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/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
70 #include <asm/sections.h>
71 #include <asm/tlbflush.h>
72 #include <asm/div64.h>
75 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
76 static DEFINE_MUTEX(pcp_batch_high_lock);
77 #define MIN_PERCPU_PAGELIST_FRACTION (8)
79 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
80 DEFINE_PER_CPU(int, numa_node);
81 EXPORT_PER_CPU_SYMBOL(numa_node);
84 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
86 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
87 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
88 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
89 * defined in <linux/topology.h>.
91 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
92 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
93 int _node_numa_mem_[MAX_NUMNODES];
96 /* work_structs for global per-cpu drains */
97 DEFINE_MUTEX(pcpu_drain_mutex);
98 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
100 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
101 volatile unsigned long latent_entropy __latent_entropy;
102 EXPORT_SYMBOL(latent_entropy);
106 * Array of node states.
108 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
109 [N_POSSIBLE] = NODE_MASK_ALL,
110 [N_ONLINE] = { { [0] = 1UL } },
112 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
113 #ifdef CONFIG_HIGHMEM
114 [N_HIGH_MEMORY] = { { [0] = 1UL } },
116 [N_MEMORY] = { { [0] = 1UL } },
117 [N_CPU] = { { [0] = 1UL } },
120 EXPORT_SYMBOL(node_states);
122 /* Protect totalram_pages and zone->managed_pages */
123 static DEFINE_SPINLOCK(managed_page_count_lock);
125 unsigned long totalram_pages __read_mostly;
126 unsigned long totalreserve_pages __read_mostly;
127 unsigned long totalcma_pages __read_mostly;
129 int percpu_pagelist_fraction;
130 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
133 * A cached value of the page's pageblock's migratetype, used when the page is
134 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
135 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
136 * Also the migratetype set in the page does not necessarily match the pcplist
137 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
138 * other index - this ensures that it will be put on the correct CMA freelist.
140 static inline int get_pcppage_migratetype(struct page *page)
145 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
147 page->index = migratetype;
150 #ifdef CONFIG_PM_SLEEP
152 * The following functions are used by the suspend/hibernate code to temporarily
153 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
154 * while devices are suspended. To avoid races with the suspend/hibernate code,
155 * they should always be called with pm_mutex held (gfp_allowed_mask also should
156 * only be modified with pm_mutex held, unless the suspend/hibernate code is
157 * guaranteed not to run in parallel with that modification).
160 static gfp_t saved_gfp_mask;
162 void pm_restore_gfp_mask(void)
164 WARN_ON(!mutex_is_locked(&pm_mutex));
165 if (saved_gfp_mask) {
166 gfp_allowed_mask = saved_gfp_mask;
171 void pm_restrict_gfp_mask(void)
173 WARN_ON(!mutex_is_locked(&pm_mutex));
174 WARN_ON(saved_gfp_mask);
175 saved_gfp_mask = gfp_allowed_mask;
176 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
179 bool pm_suspended_storage(void)
181 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
185 #endif /* CONFIG_PM_SLEEP */
187 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
188 unsigned int pageblock_order __read_mostly;
191 static void __free_pages_ok(struct page *page, unsigned int order);
194 * results with 256, 32 in the lowmem_reserve sysctl:
195 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
196 * 1G machine -> (16M dma, 784M normal, 224M high)
197 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
198 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
199 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
201 * TBD: should special case ZONE_DMA32 machines here - in those we normally
202 * don't need any ZONE_NORMAL reservation
204 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
205 #ifdef CONFIG_ZONE_DMA
208 #ifdef CONFIG_ZONE_DMA32
211 #ifdef CONFIG_HIGHMEM
217 EXPORT_SYMBOL(totalram_pages);
219 static char * const zone_names[MAX_NR_ZONES] = {
220 #ifdef CONFIG_ZONE_DMA
223 #ifdef CONFIG_ZONE_DMA32
227 #ifdef CONFIG_HIGHMEM
231 #ifdef CONFIG_ZONE_DEVICE
236 char * const migratetype_names[MIGRATE_TYPES] = {
244 #ifdef CONFIG_MEMORY_ISOLATION
249 compound_page_dtor * const compound_page_dtors[] = {
252 #ifdef CONFIG_HUGETLB_PAGE
255 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
260 int min_free_kbytes = 1024;
261 int user_min_free_kbytes = -1;
262 int watermark_scale_factor = 10;
264 static unsigned long __meminitdata nr_kernel_pages;
265 static unsigned long __meminitdata nr_all_pages;
266 static unsigned long __meminitdata dma_reserve;
268 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
269 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
270 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
271 static unsigned long __initdata required_kernelcore;
272 static unsigned long __initdata required_movablecore;
273 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
274 static bool mirrored_kernelcore;
276 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
278 EXPORT_SYMBOL(movable_zone);
279 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
282 int nr_node_ids __read_mostly = MAX_NUMNODES;
283 int nr_online_nodes __read_mostly = 1;
284 EXPORT_SYMBOL(nr_node_ids);
285 EXPORT_SYMBOL(nr_online_nodes);
288 int page_group_by_mobility_disabled __read_mostly;
290 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
291 static inline void reset_deferred_meminit(pg_data_t *pgdat)
293 unsigned long max_initialise;
294 unsigned long reserved_lowmem;
297 * Initialise at least 2G of a node but also take into account that
298 * two large system hashes that can take up 1GB for 0.25TB/node.
300 max_initialise = max(2UL << (30 - PAGE_SHIFT),
301 (pgdat->node_spanned_pages >> 8));
304 * Compensate the all the memblock reservations (e.g. crash kernel)
305 * from the initial estimation to make sure we will initialize enough
308 reserved_lowmem = memblock_reserved_memory_within(pgdat->node_start_pfn,
309 pgdat->node_start_pfn + max_initialise);
310 max_initialise += reserved_lowmem;
312 pgdat->static_init_size = min(max_initialise, pgdat->node_spanned_pages);
313 pgdat->first_deferred_pfn = ULONG_MAX;
316 /* Returns true if the struct page for the pfn is uninitialised */
317 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
319 int nid = early_pfn_to_nid(pfn);
321 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
328 * Returns false when the remaining initialisation should be deferred until
329 * later in the boot cycle when it can be parallelised.
331 static inline bool update_defer_init(pg_data_t *pgdat,
332 unsigned long pfn, unsigned long zone_end,
333 unsigned long *nr_initialised)
335 /* Always populate low zones for address-contrained allocations */
336 if (zone_end < pgdat_end_pfn(pgdat))
339 if ((*nr_initialised > pgdat->static_init_size) &&
340 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
341 pgdat->first_deferred_pfn = pfn;
348 static inline void reset_deferred_meminit(pg_data_t *pgdat)
352 static inline bool early_page_uninitialised(unsigned long pfn)
357 static inline bool update_defer_init(pg_data_t *pgdat,
358 unsigned long pfn, unsigned long zone_end,
359 unsigned long *nr_initialised)
365 /* Return a pointer to the bitmap storing bits affecting a block of pages */
366 static inline unsigned long *get_pageblock_bitmap(struct page *page,
369 #ifdef CONFIG_SPARSEMEM
370 return __pfn_to_section(pfn)->pageblock_flags;
372 return page_zone(page)->pageblock_flags;
373 #endif /* CONFIG_SPARSEMEM */
376 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
378 #ifdef CONFIG_SPARSEMEM
379 pfn &= (PAGES_PER_SECTION-1);
380 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
382 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
383 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
384 #endif /* CONFIG_SPARSEMEM */
388 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
389 * @page: The page within the block of interest
390 * @pfn: The target page frame number
391 * @end_bitidx: The last bit of interest to retrieve
392 * @mask: mask of bits that the caller is interested in
394 * Return: pageblock_bits flags
396 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
398 unsigned long end_bitidx,
401 unsigned long *bitmap;
402 unsigned long bitidx, word_bitidx;
405 bitmap = get_pageblock_bitmap(page, pfn);
406 bitidx = pfn_to_bitidx(page, pfn);
407 word_bitidx = bitidx / BITS_PER_LONG;
408 bitidx &= (BITS_PER_LONG-1);
410 word = bitmap[word_bitidx];
411 bitidx += end_bitidx;
412 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
415 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
416 unsigned long end_bitidx,
419 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
422 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
424 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
428 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
429 * @page: The page within the block of interest
430 * @flags: The flags to set
431 * @pfn: The target page frame number
432 * @end_bitidx: The last bit of interest
433 * @mask: mask of bits that the caller is interested in
435 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
437 unsigned long end_bitidx,
440 unsigned long *bitmap;
441 unsigned long bitidx, word_bitidx;
442 unsigned long old_word, word;
444 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
446 bitmap = get_pageblock_bitmap(page, pfn);
447 bitidx = pfn_to_bitidx(page, pfn);
448 word_bitidx = bitidx / BITS_PER_LONG;
449 bitidx &= (BITS_PER_LONG-1);
451 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
453 bitidx += end_bitidx;
454 mask <<= (BITS_PER_LONG - bitidx - 1);
455 flags <<= (BITS_PER_LONG - bitidx - 1);
457 word = READ_ONCE(bitmap[word_bitidx]);
459 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
460 if (word == old_word)
466 void set_pageblock_migratetype(struct page *page, int migratetype)
468 if (unlikely(page_group_by_mobility_disabled &&
469 migratetype < MIGRATE_PCPTYPES))
470 migratetype = MIGRATE_UNMOVABLE;
472 set_pageblock_flags_group(page, (unsigned long)migratetype,
473 PB_migrate, PB_migrate_end);
476 #ifdef CONFIG_DEBUG_VM
477 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
481 unsigned long pfn = page_to_pfn(page);
482 unsigned long sp, start_pfn;
485 seq = zone_span_seqbegin(zone);
486 start_pfn = zone->zone_start_pfn;
487 sp = zone->spanned_pages;
488 if (!zone_spans_pfn(zone, pfn))
490 } while (zone_span_seqretry(zone, seq));
493 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
494 pfn, zone_to_nid(zone), zone->name,
495 start_pfn, start_pfn + sp);
500 static int page_is_consistent(struct zone *zone, struct page *page)
502 if (!pfn_valid_within(page_to_pfn(page)))
504 if (zone != page_zone(page))
510 * Temporary debugging check for pages not lying within a given zone.
512 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
514 if (page_outside_zone_boundaries(zone, page))
516 if (!page_is_consistent(zone, page))
522 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
528 static void bad_page(struct page *page, const char *reason,
529 unsigned long bad_flags)
531 static unsigned long resume;
532 static unsigned long nr_shown;
533 static unsigned long nr_unshown;
536 * Allow a burst of 60 reports, then keep quiet for that minute;
537 * or allow a steady drip of one report per second.
539 if (nr_shown == 60) {
540 if (time_before(jiffies, resume)) {
546 "BUG: Bad page state: %lu messages suppressed\n",
553 resume = jiffies + 60 * HZ;
555 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
556 current->comm, page_to_pfn(page));
557 __dump_page(page, reason);
558 bad_flags &= page->flags;
560 pr_alert("bad because of flags: %#lx(%pGp)\n",
561 bad_flags, &bad_flags);
562 dump_page_owner(page);
567 /* Leave bad fields for debug, except PageBuddy could make trouble */
568 page_mapcount_reset(page); /* remove PageBuddy */
569 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
573 * Higher-order pages are called "compound pages". They are structured thusly:
575 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
577 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
578 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
580 * The first tail page's ->compound_dtor holds the offset in array of compound
581 * page destructors. See compound_page_dtors.
583 * The first tail page's ->compound_order holds the order of allocation.
584 * This usage means that zero-order pages may not be compound.
587 void free_compound_page(struct page *page)
589 __free_pages_ok(page, compound_order(page));
592 void prep_compound_page(struct page *page, unsigned int order)
595 int nr_pages = 1 << order;
597 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
598 set_compound_order(page, order);
600 for (i = 1; i < nr_pages; i++) {
601 struct page *p = page + i;
602 set_page_count(p, 0);
603 p->mapping = TAIL_MAPPING;
604 set_compound_head(p, page);
606 atomic_set(compound_mapcount_ptr(page), -1);
609 #ifdef CONFIG_DEBUG_PAGEALLOC
610 unsigned int _debug_guardpage_minorder;
611 bool _debug_pagealloc_enabled __read_mostly
612 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
613 EXPORT_SYMBOL(_debug_pagealloc_enabled);
614 bool _debug_guardpage_enabled __read_mostly;
616 static int __init early_debug_pagealloc(char *buf)
620 return kstrtobool(buf, &_debug_pagealloc_enabled);
622 early_param("debug_pagealloc", early_debug_pagealloc);
624 static bool need_debug_guardpage(void)
626 /* If we don't use debug_pagealloc, we don't need guard page */
627 if (!debug_pagealloc_enabled())
630 if (!debug_guardpage_minorder())
636 static void init_debug_guardpage(void)
638 if (!debug_pagealloc_enabled())
641 if (!debug_guardpage_minorder())
644 _debug_guardpage_enabled = true;
647 struct page_ext_operations debug_guardpage_ops = {
648 .need = need_debug_guardpage,
649 .init = init_debug_guardpage,
652 static int __init debug_guardpage_minorder_setup(char *buf)
656 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
657 pr_err("Bad debug_guardpage_minorder value\n");
660 _debug_guardpage_minorder = res;
661 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
664 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
666 static inline bool set_page_guard(struct zone *zone, struct page *page,
667 unsigned int order, int migratetype)
669 struct page_ext *page_ext;
671 if (!debug_guardpage_enabled())
674 if (order >= debug_guardpage_minorder())
677 page_ext = lookup_page_ext(page);
678 if (unlikely(!page_ext))
681 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
683 INIT_LIST_HEAD(&page->lru);
684 set_page_private(page, order);
685 /* Guard pages are not available for any usage */
686 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
691 static inline void clear_page_guard(struct zone *zone, struct page *page,
692 unsigned int order, int migratetype)
694 struct page_ext *page_ext;
696 if (!debug_guardpage_enabled())
699 page_ext = lookup_page_ext(page);
700 if (unlikely(!page_ext))
703 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
705 set_page_private(page, 0);
706 if (!is_migrate_isolate(migratetype))
707 __mod_zone_freepage_state(zone, (1 << order), migratetype);
710 struct page_ext_operations debug_guardpage_ops;
711 static inline bool set_page_guard(struct zone *zone, struct page *page,
712 unsigned int order, int migratetype) { return false; }
713 static inline void clear_page_guard(struct zone *zone, struct page *page,
714 unsigned int order, int migratetype) {}
717 static inline void set_page_order(struct page *page, unsigned int order)
719 set_page_private(page, order);
720 __SetPageBuddy(page);
723 static inline void rmv_page_order(struct page *page)
725 __ClearPageBuddy(page);
726 set_page_private(page, 0);
730 * This function checks whether a page is free && is the buddy
731 * we can do coalesce a page and its buddy if
732 * (a) the buddy is not in a hole (check before calling!) &&
733 * (b) the buddy is in the buddy system &&
734 * (c) a page and its buddy have the same order &&
735 * (d) a page and its buddy are in the same zone.
737 * For recording whether a page is in the buddy system, we set ->_mapcount
738 * PAGE_BUDDY_MAPCOUNT_VALUE.
739 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
740 * serialized by zone->lock.
742 * For recording page's order, we use page_private(page).
744 static inline int page_is_buddy(struct page *page, struct page *buddy,
747 if (page_is_guard(buddy) && page_order(buddy) == order) {
748 if (page_zone_id(page) != page_zone_id(buddy))
751 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
756 if (PageBuddy(buddy) && page_order(buddy) == order) {
758 * zone check is done late to avoid uselessly
759 * calculating zone/node ids for pages that could
762 if (page_zone_id(page) != page_zone_id(buddy))
765 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
773 * Freeing function for a buddy system allocator.
775 * The concept of a buddy system is to maintain direct-mapped table
776 * (containing bit values) for memory blocks of various "orders".
777 * The bottom level table contains the map for the smallest allocatable
778 * units of memory (here, pages), and each level above it describes
779 * pairs of units from the levels below, hence, "buddies".
780 * At a high level, all that happens here is marking the table entry
781 * at the bottom level available, and propagating the changes upward
782 * as necessary, plus some accounting needed to play nicely with other
783 * parts of the VM system.
784 * At each level, we keep a list of pages, which are heads of continuous
785 * free pages of length of (1 << order) and marked with _mapcount
786 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
788 * So when we are allocating or freeing one, we can derive the state of the
789 * other. That is, if we allocate a small block, and both were
790 * free, the remainder of the region must be split into blocks.
791 * If a block is freed, and its buddy is also free, then this
792 * triggers coalescing into a block of larger size.
797 static inline void __free_one_page(struct page *page,
799 struct zone *zone, unsigned int order,
802 unsigned long combined_pfn;
803 unsigned long uninitialized_var(buddy_pfn);
805 unsigned int max_order;
807 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
809 VM_BUG_ON(!zone_is_initialized(zone));
810 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
812 VM_BUG_ON(migratetype == -1);
813 if (likely(!is_migrate_isolate(migratetype)))
814 __mod_zone_freepage_state(zone, 1 << order, migratetype);
816 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
817 VM_BUG_ON_PAGE(bad_range(zone, page), page);
820 while (order < max_order - 1) {
821 buddy_pfn = __find_buddy_pfn(pfn, order);
822 buddy = page + (buddy_pfn - pfn);
824 if (!pfn_valid_within(buddy_pfn))
826 if (!page_is_buddy(page, buddy, order))
829 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
830 * merge with it and move up one order.
832 if (page_is_guard(buddy)) {
833 clear_page_guard(zone, buddy, order, migratetype);
835 list_del(&buddy->lru);
836 zone->free_area[order].nr_free--;
837 rmv_page_order(buddy);
839 combined_pfn = buddy_pfn & pfn;
840 page = page + (combined_pfn - pfn);
844 if (max_order < MAX_ORDER) {
845 /* If we are here, it means order is >= pageblock_order.
846 * We want to prevent merge between freepages on isolate
847 * pageblock and normal pageblock. Without this, pageblock
848 * isolation could cause incorrect freepage or CMA accounting.
850 * We don't want to hit this code for the more frequent
853 if (unlikely(has_isolate_pageblock(zone))) {
856 buddy_pfn = __find_buddy_pfn(pfn, order);
857 buddy = page + (buddy_pfn - pfn);
858 buddy_mt = get_pageblock_migratetype(buddy);
860 if (migratetype != buddy_mt
861 && (is_migrate_isolate(migratetype) ||
862 is_migrate_isolate(buddy_mt)))
866 goto continue_merging;
870 set_page_order(page, order);
873 * If this is not the largest possible page, check if the buddy
874 * of the next-highest order is free. If it is, it's possible
875 * that pages are being freed that will coalesce soon. In case,
876 * that is happening, add the free page to the tail of the list
877 * so it's less likely to be used soon and more likely to be merged
878 * as a higher order page
880 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
881 struct page *higher_page, *higher_buddy;
882 combined_pfn = buddy_pfn & pfn;
883 higher_page = page + (combined_pfn - pfn);
884 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
885 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
886 if (pfn_valid_within(buddy_pfn) &&
887 page_is_buddy(higher_page, higher_buddy, order + 1)) {
888 list_add_tail(&page->lru,
889 &zone->free_area[order].free_list[migratetype]);
894 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
896 zone->free_area[order].nr_free++;
900 * A bad page could be due to a number of fields. Instead of multiple branches,
901 * try and check multiple fields with one check. The caller must do a detailed
902 * check if necessary.
904 static inline bool page_expected_state(struct page *page,
905 unsigned long check_flags)
907 if (unlikely(atomic_read(&page->_mapcount) != -1))
910 if (unlikely((unsigned long)page->mapping |
911 page_ref_count(page) |
913 (unsigned long)page->mem_cgroup |
915 (page->flags & check_flags)))
921 static void free_pages_check_bad(struct page *page)
923 const char *bad_reason;
924 unsigned long bad_flags;
929 if (unlikely(atomic_read(&page->_mapcount) != -1))
930 bad_reason = "nonzero mapcount";
931 if (unlikely(page->mapping != NULL))
932 bad_reason = "non-NULL mapping";
933 if (unlikely(page_ref_count(page) != 0))
934 bad_reason = "nonzero _refcount";
935 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
936 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
937 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
940 if (unlikely(page->mem_cgroup))
941 bad_reason = "page still charged to cgroup";
943 bad_page(page, bad_reason, bad_flags);
946 static inline int free_pages_check(struct page *page)
948 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
951 /* Something has gone sideways, find it */
952 free_pages_check_bad(page);
956 static int free_tail_pages_check(struct page *head_page, struct page *page)
961 * We rely page->lru.next never has bit 0 set, unless the page
962 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
964 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
966 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
970 switch (page - head_page) {
972 /* the first tail page: ->mapping is compound_mapcount() */
973 if (unlikely(compound_mapcount(page))) {
974 bad_page(page, "nonzero compound_mapcount", 0);
980 * the second tail page: ->mapping is
981 * page_deferred_list().next -- ignore value.
985 if (page->mapping != TAIL_MAPPING) {
986 bad_page(page, "corrupted mapping in tail page", 0);
991 if (unlikely(!PageTail(page))) {
992 bad_page(page, "PageTail not set", 0);
995 if (unlikely(compound_head(page) != head_page)) {
996 bad_page(page, "compound_head not consistent", 0);
1001 page->mapping = NULL;
1002 clear_compound_head(page);
1006 static __always_inline bool free_pages_prepare(struct page *page,
1007 unsigned int order, bool check_free)
1011 VM_BUG_ON_PAGE(PageTail(page), page);
1013 trace_mm_page_free(page, order);
1014 kmemcheck_free_shadow(page, order);
1017 * Check tail pages before head page information is cleared to
1018 * avoid checking PageCompound for order-0 pages.
1020 if (unlikely(order)) {
1021 bool compound = PageCompound(page);
1024 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1027 ClearPageDoubleMap(page);
1028 for (i = 1; i < (1 << order); i++) {
1030 bad += free_tail_pages_check(page, page + i);
1031 if (unlikely(free_pages_check(page + i))) {
1035 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1038 if (PageMappingFlags(page))
1039 page->mapping = NULL;
1040 if (memcg_kmem_enabled() && PageKmemcg(page))
1041 memcg_kmem_uncharge(page, order);
1043 bad += free_pages_check(page);
1047 page_cpupid_reset_last(page);
1048 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1049 reset_page_owner(page, order);
1051 if (!PageHighMem(page)) {
1052 debug_check_no_locks_freed(page_address(page),
1053 PAGE_SIZE << order);
1054 debug_check_no_obj_freed(page_address(page),
1055 PAGE_SIZE << order);
1057 arch_free_page(page, order);
1058 kernel_poison_pages(page, 1 << order, 0);
1059 kernel_map_pages(page, 1 << order, 0);
1060 kasan_free_pages(page, order);
1065 #ifdef CONFIG_DEBUG_VM
1066 static inline bool free_pcp_prepare(struct page *page)
1068 return free_pages_prepare(page, 0, true);
1071 static inline bool bulkfree_pcp_prepare(struct page *page)
1076 static bool free_pcp_prepare(struct page *page)
1078 return free_pages_prepare(page, 0, false);
1081 static bool bulkfree_pcp_prepare(struct page *page)
1083 return free_pages_check(page);
1085 #endif /* CONFIG_DEBUG_VM */
1088 * Frees a number of pages from the PCP lists
1089 * Assumes all pages on list are in same zone, and of same order.
1090 * count is the number of pages to free.
1092 * If the zone was previously in an "all pages pinned" state then look to
1093 * see if this freeing clears that state.
1095 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1096 * pinned" detection logic.
1098 static void free_pcppages_bulk(struct zone *zone, int count,
1099 struct per_cpu_pages *pcp)
1101 int migratetype = 0;
1103 bool isolated_pageblocks;
1105 spin_lock(&zone->lock);
1106 isolated_pageblocks = has_isolate_pageblock(zone);
1110 struct list_head *list;
1113 * Remove pages from lists in a round-robin fashion. A
1114 * batch_free count is maintained that is incremented when an
1115 * empty list is encountered. This is so more pages are freed
1116 * off fuller lists instead of spinning excessively around empty
1121 if (++migratetype == MIGRATE_PCPTYPES)
1123 list = &pcp->lists[migratetype];
1124 } while (list_empty(list));
1126 /* This is the only non-empty list. Free them all. */
1127 if (batch_free == MIGRATE_PCPTYPES)
1131 int mt; /* migratetype of the to-be-freed page */
1133 page = list_last_entry(list, struct page, lru);
1134 /* must delete as __free_one_page list manipulates */
1135 list_del(&page->lru);
1137 mt = get_pcppage_migratetype(page);
1138 /* MIGRATE_ISOLATE page should not go to pcplists */
1139 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1140 /* Pageblock could have been isolated meanwhile */
1141 if (unlikely(isolated_pageblocks))
1142 mt = get_pageblock_migratetype(page);
1144 if (bulkfree_pcp_prepare(page))
1147 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1148 trace_mm_page_pcpu_drain(page, 0, mt);
1149 } while (--count && --batch_free && !list_empty(list));
1151 spin_unlock(&zone->lock);
1154 static void free_one_page(struct zone *zone,
1155 struct page *page, unsigned long pfn,
1159 spin_lock(&zone->lock);
1160 if (unlikely(has_isolate_pageblock(zone) ||
1161 is_migrate_isolate(migratetype))) {
1162 migratetype = get_pfnblock_migratetype(page, pfn);
1164 __free_one_page(page, pfn, zone, order, migratetype);
1165 spin_unlock(&zone->lock);
1168 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1169 unsigned long zone, int nid)
1171 set_page_links(page, zone, nid, pfn);
1172 init_page_count(page);
1173 page_mapcount_reset(page);
1174 page_cpupid_reset_last(page);
1176 INIT_LIST_HEAD(&page->lru);
1177 #ifdef WANT_PAGE_VIRTUAL
1178 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1179 if (!is_highmem_idx(zone))
1180 set_page_address(page, __va(pfn << PAGE_SHIFT));
1184 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1187 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1190 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1191 static void init_reserved_page(unsigned long pfn)
1196 if (!early_page_uninitialised(pfn))
1199 nid = early_pfn_to_nid(pfn);
1200 pgdat = NODE_DATA(nid);
1202 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1203 struct zone *zone = &pgdat->node_zones[zid];
1205 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1208 __init_single_pfn(pfn, zid, nid);
1211 static inline void init_reserved_page(unsigned long pfn)
1214 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1217 * Initialised pages do not have PageReserved set. This function is
1218 * called for each range allocated by the bootmem allocator and
1219 * marks the pages PageReserved. The remaining valid pages are later
1220 * sent to the buddy page allocator.
1222 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1224 unsigned long start_pfn = PFN_DOWN(start);
1225 unsigned long end_pfn = PFN_UP(end);
1227 for (; start_pfn < end_pfn; start_pfn++) {
1228 if (pfn_valid(start_pfn)) {
1229 struct page *page = pfn_to_page(start_pfn);
1231 init_reserved_page(start_pfn);
1233 /* Avoid false-positive PageTail() */
1234 INIT_LIST_HEAD(&page->lru);
1236 SetPageReserved(page);
1241 static void __free_pages_ok(struct page *page, unsigned int order)
1243 unsigned long flags;
1245 unsigned long pfn = page_to_pfn(page);
1247 if (!free_pages_prepare(page, order, true))
1250 migratetype = get_pfnblock_migratetype(page, pfn);
1251 local_irq_save(flags);
1252 __count_vm_events(PGFREE, 1 << order);
1253 free_one_page(page_zone(page), page, pfn, order, migratetype);
1254 local_irq_restore(flags);
1257 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1259 unsigned int nr_pages = 1 << order;
1260 struct page *p = page;
1264 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1266 __ClearPageReserved(p);
1267 set_page_count(p, 0);
1269 __ClearPageReserved(p);
1270 set_page_count(p, 0);
1272 page_zone(page)->managed_pages += nr_pages;
1273 set_page_refcounted(page);
1274 __free_pages(page, order);
1277 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1278 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1280 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1282 int __meminit early_pfn_to_nid(unsigned long pfn)
1284 static DEFINE_SPINLOCK(early_pfn_lock);
1287 spin_lock(&early_pfn_lock);
1288 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1290 nid = first_online_node;
1291 spin_unlock(&early_pfn_lock);
1297 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1298 static inline bool __meminit __maybe_unused
1299 meminit_pfn_in_nid(unsigned long pfn, int node,
1300 struct mminit_pfnnid_cache *state)
1304 nid = __early_pfn_to_nid(pfn, state);
1305 if (nid >= 0 && nid != node)
1310 /* Only safe to use early in boot when initialisation is single-threaded */
1311 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1313 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1318 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1322 static inline bool __meminit __maybe_unused
1323 meminit_pfn_in_nid(unsigned long pfn, int node,
1324 struct mminit_pfnnid_cache *state)
1331 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1334 if (early_page_uninitialised(pfn))
1336 return __free_pages_boot_core(page, order);
1340 * Check that the whole (or subset of) a pageblock given by the interval of
1341 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1342 * with the migration of free compaction scanner. The scanners then need to
1343 * use only pfn_valid_within() check for arches that allow holes within
1346 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1348 * It's possible on some configurations to have a setup like node0 node1 node0
1349 * i.e. it's possible that all pages within a zones range of pages do not
1350 * belong to a single zone. We assume that a border between node0 and node1
1351 * can occur within a single pageblock, but not a node0 node1 node0
1352 * interleaving within a single pageblock. It is therefore sufficient to check
1353 * the first and last page of a pageblock and avoid checking each individual
1354 * page in a pageblock.
1356 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1357 unsigned long end_pfn, struct zone *zone)
1359 struct page *start_page;
1360 struct page *end_page;
1362 /* end_pfn is one past the range we are checking */
1365 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1368 start_page = pfn_to_online_page(start_pfn);
1372 if (page_zone(start_page) != zone)
1375 end_page = pfn_to_page(end_pfn);
1377 /* This gives a shorter code than deriving page_zone(end_page) */
1378 if (page_zone_id(start_page) != page_zone_id(end_page))
1384 void set_zone_contiguous(struct zone *zone)
1386 unsigned long block_start_pfn = zone->zone_start_pfn;
1387 unsigned long block_end_pfn;
1389 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1390 for (; block_start_pfn < zone_end_pfn(zone);
1391 block_start_pfn = block_end_pfn,
1392 block_end_pfn += pageblock_nr_pages) {
1394 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1396 if (!__pageblock_pfn_to_page(block_start_pfn,
1397 block_end_pfn, zone))
1401 /* We confirm that there is no hole */
1402 zone->contiguous = true;
1405 void clear_zone_contiguous(struct zone *zone)
1407 zone->contiguous = false;
1410 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1411 static void __init deferred_free_range(struct page *page,
1412 unsigned long pfn, int nr_pages)
1419 /* Free a large naturally-aligned chunk if possible */
1420 if (nr_pages == pageblock_nr_pages &&
1421 (pfn & (pageblock_nr_pages - 1)) == 0) {
1422 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1423 __free_pages_boot_core(page, pageblock_order);
1427 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1428 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1429 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1430 __free_pages_boot_core(page, 0);
1434 /* Completion tracking for deferred_init_memmap() threads */
1435 static atomic_t pgdat_init_n_undone __initdata;
1436 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1438 static inline void __init pgdat_init_report_one_done(void)
1440 if (atomic_dec_and_test(&pgdat_init_n_undone))
1441 complete(&pgdat_init_all_done_comp);
1444 /* Initialise remaining memory on a node */
1445 static int __init deferred_init_memmap(void *data)
1447 pg_data_t *pgdat = data;
1448 int nid = pgdat->node_id;
1449 struct mminit_pfnnid_cache nid_init_state = { };
1450 unsigned long start = jiffies;
1451 unsigned long nr_pages = 0;
1452 unsigned long walk_start, walk_end;
1455 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1456 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1458 if (first_init_pfn == ULONG_MAX) {
1459 pgdat_init_report_one_done();
1463 /* Bind memory initialisation thread to a local node if possible */
1464 if (!cpumask_empty(cpumask))
1465 set_cpus_allowed_ptr(current, cpumask);
1467 /* Sanity check boundaries */
1468 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1469 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1470 pgdat->first_deferred_pfn = ULONG_MAX;
1472 /* Only the highest zone is deferred so find it */
1473 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1474 zone = pgdat->node_zones + zid;
1475 if (first_init_pfn < zone_end_pfn(zone))
1479 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1480 unsigned long pfn, end_pfn;
1481 struct page *page = NULL;
1482 struct page *free_base_page = NULL;
1483 unsigned long free_base_pfn = 0;
1486 end_pfn = min(walk_end, zone_end_pfn(zone));
1487 pfn = first_init_pfn;
1488 if (pfn < walk_start)
1490 if (pfn < zone->zone_start_pfn)
1491 pfn = zone->zone_start_pfn;
1493 for (; pfn < end_pfn; pfn++) {
1494 if (!pfn_valid_within(pfn))
1498 * Ensure pfn_valid is checked every
1499 * pageblock_nr_pages for memory holes
1501 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1502 if (!pfn_valid(pfn)) {
1508 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1513 /* Minimise pfn page lookups and scheduler checks */
1514 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1517 nr_pages += nr_to_free;
1518 deferred_free_range(free_base_page,
1519 free_base_pfn, nr_to_free);
1520 free_base_page = NULL;
1521 free_base_pfn = nr_to_free = 0;
1523 page = pfn_to_page(pfn);
1528 VM_BUG_ON(page_zone(page) != zone);
1532 __init_single_page(page, pfn, zid, nid);
1533 if (!free_base_page) {
1534 free_base_page = page;
1535 free_base_pfn = pfn;
1540 /* Where possible, batch up pages for a single free */
1543 /* Free the current block of pages to allocator */
1544 nr_pages += nr_to_free;
1545 deferred_free_range(free_base_page, free_base_pfn,
1547 free_base_page = NULL;
1548 free_base_pfn = nr_to_free = 0;
1550 /* Free the last block of pages to allocator */
1551 nr_pages += nr_to_free;
1552 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1554 first_init_pfn = max(end_pfn, first_init_pfn);
1557 /* Sanity check that the next zone really is unpopulated */
1558 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1560 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1561 jiffies_to_msecs(jiffies - start));
1563 pgdat_init_report_one_done();
1566 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1568 void __init page_alloc_init_late(void)
1572 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1575 /* There will be num_node_state(N_MEMORY) threads */
1576 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1577 for_each_node_state(nid, N_MEMORY) {
1578 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1581 /* Block until all are initialised */
1582 wait_for_completion(&pgdat_init_all_done_comp);
1584 /* Reinit limits that are based on free pages after the kernel is up */
1585 files_maxfiles_init();
1587 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1588 /* Discard memblock private memory */
1592 for_each_populated_zone(zone)
1593 set_zone_contiguous(zone);
1597 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1598 void __init init_cma_reserved_pageblock(struct page *page)
1600 unsigned i = pageblock_nr_pages;
1601 struct page *p = page;
1604 __ClearPageReserved(p);
1605 set_page_count(p, 0);
1608 set_pageblock_migratetype(page, MIGRATE_CMA);
1610 if (pageblock_order >= MAX_ORDER) {
1611 i = pageblock_nr_pages;
1614 set_page_refcounted(p);
1615 __free_pages(p, MAX_ORDER - 1);
1616 p += MAX_ORDER_NR_PAGES;
1617 } while (i -= MAX_ORDER_NR_PAGES);
1619 set_page_refcounted(page);
1620 __free_pages(page, pageblock_order);
1623 adjust_managed_page_count(page, pageblock_nr_pages);
1628 * The order of subdivision here is critical for the IO subsystem.
1629 * Please do not alter this order without good reasons and regression
1630 * testing. Specifically, as large blocks of memory are subdivided,
1631 * the order in which smaller blocks are delivered depends on the order
1632 * they're subdivided in this function. This is the primary factor
1633 * influencing the order in which pages are delivered to the IO
1634 * subsystem according to empirical testing, and this is also justified
1635 * by considering the behavior of a buddy system containing a single
1636 * large block of memory acted on by a series of small allocations.
1637 * This behavior is a critical factor in sglist merging's success.
1641 static inline void expand(struct zone *zone, struct page *page,
1642 int low, int high, struct free_area *area,
1645 unsigned long size = 1 << high;
1647 while (high > low) {
1651 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1654 * Mark as guard pages (or page), that will allow to
1655 * merge back to allocator when buddy will be freed.
1656 * Corresponding page table entries will not be touched,
1657 * pages will stay not present in virtual address space
1659 if (set_page_guard(zone, &page[size], high, migratetype))
1662 list_add(&page[size].lru, &area->free_list[migratetype]);
1664 set_page_order(&page[size], high);
1668 static void check_new_page_bad(struct page *page)
1670 const char *bad_reason = NULL;
1671 unsigned long bad_flags = 0;
1673 if (unlikely(atomic_read(&page->_mapcount) != -1))
1674 bad_reason = "nonzero mapcount";
1675 if (unlikely(page->mapping != NULL))
1676 bad_reason = "non-NULL mapping";
1677 if (unlikely(page_ref_count(page) != 0))
1678 bad_reason = "nonzero _count";
1679 if (unlikely(page->flags & __PG_HWPOISON)) {
1680 bad_reason = "HWPoisoned (hardware-corrupted)";
1681 bad_flags = __PG_HWPOISON;
1682 /* Don't complain about hwpoisoned pages */
1683 page_mapcount_reset(page); /* remove PageBuddy */
1686 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1687 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1688 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1691 if (unlikely(page->mem_cgroup))
1692 bad_reason = "page still charged to cgroup";
1694 bad_page(page, bad_reason, bad_flags);
1698 * This page is about to be returned from the page allocator
1700 static inline int check_new_page(struct page *page)
1702 if (likely(page_expected_state(page,
1703 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1706 check_new_page_bad(page);
1710 static inline bool free_pages_prezeroed(void)
1712 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1713 page_poisoning_enabled();
1716 #ifdef CONFIG_DEBUG_VM
1717 static bool check_pcp_refill(struct page *page)
1722 static bool check_new_pcp(struct page *page)
1724 return check_new_page(page);
1727 static bool check_pcp_refill(struct page *page)
1729 return check_new_page(page);
1731 static bool check_new_pcp(struct page *page)
1735 #endif /* CONFIG_DEBUG_VM */
1737 static bool check_new_pages(struct page *page, unsigned int order)
1740 for (i = 0; i < (1 << order); i++) {
1741 struct page *p = page + i;
1743 if (unlikely(check_new_page(p)))
1750 inline void post_alloc_hook(struct page *page, unsigned int order,
1753 set_page_private(page, 0);
1754 set_page_refcounted(page);
1756 arch_alloc_page(page, order);
1757 kernel_map_pages(page, 1 << order, 1);
1758 kernel_poison_pages(page, 1 << order, 1);
1759 kasan_alloc_pages(page, order);
1760 set_page_owner(page, order, gfp_flags);
1763 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1764 unsigned int alloc_flags)
1768 post_alloc_hook(page, order, gfp_flags);
1770 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1771 for (i = 0; i < (1 << order); i++)
1772 clear_highpage(page + i);
1774 if (order && (gfp_flags & __GFP_COMP))
1775 prep_compound_page(page, order);
1778 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1779 * allocate the page. The expectation is that the caller is taking
1780 * steps that will free more memory. The caller should avoid the page
1781 * being used for !PFMEMALLOC purposes.
1783 if (alloc_flags & ALLOC_NO_WATERMARKS)
1784 set_page_pfmemalloc(page);
1786 clear_page_pfmemalloc(page);
1790 * Go through the free lists for the given migratetype and remove
1791 * the smallest available page from the freelists
1794 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1797 unsigned int current_order;
1798 struct free_area *area;
1801 /* Find a page of the appropriate size in the preferred list */
1802 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1803 area = &(zone->free_area[current_order]);
1804 page = list_first_entry_or_null(&area->free_list[migratetype],
1808 list_del(&page->lru);
1809 rmv_page_order(page);
1811 expand(zone, page, order, current_order, area, migratetype);
1812 set_pcppage_migratetype(page, migratetype);
1821 * This array describes the order lists are fallen back to when
1822 * the free lists for the desirable migrate type are depleted
1824 static int fallbacks[MIGRATE_TYPES][4] = {
1825 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1826 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1827 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1829 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1831 #ifdef CONFIG_MEMORY_ISOLATION
1832 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1837 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1840 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1843 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1844 unsigned int order) { return NULL; }
1848 * Move the free pages in a range to the free lists of the requested type.
1849 * Note that start_page and end_pages are not aligned on a pageblock
1850 * boundary. If alignment is required, use move_freepages_block()
1852 static int move_freepages(struct zone *zone,
1853 struct page *start_page, struct page *end_page,
1854 int migratetype, int *num_movable)
1858 int pages_moved = 0;
1860 #ifndef CONFIG_HOLES_IN_ZONE
1862 * page_zone is not safe to call in this context when
1863 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1864 * anyway as we check zone boundaries in move_freepages_block().
1865 * Remove at a later date when no bug reports exist related to
1866 * grouping pages by mobility
1868 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1874 for (page = start_page; page <= end_page;) {
1875 if (!pfn_valid_within(page_to_pfn(page))) {
1880 /* Make sure we are not inadvertently changing nodes */
1881 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1883 if (!PageBuddy(page)) {
1885 * We assume that pages that could be isolated for
1886 * migration are movable. But we don't actually try
1887 * isolating, as that would be expensive.
1890 (PageLRU(page) || __PageMovable(page)))
1897 order = page_order(page);
1898 list_move(&page->lru,
1899 &zone->free_area[order].free_list[migratetype]);
1901 pages_moved += 1 << order;
1907 int move_freepages_block(struct zone *zone, struct page *page,
1908 int migratetype, int *num_movable)
1910 unsigned long start_pfn, end_pfn;
1911 struct page *start_page, *end_page;
1913 start_pfn = page_to_pfn(page);
1914 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1915 start_page = pfn_to_page(start_pfn);
1916 end_page = start_page + pageblock_nr_pages - 1;
1917 end_pfn = start_pfn + pageblock_nr_pages - 1;
1919 /* Do not cross zone boundaries */
1920 if (!zone_spans_pfn(zone, start_pfn))
1922 if (!zone_spans_pfn(zone, end_pfn))
1925 return move_freepages(zone, start_page, end_page, migratetype,
1929 static void change_pageblock_range(struct page *pageblock_page,
1930 int start_order, int migratetype)
1932 int nr_pageblocks = 1 << (start_order - pageblock_order);
1934 while (nr_pageblocks--) {
1935 set_pageblock_migratetype(pageblock_page, migratetype);
1936 pageblock_page += pageblock_nr_pages;
1941 * When we are falling back to another migratetype during allocation, try to
1942 * steal extra free pages from the same pageblocks to satisfy further
1943 * allocations, instead of polluting multiple pageblocks.
1945 * If we are stealing a relatively large buddy page, it is likely there will
1946 * be more free pages in the pageblock, so try to steal them all. For
1947 * reclaimable and unmovable allocations, we steal regardless of page size,
1948 * as fragmentation caused by those allocations polluting movable pageblocks
1949 * is worse than movable allocations stealing from unmovable and reclaimable
1952 static bool can_steal_fallback(unsigned int order, int start_mt)
1955 * Leaving this order check is intended, although there is
1956 * relaxed order check in next check. The reason is that
1957 * we can actually steal whole pageblock if this condition met,
1958 * but, below check doesn't guarantee it and that is just heuristic
1959 * so could be changed anytime.
1961 if (order >= pageblock_order)
1964 if (order >= pageblock_order / 2 ||
1965 start_mt == MIGRATE_RECLAIMABLE ||
1966 start_mt == MIGRATE_UNMOVABLE ||
1967 page_group_by_mobility_disabled)
1974 * This function implements actual steal behaviour. If order is large enough,
1975 * we can steal whole pageblock. If not, we first move freepages in this
1976 * pageblock to our migratetype and determine how many already-allocated pages
1977 * are there in the pageblock with a compatible migratetype. If at least half
1978 * of pages are free or compatible, we can change migratetype of the pageblock
1979 * itself, so pages freed in the future will be put on the correct free list.
1981 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1982 int start_type, bool whole_block)
1984 unsigned int current_order = page_order(page);
1985 struct free_area *area;
1986 int free_pages, movable_pages, alike_pages;
1989 old_block_type = get_pageblock_migratetype(page);
1992 * This can happen due to races and we want to prevent broken
1993 * highatomic accounting.
1995 if (is_migrate_highatomic(old_block_type))
1998 /* Take ownership for orders >= pageblock_order */
1999 if (current_order >= pageblock_order) {
2000 change_pageblock_range(page, current_order, start_type);
2004 /* We are not allowed to try stealing from the whole block */
2008 free_pages = move_freepages_block(zone, page, start_type,
2011 * Determine how many pages are compatible with our allocation.
2012 * For movable allocation, it's the number of movable pages which
2013 * we just obtained. For other types it's a bit more tricky.
2015 if (start_type == MIGRATE_MOVABLE) {
2016 alike_pages = movable_pages;
2019 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2020 * to MOVABLE pageblock, consider all non-movable pages as
2021 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2022 * vice versa, be conservative since we can't distinguish the
2023 * exact migratetype of non-movable pages.
2025 if (old_block_type == MIGRATE_MOVABLE)
2026 alike_pages = pageblock_nr_pages
2027 - (free_pages + movable_pages);
2032 /* moving whole block can fail due to zone boundary conditions */
2037 * If a sufficient number of pages in the block are either free or of
2038 * comparable migratability as our allocation, claim the whole block.
2040 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2041 page_group_by_mobility_disabled)
2042 set_pageblock_migratetype(page, start_type);
2047 area = &zone->free_area[current_order];
2048 list_move(&page->lru, &area->free_list[start_type]);
2052 * Check whether there is a suitable fallback freepage with requested order.
2053 * If only_stealable is true, this function returns fallback_mt only if
2054 * we can steal other freepages all together. This would help to reduce
2055 * fragmentation due to mixed migratetype pages in one pageblock.
2057 int find_suitable_fallback(struct free_area *area, unsigned int order,
2058 int migratetype, bool only_stealable, bool *can_steal)
2063 if (area->nr_free == 0)
2068 fallback_mt = fallbacks[migratetype][i];
2069 if (fallback_mt == MIGRATE_TYPES)
2072 if (list_empty(&area->free_list[fallback_mt]))
2075 if (can_steal_fallback(order, migratetype))
2078 if (!only_stealable)
2089 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2090 * there are no empty page blocks that contain a page with a suitable order
2092 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2093 unsigned int alloc_order)
2096 unsigned long max_managed, flags;
2099 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2100 * Check is race-prone but harmless.
2102 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2103 if (zone->nr_reserved_highatomic >= max_managed)
2106 spin_lock_irqsave(&zone->lock, flags);
2108 /* Recheck the nr_reserved_highatomic limit under the lock */
2109 if (zone->nr_reserved_highatomic >= max_managed)
2113 mt = get_pageblock_migratetype(page);
2114 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2115 && !is_migrate_cma(mt)) {
2116 zone->nr_reserved_highatomic += pageblock_nr_pages;
2117 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2118 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2122 spin_unlock_irqrestore(&zone->lock, flags);
2126 * Used when an allocation is about to fail under memory pressure. This
2127 * potentially hurts the reliability of high-order allocations when under
2128 * intense memory pressure but failed atomic allocations should be easier
2129 * to recover from than an OOM.
2131 * If @force is true, try to unreserve a pageblock even though highatomic
2132 * pageblock is exhausted.
2134 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2137 struct zonelist *zonelist = ac->zonelist;
2138 unsigned long flags;
2145 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2148 * Preserve at least one pageblock unless memory pressure
2151 if (!force && zone->nr_reserved_highatomic <=
2155 spin_lock_irqsave(&zone->lock, flags);
2156 for (order = 0; order < MAX_ORDER; order++) {
2157 struct free_area *area = &(zone->free_area[order]);
2159 page = list_first_entry_or_null(
2160 &area->free_list[MIGRATE_HIGHATOMIC],
2166 * In page freeing path, migratetype change is racy so
2167 * we can counter several free pages in a pageblock
2168 * in this loop althoug we changed the pageblock type
2169 * from highatomic to ac->migratetype. So we should
2170 * adjust the count once.
2172 if (is_migrate_highatomic_page(page)) {
2174 * It should never happen but changes to
2175 * locking could inadvertently allow a per-cpu
2176 * drain to add pages to MIGRATE_HIGHATOMIC
2177 * while unreserving so be safe and watch for
2180 zone->nr_reserved_highatomic -= min(
2182 zone->nr_reserved_highatomic);
2186 * Convert to ac->migratetype and avoid the normal
2187 * pageblock stealing heuristics. Minimally, the caller
2188 * is doing the work and needs the pages. More
2189 * importantly, if the block was always converted to
2190 * MIGRATE_UNMOVABLE or another type then the number
2191 * of pageblocks that cannot be completely freed
2194 set_pageblock_migratetype(page, ac->migratetype);
2195 ret = move_freepages_block(zone, page, ac->migratetype,
2198 spin_unlock_irqrestore(&zone->lock, flags);
2202 spin_unlock_irqrestore(&zone->lock, flags);
2209 * Try finding a free buddy page on the fallback list and put it on the free
2210 * list of requested migratetype, possibly along with other pages from the same
2211 * block, depending on fragmentation avoidance heuristics. Returns true if
2212 * fallback was found so that __rmqueue_smallest() can grab it.
2214 * The use of signed ints for order and current_order is a deliberate
2215 * deviation from the rest of this file, to make the for loop
2216 * condition simpler.
2219 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2221 struct free_area *area;
2228 * Find the largest available free page in the other list. This roughly
2229 * approximates finding the pageblock with the most free pages, which
2230 * would be too costly to do exactly.
2232 for (current_order = MAX_ORDER - 1; current_order >= order;
2234 area = &(zone->free_area[current_order]);
2235 fallback_mt = find_suitable_fallback(area, current_order,
2236 start_migratetype, false, &can_steal);
2237 if (fallback_mt == -1)
2241 * We cannot steal all free pages from the pageblock and the
2242 * requested migratetype is movable. In that case it's better to
2243 * steal and split the smallest available page instead of the
2244 * largest available page, because even if the next movable
2245 * allocation falls back into a different pageblock than this
2246 * one, it won't cause permanent fragmentation.
2248 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2249 && current_order > order)
2258 for (current_order = order; current_order < MAX_ORDER;
2260 area = &(zone->free_area[current_order]);
2261 fallback_mt = find_suitable_fallback(area, current_order,
2262 start_migratetype, false, &can_steal);
2263 if (fallback_mt != -1)
2268 * This should not happen - we already found a suitable fallback
2269 * when looking for the largest page.
2271 VM_BUG_ON(current_order == MAX_ORDER);
2274 page = list_first_entry(&area->free_list[fallback_mt],
2277 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2279 trace_mm_page_alloc_extfrag(page, order, current_order,
2280 start_migratetype, fallback_mt);
2287 * Do the hard work of removing an element from the buddy allocator.
2288 * Call me with the zone->lock already held.
2290 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2296 page = __rmqueue_smallest(zone, order, migratetype);
2297 if (unlikely(!page)) {
2298 if (migratetype == MIGRATE_MOVABLE)
2299 page = __rmqueue_cma_fallback(zone, order);
2301 if (!page && __rmqueue_fallback(zone, order, migratetype))
2305 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2310 * Obtain a specified number of elements from the buddy allocator, all under
2311 * a single hold of the lock, for efficiency. Add them to the supplied list.
2312 * Returns the number of new pages which were placed at *list.
2314 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2315 unsigned long count, struct list_head *list,
2316 int migratetype, bool cold)
2320 spin_lock(&zone->lock);
2321 for (i = 0; i < count; ++i) {
2322 struct page *page = __rmqueue(zone, order, migratetype);
2323 if (unlikely(page == NULL))
2326 if (unlikely(check_pcp_refill(page)))
2330 * Split buddy pages returned by expand() are received here
2331 * in physical page order. The page is added to the callers and
2332 * list and the list head then moves forward. From the callers
2333 * perspective, the linked list is ordered by page number in
2334 * some conditions. This is useful for IO devices that can
2335 * merge IO requests if the physical pages are ordered
2339 list_add(&page->lru, list);
2341 list_add_tail(&page->lru, list);
2344 if (is_migrate_cma(get_pcppage_migratetype(page)))
2345 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2350 * i pages were removed from the buddy list even if some leak due
2351 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2352 * on i. Do not confuse with 'alloced' which is the number of
2353 * pages added to the pcp list.
2355 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2356 spin_unlock(&zone->lock);
2362 * Called from the vmstat counter updater to drain pagesets of this
2363 * currently executing processor on remote nodes after they have
2366 * Note that this function must be called with the thread pinned to
2367 * a single processor.
2369 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2371 unsigned long flags;
2372 int to_drain, batch;
2374 local_irq_save(flags);
2375 batch = READ_ONCE(pcp->batch);
2376 to_drain = min(pcp->count, batch);
2378 free_pcppages_bulk(zone, to_drain, pcp);
2379 pcp->count -= to_drain;
2381 local_irq_restore(flags);
2386 * Drain pcplists of the indicated processor and zone.
2388 * The processor must either be the current processor and the
2389 * thread pinned to the current processor or a processor that
2392 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2394 unsigned long flags;
2395 struct per_cpu_pageset *pset;
2396 struct per_cpu_pages *pcp;
2398 local_irq_save(flags);
2399 pset = per_cpu_ptr(zone->pageset, cpu);
2403 free_pcppages_bulk(zone, pcp->count, pcp);
2406 local_irq_restore(flags);
2410 * Drain pcplists of all zones on the indicated processor.
2412 * The processor must either be the current processor and the
2413 * thread pinned to the current processor or a processor that
2416 static void drain_pages(unsigned int cpu)
2420 for_each_populated_zone(zone) {
2421 drain_pages_zone(cpu, zone);
2426 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2428 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2429 * the single zone's pages.
2431 void drain_local_pages(struct zone *zone)
2433 int cpu = smp_processor_id();
2436 drain_pages_zone(cpu, zone);
2441 static void drain_local_pages_wq(struct work_struct *work)
2444 * drain_all_pages doesn't use proper cpu hotplug protection so
2445 * we can race with cpu offline when the WQ can move this from
2446 * a cpu pinned worker to an unbound one. We can operate on a different
2447 * cpu which is allright but we also have to make sure to not move to
2451 drain_local_pages(NULL);
2456 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2458 * When zone parameter is non-NULL, spill just the single zone's pages.
2460 * Note that this can be extremely slow as the draining happens in a workqueue.
2462 void drain_all_pages(struct zone *zone)
2467 * Allocate in the BSS so we wont require allocation in
2468 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2470 static cpumask_t cpus_with_pcps;
2473 * Make sure nobody triggers this path before mm_percpu_wq is fully
2476 if (WARN_ON_ONCE(!mm_percpu_wq))
2479 /* Workqueues cannot recurse */
2480 if (current->flags & PF_WQ_WORKER)
2484 * Do not drain if one is already in progress unless it's specific to
2485 * a zone. Such callers are primarily CMA and memory hotplug and need
2486 * the drain to be complete when the call returns.
2488 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2491 mutex_lock(&pcpu_drain_mutex);
2495 * We don't care about racing with CPU hotplug event
2496 * as offline notification will cause the notified
2497 * cpu to drain that CPU pcps and on_each_cpu_mask
2498 * disables preemption as part of its processing
2500 for_each_online_cpu(cpu) {
2501 struct per_cpu_pageset *pcp;
2503 bool has_pcps = false;
2506 pcp = per_cpu_ptr(zone->pageset, cpu);
2510 for_each_populated_zone(z) {
2511 pcp = per_cpu_ptr(z->pageset, cpu);
2512 if (pcp->pcp.count) {
2520 cpumask_set_cpu(cpu, &cpus_with_pcps);
2522 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2525 for_each_cpu(cpu, &cpus_with_pcps) {
2526 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2527 INIT_WORK(work, drain_local_pages_wq);
2528 queue_work_on(cpu, mm_percpu_wq, work);
2530 for_each_cpu(cpu, &cpus_with_pcps)
2531 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2533 mutex_unlock(&pcpu_drain_mutex);
2536 #ifdef CONFIG_HIBERNATION
2538 void mark_free_pages(struct zone *zone)
2540 unsigned long pfn, max_zone_pfn;
2541 unsigned long flags;
2542 unsigned int order, t;
2545 if (zone_is_empty(zone))
2548 spin_lock_irqsave(&zone->lock, flags);
2550 max_zone_pfn = zone_end_pfn(zone);
2551 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2552 if (pfn_valid(pfn)) {
2553 page = pfn_to_page(pfn);
2555 if (page_zone(page) != zone)
2558 if (!swsusp_page_is_forbidden(page))
2559 swsusp_unset_page_free(page);
2562 for_each_migratetype_order(order, t) {
2563 list_for_each_entry(page,
2564 &zone->free_area[order].free_list[t], lru) {
2567 pfn = page_to_pfn(page);
2568 for (i = 0; i < (1UL << order); i++)
2569 swsusp_set_page_free(pfn_to_page(pfn + i));
2572 spin_unlock_irqrestore(&zone->lock, flags);
2574 #endif /* CONFIG_PM */
2577 * Free a 0-order page
2578 * cold == true ? free a cold page : free a hot page
2580 void free_hot_cold_page(struct page *page, bool cold)
2582 struct zone *zone = page_zone(page);
2583 struct per_cpu_pages *pcp;
2584 unsigned long flags;
2585 unsigned long pfn = page_to_pfn(page);
2588 if (!free_pcp_prepare(page))
2591 migratetype = get_pfnblock_migratetype(page, pfn);
2592 set_pcppage_migratetype(page, migratetype);
2593 local_irq_save(flags);
2594 __count_vm_event(PGFREE);
2597 * We only track unmovable, reclaimable and movable on pcp lists.
2598 * Free ISOLATE pages back to the allocator because they are being
2599 * offlined but treat HIGHATOMIC as movable pages so we can get those
2600 * areas back if necessary. Otherwise, we may have to free
2601 * excessively into the page allocator
2603 if (migratetype >= MIGRATE_PCPTYPES) {
2604 if (unlikely(is_migrate_isolate(migratetype))) {
2605 free_one_page(zone, page, pfn, 0, migratetype);
2608 migratetype = MIGRATE_MOVABLE;
2611 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2613 list_add(&page->lru, &pcp->lists[migratetype]);
2615 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2617 if (pcp->count >= pcp->high) {
2618 unsigned long batch = READ_ONCE(pcp->batch);
2619 free_pcppages_bulk(zone, batch, pcp);
2620 pcp->count -= batch;
2624 local_irq_restore(flags);
2628 * Free a list of 0-order pages
2630 void free_hot_cold_page_list(struct list_head *list, bool cold)
2632 struct page *page, *next;
2634 list_for_each_entry_safe(page, next, list, lru) {
2635 trace_mm_page_free_batched(page, cold);
2636 free_hot_cold_page(page, cold);
2641 * split_page takes a non-compound higher-order page, and splits it into
2642 * n (1<<order) sub-pages: page[0..n]
2643 * Each sub-page must be freed individually.
2645 * Note: this is probably too low level an operation for use in drivers.
2646 * Please consult with lkml before using this in your driver.
2648 void split_page(struct page *page, unsigned int order)
2652 VM_BUG_ON_PAGE(PageCompound(page), page);
2653 VM_BUG_ON_PAGE(!page_count(page), page);
2655 #ifdef CONFIG_KMEMCHECK
2657 * Split shadow pages too, because free(page[0]) would
2658 * otherwise free the whole shadow.
2660 if (kmemcheck_page_is_tracked(page))
2661 split_page(virt_to_page(page[0].shadow), order);
2664 for (i = 1; i < (1 << order); i++)
2665 set_page_refcounted(page + i);
2666 split_page_owner(page, order);
2668 EXPORT_SYMBOL_GPL(split_page);
2670 int __isolate_free_page(struct page *page, unsigned int order)
2672 unsigned long watermark;
2676 BUG_ON(!PageBuddy(page));
2678 zone = page_zone(page);
2679 mt = get_pageblock_migratetype(page);
2681 if (!is_migrate_isolate(mt)) {
2683 * Obey watermarks as if the page was being allocated. We can
2684 * emulate a high-order watermark check with a raised order-0
2685 * watermark, because we already know our high-order page
2688 watermark = min_wmark_pages(zone) + (1UL << order);
2689 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2692 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2695 /* Remove page from free list */
2696 list_del(&page->lru);
2697 zone->free_area[order].nr_free--;
2698 rmv_page_order(page);
2701 * Set the pageblock if the isolated page is at least half of a
2704 if (order >= pageblock_order - 1) {
2705 struct page *endpage = page + (1 << order) - 1;
2706 for (; page < endpage; page += pageblock_nr_pages) {
2707 int mt = get_pageblock_migratetype(page);
2708 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2709 && !is_migrate_highatomic(mt))
2710 set_pageblock_migratetype(page,
2716 return 1UL << order;
2720 * Update NUMA hit/miss statistics
2722 * Must be called with interrupts disabled.
2724 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2727 enum zone_stat_item local_stat = NUMA_LOCAL;
2729 if (z->node != numa_node_id())
2730 local_stat = NUMA_OTHER;
2732 if (z->node == preferred_zone->node)
2733 __inc_zone_state(z, NUMA_HIT);
2735 __inc_zone_state(z, NUMA_MISS);
2736 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2738 __inc_zone_state(z, local_stat);
2742 /* Remove page from the per-cpu list, caller must protect the list */
2743 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2744 bool cold, struct per_cpu_pages *pcp,
2745 struct list_head *list)
2750 if (list_empty(list)) {
2751 pcp->count += rmqueue_bulk(zone, 0,
2754 if (unlikely(list_empty(list)))
2759 page = list_last_entry(list, struct page, lru);
2761 page = list_first_entry(list, struct page, lru);
2763 list_del(&page->lru);
2765 } while (check_new_pcp(page));
2770 /* Lock and remove page from the per-cpu list */
2771 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2772 struct zone *zone, unsigned int order,
2773 gfp_t gfp_flags, int migratetype)
2775 struct per_cpu_pages *pcp;
2776 struct list_head *list;
2777 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2779 unsigned long flags;
2781 local_irq_save(flags);
2782 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2783 list = &pcp->lists[migratetype];
2784 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2786 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2787 zone_statistics(preferred_zone, zone);
2789 local_irq_restore(flags);
2794 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2797 struct page *rmqueue(struct zone *preferred_zone,
2798 struct zone *zone, unsigned int order,
2799 gfp_t gfp_flags, unsigned int alloc_flags,
2802 unsigned long flags;
2805 if (likely(order == 0)) {
2806 page = rmqueue_pcplist(preferred_zone, zone, order,
2807 gfp_flags, migratetype);
2812 * We most definitely don't want callers attempting to
2813 * allocate greater than order-1 page units with __GFP_NOFAIL.
2815 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2816 spin_lock_irqsave(&zone->lock, flags);
2820 if (alloc_flags & ALLOC_HARDER) {
2821 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2823 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2826 page = __rmqueue(zone, order, migratetype);
2827 } while (page && check_new_pages(page, order));
2828 spin_unlock(&zone->lock);
2831 __mod_zone_freepage_state(zone, -(1 << order),
2832 get_pcppage_migratetype(page));
2834 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2835 zone_statistics(preferred_zone, zone);
2836 local_irq_restore(flags);
2839 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2843 local_irq_restore(flags);
2847 #ifdef CONFIG_FAIL_PAGE_ALLOC
2850 struct fault_attr attr;
2852 bool ignore_gfp_highmem;
2853 bool ignore_gfp_reclaim;
2855 } fail_page_alloc = {
2856 .attr = FAULT_ATTR_INITIALIZER,
2857 .ignore_gfp_reclaim = true,
2858 .ignore_gfp_highmem = true,
2862 static int __init setup_fail_page_alloc(char *str)
2864 return setup_fault_attr(&fail_page_alloc.attr, str);
2866 __setup("fail_page_alloc=", setup_fail_page_alloc);
2868 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2870 if (order < fail_page_alloc.min_order)
2872 if (gfp_mask & __GFP_NOFAIL)
2874 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2876 if (fail_page_alloc.ignore_gfp_reclaim &&
2877 (gfp_mask & __GFP_DIRECT_RECLAIM))
2880 return should_fail(&fail_page_alloc.attr, 1 << order);
2883 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2885 static int __init fail_page_alloc_debugfs(void)
2887 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2890 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2891 &fail_page_alloc.attr);
2893 return PTR_ERR(dir);
2895 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2896 &fail_page_alloc.ignore_gfp_reclaim))
2898 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2899 &fail_page_alloc.ignore_gfp_highmem))
2901 if (!debugfs_create_u32("min-order", mode, dir,
2902 &fail_page_alloc.min_order))
2907 debugfs_remove_recursive(dir);
2912 late_initcall(fail_page_alloc_debugfs);
2914 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2916 #else /* CONFIG_FAIL_PAGE_ALLOC */
2918 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2923 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2926 * Return true if free base pages are above 'mark'. For high-order checks it
2927 * will return true of the order-0 watermark is reached and there is at least
2928 * one free page of a suitable size. Checking now avoids taking the zone lock
2929 * to check in the allocation paths if no pages are free.
2931 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2932 int classzone_idx, unsigned int alloc_flags,
2937 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2939 /* free_pages may go negative - that's OK */
2940 free_pages -= (1 << order) - 1;
2942 if (alloc_flags & ALLOC_HIGH)
2946 * If the caller does not have rights to ALLOC_HARDER then subtract
2947 * the high-atomic reserves. This will over-estimate the size of the
2948 * atomic reserve but it avoids a search.
2950 if (likely(!alloc_harder))
2951 free_pages -= z->nr_reserved_highatomic;
2956 /* If allocation can't use CMA areas don't use free CMA pages */
2957 if (!(alloc_flags & ALLOC_CMA))
2958 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2962 * Check watermarks for an order-0 allocation request. If these
2963 * are not met, then a high-order request also cannot go ahead
2964 * even if a suitable page happened to be free.
2966 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2969 /* If this is an order-0 request then the watermark is fine */
2973 /* For a high-order request, check at least one suitable page is free */
2974 for (o = order; o < MAX_ORDER; o++) {
2975 struct free_area *area = &z->free_area[o];
2984 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2985 if (!list_empty(&area->free_list[mt]))
2990 if ((alloc_flags & ALLOC_CMA) &&
2991 !list_empty(&area->free_list[MIGRATE_CMA])) {
2999 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3000 int classzone_idx, unsigned int alloc_flags)
3002 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3003 zone_page_state(z, NR_FREE_PAGES));
3006 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3007 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3009 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3013 /* If allocation can't use CMA areas don't use free CMA pages */
3014 if (!(alloc_flags & ALLOC_CMA))
3015 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3019 * Fast check for order-0 only. If this fails then the reserves
3020 * need to be calculated. There is a corner case where the check
3021 * passes but only the high-order atomic reserve are free. If
3022 * the caller is !atomic then it'll uselessly search the free
3023 * list. That corner case is then slower but it is harmless.
3025 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3028 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3032 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3033 unsigned long mark, int classzone_idx)
3035 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3037 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3038 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3040 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3045 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3047 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3050 #else /* CONFIG_NUMA */
3051 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3055 #endif /* CONFIG_NUMA */
3058 * get_page_from_freelist goes through the zonelist trying to allocate
3061 static struct page *
3062 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3063 const struct alloc_context *ac)
3065 struct zoneref *z = ac->preferred_zoneref;
3067 struct pglist_data *last_pgdat_dirty_limit = NULL;
3070 * Scan zonelist, looking for a zone with enough free.
3071 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3073 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3078 if (cpusets_enabled() &&
3079 (alloc_flags & ALLOC_CPUSET) &&
3080 !__cpuset_zone_allowed(zone, gfp_mask))
3083 * When allocating a page cache page for writing, we
3084 * want to get it from a node that is within its dirty
3085 * limit, such that no single node holds more than its
3086 * proportional share of globally allowed dirty pages.
3087 * The dirty limits take into account the node's
3088 * lowmem reserves and high watermark so that kswapd
3089 * should be able to balance it without having to
3090 * write pages from its LRU list.
3092 * XXX: For now, allow allocations to potentially
3093 * exceed the per-node dirty limit in the slowpath
3094 * (spread_dirty_pages unset) before going into reclaim,
3095 * which is important when on a NUMA setup the allowed
3096 * nodes are together not big enough to reach the
3097 * global limit. The proper fix for these situations
3098 * will require awareness of nodes in the
3099 * dirty-throttling and the flusher threads.
3101 if (ac->spread_dirty_pages) {
3102 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3105 if (!node_dirty_ok(zone->zone_pgdat)) {
3106 last_pgdat_dirty_limit = zone->zone_pgdat;
3111 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3112 if (!zone_watermark_fast(zone, order, mark,
3113 ac_classzone_idx(ac), alloc_flags)) {
3116 /* Checked here to keep the fast path fast */
3117 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3118 if (alloc_flags & ALLOC_NO_WATERMARKS)
3121 if (node_reclaim_mode == 0 ||
3122 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3125 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3127 case NODE_RECLAIM_NOSCAN:
3130 case NODE_RECLAIM_FULL:
3131 /* scanned but unreclaimable */
3134 /* did we reclaim enough */
3135 if (zone_watermark_ok(zone, order, mark,
3136 ac_classzone_idx(ac), alloc_flags))
3144 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3145 gfp_mask, alloc_flags, ac->migratetype);
3147 prep_new_page(page, order, gfp_mask, alloc_flags);
3150 * If this is a high-order atomic allocation then check
3151 * if the pageblock should be reserved for the future
3153 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3154 reserve_highatomic_pageblock(page, zone, order);
3164 * Large machines with many possible nodes should not always dump per-node
3165 * meminfo in irq context.
3167 static inline bool should_suppress_show_mem(void)
3172 ret = in_interrupt();
3177 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3179 unsigned int filter = SHOW_MEM_FILTER_NODES;
3180 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3182 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3186 * This documents exceptions given to allocations in certain
3187 * contexts that are allowed to allocate outside current's set
3190 if (!(gfp_mask & __GFP_NOMEMALLOC))
3191 if (test_thread_flag(TIF_MEMDIE) ||
3192 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3193 filter &= ~SHOW_MEM_FILTER_NODES;
3194 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3195 filter &= ~SHOW_MEM_FILTER_NODES;
3197 show_mem(filter, nodemask);
3200 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3202 struct va_format vaf;
3204 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3205 DEFAULT_RATELIMIT_BURST);
3207 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3210 pr_warn("%s: ", current->comm);
3212 va_start(args, fmt);
3215 pr_cont("%pV", &vaf);
3218 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3220 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3222 pr_cont("(null)\n");
3224 cpuset_print_current_mems_allowed();
3227 warn_alloc_show_mem(gfp_mask, nodemask);
3230 static inline struct page *
3231 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3232 unsigned int alloc_flags,
3233 const struct alloc_context *ac)
3237 page = get_page_from_freelist(gfp_mask, order,
3238 alloc_flags|ALLOC_CPUSET, ac);
3240 * fallback to ignore cpuset restriction if our nodes
3244 page = get_page_from_freelist(gfp_mask, order,
3250 static inline struct page *
3251 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3252 const struct alloc_context *ac, unsigned long *did_some_progress)
3254 struct oom_control oc = {
3255 .zonelist = ac->zonelist,
3256 .nodemask = ac->nodemask,
3258 .gfp_mask = gfp_mask,
3263 *did_some_progress = 0;
3266 * Acquire the oom lock. If that fails, somebody else is
3267 * making progress for us.
3269 if (!mutex_trylock(&oom_lock)) {
3270 *did_some_progress = 1;
3271 schedule_timeout_uninterruptible(1);
3276 * Go through the zonelist yet one more time, keep very high watermark
3277 * here, this is only to catch a parallel oom killing, we must fail if
3278 * we're still under heavy pressure.
3280 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3281 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3285 /* Coredumps can quickly deplete all memory reserves */
3286 if (current->flags & PF_DUMPCORE)
3288 /* The OOM killer will not help higher order allocs */
3289 if (order > PAGE_ALLOC_COSTLY_ORDER)
3292 * We have already exhausted all our reclaim opportunities without any
3293 * success so it is time to admit defeat. We will skip the OOM killer
3294 * because it is very likely that the caller has a more reasonable
3295 * fallback than shooting a random task.
3297 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3299 /* The OOM killer does not needlessly kill tasks for lowmem */
3300 if (ac->high_zoneidx < ZONE_NORMAL)
3302 if (pm_suspended_storage())
3305 * XXX: GFP_NOFS allocations should rather fail than rely on
3306 * other request to make a forward progress.
3307 * We are in an unfortunate situation where out_of_memory cannot
3308 * do much for this context but let's try it to at least get
3309 * access to memory reserved if the current task is killed (see
3310 * out_of_memory). Once filesystems are ready to handle allocation
3311 * failures more gracefully we should just bail out here.
3314 /* The OOM killer may not free memory on a specific node */
3315 if (gfp_mask & __GFP_THISNODE)
3318 /* Exhausted what can be done so it's blamo time */
3319 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3320 *did_some_progress = 1;
3323 * Help non-failing allocations by giving them access to memory
3326 if (gfp_mask & __GFP_NOFAIL)
3327 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3328 ALLOC_NO_WATERMARKS, ac);
3331 mutex_unlock(&oom_lock);
3336 * Maximum number of compaction retries wit a progress before OOM
3337 * killer is consider as the only way to move forward.
3339 #define MAX_COMPACT_RETRIES 16
3341 #ifdef CONFIG_COMPACTION
3342 /* Try memory compaction for high-order allocations before reclaim */
3343 static struct page *
3344 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3345 unsigned int alloc_flags, const struct alloc_context *ac,
3346 enum compact_priority prio, enum compact_result *compact_result)
3349 unsigned int noreclaim_flag;
3354 noreclaim_flag = memalloc_noreclaim_save();
3355 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3357 memalloc_noreclaim_restore(noreclaim_flag);
3359 if (*compact_result <= COMPACT_INACTIVE)
3363 * At least in one zone compaction wasn't deferred or skipped, so let's
3364 * count a compaction stall
3366 count_vm_event(COMPACTSTALL);
3368 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3371 struct zone *zone = page_zone(page);
3373 zone->compact_blockskip_flush = false;
3374 compaction_defer_reset(zone, order, true);
3375 count_vm_event(COMPACTSUCCESS);
3380 * It's bad if compaction run occurs and fails. The most likely reason
3381 * is that pages exist, but not enough to satisfy watermarks.
3383 count_vm_event(COMPACTFAIL);
3391 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3392 enum compact_result compact_result,
3393 enum compact_priority *compact_priority,
3394 int *compaction_retries)
3396 int max_retries = MAX_COMPACT_RETRIES;
3399 int retries = *compaction_retries;
3400 enum compact_priority priority = *compact_priority;
3405 if (compaction_made_progress(compact_result))
3406 (*compaction_retries)++;
3409 * compaction considers all the zone as desperately out of memory
3410 * so it doesn't really make much sense to retry except when the
3411 * failure could be caused by insufficient priority
3413 if (compaction_failed(compact_result))
3414 goto check_priority;
3417 * make sure the compaction wasn't deferred or didn't bail out early
3418 * due to locks contention before we declare that we should give up.
3419 * But do not retry if the given zonelist is not suitable for
3422 if (compaction_withdrawn(compact_result)) {
3423 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3428 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3429 * costly ones because they are de facto nofail and invoke OOM
3430 * killer to move on while costly can fail and users are ready
3431 * to cope with that. 1/4 retries is rather arbitrary but we
3432 * would need much more detailed feedback from compaction to
3433 * make a better decision.
3435 if (order > PAGE_ALLOC_COSTLY_ORDER)
3437 if (*compaction_retries <= max_retries) {
3443 * Make sure there are attempts at the highest priority if we exhausted
3444 * all retries or failed at the lower priorities.
3447 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3448 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3450 if (*compact_priority > min_priority) {
3451 (*compact_priority)--;
3452 *compaction_retries = 0;
3456 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3460 static inline struct page *
3461 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3462 unsigned int alloc_flags, const struct alloc_context *ac,
3463 enum compact_priority prio, enum compact_result *compact_result)
3465 *compact_result = COMPACT_SKIPPED;
3470 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3471 enum compact_result compact_result,
3472 enum compact_priority *compact_priority,
3473 int *compaction_retries)
3478 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3482 * There are setups with compaction disabled which would prefer to loop
3483 * inside the allocator rather than hit the oom killer prematurely.
3484 * Let's give them a good hope and keep retrying while the order-0
3485 * watermarks are OK.
3487 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3489 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3490 ac_classzone_idx(ac), alloc_flags))
3495 #endif /* CONFIG_COMPACTION */
3497 /* Perform direct synchronous page reclaim */
3499 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3500 const struct alloc_context *ac)
3502 struct reclaim_state reclaim_state;
3504 unsigned int noreclaim_flag;
3508 /* We now go into synchronous reclaim */
3509 cpuset_memory_pressure_bump();
3510 noreclaim_flag = memalloc_noreclaim_save();
3511 lockdep_set_current_reclaim_state(gfp_mask);
3512 reclaim_state.reclaimed_slab = 0;
3513 current->reclaim_state = &reclaim_state;
3515 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3518 current->reclaim_state = NULL;
3519 lockdep_clear_current_reclaim_state();
3520 memalloc_noreclaim_restore(noreclaim_flag);
3527 /* The really slow allocator path where we enter direct reclaim */
3528 static inline struct page *
3529 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3530 unsigned int alloc_flags, const struct alloc_context *ac,
3531 unsigned long *did_some_progress)
3533 struct page *page = NULL;
3534 bool drained = false;
3536 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3537 if (unlikely(!(*did_some_progress)))
3541 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3544 * If an allocation failed after direct reclaim, it could be because
3545 * pages are pinned on the per-cpu lists or in high alloc reserves.
3546 * Shrink them them and try again
3548 if (!page && !drained) {
3549 unreserve_highatomic_pageblock(ac, false);
3550 drain_all_pages(NULL);
3558 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3562 pg_data_t *last_pgdat = NULL;
3564 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3565 ac->high_zoneidx, ac->nodemask) {
3566 if (last_pgdat != zone->zone_pgdat)
3567 wakeup_kswapd(zone, order, ac->high_zoneidx);
3568 last_pgdat = zone->zone_pgdat;
3572 static inline unsigned int
3573 gfp_to_alloc_flags(gfp_t gfp_mask)
3575 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3577 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3578 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3581 * The caller may dip into page reserves a bit more if the caller
3582 * cannot run direct reclaim, or if the caller has realtime scheduling
3583 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3584 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3586 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3588 if (gfp_mask & __GFP_ATOMIC) {
3590 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3591 * if it can't schedule.
3593 if (!(gfp_mask & __GFP_NOMEMALLOC))
3594 alloc_flags |= ALLOC_HARDER;
3596 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3597 * comment for __cpuset_node_allowed().
3599 alloc_flags &= ~ALLOC_CPUSET;
3600 } else if (unlikely(rt_task(current)) && !in_interrupt())
3601 alloc_flags |= ALLOC_HARDER;
3604 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3605 alloc_flags |= ALLOC_CMA;
3610 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3612 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3615 if (gfp_mask & __GFP_MEMALLOC)
3617 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3619 if (!in_interrupt() &&
3620 ((current->flags & PF_MEMALLOC) ||
3621 unlikely(test_thread_flag(TIF_MEMDIE))))
3628 * Checks whether it makes sense to retry the reclaim to make a forward progress
3629 * for the given allocation request.
3631 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3632 * without success, or when we couldn't even meet the watermark if we
3633 * reclaimed all remaining pages on the LRU lists.
3635 * Returns true if a retry is viable or false to enter the oom path.
3638 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3639 struct alloc_context *ac, int alloc_flags,
3640 bool did_some_progress, int *no_progress_loops)
3646 * Costly allocations might have made a progress but this doesn't mean
3647 * their order will become available due to high fragmentation so
3648 * always increment the no progress counter for them
3650 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3651 *no_progress_loops = 0;
3653 (*no_progress_loops)++;
3656 * Make sure we converge to OOM if we cannot make any progress
3657 * several times in the row.
3659 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3660 /* Before OOM, exhaust highatomic_reserve */
3661 return unreserve_highatomic_pageblock(ac, true);
3665 * Keep reclaiming pages while there is a chance this will lead
3666 * somewhere. If none of the target zones can satisfy our allocation
3667 * request even if all reclaimable pages are considered then we are
3668 * screwed and have to go OOM.
3670 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3672 unsigned long available;
3673 unsigned long reclaimable;
3674 unsigned long min_wmark = min_wmark_pages(zone);
3677 available = reclaimable = zone_reclaimable_pages(zone);
3678 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3681 * Would the allocation succeed if we reclaimed all
3682 * reclaimable pages?
3684 wmark = __zone_watermark_ok(zone, order, min_wmark,
3685 ac_classzone_idx(ac), alloc_flags, available);
3686 trace_reclaim_retry_zone(z, order, reclaimable,
3687 available, min_wmark, *no_progress_loops, wmark);
3690 * If we didn't make any progress and have a lot of
3691 * dirty + writeback pages then we should wait for
3692 * an IO to complete to slow down the reclaim and
3693 * prevent from pre mature OOM
3695 if (!did_some_progress) {
3696 unsigned long write_pending;
3698 write_pending = zone_page_state_snapshot(zone,
3699 NR_ZONE_WRITE_PENDING);
3701 if (2 * write_pending > reclaimable) {
3702 congestion_wait(BLK_RW_ASYNC, HZ/10);
3708 * Memory allocation/reclaim might be called from a WQ
3709 * context and the current implementation of the WQ
3710 * concurrency control doesn't recognize that
3711 * a particular WQ is congested if the worker thread is
3712 * looping without ever sleeping. Therefore we have to
3713 * do a short sleep here rather than calling
3716 if (current->flags & PF_WQ_WORKER)
3717 schedule_timeout_uninterruptible(1);
3729 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3732 * It's possible that cpuset's mems_allowed and the nodemask from
3733 * mempolicy don't intersect. This should be normally dealt with by
3734 * policy_nodemask(), but it's possible to race with cpuset update in
3735 * such a way the check therein was true, and then it became false
3736 * before we got our cpuset_mems_cookie here.
3737 * This assumes that for all allocations, ac->nodemask can come only
3738 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3739 * when it does not intersect with the cpuset restrictions) or the
3740 * caller can deal with a violated nodemask.
3742 if (cpusets_enabled() && ac->nodemask &&
3743 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3744 ac->nodemask = NULL;
3749 * When updating a task's mems_allowed or mempolicy nodemask, it is
3750 * possible to race with parallel threads in such a way that our
3751 * allocation can fail while the mask is being updated. If we are about
3752 * to fail, check if the cpuset changed during allocation and if so,
3755 if (read_mems_allowed_retry(cpuset_mems_cookie))
3761 static inline struct page *
3762 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3763 struct alloc_context *ac)
3765 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3766 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3767 struct page *page = NULL;
3768 unsigned int alloc_flags;
3769 unsigned long did_some_progress;
3770 enum compact_priority compact_priority;
3771 enum compact_result compact_result;
3772 int compaction_retries;
3773 int no_progress_loops;
3774 unsigned long alloc_start = jiffies;
3775 unsigned int stall_timeout = 10 * HZ;
3776 unsigned int cpuset_mems_cookie;
3779 * In the slowpath, we sanity check order to avoid ever trying to
3780 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3781 * be using allocators in order of preference for an area that is
3784 if (order >= MAX_ORDER) {
3785 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3790 * We also sanity check to catch abuse of atomic reserves being used by
3791 * callers that are not in atomic context.
3793 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3794 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3795 gfp_mask &= ~__GFP_ATOMIC;
3798 compaction_retries = 0;
3799 no_progress_loops = 0;
3800 compact_priority = DEF_COMPACT_PRIORITY;
3801 cpuset_mems_cookie = read_mems_allowed_begin();
3804 * The fast path uses conservative alloc_flags to succeed only until
3805 * kswapd needs to be woken up, and to avoid the cost of setting up
3806 * alloc_flags precisely. So we do that now.
3808 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3811 * We need to recalculate the starting point for the zonelist iterator
3812 * because we might have used different nodemask in the fast path, or
3813 * there was a cpuset modification and we are retrying - otherwise we
3814 * could end up iterating over non-eligible zones endlessly.
3816 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3817 ac->high_zoneidx, ac->nodemask);
3818 if (!ac->preferred_zoneref->zone)
3821 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3822 wake_all_kswapds(order, ac);
3825 * The adjusted alloc_flags might result in immediate success, so try
3828 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3833 * For costly allocations, try direct compaction first, as it's likely
3834 * that we have enough base pages and don't need to reclaim. For non-
3835 * movable high-order allocations, do that as well, as compaction will
3836 * try prevent permanent fragmentation by migrating from blocks of the
3838 * Don't try this for allocations that are allowed to ignore
3839 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3841 if (can_direct_reclaim &&
3843 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3844 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3845 page = __alloc_pages_direct_compact(gfp_mask, order,
3847 INIT_COMPACT_PRIORITY,
3853 * Checks for costly allocations with __GFP_NORETRY, which
3854 * includes THP page fault allocations
3856 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
3858 * If compaction is deferred for high-order allocations,
3859 * it is because sync compaction recently failed. If
3860 * this is the case and the caller requested a THP
3861 * allocation, we do not want to heavily disrupt the
3862 * system, so we fail the allocation instead of entering
3865 if (compact_result == COMPACT_DEFERRED)
3869 * Looks like reclaim/compaction is worth trying, but
3870 * sync compaction could be very expensive, so keep
3871 * using async compaction.
3873 compact_priority = INIT_COMPACT_PRIORITY;
3878 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3879 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3880 wake_all_kswapds(order, ac);
3882 if (gfp_pfmemalloc_allowed(gfp_mask))
3883 alloc_flags = ALLOC_NO_WATERMARKS;
3886 * Reset the zonelist iterators if memory policies can be ignored.
3887 * These allocations are high priority and system rather than user
3890 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3891 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3892 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3893 ac->high_zoneidx, ac->nodemask);
3896 /* Attempt with potentially adjusted zonelist and alloc_flags */
3897 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3901 /* Caller is not willing to reclaim, we can't balance anything */
3902 if (!can_direct_reclaim)
3905 /* Make sure we know about allocations which stall for too long */
3906 if (time_after(jiffies, alloc_start + stall_timeout)) {
3907 warn_alloc(gfp_mask & ~__GFP_NOWARN, ac->nodemask,
3908 "page allocation stalls for %ums, order:%u",
3909 jiffies_to_msecs(jiffies-alloc_start), order);
3910 stall_timeout += 10 * HZ;
3913 /* Avoid recursion of direct reclaim */
3914 if (current->flags & PF_MEMALLOC)
3917 /* Try direct reclaim and then allocating */
3918 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3919 &did_some_progress);
3923 /* Try direct compaction and then allocating */
3924 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3925 compact_priority, &compact_result);
3929 /* Do not loop if specifically requested */
3930 if (gfp_mask & __GFP_NORETRY)
3934 * Do not retry costly high order allocations unless they are
3935 * __GFP_RETRY_MAYFAIL
3937 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
3940 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3941 did_some_progress > 0, &no_progress_loops))
3945 * It doesn't make any sense to retry for the compaction if the order-0
3946 * reclaim is not able to make any progress because the current
3947 * implementation of the compaction depends on the sufficient amount
3948 * of free memory (see __compaction_suitable)
3950 if (did_some_progress > 0 &&
3951 should_compact_retry(ac, order, alloc_flags,
3952 compact_result, &compact_priority,
3953 &compaction_retries))
3957 /* Deal with possible cpuset update races before we start OOM killing */
3958 if (check_retry_cpuset(cpuset_mems_cookie, ac))
3961 /* Reclaim has failed us, start killing things */
3962 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3966 /* Avoid allocations with no watermarks from looping endlessly */
3967 if (test_thread_flag(TIF_MEMDIE) &&
3968 (alloc_flags == ALLOC_NO_WATERMARKS ||
3969 (gfp_mask & __GFP_NOMEMALLOC)))
3972 /* Retry as long as the OOM killer is making progress */
3973 if (did_some_progress) {
3974 no_progress_loops = 0;
3979 /* Deal with possible cpuset update races before we fail */
3980 if (check_retry_cpuset(cpuset_mems_cookie, ac))
3984 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
3987 if (gfp_mask & __GFP_NOFAIL) {
3989 * All existing users of the __GFP_NOFAIL are blockable, so warn
3990 * of any new users that actually require GFP_NOWAIT
3992 if (WARN_ON_ONCE(!can_direct_reclaim))
3996 * PF_MEMALLOC request from this context is rather bizarre
3997 * because we cannot reclaim anything and only can loop waiting
3998 * for somebody to do a work for us
4000 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4003 * non failing costly orders are a hard requirement which we
4004 * are not prepared for much so let's warn about these users
4005 * so that we can identify them and convert them to something
4008 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4011 * Help non-failing allocations by giving them access to memory
4012 * reserves but do not use ALLOC_NO_WATERMARKS because this
4013 * could deplete whole memory reserves which would just make
4014 * the situation worse
4016 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4024 warn_alloc(gfp_mask, ac->nodemask,
4025 "page allocation failure: order:%u", order);
4030 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4031 int preferred_nid, nodemask_t *nodemask,
4032 struct alloc_context *ac, gfp_t *alloc_mask,
4033 unsigned int *alloc_flags)
4035 ac->high_zoneidx = gfp_zone(gfp_mask);
4036 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4037 ac->nodemask = nodemask;
4038 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4040 if (cpusets_enabled()) {
4041 *alloc_mask |= __GFP_HARDWALL;
4043 ac->nodemask = &cpuset_current_mems_allowed;
4045 *alloc_flags |= ALLOC_CPUSET;
4048 lockdep_trace_alloc(gfp_mask);
4050 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4052 if (should_fail_alloc_page(gfp_mask, order))
4055 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4056 *alloc_flags |= ALLOC_CMA;
4061 /* Determine whether to spread dirty pages and what the first usable zone */
4062 static inline void finalise_ac(gfp_t gfp_mask,
4063 unsigned int order, struct alloc_context *ac)
4065 /* Dirty zone balancing only done in the fast path */
4066 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4069 * The preferred zone is used for statistics but crucially it is
4070 * also used as the starting point for the zonelist iterator. It
4071 * may get reset for allocations that ignore memory policies.
4073 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4074 ac->high_zoneidx, ac->nodemask);
4078 * This is the 'heart' of the zoned buddy allocator.
4081 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4082 nodemask_t *nodemask)
4085 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4086 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
4087 struct alloc_context ac = { };
4089 gfp_mask &= gfp_allowed_mask;
4090 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4093 finalise_ac(gfp_mask, order, &ac);
4095 /* First allocation attempt */
4096 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4101 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4102 * resp. GFP_NOIO which has to be inherited for all allocation requests
4103 * from a particular context which has been marked by
4104 * memalloc_no{fs,io}_{save,restore}.
4106 alloc_mask = current_gfp_context(gfp_mask);
4107 ac.spread_dirty_pages = false;
4110 * Restore the original nodemask if it was potentially replaced with
4111 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4113 if (unlikely(ac.nodemask != nodemask))
4114 ac.nodemask = nodemask;
4116 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4119 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4120 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4121 __free_pages(page, order);
4125 if (kmemcheck_enabled && page)
4126 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
4128 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4132 EXPORT_SYMBOL(__alloc_pages_nodemask);
4135 * Common helper functions.
4137 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4142 * __get_free_pages() returns a 32-bit address, which cannot represent
4145 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4147 page = alloc_pages(gfp_mask, order);
4150 return (unsigned long) page_address(page);
4152 EXPORT_SYMBOL(__get_free_pages);
4154 unsigned long get_zeroed_page(gfp_t gfp_mask)
4156 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4158 EXPORT_SYMBOL(get_zeroed_page);
4160 void __free_pages(struct page *page, unsigned int order)
4162 if (put_page_testzero(page)) {
4164 free_hot_cold_page(page, false);
4166 __free_pages_ok(page, order);
4170 EXPORT_SYMBOL(__free_pages);
4172 void free_pages(unsigned long addr, unsigned int order)
4175 VM_BUG_ON(!virt_addr_valid((void *)addr));
4176 __free_pages(virt_to_page((void *)addr), order);
4180 EXPORT_SYMBOL(free_pages);
4184 * An arbitrary-length arbitrary-offset area of memory which resides
4185 * within a 0 or higher order page. Multiple fragments within that page
4186 * are individually refcounted, in the page's reference counter.
4188 * The page_frag functions below provide a simple allocation framework for
4189 * page fragments. This is used by the network stack and network device
4190 * drivers to provide a backing region of memory for use as either an
4191 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4193 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4196 struct page *page = NULL;
4197 gfp_t gfp = gfp_mask;
4199 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4200 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4202 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4203 PAGE_FRAG_CACHE_MAX_ORDER);
4204 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4206 if (unlikely(!page))
4207 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4209 nc->va = page ? page_address(page) : NULL;
4214 void __page_frag_cache_drain(struct page *page, unsigned int count)
4216 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4218 if (page_ref_sub_and_test(page, count)) {
4219 unsigned int order = compound_order(page);
4222 free_hot_cold_page(page, false);
4224 __free_pages_ok(page, order);
4227 EXPORT_SYMBOL(__page_frag_cache_drain);
4229 void *page_frag_alloc(struct page_frag_cache *nc,
4230 unsigned int fragsz, gfp_t gfp_mask)
4232 unsigned int size = PAGE_SIZE;
4236 if (unlikely(!nc->va)) {
4238 page = __page_frag_cache_refill(nc, gfp_mask);
4242 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4243 /* if size can vary use size else just use PAGE_SIZE */
4246 /* Even if we own the page, we do not use atomic_set().
4247 * This would break get_page_unless_zero() users.
4249 page_ref_add(page, size - 1);
4251 /* reset page count bias and offset to start of new frag */
4252 nc->pfmemalloc = page_is_pfmemalloc(page);
4253 nc->pagecnt_bias = size;
4257 offset = nc->offset - fragsz;
4258 if (unlikely(offset < 0)) {
4259 page = virt_to_page(nc->va);
4261 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4264 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4265 /* if size can vary use size else just use PAGE_SIZE */
4268 /* OK, page count is 0, we can safely set it */
4269 set_page_count(page, size);
4271 /* reset page count bias and offset to start of new frag */
4272 nc->pagecnt_bias = size;
4273 offset = size - fragsz;
4277 nc->offset = offset;
4279 return nc->va + offset;
4281 EXPORT_SYMBOL(page_frag_alloc);
4284 * Frees a page fragment allocated out of either a compound or order 0 page.
4286 void page_frag_free(void *addr)
4288 struct page *page = virt_to_head_page(addr);
4290 if (unlikely(put_page_testzero(page)))
4291 __free_pages_ok(page, compound_order(page));
4293 EXPORT_SYMBOL(page_frag_free);
4295 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4299 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4300 unsigned long used = addr + PAGE_ALIGN(size);
4302 split_page(virt_to_page((void *)addr), order);
4303 while (used < alloc_end) {
4308 return (void *)addr;
4312 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4313 * @size: the number of bytes to allocate
4314 * @gfp_mask: GFP flags for the allocation
4316 * This function is similar to alloc_pages(), except that it allocates the
4317 * minimum number of pages to satisfy the request. alloc_pages() can only
4318 * allocate memory in power-of-two pages.
4320 * This function is also limited by MAX_ORDER.
4322 * Memory allocated by this function must be released by free_pages_exact().
4324 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4326 unsigned int order = get_order(size);
4329 addr = __get_free_pages(gfp_mask, order);
4330 return make_alloc_exact(addr, order, size);
4332 EXPORT_SYMBOL(alloc_pages_exact);
4335 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4337 * @nid: the preferred node ID where memory should be allocated
4338 * @size: the number of bytes to allocate
4339 * @gfp_mask: GFP flags for the allocation
4341 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4344 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4346 unsigned int order = get_order(size);
4347 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4350 return make_alloc_exact((unsigned long)page_address(p), order, size);
4354 * free_pages_exact - release memory allocated via alloc_pages_exact()
4355 * @virt: the value returned by alloc_pages_exact.
4356 * @size: size of allocation, same value as passed to alloc_pages_exact().
4358 * Release the memory allocated by a previous call to alloc_pages_exact.
4360 void free_pages_exact(void *virt, size_t size)
4362 unsigned long addr = (unsigned long)virt;
4363 unsigned long end = addr + PAGE_ALIGN(size);
4365 while (addr < end) {
4370 EXPORT_SYMBOL(free_pages_exact);
4373 * nr_free_zone_pages - count number of pages beyond high watermark
4374 * @offset: The zone index of the highest zone
4376 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4377 * high watermark within all zones at or below a given zone index. For each
4378 * zone, the number of pages is calculated as:
4380 * nr_free_zone_pages = managed_pages - high_pages
4382 static unsigned long nr_free_zone_pages(int offset)
4387 /* Just pick one node, since fallback list is circular */
4388 unsigned long sum = 0;
4390 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4392 for_each_zone_zonelist(zone, z, zonelist, offset) {
4393 unsigned long size = zone->managed_pages;
4394 unsigned long high = high_wmark_pages(zone);
4403 * nr_free_buffer_pages - count number of pages beyond high watermark
4405 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4406 * watermark within ZONE_DMA and ZONE_NORMAL.
4408 unsigned long nr_free_buffer_pages(void)
4410 return nr_free_zone_pages(gfp_zone(GFP_USER));
4412 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4415 * nr_free_pagecache_pages - count number of pages beyond high watermark
4417 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4418 * high watermark within all zones.
4420 unsigned long nr_free_pagecache_pages(void)
4422 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4425 static inline void show_node(struct zone *zone)
4427 if (IS_ENABLED(CONFIG_NUMA))
4428 printk("Node %d ", zone_to_nid(zone));
4431 long si_mem_available(void)
4434 unsigned long pagecache;
4435 unsigned long wmark_low = 0;
4436 unsigned long pages[NR_LRU_LISTS];
4440 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4441 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4444 wmark_low += zone->watermark[WMARK_LOW];
4447 * Estimate the amount of memory available for userspace allocations,
4448 * without causing swapping.
4450 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4453 * Not all the page cache can be freed, otherwise the system will
4454 * start swapping. Assume at least half of the page cache, or the
4455 * low watermark worth of cache, needs to stay.
4457 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4458 pagecache -= min(pagecache / 2, wmark_low);
4459 available += pagecache;
4462 * Part of the reclaimable slab consists of items that are in use,
4463 * and cannot be freed. Cap this estimate at the low watermark.
4465 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4466 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4473 EXPORT_SYMBOL_GPL(si_mem_available);
4475 void si_meminfo(struct sysinfo *val)
4477 val->totalram = totalram_pages;
4478 val->sharedram = global_node_page_state(NR_SHMEM);
4479 val->freeram = global_page_state(NR_FREE_PAGES);
4480 val->bufferram = nr_blockdev_pages();
4481 val->totalhigh = totalhigh_pages;
4482 val->freehigh = nr_free_highpages();
4483 val->mem_unit = PAGE_SIZE;
4486 EXPORT_SYMBOL(si_meminfo);
4489 void si_meminfo_node(struct sysinfo *val, int nid)
4491 int zone_type; /* needs to be signed */
4492 unsigned long managed_pages = 0;
4493 unsigned long managed_highpages = 0;
4494 unsigned long free_highpages = 0;
4495 pg_data_t *pgdat = NODE_DATA(nid);
4497 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4498 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4499 val->totalram = managed_pages;
4500 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4501 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4502 #ifdef CONFIG_HIGHMEM
4503 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4504 struct zone *zone = &pgdat->node_zones[zone_type];
4506 if (is_highmem(zone)) {
4507 managed_highpages += zone->managed_pages;
4508 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4511 val->totalhigh = managed_highpages;
4512 val->freehigh = free_highpages;
4514 val->totalhigh = managed_highpages;
4515 val->freehigh = free_highpages;
4517 val->mem_unit = PAGE_SIZE;
4522 * Determine whether the node should be displayed or not, depending on whether
4523 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4525 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4527 if (!(flags & SHOW_MEM_FILTER_NODES))
4531 * no node mask - aka implicit memory numa policy. Do not bother with
4532 * the synchronization - read_mems_allowed_begin - because we do not
4533 * have to be precise here.
4536 nodemask = &cpuset_current_mems_allowed;
4538 return !node_isset(nid, *nodemask);
4541 #define K(x) ((x) << (PAGE_SHIFT-10))
4543 static void show_migration_types(unsigned char type)
4545 static const char types[MIGRATE_TYPES] = {
4546 [MIGRATE_UNMOVABLE] = 'U',
4547 [MIGRATE_MOVABLE] = 'M',
4548 [MIGRATE_RECLAIMABLE] = 'E',
4549 [MIGRATE_HIGHATOMIC] = 'H',
4551 [MIGRATE_CMA] = 'C',
4553 #ifdef CONFIG_MEMORY_ISOLATION
4554 [MIGRATE_ISOLATE] = 'I',
4557 char tmp[MIGRATE_TYPES + 1];
4561 for (i = 0; i < MIGRATE_TYPES; i++) {
4562 if (type & (1 << i))
4567 printk(KERN_CONT "(%s) ", tmp);
4571 * Show free area list (used inside shift_scroll-lock stuff)
4572 * We also calculate the percentage fragmentation. We do this by counting the
4573 * memory on each free list with the exception of the first item on the list.
4576 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4579 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4581 unsigned long free_pcp = 0;
4586 for_each_populated_zone(zone) {
4587 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4590 for_each_online_cpu(cpu)
4591 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4594 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4595 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4596 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4597 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4598 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4599 " free:%lu free_pcp:%lu free_cma:%lu\n",
4600 global_node_page_state(NR_ACTIVE_ANON),
4601 global_node_page_state(NR_INACTIVE_ANON),
4602 global_node_page_state(NR_ISOLATED_ANON),
4603 global_node_page_state(NR_ACTIVE_FILE),
4604 global_node_page_state(NR_INACTIVE_FILE),
4605 global_node_page_state(NR_ISOLATED_FILE),
4606 global_node_page_state(NR_UNEVICTABLE),
4607 global_node_page_state(NR_FILE_DIRTY),
4608 global_node_page_state(NR_WRITEBACK),
4609 global_node_page_state(NR_UNSTABLE_NFS),
4610 global_node_page_state(NR_SLAB_RECLAIMABLE),
4611 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4612 global_node_page_state(NR_FILE_MAPPED),
4613 global_node_page_state(NR_SHMEM),
4614 global_page_state(NR_PAGETABLE),
4615 global_page_state(NR_BOUNCE),
4616 global_page_state(NR_FREE_PAGES),
4618 global_page_state(NR_FREE_CMA_PAGES));
4620 for_each_online_pgdat(pgdat) {
4621 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4625 " active_anon:%lukB"
4626 " inactive_anon:%lukB"
4627 " active_file:%lukB"
4628 " inactive_file:%lukB"
4629 " unevictable:%lukB"
4630 " isolated(anon):%lukB"
4631 " isolated(file):%lukB"
4636 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4638 " shmem_pmdmapped: %lukB"
4641 " writeback_tmp:%lukB"
4643 " all_unreclaimable? %s"
4646 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4647 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4648 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4649 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4650 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4651 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4652 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4653 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4654 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4655 K(node_page_state(pgdat, NR_WRITEBACK)),
4656 K(node_page_state(pgdat, NR_SHMEM)),
4657 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4658 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4659 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4661 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4663 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4664 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4665 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4669 for_each_populated_zone(zone) {
4672 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4676 for_each_online_cpu(cpu)
4677 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4686 " active_anon:%lukB"
4687 " inactive_anon:%lukB"
4688 " active_file:%lukB"
4689 " inactive_file:%lukB"
4690 " unevictable:%lukB"
4691 " writepending:%lukB"
4695 " kernel_stack:%lukB"
4703 K(zone_page_state(zone, NR_FREE_PAGES)),
4704 K(min_wmark_pages(zone)),
4705 K(low_wmark_pages(zone)),
4706 K(high_wmark_pages(zone)),
4707 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4708 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4709 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4710 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4711 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4712 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4713 K(zone->present_pages),
4714 K(zone->managed_pages),
4715 K(zone_page_state(zone, NR_MLOCK)),
4716 zone_page_state(zone, NR_KERNEL_STACK_KB),
4717 K(zone_page_state(zone, NR_PAGETABLE)),
4718 K(zone_page_state(zone, NR_BOUNCE)),
4720 K(this_cpu_read(zone->pageset->pcp.count)),
4721 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4722 printk("lowmem_reserve[]:");
4723 for (i = 0; i < MAX_NR_ZONES; i++)
4724 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4725 printk(KERN_CONT "\n");
4728 for_each_populated_zone(zone) {
4730 unsigned long nr[MAX_ORDER], flags, total = 0;
4731 unsigned char types[MAX_ORDER];
4733 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4736 printk(KERN_CONT "%s: ", zone->name);
4738 spin_lock_irqsave(&zone->lock, flags);
4739 for (order = 0; order < MAX_ORDER; order++) {
4740 struct free_area *area = &zone->free_area[order];
4743 nr[order] = area->nr_free;
4744 total += nr[order] << order;
4747 for (type = 0; type < MIGRATE_TYPES; type++) {
4748 if (!list_empty(&area->free_list[type]))
4749 types[order] |= 1 << type;
4752 spin_unlock_irqrestore(&zone->lock, flags);
4753 for (order = 0; order < MAX_ORDER; order++) {
4754 printk(KERN_CONT "%lu*%lukB ",
4755 nr[order], K(1UL) << order);
4757 show_migration_types(types[order]);
4759 printk(KERN_CONT "= %lukB\n", K(total));
4762 hugetlb_show_meminfo();
4764 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4766 show_swap_cache_info();
4769 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4771 zoneref->zone = zone;
4772 zoneref->zone_idx = zone_idx(zone);
4776 * Builds allocation fallback zone lists.
4778 * Add all populated zones of a node to the zonelist.
4780 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4784 enum zone_type zone_type = MAX_NR_ZONES;
4788 zone = pgdat->node_zones + zone_type;
4789 if (managed_zone(zone)) {
4790 zoneref_set_zone(zone,
4791 &zonelist->_zonerefs[nr_zones++]);
4792 check_highest_zone(zone_type);
4794 } while (zone_type);
4802 * 0 = automatic detection of better ordering.
4803 * 1 = order by ([node] distance, -zonetype)
4804 * 2 = order by (-zonetype, [node] distance)
4806 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4807 * the same zonelist. So only NUMA can configure this param.
4809 #define ZONELIST_ORDER_DEFAULT 0
4810 #define ZONELIST_ORDER_NODE 1
4811 #define ZONELIST_ORDER_ZONE 2
4813 /* zonelist order in the kernel.
4814 * set_zonelist_order() will set this to NODE or ZONE.
4816 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4817 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4821 /* The value user specified ....changed by config */
4822 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4823 /* string for sysctl */
4824 #define NUMA_ZONELIST_ORDER_LEN 16
4825 char numa_zonelist_order[16] = "default";
4828 * interface for configure zonelist ordering.
4829 * command line option "numa_zonelist_order"
4830 * = "[dD]efault - default, automatic configuration.
4831 * = "[nN]ode - order by node locality, then by zone within node
4832 * = "[zZ]one - order by zone, then by locality within zone
4835 static int __parse_numa_zonelist_order(char *s)
4837 if (*s == 'd' || *s == 'D') {
4838 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4839 } else if (*s == 'n' || *s == 'N') {
4840 user_zonelist_order = ZONELIST_ORDER_NODE;
4841 } else if (*s == 'z' || *s == 'Z') {
4842 user_zonelist_order = ZONELIST_ORDER_ZONE;
4844 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4850 static __init int setup_numa_zonelist_order(char *s)
4857 ret = __parse_numa_zonelist_order(s);
4859 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4863 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4866 * sysctl handler for numa_zonelist_order
4868 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4869 void __user *buffer, size_t *length,
4872 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4874 static DEFINE_MUTEX(zl_order_mutex);
4876 mutex_lock(&zl_order_mutex);
4878 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4882 strcpy(saved_string, (char *)table->data);
4884 ret = proc_dostring(table, write, buffer, length, ppos);
4888 int oldval = user_zonelist_order;
4890 ret = __parse_numa_zonelist_order((char *)table->data);
4893 * bogus value. restore saved string
4895 strncpy((char *)table->data, saved_string,
4896 NUMA_ZONELIST_ORDER_LEN);
4897 user_zonelist_order = oldval;
4898 } else if (oldval != user_zonelist_order) {
4899 mem_hotplug_begin();
4900 mutex_lock(&zonelists_mutex);
4901 build_all_zonelists(NULL, NULL);
4902 mutex_unlock(&zonelists_mutex);
4907 mutex_unlock(&zl_order_mutex);
4912 #define MAX_NODE_LOAD (nr_online_nodes)
4913 static int node_load[MAX_NUMNODES];
4916 * find_next_best_node - find the next node that should appear in a given node's fallback list
4917 * @node: node whose fallback list we're appending
4918 * @used_node_mask: nodemask_t of already used nodes
4920 * We use a number of factors to determine which is the next node that should
4921 * appear on a given node's fallback list. The node should not have appeared
4922 * already in @node's fallback list, and it should be the next closest node
4923 * according to the distance array (which contains arbitrary distance values
4924 * from each node to each node in the system), and should also prefer nodes
4925 * with no CPUs, since presumably they'll have very little allocation pressure
4926 * on them otherwise.
4927 * It returns -1 if no node is found.
4929 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4932 int min_val = INT_MAX;
4933 int best_node = NUMA_NO_NODE;
4934 const struct cpumask *tmp = cpumask_of_node(0);
4936 /* Use the local node if we haven't already */
4937 if (!node_isset(node, *used_node_mask)) {
4938 node_set(node, *used_node_mask);
4942 for_each_node_state(n, N_MEMORY) {
4944 /* Don't want a node to appear more than once */
4945 if (node_isset(n, *used_node_mask))
4948 /* Use the distance array to find the distance */
4949 val = node_distance(node, n);
4951 /* Penalize nodes under us ("prefer the next node") */
4954 /* Give preference to headless and unused nodes */
4955 tmp = cpumask_of_node(n);
4956 if (!cpumask_empty(tmp))
4957 val += PENALTY_FOR_NODE_WITH_CPUS;
4959 /* Slight preference for less loaded node */
4960 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4961 val += node_load[n];
4963 if (val < min_val) {
4970 node_set(best_node, *used_node_mask);
4977 * Build zonelists ordered by node and zones within node.
4978 * This results in maximum locality--normal zone overflows into local
4979 * DMA zone, if any--but risks exhausting DMA zone.
4981 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4984 struct zonelist *zonelist;
4986 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4987 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4989 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4990 zonelist->_zonerefs[j].zone = NULL;
4991 zonelist->_zonerefs[j].zone_idx = 0;
4995 * Build gfp_thisnode zonelists
4997 static void build_thisnode_zonelists(pg_data_t *pgdat)
5000 struct zonelist *zonelist;
5002 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
5003 j = build_zonelists_node(pgdat, zonelist, 0);
5004 zonelist->_zonerefs[j].zone = NULL;
5005 zonelist->_zonerefs[j].zone_idx = 0;
5009 * Build zonelists ordered by zone and nodes within zones.
5010 * This results in conserving DMA zone[s] until all Normal memory is
5011 * exhausted, but results in overflowing to remote node while memory
5012 * may still exist in local DMA zone.
5014 static int node_order[MAX_NUMNODES];
5016 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
5019 int zone_type; /* needs to be signed */
5021 struct zonelist *zonelist;
5023 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5025 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
5026 for (j = 0; j < nr_nodes; j++) {
5027 node = node_order[j];
5028 z = &NODE_DATA(node)->node_zones[zone_type];
5029 if (managed_zone(z)) {
5031 &zonelist->_zonerefs[pos++]);
5032 check_highest_zone(zone_type);
5036 zonelist->_zonerefs[pos].zone = NULL;
5037 zonelist->_zonerefs[pos].zone_idx = 0;
5040 #if defined(CONFIG_64BIT)
5042 * Devices that require DMA32/DMA are relatively rare and do not justify a
5043 * penalty to every machine in case the specialised case applies. Default
5044 * to Node-ordering on 64-bit NUMA machines
5046 static int default_zonelist_order(void)
5048 return ZONELIST_ORDER_NODE;
5052 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
5053 * by the kernel. If processes running on node 0 deplete the low memory zone
5054 * then reclaim will occur more frequency increasing stalls and potentially
5055 * be easier to OOM if a large percentage of the zone is under writeback or
5056 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
5057 * Hence, default to zone ordering on 32-bit.
5059 static int default_zonelist_order(void)
5061 return ZONELIST_ORDER_ZONE;
5063 #endif /* CONFIG_64BIT */
5065 static void set_zonelist_order(void)
5067 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
5068 current_zonelist_order = default_zonelist_order();
5070 current_zonelist_order = user_zonelist_order;
5073 static void build_zonelists(pg_data_t *pgdat)
5076 nodemask_t used_mask;
5077 int local_node, prev_node;
5078 struct zonelist *zonelist;
5079 unsigned int order = current_zonelist_order;
5081 /* initialize zonelists */
5082 for (i = 0; i < MAX_ZONELISTS; i++) {
5083 zonelist = pgdat->node_zonelists + i;
5084 zonelist->_zonerefs[0].zone = NULL;
5085 zonelist->_zonerefs[0].zone_idx = 0;
5088 /* NUMA-aware ordering of nodes */
5089 local_node = pgdat->node_id;
5090 load = nr_online_nodes;
5091 prev_node = local_node;
5092 nodes_clear(used_mask);
5094 memset(node_order, 0, sizeof(node_order));
5097 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5099 * We don't want to pressure a particular node.
5100 * So adding penalty to the first node in same
5101 * distance group to make it round-robin.
5103 if (node_distance(local_node, node) !=
5104 node_distance(local_node, prev_node))
5105 node_load[node] = load;
5109 if (order == ZONELIST_ORDER_NODE)
5110 build_zonelists_in_node_order(pgdat, node);
5112 node_order[i++] = node; /* remember order */
5115 if (order == ZONELIST_ORDER_ZONE) {
5116 /* calculate node order -- i.e., DMA last! */
5117 build_zonelists_in_zone_order(pgdat, i);
5120 build_thisnode_zonelists(pgdat);
5123 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5125 * Return node id of node used for "local" allocations.
5126 * I.e., first node id of first zone in arg node's generic zonelist.
5127 * Used for initializing percpu 'numa_mem', which is used primarily
5128 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5130 int local_memory_node(int node)
5134 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5135 gfp_zone(GFP_KERNEL),
5137 return z->zone->node;
5141 static void setup_min_unmapped_ratio(void);
5142 static void setup_min_slab_ratio(void);
5143 #else /* CONFIG_NUMA */
5145 static void set_zonelist_order(void)
5147 current_zonelist_order = ZONELIST_ORDER_ZONE;
5150 static void build_zonelists(pg_data_t *pgdat)
5152 int node, local_node;
5154 struct zonelist *zonelist;
5156 local_node = pgdat->node_id;
5158 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5159 j = build_zonelists_node(pgdat, zonelist, 0);
5162 * Now we build the zonelist so that it contains the zones
5163 * of all the other nodes.
5164 * We don't want to pressure a particular node, so when
5165 * building the zones for node N, we make sure that the
5166 * zones coming right after the local ones are those from
5167 * node N+1 (modulo N)
5169 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5170 if (!node_online(node))
5172 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5174 for (node = 0; node < local_node; node++) {
5175 if (!node_online(node))
5177 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5180 zonelist->_zonerefs[j].zone = NULL;
5181 zonelist->_zonerefs[j].zone_idx = 0;
5184 #endif /* CONFIG_NUMA */
5187 * Boot pageset table. One per cpu which is going to be used for all
5188 * zones and all nodes. The parameters will be set in such a way
5189 * that an item put on a list will immediately be handed over to
5190 * the buddy list. This is safe since pageset manipulation is done
5191 * with interrupts disabled.
5193 * The boot_pagesets must be kept even after bootup is complete for
5194 * unused processors and/or zones. They do play a role for bootstrapping
5195 * hotplugged processors.
5197 * zoneinfo_show() and maybe other functions do
5198 * not check if the processor is online before following the pageset pointer.
5199 * Other parts of the kernel may not check if the zone is available.
5201 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5202 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5203 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5204 static void setup_zone_pageset(struct zone *zone);
5207 * Global mutex to protect against size modification of zonelists
5208 * as well as to serialize pageset setup for the new populated zone.
5210 DEFINE_MUTEX(zonelists_mutex);
5212 /* return values int ....just for stop_machine() */
5213 static int __build_all_zonelists(void *data)
5217 pg_data_t *self = data;
5220 memset(node_load, 0, sizeof(node_load));
5223 if (self && !node_online(self->node_id)) {
5224 build_zonelists(self);
5227 for_each_online_node(nid) {
5228 pg_data_t *pgdat = NODE_DATA(nid);
5230 build_zonelists(pgdat);
5234 * Initialize the boot_pagesets that are going to be used
5235 * for bootstrapping processors. The real pagesets for
5236 * each zone will be allocated later when the per cpu
5237 * allocator is available.
5239 * boot_pagesets are used also for bootstrapping offline
5240 * cpus if the system is already booted because the pagesets
5241 * are needed to initialize allocators on a specific cpu too.
5242 * F.e. the percpu allocator needs the page allocator which
5243 * needs the percpu allocator in order to allocate its pagesets
5244 * (a chicken-egg dilemma).
5246 for_each_possible_cpu(cpu) {
5247 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5249 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5251 * We now know the "local memory node" for each node--
5252 * i.e., the node of the first zone in the generic zonelist.
5253 * Set up numa_mem percpu variable for on-line cpus. During
5254 * boot, only the boot cpu should be on-line; we'll init the
5255 * secondary cpus' numa_mem as they come on-line. During
5256 * node/memory hotplug, we'll fixup all on-line cpus.
5258 if (cpu_online(cpu))
5259 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5266 static noinline void __init
5267 build_all_zonelists_init(void)
5269 __build_all_zonelists(NULL);
5270 mminit_verify_zonelist();
5271 cpuset_init_current_mems_allowed();
5275 * Called with zonelists_mutex held always
5276 * unless system_state == SYSTEM_BOOTING.
5278 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5279 * [we're only called with non-NULL zone through __meminit paths] and
5280 * (2) call of __init annotated helper build_all_zonelists_init
5281 * [protected by SYSTEM_BOOTING].
5283 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5285 set_zonelist_order();
5287 if (system_state == SYSTEM_BOOTING) {
5288 build_all_zonelists_init();
5290 #ifdef CONFIG_MEMORY_HOTPLUG
5292 setup_zone_pageset(zone);
5294 /* we have to stop all cpus to guarantee there is no user
5296 stop_machine_cpuslocked(__build_all_zonelists, pgdat, NULL);
5297 /* cpuset refresh routine should be here */
5299 vm_total_pages = nr_free_pagecache_pages();
5301 * Disable grouping by mobility if the number of pages in the
5302 * system is too low to allow the mechanism to work. It would be
5303 * more accurate, but expensive to check per-zone. This check is
5304 * made on memory-hotadd so a system can start with mobility
5305 * disabled and enable it later
5307 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5308 page_group_by_mobility_disabled = 1;
5310 page_group_by_mobility_disabled = 0;
5312 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5314 zonelist_order_name[current_zonelist_order],
5315 page_group_by_mobility_disabled ? "off" : "on",
5318 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5323 * Initially all pages are reserved - free ones are freed
5324 * up by free_all_bootmem() once the early boot process is
5325 * done. Non-atomic initialization, single-pass.
5327 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5328 unsigned long start_pfn, enum memmap_context context)
5330 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5331 unsigned long end_pfn = start_pfn + size;
5332 pg_data_t *pgdat = NODE_DATA(nid);
5334 unsigned long nr_initialised = 0;
5335 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5336 struct memblock_region *r = NULL, *tmp;
5339 if (highest_memmap_pfn < end_pfn - 1)
5340 highest_memmap_pfn = end_pfn - 1;
5343 * Honor reservation requested by the driver for this ZONE_DEVICE
5346 if (altmap && start_pfn == altmap->base_pfn)
5347 start_pfn += altmap->reserve;
5349 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5351 * There can be holes in boot-time mem_map[]s handed to this
5352 * function. They do not exist on hotplugged memory.
5354 if (context != MEMMAP_EARLY)
5357 if (!early_pfn_valid(pfn)) {
5358 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5360 * Skip to the pfn preceding the next valid one (or
5361 * end_pfn), such that we hit a valid pfn (or end_pfn)
5362 * on our next iteration of the loop.
5364 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5368 if (!early_pfn_in_nid(pfn, nid))
5370 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5373 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5375 * Check given memblock attribute by firmware which can affect
5376 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5377 * mirrored, it's an overlapped memmap init. skip it.
5379 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5380 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5381 for_each_memblock(memory, tmp)
5382 if (pfn < memblock_region_memory_end_pfn(tmp))
5386 if (pfn >= memblock_region_memory_base_pfn(r) &&
5387 memblock_is_mirror(r)) {
5388 /* already initialized as NORMAL */
5389 pfn = memblock_region_memory_end_pfn(r);
5397 * Mark the block movable so that blocks are reserved for
5398 * movable at startup. This will force kernel allocations
5399 * to reserve their blocks rather than leaking throughout
5400 * the address space during boot when many long-lived
5401 * kernel allocations are made.
5403 * bitmap is created for zone's valid pfn range. but memmap
5404 * can be created for invalid pages (for alignment)
5405 * check here not to call set_pageblock_migratetype() against
5408 if (!(pfn & (pageblock_nr_pages - 1))) {
5409 struct page *page = pfn_to_page(pfn);
5411 __init_single_page(page, pfn, zone, nid);
5412 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5414 __init_single_pfn(pfn, zone, nid);
5419 static void __meminit zone_init_free_lists(struct zone *zone)
5421 unsigned int order, t;
5422 for_each_migratetype_order(order, t) {
5423 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5424 zone->free_area[order].nr_free = 0;
5428 #ifndef __HAVE_ARCH_MEMMAP_INIT
5429 #define memmap_init(size, nid, zone, start_pfn) \
5430 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5433 static int zone_batchsize(struct zone *zone)
5439 * The per-cpu-pages pools are set to around 1000th of the
5440 * size of the zone. But no more than 1/2 of a meg.
5442 * OK, so we don't know how big the cache is. So guess.
5444 batch = zone->managed_pages / 1024;
5445 if (batch * PAGE_SIZE > 512 * 1024)
5446 batch = (512 * 1024) / PAGE_SIZE;
5447 batch /= 4; /* We effectively *= 4 below */
5452 * Clamp the batch to a 2^n - 1 value. Having a power
5453 * of 2 value was found to be more likely to have
5454 * suboptimal cache aliasing properties in some cases.
5456 * For example if 2 tasks are alternately allocating
5457 * batches of pages, one task can end up with a lot
5458 * of pages of one half of the possible page colors
5459 * and the other with pages of the other colors.
5461 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5466 /* The deferral and batching of frees should be suppressed under NOMMU
5469 * The problem is that NOMMU needs to be able to allocate large chunks
5470 * of contiguous memory as there's no hardware page translation to
5471 * assemble apparent contiguous memory from discontiguous pages.
5473 * Queueing large contiguous runs of pages for batching, however,
5474 * causes the pages to actually be freed in smaller chunks. As there
5475 * can be a significant delay between the individual batches being
5476 * recycled, this leads to the once large chunks of space being
5477 * fragmented and becoming unavailable for high-order allocations.
5484 * pcp->high and pcp->batch values are related and dependent on one another:
5485 * ->batch must never be higher then ->high.
5486 * The following function updates them in a safe manner without read side
5489 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5490 * those fields changing asynchronously (acording the the above rule).
5492 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5493 * outside of boot time (or some other assurance that no concurrent updaters
5496 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5497 unsigned long batch)
5499 /* start with a fail safe value for batch */
5503 /* Update high, then batch, in order */
5510 /* a companion to pageset_set_high() */
5511 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5513 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5516 static void pageset_init(struct per_cpu_pageset *p)
5518 struct per_cpu_pages *pcp;
5521 memset(p, 0, sizeof(*p));
5525 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5526 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5529 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5532 pageset_set_batch(p, batch);
5536 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5537 * to the value high for the pageset p.
5539 static void pageset_set_high(struct per_cpu_pageset *p,
5542 unsigned long batch = max(1UL, high / 4);
5543 if ((high / 4) > (PAGE_SHIFT * 8))
5544 batch = PAGE_SHIFT * 8;
5546 pageset_update(&p->pcp, high, batch);
5549 static void pageset_set_high_and_batch(struct zone *zone,
5550 struct per_cpu_pageset *pcp)
5552 if (percpu_pagelist_fraction)
5553 pageset_set_high(pcp,
5554 (zone->managed_pages /
5555 percpu_pagelist_fraction));
5557 pageset_set_batch(pcp, zone_batchsize(zone));
5560 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5562 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5565 pageset_set_high_and_batch(zone, pcp);
5568 static void __meminit setup_zone_pageset(struct zone *zone)
5571 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5572 for_each_possible_cpu(cpu)
5573 zone_pageset_init(zone, cpu);
5577 * Allocate per cpu pagesets and initialize them.
5578 * Before this call only boot pagesets were available.
5580 void __init setup_per_cpu_pageset(void)
5582 struct pglist_data *pgdat;
5585 for_each_populated_zone(zone)
5586 setup_zone_pageset(zone);
5588 for_each_online_pgdat(pgdat)
5589 pgdat->per_cpu_nodestats =
5590 alloc_percpu(struct per_cpu_nodestat);
5593 static __meminit void zone_pcp_init(struct zone *zone)
5596 * per cpu subsystem is not up at this point. The following code
5597 * relies on the ability of the linker to provide the
5598 * offset of a (static) per cpu variable into the per cpu area.
5600 zone->pageset = &boot_pageset;
5602 if (populated_zone(zone))
5603 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5604 zone->name, zone->present_pages,
5605 zone_batchsize(zone));
5608 void __meminit init_currently_empty_zone(struct zone *zone,
5609 unsigned long zone_start_pfn,
5612 struct pglist_data *pgdat = zone->zone_pgdat;
5614 pgdat->nr_zones = zone_idx(zone) + 1;
5616 zone->zone_start_pfn = zone_start_pfn;
5618 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5619 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5621 (unsigned long)zone_idx(zone),
5622 zone_start_pfn, (zone_start_pfn + size));
5624 zone_init_free_lists(zone);
5625 zone->initialized = 1;
5628 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5629 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5632 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5634 int __meminit __early_pfn_to_nid(unsigned long pfn,
5635 struct mminit_pfnnid_cache *state)
5637 unsigned long start_pfn, end_pfn;
5640 if (state->last_start <= pfn && pfn < state->last_end)
5641 return state->last_nid;
5643 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5645 state->last_start = start_pfn;
5646 state->last_end = end_pfn;
5647 state->last_nid = nid;
5652 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5655 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5656 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5657 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5659 * If an architecture guarantees that all ranges registered contain no holes
5660 * and may be freed, this this function may be used instead of calling
5661 * memblock_free_early_nid() manually.
5663 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5665 unsigned long start_pfn, end_pfn;
5668 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5669 start_pfn = min(start_pfn, max_low_pfn);
5670 end_pfn = min(end_pfn, max_low_pfn);
5672 if (start_pfn < end_pfn)
5673 memblock_free_early_nid(PFN_PHYS(start_pfn),
5674 (end_pfn - start_pfn) << PAGE_SHIFT,
5680 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5681 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5683 * If an architecture guarantees that all ranges registered contain no holes and may
5684 * be freed, this function may be used instead of calling memory_present() manually.
5686 void __init sparse_memory_present_with_active_regions(int nid)
5688 unsigned long start_pfn, end_pfn;
5691 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5692 memory_present(this_nid, start_pfn, end_pfn);
5696 * get_pfn_range_for_nid - Return the start and end page frames for a node
5697 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5698 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5699 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5701 * It returns the start and end page frame of a node based on information
5702 * provided by memblock_set_node(). If called for a node
5703 * with no available memory, a warning is printed and the start and end
5706 void __meminit get_pfn_range_for_nid(unsigned int nid,
5707 unsigned long *start_pfn, unsigned long *end_pfn)
5709 unsigned long this_start_pfn, this_end_pfn;
5715 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5716 *start_pfn = min(*start_pfn, this_start_pfn);
5717 *end_pfn = max(*end_pfn, this_end_pfn);
5720 if (*start_pfn == -1UL)
5725 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5726 * assumption is made that zones within a node are ordered in monotonic
5727 * increasing memory addresses so that the "highest" populated zone is used
5729 static void __init find_usable_zone_for_movable(void)
5732 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5733 if (zone_index == ZONE_MOVABLE)
5736 if (arch_zone_highest_possible_pfn[zone_index] >
5737 arch_zone_lowest_possible_pfn[zone_index])
5741 VM_BUG_ON(zone_index == -1);
5742 movable_zone = zone_index;
5746 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5747 * because it is sized independent of architecture. Unlike the other zones,
5748 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5749 * in each node depending on the size of each node and how evenly kernelcore
5750 * is distributed. This helper function adjusts the zone ranges
5751 * provided by the architecture for a given node by using the end of the
5752 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5753 * zones within a node are in order of monotonic increases memory addresses
5755 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5756 unsigned long zone_type,
5757 unsigned long node_start_pfn,
5758 unsigned long node_end_pfn,
5759 unsigned long *zone_start_pfn,
5760 unsigned long *zone_end_pfn)
5762 /* Only adjust if ZONE_MOVABLE is on this node */
5763 if (zone_movable_pfn[nid]) {
5764 /* Size ZONE_MOVABLE */
5765 if (zone_type == ZONE_MOVABLE) {
5766 *zone_start_pfn = zone_movable_pfn[nid];
5767 *zone_end_pfn = min(node_end_pfn,
5768 arch_zone_highest_possible_pfn[movable_zone]);
5770 /* Adjust for ZONE_MOVABLE starting within this range */
5771 } else if (!mirrored_kernelcore &&
5772 *zone_start_pfn < zone_movable_pfn[nid] &&
5773 *zone_end_pfn > zone_movable_pfn[nid]) {
5774 *zone_end_pfn = zone_movable_pfn[nid];
5776 /* Check if this whole range is within ZONE_MOVABLE */
5777 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5778 *zone_start_pfn = *zone_end_pfn;
5783 * Return the number of pages a zone spans in a node, including holes
5784 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5786 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5787 unsigned long zone_type,
5788 unsigned long node_start_pfn,
5789 unsigned long node_end_pfn,
5790 unsigned long *zone_start_pfn,
5791 unsigned long *zone_end_pfn,
5792 unsigned long *ignored)
5794 /* When hotadd a new node from cpu_up(), the node should be empty */
5795 if (!node_start_pfn && !node_end_pfn)
5798 /* Get the start and end of the zone */
5799 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5800 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5801 adjust_zone_range_for_zone_movable(nid, zone_type,
5802 node_start_pfn, node_end_pfn,
5803 zone_start_pfn, zone_end_pfn);
5805 /* Check that this node has pages within the zone's required range */
5806 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5809 /* Move the zone boundaries inside the node if necessary */
5810 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5811 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5813 /* Return the spanned pages */
5814 return *zone_end_pfn - *zone_start_pfn;
5818 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5819 * then all holes in the requested range will be accounted for.
5821 unsigned long __meminit __absent_pages_in_range(int nid,
5822 unsigned long range_start_pfn,
5823 unsigned long range_end_pfn)
5825 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5826 unsigned long start_pfn, end_pfn;
5829 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5830 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5831 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5832 nr_absent -= end_pfn - start_pfn;
5838 * absent_pages_in_range - Return number of page frames in holes within a range
5839 * @start_pfn: The start PFN to start searching for holes
5840 * @end_pfn: The end PFN to stop searching for holes
5842 * It returns the number of pages frames in memory holes within a range.
5844 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5845 unsigned long end_pfn)
5847 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5850 /* Return the number of page frames in holes in a zone on a node */
5851 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5852 unsigned long zone_type,
5853 unsigned long node_start_pfn,
5854 unsigned long node_end_pfn,
5855 unsigned long *ignored)
5857 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5858 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5859 unsigned long zone_start_pfn, zone_end_pfn;
5860 unsigned long nr_absent;
5862 /* When hotadd a new node from cpu_up(), the node should be empty */
5863 if (!node_start_pfn && !node_end_pfn)
5866 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5867 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5869 adjust_zone_range_for_zone_movable(nid, zone_type,
5870 node_start_pfn, node_end_pfn,
5871 &zone_start_pfn, &zone_end_pfn);
5872 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5875 * ZONE_MOVABLE handling.
5876 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5879 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5880 unsigned long start_pfn, end_pfn;
5881 struct memblock_region *r;
5883 for_each_memblock(memory, r) {
5884 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5885 zone_start_pfn, zone_end_pfn);
5886 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5887 zone_start_pfn, zone_end_pfn);
5889 if (zone_type == ZONE_MOVABLE &&
5890 memblock_is_mirror(r))
5891 nr_absent += end_pfn - start_pfn;
5893 if (zone_type == ZONE_NORMAL &&
5894 !memblock_is_mirror(r))
5895 nr_absent += end_pfn - start_pfn;
5902 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5903 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5904 unsigned long zone_type,
5905 unsigned long node_start_pfn,
5906 unsigned long node_end_pfn,
5907 unsigned long *zone_start_pfn,
5908 unsigned long *zone_end_pfn,
5909 unsigned long *zones_size)
5913 *zone_start_pfn = node_start_pfn;
5914 for (zone = 0; zone < zone_type; zone++)
5915 *zone_start_pfn += zones_size[zone];
5917 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5919 return zones_size[zone_type];
5922 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5923 unsigned long zone_type,
5924 unsigned long node_start_pfn,
5925 unsigned long node_end_pfn,
5926 unsigned long *zholes_size)
5931 return zholes_size[zone_type];
5934 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5936 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5937 unsigned long node_start_pfn,
5938 unsigned long node_end_pfn,
5939 unsigned long *zones_size,
5940 unsigned long *zholes_size)
5942 unsigned long realtotalpages = 0, totalpages = 0;
5945 for (i = 0; i < MAX_NR_ZONES; i++) {
5946 struct zone *zone = pgdat->node_zones + i;
5947 unsigned long zone_start_pfn, zone_end_pfn;
5948 unsigned long size, real_size;
5950 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5956 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5957 node_start_pfn, node_end_pfn,
5960 zone->zone_start_pfn = zone_start_pfn;
5962 zone->zone_start_pfn = 0;
5963 zone->spanned_pages = size;
5964 zone->present_pages = real_size;
5967 realtotalpages += real_size;
5970 pgdat->node_spanned_pages = totalpages;
5971 pgdat->node_present_pages = realtotalpages;
5972 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5976 #ifndef CONFIG_SPARSEMEM
5978 * Calculate the size of the zone->blockflags rounded to an unsigned long
5979 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5980 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5981 * round what is now in bits to nearest long in bits, then return it in
5984 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5986 unsigned long usemapsize;
5988 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5989 usemapsize = roundup(zonesize, pageblock_nr_pages);
5990 usemapsize = usemapsize >> pageblock_order;
5991 usemapsize *= NR_PAGEBLOCK_BITS;
5992 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5994 return usemapsize / 8;
5997 static void __init setup_usemap(struct pglist_data *pgdat,
5999 unsigned long zone_start_pfn,
6000 unsigned long zonesize)
6002 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6003 zone->pageblock_flags = NULL;
6005 zone->pageblock_flags =
6006 memblock_virt_alloc_node_nopanic(usemapsize,
6010 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6011 unsigned long zone_start_pfn, unsigned long zonesize) {}
6012 #endif /* CONFIG_SPARSEMEM */
6014 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6016 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6017 void __paginginit set_pageblock_order(void)
6021 /* Check that pageblock_nr_pages has not already been setup */
6022 if (pageblock_order)
6025 if (HPAGE_SHIFT > PAGE_SHIFT)
6026 order = HUGETLB_PAGE_ORDER;
6028 order = MAX_ORDER - 1;
6031 * Assume the largest contiguous order of interest is a huge page.
6032 * This value may be variable depending on boot parameters on IA64 and
6035 pageblock_order = order;
6037 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6040 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6041 * is unused as pageblock_order is set at compile-time. See
6042 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6045 void __paginginit set_pageblock_order(void)
6049 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6051 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6052 unsigned long present_pages)
6054 unsigned long pages = spanned_pages;
6057 * Provide a more accurate estimation if there are holes within
6058 * the zone and SPARSEMEM is in use. If there are holes within the
6059 * zone, each populated memory region may cost us one or two extra
6060 * memmap pages due to alignment because memmap pages for each
6061 * populated regions may not be naturally aligned on page boundary.
6062 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6064 if (spanned_pages > present_pages + (present_pages >> 4) &&
6065 IS_ENABLED(CONFIG_SPARSEMEM))
6066 pages = present_pages;
6068 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6072 * Set up the zone data structures:
6073 * - mark all pages reserved
6074 * - mark all memory queues empty
6075 * - clear the memory bitmaps
6077 * NOTE: pgdat should get zeroed by caller.
6079 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6082 int nid = pgdat->node_id;
6084 pgdat_resize_init(pgdat);
6085 #ifdef CONFIG_NUMA_BALANCING
6086 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6087 pgdat->numabalancing_migrate_nr_pages = 0;
6088 pgdat->numabalancing_migrate_next_window = jiffies;
6090 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6091 spin_lock_init(&pgdat->split_queue_lock);
6092 INIT_LIST_HEAD(&pgdat->split_queue);
6093 pgdat->split_queue_len = 0;
6095 init_waitqueue_head(&pgdat->kswapd_wait);
6096 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6097 #ifdef CONFIG_COMPACTION
6098 init_waitqueue_head(&pgdat->kcompactd_wait);
6100 pgdat_page_ext_init(pgdat);
6101 spin_lock_init(&pgdat->lru_lock);
6102 lruvec_init(node_lruvec(pgdat));
6104 pgdat->per_cpu_nodestats = &boot_nodestats;
6106 for (j = 0; j < MAX_NR_ZONES; j++) {
6107 struct zone *zone = pgdat->node_zones + j;
6108 unsigned long size, realsize, freesize, memmap_pages;
6109 unsigned long zone_start_pfn = zone->zone_start_pfn;
6111 size = zone->spanned_pages;
6112 realsize = freesize = zone->present_pages;
6115 * Adjust freesize so that it accounts for how much memory
6116 * is used by this zone for memmap. This affects the watermark
6117 * and per-cpu initialisations
6119 memmap_pages = calc_memmap_size(size, realsize);
6120 if (!is_highmem_idx(j)) {
6121 if (freesize >= memmap_pages) {
6122 freesize -= memmap_pages;
6125 " %s zone: %lu pages used for memmap\n",
6126 zone_names[j], memmap_pages);
6128 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6129 zone_names[j], memmap_pages, freesize);
6132 /* Account for reserved pages */
6133 if (j == 0 && freesize > dma_reserve) {
6134 freesize -= dma_reserve;
6135 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6136 zone_names[0], dma_reserve);
6139 if (!is_highmem_idx(j))
6140 nr_kernel_pages += freesize;
6141 /* Charge for highmem memmap if there are enough kernel pages */
6142 else if (nr_kernel_pages > memmap_pages * 2)
6143 nr_kernel_pages -= memmap_pages;
6144 nr_all_pages += freesize;
6147 * Set an approximate value for lowmem here, it will be adjusted
6148 * when the bootmem allocator frees pages into the buddy system.
6149 * And all highmem pages will be managed by the buddy system.
6151 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6155 zone->name = zone_names[j];
6156 zone->zone_pgdat = pgdat;
6157 spin_lock_init(&zone->lock);
6158 zone_seqlock_init(zone);
6159 zone_pcp_init(zone);
6164 set_pageblock_order();
6165 setup_usemap(pgdat, zone, zone_start_pfn, size);
6166 init_currently_empty_zone(zone, zone_start_pfn, size);
6167 memmap_init(size, nid, j, zone_start_pfn);
6171 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6173 unsigned long __maybe_unused start = 0;
6174 unsigned long __maybe_unused offset = 0;
6176 /* Skip empty nodes */
6177 if (!pgdat->node_spanned_pages)
6180 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6181 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6182 offset = pgdat->node_start_pfn - start;
6183 /* ia64 gets its own node_mem_map, before this, without bootmem */
6184 if (!pgdat->node_mem_map) {
6185 unsigned long size, end;
6189 * The zone's endpoints aren't required to be MAX_ORDER
6190 * aligned but the node_mem_map endpoints must be in order
6191 * for the buddy allocator to function correctly.
6193 end = pgdat_end_pfn(pgdat);
6194 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6195 size = (end - start) * sizeof(struct page);
6196 map = alloc_remap(pgdat->node_id, size);
6198 map = memblock_virt_alloc_node_nopanic(size,
6200 pgdat->node_mem_map = map + offset;
6202 #ifndef CONFIG_NEED_MULTIPLE_NODES
6204 * With no DISCONTIG, the global mem_map is just set as node 0's
6206 if (pgdat == NODE_DATA(0)) {
6207 mem_map = NODE_DATA(0)->node_mem_map;
6208 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6209 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6211 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6214 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6217 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6218 unsigned long node_start_pfn, unsigned long *zholes_size)
6220 pg_data_t *pgdat = NODE_DATA(nid);
6221 unsigned long start_pfn = 0;
6222 unsigned long end_pfn = 0;
6224 /* pg_data_t should be reset to zero when it's allocated */
6225 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6227 pgdat->node_id = nid;
6228 pgdat->node_start_pfn = node_start_pfn;
6229 pgdat->per_cpu_nodestats = NULL;
6230 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6231 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6232 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6233 (u64)start_pfn << PAGE_SHIFT,
6234 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6236 start_pfn = node_start_pfn;
6238 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6239 zones_size, zholes_size);
6241 alloc_node_mem_map(pgdat);
6242 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6243 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6244 nid, (unsigned long)pgdat,
6245 (unsigned long)pgdat->node_mem_map);
6248 reset_deferred_meminit(pgdat);
6249 free_area_init_core(pgdat);
6252 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6254 #if MAX_NUMNODES > 1
6256 * Figure out the number of possible node ids.
6258 void __init setup_nr_node_ids(void)
6260 unsigned int highest;
6262 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6263 nr_node_ids = highest + 1;
6268 * node_map_pfn_alignment - determine the maximum internode alignment
6270 * This function should be called after node map is populated and sorted.
6271 * It calculates the maximum power of two alignment which can distinguish
6274 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6275 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6276 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6277 * shifted, 1GiB is enough and this function will indicate so.
6279 * This is used to test whether pfn -> nid mapping of the chosen memory
6280 * model has fine enough granularity to avoid incorrect mapping for the
6281 * populated node map.
6283 * Returns the determined alignment in pfn's. 0 if there is no alignment
6284 * requirement (single node).
6286 unsigned long __init node_map_pfn_alignment(void)
6288 unsigned long accl_mask = 0, last_end = 0;
6289 unsigned long start, end, mask;
6293 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6294 if (!start || last_nid < 0 || last_nid == nid) {
6301 * Start with a mask granular enough to pin-point to the
6302 * start pfn and tick off bits one-by-one until it becomes
6303 * too coarse to separate the current node from the last.
6305 mask = ~((1 << __ffs(start)) - 1);
6306 while (mask && last_end <= (start & (mask << 1)))
6309 /* accumulate all internode masks */
6313 /* convert mask to number of pages */
6314 return ~accl_mask + 1;
6317 /* Find the lowest pfn for a node */
6318 static unsigned long __init find_min_pfn_for_node(int nid)
6320 unsigned long min_pfn = ULONG_MAX;
6321 unsigned long start_pfn;
6324 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6325 min_pfn = min(min_pfn, start_pfn);
6327 if (min_pfn == ULONG_MAX) {
6328 pr_warn("Could not find start_pfn for node %d\n", nid);
6336 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6338 * It returns the minimum PFN based on information provided via
6339 * memblock_set_node().
6341 unsigned long __init find_min_pfn_with_active_regions(void)
6343 return find_min_pfn_for_node(MAX_NUMNODES);
6347 * early_calculate_totalpages()
6348 * Sum pages in active regions for movable zone.
6349 * Populate N_MEMORY for calculating usable_nodes.
6351 static unsigned long __init early_calculate_totalpages(void)
6353 unsigned long totalpages = 0;
6354 unsigned long start_pfn, end_pfn;
6357 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6358 unsigned long pages = end_pfn - start_pfn;
6360 totalpages += pages;
6362 node_set_state(nid, N_MEMORY);
6368 * Find the PFN the Movable zone begins in each node. Kernel memory
6369 * is spread evenly between nodes as long as the nodes have enough
6370 * memory. When they don't, some nodes will have more kernelcore than
6373 static void __init find_zone_movable_pfns_for_nodes(void)
6376 unsigned long usable_startpfn;
6377 unsigned long kernelcore_node, kernelcore_remaining;
6378 /* save the state before borrow the nodemask */
6379 nodemask_t saved_node_state = node_states[N_MEMORY];
6380 unsigned long totalpages = early_calculate_totalpages();
6381 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6382 struct memblock_region *r;
6384 /* Need to find movable_zone earlier when movable_node is specified. */
6385 find_usable_zone_for_movable();
6388 * If movable_node is specified, ignore kernelcore and movablecore
6391 if (movable_node_is_enabled()) {
6392 for_each_memblock(memory, r) {
6393 if (!memblock_is_hotpluggable(r))
6398 usable_startpfn = PFN_DOWN(r->base);
6399 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6400 min(usable_startpfn, zone_movable_pfn[nid]) :
6408 * If kernelcore=mirror is specified, ignore movablecore option
6410 if (mirrored_kernelcore) {
6411 bool mem_below_4gb_not_mirrored = false;
6413 for_each_memblock(memory, r) {
6414 if (memblock_is_mirror(r))
6419 usable_startpfn = memblock_region_memory_base_pfn(r);
6421 if (usable_startpfn < 0x100000) {
6422 mem_below_4gb_not_mirrored = true;
6426 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6427 min(usable_startpfn, zone_movable_pfn[nid]) :
6431 if (mem_below_4gb_not_mirrored)
6432 pr_warn("This configuration results in unmirrored kernel memory.");
6438 * If movablecore=nn[KMG] was specified, calculate what size of
6439 * kernelcore that corresponds so that memory usable for
6440 * any allocation type is evenly spread. If both kernelcore
6441 * and movablecore are specified, then the value of kernelcore
6442 * will be used for required_kernelcore if it's greater than
6443 * what movablecore would have allowed.
6445 if (required_movablecore) {
6446 unsigned long corepages;
6449 * Round-up so that ZONE_MOVABLE is at least as large as what
6450 * was requested by the user
6452 required_movablecore =
6453 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6454 required_movablecore = min(totalpages, required_movablecore);
6455 corepages = totalpages - required_movablecore;
6457 required_kernelcore = max(required_kernelcore, corepages);
6461 * If kernelcore was not specified or kernelcore size is larger
6462 * than totalpages, there is no ZONE_MOVABLE.
6464 if (!required_kernelcore || required_kernelcore >= totalpages)
6467 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6468 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6471 /* Spread kernelcore memory as evenly as possible throughout nodes */
6472 kernelcore_node = required_kernelcore / usable_nodes;
6473 for_each_node_state(nid, N_MEMORY) {
6474 unsigned long start_pfn, end_pfn;
6477 * Recalculate kernelcore_node if the division per node
6478 * now exceeds what is necessary to satisfy the requested
6479 * amount of memory for the kernel
6481 if (required_kernelcore < kernelcore_node)
6482 kernelcore_node = required_kernelcore / usable_nodes;
6485 * As the map is walked, we track how much memory is usable
6486 * by the kernel using kernelcore_remaining. When it is
6487 * 0, the rest of the node is usable by ZONE_MOVABLE
6489 kernelcore_remaining = kernelcore_node;
6491 /* Go through each range of PFNs within this node */
6492 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6493 unsigned long size_pages;
6495 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6496 if (start_pfn >= end_pfn)
6499 /* Account for what is only usable for kernelcore */
6500 if (start_pfn < usable_startpfn) {
6501 unsigned long kernel_pages;
6502 kernel_pages = min(end_pfn, usable_startpfn)
6505 kernelcore_remaining -= min(kernel_pages,
6506 kernelcore_remaining);
6507 required_kernelcore -= min(kernel_pages,
6508 required_kernelcore);
6510 /* Continue if range is now fully accounted */
6511 if (end_pfn <= usable_startpfn) {
6514 * Push zone_movable_pfn to the end so
6515 * that if we have to rebalance
6516 * kernelcore across nodes, we will
6517 * not double account here
6519 zone_movable_pfn[nid] = end_pfn;
6522 start_pfn = usable_startpfn;
6526 * The usable PFN range for ZONE_MOVABLE is from
6527 * start_pfn->end_pfn. Calculate size_pages as the
6528 * number of pages used as kernelcore
6530 size_pages = end_pfn - start_pfn;
6531 if (size_pages > kernelcore_remaining)
6532 size_pages = kernelcore_remaining;
6533 zone_movable_pfn[nid] = start_pfn + size_pages;
6536 * Some kernelcore has been met, update counts and
6537 * break if the kernelcore for this node has been
6540 required_kernelcore -= min(required_kernelcore,
6542 kernelcore_remaining -= size_pages;
6543 if (!kernelcore_remaining)
6549 * If there is still required_kernelcore, we do another pass with one
6550 * less node in the count. This will push zone_movable_pfn[nid] further
6551 * along on the nodes that still have memory until kernelcore is
6555 if (usable_nodes && required_kernelcore > usable_nodes)
6559 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6560 for (nid = 0; nid < MAX_NUMNODES; nid++)
6561 zone_movable_pfn[nid] =
6562 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6565 /* restore the node_state */
6566 node_states[N_MEMORY] = saved_node_state;
6569 /* Any regular or high memory on that node ? */
6570 static void check_for_memory(pg_data_t *pgdat, int nid)
6572 enum zone_type zone_type;
6574 if (N_MEMORY == N_NORMAL_MEMORY)
6577 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6578 struct zone *zone = &pgdat->node_zones[zone_type];
6579 if (populated_zone(zone)) {
6580 node_set_state(nid, N_HIGH_MEMORY);
6581 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6582 zone_type <= ZONE_NORMAL)
6583 node_set_state(nid, N_NORMAL_MEMORY);
6590 * free_area_init_nodes - Initialise all pg_data_t and zone data
6591 * @max_zone_pfn: an array of max PFNs for each zone
6593 * This will call free_area_init_node() for each active node in the system.
6594 * Using the page ranges provided by memblock_set_node(), the size of each
6595 * zone in each node and their holes is calculated. If the maximum PFN
6596 * between two adjacent zones match, it is assumed that the zone is empty.
6597 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6598 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6599 * starts where the previous one ended. For example, ZONE_DMA32 starts
6600 * at arch_max_dma_pfn.
6602 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6604 unsigned long start_pfn, end_pfn;
6607 /* Record where the zone boundaries are */
6608 memset(arch_zone_lowest_possible_pfn, 0,
6609 sizeof(arch_zone_lowest_possible_pfn));
6610 memset(arch_zone_highest_possible_pfn, 0,
6611 sizeof(arch_zone_highest_possible_pfn));
6613 start_pfn = find_min_pfn_with_active_regions();
6615 for (i = 0; i < MAX_NR_ZONES; i++) {
6616 if (i == ZONE_MOVABLE)
6619 end_pfn = max(max_zone_pfn[i], start_pfn);
6620 arch_zone_lowest_possible_pfn[i] = start_pfn;
6621 arch_zone_highest_possible_pfn[i] = end_pfn;
6623 start_pfn = end_pfn;
6626 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6627 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6628 find_zone_movable_pfns_for_nodes();
6630 /* Print out the zone ranges */
6631 pr_info("Zone ranges:\n");
6632 for (i = 0; i < MAX_NR_ZONES; i++) {
6633 if (i == ZONE_MOVABLE)
6635 pr_info(" %-8s ", zone_names[i]);
6636 if (arch_zone_lowest_possible_pfn[i] ==
6637 arch_zone_highest_possible_pfn[i])
6640 pr_cont("[mem %#018Lx-%#018Lx]\n",
6641 (u64)arch_zone_lowest_possible_pfn[i]
6643 ((u64)arch_zone_highest_possible_pfn[i]
6644 << PAGE_SHIFT) - 1);
6647 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6648 pr_info("Movable zone start for each node\n");
6649 for (i = 0; i < MAX_NUMNODES; i++) {
6650 if (zone_movable_pfn[i])
6651 pr_info(" Node %d: %#018Lx\n", i,
6652 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6655 /* Print out the early node map */
6656 pr_info("Early memory node ranges\n");
6657 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6658 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6659 (u64)start_pfn << PAGE_SHIFT,
6660 ((u64)end_pfn << PAGE_SHIFT) - 1);
6662 /* Initialise every node */
6663 mminit_verify_pageflags_layout();
6664 setup_nr_node_ids();
6665 for_each_online_node(nid) {
6666 pg_data_t *pgdat = NODE_DATA(nid);
6667 free_area_init_node(nid, NULL,
6668 find_min_pfn_for_node(nid), NULL);
6670 /* Any memory on that node */
6671 if (pgdat->node_present_pages)
6672 node_set_state(nid, N_MEMORY);
6673 check_for_memory(pgdat, nid);
6677 static int __init cmdline_parse_core(char *p, unsigned long *core)
6679 unsigned long long coremem;
6683 coremem = memparse(p, &p);
6684 *core = coremem >> PAGE_SHIFT;
6686 /* Paranoid check that UL is enough for the coremem value */
6687 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6693 * kernelcore=size sets the amount of memory for use for allocations that
6694 * cannot be reclaimed or migrated.
6696 static int __init cmdline_parse_kernelcore(char *p)
6698 /* parse kernelcore=mirror */
6699 if (parse_option_str(p, "mirror")) {
6700 mirrored_kernelcore = true;
6704 return cmdline_parse_core(p, &required_kernelcore);
6708 * movablecore=size sets the amount of memory for use for allocations that
6709 * can be reclaimed or migrated.
6711 static int __init cmdline_parse_movablecore(char *p)
6713 return cmdline_parse_core(p, &required_movablecore);
6716 early_param("kernelcore", cmdline_parse_kernelcore);
6717 early_param("movablecore", cmdline_parse_movablecore);
6719 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6721 void adjust_managed_page_count(struct page *page, long count)
6723 spin_lock(&managed_page_count_lock);
6724 page_zone(page)->managed_pages += count;
6725 totalram_pages += count;
6726 #ifdef CONFIG_HIGHMEM
6727 if (PageHighMem(page))
6728 totalhigh_pages += count;
6730 spin_unlock(&managed_page_count_lock);
6732 EXPORT_SYMBOL(adjust_managed_page_count);
6734 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6737 unsigned long pages = 0;
6739 start = (void *)PAGE_ALIGN((unsigned long)start);
6740 end = (void *)((unsigned long)end & PAGE_MASK);
6741 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6742 if ((unsigned int)poison <= 0xFF)
6743 memset(pos, poison, PAGE_SIZE);
6744 free_reserved_page(virt_to_page(pos));
6748 pr_info("Freeing %s memory: %ldK\n",
6749 s, pages << (PAGE_SHIFT - 10));
6753 EXPORT_SYMBOL(free_reserved_area);
6755 #ifdef CONFIG_HIGHMEM
6756 void free_highmem_page(struct page *page)
6758 __free_reserved_page(page);
6760 page_zone(page)->managed_pages++;
6766 void __init mem_init_print_info(const char *str)
6768 unsigned long physpages, codesize, datasize, rosize, bss_size;
6769 unsigned long init_code_size, init_data_size;
6771 physpages = get_num_physpages();
6772 codesize = _etext - _stext;
6773 datasize = _edata - _sdata;
6774 rosize = __end_rodata - __start_rodata;
6775 bss_size = __bss_stop - __bss_start;
6776 init_data_size = __init_end - __init_begin;
6777 init_code_size = _einittext - _sinittext;
6780 * Detect special cases and adjust section sizes accordingly:
6781 * 1) .init.* may be embedded into .data sections
6782 * 2) .init.text.* may be out of [__init_begin, __init_end],
6783 * please refer to arch/tile/kernel/vmlinux.lds.S.
6784 * 3) .rodata.* may be embedded into .text or .data sections.
6786 #define adj_init_size(start, end, size, pos, adj) \
6788 if (start <= pos && pos < end && size > adj) \
6792 adj_init_size(__init_begin, __init_end, init_data_size,
6793 _sinittext, init_code_size);
6794 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6795 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6796 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6797 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6799 #undef adj_init_size
6801 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6802 #ifdef CONFIG_HIGHMEM
6806 nr_free_pages() << (PAGE_SHIFT - 10),
6807 physpages << (PAGE_SHIFT - 10),
6808 codesize >> 10, datasize >> 10, rosize >> 10,
6809 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6810 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6811 totalcma_pages << (PAGE_SHIFT - 10),
6812 #ifdef CONFIG_HIGHMEM
6813 totalhigh_pages << (PAGE_SHIFT - 10),
6815 str ? ", " : "", str ? str : "");
6819 * set_dma_reserve - set the specified number of pages reserved in the first zone
6820 * @new_dma_reserve: The number of pages to mark reserved
6822 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6823 * In the DMA zone, a significant percentage may be consumed by kernel image
6824 * and other unfreeable allocations which can skew the watermarks badly. This
6825 * function may optionally be used to account for unfreeable pages in the
6826 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6827 * smaller per-cpu batchsize.
6829 void __init set_dma_reserve(unsigned long new_dma_reserve)
6831 dma_reserve = new_dma_reserve;
6834 void __init free_area_init(unsigned long *zones_size)
6836 free_area_init_node(0, zones_size,
6837 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6840 static int page_alloc_cpu_dead(unsigned int cpu)
6843 lru_add_drain_cpu(cpu);
6847 * Spill the event counters of the dead processor
6848 * into the current processors event counters.
6849 * This artificially elevates the count of the current
6852 vm_events_fold_cpu(cpu);
6855 * Zero the differential counters of the dead processor
6856 * so that the vm statistics are consistent.
6858 * This is only okay since the processor is dead and cannot
6859 * race with what we are doing.
6861 cpu_vm_stats_fold(cpu);
6865 void __init page_alloc_init(void)
6869 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6870 "mm/page_alloc:dead", NULL,
6871 page_alloc_cpu_dead);
6876 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6877 * or min_free_kbytes changes.
6879 static void calculate_totalreserve_pages(void)
6881 struct pglist_data *pgdat;
6882 unsigned long reserve_pages = 0;
6883 enum zone_type i, j;
6885 for_each_online_pgdat(pgdat) {
6887 pgdat->totalreserve_pages = 0;
6889 for (i = 0; i < MAX_NR_ZONES; i++) {
6890 struct zone *zone = pgdat->node_zones + i;
6893 /* Find valid and maximum lowmem_reserve in the zone */
6894 for (j = i; j < MAX_NR_ZONES; j++) {
6895 if (zone->lowmem_reserve[j] > max)
6896 max = zone->lowmem_reserve[j];
6899 /* we treat the high watermark as reserved pages. */
6900 max += high_wmark_pages(zone);
6902 if (max > zone->managed_pages)
6903 max = zone->managed_pages;
6905 pgdat->totalreserve_pages += max;
6907 reserve_pages += max;
6910 totalreserve_pages = reserve_pages;
6914 * setup_per_zone_lowmem_reserve - called whenever
6915 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6916 * has a correct pages reserved value, so an adequate number of
6917 * pages are left in the zone after a successful __alloc_pages().
6919 static void setup_per_zone_lowmem_reserve(void)
6921 struct pglist_data *pgdat;
6922 enum zone_type j, idx;
6924 for_each_online_pgdat(pgdat) {
6925 for (j = 0; j < MAX_NR_ZONES; j++) {
6926 struct zone *zone = pgdat->node_zones + j;
6927 unsigned long managed_pages = zone->managed_pages;
6929 zone->lowmem_reserve[j] = 0;
6933 struct zone *lower_zone;
6937 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6938 sysctl_lowmem_reserve_ratio[idx] = 1;
6940 lower_zone = pgdat->node_zones + idx;
6941 lower_zone->lowmem_reserve[j] = managed_pages /
6942 sysctl_lowmem_reserve_ratio[idx];
6943 managed_pages += lower_zone->managed_pages;
6948 /* update totalreserve_pages */
6949 calculate_totalreserve_pages();
6952 static void __setup_per_zone_wmarks(void)
6954 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6955 unsigned long lowmem_pages = 0;
6957 unsigned long flags;
6959 /* Calculate total number of !ZONE_HIGHMEM pages */
6960 for_each_zone(zone) {
6961 if (!is_highmem(zone))
6962 lowmem_pages += zone->managed_pages;
6965 for_each_zone(zone) {
6968 spin_lock_irqsave(&zone->lock, flags);
6969 tmp = (u64)pages_min * zone->managed_pages;
6970 do_div(tmp, lowmem_pages);
6971 if (is_highmem(zone)) {
6973 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6974 * need highmem pages, so cap pages_min to a small
6977 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6978 * deltas control asynch page reclaim, and so should
6979 * not be capped for highmem.
6981 unsigned long min_pages;
6983 min_pages = zone->managed_pages / 1024;
6984 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6985 zone->watermark[WMARK_MIN] = min_pages;
6988 * If it's a lowmem zone, reserve a number of pages
6989 * proportionate to the zone's size.
6991 zone->watermark[WMARK_MIN] = tmp;
6995 * Set the kswapd watermarks distance according to the
6996 * scale factor in proportion to available memory, but
6997 * ensure a minimum size on small systems.
6999 tmp = max_t(u64, tmp >> 2,
7000 mult_frac(zone->managed_pages,
7001 watermark_scale_factor, 10000));
7003 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7004 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7006 spin_unlock_irqrestore(&zone->lock, flags);
7009 /* update totalreserve_pages */
7010 calculate_totalreserve_pages();
7014 * setup_per_zone_wmarks - called when min_free_kbytes changes
7015 * or when memory is hot-{added|removed}
7017 * Ensures that the watermark[min,low,high] values for each zone are set
7018 * correctly with respect to min_free_kbytes.
7020 void setup_per_zone_wmarks(void)
7022 mutex_lock(&zonelists_mutex);
7023 __setup_per_zone_wmarks();
7024 mutex_unlock(&zonelists_mutex);
7028 * Initialise min_free_kbytes.
7030 * For small machines we want it small (128k min). For large machines
7031 * we want it large (64MB max). But it is not linear, because network
7032 * bandwidth does not increase linearly with machine size. We use
7034 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7035 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7051 int __meminit init_per_zone_wmark_min(void)
7053 unsigned long lowmem_kbytes;
7054 int new_min_free_kbytes;
7056 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7057 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7059 if (new_min_free_kbytes > user_min_free_kbytes) {
7060 min_free_kbytes = new_min_free_kbytes;
7061 if (min_free_kbytes < 128)
7062 min_free_kbytes = 128;
7063 if (min_free_kbytes > 65536)
7064 min_free_kbytes = 65536;
7066 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7067 new_min_free_kbytes, user_min_free_kbytes);
7069 setup_per_zone_wmarks();
7070 refresh_zone_stat_thresholds();
7071 setup_per_zone_lowmem_reserve();
7074 setup_min_unmapped_ratio();
7075 setup_min_slab_ratio();
7080 core_initcall(init_per_zone_wmark_min)
7083 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7084 * that we can call two helper functions whenever min_free_kbytes
7087 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7088 void __user *buffer, size_t *length, loff_t *ppos)
7092 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7097 user_min_free_kbytes = min_free_kbytes;
7098 setup_per_zone_wmarks();
7103 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7104 void __user *buffer, size_t *length, loff_t *ppos)
7108 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7113 setup_per_zone_wmarks();
7119 static void setup_min_unmapped_ratio(void)
7124 for_each_online_pgdat(pgdat)
7125 pgdat->min_unmapped_pages = 0;
7128 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7129 sysctl_min_unmapped_ratio) / 100;
7133 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7134 void __user *buffer, size_t *length, loff_t *ppos)
7138 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7142 setup_min_unmapped_ratio();
7147 static void setup_min_slab_ratio(void)
7152 for_each_online_pgdat(pgdat)
7153 pgdat->min_slab_pages = 0;
7156 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7157 sysctl_min_slab_ratio) / 100;
7160 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7161 void __user *buffer, size_t *length, loff_t *ppos)
7165 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7169 setup_min_slab_ratio();
7176 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7177 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7178 * whenever sysctl_lowmem_reserve_ratio changes.
7180 * The reserve ratio obviously has absolutely no relation with the
7181 * minimum watermarks. The lowmem reserve ratio can only make sense
7182 * if in function of the boot time zone sizes.
7184 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7185 void __user *buffer, size_t *length, loff_t *ppos)
7187 proc_dointvec_minmax(table, write, buffer, length, ppos);
7188 setup_per_zone_lowmem_reserve();
7193 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7194 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7195 * pagelist can have before it gets flushed back to buddy allocator.
7197 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7198 void __user *buffer, size_t *length, loff_t *ppos)
7201 int old_percpu_pagelist_fraction;
7204 mutex_lock(&pcp_batch_high_lock);
7205 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7207 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7208 if (!write || ret < 0)
7211 /* Sanity checking to avoid pcp imbalance */
7212 if (percpu_pagelist_fraction &&
7213 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7214 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7220 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7223 for_each_populated_zone(zone) {
7226 for_each_possible_cpu(cpu)
7227 pageset_set_high_and_batch(zone,
7228 per_cpu_ptr(zone->pageset, cpu));
7231 mutex_unlock(&pcp_batch_high_lock);
7236 int hashdist = HASHDIST_DEFAULT;
7238 static int __init set_hashdist(char *str)
7242 hashdist = simple_strtoul(str, &str, 0);
7245 __setup("hashdist=", set_hashdist);
7248 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7250 * Returns the number of pages that arch has reserved but
7251 * is not known to alloc_large_system_hash().
7253 static unsigned long __init arch_reserved_kernel_pages(void)
7260 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7261 * machines. As memory size is increased the scale is also increased but at
7262 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7263 * quadruples the scale is increased by one, which means the size of hash table
7264 * only doubles, instead of quadrupling as well.
7265 * Because 32-bit systems cannot have large physical memory, where this scaling
7266 * makes sense, it is disabled on such platforms.
7268 #if __BITS_PER_LONG > 32
7269 #define ADAPT_SCALE_BASE (64ul << 30)
7270 #define ADAPT_SCALE_SHIFT 2
7271 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7275 * allocate a large system hash table from bootmem
7276 * - it is assumed that the hash table must contain an exact power-of-2
7277 * quantity of entries
7278 * - limit is the number of hash buckets, not the total allocation size
7280 void *__init alloc_large_system_hash(const char *tablename,
7281 unsigned long bucketsize,
7282 unsigned long numentries,
7285 unsigned int *_hash_shift,
7286 unsigned int *_hash_mask,
7287 unsigned long low_limit,
7288 unsigned long high_limit)
7290 unsigned long long max = high_limit;
7291 unsigned long log2qty, size;
7295 /* allow the kernel cmdline to have a say */
7297 /* round applicable memory size up to nearest megabyte */
7298 numentries = nr_kernel_pages;
7299 numentries -= arch_reserved_kernel_pages();
7301 /* It isn't necessary when PAGE_SIZE >= 1MB */
7302 if (PAGE_SHIFT < 20)
7303 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7305 #if __BITS_PER_LONG > 32
7307 unsigned long adapt;
7309 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7310 adapt <<= ADAPT_SCALE_SHIFT)
7315 /* limit to 1 bucket per 2^scale bytes of low memory */
7316 if (scale > PAGE_SHIFT)
7317 numentries >>= (scale - PAGE_SHIFT);
7319 numentries <<= (PAGE_SHIFT - scale);
7321 /* Make sure we've got at least a 0-order allocation.. */
7322 if (unlikely(flags & HASH_SMALL)) {
7323 /* Makes no sense without HASH_EARLY */
7324 WARN_ON(!(flags & HASH_EARLY));
7325 if (!(numentries >> *_hash_shift)) {
7326 numentries = 1UL << *_hash_shift;
7327 BUG_ON(!numentries);
7329 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7330 numentries = PAGE_SIZE / bucketsize;
7332 numentries = roundup_pow_of_two(numentries);
7334 /* limit allocation size to 1/16 total memory by default */
7336 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7337 do_div(max, bucketsize);
7339 max = min(max, 0x80000000ULL);
7341 if (numentries < low_limit)
7342 numentries = low_limit;
7343 if (numentries > max)
7346 log2qty = ilog2(numentries);
7349 * memblock allocator returns zeroed memory already, so HASH_ZERO is
7350 * currently not used when HASH_EARLY is specified.
7352 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7354 size = bucketsize << log2qty;
7355 if (flags & HASH_EARLY)
7356 table = memblock_virt_alloc_nopanic(size, 0);
7358 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7361 * If bucketsize is not a power-of-two, we may free
7362 * some pages at the end of hash table which
7363 * alloc_pages_exact() automatically does
7365 if (get_order(size) < MAX_ORDER) {
7366 table = alloc_pages_exact(size, gfp_flags);
7367 kmemleak_alloc(table, size, 1, gfp_flags);
7370 } while (!table && size > PAGE_SIZE && --log2qty);
7373 panic("Failed to allocate %s hash table\n", tablename);
7375 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7376 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7379 *_hash_shift = log2qty;
7381 *_hash_mask = (1 << log2qty) - 1;
7387 * This function checks whether pageblock includes unmovable pages or not.
7388 * If @count is not zero, it is okay to include less @count unmovable pages
7390 * PageLRU check without isolation or lru_lock could race so that
7391 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7392 * check without lock_page also may miss some movable non-lru pages at
7393 * race condition. So you can't expect this function should be exact.
7395 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7396 bool skip_hwpoisoned_pages)
7398 unsigned long pfn, iter, found;
7402 * For avoiding noise data, lru_add_drain_all() should be called
7403 * If ZONE_MOVABLE, the zone never contains unmovable pages
7405 if (zone_idx(zone) == ZONE_MOVABLE)
7407 mt = get_pageblock_migratetype(page);
7408 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7411 pfn = page_to_pfn(page);
7412 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7413 unsigned long check = pfn + iter;
7415 if (!pfn_valid_within(check))
7418 page = pfn_to_page(check);
7421 * Hugepages are not in LRU lists, but they're movable.
7422 * We need not scan over tail pages bacause we don't
7423 * handle each tail page individually in migration.
7425 if (PageHuge(page)) {
7426 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7431 * We can't use page_count without pin a page
7432 * because another CPU can free compound page.
7433 * This check already skips compound tails of THP
7434 * because their page->_refcount is zero at all time.
7436 if (!page_ref_count(page)) {
7437 if (PageBuddy(page))
7438 iter += (1 << page_order(page)) - 1;
7443 * The HWPoisoned page may be not in buddy system, and
7444 * page_count() is not 0.
7446 if (skip_hwpoisoned_pages && PageHWPoison(page))
7449 if (__PageMovable(page))
7455 * If there are RECLAIMABLE pages, we need to check
7456 * it. But now, memory offline itself doesn't call
7457 * shrink_node_slabs() and it still to be fixed.
7460 * If the page is not RAM, page_count()should be 0.
7461 * we don't need more check. This is an _used_ not-movable page.
7463 * The problematic thing here is PG_reserved pages. PG_reserved
7464 * is set to both of a memory hole page and a _used_ kernel
7473 bool is_pageblock_removable_nolock(struct page *page)
7479 * We have to be careful here because we are iterating over memory
7480 * sections which are not zone aware so we might end up outside of
7481 * the zone but still within the section.
7482 * We have to take care about the node as well. If the node is offline
7483 * its NODE_DATA will be NULL - see page_zone.
7485 if (!node_online(page_to_nid(page)))
7488 zone = page_zone(page);
7489 pfn = page_to_pfn(page);
7490 if (!zone_spans_pfn(zone, pfn))
7493 return !has_unmovable_pages(zone, page, 0, true);
7496 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7498 static unsigned long pfn_max_align_down(unsigned long pfn)
7500 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7501 pageblock_nr_pages) - 1);
7504 static unsigned long pfn_max_align_up(unsigned long pfn)
7506 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7507 pageblock_nr_pages));
7510 /* [start, end) must belong to a single zone. */
7511 static int __alloc_contig_migrate_range(struct compact_control *cc,
7512 unsigned long start, unsigned long end)
7514 /* This function is based on compact_zone() from compaction.c. */
7515 unsigned long nr_reclaimed;
7516 unsigned long pfn = start;
7517 unsigned int tries = 0;
7522 while (pfn < end || !list_empty(&cc->migratepages)) {
7523 if (fatal_signal_pending(current)) {
7528 if (list_empty(&cc->migratepages)) {
7529 cc->nr_migratepages = 0;
7530 pfn = isolate_migratepages_range(cc, pfn, end);
7536 } else if (++tries == 5) {
7537 ret = ret < 0 ? ret : -EBUSY;
7541 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7543 cc->nr_migratepages -= nr_reclaimed;
7545 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7546 NULL, 0, cc->mode, MR_CMA);
7549 putback_movable_pages(&cc->migratepages);
7556 * alloc_contig_range() -- tries to allocate given range of pages
7557 * @start: start PFN to allocate
7558 * @end: one-past-the-last PFN to allocate
7559 * @migratetype: migratetype of the underlaying pageblocks (either
7560 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7561 * in range must have the same migratetype and it must
7562 * be either of the two.
7563 * @gfp_mask: GFP mask to use during compaction
7565 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7566 * aligned, however it's the caller's responsibility to guarantee that
7567 * we are the only thread that changes migrate type of pageblocks the
7570 * The PFN range must belong to a single zone.
7572 * Returns zero on success or negative error code. On success all
7573 * pages which PFN is in [start, end) are allocated for the caller and
7574 * need to be freed with free_contig_range().
7576 int alloc_contig_range(unsigned long start, unsigned long end,
7577 unsigned migratetype, gfp_t gfp_mask)
7579 unsigned long outer_start, outer_end;
7583 struct compact_control cc = {
7584 .nr_migratepages = 0,
7586 .zone = page_zone(pfn_to_page(start)),
7587 .mode = MIGRATE_SYNC,
7588 .ignore_skip_hint = true,
7589 .gfp_mask = current_gfp_context(gfp_mask),
7591 INIT_LIST_HEAD(&cc.migratepages);
7594 * What we do here is we mark all pageblocks in range as
7595 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7596 * have different sizes, and due to the way page allocator
7597 * work, we align the range to biggest of the two pages so
7598 * that page allocator won't try to merge buddies from
7599 * different pageblocks and change MIGRATE_ISOLATE to some
7600 * other migration type.
7602 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7603 * migrate the pages from an unaligned range (ie. pages that
7604 * we are interested in). This will put all the pages in
7605 * range back to page allocator as MIGRATE_ISOLATE.
7607 * When this is done, we take the pages in range from page
7608 * allocator removing them from the buddy system. This way
7609 * page allocator will never consider using them.
7611 * This lets us mark the pageblocks back as
7612 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7613 * aligned range but not in the unaligned, original range are
7614 * put back to page allocator so that buddy can use them.
7617 ret = start_isolate_page_range(pfn_max_align_down(start),
7618 pfn_max_align_up(end), migratetype,
7624 * In case of -EBUSY, we'd like to know which page causes problem.
7625 * So, just fall through. We will check it in test_pages_isolated().
7627 ret = __alloc_contig_migrate_range(&cc, start, end);
7628 if (ret && ret != -EBUSY)
7632 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7633 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7634 * more, all pages in [start, end) are free in page allocator.
7635 * What we are going to do is to allocate all pages from
7636 * [start, end) (that is remove them from page allocator).
7638 * The only problem is that pages at the beginning and at the
7639 * end of interesting range may be not aligned with pages that
7640 * page allocator holds, ie. they can be part of higher order
7641 * pages. Because of this, we reserve the bigger range and
7642 * once this is done free the pages we are not interested in.
7644 * We don't have to hold zone->lock here because the pages are
7645 * isolated thus they won't get removed from buddy.
7648 lru_add_drain_all();
7649 drain_all_pages(cc.zone);
7652 outer_start = start;
7653 while (!PageBuddy(pfn_to_page(outer_start))) {
7654 if (++order >= MAX_ORDER) {
7655 outer_start = start;
7658 outer_start &= ~0UL << order;
7661 if (outer_start != start) {
7662 order = page_order(pfn_to_page(outer_start));
7665 * outer_start page could be small order buddy page and
7666 * it doesn't include start page. Adjust outer_start
7667 * in this case to report failed page properly
7668 * on tracepoint in test_pages_isolated()
7670 if (outer_start + (1UL << order) <= start)
7671 outer_start = start;
7674 /* Make sure the range is really isolated. */
7675 if (test_pages_isolated(outer_start, end, false)) {
7676 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7677 __func__, outer_start, end);
7682 /* Grab isolated pages from freelists. */
7683 outer_end = isolate_freepages_range(&cc, outer_start, end);
7689 /* Free head and tail (if any) */
7690 if (start != outer_start)
7691 free_contig_range(outer_start, start - outer_start);
7692 if (end != outer_end)
7693 free_contig_range(end, outer_end - end);
7696 undo_isolate_page_range(pfn_max_align_down(start),
7697 pfn_max_align_up(end), migratetype);
7701 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7703 unsigned int count = 0;
7705 for (; nr_pages--; pfn++) {
7706 struct page *page = pfn_to_page(pfn);
7708 count += page_count(page) != 1;
7711 WARN(count != 0, "%d pages are still in use!\n", count);
7715 #ifdef CONFIG_MEMORY_HOTPLUG
7717 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7718 * page high values need to be recalulated.
7720 void __meminit zone_pcp_update(struct zone *zone)
7723 mutex_lock(&pcp_batch_high_lock);
7724 for_each_possible_cpu(cpu)
7725 pageset_set_high_and_batch(zone,
7726 per_cpu_ptr(zone->pageset, cpu));
7727 mutex_unlock(&pcp_batch_high_lock);
7731 void zone_pcp_reset(struct zone *zone)
7733 unsigned long flags;
7735 struct per_cpu_pageset *pset;
7737 /* avoid races with drain_pages() */
7738 local_irq_save(flags);
7739 if (zone->pageset != &boot_pageset) {
7740 for_each_online_cpu(cpu) {
7741 pset = per_cpu_ptr(zone->pageset, cpu);
7742 drain_zonestat(zone, pset);
7744 free_percpu(zone->pageset);
7745 zone->pageset = &boot_pageset;
7747 local_irq_restore(flags);
7750 #ifdef CONFIG_MEMORY_HOTREMOVE
7752 * All pages in the range must be in a single zone and isolated
7753 * before calling this.
7756 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7760 unsigned int order, i;
7762 unsigned long flags;
7763 /* find the first valid pfn */
7764 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7769 offline_mem_sections(pfn, end_pfn);
7770 zone = page_zone(pfn_to_page(pfn));
7771 spin_lock_irqsave(&zone->lock, flags);
7773 while (pfn < end_pfn) {
7774 if (!pfn_valid(pfn)) {
7778 page = pfn_to_page(pfn);
7780 * The HWPoisoned page may be not in buddy system, and
7781 * page_count() is not 0.
7783 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7785 SetPageReserved(page);
7789 BUG_ON(page_count(page));
7790 BUG_ON(!PageBuddy(page));
7791 order = page_order(page);
7792 #ifdef CONFIG_DEBUG_VM
7793 pr_info("remove from free list %lx %d %lx\n",
7794 pfn, 1 << order, end_pfn);
7796 list_del(&page->lru);
7797 rmv_page_order(page);
7798 zone->free_area[order].nr_free--;
7799 for (i = 0; i < (1 << order); i++)
7800 SetPageReserved((page+i));
7801 pfn += (1 << order);
7803 spin_unlock_irqrestore(&zone->lock, flags);
7807 bool is_free_buddy_page(struct page *page)
7809 struct zone *zone = page_zone(page);
7810 unsigned long pfn = page_to_pfn(page);
7811 unsigned long flags;
7814 spin_lock_irqsave(&zone->lock, flags);
7815 for (order = 0; order < MAX_ORDER; order++) {
7816 struct page *page_head = page - (pfn & ((1 << order) - 1));
7818 if (PageBuddy(page_head) && page_order(page_head) >= order)
7821 spin_unlock_irqrestore(&zone->lock, flags);
7823 return order < MAX_ORDER;