at this point. The following code
5799 * relies on the ability of the linker to provide the
5800 * offset of a (static) per cpu variable into the per cpu area.
5802 zone->pageset = &boot_pageset;
5804 if (populated_zone(zone))
5805 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5806 zone->name, zone->present_pages,
5807 zone_batchsize(zone));
5810 void __meminit init_currently_empty_zone(struct zone *zone,
5811 unsigned long zone_start_pfn,
5814 struct pglist_data *pgdat = zone->zone_pgdat;
5815 int zone_idx = zone_idx(zone) + 1;
5817 if (zone_idx > pgdat->nr_zones)
5818 pgdat->nr_zones = zone_idx;
5820 zone->zone_start_pfn = zone_start_pfn;
5822 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5823 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5825 (unsigned long)zone_idx(zone),
5826 zone_start_pfn, (zone_start_pfn + size));
5828 zone_init_free_lists(zone);
5829 zone->initialized = 1;
5832 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5833 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5836 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5838 int __meminit __early_pfn_to_nid(unsigned long pfn,
5839 struct mminit_pfnnid_cache *state)
5841 unsigned long start_pfn, end_pfn;
5844 if (state->last_start <= pfn && pfn < state->last_end)
5845 return state->last_nid;
5847 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5849 state->last_start = start_pfn;
5850 state->last_end = end_pfn;
5851 state->last_nid = nid;
5856 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5859 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5860 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5861 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5863 * If an architecture guarantees that all ranges registered contain no holes
5864 * and may be freed, this this function may be used instead of calling
5865 * memblock_free_early_nid() manually.
5867 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5869 unsigned long start_pfn, end_pfn;
5872 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5873 start_pfn = min(start_pfn, max_low_pfn);
5874 end_pfn = min(end_pfn, max_low_pfn);
5876 if (start_pfn < end_pfn)
5877 memblock_free_early_nid(PFN_PHYS(start_pfn),
5878 (end_pfn - start_pfn) << PAGE_SHIFT,
5884 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5885 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5887 * If an architecture guarantees that all ranges registered contain no holes and may
5888 * be freed, this function may be used instead of calling memory_present() manually.
5890 void __init sparse_memory_present_with_active_regions(int nid)
5892 unsigned long start_pfn, end_pfn;
5895 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5896 memory_present(this_nid, start_pfn, end_pfn);
5900 * get_pfn_range_for_nid - Return the start and end page frames for a node
5901 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5902 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5903 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5905 * It returns the start and end page frame of a node based on information
5906 * provided by memblock_set_node(). If called for a node
5907 * with no available memory, a warning is printed and the start and end
5910 void __meminit get_pfn_range_for_nid(unsigned int nid,
5911 unsigned long *start_pfn, unsigned long *end_pfn)
5913 unsigned long this_start_pfn, this_end_pfn;
5919 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5920 *start_pfn = min(*start_pfn, this_start_pfn);
5921 *end_pfn = max(*end_pfn, this_end_pfn);
5924 if (*start_pfn == -1UL)
5929 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5930 * assumption is made that zones within a node are ordered in monotonic
5931 * increasing memory addresses so that the "highest" populated zone is used
5933 static void __init find_usable_zone_for_movable(void)
5936 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5937 if (zone_index == ZONE_MOVABLE)
5940 if (arch_zone_highest_possible_pfn[zone_index] >
5941 arch_zone_lowest_possible_pfn[zone_index])
5945 VM_BUG_ON(zone_index == -1);
5946 movable_zone = zone_index;
5950 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5951 * because it is sized independent of architecture. Unlike the other zones,
5952 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5953 * in each node depending on the size of each node and how evenly kernelcore
5954 * is distributed. This helper function adjusts the zone ranges
5955 * provided by the architecture for a given node by using the end of the
5956 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5957 * zones within a node are in order of monotonic increases memory addresses
5959 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5960 unsigned long zone_type,
5961 unsigned long node_start_pfn,
5962 unsigned long node_end_pfn,
5963 unsigned long *zone_start_pfn,
5964 unsigned long *zone_end_pfn)
5966 /* Only adjust if ZONE_MOVABLE is on this node */
5967 if (zone_movable_pfn[nid]) {
5968 /* Size ZONE_MOVABLE */
5969 if (zone_type == ZONE_MOVABLE) {
5970 *zone_start_pfn = zone_movable_pfn[nid];
5971 *zone_end_pfn = min(node_end_pfn,
5972 arch_zone_highest_possible_pfn[movable_zone]);
5974 /* Adjust for ZONE_MOVABLE starting within this range */
5975 } else if (!mirrored_kernelcore &&
5976 *zone_start_pfn < zone_movable_pfn[nid] &&
5977 *zone_end_pfn > zone_movable_pfn[nid]) {
5978 *zone_end_pfn = zone_movable_pfn[nid];
5980 /* Check if this whole range is within ZONE_MOVABLE */
5981 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5982 *zone_start_pfn = *zone_end_pfn;
5987 * Return the number of pages a zone spans in a node, including holes
5988 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5990 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5991 unsigned long zone_type,
5992 unsigned long node_start_pfn,
5993 unsigned long node_end_pfn,
5994 unsigned long *zone_start_pfn,
5995 unsigned long *zone_end_pfn,
5996 unsigned long *ignored)
5998 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5999 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6000 /* When hotadd a new node from cpu_up(), the node should be empty */
6001 if (!node_start_pfn && !node_end_pfn)
6004 /* Get the start and end of the zone */
6005 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6006 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6007 adjust_zone_range_for_zone_movable(nid, zone_type,
6008 node_start_pfn, node_end_pfn,
6009 zone_start_pfn, zone_end_pfn);
6011 /* Check that this node has pages within the zone's required range */
6012 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6015 /* Move the zone boundaries inside the node if necessary */
6016 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6017 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6019 /* Return the spanned pages */
6020 return *zone_end_pfn - *zone_start_pfn;
6024 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6025 * then all holes in the requested range will be accounted for.
6027 unsigned long __meminit __absent_pages_in_range(int nid,
6028 unsigned long range_start_pfn,
6029 unsigned long range_end_pfn)
6031 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6032 unsigned long start_pfn, end_pfn;
6035 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6036 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6037 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6038 nr_absent -= end_pfn - start_pfn;
6044 * absent_pages_in_range - Return number of page frames in holes within a range
6045 * @start_pfn: The start PFN to start searching for holes
6046 * @end_pfn: The end PFN to stop searching for holes
6048 * It returns the number of pages frames in memory holes within a range.
6050 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6051 unsigned long end_pfn)
6053 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6056 /* Return the number of page frames in holes in a zone on a node */
6057 static unsigned long __meminit zone_absent_pages_in_node(int nid,
6058 unsigned long zone_type,
6059 unsigned long node_start_pfn,
6060 unsigned long node_end_pfn,
6061 unsigned long *ignored)
6063 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6064 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6065 unsigned long zone_start_pfn, zone_end_pfn;
6066 unsigned long nr_absent;
6068 /* When hotadd a new node from cpu_up(), the node should be empty */
6069 if (!node_start_pfn && !node_end_pfn)
6072 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6073 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6075 adjust_zone_range_for_zone_movable(nid, zone_type,
6076 node_start_pfn, node_end_pfn,
6077 &zone_start_pfn, &zone_end_pfn);
6078 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6081 * ZONE_MOVABLE handling.
6082 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6085 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6086 unsigned long start_pfn, end_pfn;
6087 struct memblock_region *r;
6089 for_each_memblock(memory, r) {
6090 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6091 zone_start_pfn, zone_end_pfn);
6092 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6093 zone_start_pfn, zone_end_pfn);
6095 if (zone_type == ZONE_MOVABLE &&
6096 memblock_is_mirror(r))
6097 nr_absent += end_pfn - start_pfn;
6099 if (zone_type == ZONE_NORMAL &&
6100 !memblock_is_mirror(r))
6101 nr_absent += end_pfn - start_pfn;
6108 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6109 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
6110 unsigned long zone_type,
6111 unsigned long node_start_pfn,
6112 unsigned long node_end_pfn,
6113 unsigned long *zone_start_pfn,
6114 unsigned long *zone_end_pfn,
6115 unsigned long *zones_size)
6119 *zone_start_pfn = node_start_pfn;
6120 for (zone = 0; zone < zone_type; zone++)
6121 *zone_start_pfn += zones_size[zone];
6123 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6125 return zones_size[zone_type];
6128 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
6129 unsigned long zone_type,
6130 unsigned long node_start_pfn,
6131 unsigned long node_end_pfn,
6132 unsigned long *zholes_size)
6137 return zholes_size[zone_type];
6140 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6142 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
6143 unsigned long node_start_pfn,
6144 unsigned long node_end_pfn,
6145 unsigned long *zones_size,
6146 unsigned long *zholes_size)
6148 unsigned long realtotalpages = 0, totalpages = 0;
6151 for (i = 0; i < MAX_NR_ZONES; i++) {
6152 struct zone *zone = pgdat->node_zones + i;
6153 unsigned long zone_start_pfn, zone_end_pfn;
6154 unsigned long size, real_size;
6156 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6162 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6163 node_start_pfn, node_end_pfn,
6166 zone->zone_start_pfn = zone_start_pfn;
6168 zone->zone_start_pfn = 0;
6169 zone->spanned_pages = size;
6170 zone->present_pages = real_size;
6173 realtotalpages += real_size;
6176 pgdat->node_spanned_pages = totalpages;
6177 pgdat->node_present_pages = realtotalpages;
6178 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6182 #ifndef CONFIG_SPARSEMEM
6184 * Calculate the size of the zone->blockflags rounded to an unsigned long
6185 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6186 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6187 * round what is now in bits to nearest long in bits, then return it in
6190 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6192 unsigned long usemapsize;
6194 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6195 usemapsize = roundup(zonesize, pageblock_nr_pages);
6196 usemapsize = usemapsize >> pageblock_order;
6197 usemapsize *= NR_PAGEBLOCK_BITS;
6198 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6200 return usemapsize / 8;
6203 static void __ref setup_usemap(struct pglist_data *pgdat,
6205 unsigned long zone_start_pfn,
6206 unsigned long zonesize)
6208 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6209 zone->pageblock_flags = NULL;
6211 zone->pageblock_flags =
6212 memblock_virt_alloc_node_nopanic(usemapsize,
6216 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6217 unsigned long zone_start_pfn, unsigned long zonesize) {}
6218 #endif /* CONFIG_SPARSEMEM */
6220 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6222 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6223 void __init set_pageblock_order(void)
6227 /* Check that pageblock_nr_pages has not already been setup */
6228 if (pageblock_order)
6231 if (HPAGE_SHIFT > PAGE_SHIFT)
6232 order = HUGETLB_PAGE_ORDER;
6234 order = MAX_ORDER - 1;
6237 * Assume the largest contiguous order of interest is a huge page.
6238 * This value may be variable depending on boot parameters on IA64 and
6241 pageblock_order = order;
6243 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6246 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6247 * is unused as pageblock_order is set at compile-time. See
6248 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6251 void __init set_pageblock_order(void)
6255 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6257 static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6258 unsigned long present_pages)
6260 unsigned long pages = spanned_pages;
6263 * Provide a more accurate estimation if there are holes within
6264 * the zone and SPARSEMEM is in use. If there are holes within the
6265 * zone, each populated memory region may cost us one or two extra
6266 * memmap pages due to alignment because memmap pages for each
6267 * populated regions may not be naturally aligned on page boundary.
6268 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6270 if (spanned_pages > present_pages + (present_pages >> 4) &&
6271 IS_ENABLED(CONFIG_SPARSEMEM))
6272 pages = present_pages;
6274 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6277 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6278 static void pgdat_init_split_queue(struct pglist_data *pgdat)
6280 spin_lock_init(&pgdat->split_queue_lock);
6281 INIT_LIST_HEAD(&pgdat->split_queue);
6282 pgdat->split_queue_len = 0;
6285 static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6288 #ifdef CONFIG_COMPACTION
6289 static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6291 init_waitqueue_head(&pgdat->kcompactd_wait);
6294 static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6297 static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6299 pgdat_resize_init(pgdat);
6301 pgdat_init_split_queue(pgdat);
6302 pgdat_init_kcompactd(pgdat);
6304 init_waitqueue_head(&pgdat->kswapd_wait);
6305 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6307 pgdat_page_ext_init(pgdat);
6308 spin_lock_init(&pgdat->lru_lock);
6309 lruvec_init(node_lruvec(pgdat));
6312 static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6313 unsigned long remaining_pages)
6315 zone->managed_pages = remaining_pages;
6316 zone_set_nid(zone, nid);
6317 zone->name = zone_names[idx];
6318 zone->zone_pgdat = NODE_DATA(nid);
6319 spin_lock_init(&zone->lock);
6320 zone_seqlock_init(zone);
6321 zone_pcp_init(zone);
6325 * Set up the zone data structures
6326 * - init pgdat internals
6327 * - init all zones belonging to this node
6329 * NOTE: this function is only called during memory hotplug
6331 #ifdef CONFIG_MEMORY_HOTPLUG
6332 void __ref free_area_init_core_hotplug(int nid)
6335 pg_data_t *pgdat = NODE_DATA(nid);
6337 pgdat_init_internals(pgdat);
6338 for (z = 0; z < MAX_NR_ZONES; z++)
6339 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6344 * Set up the zone data structures:
6345 * - mark all pages reserved
6346 * - mark all memory queues empty
6347 * - clear the memory bitmaps
6349 * NOTE: pgdat should get zeroed by caller.
6350 * NOTE: this function is only called during early init.
6352 static void __init free_area_init_core(struct pglist_data *pgdat)
6355 int nid = pgdat->node_id;
6357 pgdat_init_internals(pgdat);
6358 pgdat->per_cpu_nodestats = &boot_nodestats;
6360 for (j = 0; j < MAX_NR_ZONES; j++) {
6361 struct zone *zone = pgdat->node_zones + j;
6362 unsigned long size, freesize, memmap_pages;
6363 unsigned long zone_start_pfn = zone->zone_start_pfn;
6365 size = zone->spanned_pages;
6366 freesize = zone->present_pages;
6369 * Adjust freesize so that it accounts for how much memory
6370 * is used by this zone for memmap. This affects the watermark
6371 * and per-cpu initialisations
6373 memmap_pages = calc_memmap_size(size, freesize);
6374 if (!is_highmem_idx(j)) {
6375 if (freesize >= memmap_pages) {
6376 freesize -= memmap_pages;
6379 " %s zone: %lu pages used for memmap\n",
6380 zone_names[j], memmap_pages);
6382 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6383 zone_names[j], memmap_pages, freesize);
6386 /* Account for reserved pages */
6387 if (j == 0 && freesize > dma_reserve) {
6388 freesize -= dma_reserve;
6389 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6390 zone_names[0], dma_reserve);
6393 if (!is_highmem_idx(j))
6394 nr_kernel_pages += freesize;
6395 /* Charge for highmem memmap if there are enough kernel pages */
6396 else if (nr_kernel_pages > memmap_pages * 2)
6397 nr_kernel_pages -= memmap_pages;
6398 nr_all_pages += freesize;
6401 * Set an approximate value for lowmem here, it will be adjusted
6402 * when the bootmem allocator frees pages into the buddy system.
6403 * And all highmem pages will be managed by the buddy system.
6405 zone_init_internals(zone, j, nid, freesize);
6410 set_pageblock_order();
6411 setup_usemap(pgdat, zone, zone_start_pfn, size);
6412 init_currently_empty_zone(zone, zone_start_pfn, size);
6413 memmap_init(size, nid, j, zone_start_pfn);
6417 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6418 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6420 unsigned long __maybe_unused start = 0;
6421 unsigned long __maybe_unused offset = 0;
6423 /* Skip empty nodes */
6424 if (!pgdat->node_spanned_pages)
6427 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6428 offset = pgdat->node_start_pfn - start;
6429 /* ia64 gets its own node_mem_map, before this, without bootmem */
6430 if (!pgdat->node_mem_map) {
6431 unsigned long size, end;
6435 * The zone's endpoints aren't required to be MAX_ORDER
6436 * aligned but the node_mem_map endpoints must be in order
6437 * for the buddy allocator to function correctly.
6439 end = pgdat_end_pfn(pgdat);
6440 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6441 size = (end - start) * sizeof(struct page);
6442 map = memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
6443 pgdat->node_mem_map = map + offset;
6445 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6446 __func__, pgdat->node_id, (unsigned long)pgdat,
6447 (unsigned long)pgdat->node_mem_map);
6448 #ifndef CONFIG_NEED_MULTIPLE_NODES
6450 * With no DISCONTIG, the global mem_map is just set as node 0's
6452 if (pgdat == NODE_DATA(0)) {
6453 mem_map = NODE_DATA(0)->node_mem_map;
6454 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6455 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6457 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6462 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6463 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6465 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6466 static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6469 * We start only with one section of pages, more pages are added as
6470 * needed until the rest of deferred pages are initialized.
6472 pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
6473 pgdat->node_spanned_pages);
6474 pgdat->first_deferred_pfn = ULONG_MAX;
6477 static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6480 void __init free_area_init_node(int nid, unsigned long *zones_size,
6481 unsigned long node_start_pfn,
6482 unsigned long *zholes_size)
6484 pg_data_t *pgdat = NODE_DATA(nid);
6485 unsigned long start_pfn = 0;
6486 unsigned long end_pfn = 0;
6488 /* pg_data_t should be reset to zero when it's allocated */
6489 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6491 pgdat->node_id = nid;
6492 pgdat->node_start_pfn = node_start_pfn;
6493 pgdat->per_cpu_nodestats = NULL;
6494 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6495 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6496 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6497 (u64)start_pfn << PAGE_SHIFT,
6498 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6500 start_pfn = node_start_pfn;
6502 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6503 zones_size, zholes_size);
6505 alloc_node_mem_map(pgdat);
6506 pgdat_set_deferred_range(pgdat);
6508 free_area_init_core(pgdat);
6511 #if defined(CONFIG_HAVE_MEMBLOCK) && !defined(CONFIG_FLAT_NODE_MEM_MAP)
6514 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6517 static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6522 for (pfn = spfn; pfn < epfn; pfn++) {
6523 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6524 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6525 + pageblock_nr_pages - 1;
6528 mm_zero_struct_page(pfn_to_page(pfn));
6536 * Only struct pages that are backed by physical memory are zeroed and
6537 * initialized by going through __init_single_page(). But, there are some
6538 * struct pages which are reserved in memblock allocator and their fields
6539 * may be accessed (for example page_to_pfn() on some configuration accesses
6540 * flags). We must explicitly zero those struct pages.
6542 * This function also addresses a similar issue where struct pages are left
6543 * uninitialized because the physical address range is not covered by
6544 * memblock.memory or memblock.reserved. That could happen when memblock
6545 * layout is manually configured via memmap=, or when the highest physical
6546 * address (max_pfn) does not end on a section boundary.
6548 void __init zero_resv_unavail(void)
6550 phys_addr_t start, end;
6552 phys_addr_t next = 0;
6555 * Loop through unavailable ranges not covered by memblock.memory.
6558 for_each_mem_range(i, &memblock.memory, NULL,
6559 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6561 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6566 * Early sections always have a fully populated memmap for the whole
6567 * section - see pfn_valid(). If the last section has holes at the
6568 * end and that section is marked "online", the memmap will be
6569 * considered initialized. Make sure that memmap has a well defined
6572 pgcnt += zero_pfn_range(PFN_DOWN(next),
6573 round_up(max_pfn, PAGES_PER_SECTION));
6576 * Struct pages that do not have backing memory. This could be because
6577 * firmware is using some of this memory, or for some other reasons.
6580 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6582 #endif /* CONFIG_HAVE_MEMBLOCK && !CONFIG_FLAT_NODE_MEM_MAP */
6584 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6586 #if MAX_NUMNODES > 1
6588 * Figure out the number of possible node ids.
6590 void __init setup_nr_node_ids(void)
6592 unsigned int highest;
6594 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6595 nr_node_ids = highest + 1;
6600 * node_map_pfn_alignment - determine the maximum internode alignment
6602 * This function should be called after node map is populated and sorted.
6603 * It calculates the maximum power of two alignment which can distinguish
6606 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6607 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6608 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6609 * shifted, 1GiB is enough and this function will indicate so.
6611 * This is used to test whether pfn -> nid mapping of the chosen memory
6612 * model has fine enough granularity to avoid incorrect mapping for the
6613 * populated node map.
6615 * Returns the determined alignment in pfn's. 0 if there is no alignment
6616 * requirement (single node).
6618 unsigned long __init node_map_pfn_alignment(void)
6620 unsigned long accl_mask = 0, last_end = 0;
6621 unsigned long start, end, mask;
6625 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6626 if (!start || last_nid < 0 || last_nid == nid) {
6633 * Start with a mask granular enough to pin-point to the
6634 * start pfn and tick off bits one-by-one until it becomes
6635 * too coarse to separate the current node from the last.
6637 mask = ~((1 << __ffs(start)) - 1);
6638 while (mask && last_end <= (start & (mask << 1)))
6641 /* accumulate all internode masks */
6645 /* convert mask to number of pages */
6646 return ~accl_mask + 1;
6649 /* Find the lowest pfn for a node */
6650 static unsigned long __init find_min_pfn_for_node(int nid)
6652 unsigned long min_pfn = ULONG_MAX;
6653 unsigned long start_pfn;
6656 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6657 min_pfn = min(min_pfn, start_pfn);
6659 if (min_pfn == ULONG_MAX) {
6660 pr_warn("Could not find start_pfn for node %d\n", nid);
6668 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6670 * It returns the minimum PFN based on information provided via
6671 * memblock_set_node().
6673 unsigned long __init find_min_pfn_with_active_regions(void)
6675 return find_min_pfn_for_node(MAX_NUMNODES);
6679 * early_calculate_totalpages()
6680 * Sum pages in active regions for movable zone.
6681 * Populate N_MEMORY for calculating usable_nodes.
6683 static unsigned long __init early_calculate_totalpages(void)
6685 unsigned long totalpages = 0;
6686 unsigned long start_pfn, end_pfn;
6689 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6690 unsigned long pages = end_pfn - start_pfn;
6692 totalpages += pages;
6694 node_set_state(nid, N_MEMORY);
6700 * Find the PFN the Movable zone begins in each node. Kernel memory
6701 * is spread evenly between nodes as long as the nodes have enough
6702 * memory. When they don't, some nodes will have more kernelcore than
6705 static void __init find_zone_movable_pfns_for_nodes(void)
6708 unsigned long usable_startpfn;
6709 unsigned long kernelcore_node, kernelcore_remaining;
6710 /* save the state before borrow the nodemask */
6711 nodemask_t saved_node_state = node_states[N_MEMORY];
6712 unsigned long totalpages = early_calculate_totalpages();
6713 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6714 struct memblock_region *r;
6716 /* Need to find movable_zone earlier when movable_node is specified. */
6717 find_usable_zone_for_movable();
6720 * If movable_node is specified, ignore kernelcore and movablecore
6723 if (movable_node_is_enabled()) {
6724 for_each_memblock(memory, r) {
6725 if (!memblock_is_hotpluggable(r))
6730 usable_startpfn = PFN_DOWN(r->base);
6731 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6732 min(usable_startpfn, zone_movable_pfn[nid]) :
6740 * If kernelcore=mirror is specified, ignore movablecore option
6742 if (mirrored_kernelcore) {
6743 bool mem_below_4gb_not_mirrored = false;
6745 for_each_memblock(memory, r) {
6746 if (memblock_is_mirror(r))
6751 usable_startpfn = memblock_region_memory_base_pfn(r);
6753 if (usable_startpfn < 0x100000) {
6754 mem_below_4gb_not_mirrored = true;
6758 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6759 min(usable_startpfn, zone_movable_pfn[nid]) :
6763 if (mem_below_4gb_not_mirrored)
6764 pr_warn("This configuration results in unmirrored kernel memory.");
6770 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6771 * amount of necessary memory.
6773 if (required_kernelcore_percent)
6774 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6776 if (required_movablecore_percent)
6777 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6781 * If movablecore= was specified, calculate what size of
6782 * kernelcore that corresponds so that memory usable for
6783 * any allocation type is evenly spread. If both kernelcore
6784 * and movablecore are specified, then the value of kernelcore
6785 * will be used for required_kernelcore if it's greater than
6786 * what movablecore would have allowed.
6788 if (required_movablecore) {
6789 unsigned long corepages;
6792 * Round-up so that ZONE_MOVABLE is at least as large as what
6793 * was requested by the user
6795 required_movablecore =
6796 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6797 required_movablecore = min(totalpages, required_movablecore);
6798 corepages = totalpages - required_movablecore;
6800 required_kernelcore = max(required_kernelcore, corepages);
6804 * If kernelcore was not specified or kernelcore size is larger
6805 * than totalpages, there is no ZONE_MOVABLE.
6807 if (!required_kernelcore || required_kernelcore >= totalpages)
6810 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6811 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6814 /* Spread kernelcore memory as evenly as possible throughout nodes */
6815 kernelcore_node = required_kernelcore / usable_nodes;
6816 for_each_node_state(nid, N_MEMORY) {
6817 unsigned long start_pfn, end_pfn;
6820 * Recalculate kernelcore_node if the division per node
6821 * now exceeds what is necessary to satisfy the requested
6822 * amount of memory for the kernel
6824 if (required_kernelcore < kernelcore_node)
6825 kernelcore_node = required_kernelcore / usable_nodes;
6828 * As the map is walked, we track how much memory is usable
6829 * by the kernel using kernelcore_remaining. When it is
6830 * 0, the rest of the node is usable by ZONE_MOVABLE
6832 kernelcore_remaining = kernelcore_node;
6834 /* Go through each range of PFNs within this node */
6835 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6836 unsigned long size_pages;
6838 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6839 if (start_pfn >= end_pfn)
6842 /* Account for what is only usable for kernelcore */
6843 if (start_pfn < usable_startpfn) {
6844 unsigned long kernel_pages;
6845 kernel_pages = min(end_pfn, usable_startpfn)
6848 kernelcore_remaining -= min(kernel_pages,
6849 kernelcore_remaining);
6850 required_kernelcore -= min(kernel_pages,
6851 required_kernelcore);
6853 /* Continue if range is now fully accounted */
6854 if (end_pfn <= usable_startpfn) {
6857 * Push zone_movable_pfn to the end so
6858 * that if we have to rebalance
6859 * kernelcore across nodes, we will
6860 * not double account here
6862 zone_movable_pfn[nid] = end_pfn;
6865 start_pfn = usable_startpfn;
6869 * The usable PFN range for ZONE_MOVABLE is from
6870 * start_pfn->end_pfn. Calculate size_pages as the
6871 * number of pages used as kernelcore
6873 size_pages = end_pfn - start_pfn;
6874 if (size_pages > kernelcore_remaining)
6875 size_pages = kernelcore_remaining;
6876 zone_movable_pfn[nid] = start_pfn + size_pages;
6879 * Some kernelcore has been met, update counts and
6880 * break if the kernelcore for this node has been
6883 required_kernelcore -= min(required_kernelcore,
6885 kernelcore_remaining -= size_pages;
6886 if (!kernelcore_remaining)
6892 * If there is still required_kernelcore, we do another pass with one
6893 * less node in the count. This will push zone_movable_pfn[nid] further
6894 * along on the nodes that still have memory until kernelcore is
6898 if (usable_nodes && required_kernelcore > usable_nodes)
6902 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6903 for (nid = 0; nid < MAX_NUMNODES; nid++) {
6904 unsigned long start_pfn, end_pfn;
6906 zone_movable_pfn[nid] =
6907 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6909 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6910 if (zone_movable_pfn[nid] >= end_pfn)
6911 zone_movable_pfn[nid] = 0;
6915 /* restore the node_state */
6916 node_states[N_MEMORY] = saved_node_state;
6919 /* Any regular or high memory on that node ? */
6920 static void check_for_memory(pg_data_t *pgdat, int nid)
6922 enum zone_type zone_type;
6924 if (N_MEMORY == N_NORMAL_MEMORY)
6927 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6928 struct zone *zone = &pgdat->node_zones[zone_type];
6929 if (populated_zone(zone)) {
6930 node_set_state(nid, N_HIGH_MEMORY);
6931 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6932 zone_type <= ZONE_NORMAL)
6933 node_set_state(nid, N_NORMAL_MEMORY);
6940 * free_area_init_nodes - Initialise all pg_data_t and zone data
6941 * @max_zone_pfn: an array of max PFNs for each zone
6943 * This will call free_area_init_node() for each active node in the system.
6944 * Using the page ranges provided by memblock_set_node(), the size of each
6945 * zone in each node and their holes is calculated. If the maximum PFN
6946 * between two adjacent zones match, it is assumed that the zone is empty.
6947 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6948 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6949 * starts where the previous one ended. For example, ZONE_DMA32 starts
6950 * at arch_max_dma_pfn.
6952 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6954 unsigned long start_pfn, end_pfn;
6957 /* Record where the zone boundaries are */
6958 memset(arch_zone_lowest_possible_pfn, 0,
6959 sizeof(arch_zone_lowest_possible_pfn));
6960 memset(arch_zone_highest_possible_pfn, 0,
6961 sizeof(arch_zone_highest_possible_pfn));
6963 start_pfn = find_min_pfn_with_active_regions();
6965 for (i = 0; i < MAX_NR_ZONES; i++) {
6966 if (i == ZONE_MOVABLE)
6969 end_pfn = max(max_zone_pfn[i], start_pfn);
6970 arch_zone_lowest_possible_pfn[i] = start_pfn;
6971 arch_zone_highest_possible_pfn[i] = end_pfn;
6973 start_pfn = end_pfn;
6976 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6977 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6978 find_zone_movable_pfns_for_nodes();
6980 /* Print out the zone ranges */
6981 pr_info("Zone ranges:\n");
6982 for (i = 0; i < MAX_NR_ZONES; i++) {
6983 if (i == ZONE_MOVABLE)
6985 pr_info(" %-8s ", zone_names[i]);
6986 if (arch_zone_lowest_possible_pfn[i] ==
6987 arch_zone_highest_possible_pfn[i])
6990 pr_cont("[mem %#018Lx-%#018Lx]\n",
6991 (u64)arch_zone_lowest_possible_pfn[i]
6993 ((u64)arch_zone_highest_possible_pfn[i]
6994 << PAGE_SHIFT) - 1);
6997 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6998 pr_info("Movable zone start for each node\n");
6999 for (i = 0; i < MAX_NUMNODES; i++) {
7000 if (zone_movable_pfn[i])
7001 pr_info(" Node %d: %#018Lx\n", i,
7002 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7005 /* Print out the early node map */
7006 pr_info("Early memory node ranges\n");
7007 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
7008 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7009 (u64)start_pfn << PAGE_SHIFT,
7010 ((u64)end_pfn << PAGE_SHIFT) - 1);
7012 /* Initialise every node */
7013 mminit_verify_pageflags_layout();
7014 setup_nr_node_ids();
7015 zero_resv_unavail();
7016 for_each_online_node(nid) {
7017 pg_data_t *pgdat = NODE_DATA(nid);
7018 free_area_init_node(nid, NULL,
7019 find_min_pfn_for_node(nid), NULL);
7021 /* Any memory on that node */
7022 if (pgdat->node_present_pages)
7023 node_set_state(nid, N_MEMORY);
7024 check_for_memory(pgdat, nid);
7028 static int __init cmdline_parse_core(char *p, unsigned long *core,
7029 unsigned long *percent)
7031 unsigned long long coremem;
7037 /* Value may be a percentage of total memory, otherwise bytes */
7038 coremem = simple_strtoull(p, &endptr, 0);
7039 if (*endptr == '%') {
7040 /* Paranoid check for percent values greater than 100 */
7041 WARN_ON(coremem > 100);
7045 coremem = memparse(p, &p);
7046 /* Paranoid check that UL is enough for the coremem value */
7047 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7049 *core = coremem >> PAGE_SHIFT;
7056 * kernelcore=size sets the amount of memory for use for allocations that
7057 * cannot be reclaimed or migrated.
7059 static int __init cmdline_parse_kernelcore(char *p)
7061 /* parse kernelcore=mirror */
7062 if (parse_option_str(p, "mirror")) {
7063 mirrored_kernelcore = true;
7067 return cmdline_parse_core(p, &required_kernelcore,
7068 &required_kernelcore_percent);
7072 * movablecore=size sets the amount of memory for use for allocations that
7073 * can be reclaimed or migrated.
7075 static int __init cmdline_parse_movablecore(char *p)
7077 return cmdline_parse_core(p, &required_movablecore,
7078 &required_movablecore_percent);
7081 early_param("kernelcore", cmdline_parse_kernelcore);
7082 early_param("movablecore", cmdline_parse_movablecore);
7084 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7086 void adjust_managed_page_count(struct page *page, long count)
7088 spin_lock(&managed_page_count_lock);
7089 page_zone(page)->managed_pages += count;
7090 totalram_pages += count;
7091 #ifdef CONFIG_HIGHMEM
7092 if (PageHighMem(page))
7093 totalhigh_pages += count;
7095 spin_unlock(&managed_page_count_lock);
7097 EXPORT_SYMBOL(adjust_managed_page_count);
7099 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
7102 unsigned long pages = 0;
7104 start = (void *)PAGE_ALIGN((unsigned long)start);
7105 end = (void *)((unsigned long)end & PAGE_MASK);
7106 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7107 struct page *page = virt_to_page(pos);
7108 void *direct_map_addr;
7111 * 'direct_map_addr' might be different from 'pos'
7112 * because some architectures' virt_to_page()
7113 * work with aliases. Getting the direct map
7114 * address ensures that we get a _writeable_
7115 * alias for the memset().
7117 direct_map_addr = page_address(page);
7118 if ((unsigned int)poison <= 0xFF)
7119 memset(direct_map_addr, poison, PAGE_SIZE);
7121 free_reserved_page(page);
7125 pr_info("Freeing %s memory: %ldK\n",
7126 s, pages << (PAGE_SHIFT - 10));
7130 EXPORT_SYMBOL(free_reserved_area);
7132 #ifdef CONFIG_HIGHMEM
7133 void free_highmem_page(struct page *page)
7135 __free_reserved_page(page);
7137 page_zone(page)->managed_pages++;
7143 void __init mem_init_print_info(const char *str)
7145 unsigned long physpages, codesize, datasize, rosize, bss_size;
7146 unsigned long init_code_size, init_data_size;
7148 physpages = get_num_physpages();
7149 codesize = _etext - _stext;
7150 datasize = _edata - _sdata;
7151 rosize = __end_rodata - __start_rodata;
7152 bss_size = __bss_stop - __bss_start;
7153 init_data_size = __init_end - __init_begin;
7154 init_code_size = _einittext - _sinittext;
7157 * Detect special cases and adjust section sizes accordingly:
7158 * 1) .init.* may be embedded into .data sections
7159 * 2) .init.text.* may be out of [__init_begin, __init_end],
7160 * please refer to arch/tile/kernel/vmlinux.lds.S.
7161 * 3) .rodata.* may be embedded into .text or .data sections.
7163 #define adj_init_size(start, end, size, pos, adj) \
7165 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
7169 adj_init_size(__init_begin, __init_end, init_data_size,
7170 _sinittext, init_code_size);
7171 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7172 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7173 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7174 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7176 #undef adj_init_size
7178 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7179 #ifdef CONFIG_HIGHMEM
7183 nr_free_pages() << (PAGE_SHIFT - 10),
7184 physpages << (PAGE_SHIFT - 10),
7185 codesize >> 10, datasize >> 10, rosize >> 10,
7186 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7187 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
7188 totalcma_pages << (PAGE_SHIFT - 10),
7189 #ifdef CONFIG_HIGHMEM
7190 totalhigh_pages << (PAGE_SHIFT - 10),
7192 str ? ", " : "", str ? str : "");
7196 * set_dma_reserve - set the specified number of pages reserved in the first zone
7197 * @new_dma_reserve: The number of pages to mark reserved
7199 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7200 * In the DMA zone, a significant percentage may be consumed by kernel image
7201 * and other unfreeable allocations which can skew the watermarks badly. This
7202 * function may optionally be used to account for unfreeable pages in the
7203 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7204 * smaller per-cpu batchsize.
7206 void __init set_dma_reserve(unsigned long new_dma_reserve)
7208 dma_reserve = new_dma_reserve;
7211 void __init free_area_init(unsigned long *zones_size)
7213 zero_resv_unavail();
7214 free_area_init_node(0, zones_size,
7215 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7218 static int page_alloc_cpu_dead(unsigned int cpu)
7221 lru_add_drain_cpu(cpu);
7225 * Spill the event counters of the dead processor
7226 * into the current processors event counters.
7227 * This artificially elevates the count of the current
7230 vm_events_fold_cpu(cpu);
7233 * Zero the differential counters of the dead processor
7234 * so that the vm statistics are consistent.
7236 * This is only okay since the processor is dead and cannot
7237 * race with what we are doing.
7239 cpu_vm_stats_fold(cpu);
7243 void __init page_alloc_init(void)
7247 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7248 "mm/page_alloc:dead", NULL,
7249 page_alloc_cpu_dead);
7254 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7255 * or min_free_kbytes changes.
7257 static void calculate_totalreserve_pages(void)
7259 struct pglist_data *pgdat;
7260 unsigned long reserve_pages = 0;
7261 enum zone_type i, j;
7263 for_each_online_pgdat(pgdat) {
7265 pgdat->totalreserve_pages = 0;
7267 for (i = 0; i < MAX_NR_ZONES; i++) {
7268 struct zone *zone = pgdat->node_zones + i;
7271 /* Find valid and maximum lowmem_reserve in the zone */
7272 for (j = i; j < MAX_NR_ZONES; j++) {
7273 if (zone->lowmem_reserve[j] > max)
7274 max = zone->lowmem_reserve[j];
7277 /* we treat the high watermark as reserved pages. */
7278 max += high_wmark_pages(zone);
7280 if (max > zone->managed_pages)
7281 max = zone->managed_pages;
7283 pgdat->totalreserve_pages += max;
7285 reserve_pages += max;
7288 totalreserve_pages = reserve_pages;
7292 * setup_per_zone_lowmem_reserve - called whenever
7293 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7294 * has a correct pages reserved value, so an adequate number of
7295 * pages are left in the zone after a successful __alloc_pages().
7297 static void setup_per_zone_lowmem_reserve(void)
7299 struct pglist_data *pgdat;
7300 enum zone_type j, idx;
7302 for_each_online_pgdat(pgdat) {
7303 for (j = 0; j < MAX_NR_ZONES; j++) {
7304 struct zone *zone = pgdat->node_zones + j;
7305 unsigned long managed_pages = zone->managed_pages;
7307 zone->lowmem_reserve[j] = 0;
7311 struct zone *lower_zone;
7314 lower_zone = pgdat->node_zones + idx;
7316 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7317 sysctl_lowmem_reserve_ratio[idx] = 0;
7318 lower_zone->lowmem_reserve[j] = 0;
7320 lower_zone->lowmem_reserve[j] =
7321 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7323 managed_pages += lower_zone->managed_pages;
7328 /* update totalreserve_pages */
7329 calculate_totalreserve_pages();
7332 static void __setup_per_zone_wmarks(void)
7334 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7335 unsigned long lowmem_pages = 0;
7337 unsigned long flags;
7339 /* Calculate total number of !ZONE_HIGHMEM pages */
7340 for_each_zone(zone) {
7341 if (!is_highmem(zone))
7342 lowmem_pages += zone->managed_pages;
7345 for_each_zone(zone) {
7348 spin_lock_irqsave(&zone->lock, flags);
7349 tmp = (u64)pages_min * zone->managed_pages;
7350 do_div(tmp, lowmem_pages);
7351 if (is_highmem(zone)) {
7353 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7354 * need highmem pages, so cap pages_min to a small
7357 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7358 * deltas control asynch page reclaim, and so should
7359 * not be capped for highmem.
7361 unsigned long min_pages;
7363 min_pages = zone->managed_pages / 1024;
7364 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7365 zone->watermark[WMARK_MIN] = min_pages;
7368 * If it's a lowmem zone, reserve a number of pages
7369 * proportionate to the zone's size.
7371 zone->watermark[WMARK_MIN] = tmp;
7375 * Set the kswapd watermarks distance according to the
7376 * scale factor in proportion to available memory, but
7377 * ensure a minimum size on small systems.
7379 tmp = max_t(u64, tmp >> 2,
7380 mult_frac(zone->managed_pages,
7381 watermark_scale_factor, 10000));
7383 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7384 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7386 spin_unlock_irqrestore(&zone->lock, flags);
7389 /* update totalreserve_pages */
7390 calculate_totalreserve_pages();
7394 * setup_per_zone_wmarks - called when min_free_kbytes changes
7395 * or when memory is hot-{added|removed}
7397 * Ensures that the watermark[min,low,high] values for each zone are set
7398 * correctly with respect to min_free_kbytes.
7400 void setup_per_zone_wmarks(void)
7402 static DEFINE_SPINLOCK(lock);
7405 __setup_per_zone_wmarks();
7410 * Initialise min_free_kbytes.
7412 * For small machines we want it small (128k min). For large machines
7413 * we want it large (64MB max). But it is not linear, because network
7414 * bandwidth does not increase linearly with machine size. We use
7416 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7417 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7433 int __meminit init_per_zone_wmark_min(void)
7435 unsigned long lowmem_kbytes;
7436 int new_min_free_kbytes;
7438 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7439 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7441 if (new_min_free_kbytes > user_min_free_kbytes) {
7442 min_free_kbytes = new_min_free_kbytes;
7443 if (min_free_kbytes < 128)
7444 min_free_kbytes = 128;
7445 if (min_free_kbytes > 65536)
7446 min_free_kbytes = 65536;
7448 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7449 new_min_free_kbytes, user_min_free_kbytes);
7451 setup_per_zone_wmarks();
7452 refresh_zone_stat_thresholds();
7453 setup_per_zone_lowmem_reserve();
7456 setup_min_unmapped_ratio();
7457 setup_min_slab_ratio();
7460 khugepaged_min_free_kbytes_update();
7464 postcore_initcall(init_per_zone_wmark_min)
7467 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7468 * that we can call two helper functions whenever min_free_kbytes
7471 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7472 void __user *buffer, size_t *length, loff_t *ppos)
7476 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7481 user_min_free_kbytes = min_free_kbytes;
7482 setup_per_zone_wmarks();
7487 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7488 void __user *buffer, size_t *length, loff_t *ppos)
7492 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7497 setup_per_zone_wmarks();
7503 static void setup_min_unmapped_ratio(void)
7508 for_each_online_pgdat(pgdat)
7509 pgdat->min_unmapped_pages = 0;
7512 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7513 sysctl_min_unmapped_ratio) / 100;
7517 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7518 void __user *buffer, size_t *length, loff_t *ppos)
7522 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7526 setup_min_unmapped_ratio();
7531 static void setup_min_slab_ratio(void)
7536 for_each_online_pgdat(pgdat)
7537 pgdat->min_slab_pages = 0;
7540 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7541 sysctl_min_slab_ratio) / 100;
7544 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7545 void __user *buffer, size_t *length, loff_t *ppos)
7549 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7553 setup_min_slab_ratio();
7560 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7561 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7562 * whenever sysctl_lowmem_reserve_ratio changes.
7564 * The reserve ratio obviously has absolutely no relation with the
7565 * minimum watermarks. The lowmem reserve ratio can only make sense
7566 * if in function of the boot time zone sizes.
7568 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7569 void __user *buffer, size_t *length, loff_t *ppos)
7571 proc_dointvec_minmax(table, write, buffer, length, ppos);
7572 setup_per_zone_lowmem_reserve();
7577 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7578 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7579 * pagelist can have before it gets flushed back to buddy allocator.
7581 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7582 void __user *buffer, size_t *length, loff_t *ppos)
7585 int old_percpu_pagelist_fraction;
7588 mutex_lock(&pcp_batch_high_lock);
7589 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7591 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7592 if (!write || ret < 0)
7595 /* Sanity checking to avoid pcp imbalance */
7596 if (percpu_pagelist_fraction &&
7597 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7598 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7604 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7607 for_each_populated_zone(zone) {
7610 for_each_possible_cpu(cpu)
7611 pageset_set_high_and_batch(zone,
7612 per_cpu_ptr(zone->pageset, cpu));
7615 mutex_unlock(&pcp_batch_high_lock);
7620 int hashdist = HASHDIST_DEFAULT;
7622 static int __init set_hashdist(char *str)
7626 hashdist = simple_strtoul(str, &str, 0);
7629 __setup("hashdist=", set_hashdist);
7632 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7634 * Returns the number of pages that arch has reserved but
7635 * is not known to alloc_large_system_hash().
7637 static unsigned long __init arch_reserved_kernel_pages(void)
7644 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7645 * machines. As memory size is increased the scale is also increased but at
7646 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7647 * quadruples the scale is increased by one, which means the size of hash table
7648 * only doubles, instead of quadrupling as well.
7649 * Because 32-bit systems cannot have large physical memory, where this scaling
7650 * makes sense, it is disabled on such platforms.
7652 #if __BITS_PER_LONG > 32
7653 #define ADAPT_SCALE_BASE (64ul << 30)
7654 #define ADAPT_SCALE_SHIFT 2
7655 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7659 * allocate a large system hash table from bootmem
7660 * - it is assumed that the hash table must contain an exact power-of-2
7661 * quantity of entries
7662 * - limit is the number of hash buckets, not the total allocation size
7664 void *__init alloc_large_system_hash(const char *tablename,
7665 unsigned long bucketsize,
7666 unsigned long numentries,
7669 unsigned int *_hash_shift,
7670 unsigned int *_hash_mask,
7671 unsigned long low_limit,
7672 unsigned long high_limit)
7674 unsigned long long max = high_limit;
7675 unsigned long log2qty, size;
7679 /* allow the kernel cmdline to have a say */
7681 /* round applicable memory size up to nearest megabyte */
7682 numentries = nr_kernel_pages;
7683 numentries -= arch_reserved_kernel_pages();
7685 /* It isn't necessary when PAGE_SIZE >= 1MB */
7686 if (PAGE_SHIFT < 20)
7687 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7689 #if __BITS_PER_LONG > 32
7691 unsigned long adapt;
7693 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7694 adapt <<= ADAPT_SCALE_SHIFT)
7699 /* limit to 1 bucket per 2^scale bytes of low memory */
7700 if (scale > PAGE_SHIFT)
7701 numentries >>= (scale - PAGE_SHIFT);
7703 numentries <<= (PAGE_SHIFT - scale);
7705 /* Make sure we've got at least a 0-order allocation.. */
7706 if (unlikely(flags & HASH_SMALL)) {
7707 /* Makes no sense without HASH_EARLY */
7708 WARN_ON(!(flags & HASH_EARLY));
7709 if (!(numentries >> *_hash_shift)) {
7710 numentries = 1UL << *_hash_shift;
7711 BUG_ON(!numentries);
7713 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7714 numentries = PAGE_SIZE / bucketsize;
7716 numentries = roundup_pow_of_two(numentries);
7718 /* limit allocation size to 1/16 total memory by default */
7720 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7721 do_div(max, bucketsize);
7723 max = min(max, 0x80000000ULL);
7725 if (numentries < low_limit)
7726 numentries = low_limit;
7727 if (numentries > max)
7730 log2qty = ilog2(numentries);
7732 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7734 size = bucketsize << log2qty;
7735 if (flags & HASH_EARLY) {
7736 if (flags & HASH_ZERO)
7737 table = memblock_virt_alloc_nopanic(size, 0);
7739 table = memblock_virt_alloc_raw(size, 0);
7740 } else if (hashdist) {
7741 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7744 * If bucketsize is not a power-of-two, we may free
7745 * some pages at the end of hash table which
7746 * alloc_pages_exact() automatically does
7748 if (get_order(size) < MAX_ORDER) {
7749 table = alloc_pages_exact(size, gfp_flags);
7750 kmemleak_alloc(table, size, 1, gfp_flags);
7753 } while (!table && size > PAGE_SIZE && --log2qty);
7756 panic("Failed to allocate %s hash table\n", tablename);
7758 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7759 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7762 *_hash_shift = log2qty;
7764 *_hash_mask = (1 << log2qty) - 1;
7770 * This function checks whether pageblock includes unmovable pages or not.
7771 * If @count is not zero, it is okay to include less @count unmovable pages
7773 * PageLRU check without isolation or lru_lock could race so that
7774 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7775 * check without lock_page also may miss some movable non-lru pages at
7776 * race condition. So you can't expect this function should be exact.
7778 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7780 bool skip_hwpoisoned_pages)
7782 unsigned long pfn, iter, found;
7785 * TODO we could make this much more efficient by not checking every
7786 * page in the range if we know all of them are in MOVABLE_ZONE and
7787 * that the movable zone guarantees that pages are migratable but
7788 * the later is not the case right now unfortunatelly. E.g. movablecore
7789 * can still lead to having bootmem allocations in zone_movable.
7793 * CMA allocations (alloc_contig_range) really need to mark isolate
7794 * CMA pageblocks even when they are not movable in fact so consider
7795 * them movable here.
7797 if (is_migrate_cma(migratetype) &&
7798 is_migrate_cma(get_pageblock_migratetype(page)))
7801 pfn = page_to_pfn(page);
7802 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7803 unsigned long check = pfn + iter;
7805 if (!pfn_valid_within(check))
7808 page = pfn_to_page(check);
7810 if (PageReserved(page))
7814 * If the zone is movable and we have ruled out all reserved
7815 * pages then it should be reasonably safe to assume the rest
7818 if (zone_idx(zone) == ZONE_MOVABLE)
7822 * Hugepages are not in LRU lists, but they're movable.
7823 * We need not scan over tail pages bacause we don't
7824 * handle each tail page individually in migration.
7826 if (PageHuge(page)) {
7827 struct page *head = compound_head(page);
7828 unsigned int skip_pages;
7830 if (!hugepage_migration_supported(page_hstate(head)))
7833 skip_pages = (1 << compound_order(head)) - (page - head);
7834 iter += skip_pages - 1;
7839 * We can't use page_count without pin a page
7840 * because another CPU can free compound page.
7841 * This check already skips compound tails of THP
7842 * because their page->_refcount is zero at all time.
7844 if (!page_ref_count(page)) {
7845 if (PageBuddy(page))
7846 iter += (1 << page_order(page)) - 1;
7851 * The HWPoisoned page may be not in buddy system, and
7852 * page_count() is not 0.
7854 if (skip_hwpoisoned_pages && PageHWPoison(page))
7857 if (__PageMovable(page))
7863 * If there are RECLAIMABLE pages, we need to check
7864 * it. But now, memory offline itself doesn't call
7865 * shrink_node_slabs() and it still to be fixed.
7868 * If the page is not RAM, page_count()should be 0.
7869 * we don't need more check. This is an _used_ not-movable page.
7871 * The problematic thing here is PG_reserved pages. PG_reserved
7872 * is set to both of a memory hole page and a _used_ kernel
7880 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
7884 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7886 static unsigned long pfn_max_align_down(unsigned long pfn)
7888 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7889 pageblock_nr_pages) - 1);
7892 static unsigned long pfn_max_align_up(unsigned long pfn)
7894 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7895 pageblock_nr_pages));
7898 /* [start, end) must belong to a single zone. */
7899 static int __alloc_contig_migrate_range(struct compact_control *cc,
7900 unsigned long start, unsigned long end)
7902 /* This function is based on compact_zone() from compaction.c. */
7903 unsigned long nr_reclaimed;
7904 unsigned long pfn = start;
7905 unsigned int tries = 0;
7910 while (pfn < end || !list_empty(&cc->migratepages)) {
7911 if (fatal_signal_pending(current)) {
7916 if (list_empty(&cc->migratepages)) {
7917 cc->nr_migratepages = 0;
7918 pfn = isolate_migratepages_range(cc, pfn, end);
7924 } else if (++tries == 5) {
7925 ret = ret < 0 ? ret : -EBUSY;
7929 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7931 cc->nr_migratepages -= nr_reclaimed;
7933 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7934 NULL, 0, cc->mode, MR_CONTIG_RANGE);
7937 putback_movable_pages(&cc->migratepages);
7944 * alloc_contig_range() -- tries to allocate given range of pages
7945 * @start: start PFN to allocate
7946 * @end: one-past-the-last PFN to allocate
7947 * @migratetype: migratetype of the underlaying pageblocks (either
7948 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7949 * in range must have the same migratetype and it must
7950 * be either of the two.
7951 * @gfp_mask: GFP mask to use during compaction
7953 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7954 * aligned. The PFN range must belong to a single zone.
7956 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
7957 * pageblocks in the range. Once isolated, the pageblocks should not
7958 * be modified by others.
7960 * Returns zero on success or negative error code. On success all
7961 * pages which PFN is in [start, end) are allocated for the caller and
7962 * need to be freed with free_contig_range().
7964 int alloc_contig_range(unsigned long start, unsigned long end,
7965 unsigned migratetype, gfp_t gfp_mask)
7967 unsigned long outer_start, outer_end;
7971 struct compact_control cc = {
7972 .nr_migratepages = 0,
7974 .zone = page_zone(pfn_to_page(start)),
7975 .mode = MIGRATE_SYNC,
7976 .ignore_skip_hint = true,
7977 .no_set_skip_hint = true,
7978 .gfp_mask = current_gfp_context(gfp_mask),
7980 INIT_LIST_HEAD(&cc.migratepages);
7983 * What we do here is we mark all pageblocks in range as
7984 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7985 * have different sizes, and due to the way page allocator
7986 * work, we align the range to biggest of the two pages so
7987 * that page allocator won't try to merge buddies from
7988 * different pageblocks and change MIGRATE_ISOLATE to some
7989 * other migration type.
7991 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7992 * migrate the pages from an unaligned range (ie. pages that
7993 * we are interested in). This will put all the pages in
7994 * range back to page allocator as MIGRATE_ISOLATE.
7996 * When this is done, we take the pages in range from page
7997 * allocator removing them from the buddy system. This way
7998 * page allocator will never consider using them.
8000 * This lets us mark the pageblocks back as
8001 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8002 * aligned range but not in the unaligned, original range are
8003 * put back to page allocator so that buddy can use them.
8006 ret = start_isolate_page_range(pfn_max_align_down(start),
8007 pfn_max_align_up(end), migratetype,
8013 * In case of -EBUSY, we'd like to know which page causes problem.
8014 * So, just fall through. test_pages_isolated() has a tracepoint
8015 * which will report the busy page.
8017 * It is possible that busy pages could become available before
8018 * the call to test_pages_isolated, and the range will actually be
8019 * allocated. So, if we fall through be sure to clear ret so that
8020 * -EBUSY is not accidentally used or returned to caller.
8022 ret = __alloc_contig_migrate_range(&cc, start, end);
8023 if (ret && ret != -EBUSY)
8028 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8029 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8030 * more, all pages in [start, end) are free in page allocator.
8031 * What we are going to do is to allocate all pages from
8032 * [start, end) (that is remove them from page allocator).
8034 * The only problem is that pages at the beginning and at the
8035 * end of interesting range may be not aligned with pages that
8036 * page allocator holds, ie. they can be part of higher order
8037 * pages. Because of this, we reserve the bigger range and
8038 * once this is done free the pages we are not interested in.
8040 * We don't have to hold zone->lock here because the pages are
8041 * isolated thus they won't get removed from buddy.
8044 lru_add_drain_all();
8045 drain_all_pages(cc.zone);
8048 outer_start = start;
8049 while (!PageBuddy(pfn_to_page(outer_start))) {
8050 if (++order >= MAX_ORDER) {
8051 outer_start = start;
8054 outer_start &= ~0UL << order;
8057 if (outer_start != start) {
8058 order = page_order(pfn_to_page(outer_start));
8061 * outer_start page could be small order buddy page and
8062 * it doesn't include start page. Adjust outer_start
8063 * in this case to report failed page properly
8064 * on tracepoint in test_pages_isolated()
8066 if (outer_start + (1UL << order) <= start)
8067 outer_start = start;
8070 /* Make sure the range is really isolated. */
8071 if (test_pages_isolated(outer_start, end, false)) {
8072 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8073 __func__, outer_start, end);
8078 /* Grab isolated pages from freelists. */
8079 outer_end = isolate_freepages_range(&cc, outer_start, end);
8085 /* Free head and tail (if any) */
8086 if (start != outer_start)
8087 free_contig_range(outer_start, start - outer_start);
8088 if (end != outer_end)
8089 free_contig_range(end, outer_end - end);
8092 undo_isolate_page_range(pfn_max_align_down(start),
8093 pfn_max_align_up(end), migratetype);
8097 void free_contig_range(unsigned long pfn, unsigned nr_pages)
8099 unsigned int count = 0;
8101 for (; nr_pages--; pfn++) {
8102 struct page *page = pfn_to_page(pfn);
8104 count += page_count(page) != 1;
8107 WARN(count != 0, "%d pages are still in use!\n", count);
8112 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8113 * page high values need to be recalulated.
8115 void __meminit zone_pcp_update(struct zone *zone)
8118 mutex_lock(&pcp_batch_high_lock);
8119 for_each_possible_cpu(cpu)
8120 pageset_set_high_and_batch(zone,
8121 per_cpu_ptr(zone->pageset, cpu));
8122 mutex_unlock(&pcp_batch_high_lock);
8125 void zone_pcp_reset(struct zone *zone)
8127 unsigned long flags;
8129 struct per_cpu_pageset *pset;
8131 /* avoid races with drain_pages() */
8132 local_irq_save(flags);
8133 if (zone->pageset != &boot_pageset) {
8134 for_each_online_cpu(cpu) {
8135 pset = per_cpu_ptr(zone->pageset, cpu);
8136 drain_zonestat(zone, pset);
8138 free_percpu(zone->pageset);
8139 zone->pageset = &boot_pageset;
8141 local_irq_restore(flags);
8144 #ifdef CONFIG_MEMORY_HOTREMOVE
8146 * All pages in the range must be in a single zone and isolated
8147 * before calling this.
8150 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8154 unsigned int order, i;
8156 unsigned long flags;
8157 /* find the first valid pfn */
8158 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8163 offline_mem_sections(pfn, end_pfn);
8164 zone = page_zone(pfn_to_page(pfn));
8165 spin_lock_irqsave(&zone->lock, flags);
8167 while (pfn < end_pfn) {
8168 if (!pfn_valid(pfn)) {
8172 page = pfn_to_page(pfn);
8174 * The HWPoisoned page may be not in buddy system, and
8175 * page_count() is not 0.
8177 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8179 SetPageReserved(page);
8183 BUG_ON(page_count(page));
8184 BUG_ON(!PageBuddy(page));
8185 order = page_order(page);
8186 #ifdef CONFIG_DEBUG_VM
8187 pr_info("remove from free list %lx %d %lx\n",
8188 pfn, 1 << order, end_pfn);
8190 list_del(&page->lru);
8191 rmv_page_order(page);
8192 zone->free_area[order].nr_free--;
8193 for (i = 0; i < (1 << order); i++)
8194 SetPageReserved((page+i));
8195 pfn += (1 << order);
8197 spin_unlock_irqrestore(&zone->lock, flags);
8201 bool is_free_buddy_page(struct page *page)
8203 struct zone *zone = page_zone(page);
8204 unsigned long pfn = page_to_pfn(page);
8205 unsigned long flags;
8208 spin_lock_irqsave(&zone->lock, flags);
8209 for (order = 0; order < MAX_ORDER; order++) {
8210 struct page *page_head = page - (pfn & ((1 << order) - 1));
8212 if (PageBuddy(page_head) && page_order(page_head) >= order)
8215 spin_unlock_irqrestore(&zone->lock, flags);
8217 return order < MAX_ORDER;
8220 #ifdef CONFIG_MEMORY_FAILURE
8222 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8223 * test is performed under the zone lock to prevent a race against page
8226 bool set_hwpoison_free_buddy_page(struct page *page)
8228 struct zone *zone = page_zone(page);
8229 unsigned long pfn = page_to_pfn(page);
8230 unsigned long flags;
8232 bool hwpoisoned = false;
8234 spin_lock_irqsave(&zone->lock, flags);
8235 for (order = 0; order < MAX_ORDER; order++) {
8236 struct page *page_head = page - (pfn & ((1 << order) - 1));
8238 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8239 if (!TestSetPageHWPoison(page))
8244 spin_unlock_irqrestore(&zone->lock, flags);