2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #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>
69 #include <linux/lockdep.h>
70 #include <linux/nmi.h>
72 #include <asm/sections.h>
73 #include <asm/tlbflush.h>
74 #include <asm/div64.h>
77 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
78 static DEFINE_MUTEX(pcp_batch_high_lock);
79 #define MIN_PERCPU_PAGELIST_FRACTION (8)
81 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
82 DEFINE_PER_CPU(int, numa_node);
83 EXPORT_PER_CPU_SYMBOL(numa_node);
86 DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
88 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
90 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
91 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
92 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
93 * defined in <linux/topology.h>.
95 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
96 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
97 int _node_numa_mem_[MAX_NUMNODES];
100 /* work_structs for global per-cpu drains */
101 DEFINE_MUTEX(pcpu_drain_mutex);
102 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
104 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
105 volatile unsigned long latent_entropy __latent_entropy;
106 EXPORT_SYMBOL(latent_entropy);
110 * Array of node states.
112 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
113 [N_POSSIBLE] = NODE_MASK_ALL,
114 [N_ONLINE] = { { [0] = 1UL } },
116 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
117 #ifdef CONFIG_HIGHMEM
118 [N_HIGH_MEMORY] = { { [0] = 1UL } },
120 [N_MEMORY] = { { [0] = 1UL } },
121 [N_CPU] = { { [0] = 1UL } },
124 EXPORT_SYMBOL(node_states);
126 /* Protect totalram_pages and zone->managed_pages */
127 static DEFINE_SPINLOCK(managed_page_count_lock);
129 unsigned long totalram_pages __read_mostly;
130 unsigned long totalreserve_pages __read_mostly;
131 unsigned long totalcma_pages __read_mostly;
133 int percpu_pagelist_fraction;
134 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
137 * A cached value of the page's pageblock's migratetype, used when the page is
138 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
139 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
140 * Also the migratetype set in the page does not necessarily match the pcplist
141 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
142 * other index - this ensures that it will be put on the correct CMA freelist.
144 static inline int get_pcppage_migratetype(struct page *page)
149 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
151 page->index = migratetype;
154 #ifdef CONFIG_PM_SLEEP
156 * The following functions are used by the suspend/hibernate code to temporarily
157 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
158 * while devices are suspended. To avoid races with the suspend/hibernate code,
159 * they should always be called with pm_mutex held (gfp_allowed_mask also should
160 * only be modified with pm_mutex held, unless the suspend/hibernate code is
161 * guaranteed not to run in parallel with that modification).
164 static gfp_t saved_gfp_mask;
166 void pm_restore_gfp_mask(void)
168 WARN_ON(!mutex_is_locked(&pm_mutex));
169 if (saved_gfp_mask) {
170 gfp_allowed_mask = saved_gfp_mask;
175 void pm_restrict_gfp_mask(void)
177 WARN_ON(!mutex_is_locked(&pm_mutex));
178 WARN_ON(saved_gfp_mask);
179 saved_gfp_mask = gfp_allowed_mask;
180 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
183 bool pm_suspended_storage(void)
185 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
189 #endif /* CONFIG_PM_SLEEP */
191 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
192 unsigned int pageblock_order __read_mostly;
195 static void __free_pages_ok(struct page *page, unsigned int order);
198 * results with 256, 32 in the lowmem_reserve sysctl:
199 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
200 * 1G machine -> (16M dma, 784M normal, 224M high)
201 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
202 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
203 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
205 * TBD: should special case ZONE_DMA32 machines here - in those we normally
206 * don't need any ZONE_NORMAL reservation
208 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
209 #ifdef CONFIG_ZONE_DMA
212 #ifdef CONFIG_ZONE_DMA32
215 #ifdef CONFIG_HIGHMEM
221 EXPORT_SYMBOL(totalram_pages);
223 static char * const zone_names[MAX_NR_ZONES] = {
224 #ifdef CONFIG_ZONE_DMA
227 #ifdef CONFIG_ZONE_DMA32
231 #ifdef CONFIG_HIGHMEM
235 #ifdef CONFIG_ZONE_DEVICE
240 char * const migratetype_names[MIGRATE_TYPES] = {
248 #ifdef CONFIG_MEMORY_ISOLATION
253 compound_page_dtor * const compound_page_dtors[] = {
256 #ifdef CONFIG_HUGETLB_PAGE
259 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
264 int min_free_kbytes = 1024;
265 int user_min_free_kbytes = -1;
266 int watermark_scale_factor = 10;
268 static unsigned long __meminitdata nr_kernel_pages;
269 static unsigned long __meminitdata nr_all_pages;
270 static unsigned long __meminitdata dma_reserve;
272 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
274 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
275 static unsigned long __initdata required_kernelcore;
276 static unsigned long __initdata required_movablecore;
277 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
278 static bool mirrored_kernelcore;
280 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
282 EXPORT_SYMBOL(movable_zone);
283 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
286 int nr_node_ids __read_mostly = MAX_NUMNODES;
287 int nr_online_nodes __read_mostly = 1;
288 EXPORT_SYMBOL(nr_node_ids);
289 EXPORT_SYMBOL(nr_online_nodes);
292 int page_group_by_mobility_disabled __read_mostly;
294 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
297 * Determine how many pages need to be initialized during early boot
298 * (non-deferred initialization).
299 * The value of first_deferred_pfn will be set later, once non-deferred pages
300 * are initialized, but for now set it ULONG_MAX.
302 static inline void reset_deferred_meminit(pg_data_t *pgdat)
304 phys_addr_t start_addr, end_addr;
305 unsigned long max_pgcnt;
306 unsigned long reserved;
309 * Initialise at least 2G of a node but also take into account that
310 * two large system hashes that can take up 1GB for 0.25TB/node.
312 max_pgcnt = max(2UL << (30 - PAGE_SHIFT),
313 (pgdat->node_spanned_pages >> 8));
316 * Compensate the all the memblock reservations (e.g. crash kernel)
317 * from the initial estimation to make sure we will initialize enough
320 start_addr = PFN_PHYS(pgdat->node_start_pfn);
321 end_addr = PFN_PHYS(pgdat->node_start_pfn + max_pgcnt);
322 reserved = memblock_reserved_memory_within(start_addr, end_addr);
323 max_pgcnt += PHYS_PFN(reserved);
325 pgdat->static_init_pgcnt = min(max_pgcnt, pgdat->node_spanned_pages);
326 pgdat->first_deferred_pfn = ULONG_MAX;
329 /* Returns true if the struct page for the pfn is uninitialised */
330 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
332 int nid = early_pfn_to_nid(pfn);
334 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
341 * Returns false when the remaining initialisation should be deferred until
342 * later in the boot cycle when it can be parallelised.
344 static inline bool update_defer_init(pg_data_t *pgdat,
345 unsigned long pfn, unsigned long zone_end,
346 unsigned long *nr_initialised)
348 /* Always populate low zones for address-constrained allocations */
349 if (zone_end < pgdat_end_pfn(pgdat))
351 /* Xen PV domains need page structures early */
355 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
356 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
357 pgdat->first_deferred_pfn = pfn;
364 static inline void reset_deferred_meminit(pg_data_t *pgdat)
368 static inline bool early_page_uninitialised(unsigned long pfn)
373 static inline bool update_defer_init(pg_data_t *pgdat,
374 unsigned long pfn, unsigned long zone_end,
375 unsigned long *nr_initialised)
381 /* Return a pointer to the bitmap storing bits affecting a block of pages */
382 static inline unsigned long *get_pageblock_bitmap(struct page *page,
385 #ifdef CONFIG_SPARSEMEM
386 return __pfn_to_section(pfn)->pageblock_flags;
388 return page_zone(page)->pageblock_flags;
389 #endif /* CONFIG_SPARSEMEM */
392 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
394 #ifdef CONFIG_SPARSEMEM
395 pfn &= (PAGES_PER_SECTION-1);
396 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
398 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
399 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
400 #endif /* CONFIG_SPARSEMEM */
404 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
405 * @page: The page within the block of interest
406 * @pfn: The target page frame number
407 * @end_bitidx: The last bit of interest to retrieve
408 * @mask: mask of bits that the caller is interested in
410 * Return: pageblock_bits flags
412 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
414 unsigned long end_bitidx,
417 unsigned long *bitmap;
418 unsigned long bitidx, word_bitidx;
421 bitmap = get_pageblock_bitmap(page, pfn);
422 bitidx = pfn_to_bitidx(page, pfn);
423 word_bitidx = bitidx / BITS_PER_LONG;
424 bitidx &= (BITS_PER_LONG-1);
426 word = bitmap[word_bitidx];
427 bitidx += end_bitidx;
428 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
431 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
432 unsigned long end_bitidx,
435 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
438 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
440 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
444 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
445 * @page: The page within the block of interest
446 * @flags: The flags to set
447 * @pfn: The target page frame number
448 * @end_bitidx: The last bit of interest
449 * @mask: mask of bits that the caller is interested in
451 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
453 unsigned long end_bitidx,
456 unsigned long *bitmap;
457 unsigned long bitidx, word_bitidx;
458 unsigned long old_word, word;
460 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
462 bitmap = get_pageblock_bitmap(page, pfn);
463 bitidx = pfn_to_bitidx(page, pfn);
464 word_bitidx = bitidx / BITS_PER_LONG;
465 bitidx &= (BITS_PER_LONG-1);
467 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
469 bitidx += end_bitidx;
470 mask <<= (BITS_PER_LONG - bitidx - 1);
471 flags <<= (BITS_PER_LONG - bitidx - 1);
473 word = READ_ONCE(bitmap[word_bitidx]);
475 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
476 if (word == old_word)
482 void set_pageblock_migratetype(struct page *page, int migratetype)
484 if (unlikely(page_group_by_mobility_disabled &&
485 migratetype < MIGRATE_PCPTYPES))
486 migratetype = MIGRATE_UNMOVABLE;
488 set_pageblock_flags_group(page, (unsigned long)migratetype,
489 PB_migrate, PB_migrate_end);
492 #ifdef CONFIG_DEBUG_VM
493 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
497 unsigned long pfn = page_to_pfn(page);
498 unsigned long sp, start_pfn;
501 seq = zone_span_seqbegin(zone);
502 start_pfn = zone->zone_start_pfn;
503 sp = zone->spanned_pages;
504 if (!zone_spans_pfn(zone, pfn))
506 } while (zone_span_seqretry(zone, seq));
509 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
510 pfn, zone_to_nid(zone), zone->name,
511 start_pfn, start_pfn + sp);
516 static int page_is_consistent(struct zone *zone, struct page *page)
518 if (!pfn_valid_within(page_to_pfn(page)))
520 if (zone != page_zone(page))
526 * Temporary debugging check for pages not lying within a given zone.
528 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
530 if (page_outside_zone_boundaries(zone, page))
532 if (!page_is_consistent(zone, page))
538 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
544 static void bad_page(struct page *page, const char *reason,
545 unsigned long bad_flags)
547 static unsigned long resume;
548 static unsigned long nr_shown;
549 static unsigned long nr_unshown;
552 * Allow a burst of 60 reports, then keep quiet for that minute;
553 * or allow a steady drip of one report per second.
555 if (nr_shown == 60) {
556 if (time_before(jiffies, resume)) {
562 "BUG: Bad page state: %lu messages suppressed\n",
569 resume = jiffies + 60 * HZ;
571 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
572 current->comm, page_to_pfn(page));
573 __dump_page(page, reason);
574 bad_flags &= page->flags;
576 pr_alert("bad because of flags: %#lx(%pGp)\n",
577 bad_flags, &bad_flags);
578 dump_page_owner(page);
583 /* Leave bad fields for debug, except PageBuddy could make trouble */
584 page_mapcount_reset(page); /* remove PageBuddy */
585 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
589 * Higher-order pages are called "compound pages". They are structured thusly:
591 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
593 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
594 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
596 * The first tail page's ->compound_dtor holds the offset in array of compound
597 * page destructors. See compound_page_dtors.
599 * The first tail page's ->compound_order holds the order of allocation.
600 * This usage means that zero-order pages may not be compound.
603 void free_compound_page(struct page *page)
605 __free_pages_ok(page, compound_order(page));
608 void prep_compound_page(struct page *page, unsigned int order)
611 int nr_pages = 1 << order;
613 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
614 set_compound_order(page, order);
616 for (i = 1; i < nr_pages; i++) {
617 struct page *p = page + i;
618 set_page_count(p, 0);
619 p->mapping = TAIL_MAPPING;
620 set_compound_head(p, page);
622 atomic_set(compound_mapcount_ptr(page), -1);
625 #ifdef CONFIG_DEBUG_PAGEALLOC
626 unsigned int _debug_guardpage_minorder;
627 bool _debug_pagealloc_enabled __read_mostly
628 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
629 EXPORT_SYMBOL(_debug_pagealloc_enabled);
630 bool _debug_guardpage_enabled __read_mostly;
632 static int __init early_debug_pagealloc(char *buf)
636 return kstrtobool(buf, &_debug_pagealloc_enabled);
638 early_param("debug_pagealloc", early_debug_pagealloc);
640 static bool need_debug_guardpage(void)
642 /* If we don't use debug_pagealloc, we don't need guard page */
643 if (!debug_pagealloc_enabled())
646 if (!debug_guardpage_minorder())
652 static void init_debug_guardpage(void)
654 if (!debug_pagealloc_enabled())
657 if (!debug_guardpage_minorder())
660 _debug_guardpage_enabled = true;
663 struct page_ext_operations debug_guardpage_ops = {
664 .need = need_debug_guardpage,
665 .init = init_debug_guardpage,
668 static int __init debug_guardpage_minorder_setup(char *buf)
672 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
673 pr_err("Bad debug_guardpage_minorder value\n");
676 _debug_guardpage_minorder = res;
677 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
680 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
682 static inline bool set_page_guard(struct zone *zone, struct page *page,
683 unsigned int order, int migratetype)
685 struct page_ext *page_ext;
687 if (!debug_guardpage_enabled())
690 if (order >= debug_guardpage_minorder())
693 page_ext = lookup_page_ext(page);
694 if (unlikely(!page_ext))
697 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
699 INIT_LIST_HEAD(&page->lru);
700 set_page_private(page, order);
701 /* Guard pages are not available for any usage */
702 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
707 static inline void clear_page_guard(struct zone *zone, struct page *page,
708 unsigned int order, int migratetype)
710 struct page_ext *page_ext;
712 if (!debug_guardpage_enabled())
715 page_ext = lookup_page_ext(page);
716 if (unlikely(!page_ext))
719 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
721 set_page_private(page, 0);
722 if (!is_migrate_isolate(migratetype))
723 __mod_zone_freepage_state(zone, (1 << order), migratetype);
726 struct page_ext_operations debug_guardpage_ops;
727 static inline bool set_page_guard(struct zone *zone, struct page *page,
728 unsigned int order, int migratetype) { return false; }
729 static inline void clear_page_guard(struct zone *zone, struct page *page,
730 unsigned int order, int migratetype) {}
733 static inline void set_page_order(struct page *page, unsigned int order)
735 set_page_private(page, order);
736 __SetPageBuddy(page);
739 static inline void rmv_page_order(struct page *page)
741 __ClearPageBuddy(page);
742 set_page_private(page, 0);
746 * This function checks whether a page is free && is the buddy
747 * we can do coalesce a page and its buddy if
748 * (a) the buddy is not in a hole (check before calling!) &&
749 * (b) the buddy is in the buddy system &&
750 * (c) a page and its buddy have the same order &&
751 * (d) a page and its buddy are in the same zone.
753 * For recording whether a page is in the buddy system, we set ->_mapcount
754 * PAGE_BUDDY_MAPCOUNT_VALUE.
755 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
756 * serialized by zone->lock.
758 * For recording page's order, we use page_private(page).
760 static inline int page_is_buddy(struct page *page, struct page *buddy,
763 if (page_is_guard(buddy) && page_order(buddy) == order) {
764 if (page_zone_id(page) != page_zone_id(buddy))
767 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
772 if (PageBuddy(buddy) && page_order(buddy) == order) {
774 * zone check is done late to avoid uselessly
775 * calculating zone/node ids for pages that could
778 if (page_zone_id(page) != page_zone_id(buddy))
781 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
789 * Freeing function for a buddy system allocator.
791 * The concept of a buddy system is to maintain direct-mapped table
792 * (containing bit values) for memory blocks of various "orders".
793 * The bottom level table contains the map for the smallest allocatable
794 * units of memory (here, pages), and each level above it describes
795 * pairs of units from the levels below, hence, "buddies".
796 * At a high level, all that happens here is marking the table entry
797 * at the bottom level available, and propagating the changes upward
798 * as necessary, plus some accounting needed to play nicely with other
799 * parts of the VM system.
800 * At each level, we keep a list of pages, which are heads of continuous
801 * free pages of length of (1 << order) and marked with _mapcount
802 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
804 * So when we are allocating or freeing one, we can derive the state of the
805 * other. That is, if we allocate a small block, and both were
806 * free, the remainder of the region must be split into blocks.
807 * If a block is freed, and its buddy is also free, then this
808 * triggers coalescing into a block of larger size.
813 static inline void __free_one_page(struct page *page,
815 struct zone *zone, unsigned int order,
818 unsigned long combined_pfn;
819 unsigned long uninitialized_var(buddy_pfn);
821 unsigned int max_order;
823 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
825 VM_BUG_ON(!zone_is_initialized(zone));
826 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
828 VM_BUG_ON(migratetype == -1);
829 if (likely(!is_migrate_isolate(migratetype)))
830 __mod_zone_freepage_state(zone, 1 << order, migratetype);
832 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
833 VM_BUG_ON_PAGE(bad_range(zone, page), page);
836 while (order < max_order - 1) {
837 buddy_pfn = __find_buddy_pfn(pfn, order);
838 buddy = page + (buddy_pfn - pfn);
840 if (!pfn_valid_within(buddy_pfn))
842 if (!page_is_buddy(page, buddy, order))
845 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
846 * merge with it and move up one order.
848 if (page_is_guard(buddy)) {
849 clear_page_guard(zone, buddy, order, migratetype);
851 list_del(&buddy->lru);
852 zone->free_area[order].nr_free--;
853 rmv_page_order(buddy);
855 combined_pfn = buddy_pfn & pfn;
856 page = page + (combined_pfn - pfn);
860 if (max_order < MAX_ORDER) {
861 /* If we are here, it means order is >= pageblock_order.
862 * We want to prevent merge between freepages on isolate
863 * pageblock and normal pageblock. Without this, pageblock
864 * isolation could cause incorrect freepage or CMA accounting.
866 * We don't want to hit this code for the more frequent
869 if (unlikely(has_isolate_pageblock(zone))) {
872 buddy_pfn = __find_buddy_pfn(pfn, order);
873 buddy = page + (buddy_pfn - pfn);
874 buddy_mt = get_pageblock_migratetype(buddy);
876 if (migratetype != buddy_mt
877 && (is_migrate_isolate(migratetype) ||
878 is_migrate_isolate(buddy_mt)))
882 goto continue_merging;
886 set_page_order(page, order);
889 * If this is not the largest possible page, check if the buddy
890 * of the next-highest order is free. If it is, it's possible
891 * that pages are being freed that will coalesce soon. In case,
892 * that is happening, add the free page to the tail of the list
893 * so it's less likely to be used soon and more likely to be merged
894 * as a higher order page
896 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
897 struct page *higher_page, *higher_buddy;
898 combined_pfn = buddy_pfn & pfn;
899 higher_page = page + (combined_pfn - pfn);
900 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
901 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
902 if (pfn_valid_within(buddy_pfn) &&
903 page_is_buddy(higher_page, higher_buddy, order + 1)) {
904 list_add_tail(&page->lru,
905 &zone->free_area[order].free_list[migratetype]);
910 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
912 zone->free_area[order].nr_free++;
916 * A bad page could be due to a number of fields. Instead of multiple branches,
917 * try and check multiple fields with one check. The caller must do a detailed
918 * check if necessary.
920 static inline bool page_expected_state(struct page *page,
921 unsigned long check_flags)
923 if (unlikely(atomic_read(&page->_mapcount) != -1))
926 if (unlikely((unsigned long)page->mapping |
927 page_ref_count(page) |
929 (unsigned long)page->mem_cgroup |
931 (page->flags & check_flags)))
937 static void free_pages_check_bad(struct page *page)
939 const char *bad_reason;
940 unsigned long bad_flags;
945 if (unlikely(atomic_read(&page->_mapcount) != -1))
946 bad_reason = "nonzero mapcount";
947 if (unlikely(page->mapping != NULL))
948 bad_reason = "non-NULL mapping";
949 if (unlikely(page_ref_count(page) != 0))
950 bad_reason = "nonzero _refcount";
951 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
952 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
953 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
956 if (unlikely(page->mem_cgroup))
957 bad_reason = "page still charged to cgroup";
959 bad_page(page, bad_reason, bad_flags);
962 static inline int free_pages_check(struct page *page)
964 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
967 /* Something has gone sideways, find it */
968 free_pages_check_bad(page);
972 static int free_tail_pages_check(struct page *head_page, struct page *page)
977 * We rely page->lru.next never has bit 0 set, unless the page
978 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
980 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
982 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
986 switch (page - head_page) {
988 /* the first tail page: ->mapping is compound_mapcount() */
989 if (unlikely(compound_mapcount(page))) {
990 bad_page(page, "nonzero compound_mapcount", 0);
996 * the second tail page: ->mapping is
997 * page_deferred_list().next -- ignore value.
1001 if (page->mapping != TAIL_MAPPING) {
1002 bad_page(page, "corrupted mapping in tail page", 0);
1007 if (unlikely(!PageTail(page))) {
1008 bad_page(page, "PageTail not set", 0);
1011 if (unlikely(compound_head(page) != head_page)) {
1012 bad_page(page, "compound_head not consistent", 0);
1017 page->mapping = NULL;
1018 clear_compound_head(page);
1022 static __always_inline bool free_pages_prepare(struct page *page,
1023 unsigned int order, bool check_free)
1027 VM_BUG_ON_PAGE(PageTail(page), page);
1029 trace_mm_page_free(page, order);
1032 * Check tail pages before head page information is cleared to
1033 * avoid checking PageCompound for order-0 pages.
1035 if (unlikely(order)) {
1036 bool compound = PageCompound(page);
1039 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1042 ClearPageDoubleMap(page);
1043 for (i = 1; i < (1 << order); i++) {
1045 bad += free_tail_pages_check(page, page + i);
1046 if (unlikely(free_pages_check(page + i))) {
1050 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1053 if (PageMappingFlags(page))
1054 page->mapping = NULL;
1055 if (memcg_kmem_enabled() && PageKmemcg(page))
1056 memcg_kmem_uncharge(page, order);
1058 bad += free_pages_check(page);
1062 page_cpupid_reset_last(page);
1063 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1064 reset_page_owner(page, order);
1066 if (!PageHighMem(page)) {
1067 debug_check_no_locks_freed(page_address(page),
1068 PAGE_SIZE << order);
1069 debug_check_no_obj_freed(page_address(page),
1070 PAGE_SIZE << order);
1072 arch_free_page(page, order);
1073 kernel_poison_pages(page, 1 << order, 0);
1074 kernel_map_pages(page, 1 << order, 0);
1075 kasan_free_pages(page, order);
1080 #ifdef CONFIG_DEBUG_VM
1081 static inline bool free_pcp_prepare(struct page *page)
1083 return free_pages_prepare(page, 0, true);
1086 static inline bool bulkfree_pcp_prepare(struct page *page)
1091 static bool free_pcp_prepare(struct page *page)
1093 return free_pages_prepare(page, 0, false);
1096 static bool bulkfree_pcp_prepare(struct page *page)
1098 return free_pages_check(page);
1100 #endif /* CONFIG_DEBUG_VM */
1103 * Frees a number of pages from the PCP lists
1104 * Assumes all pages on list are in same zone, and of same order.
1105 * count is the number of pages to free.
1107 * If the zone was previously in an "all pages pinned" state then look to
1108 * see if this freeing clears that state.
1110 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1111 * pinned" detection logic.
1113 static void free_pcppages_bulk(struct zone *zone, int count,
1114 struct per_cpu_pages *pcp)
1116 int migratetype = 0;
1118 bool isolated_pageblocks;
1120 spin_lock(&zone->lock);
1121 isolated_pageblocks = has_isolate_pageblock(zone);
1125 struct list_head *list;
1128 * Remove pages from lists in a round-robin fashion. A
1129 * batch_free count is maintained that is incremented when an
1130 * empty list is encountered. This is so more pages are freed
1131 * off fuller lists instead of spinning excessively around empty
1136 if (++migratetype == MIGRATE_PCPTYPES)
1138 list = &pcp->lists[migratetype];
1139 } while (list_empty(list));
1141 /* This is the only non-empty list. Free them all. */
1142 if (batch_free == MIGRATE_PCPTYPES)
1146 int mt; /* migratetype of the to-be-freed page */
1148 page = list_last_entry(list, struct page, lru);
1149 /* must delete as __free_one_page list manipulates */
1150 list_del(&page->lru);
1152 mt = get_pcppage_migratetype(page);
1153 /* MIGRATE_ISOLATE page should not go to pcplists */
1154 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1155 /* Pageblock could have been isolated meanwhile */
1156 if (unlikely(isolated_pageblocks))
1157 mt = get_pageblock_migratetype(page);
1159 if (bulkfree_pcp_prepare(page))
1162 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1163 trace_mm_page_pcpu_drain(page, 0, mt);
1164 } while (--count && --batch_free && !list_empty(list));
1166 spin_unlock(&zone->lock);
1169 static void free_one_page(struct zone *zone,
1170 struct page *page, unsigned long pfn,
1174 spin_lock(&zone->lock);
1175 if (unlikely(has_isolate_pageblock(zone) ||
1176 is_migrate_isolate(migratetype))) {
1177 migratetype = get_pfnblock_migratetype(page, pfn);
1179 __free_one_page(page, pfn, zone, order, migratetype);
1180 spin_unlock(&zone->lock);
1183 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1184 unsigned long zone, int nid, bool zero)
1187 mm_zero_struct_page(page);
1188 set_page_links(page, zone, nid, pfn);
1189 init_page_count(page);
1190 page_mapcount_reset(page);
1191 page_cpupid_reset_last(page);
1193 INIT_LIST_HEAD(&page->lru);
1194 #ifdef WANT_PAGE_VIRTUAL
1195 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1196 if (!is_highmem_idx(zone))
1197 set_page_address(page, __va(pfn << PAGE_SHIFT));
1201 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1204 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid, zero);
1207 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1208 static void __meminit init_reserved_page(unsigned long pfn)
1213 if (!early_page_uninitialised(pfn))
1216 nid = early_pfn_to_nid(pfn);
1217 pgdat = NODE_DATA(nid);
1219 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1220 struct zone *zone = &pgdat->node_zones[zid];
1222 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1225 __init_single_pfn(pfn, zid, nid, true);
1228 static inline void init_reserved_page(unsigned long pfn)
1231 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1234 * Initialised pages do not have PageReserved set. This function is
1235 * called for each range allocated by the bootmem allocator and
1236 * marks the pages PageReserved. The remaining valid pages are later
1237 * sent to the buddy page allocator.
1239 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1241 unsigned long start_pfn = PFN_DOWN(start);
1242 unsigned long end_pfn = PFN_UP(end);
1244 for (; start_pfn < end_pfn; start_pfn++) {
1245 if (pfn_valid(start_pfn)) {
1246 struct page *page = pfn_to_page(start_pfn);
1248 init_reserved_page(start_pfn);
1250 /* Avoid false-positive PageTail() */
1251 INIT_LIST_HEAD(&page->lru);
1253 SetPageReserved(page);
1258 static void __free_pages_ok(struct page *page, unsigned int order)
1260 unsigned long flags;
1262 unsigned long pfn = page_to_pfn(page);
1264 if (!free_pages_prepare(page, order, true))
1267 migratetype = get_pfnblock_migratetype(page, pfn);
1268 local_irq_save(flags);
1269 __count_vm_events(PGFREE, 1 << order);
1270 free_one_page(page_zone(page), page, pfn, order, migratetype);
1271 local_irq_restore(flags);
1274 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1276 unsigned int nr_pages = 1 << order;
1277 struct page *p = page;
1281 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1283 __ClearPageReserved(p);
1284 set_page_count(p, 0);
1286 __ClearPageReserved(p);
1287 set_page_count(p, 0);
1289 page_zone(page)->managed_pages += nr_pages;
1290 set_page_refcounted(page);
1291 __free_pages(page, order);
1294 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1295 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1297 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1299 int __meminit early_pfn_to_nid(unsigned long pfn)
1301 static DEFINE_SPINLOCK(early_pfn_lock);
1304 spin_lock(&early_pfn_lock);
1305 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1307 nid = first_online_node;
1308 spin_unlock(&early_pfn_lock);
1314 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1315 static inline bool __meminit __maybe_unused
1316 meminit_pfn_in_nid(unsigned long pfn, int node,
1317 struct mminit_pfnnid_cache *state)
1321 nid = __early_pfn_to_nid(pfn, state);
1322 if (nid >= 0 && nid != node)
1327 /* Only safe to use early in boot when initialisation is single-threaded */
1328 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1330 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1335 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1339 static inline bool __meminit __maybe_unused
1340 meminit_pfn_in_nid(unsigned long pfn, int node,
1341 struct mminit_pfnnid_cache *state)
1348 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1351 if (early_page_uninitialised(pfn))
1353 return __free_pages_boot_core(page, order);
1357 * Check that the whole (or subset of) a pageblock given by the interval of
1358 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1359 * with the migration of free compaction scanner. The scanners then need to
1360 * use only pfn_valid_within() check for arches that allow holes within
1363 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1365 * It's possible on some configurations to have a setup like node0 node1 node0
1366 * i.e. it's possible that all pages within a zones range of pages do not
1367 * belong to a single zone. We assume that a border between node0 and node1
1368 * can occur within a single pageblock, but not a node0 node1 node0
1369 * interleaving within a single pageblock. It is therefore sufficient to check
1370 * the first and last page of a pageblock and avoid checking each individual
1371 * page in a pageblock.
1373 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1374 unsigned long end_pfn, struct zone *zone)
1376 struct page *start_page;
1377 struct page *end_page;
1379 /* end_pfn is one past the range we are checking */
1382 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1385 start_page = pfn_to_online_page(start_pfn);
1389 if (page_zone(start_page) != zone)
1392 end_page = pfn_to_page(end_pfn);
1394 /* This gives a shorter code than deriving page_zone(end_page) */
1395 if (page_zone_id(start_page) != page_zone_id(end_page))
1401 void set_zone_contiguous(struct zone *zone)
1403 unsigned long block_start_pfn = zone->zone_start_pfn;
1404 unsigned long block_end_pfn;
1406 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1407 for (; block_start_pfn < zone_end_pfn(zone);
1408 block_start_pfn = block_end_pfn,
1409 block_end_pfn += pageblock_nr_pages) {
1411 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1413 if (!__pageblock_pfn_to_page(block_start_pfn,
1414 block_end_pfn, zone))
1418 /* We confirm that there is no hole */
1419 zone->contiguous = true;
1422 void clear_zone_contiguous(struct zone *zone)
1424 zone->contiguous = false;
1427 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1428 static void __init deferred_free_range(unsigned long pfn,
1429 unsigned long nr_pages)
1437 page = pfn_to_page(pfn);
1439 /* Free a large naturally-aligned chunk if possible */
1440 if (nr_pages == pageblock_nr_pages &&
1441 (pfn & (pageblock_nr_pages - 1)) == 0) {
1442 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1443 __free_pages_boot_core(page, pageblock_order);
1447 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1448 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1449 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1450 __free_pages_boot_core(page, 0);
1454 /* Completion tracking for deferred_init_memmap() threads */
1455 static atomic_t pgdat_init_n_undone __initdata;
1456 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1458 static inline void __init pgdat_init_report_one_done(void)
1460 if (atomic_dec_and_test(&pgdat_init_n_undone))
1461 complete(&pgdat_init_all_done_comp);
1465 * Returns true if page needs to be initialized or freed to buddy allocator.
1467 * First we check if pfn is valid on architectures where it is possible to have
1468 * holes within pageblock_nr_pages. On systems where it is not possible, this
1469 * function is optimized out.
1471 * Then, we check if a current large page is valid by only checking the validity
1474 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1475 * within a node: a pfn is between start and end of a node, but does not belong
1476 * to this memory node.
1478 static inline bool __init
1479 deferred_pfn_valid(int nid, unsigned long pfn,
1480 struct mminit_pfnnid_cache *nid_init_state)
1482 if (!pfn_valid_within(pfn))
1484 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1486 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1492 * Free pages to buddy allocator. Try to free aligned pages in
1493 * pageblock_nr_pages sizes.
1495 static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1496 unsigned long end_pfn)
1498 struct mminit_pfnnid_cache nid_init_state = { };
1499 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1500 unsigned long nr_free = 0;
1502 for (; pfn < end_pfn; pfn++) {
1503 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1504 deferred_free_range(pfn - nr_free, nr_free);
1506 } else if (!(pfn & nr_pgmask)) {
1507 deferred_free_range(pfn - nr_free, nr_free);
1514 /* Free the last block of pages to allocator */
1515 deferred_free_range(pfn - nr_free, nr_free);
1519 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1520 * by performing it only once every pageblock_nr_pages.
1521 * Return number of pages initialized.
1523 static unsigned long __init deferred_init_pages(int nid, int zid,
1525 unsigned long end_pfn)
1527 struct mminit_pfnnid_cache nid_init_state = { };
1528 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1529 unsigned long nr_pages = 0;
1530 struct page *page = NULL;
1532 for (; pfn < end_pfn; pfn++) {
1533 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1536 } else if (!page || !(pfn & nr_pgmask)) {
1537 page = pfn_to_page(pfn);
1542 __init_single_page(page, pfn, zid, nid, true);
1548 /* Initialise remaining memory on a node */
1549 static int __init deferred_init_memmap(void *data)
1551 pg_data_t *pgdat = data;
1552 int nid = pgdat->node_id;
1553 unsigned long start = jiffies;
1554 unsigned long nr_pages = 0;
1555 unsigned long spfn, epfn;
1556 phys_addr_t spa, epa;
1559 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1560 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1563 if (first_init_pfn == ULONG_MAX) {
1564 pgdat_init_report_one_done();
1568 /* Bind memory initialisation thread to a local node if possible */
1569 if (!cpumask_empty(cpumask))
1570 set_cpus_allowed_ptr(current, cpumask);
1572 /* Sanity check boundaries */
1573 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1574 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1575 pgdat->first_deferred_pfn = ULONG_MAX;
1577 /* Only the highest zone is deferred so find it */
1578 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1579 zone = pgdat->node_zones + zid;
1580 if (first_init_pfn < zone_end_pfn(zone))
1583 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1586 * Initialize and free pages. We do it in two loops: first we initialize
1587 * struct page, than free to buddy allocator, because while we are
1588 * freeing pages we can access pages that are ahead (computing buddy
1589 * page in __free_one_page()).
1591 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1592 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1593 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1594 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1596 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1597 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1598 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1599 deferred_free_pages(nid, zid, spfn, epfn);
1602 /* Sanity check that the next zone really is unpopulated */
1603 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1605 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1606 jiffies_to_msecs(jiffies - start));
1608 pgdat_init_report_one_done();
1611 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1613 void __init page_alloc_init_late(void)
1617 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1620 /* There will be num_node_state(N_MEMORY) threads */
1621 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1622 for_each_node_state(nid, N_MEMORY) {
1623 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1626 /* Block until all are initialised */
1627 wait_for_completion(&pgdat_init_all_done_comp);
1629 /* Reinit limits that are based on free pages after the kernel is up */
1630 files_maxfiles_init();
1632 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1633 /* Discard memblock private memory */
1637 for_each_populated_zone(zone)
1638 set_zone_contiguous(zone);
1642 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1643 void __init init_cma_reserved_pageblock(struct page *page)
1645 unsigned i = pageblock_nr_pages;
1646 struct page *p = page;
1649 __ClearPageReserved(p);
1650 set_page_count(p, 0);
1653 set_pageblock_migratetype(page, MIGRATE_CMA);
1655 if (pageblock_order >= MAX_ORDER) {
1656 i = pageblock_nr_pages;
1659 set_page_refcounted(p);
1660 __free_pages(p, MAX_ORDER - 1);
1661 p += MAX_ORDER_NR_PAGES;
1662 } while (i -= MAX_ORDER_NR_PAGES);
1664 set_page_refcounted(page);
1665 __free_pages(page, pageblock_order);
1668 adjust_managed_page_count(page, pageblock_nr_pages);
1673 * The order of subdivision here is critical for the IO subsystem.
1674 * Please do not alter this order without good reasons and regression
1675 * testing. Specifically, as large blocks of memory are subdivided,
1676 * the order in which smaller blocks are delivered depends on the order
1677 * they're subdivided in this function. This is the primary factor
1678 * influencing the order in which pages are delivered to the IO
1679 * subsystem according to empirical testing, and this is also justified
1680 * by considering the behavior of a buddy system containing a single
1681 * large block of memory acted on by a series of small allocations.
1682 * This behavior is a critical factor in sglist merging's success.
1686 static inline void expand(struct zone *zone, struct page *page,
1687 int low, int high, struct free_area *area,
1690 unsigned long size = 1 << high;
1692 while (high > low) {
1696 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1699 * Mark as guard pages (or page), that will allow to
1700 * merge back to allocator when buddy will be freed.
1701 * Corresponding page table entries will not be touched,
1702 * pages will stay not present in virtual address space
1704 if (set_page_guard(zone, &page[size], high, migratetype))
1707 list_add(&page[size].lru, &area->free_list[migratetype]);
1709 set_page_order(&page[size], high);
1713 static void check_new_page_bad(struct page *page)
1715 const char *bad_reason = NULL;
1716 unsigned long bad_flags = 0;
1718 if (unlikely(atomic_read(&page->_mapcount) != -1))
1719 bad_reason = "nonzero mapcount";
1720 if (unlikely(page->mapping != NULL))
1721 bad_reason = "non-NULL mapping";
1722 if (unlikely(page_ref_count(page) != 0))
1723 bad_reason = "nonzero _count";
1724 if (unlikely(page->flags & __PG_HWPOISON)) {
1725 bad_reason = "HWPoisoned (hardware-corrupted)";
1726 bad_flags = __PG_HWPOISON;
1727 /* Don't complain about hwpoisoned pages */
1728 page_mapcount_reset(page); /* remove PageBuddy */
1731 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1732 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1733 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1736 if (unlikely(page->mem_cgroup))
1737 bad_reason = "page still charged to cgroup";
1739 bad_page(page, bad_reason, bad_flags);
1743 * This page is about to be returned from the page allocator
1745 static inline int check_new_page(struct page *page)
1747 if (likely(page_expected_state(page,
1748 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1751 check_new_page_bad(page);
1755 static inline bool free_pages_prezeroed(void)
1757 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1758 page_poisoning_enabled();
1761 #ifdef CONFIG_DEBUG_VM
1762 static bool check_pcp_refill(struct page *page)
1767 static bool check_new_pcp(struct page *page)
1769 return check_new_page(page);
1772 static bool check_pcp_refill(struct page *page)
1774 return check_new_page(page);
1776 static bool check_new_pcp(struct page *page)
1780 #endif /* CONFIG_DEBUG_VM */
1782 static bool check_new_pages(struct page *page, unsigned int order)
1785 for (i = 0; i < (1 << order); i++) {
1786 struct page *p = page + i;
1788 if (unlikely(check_new_page(p)))
1795 inline void post_alloc_hook(struct page *page, unsigned int order,
1798 set_page_private(page, 0);
1799 set_page_refcounted(page);
1801 arch_alloc_page(page, order);
1802 kernel_map_pages(page, 1 << order, 1);
1803 kernel_poison_pages(page, 1 << order, 1);
1804 kasan_alloc_pages(page, order);
1805 set_page_owner(page, order, gfp_flags);
1808 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1809 unsigned int alloc_flags)
1813 post_alloc_hook(page, order, gfp_flags);
1815 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1816 for (i = 0; i < (1 << order); i++)
1817 clear_highpage(page + i);
1819 if (order && (gfp_flags & __GFP_COMP))
1820 prep_compound_page(page, order);
1823 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1824 * allocate the page. The expectation is that the caller is taking
1825 * steps that will free more memory. The caller should avoid the page
1826 * being used for !PFMEMALLOC purposes.
1828 if (alloc_flags & ALLOC_NO_WATERMARKS)
1829 set_page_pfmemalloc(page);
1831 clear_page_pfmemalloc(page);
1835 * Go through the free lists for the given migratetype and remove
1836 * the smallest available page from the freelists
1838 static __always_inline
1839 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1842 unsigned int current_order;
1843 struct free_area *area;
1846 /* Find a page of the appropriate size in the preferred list */
1847 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1848 area = &(zone->free_area[current_order]);
1849 page = list_first_entry_or_null(&area->free_list[migratetype],
1853 list_del(&page->lru);
1854 rmv_page_order(page);
1856 expand(zone, page, order, current_order, area, migratetype);
1857 set_pcppage_migratetype(page, migratetype);
1866 * This array describes the order lists are fallen back to when
1867 * the free lists for the desirable migrate type are depleted
1869 static int fallbacks[MIGRATE_TYPES][4] = {
1870 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1871 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1872 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1874 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1876 #ifdef CONFIG_MEMORY_ISOLATION
1877 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1882 static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1885 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1888 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1889 unsigned int order) { return NULL; }
1893 * Move the free pages in a range to the free lists of the requested type.
1894 * Note that start_page and end_pages are not aligned on a pageblock
1895 * boundary. If alignment is required, use move_freepages_block()
1897 static int move_freepages(struct zone *zone,
1898 struct page *start_page, struct page *end_page,
1899 int migratetype, int *num_movable)
1903 int pages_moved = 0;
1905 #ifndef CONFIG_HOLES_IN_ZONE
1907 * page_zone is not safe to call in this context when
1908 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1909 * anyway as we check zone boundaries in move_freepages_block().
1910 * Remove at a later date when no bug reports exist related to
1911 * grouping pages by mobility
1913 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
1914 pfn_valid(page_to_pfn(end_page)) &&
1915 page_zone(start_page) != page_zone(end_page));
1921 for (page = start_page; page <= end_page;) {
1922 if (!pfn_valid_within(page_to_pfn(page))) {
1927 /* Make sure we are not inadvertently changing nodes */
1928 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1930 if (!PageBuddy(page)) {
1932 * We assume that pages that could be isolated for
1933 * migration are movable. But we don't actually try
1934 * isolating, as that would be expensive.
1937 (PageLRU(page) || __PageMovable(page)))
1944 order = page_order(page);
1945 list_move(&page->lru,
1946 &zone->free_area[order].free_list[migratetype]);
1948 pages_moved += 1 << order;
1954 int move_freepages_block(struct zone *zone, struct page *page,
1955 int migratetype, int *num_movable)
1957 unsigned long start_pfn, end_pfn;
1958 struct page *start_page, *end_page;
1960 start_pfn = page_to_pfn(page);
1961 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1962 start_page = pfn_to_page(start_pfn);
1963 end_page = start_page + pageblock_nr_pages - 1;
1964 end_pfn = start_pfn + pageblock_nr_pages - 1;
1966 /* Do not cross zone boundaries */
1967 if (!zone_spans_pfn(zone, start_pfn))
1969 if (!zone_spans_pfn(zone, end_pfn))
1972 return move_freepages(zone, start_page, end_page, migratetype,
1976 static void change_pageblock_range(struct page *pageblock_page,
1977 int start_order, int migratetype)
1979 int nr_pageblocks = 1 << (start_order - pageblock_order);
1981 while (nr_pageblocks--) {
1982 set_pageblock_migratetype(pageblock_page, migratetype);
1983 pageblock_page += pageblock_nr_pages;
1988 * When we are falling back to another migratetype during allocation, try to
1989 * steal extra free pages from the same pageblocks to satisfy further
1990 * allocations, instead of polluting multiple pageblocks.
1992 * If we are stealing a relatively large buddy page, it is likely there will
1993 * be more free pages in the pageblock, so try to steal them all. For
1994 * reclaimable and unmovable allocations, we steal regardless of page size,
1995 * as fragmentation caused by those allocations polluting movable pageblocks
1996 * is worse than movable allocations stealing from unmovable and reclaimable
1999 static bool can_steal_fallback(unsigned int order, int start_mt)
2002 * Leaving this order check is intended, although there is
2003 * relaxed order check in next check. The reason is that
2004 * we can actually steal whole pageblock if this condition met,
2005 * but, below check doesn't guarantee it and that is just heuristic
2006 * so could be changed anytime.
2008 if (order >= pageblock_order)
2011 if (order >= pageblock_order / 2 ||
2012 start_mt == MIGRATE_RECLAIMABLE ||
2013 start_mt == MIGRATE_UNMOVABLE ||
2014 page_group_by_mobility_disabled)
2021 * This function implements actual steal behaviour. If order is large enough,
2022 * we can steal whole pageblock. If not, we first move freepages in this
2023 * pageblock to our migratetype and determine how many already-allocated pages
2024 * are there in the pageblock with a compatible migratetype. If at least half
2025 * of pages are free or compatible, we can change migratetype of the pageblock
2026 * itself, so pages freed in the future will be put on the correct free list.
2028 static void steal_suitable_fallback(struct zone *zone, struct page *page,
2029 int start_type, bool whole_block)
2031 unsigned int current_order = page_order(page);
2032 struct free_area *area;
2033 int free_pages, movable_pages, alike_pages;
2036 old_block_type = get_pageblock_migratetype(page);
2039 * This can happen due to races and we want to prevent broken
2040 * highatomic accounting.
2042 if (is_migrate_highatomic(old_block_type))
2045 /* Take ownership for orders >= pageblock_order */
2046 if (current_order >= pageblock_order) {
2047 change_pageblock_range(page, current_order, start_type);
2051 /* We are not allowed to try stealing from the whole block */
2055 free_pages = move_freepages_block(zone, page, start_type,
2058 * Determine how many pages are compatible with our allocation.
2059 * For movable allocation, it's the number of movable pages which
2060 * we just obtained. For other types it's a bit more tricky.
2062 if (start_type == MIGRATE_MOVABLE) {
2063 alike_pages = movable_pages;
2066 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2067 * to MOVABLE pageblock, consider all non-movable pages as
2068 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2069 * vice versa, be conservative since we can't distinguish the
2070 * exact migratetype of non-movable pages.
2072 if (old_block_type == MIGRATE_MOVABLE)
2073 alike_pages = pageblock_nr_pages
2074 - (free_pages + movable_pages);
2079 /* moving whole block can fail due to zone boundary conditions */
2084 * If a sufficient number of pages in the block are either free or of
2085 * comparable migratability as our allocation, claim the whole block.
2087 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2088 page_group_by_mobility_disabled)
2089 set_pageblock_migratetype(page, start_type);
2094 area = &zone->free_area[current_order];
2095 list_move(&page->lru, &area->free_list[start_type]);
2099 * Check whether there is a suitable fallback freepage with requested order.
2100 * If only_stealable is true, this function returns fallback_mt only if
2101 * we can steal other freepages all together. This would help to reduce
2102 * fragmentation due to mixed migratetype pages in one pageblock.
2104 int find_suitable_fallback(struct free_area *area, unsigned int order,
2105 int migratetype, bool only_stealable, bool *can_steal)
2110 if (area->nr_free == 0)
2115 fallback_mt = fallbacks[migratetype][i];
2116 if (fallback_mt == MIGRATE_TYPES)
2119 if (list_empty(&area->free_list[fallback_mt]))
2122 if (can_steal_fallback(order, migratetype))
2125 if (!only_stealable)
2136 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2137 * there are no empty page blocks that contain a page with a suitable order
2139 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2140 unsigned int alloc_order)
2143 unsigned long max_managed, flags;
2146 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2147 * Check is race-prone but harmless.
2149 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2150 if (zone->nr_reserved_highatomic >= max_managed)
2153 spin_lock_irqsave(&zone->lock, flags);
2155 /* Recheck the nr_reserved_highatomic limit under the lock */
2156 if (zone->nr_reserved_highatomic >= max_managed)
2160 mt = get_pageblock_migratetype(page);
2161 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2162 && !is_migrate_cma(mt)) {
2163 zone->nr_reserved_highatomic += pageblock_nr_pages;
2164 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2165 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2169 spin_unlock_irqrestore(&zone->lock, flags);
2173 * Used when an allocation is about to fail under memory pressure. This
2174 * potentially hurts the reliability of high-order allocations when under
2175 * intense memory pressure but failed atomic allocations should be easier
2176 * to recover from than an OOM.
2178 * If @force is true, try to unreserve a pageblock even though highatomic
2179 * pageblock is exhausted.
2181 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2184 struct zonelist *zonelist = ac->zonelist;
2185 unsigned long flags;
2192 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2195 * Preserve at least one pageblock unless memory pressure
2198 if (!force && zone->nr_reserved_highatomic <=
2202 spin_lock_irqsave(&zone->lock, flags);
2203 for (order = 0; order < MAX_ORDER; order++) {
2204 struct free_area *area = &(zone->free_area[order]);
2206 page = list_first_entry_or_null(
2207 &area->free_list[MIGRATE_HIGHATOMIC],
2213 * In page freeing path, migratetype change is racy so
2214 * we can counter several free pages in a pageblock
2215 * in this loop althoug we changed the pageblock type
2216 * from highatomic to ac->migratetype. So we should
2217 * adjust the count once.
2219 if (is_migrate_highatomic_page(page)) {
2221 * It should never happen but changes to
2222 * locking could inadvertently allow a per-cpu
2223 * drain to add pages to MIGRATE_HIGHATOMIC
2224 * while unreserving so be safe and watch for
2227 zone->nr_reserved_highatomic -= min(
2229 zone->nr_reserved_highatomic);
2233 * Convert to ac->migratetype and avoid the normal
2234 * pageblock stealing heuristics. Minimally, the caller
2235 * is doing the work and needs the pages. More
2236 * importantly, if the block was always converted to
2237 * MIGRATE_UNMOVABLE or another type then the number
2238 * of pageblocks that cannot be completely freed
2241 set_pageblock_migratetype(page, ac->migratetype);
2242 ret = move_freepages_block(zone, page, ac->migratetype,
2245 spin_unlock_irqrestore(&zone->lock, flags);
2249 spin_unlock_irqrestore(&zone->lock, flags);
2256 * Try finding a free buddy page on the fallback list and put it on the free
2257 * list of requested migratetype, possibly along with other pages from the same
2258 * block, depending on fragmentation avoidance heuristics. Returns true if
2259 * fallback was found so that __rmqueue_smallest() can grab it.
2261 * The use of signed ints for order and current_order is a deliberate
2262 * deviation from the rest of this file, to make the for loop
2263 * condition simpler.
2265 static __always_inline bool
2266 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2268 struct free_area *area;
2275 * Find the largest available free page in the other list. This roughly
2276 * approximates finding the pageblock with the most free pages, which
2277 * would be too costly to do exactly.
2279 for (current_order = MAX_ORDER - 1; current_order >= order;
2281 area = &(zone->free_area[current_order]);
2282 fallback_mt = find_suitable_fallback(area, current_order,
2283 start_migratetype, false, &can_steal);
2284 if (fallback_mt == -1)
2288 * We cannot steal all free pages from the pageblock and the
2289 * requested migratetype is movable. In that case it's better to
2290 * steal and split the smallest available page instead of the
2291 * largest available page, because even if the next movable
2292 * allocation falls back into a different pageblock than this
2293 * one, it won't cause permanent fragmentation.
2295 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2296 && current_order > order)
2305 for (current_order = order; current_order < MAX_ORDER;
2307 area = &(zone->free_area[current_order]);
2308 fallback_mt = find_suitable_fallback(area, current_order,
2309 start_migratetype, false, &can_steal);
2310 if (fallback_mt != -1)
2315 * This should not happen - we already found a suitable fallback
2316 * when looking for the largest page.
2318 VM_BUG_ON(current_order == MAX_ORDER);
2321 page = list_first_entry(&area->free_list[fallback_mt],
2324 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2326 trace_mm_page_alloc_extfrag(page, order, current_order,
2327 start_migratetype, fallback_mt);
2334 * Do the hard work of removing an element from the buddy allocator.
2335 * Call me with the zone->lock already held.
2337 static __always_inline struct page *
2338 __rmqueue(struct zone *zone, unsigned int order, int migratetype)
2343 page = __rmqueue_smallest(zone, order, migratetype);
2344 if (unlikely(!page)) {
2345 if (migratetype == MIGRATE_MOVABLE)
2346 page = __rmqueue_cma_fallback(zone, order);
2348 if (!page && __rmqueue_fallback(zone, order, migratetype))
2352 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2357 * Obtain a specified number of elements from the buddy allocator, all under
2358 * a single hold of the lock, for efficiency. Add them to the supplied list.
2359 * Returns the number of new pages which were placed at *list.
2361 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2362 unsigned long count, struct list_head *list,
2367 spin_lock(&zone->lock);
2368 for (i = 0; i < count; ++i) {
2369 struct page *page = __rmqueue(zone, order, migratetype);
2370 if (unlikely(page == NULL))
2373 if (unlikely(check_pcp_refill(page)))
2377 * Split buddy pages returned by expand() are received here in
2378 * physical page order. The page is added to the tail of
2379 * caller's list. From the callers perspective, the linked list
2380 * is ordered by page number under some conditions. This is
2381 * useful for IO devices that can forward direction from the
2382 * head, thus also in the physical page order. This is useful
2383 * for IO devices that can merge IO requests if the physical
2384 * pages are ordered properly.
2386 list_add_tail(&page->lru, list);
2388 if (is_migrate_cma(get_pcppage_migratetype(page)))
2389 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2394 * i pages were removed from the buddy list even if some leak due
2395 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2396 * on i. Do not confuse with 'alloced' which is the number of
2397 * pages added to the pcp list.
2399 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2400 spin_unlock(&zone->lock);
2406 * Called from the vmstat counter updater to drain pagesets of this
2407 * currently executing processor on remote nodes after they have
2410 * Note that this function must be called with the thread pinned to
2411 * a single processor.
2413 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2415 unsigned long flags;
2416 int to_drain, batch;
2418 local_irq_save(flags);
2419 batch = READ_ONCE(pcp->batch);
2420 to_drain = min(pcp->count, batch);
2422 free_pcppages_bulk(zone, to_drain, pcp);
2423 pcp->count -= to_drain;
2425 local_irq_restore(flags);
2430 * Drain pcplists of the indicated processor and zone.
2432 * The processor must either be the current processor and the
2433 * thread pinned to the current processor or a processor that
2436 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2438 unsigned long flags;
2439 struct per_cpu_pageset *pset;
2440 struct per_cpu_pages *pcp;
2442 local_irq_save(flags);
2443 pset = per_cpu_ptr(zone->pageset, cpu);
2447 free_pcppages_bulk(zone, pcp->count, pcp);
2450 local_irq_restore(flags);
2454 * Drain pcplists of all zones on the indicated processor.
2456 * The processor must either be the current processor and the
2457 * thread pinned to the current processor or a processor that
2460 static void drain_pages(unsigned int cpu)
2464 for_each_populated_zone(zone) {
2465 drain_pages_zone(cpu, zone);
2470 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2472 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2473 * the single zone's pages.
2475 void drain_local_pages(struct zone *zone)
2477 int cpu = smp_processor_id();
2480 drain_pages_zone(cpu, zone);
2485 static void drain_local_pages_wq(struct work_struct *work)
2488 * drain_all_pages doesn't use proper cpu hotplug protection so
2489 * we can race with cpu offline when the WQ can move this from
2490 * a cpu pinned worker to an unbound one. We can operate on a different
2491 * cpu which is allright but we also have to make sure to not move to
2495 drain_local_pages(NULL);
2500 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2502 * When zone parameter is non-NULL, spill just the single zone's pages.
2504 * Note that this can be extremely slow as the draining happens in a workqueue.
2506 void drain_all_pages(struct zone *zone)
2511 * Allocate in the BSS so we wont require allocation in
2512 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2514 static cpumask_t cpus_with_pcps;
2517 * Make sure nobody triggers this path before mm_percpu_wq is fully
2520 if (WARN_ON_ONCE(!mm_percpu_wq))
2524 * Do not drain if one is already in progress unless it's specific to
2525 * a zone. Such callers are primarily CMA and memory hotplug and need
2526 * the drain to be complete when the call returns.
2528 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2531 mutex_lock(&pcpu_drain_mutex);
2535 * We don't care about racing with CPU hotplug event
2536 * as offline notification will cause the notified
2537 * cpu to drain that CPU pcps and on_each_cpu_mask
2538 * disables preemption as part of its processing
2540 for_each_online_cpu(cpu) {
2541 struct per_cpu_pageset *pcp;
2543 bool has_pcps = false;
2546 pcp = per_cpu_ptr(zone->pageset, cpu);
2550 for_each_populated_zone(z) {
2551 pcp = per_cpu_ptr(z->pageset, cpu);
2552 if (pcp->pcp.count) {
2560 cpumask_set_cpu(cpu, &cpus_with_pcps);
2562 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2565 for_each_cpu(cpu, &cpus_with_pcps) {
2566 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2567 INIT_WORK(work, drain_local_pages_wq);
2568 queue_work_on(cpu, mm_percpu_wq, work);
2570 for_each_cpu(cpu, &cpus_with_pcps)
2571 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2573 mutex_unlock(&pcpu_drain_mutex);
2576 #ifdef CONFIG_HIBERNATION
2579 * Touch the watchdog for every WD_PAGE_COUNT pages.
2581 #define WD_PAGE_COUNT (128*1024)
2583 void mark_free_pages(struct zone *zone)
2585 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2586 unsigned long flags;
2587 unsigned int order, t;
2590 if (zone_is_empty(zone))
2593 spin_lock_irqsave(&zone->lock, flags);
2595 max_zone_pfn = zone_end_pfn(zone);
2596 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2597 if (pfn_valid(pfn)) {
2598 page = pfn_to_page(pfn);
2600 if (!--page_count) {
2601 touch_nmi_watchdog();
2602 page_count = WD_PAGE_COUNT;
2605 if (page_zone(page) != zone)
2608 if (!swsusp_page_is_forbidden(page))
2609 swsusp_unset_page_free(page);
2612 for_each_migratetype_order(order, t) {
2613 list_for_each_entry(page,
2614 &zone->free_area[order].free_list[t], lru) {
2617 pfn = page_to_pfn(page);
2618 for (i = 0; i < (1UL << order); i++) {
2619 if (!--page_count) {
2620 touch_nmi_watchdog();
2621 page_count = WD_PAGE_COUNT;
2623 swsusp_set_page_free(pfn_to_page(pfn + i));
2627 spin_unlock_irqrestore(&zone->lock, flags);
2629 #endif /* CONFIG_PM */
2631 static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2635 if (!free_pcp_prepare(page))
2638 migratetype = get_pfnblock_migratetype(page, pfn);
2639 set_pcppage_migratetype(page, migratetype);
2643 static void free_unref_page_commit(struct page *page, unsigned long pfn)
2645 struct zone *zone = page_zone(page);
2646 struct per_cpu_pages *pcp;
2649 migratetype = get_pcppage_migratetype(page);
2650 __count_vm_event(PGFREE);
2653 * We only track unmovable, reclaimable and movable on pcp lists.
2654 * Free ISOLATE pages back to the allocator because they are being
2655 * offlined but treat HIGHATOMIC as movable pages so we can get those
2656 * areas back if necessary. Otherwise, we may have to free
2657 * excessively into the page allocator
2659 if (migratetype >= MIGRATE_PCPTYPES) {
2660 if (unlikely(is_migrate_isolate(migratetype))) {
2661 free_one_page(zone, page, pfn, 0, migratetype);
2664 migratetype = MIGRATE_MOVABLE;
2667 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2668 list_add(&page->lru, &pcp->lists[migratetype]);
2670 if (pcp->count >= pcp->high) {
2671 unsigned long batch = READ_ONCE(pcp->batch);
2672 free_pcppages_bulk(zone, batch, pcp);
2673 pcp->count -= batch;
2678 * Free a 0-order page
2680 void free_unref_page(struct page *page)
2682 unsigned long flags;
2683 unsigned long pfn = page_to_pfn(page);
2685 if (!free_unref_page_prepare(page, pfn))
2688 local_irq_save(flags);
2689 free_unref_page_commit(page, pfn);
2690 local_irq_restore(flags);
2694 * Free a list of 0-order pages
2696 void free_unref_page_list(struct list_head *list)
2698 struct page *page, *next;
2699 unsigned long flags, pfn;
2700 int batch_count = 0;
2702 /* Prepare pages for freeing */
2703 list_for_each_entry_safe(page, next, list, lru) {
2704 pfn = page_to_pfn(page);
2705 if (!free_unref_page_prepare(page, pfn))
2706 list_del(&page->lru);
2707 set_page_private(page, pfn);
2710 local_irq_save(flags);
2711 list_for_each_entry_safe(page, next, list, lru) {
2712 unsigned long pfn = page_private(page);
2714 set_page_private(page, 0);
2715 trace_mm_page_free_batched(page);
2716 free_unref_page_commit(page, pfn);
2719 * Guard against excessive IRQ disabled times when we get
2720 * a large list of pages to free.
2722 if (++batch_count == SWAP_CLUSTER_MAX) {
2723 local_irq_restore(flags);
2725 local_irq_save(flags);
2728 local_irq_restore(flags);
2732 * split_page takes a non-compound higher-order page, and splits it into
2733 * n (1<<order) sub-pages: page[0..n]
2734 * Each sub-page must be freed individually.
2736 * Note: this is probably too low level an operation for use in drivers.
2737 * Please consult with lkml before using this in your driver.
2739 void split_page(struct page *page, unsigned int order)
2743 VM_BUG_ON_PAGE(PageCompound(page), page);
2744 VM_BUG_ON_PAGE(!page_count(page), page);
2746 for (i = 1; i < (1 << order); i++)
2747 set_page_refcounted(page + i);
2748 split_page_owner(page, order);
2750 EXPORT_SYMBOL_GPL(split_page);
2752 int __isolate_free_page(struct page *page, unsigned int order)
2754 unsigned long watermark;
2758 BUG_ON(!PageBuddy(page));
2760 zone = page_zone(page);
2761 mt = get_pageblock_migratetype(page);
2763 if (!is_migrate_isolate(mt)) {
2765 * Obey watermarks as if the page was being allocated. We can
2766 * emulate a high-order watermark check with a raised order-0
2767 * watermark, because we already know our high-order page
2770 watermark = min_wmark_pages(zone) + (1UL << order);
2771 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2774 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2777 /* Remove page from free list */
2778 list_del(&page->lru);
2779 zone->free_area[order].nr_free--;
2780 rmv_page_order(page);
2783 * Set the pageblock if the isolated page is at least half of a
2786 if (order >= pageblock_order - 1) {
2787 struct page *endpage = page + (1 << order) - 1;
2788 for (; page < endpage; page += pageblock_nr_pages) {
2789 int mt = get_pageblock_migratetype(page);
2790 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2791 && !is_migrate_highatomic(mt))
2792 set_pageblock_migratetype(page,
2798 return 1UL << order;
2802 * Update NUMA hit/miss statistics
2804 * Must be called with interrupts disabled.
2806 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2809 enum numa_stat_item local_stat = NUMA_LOCAL;
2811 /* skip numa counters update if numa stats is disabled */
2812 if (!static_branch_likely(&vm_numa_stat_key))
2815 if (z->node != numa_node_id())
2816 local_stat = NUMA_OTHER;
2818 if (z->node == preferred_zone->node)
2819 __inc_numa_state(z, NUMA_HIT);
2821 __inc_numa_state(z, NUMA_MISS);
2822 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2824 __inc_numa_state(z, local_stat);
2828 /* Remove page from the per-cpu list, caller must protect the list */
2829 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2830 struct per_cpu_pages *pcp,
2831 struct list_head *list)
2836 if (list_empty(list)) {
2837 pcp->count += rmqueue_bulk(zone, 0,
2840 if (unlikely(list_empty(list)))
2844 page = list_first_entry(list, struct page, lru);
2845 list_del(&page->lru);
2847 } while (check_new_pcp(page));
2852 /* Lock and remove page from the per-cpu list */
2853 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2854 struct zone *zone, unsigned int order,
2855 gfp_t gfp_flags, int migratetype)
2857 struct per_cpu_pages *pcp;
2858 struct list_head *list;
2860 unsigned long flags;
2862 local_irq_save(flags);
2863 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2864 list = &pcp->lists[migratetype];
2865 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2867 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2868 zone_statistics(preferred_zone, zone);
2870 local_irq_restore(flags);
2875 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2878 struct page *rmqueue(struct zone *preferred_zone,
2879 struct zone *zone, unsigned int order,
2880 gfp_t gfp_flags, unsigned int alloc_flags,
2883 unsigned long flags;
2886 if (likely(order == 0)) {
2887 page = rmqueue_pcplist(preferred_zone, zone, order,
2888 gfp_flags, migratetype);
2893 * We most definitely don't want callers attempting to
2894 * allocate greater than order-1 page units with __GFP_NOFAIL.
2896 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2897 spin_lock_irqsave(&zone->lock, flags);
2901 if (alloc_flags & ALLOC_HARDER) {
2902 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2904 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2907 page = __rmqueue(zone, order, migratetype);
2908 } while (page && check_new_pages(page, order));
2909 spin_unlock(&zone->lock);
2912 __mod_zone_freepage_state(zone, -(1 << order),
2913 get_pcppage_migratetype(page));
2915 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2916 zone_statistics(preferred_zone, zone);
2917 local_irq_restore(flags);
2920 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2924 local_irq_restore(flags);
2928 #ifdef CONFIG_FAIL_PAGE_ALLOC
2931 struct fault_attr attr;
2933 bool ignore_gfp_highmem;
2934 bool ignore_gfp_reclaim;
2936 } fail_page_alloc = {
2937 .attr = FAULT_ATTR_INITIALIZER,
2938 .ignore_gfp_reclaim = true,
2939 .ignore_gfp_highmem = true,
2943 static int __init setup_fail_page_alloc(char *str)
2945 return setup_fault_attr(&fail_page_alloc.attr, str);
2947 __setup("fail_page_alloc=", setup_fail_page_alloc);
2949 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2951 if (order < fail_page_alloc.min_order)
2953 if (gfp_mask & __GFP_NOFAIL)
2955 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2957 if (fail_page_alloc.ignore_gfp_reclaim &&
2958 (gfp_mask & __GFP_DIRECT_RECLAIM))
2961 return should_fail(&fail_page_alloc.attr, 1 << order);
2964 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2966 static int __init fail_page_alloc_debugfs(void)
2968 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2971 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2972 &fail_page_alloc.attr);
2974 return PTR_ERR(dir);
2976 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2977 &fail_page_alloc.ignore_gfp_reclaim))
2979 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2980 &fail_page_alloc.ignore_gfp_highmem))
2982 if (!debugfs_create_u32("min-order", mode, dir,
2983 &fail_page_alloc.min_order))
2988 debugfs_remove_recursive(dir);
2993 late_initcall(fail_page_alloc_debugfs);
2995 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2997 #else /* CONFIG_FAIL_PAGE_ALLOC */
2999 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3004 #endif /* CONFIG_FAIL_PAGE_ALLOC */
3007 * Return true if free base pages are above 'mark'. For high-order checks it
3008 * will return true of the order-0 watermark is reached and there is at least
3009 * one free page of a suitable size. Checking now avoids taking the zone lock
3010 * to check in the allocation paths if no pages are free.
3012 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3013 int classzone_idx, unsigned int alloc_flags,
3018 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3020 /* free_pages may go negative - that's OK */
3021 free_pages -= (1 << order) - 1;
3023 if (alloc_flags & ALLOC_HIGH)
3027 * If the caller does not have rights to ALLOC_HARDER then subtract
3028 * the high-atomic reserves. This will over-estimate the size of the
3029 * atomic reserve but it avoids a search.
3031 if (likely(!alloc_harder)) {
3032 free_pages -= z->nr_reserved_highatomic;
3035 * OOM victims can try even harder than normal ALLOC_HARDER
3036 * users on the grounds that it's definitely going to be in
3037 * the exit path shortly and free memory. Any allocation it
3038 * makes during the free path will be small and short-lived.
3040 if (alloc_flags & ALLOC_OOM)
3048 /* If allocation can't use CMA areas don't use free CMA pages */
3049 if (!(alloc_flags & ALLOC_CMA))
3050 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3054 * Check watermarks for an order-0 allocation request. If these
3055 * are not met, then a high-order request also cannot go ahead
3056 * even if a suitable page happened to be free.
3058 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3061 /* If this is an order-0 request then the watermark is fine */
3065 /* For a high-order request, check at least one suitable page is free */
3066 for (o = order; o < MAX_ORDER; o++) {
3067 struct free_area *area = &z->free_area[o];
3073 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3074 if (!list_empty(&area->free_list[mt]))
3079 if ((alloc_flags & ALLOC_CMA) &&
3080 !list_empty(&area->free_list[MIGRATE_CMA])) {
3085 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3091 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3092 int classzone_idx, unsigned int alloc_flags)
3094 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3095 zone_page_state(z, NR_FREE_PAGES));
3098 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3099 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3101 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3105 /* If allocation can't use CMA areas don't use free CMA pages */
3106 if (!(alloc_flags & ALLOC_CMA))
3107 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3111 * Fast check for order-0 only. If this fails then the reserves
3112 * need to be calculated. There is a corner case where the check
3113 * passes but only the high-order atomic reserve are free. If
3114 * the caller is !atomic then it'll uselessly search the free
3115 * list. That corner case is then slower but it is harmless.
3117 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3120 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3124 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3125 unsigned long mark, int classzone_idx)
3127 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3129 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3130 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3132 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3137 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3139 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3142 #else /* CONFIG_NUMA */
3143 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3147 #endif /* CONFIG_NUMA */
3150 * get_page_from_freelist goes through the zonelist trying to allocate
3153 static struct page *
3154 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3155 const struct alloc_context *ac)
3157 struct zoneref *z = ac->preferred_zoneref;
3159 struct pglist_data *last_pgdat_dirty_limit = NULL;
3162 * Scan zonelist, looking for a zone with enough free.
3163 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3165 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3170 if (cpusets_enabled() &&
3171 (alloc_flags & ALLOC_CPUSET) &&
3172 !__cpuset_zone_allowed(zone, gfp_mask))
3175 * When allocating a page cache page for writing, we
3176 * want to get it from a node that is within its dirty
3177 * limit, such that no single node holds more than its
3178 * proportional share of globally allowed dirty pages.
3179 * The dirty limits take into account the node's
3180 * lowmem reserves and high watermark so that kswapd
3181 * should be able to balance it without having to
3182 * write pages from its LRU list.
3184 * XXX: For now, allow allocations to potentially
3185 * exceed the per-node dirty limit in the slowpath
3186 * (spread_dirty_pages unset) before going into reclaim,
3187 * which is important when on a NUMA setup the allowed
3188 * nodes are together not big enough to reach the
3189 * global limit. The proper fix for these situations
3190 * will require awareness of nodes in the
3191 * dirty-throttling and the flusher threads.
3193 if (ac->spread_dirty_pages) {
3194 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3197 if (!node_dirty_ok(zone->zone_pgdat)) {
3198 last_pgdat_dirty_limit = zone->zone_pgdat;
3203 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3204 if (!zone_watermark_fast(zone, order, mark,
3205 ac_classzone_idx(ac), alloc_flags)) {
3208 /* Checked here to keep the fast path fast */
3209 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3210 if (alloc_flags & ALLOC_NO_WATERMARKS)
3213 if (node_reclaim_mode == 0 ||
3214 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3217 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3219 case NODE_RECLAIM_NOSCAN:
3222 case NODE_RECLAIM_FULL:
3223 /* scanned but unreclaimable */
3226 /* did we reclaim enough */
3227 if (zone_watermark_ok(zone, order, mark,
3228 ac_classzone_idx(ac), alloc_flags))
3236 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3237 gfp_mask, alloc_flags, ac->migratetype);
3239 prep_new_page(page, order, gfp_mask, alloc_flags);
3242 * If this is a high-order atomic allocation then check
3243 * if the pageblock should be reserved for the future
3245 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3246 reserve_highatomic_pageblock(page, zone, order);
3256 * Large machines with many possible nodes should not always dump per-node
3257 * meminfo in irq context.
3259 static inline bool should_suppress_show_mem(void)
3264 ret = in_interrupt();
3269 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3271 unsigned int filter = SHOW_MEM_FILTER_NODES;
3272 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3274 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3278 * This documents exceptions given to allocations in certain
3279 * contexts that are allowed to allocate outside current's set
3282 if (!(gfp_mask & __GFP_NOMEMALLOC))
3283 if (tsk_is_oom_victim(current) ||
3284 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3285 filter &= ~SHOW_MEM_FILTER_NODES;
3286 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3287 filter &= ~SHOW_MEM_FILTER_NODES;
3289 show_mem(filter, nodemask);
3292 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3294 struct va_format vaf;
3296 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3297 DEFAULT_RATELIMIT_BURST);
3299 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3302 va_start(args, fmt);
3305 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3306 current->comm, &vaf, gfp_mask, &gfp_mask,
3307 nodemask_pr_args(nodemask));
3310 cpuset_print_current_mems_allowed();
3313 warn_alloc_show_mem(gfp_mask, nodemask);
3316 static inline struct page *
3317 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3318 unsigned int alloc_flags,
3319 const struct alloc_context *ac)
3323 page = get_page_from_freelist(gfp_mask, order,
3324 alloc_flags|ALLOC_CPUSET, ac);
3326 * fallback to ignore cpuset restriction if our nodes
3330 page = get_page_from_freelist(gfp_mask, order,
3336 static inline struct page *
3337 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3338 const struct alloc_context *ac, unsigned long *did_some_progress)
3340 struct oom_control oc = {
3341 .zonelist = ac->zonelist,
3342 .nodemask = ac->nodemask,
3344 .gfp_mask = gfp_mask,
3349 *did_some_progress = 0;
3352 * Acquire the oom lock. If that fails, somebody else is
3353 * making progress for us.
3355 if (!mutex_trylock(&oom_lock)) {
3356 *did_some_progress = 1;
3357 schedule_timeout_uninterruptible(1);
3362 * Go through the zonelist yet one more time, keep very high watermark
3363 * here, this is only to catch a parallel oom killing, we must fail if
3364 * we're still under heavy pressure. But make sure that this reclaim
3365 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3366 * allocation which will never fail due to oom_lock already held.
3368 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3369 ~__GFP_DIRECT_RECLAIM, order,
3370 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3374 /* Coredumps can quickly deplete all memory reserves */
3375 if (current->flags & PF_DUMPCORE)
3377 /* The OOM killer will not help higher order allocs */
3378 if (order > PAGE_ALLOC_COSTLY_ORDER)
3381 * We have already exhausted all our reclaim opportunities without any
3382 * success so it is time to admit defeat. We will skip the OOM killer
3383 * because it is very likely that the caller has a more reasonable
3384 * fallback than shooting a random task.
3386 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3388 /* The OOM killer does not needlessly kill tasks for lowmem */
3389 if (ac->high_zoneidx < ZONE_NORMAL)
3391 if (pm_suspended_storage())
3394 * XXX: GFP_NOFS allocations should rather fail than rely on
3395 * other request to make a forward progress.
3396 * We are in an unfortunate situation where out_of_memory cannot
3397 * do much for this context but let's try it to at least get
3398 * access to memory reserved if the current task is killed (see
3399 * out_of_memory). Once filesystems are ready to handle allocation
3400 * failures more gracefully we should just bail out here.
3403 /* The OOM killer may not free memory on a specific node */
3404 if (gfp_mask & __GFP_THISNODE)
3407 /* Exhausted what can be done so it's blame time */
3408 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3409 *did_some_progress = 1;
3412 * Help non-failing allocations by giving them access to memory
3415 if (gfp_mask & __GFP_NOFAIL)
3416 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3417 ALLOC_NO_WATERMARKS, ac);
3420 mutex_unlock(&oom_lock);
3425 * Maximum number of compaction retries wit a progress before OOM
3426 * killer is consider as the only way to move forward.
3428 #define MAX_COMPACT_RETRIES 16
3430 #ifdef CONFIG_COMPACTION
3431 /* Try memory compaction for high-order allocations before reclaim */
3432 static struct page *
3433 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3434 unsigned int alloc_flags, const struct alloc_context *ac,
3435 enum compact_priority prio, enum compact_result *compact_result)
3438 unsigned int noreclaim_flag;
3443 noreclaim_flag = memalloc_noreclaim_save();
3444 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3446 memalloc_noreclaim_restore(noreclaim_flag);
3448 if (*compact_result <= COMPACT_INACTIVE)
3452 * At least in one zone compaction wasn't deferred or skipped, so let's
3453 * count a compaction stall
3455 count_vm_event(COMPACTSTALL);
3457 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3460 struct zone *zone = page_zone(page);
3462 zone->compact_blockskip_flush = false;
3463 compaction_defer_reset(zone, order, true);
3464 count_vm_event(COMPACTSUCCESS);
3469 * It's bad if compaction run occurs and fails. The most likely reason
3470 * is that pages exist, but not enough to satisfy watermarks.
3472 count_vm_event(COMPACTFAIL);
3480 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3481 enum compact_result compact_result,
3482 enum compact_priority *compact_priority,
3483 int *compaction_retries)
3485 int max_retries = MAX_COMPACT_RETRIES;
3488 int retries = *compaction_retries;
3489 enum compact_priority priority = *compact_priority;
3494 if (compaction_made_progress(compact_result))
3495 (*compaction_retries)++;
3498 * compaction considers all the zone as desperately out of memory
3499 * so it doesn't really make much sense to retry except when the
3500 * failure could be caused by insufficient priority
3502 if (compaction_failed(compact_result))
3503 goto check_priority;
3506 * make sure the compaction wasn't deferred or didn't bail out early
3507 * due to locks contention before we declare that we should give up.
3508 * But do not retry if the given zonelist is not suitable for
3511 if (compaction_withdrawn(compact_result)) {
3512 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3517 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3518 * costly ones because they are de facto nofail and invoke OOM
3519 * killer to move on while costly can fail and users are ready
3520 * to cope with that. 1/4 retries is rather arbitrary but we
3521 * would need much more detailed feedback from compaction to
3522 * make a better decision.
3524 if (order > PAGE_ALLOC_COSTLY_ORDER)
3526 if (*compaction_retries <= max_retries) {
3532 * Make sure there are attempts at the highest priority if we exhausted
3533 * all retries or failed at the lower priorities.
3536 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3537 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3539 if (*compact_priority > min_priority) {
3540 (*compact_priority)--;
3541 *compaction_retries = 0;
3545 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3549 static inline struct page *
3550 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3551 unsigned int alloc_flags, const struct alloc_context *ac,
3552 enum compact_priority prio, enum compact_result *compact_result)
3554 *compact_result = COMPACT_SKIPPED;
3559 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3560 enum compact_result compact_result,
3561 enum compact_priority *compact_priority,
3562 int *compaction_retries)
3567 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3571 * There are setups with compaction disabled which would prefer to loop
3572 * inside the allocator rather than hit the oom killer prematurely.
3573 * Let's give them a good hope and keep retrying while the order-0
3574 * watermarks are OK.
3576 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3578 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3579 ac_classzone_idx(ac), alloc_flags))
3584 #endif /* CONFIG_COMPACTION */
3586 #ifdef CONFIG_LOCKDEP
3587 struct lockdep_map __fs_reclaim_map =
3588 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3590 static bool __need_fs_reclaim(gfp_t gfp_mask)
3592 gfp_mask = current_gfp_context(gfp_mask);
3594 /* no reclaim without waiting on it */
3595 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3598 /* this guy won't enter reclaim */
3599 if (current->flags & PF_MEMALLOC)
3602 /* We're only interested __GFP_FS allocations for now */
3603 if (!(gfp_mask & __GFP_FS))
3606 if (gfp_mask & __GFP_NOLOCKDEP)
3612 void fs_reclaim_acquire(gfp_t gfp_mask)
3614 if (__need_fs_reclaim(gfp_mask))
3615 lock_map_acquire(&__fs_reclaim_map);
3617 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3619 void fs_reclaim_release(gfp_t gfp_mask)
3621 if (__need_fs_reclaim(gfp_mask))
3622 lock_map_release(&__fs_reclaim_map);
3624 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3627 /* Perform direct synchronous page reclaim */
3629 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3630 const struct alloc_context *ac)
3632 struct reclaim_state reclaim_state;
3634 unsigned int noreclaim_flag;
3638 /* We now go into synchronous reclaim */
3639 cpuset_memory_pressure_bump();
3640 noreclaim_flag = memalloc_noreclaim_save();
3641 fs_reclaim_acquire(gfp_mask);
3642 reclaim_state.reclaimed_slab = 0;
3643 current->reclaim_state = &reclaim_state;
3645 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3648 current->reclaim_state = NULL;
3649 fs_reclaim_release(gfp_mask);
3650 memalloc_noreclaim_restore(noreclaim_flag);
3657 /* The really slow allocator path where we enter direct reclaim */
3658 static inline struct page *
3659 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3660 unsigned int alloc_flags, const struct alloc_context *ac,
3661 unsigned long *did_some_progress)
3663 struct page *page = NULL;
3664 bool drained = false;
3666 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3667 if (unlikely(!(*did_some_progress)))
3671 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3674 * If an allocation failed after direct reclaim, it could be because
3675 * pages are pinned on the per-cpu lists or in high alloc reserves.
3676 * Shrink them them and try again
3678 if (!page && !drained) {
3679 unreserve_highatomic_pageblock(ac, false);
3680 drain_all_pages(NULL);
3688 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3692 pg_data_t *last_pgdat = NULL;
3694 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3695 ac->high_zoneidx, ac->nodemask) {
3696 if (last_pgdat != zone->zone_pgdat)
3697 wakeup_kswapd(zone, order, ac->high_zoneidx);
3698 last_pgdat = zone->zone_pgdat;
3702 static inline unsigned int
3703 gfp_to_alloc_flags(gfp_t gfp_mask)
3705 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3707 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3708 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3711 * The caller may dip into page reserves a bit more if the caller
3712 * cannot run direct reclaim, or if the caller has realtime scheduling
3713 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3714 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3716 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3718 if (gfp_mask & __GFP_ATOMIC) {
3720 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3721 * if it can't schedule.
3723 if (!(gfp_mask & __GFP_NOMEMALLOC))
3724 alloc_flags |= ALLOC_HARDER;
3726 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3727 * comment for __cpuset_node_allowed().
3729 alloc_flags &= ~ALLOC_CPUSET;
3730 } else if (unlikely(rt_task(current)) && !in_interrupt())
3731 alloc_flags |= ALLOC_HARDER;
3734 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3735 alloc_flags |= ALLOC_CMA;
3740 static bool oom_reserves_allowed(struct task_struct *tsk)
3742 if (!tsk_is_oom_victim(tsk))
3746 * !MMU doesn't have oom reaper so give access to memory reserves
3747 * only to the thread with TIF_MEMDIE set
3749 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3756 * Distinguish requests which really need access to full memory
3757 * reserves from oom victims which can live with a portion of it
3759 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3761 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3763 if (gfp_mask & __GFP_MEMALLOC)
3764 return ALLOC_NO_WATERMARKS;
3765 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3766 return ALLOC_NO_WATERMARKS;
3767 if (!in_interrupt()) {
3768 if (current->flags & PF_MEMALLOC)
3769 return ALLOC_NO_WATERMARKS;
3770 else if (oom_reserves_allowed(current))
3777 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3779 return !!__gfp_pfmemalloc_flags(gfp_mask);
3783 * Checks whether it makes sense to retry the reclaim to make a forward progress
3784 * for the given allocation request.
3786 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3787 * without success, or when we couldn't even meet the watermark if we
3788 * reclaimed all remaining pages on the LRU lists.
3790 * Returns true if a retry is viable or false to enter the oom path.
3793 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3794 struct alloc_context *ac, int alloc_flags,
3795 bool did_some_progress, int *no_progress_loops)
3801 * Costly allocations might have made a progress but this doesn't mean
3802 * their order will become available due to high fragmentation so
3803 * always increment the no progress counter for them
3805 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3806 *no_progress_loops = 0;
3808 (*no_progress_loops)++;
3811 * Make sure we converge to OOM if we cannot make any progress
3812 * several times in the row.
3814 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3815 /* Before OOM, exhaust highatomic_reserve */
3816 return unreserve_highatomic_pageblock(ac, true);
3820 * Keep reclaiming pages while there is a chance this will lead
3821 * somewhere. If none of the target zones can satisfy our allocation
3822 * request even if all reclaimable pages are considered then we are
3823 * screwed and have to go OOM.
3825 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3827 unsigned long available;
3828 unsigned long reclaimable;
3829 unsigned long min_wmark = min_wmark_pages(zone);
3832 available = reclaimable = zone_reclaimable_pages(zone);
3833 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3836 * Would the allocation succeed if we reclaimed all
3837 * reclaimable pages?
3839 wmark = __zone_watermark_ok(zone, order, min_wmark,
3840 ac_classzone_idx(ac), alloc_flags, available);
3841 trace_reclaim_retry_zone(z, order, reclaimable,
3842 available, min_wmark, *no_progress_loops, wmark);
3845 * If we didn't make any progress and have a lot of
3846 * dirty + writeback pages then we should wait for
3847 * an IO to complete to slow down the reclaim and
3848 * prevent from pre mature OOM
3850 if (!did_some_progress) {
3851 unsigned long write_pending;
3853 write_pending = zone_page_state_snapshot(zone,
3854 NR_ZONE_WRITE_PENDING);
3856 if (2 * write_pending > reclaimable) {
3857 congestion_wait(BLK_RW_ASYNC, HZ/10);
3863 * Memory allocation/reclaim might be called from a WQ
3864 * context and the current implementation of the WQ
3865 * concurrency control doesn't recognize that
3866 * a particular WQ is congested if the worker thread is
3867 * looping without ever sleeping. Therefore we have to
3868 * do a short sleep here rather than calling
3871 if (current->flags & PF_WQ_WORKER)
3872 schedule_timeout_uninterruptible(1);
3884 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3887 * It's possible that cpuset's mems_allowed and the nodemask from
3888 * mempolicy don't intersect. This should be normally dealt with by
3889 * policy_nodemask(), but it's possible to race with cpuset update in
3890 * such a way the check therein was true, and then it became false
3891 * before we got our cpuset_mems_cookie here.
3892 * This assumes that for all allocations, ac->nodemask can come only
3893 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3894 * when it does not intersect with the cpuset restrictions) or the
3895 * caller can deal with a violated nodemask.
3897 if (cpusets_enabled() && ac->nodemask &&
3898 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3899 ac->nodemask = NULL;
3904 * When updating a task's mems_allowed or mempolicy nodemask, it is
3905 * possible to race with parallel threads in such a way that our
3906 * allocation can fail while the mask is being updated. If we are about
3907 * to fail, check if the cpuset changed during allocation and if so,
3910 if (read_mems_allowed_retry(cpuset_mems_cookie))
3916 static inline struct page *
3917 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3918 struct alloc_context *ac)
3920 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3921 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3922 struct page *page = NULL;
3923 unsigned int alloc_flags;
3924 unsigned long did_some_progress;
3925 enum compact_priority compact_priority;
3926 enum compact_result compact_result;
3927 int compaction_retries;
3928 int no_progress_loops;
3929 unsigned int cpuset_mems_cookie;
3933 * In the slowpath, we sanity check order to avoid ever trying to
3934 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3935 * be using allocators in order of preference for an area that is
3938 if (order >= MAX_ORDER) {
3939 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3944 * We also sanity check to catch abuse of atomic reserves being used by
3945 * callers that are not in atomic context.
3947 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3948 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3949 gfp_mask &= ~__GFP_ATOMIC;
3952 compaction_retries = 0;
3953 no_progress_loops = 0;
3954 compact_priority = DEF_COMPACT_PRIORITY;
3955 cpuset_mems_cookie = read_mems_allowed_begin();
3958 * The fast path uses conservative alloc_flags to succeed only until
3959 * kswapd needs to be woken up, and to avoid the cost of setting up
3960 * alloc_flags precisely. So we do that now.
3962 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3965 * We need to recalculate the starting point for the zonelist iterator
3966 * because we might have used different nodemask in the fast path, or
3967 * there was a cpuset modification and we are retrying - otherwise we
3968 * could end up iterating over non-eligible zones endlessly.
3970 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3971 ac->high_zoneidx, ac->nodemask);
3972 if (!ac->preferred_zoneref->zone)
3975 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3976 wake_all_kswapds(order, ac);
3979 * The adjusted alloc_flags might result in immediate success, so try
3982 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3987 * For costly allocations, try direct compaction first, as it's likely
3988 * that we have enough base pages and don't need to reclaim. For non-
3989 * movable high-order allocations, do that as well, as compaction will
3990 * try prevent permanent fragmentation by migrating from blocks of the
3992 * Don't try this for allocations that are allowed to ignore
3993 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3995 if (can_direct_reclaim &&
3997 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3998 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3999 page = __alloc_pages_direct_compact(gfp_mask, order,
4001 INIT_COMPACT_PRIORITY,
4007 * Checks for costly allocations with __GFP_NORETRY, which
4008 * includes THP page fault allocations
4010 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4012 * If compaction is deferred for high-order allocations,
4013 * it is because sync compaction recently failed. If
4014 * this is the case and the caller requested a THP
4015 * allocation, we do not want to heavily disrupt the
4016 * system, so we fail the allocation instead of entering
4019 if (compact_result == COMPACT_DEFERRED)
4023 * Looks like reclaim/compaction is worth trying, but
4024 * sync compaction could be very expensive, so keep
4025 * using async compaction.
4027 compact_priority = INIT_COMPACT_PRIORITY;
4032 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4033 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4034 wake_all_kswapds(order, ac);
4036 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4038 alloc_flags = reserve_flags;
4041 * Reset the zonelist iterators if memory policies can be ignored.
4042 * These allocations are high priority and system rather than user
4045 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4046 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
4047 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4048 ac->high_zoneidx, ac->nodemask);
4051 /* Attempt with potentially adjusted zonelist and alloc_flags */
4052 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4056 /* Caller is not willing to reclaim, we can't balance anything */
4057 if (!can_direct_reclaim)
4060 /* Avoid recursion of direct reclaim */
4061 if (current->flags & PF_MEMALLOC)
4064 /* Try direct reclaim and then allocating */
4065 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4066 &did_some_progress);
4070 /* Try direct compaction and then allocating */
4071 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4072 compact_priority, &compact_result);
4076 /* Do not loop if specifically requested */
4077 if (gfp_mask & __GFP_NORETRY)
4081 * Do not retry costly high order allocations unless they are
4082 * __GFP_RETRY_MAYFAIL
4084 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4087 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4088 did_some_progress > 0, &no_progress_loops))
4092 * It doesn't make any sense to retry for the compaction if the order-0
4093 * reclaim is not able to make any progress because the current
4094 * implementation of the compaction depends on the sufficient amount
4095 * of free memory (see __compaction_suitable)
4097 if (did_some_progress > 0 &&
4098 should_compact_retry(ac, order, alloc_flags,
4099 compact_result, &compact_priority,
4100 &compaction_retries))
4104 /* Deal with possible cpuset update races before we start OOM killing */
4105 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4108 /* Reclaim has failed us, start killing things */
4109 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4113 /* Avoid allocations with no watermarks from looping endlessly */
4114 if (tsk_is_oom_victim(current) &&
4115 (alloc_flags == ALLOC_OOM ||
4116 (gfp_mask & __GFP_NOMEMALLOC)))
4119 /* Retry as long as the OOM killer is making progress */
4120 if (did_some_progress) {
4121 no_progress_loops = 0;
4126 /* Deal with possible cpuset update races before we fail */
4127 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4131 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4134 if (gfp_mask & __GFP_NOFAIL) {
4136 * All existing users of the __GFP_NOFAIL are blockable, so warn
4137 * of any new users that actually require GFP_NOWAIT
4139 if (WARN_ON_ONCE(!can_direct_reclaim))
4143 * PF_MEMALLOC request from this context is rather bizarre
4144 * because we cannot reclaim anything and only can loop waiting
4145 * for somebody to do a work for us
4147 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4150 * non failing costly orders are a hard requirement which we
4151 * are not prepared for much so let's warn about these users
4152 * so that we can identify them and convert them to something
4155 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4158 * Help non-failing allocations by giving them access to memory
4159 * reserves but do not use ALLOC_NO_WATERMARKS because this
4160 * could deplete whole memory reserves which would just make
4161 * the situation worse
4163 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4171 warn_alloc(gfp_mask, ac->nodemask,
4172 "page allocation failure: order:%u", order);
4177 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4178 int preferred_nid, nodemask_t *nodemask,
4179 struct alloc_context *ac, gfp_t *alloc_mask,
4180 unsigned int *alloc_flags)
4182 ac->high_zoneidx = gfp_zone(gfp_mask);
4183 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4184 ac->nodemask = nodemask;
4185 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4187 if (cpusets_enabled()) {
4188 *alloc_mask |= __GFP_HARDWALL;
4190 ac->nodemask = &cpuset_current_mems_allowed;
4192 *alloc_flags |= ALLOC_CPUSET;
4195 fs_reclaim_acquire(gfp_mask);
4196 fs_reclaim_release(gfp_mask);
4198 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4200 if (should_fail_alloc_page(gfp_mask, order))
4203 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4204 *alloc_flags |= ALLOC_CMA;
4209 /* Determine whether to spread dirty pages and what the first usable zone */
4210 static inline void finalise_ac(gfp_t gfp_mask,
4211 unsigned int order, struct alloc_context *ac)
4213 /* Dirty zone balancing only done in the fast path */
4214 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4217 * The preferred zone is used for statistics but crucially it is
4218 * also used as the starting point for the zonelist iterator. It
4219 * may get reset for allocations that ignore memory policies.
4221 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4222 ac->high_zoneidx, ac->nodemask);
4226 * This is the 'heart' of the zoned buddy allocator.
4229 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4230 nodemask_t *nodemask)
4233 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4234 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4235 struct alloc_context ac = { };
4237 gfp_mask &= gfp_allowed_mask;
4238 alloc_mask = gfp_mask;
4239 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4242 finalise_ac(gfp_mask, order, &ac);
4244 /* First allocation attempt */
4245 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4250 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4251 * resp. GFP_NOIO which has to be inherited for all allocation requests
4252 * from a particular context which has been marked by
4253 * memalloc_no{fs,io}_{save,restore}.
4255 alloc_mask = current_gfp_context(gfp_mask);
4256 ac.spread_dirty_pages = false;
4259 * Restore the original nodemask if it was potentially replaced with
4260 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4262 if (unlikely(ac.nodemask != nodemask))
4263 ac.nodemask = nodemask;
4265 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4268 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4269 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4270 __free_pages(page, order);
4274 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4278 EXPORT_SYMBOL(__alloc_pages_nodemask);
4281 * Common helper functions.
4283 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4288 * __get_free_pages() returns a virtual address, which cannot represent
4291 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4293 page = alloc_pages(gfp_mask, order);
4296 return (unsigned long) page_address(page);
4298 EXPORT_SYMBOL(__get_free_pages);
4300 unsigned long get_zeroed_page(gfp_t gfp_mask)
4302 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4304 EXPORT_SYMBOL(get_zeroed_page);
4306 void __free_pages(struct page *page, unsigned int order)
4308 if (put_page_testzero(page)) {
4310 free_unref_page(page);
4312 __free_pages_ok(page, order);
4316 EXPORT_SYMBOL(__free_pages);
4318 void free_pages(unsigned long addr, unsigned int order)
4321 VM_BUG_ON(!virt_addr_valid((void *)addr));
4322 __free_pages(virt_to_page((void *)addr), order);
4326 EXPORT_SYMBOL(free_pages);
4330 * An arbitrary-length arbitrary-offset area of memory which resides
4331 * within a 0 or higher order page. Multiple fragments within that page
4332 * are individually refcounted, in the page's reference counter.
4334 * The page_frag functions below provide a simple allocation framework for
4335 * page fragments. This is used by the network stack and network device
4336 * drivers to provide a backing region of memory for use as either an
4337 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4339 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4342 struct page *page = NULL;
4343 gfp_t gfp = gfp_mask;
4345 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4346 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4348 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4349 PAGE_FRAG_CACHE_MAX_ORDER);
4350 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4352 if (unlikely(!page))
4353 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4355 nc->va = page ? page_address(page) : NULL;
4360 void __page_frag_cache_drain(struct page *page, unsigned int count)
4362 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4364 if (page_ref_sub_and_test(page, count)) {
4365 unsigned int order = compound_order(page);
4368 free_unref_page(page);
4370 __free_pages_ok(page, order);
4373 EXPORT_SYMBOL(__page_frag_cache_drain);
4375 void *page_frag_alloc(struct page_frag_cache *nc,
4376 unsigned int fragsz, gfp_t gfp_mask)
4378 unsigned int size = PAGE_SIZE;
4382 if (unlikely(!nc->va)) {
4384 page = __page_frag_cache_refill(nc, gfp_mask);
4388 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4389 /* if size can vary use size else just use PAGE_SIZE */
4392 /* Even if we own the page, we do not use atomic_set().
4393 * This would break get_page_unless_zero() users.
4395 page_ref_add(page, size - 1);
4397 /* reset page count bias and offset to start of new frag */
4398 nc->pfmemalloc = page_is_pfmemalloc(page);
4399 nc->pagecnt_bias = size;
4403 offset = nc->offset - fragsz;
4404 if (unlikely(offset < 0)) {
4405 page = virt_to_page(nc->va);
4407 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4410 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4411 /* if size can vary use size else just use PAGE_SIZE */
4414 /* OK, page count is 0, we can safely set it */
4415 set_page_count(page, size);
4417 /* reset page count bias and offset to start of new frag */
4418 nc->pagecnt_bias = size;
4419 offset = size - fragsz;
4423 nc->offset = offset;
4425 return nc->va + offset;
4427 EXPORT_SYMBOL(page_frag_alloc);
4430 * Frees a page fragment allocated out of either a compound or order 0 page.
4432 void page_frag_free(void *addr)
4434 struct page *page = virt_to_head_page(addr);
4436 if (unlikely(put_page_testzero(page)))
4437 __free_pages_ok(page, compound_order(page));
4439 EXPORT_SYMBOL(page_frag_free);
4441 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4445 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4446 unsigned long used = addr + PAGE_ALIGN(size);
4448 split_page(virt_to_page((void *)addr), order);
4449 while (used < alloc_end) {
4454 return (void *)addr;
4458 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4459 * @size: the number of bytes to allocate
4460 * @gfp_mask: GFP flags for the allocation
4462 * This function is similar to alloc_pages(), except that it allocates the
4463 * minimum number of pages to satisfy the request. alloc_pages() can only
4464 * allocate memory in power-of-two pages.
4466 * This function is also limited by MAX_ORDER.
4468 * Memory allocated by this function must be released by free_pages_exact().
4470 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4472 unsigned int order = get_order(size);
4475 addr = __get_free_pages(gfp_mask, order);
4476 return make_alloc_exact(addr, order, size);
4478 EXPORT_SYMBOL(alloc_pages_exact);
4481 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4483 * @nid: the preferred node ID where memory should be allocated
4484 * @size: the number of bytes to allocate
4485 * @gfp_mask: GFP flags for the allocation
4487 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4490 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4492 unsigned int order = get_order(size);
4493 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4496 return make_alloc_exact((unsigned long)page_address(p), order, size);
4500 * free_pages_exact - release memory allocated via alloc_pages_exact()
4501 * @virt: the value returned by alloc_pages_exact.
4502 * @size: size of allocation, same value as passed to alloc_pages_exact().
4504 * Release the memory allocated by a previous call to alloc_pages_exact.
4506 void free_pages_exact(void *virt, size_t size)
4508 unsigned long addr = (unsigned long)virt;
4509 unsigned long end = addr + PAGE_ALIGN(size);
4511 while (addr < end) {
4516 EXPORT_SYMBOL(free_pages_exact);
4519 * nr_free_zone_pages - count number of pages beyond high watermark
4520 * @offset: The zone index of the highest zone
4522 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4523 * high watermark within all zones at or below a given zone index. For each
4524 * zone, the number of pages is calculated as:
4526 * nr_free_zone_pages = managed_pages - high_pages
4528 static unsigned long nr_free_zone_pages(int offset)
4533 /* Just pick one node, since fallback list is circular */
4534 unsigned long sum = 0;
4536 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4538 for_each_zone_zonelist(zone, z, zonelist, offset) {
4539 unsigned long size = zone->managed_pages;
4540 unsigned long high = high_wmark_pages(zone);
4549 * nr_free_buffer_pages - count number of pages beyond high watermark
4551 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4552 * watermark within ZONE_DMA and ZONE_NORMAL.
4554 unsigned long nr_free_buffer_pages(void)
4556 return nr_free_zone_pages(gfp_zone(GFP_USER));
4558 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4561 * nr_free_pagecache_pages - count number of pages beyond high watermark
4563 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4564 * high watermark within all zones.
4566 unsigned long nr_free_pagecache_pages(void)
4568 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4571 static inline void show_node(struct zone *zone)
4573 if (IS_ENABLED(CONFIG_NUMA))
4574 printk("Node %d ", zone_to_nid(zone));
4577 long si_mem_available(void)
4580 unsigned long pagecache;
4581 unsigned long wmark_low = 0;
4582 unsigned long pages[NR_LRU_LISTS];
4586 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4587 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4590 wmark_low += zone->watermark[WMARK_LOW];
4593 * Estimate the amount of memory available for userspace allocations,
4594 * without causing swapping.
4596 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4599 * Not all the page cache can be freed, otherwise the system will
4600 * start swapping. Assume at least half of the page cache, or the
4601 * low watermark worth of cache, needs to stay.
4603 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4604 pagecache -= min(pagecache / 2, wmark_low);
4605 available += pagecache;
4608 * Part of the reclaimable slab consists of items that are in use,
4609 * and cannot be freed. Cap this estimate at the low watermark.
4611 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4612 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4619 EXPORT_SYMBOL_GPL(si_mem_available);
4621 void si_meminfo(struct sysinfo *val)
4623 val->totalram = totalram_pages;
4624 val->sharedram = global_node_page_state(NR_SHMEM);
4625 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4626 val->bufferram = nr_blockdev_pages();
4627 val->totalhigh = totalhigh_pages;
4628 val->freehigh = nr_free_highpages();
4629 val->mem_unit = PAGE_SIZE;
4632 EXPORT_SYMBOL(si_meminfo);
4635 void si_meminfo_node(struct sysinfo *val, int nid)
4637 int zone_type; /* needs to be signed */
4638 unsigned long managed_pages = 0;
4639 unsigned long managed_highpages = 0;
4640 unsigned long free_highpages = 0;
4641 pg_data_t *pgdat = NODE_DATA(nid);
4643 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4644 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4645 val->totalram = managed_pages;
4646 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4647 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4648 #ifdef CONFIG_HIGHMEM
4649 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4650 struct zone *zone = &pgdat->node_zones[zone_type];
4652 if (is_highmem(zone)) {
4653 managed_highpages += zone->managed_pages;
4654 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4657 val->totalhigh = managed_highpages;
4658 val->freehigh = free_highpages;
4660 val->totalhigh = managed_highpages;
4661 val->freehigh = free_highpages;
4663 val->mem_unit = PAGE_SIZE;
4668 * Determine whether the node should be displayed or not, depending on whether
4669 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4671 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4673 if (!(flags & SHOW_MEM_FILTER_NODES))
4677 * no node mask - aka implicit memory numa policy. Do not bother with
4678 * the synchronization - read_mems_allowed_begin - because we do not
4679 * have to be precise here.
4682 nodemask = &cpuset_current_mems_allowed;
4684 return !node_isset(nid, *nodemask);
4687 #define K(x) ((x) << (PAGE_SHIFT-10))
4689 static void show_migration_types(unsigned char type)
4691 static const char types[MIGRATE_TYPES] = {
4692 [MIGRATE_UNMOVABLE] = 'U',
4693 [MIGRATE_MOVABLE] = 'M',
4694 [MIGRATE_RECLAIMABLE] = 'E',
4695 [MIGRATE_HIGHATOMIC] = 'H',
4697 [MIGRATE_CMA] = 'C',
4699 #ifdef CONFIG_MEMORY_ISOLATION
4700 [MIGRATE_ISOLATE] = 'I',
4703 char tmp[MIGRATE_TYPES + 1];
4707 for (i = 0; i < MIGRATE_TYPES; i++) {
4708 if (type & (1 << i))
4713 printk(KERN_CONT "(%s) ", tmp);
4717 * Show free area list (used inside shift_scroll-lock stuff)
4718 * We also calculate the percentage fragmentation. We do this by counting the
4719 * memory on each free list with the exception of the first item on the list.
4722 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4725 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4727 unsigned long free_pcp = 0;
4732 for_each_populated_zone(zone) {
4733 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4736 for_each_online_cpu(cpu)
4737 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4740 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4741 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4742 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4743 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4744 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4745 " free:%lu free_pcp:%lu free_cma:%lu\n",
4746 global_node_page_state(NR_ACTIVE_ANON),
4747 global_node_page_state(NR_INACTIVE_ANON),
4748 global_node_page_state(NR_ISOLATED_ANON),
4749 global_node_page_state(NR_ACTIVE_FILE),
4750 global_node_page_state(NR_INACTIVE_FILE),
4751 global_node_page_state(NR_ISOLATED_FILE),
4752 global_node_page_state(NR_UNEVICTABLE),
4753 global_node_page_state(NR_FILE_DIRTY),
4754 global_node_page_state(NR_WRITEBACK),
4755 global_node_page_state(NR_UNSTABLE_NFS),
4756 global_node_page_state(NR_SLAB_RECLAIMABLE),
4757 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4758 global_node_page_state(NR_FILE_MAPPED),
4759 global_node_page_state(NR_SHMEM),
4760 global_zone_page_state(NR_PAGETABLE),
4761 global_zone_page_state(NR_BOUNCE),
4762 global_zone_page_state(NR_FREE_PAGES),
4764 global_zone_page_state(NR_FREE_CMA_PAGES));
4766 for_each_online_pgdat(pgdat) {
4767 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4771 " active_anon:%lukB"
4772 " inactive_anon:%lukB"
4773 " active_file:%lukB"
4774 " inactive_file:%lukB"
4775 " unevictable:%lukB"
4776 " isolated(anon):%lukB"
4777 " isolated(file):%lukB"
4782 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4784 " shmem_pmdmapped: %lukB"
4787 " writeback_tmp:%lukB"
4789 " all_unreclaimable? %s"
4792 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4793 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4794 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4795 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4796 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4797 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4798 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4799 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4800 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4801 K(node_page_state(pgdat, NR_WRITEBACK)),
4802 K(node_page_state(pgdat, NR_SHMEM)),
4803 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4804 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4805 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4807 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4809 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4810 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4811 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4815 for_each_populated_zone(zone) {
4818 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4822 for_each_online_cpu(cpu)
4823 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4832 " active_anon:%lukB"
4833 " inactive_anon:%lukB"
4834 " active_file:%lukB"
4835 " inactive_file:%lukB"
4836 " unevictable:%lukB"
4837 " writepending:%lukB"
4841 " kernel_stack:%lukB"
4849 K(zone_page_state(zone, NR_FREE_PAGES)),
4850 K(min_wmark_pages(zone)),
4851 K(low_wmark_pages(zone)),
4852 K(high_wmark_pages(zone)),
4853 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4854 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4855 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4856 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4857 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4858 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4859 K(zone->present_pages),
4860 K(zone->managed_pages),
4861 K(zone_page_state(zone, NR_MLOCK)),
4862 zone_page_state(zone, NR_KERNEL_STACK_KB),
4863 K(zone_page_state(zone, NR_PAGETABLE)),
4864 K(zone_page_state(zone, NR_BOUNCE)),
4866 K(this_cpu_read(zone->pageset->pcp.count)),
4867 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4868 printk("lowmem_reserve[]:");
4869 for (i = 0; i < MAX_NR_ZONES; i++)
4870 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4871 printk(KERN_CONT "\n");
4874 for_each_populated_zone(zone) {
4876 unsigned long nr[MAX_ORDER], flags, total = 0;
4877 unsigned char types[MAX_ORDER];
4879 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4882 printk(KERN_CONT "%s: ", zone->name);
4884 spin_lock_irqsave(&zone->lock, flags);
4885 for (order = 0; order < MAX_ORDER; order++) {
4886 struct free_area *area = &zone->free_area[order];
4889 nr[order] = area->nr_free;
4890 total += nr[order] << order;
4893 for (type = 0; type < MIGRATE_TYPES; type++) {
4894 if (!list_empty(&area->free_list[type]))
4895 types[order] |= 1 << type;
4898 spin_unlock_irqrestore(&zone->lock, flags);
4899 for (order = 0; order < MAX_ORDER; order++) {
4900 printk(KERN_CONT "%lu*%lukB ",
4901 nr[order], K(1UL) << order);
4903 show_migration_types(types[order]);
4905 printk(KERN_CONT "= %lukB\n", K(total));
4908 hugetlb_show_meminfo();
4910 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4912 show_swap_cache_info();
4915 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4917 zoneref->zone = zone;
4918 zoneref->zone_idx = zone_idx(zone);
4922 * Builds allocation fallback zone lists.
4924 * Add all populated zones of a node to the zonelist.
4926 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4929 enum zone_type zone_type = MAX_NR_ZONES;
4934 zone = pgdat->node_zones + zone_type;
4935 if (managed_zone(zone)) {
4936 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4937 check_highest_zone(zone_type);
4939 } while (zone_type);
4946 static int __parse_numa_zonelist_order(char *s)
4949 * We used to support different zonlists modes but they turned
4950 * out to be just not useful. Let's keep the warning in place
4951 * if somebody still use the cmd line parameter so that we do
4952 * not fail it silently
4954 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4955 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4961 static __init int setup_numa_zonelist_order(char *s)
4966 return __parse_numa_zonelist_order(s);
4968 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4970 char numa_zonelist_order[] = "Node";
4973 * sysctl handler for numa_zonelist_order
4975 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4976 void __user *buffer, size_t *length,
4983 return proc_dostring(table, write, buffer, length, ppos);
4984 str = memdup_user_nul(buffer, 16);
4986 return PTR_ERR(str);
4988 ret = __parse_numa_zonelist_order(str);
4994 #define MAX_NODE_LOAD (nr_online_nodes)
4995 static int node_load[MAX_NUMNODES];
4998 * find_next_best_node - find the next node that should appear in a given node's fallback list
4999 * @node: node whose fallback list we're appending
5000 * @used_node_mask: nodemask_t of already used nodes
5002 * We use a number of factors to determine which is the next node that should
5003 * appear on a given node's fallback list. The node should not have appeared
5004 * already in @node's fallback list, and it should be the next closest node
5005 * according to the distance array (which contains arbitrary distance values
5006 * from each node to each node in the system), and should also prefer nodes
5007 * with no CPUs, since presumably they'll have very little allocation pressure
5008 * on them otherwise.
5009 * It returns -1 if no node is found.
5011 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5014 int min_val = INT_MAX;
5015 int best_node = NUMA_NO_NODE;
5016 const struct cpumask *tmp = cpumask_of_node(0);
5018 /* Use the local node if we haven't already */
5019 if (!node_isset(node, *used_node_mask)) {
5020 node_set(node, *used_node_mask);
5024 for_each_node_state(n, N_MEMORY) {
5026 /* Don't want a node to appear more than once */
5027 if (node_isset(n, *used_node_mask))
5030 /* Use the distance array to find the distance */
5031 val = node_distance(node, n);
5033 /* Penalize nodes under us ("prefer the next node") */
5036 /* Give preference to headless and unused nodes */
5037 tmp = cpumask_of_node(n);
5038 if (!cpumask_empty(tmp))
5039 val += PENALTY_FOR_NODE_WITH_CPUS;
5041 /* Slight preference for less loaded node */
5042 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5043 val += node_load[n];
5045 if (val < min_val) {
5052 node_set(best_node, *used_node_mask);
5059 * Build zonelists ordered by node and zones within node.
5060 * This results in maximum locality--normal zone overflows into local
5061 * DMA zone, if any--but risks exhausting DMA zone.
5063 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5066 struct zoneref *zonerefs;
5069 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5071 for (i = 0; i < nr_nodes; i++) {
5074 pg_data_t *node = NODE_DATA(node_order[i]);
5076 nr_zones = build_zonerefs_node(node, zonerefs);
5077 zonerefs += nr_zones;
5079 zonerefs->zone = NULL;
5080 zonerefs->zone_idx = 0;
5084 * Build gfp_thisnode zonelists
5086 static void build_thisnode_zonelists(pg_data_t *pgdat)
5088 struct zoneref *zonerefs;
5091 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5092 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5093 zonerefs += nr_zones;
5094 zonerefs->zone = NULL;
5095 zonerefs->zone_idx = 0;
5099 * Build zonelists ordered by zone and nodes within zones.
5100 * This results in conserving DMA zone[s] until all Normal memory is
5101 * exhausted, but results in overflowing to remote node while memory
5102 * may still exist in local DMA zone.
5105 static void build_zonelists(pg_data_t *pgdat)
5107 static int node_order[MAX_NUMNODES];
5108 int node, load, nr_nodes = 0;
5109 nodemask_t used_mask;
5110 int local_node, prev_node;
5112 /* NUMA-aware ordering of nodes */
5113 local_node = pgdat->node_id;
5114 load = nr_online_nodes;
5115 prev_node = local_node;
5116 nodes_clear(used_mask);
5118 memset(node_order, 0, sizeof(node_order));
5119 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5121 * We don't want to pressure a particular node.
5122 * So adding penalty to the first node in same
5123 * distance group to make it round-robin.
5125 if (node_distance(local_node, node) !=
5126 node_distance(local_node, prev_node))
5127 node_load[node] = load;
5129 node_order[nr_nodes++] = node;
5134 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5135 build_thisnode_zonelists(pgdat);
5138 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5140 * Return node id of node used for "local" allocations.
5141 * I.e., first node id of first zone in arg node's generic zonelist.
5142 * Used for initializing percpu 'numa_mem', which is used primarily
5143 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5145 int local_memory_node(int node)
5149 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5150 gfp_zone(GFP_KERNEL),
5152 return z->zone->node;
5156 static void setup_min_unmapped_ratio(void);
5157 static void setup_min_slab_ratio(void);
5158 #else /* CONFIG_NUMA */
5160 static void build_zonelists(pg_data_t *pgdat)
5162 int node, local_node;
5163 struct zoneref *zonerefs;
5166 local_node = pgdat->node_id;
5168 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5169 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5170 zonerefs += nr_zones;
5173 * Now we build the zonelist so that it contains the zones
5174 * of all the other nodes.
5175 * We don't want to pressure a particular node, so when
5176 * building the zones for node N, we make sure that the
5177 * zones coming right after the local ones are those from
5178 * node N+1 (modulo N)
5180 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5181 if (!node_online(node))
5183 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5184 zonerefs += nr_zones;
5186 for (node = 0; node < local_node; node++) {
5187 if (!node_online(node))
5189 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5190 zonerefs += nr_zones;
5193 zonerefs->zone = NULL;
5194 zonerefs->zone_idx = 0;
5197 #endif /* CONFIG_NUMA */
5200 * Boot pageset table. One per cpu which is going to be used for all
5201 * zones and all nodes. The parameters will be set in such a way
5202 * that an item put on a list will immediately be handed over to
5203 * the buddy list. This is safe since pageset manipulation is done
5204 * with interrupts disabled.
5206 * The boot_pagesets must be kept even after bootup is complete for
5207 * unused processors and/or zones. They do play a role for bootstrapping
5208 * hotplugged processors.
5210 * zoneinfo_show() and maybe other functions do
5211 * not check if the processor is online before following the pageset pointer.
5212 * Other parts of the kernel may not check if the zone is available.
5214 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5215 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5216 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5218 static void __build_all_zonelists(void *data)
5221 int __maybe_unused cpu;
5222 pg_data_t *self = data;
5223 static DEFINE_SPINLOCK(lock);
5228 memset(node_load, 0, sizeof(node_load));
5232 * This node is hotadded and no memory is yet present. So just
5233 * building zonelists is fine - no need to touch other nodes.
5235 if (self && !node_online(self->node_id)) {
5236 build_zonelists(self);
5238 for_each_online_node(nid) {
5239 pg_data_t *pgdat = NODE_DATA(nid);
5241 build_zonelists(pgdat);
5244 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5246 * We now know the "local memory node" for each node--
5247 * i.e., the node of the first zone in the generic zonelist.
5248 * Set up numa_mem percpu variable for on-line cpus. During
5249 * boot, only the boot cpu should be on-line; we'll init the
5250 * secondary cpus' numa_mem as they come on-line. During
5251 * node/memory hotplug, we'll fixup all on-line cpus.
5253 for_each_online_cpu(cpu)
5254 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5261 static noinline void __init
5262 build_all_zonelists_init(void)
5266 __build_all_zonelists(NULL);
5269 * Initialize the boot_pagesets that are going to be used
5270 * for bootstrapping processors. The real pagesets for
5271 * each zone will be allocated later when the per cpu
5272 * allocator is available.
5274 * boot_pagesets are used also for bootstrapping offline
5275 * cpus if the system is already booted because the pagesets
5276 * are needed to initialize allocators on a specific cpu too.
5277 * F.e. the percpu allocator needs the page allocator which
5278 * needs the percpu allocator in order to allocate its pagesets
5279 * (a chicken-egg dilemma).
5281 for_each_possible_cpu(cpu)
5282 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5284 mminit_verify_zonelist();
5285 cpuset_init_current_mems_allowed();
5289 * unless system_state == SYSTEM_BOOTING.
5291 * __ref due to call of __init annotated helper build_all_zonelists_init
5292 * [protected by SYSTEM_BOOTING].
5294 void __ref build_all_zonelists(pg_data_t *pgdat)
5296 if (system_state == SYSTEM_BOOTING) {
5297 build_all_zonelists_init();
5299 __build_all_zonelists(pgdat);
5300 /* cpuset refresh routine should be here */
5302 vm_total_pages = nr_free_pagecache_pages();
5304 * Disable grouping by mobility if the number of pages in the
5305 * system is too low to allow the mechanism to work. It would be
5306 * more accurate, but expensive to check per-zone. This check is
5307 * made on memory-hotadd so a system can start with mobility
5308 * disabled and enable it later
5310 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5311 page_group_by_mobility_disabled = 1;
5313 page_group_by_mobility_disabled = 0;
5315 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5317 page_group_by_mobility_disabled ? "off" : "on",
5320 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5325 * Initially all pages are reserved - free ones are freed
5326 * up by free_all_bootmem() once the early boot process is
5327 * done. Non-atomic initialization, single-pass.
5329 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5330 unsigned long start_pfn, enum memmap_context context,
5331 struct vmem_altmap *altmap)
5333 unsigned long end_pfn = start_pfn + size;
5334 pg_data_t *pgdat = NODE_DATA(nid);
5336 unsigned long nr_initialised = 0;
5337 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5338 struct memblock_region *r = NULL, *tmp;
5341 if (highest_memmap_pfn < end_pfn - 1)
5342 highest_memmap_pfn = end_pfn - 1;
5345 * Honor reservation requested by the driver for this ZONE_DEVICE
5348 if (altmap && start_pfn == altmap->base_pfn)
5349 start_pfn += altmap->reserve;
5351 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5353 * There can be holes in boot-time mem_map[]s handed to this
5354 * function. They do not exist on hotplugged memory.
5356 if (context != MEMMAP_EARLY)
5359 if (!early_pfn_valid(pfn))
5361 if (!early_pfn_in_nid(pfn, nid))
5363 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5366 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5368 * Check given memblock attribute by firmware which can affect
5369 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5370 * mirrored, it's an overlapped memmap init. skip it.
5372 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5373 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5374 for_each_memblock(memory, tmp)
5375 if (pfn < memblock_region_memory_end_pfn(tmp))
5379 if (pfn >= memblock_region_memory_base_pfn(r) &&
5380 memblock_is_mirror(r)) {
5381 /* already initialized as NORMAL */
5382 pfn = memblock_region_memory_end_pfn(r);
5390 * Mark the block movable so that blocks are reserved for
5391 * movable at startup. This will force kernel allocations
5392 * to reserve their blocks rather than leaking throughout
5393 * the address space during boot when many long-lived
5394 * kernel allocations are made.
5396 * bitmap is created for zone's valid pfn range. but memmap
5397 * can be created for invalid pages (for alignment)
5398 * check here not to call set_pageblock_migratetype() against
5401 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5402 * because this is done early in sparse_add_one_section
5404 if (!(pfn & (pageblock_nr_pages - 1))) {
5405 struct page *page = pfn_to_page(pfn);
5407 __init_single_page(page, pfn, zone, nid,
5408 context != MEMMAP_HOTPLUG);
5409 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5412 __init_single_pfn(pfn, zone, nid,
5413 context != MEMMAP_HOTPLUG);
5418 static void __meminit zone_init_free_lists(struct zone *zone)
5420 unsigned int order, t;
5421 for_each_migratetype_order(order, t) {
5422 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5423 zone->free_area[order].nr_free = 0;
5427 #ifndef __HAVE_ARCH_MEMMAP_INIT
5428 #define memmap_init(size, nid, zone, start_pfn) \
5429 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY, NULL)
5432 static int zone_batchsize(struct zone *zone)
5438 * The per-cpu-pages pools are set to around 1000th of the
5439 * size of the zone. But no more than 1/2 of a meg.
5441 * OK, so we don't know how big the cache is. So guess.
5443 batch = zone->managed_pages / 1024;
5444 if (batch * PAGE_SIZE > 512 * 1024)
5445 batch = (512 * 1024) / PAGE_SIZE;
5446 batch /= 4; /* We effectively *= 4 below */
5451 * Clamp the batch to a 2^n - 1 value. Having a power
5452 * of 2 value was found to be more likely to have
5453 * suboptimal cache aliasing properties in some cases.
5455 * For example if 2 tasks are alternately allocating
5456 * batches of pages, one task can end up with a lot
5457 * of pages of one half of the possible page colors
5458 * and the other with pages of the other colors.
5460 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5465 /* The deferral and batching of frees should be suppressed under NOMMU
5468 * The problem is that NOMMU needs to be able to allocate large chunks
5469 * of contiguous memory as there's no hardware page translation to
5470 * assemble apparent contiguous memory from discontiguous pages.
5472 * Queueing large contiguous runs of pages for batching, however,
5473 * causes the pages to actually be freed in smaller chunks. As there
5474 * can be a significant delay between the individual batches being
5475 * recycled, this leads to the once large chunks of space being
5476 * fragmented and becoming unavailable for high-order allocations.
5483 * pcp->high and pcp->batch values are related and dependent on one another:
5484 * ->batch must never be higher then ->high.
5485 * The following function updates them in a safe manner without read side
5488 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5489 * those fields changing asynchronously (acording the the above rule).
5491 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5492 * outside of boot time (or some other assurance that no concurrent updaters
5495 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5496 unsigned long batch)
5498 /* start with a fail safe value for batch */
5502 /* Update high, then batch, in order */
5509 /* a companion to pageset_set_high() */
5510 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5512 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5515 static void pageset_init(struct per_cpu_pageset *p)
5517 struct per_cpu_pages *pcp;
5520 memset(p, 0, sizeof(*p));
5524 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5525 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5528 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5531 pageset_set_batch(p, batch);
5535 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5536 * to the value high for the pageset p.
5538 static void pageset_set_high(struct per_cpu_pageset *p,
5541 unsigned long batch = max(1UL, high / 4);
5542 if ((high / 4) > (PAGE_SHIFT * 8))
5543 batch = PAGE_SHIFT * 8;
5545 pageset_update(&p->pcp, high, batch);
5548 static void pageset_set_high_and_batch(struct zone *zone,
5549 struct per_cpu_pageset *pcp)
5551 if (percpu_pagelist_fraction)
5552 pageset_set_high(pcp,
5553 (zone->managed_pages /
5554 percpu_pagelist_fraction));
5556 pageset_set_batch(pcp, zone_batchsize(zone));
5559 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5561 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5564 pageset_set_high_and_batch(zone, pcp);
5567 void __meminit setup_zone_pageset(struct zone *zone)
5570 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5571 for_each_possible_cpu(cpu)
5572 zone_pageset_init(zone, cpu);
5576 * Allocate per cpu pagesets and initialize them.
5577 * Before this call only boot pagesets were available.
5579 void __init setup_per_cpu_pageset(void)
5581 struct pglist_data *pgdat;
5584 for_each_populated_zone(zone)
5585 setup_zone_pageset(zone);
5587 for_each_online_pgdat(pgdat)
5588 pgdat->per_cpu_nodestats =
5589 alloc_percpu(struct per_cpu_nodestat);
5592 static __meminit void zone_pcp_init(struct zone *zone)
5595 * per cpu subsystem is not up at this point. The following code
5596 * relies on the ability of the linker to provide the
5597 * offset of a (static) per cpu variable into the per cpu area.
5599 zone->pageset = &boot_pageset;
5601 if (populated_zone(zone))
5602 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5603 zone->name, zone->present_pages,
5604 zone_batchsize(zone));
5607 void __meminit init_currently_empty_zone(struct zone *zone,
5608 unsigned long zone_start_pfn,
5611 struct pglist_data *pgdat = zone->zone_pgdat;
5613 pgdat->nr_zones = zone_idx(zone) + 1;
5615 zone->zone_start_pfn = zone_start_pfn;
5617 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5618 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5620 (unsigned long)zone_idx(zone),
5621 zone_start_pfn, (zone_start_pfn + size));
5623 zone_init_free_lists(zone);
5624 zone->initialized = 1;
5627 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5628 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5631 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5633 int __meminit __early_pfn_to_nid(unsigned long pfn,
5634 struct mminit_pfnnid_cache *state)
5636 unsigned long start_pfn, end_pfn;
5639 if (state->last_start <= pfn && pfn < state->last_end)
5640 return state->last_nid;
5642 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5644 state->last_start = start_pfn;
5645 state->last_end = end_pfn;
5646 state->last_nid = nid;
5651 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5654 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5655 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5656 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5658 * If an architecture guarantees that all ranges registered contain no holes
5659 * and may be freed, this this function may be used instead of calling
5660 * memblock_free_early_nid() manually.
5662 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5664 unsigned long start_pfn, end_pfn;
5667 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5668 start_pfn = min(start_pfn, max_low_pfn);
5669 end_pfn = min(end_pfn, max_low_pfn);
5671 if (start_pfn < end_pfn)
5672 memblock_free_early_nid(PFN_PHYS(start_pfn),
5673 (end_pfn - start_pfn) << PAGE_SHIFT,
5679 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5680 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5682 * If an architecture guarantees that all ranges registered contain no holes and may
5683 * be freed, this function may be used instead of calling memory_present() manually.
5685 void __init sparse_memory_present_with_active_regions(int nid)
5687 unsigned long start_pfn, end_pfn;
5690 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5691 memory_present(this_nid, start_pfn, end_pfn);
5695 * get_pfn_range_for_nid - Return the start and end page frames for a node
5696 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5697 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5698 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5700 * It returns the start and end page frame of a node based on information
5701 * provided by memblock_set_node(). If called for a node
5702 * with no available memory, a warning is printed and the start and end
5705 void __meminit get_pfn_range_for_nid(unsigned int nid,
5706 unsigned long *start_pfn, unsigned long *end_pfn)
5708 unsigned long this_start_pfn, this_end_pfn;
5714 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5715 *start_pfn = min(*start_pfn, this_start_pfn);
5716 *end_pfn = max(*end_pfn, this_end_pfn);
5719 if (*start_pfn == -1UL)
5724 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5725 * assumption is made that zones within a node are ordered in monotonic
5726 * increasing memory addresses so that the "highest" populated zone is used
5728 static void __init find_usable_zone_for_movable(void)
5731 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5732 if (zone_index == ZONE_MOVABLE)
5735 if (arch_zone_highest_possible_pfn[zone_index] >
5736 arch_zone_lowest_possible_pfn[zone_index])
5740 VM_BUG_ON(zone_index == -1);
5741 movable_zone = zone_index;
5745 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5746 * because it is sized independent of architecture. Unlike the other zones,
5747 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5748 * in each node depending on the size of each node and how evenly kernelcore
5749 * is distributed. This helper function adjusts the zone ranges
5750 * provided by the architecture for a given node by using the end of the
5751 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5752 * zones within a node are in order of monotonic increases memory addresses
5754 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5755 unsigned long zone_type,
5756 unsigned long node_start_pfn,
5757 unsigned long node_end_pfn,
5758 unsigned long *zone_start_pfn,
5759 unsigned long *zone_end_pfn)
5761 /* Only adjust if ZONE_MOVABLE is on this node */
5762 if (zone_movable_pfn[nid]) {
5763 /* Size ZONE_MOVABLE */
5764 if (zone_type == ZONE_MOVABLE) {
5765 *zone_start_pfn = zone_movable_pfn[nid];
5766 *zone_end_pfn = min(node_end_pfn,
5767 arch_zone_highest_possible_pfn[movable_zone]);
5769 /* Adjust for ZONE_MOVABLE starting within this range */
5770 } else if (!mirrored_kernelcore &&
5771 *zone_start_pfn < zone_movable_pfn[nid] &&
5772 *zone_end_pfn > zone_movable_pfn[nid]) {
5773 *zone_end_pfn = zone_movable_pfn[nid];
5775 /* Check if this whole range is within ZONE_MOVABLE */
5776 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5777 *zone_start_pfn = *zone_end_pfn;
5782 * Return the number of pages a zone spans in a node, including holes
5783 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5785 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5786 unsigned long zone_type,
5787 unsigned long node_start_pfn,
5788 unsigned long node_end_pfn,
5789 unsigned long *zone_start_pfn,
5790 unsigned long *zone_end_pfn,
5791 unsigned long *ignored)
5793 /* When hotadd a new node from cpu_up(), the node should be empty */
5794 if (!node_start_pfn && !node_end_pfn)
5797 /* Get the start and end of the zone */
5798 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5799 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5800 adjust_zone_range_for_zone_movable(nid, zone_type,
5801 node_start_pfn, node_end_pfn,
5802 zone_start_pfn, zone_end_pfn);
5804 /* Check that this node has pages within the zone's required range */
5805 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5808 /* Move the zone boundaries inside the node if necessary */
5809 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5810 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5812 /* Return the spanned pages */
5813 return *zone_end_pfn - *zone_start_pfn;
5817 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5818 * then all holes in the requested range will be accounted for.
5820 unsigned long __meminit __absent_pages_in_range(int nid,
5821 unsigned long range_start_pfn,
5822 unsigned long range_end_pfn)
5824 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5825 unsigned long start_pfn, end_pfn;
5828 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5829 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5830 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5831 nr_absent -= end_pfn - start_pfn;
5837 * absent_pages_in_range - Return number of page frames in holes within a range
5838 * @start_pfn: The start PFN to start searching for holes
5839 * @end_pfn: The end PFN to stop searching for holes
5841 * It returns the number of pages frames in memory holes within a range.
5843 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5844 unsigned long end_pfn)
5846 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5849 /* Return the number of page frames in holes in a zone on a node */
5850 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5851 unsigned long zone_type,
5852 unsigned long node_start_pfn,
5853 unsigned long node_end_pfn,
5854 unsigned long *ignored)
5856 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5857 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5858 unsigned long zone_start_pfn, zone_end_pfn;
5859 unsigned long nr_absent;
5861 /* When hotadd a new node from cpu_up(), the node should be empty */
5862 if (!node_start_pfn && !node_end_pfn)
5865 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5866 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5868 adjust_zone_range_for_zone_movable(nid, zone_type,
5869 node_start_pfn, node_end_pfn,
5870 &zone_start_pfn, &zone_end_pfn);
5871 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5874 * ZONE_MOVABLE handling.
5875 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5878 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5879 unsigned long start_pfn, end_pfn;
5880 struct memblock_region *r;
5882 for_each_memblock(memory, r) {
5883 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5884 zone_start_pfn, zone_end_pfn);
5885 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5886 zone_start_pfn, zone_end_pfn);
5888 if (zone_type == ZONE_MOVABLE &&
5889 memblock_is_mirror(r))
5890 nr_absent += end_pfn - start_pfn;
5892 if (zone_type == ZONE_NORMAL &&
5893 !memblock_is_mirror(r))
5894 nr_absent += end_pfn - start_pfn;
5901 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5902 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5903 unsigned long zone_type,
5904 unsigned long node_start_pfn,
5905 unsigned long node_end_pfn,
5906 unsigned long *zone_start_pfn,
5907 unsigned long *zone_end_pfn,
5908 unsigned long *zones_size)
5912 *zone_start_pfn = node_start_pfn;
5913 for (zone = 0; zone < zone_type; zone++)
5914 *zone_start_pfn += zones_size[zone];
5916 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5918 return zones_size[zone_type];
5921 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5922 unsigned long zone_type,
5923 unsigned long node_start_pfn,
5924 unsigned long node_end_pfn,
5925 unsigned long *zholes_size)
5930 return zholes_size[zone_type];
5933 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5935 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5936 unsigned long node_start_pfn,
5937 unsigned long node_end_pfn,
5938 unsigned long *zones_size,
5939 unsigned long *zholes_size)
5941 unsigned long realtotalpages = 0, totalpages = 0;
5944 for (i = 0; i < MAX_NR_ZONES; i++) {
5945 struct zone *zone = pgdat->node_zones + i;
5946 unsigned long zone_start_pfn, zone_end_pfn;
5947 unsigned long size, real_size;
5949 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5955 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5956 node_start_pfn, node_end_pfn,
5959 zone->zone_start_pfn = zone_start_pfn;
5961 zone->zone_start_pfn = 0;
5962 zone->spanned_pages = size;
5963 zone->present_pages = real_size;
5966 realtotalpages += real_size;
5969 pgdat->node_spanned_pages = totalpages;
5970 pgdat->node_present_pages = realtotalpages;
5971 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5975 #ifndef CONFIG_SPARSEMEM
5977 * Calculate the size of the zone->blockflags rounded to an unsigned long
5978 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5979 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5980 * round what is now in bits to nearest long in bits, then return it in
5983 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5985 unsigned long usemapsize;
5987 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5988 usemapsize = roundup(zonesize, pageblock_nr_pages);
5989 usemapsize = usemapsize >> pageblock_order;
5990 usemapsize *= NR_PAGEBLOCK_BITS;
5991 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5993 return usemapsize / 8;
5996 static void __init setup_usemap(struct pglist_data *pgdat,
5998 unsigned long zone_start_pfn,
5999 unsigned long zonesize)
6001 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6002 zone->pageblock_flags = NULL;
6004 zone->pageblock_flags =
6005 memblock_virt_alloc_node_nopanic(usemapsize,
6009 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6010 unsigned long zone_start_pfn, unsigned long zonesize) {}
6011 #endif /* CONFIG_SPARSEMEM */
6013 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6015 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6016 void __paginginit set_pageblock_order(void)
6020 /* Check that pageblock_nr_pages has not already been setup */
6021 if (pageblock_order)
6024 if (HPAGE_SHIFT > PAGE_SHIFT)
6025 order = HUGETLB_PAGE_ORDER;
6027 order = MAX_ORDER - 1;
6030 * Assume the largest contiguous order of interest is a huge page.
6031 * This value may be variable depending on boot parameters on IA64 and
6034 pageblock_order = order;
6036 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6039 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6040 * is unused as pageblock_order is set at compile-time. See
6041 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6044 void __paginginit set_pageblock_order(void)
6048 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6050 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6051 unsigned long present_pages)
6053 unsigned long pages = spanned_pages;
6056 * Provide a more accurate estimation if there are holes within
6057 * the zone and SPARSEMEM is in use. If there are holes within the
6058 * zone, each populated memory region may cost us one or two extra
6059 * memmap pages due to alignment because memmap pages for each
6060 * populated regions may not be naturally aligned on page boundary.
6061 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6063 if (spanned_pages > present_pages + (present_pages >> 4) &&
6064 IS_ENABLED(CONFIG_SPARSEMEM))
6065 pages = present_pages;
6067 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6071 * Set up the zone data structures:
6072 * - mark all pages reserved
6073 * - mark all memory queues empty
6074 * - clear the memory bitmaps
6076 * NOTE: pgdat should get zeroed by caller.
6078 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6081 int nid = pgdat->node_id;
6083 pgdat_resize_init(pgdat);
6084 #ifdef CONFIG_NUMA_BALANCING
6085 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6086 pgdat->numabalancing_migrate_nr_pages = 0;
6087 pgdat->numabalancing_migrate_next_window = jiffies;
6089 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6090 spin_lock_init(&pgdat->split_queue_lock);
6091 INIT_LIST_HEAD(&pgdat->split_queue);
6092 pgdat->split_queue_len = 0;
6094 init_waitqueue_head(&pgdat->kswapd_wait);
6095 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6096 #ifdef CONFIG_COMPACTION
6097 init_waitqueue_head(&pgdat->kcompactd_wait);
6099 pgdat_page_ext_init(pgdat);
6100 spin_lock_init(&pgdat->lru_lock);
6101 lruvec_init(node_lruvec(pgdat));
6103 pgdat->per_cpu_nodestats = &boot_nodestats;
6105 for (j = 0; j < MAX_NR_ZONES; j++) {
6106 struct zone *zone = pgdat->node_zones + j;
6107 unsigned long size, realsize, freesize, memmap_pages;
6108 unsigned long zone_start_pfn = zone->zone_start_pfn;
6110 size = zone->spanned_pages;
6111 realsize = freesize = zone->present_pages;
6114 * Adjust freesize so that it accounts for how much memory
6115 * is used by this zone for memmap. This affects the watermark
6116 * and per-cpu initialisations
6118 memmap_pages = calc_memmap_size(size, realsize);
6119 if (!is_highmem_idx(j)) {
6120 if (freesize >= memmap_pages) {
6121 freesize -= memmap_pages;
6124 " %s zone: %lu pages used for memmap\n",
6125 zone_names[j], memmap_pages);
6127 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6128 zone_names[j], memmap_pages, freesize);
6131 /* Account for reserved pages */
6132 if (j == 0 && freesize > dma_reserve) {
6133 freesize -= dma_reserve;
6134 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6135 zone_names[0], dma_reserve);
6138 if (!is_highmem_idx(j))
6139 nr_kernel_pages += freesize;
6140 /* Charge for highmem memmap if there are enough kernel pages */
6141 else if (nr_kernel_pages > memmap_pages * 2)
6142 nr_kernel_pages -= memmap_pages;
6143 nr_all_pages += freesize;
6146 * Set an approximate value for lowmem here, it will be adjusted
6147 * when the bootmem allocator frees pages into the buddy system.
6148 * And all highmem pages will be managed by the buddy system.
6150 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6154 zone->name = zone_names[j];
6155 zone->zone_pgdat = pgdat;
6156 spin_lock_init(&zone->lock);
6157 zone_seqlock_init(zone);
6158 zone_pcp_init(zone);
6163 set_pageblock_order();
6164 setup_usemap(pgdat, zone, zone_start_pfn, size);
6165 init_currently_empty_zone(zone, zone_start_pfn, size);
6166 memmap_init(size, nid, j, zone_start_pfn);
6170 #ifdef CONFIG_FLAT_NODE_MEM_MAP
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 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6181 offset = pgdat->node_start_pfn - start;
6182 /* ia64 gets its own node_mem_map, before this, without bootmem */
6183 if (!pgdat->node_mem_map) {
6184 unsigned long size, end;
6188 * The zone's endpoints aren't required to be MAX_ORDER
6189 * aligned but the node_mem_map endpoints must be in order
6190 * for the buddy allocator to function correctly.
6192 end = pgdat_end_pfn(pgdat);
6193 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6194 size = (end - start) * sizeof(struct page);
6195 map = alloc_remap(pgdat->node_id, size);
6197 map = memblock_virt_alloc_node_nopanic(size,
6199 pgdat->node_mem_map = map + offset;
6201 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6202 __func__, pgdat->node_id, (unsigned long)pgdat,
6203 (unsigned long)pgdat->node_mem_map);
6204 #ifndef CONFIG_NEED_MULTIPLE_NODES
6206 * With no DISCONTIG, the global mem_map is just set as node 0's
6208 if (pgdat == NODE_DATA(0)) {
6209 mem_map = NODE_DATA(0)->node_mem_map;
6210 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6211 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6213 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6218 static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6219 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6221 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6222 unsigned long node_start_pfn, unsigned long *zholes_size)
6224 pg_data_t *pgdat = NODE_DATA(nid);
6225 unsigned long start_pfn = 0;
6226 unsigned long end_pfn = 0;
6228 /* pg_data_t should be reset to zero when it's allocated */
6229 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6231 pgdat->node_id = nid;
6232 pgdat->node_start_pfn = node_start_pfn;
6233 pgdat->per_cpu_nodestats = NULL;
6234 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6235 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6236 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6237 (u64)start_pfn << PAGE_SHIFT,
6238 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6240 start_pfn = node_start_pfn;
6242 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6243 zones_size, zholes_size);
6245 alloc_node_mem_map(pgdat);
6247 reset_deferred_meminit(pgdat);
6248 free_area_init_core(pgdat);
6251 #ifdef CONFIG_HAVE_MEMBLOCK
6253 * Only struct pages that are backed by physical memory are zeroed and
6254 * initialized by going through __init_single_page(). But, there are some
6255 * struct pages which are reserved in memblock allocator and their fields
6256 * may be accessed (for example page_to_pfn() on some configuration accesses
6257 * flags). We must explicitly zero those struct pages.
6259 void __paginginit zero_resv_unavail(void)
6261 phys_addr_t start, end;
6266 * Loop through ranges that are reserved, but do not have reported
6267 * physical memory backing.
6270 for_each_resv_unavail_range(i, &start, &end) {
6271 for (pfn = PFN_DOWN(start); pfn < PFN_UP(end); pfn++) {
6272 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages)))
6274 mm_zero_struct_page(pfn_to_page(pfn));
6280 * Struct pages that do not have backing memory. This could be because
6281 * firmware is using some of this memory, or for some other reasons.
6282 * Once memblock is changed so such behaviour is not allowed: i.e.
6283 * list of "reserved" memory must be a subset of list of "memory", then
6284 * this code can be removed.
6287 pr_info("Reserved but unavailable: %lld pages", pgcnt);
6289 #endif /* CONFIG_HAVE_MEMBLOCK */
6291 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6293 #if MAX_NUMNODES > 1
6295 * Figure out the number of possible node ids.
6297 void __init setup_nr_node_ids(void)
6299 unsigned int highest;
6301 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6302 nr_node_ids = highest + 1;
6307 * node_map_pfn_alignment - determine the maximum internode alignment
6309 * This function should be called after node map is populated and sorted.
6310 * It calculates the maximum power of two alignment which can distinguish
6313 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6314 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6315 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6316 * shifted, 1GiB is enough and this function will indicate so.
6318 * This is used to test whether pfn -> nid mapping of the chosen memory
6319 * model has fine enough granularity to avoid incorrect mapping for the
6320 * populated node map.
6322 * Returns the determined alignment in pfn's. 0 if there is no alignment
6323 * requirement (single node).
6325 unsigned long __init node_map_pfn_alignment(void)
6327 unsigned long accl_mask = 0, last_end = 0;
6328 unsigned long start, end, mask;
6332 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6333 if (!start || last_nid < 0 || last_nid == nid) {
6340 * Start with a mask granular enough to pin-point to the
6341 * start pfn and tick off bits one-by-one until it becomes
6342 * too coarse to separate the current node from the last.
6344 mask = ~((1 << __ffs(start)) - 1);
6345 while (mask && last_end <= (start & (mask << 1)))
6348 /* accumulate all internode masks */
6352 /* convert mask to number of pages */
6353 return ~accl_mask + 1;
6356 /* Find the lowest pfn for a node */
6357 static unsigned long __init find_min_pfn_for_node(int nid)
6359 unsigned long min_pfn = ULONG_MAX;
6360 unsigned long start_pfn;
6363 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6364 min_pfn = min(min_pfn, start_pfn);
6366 if (min_pfn == ULONG_MAX) {
6367 pr_warn("Could not find start_pfn for node %d\n", nid);
6375 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6377 * It returns the minimum PFN based on information provided via
6378 * memblock_set_node().
6380 unsigned long __init find_min_pfn_with_active_regions(void)
6382 return find_min_pfn_for_node(MAX_NUMNODES);
6386 * early_calculate_totalpages()
6387 * Sum pages in active regions for movable zone.
6388 * Populate N_MEMORY for calculating usable_nodes.
6390 static unsigned long __init early_calculate_totalpages(void)
6392 unsigned long totalpages = 0;
6393 unsigned long start_pfn, end_pfn;
6396 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6397 unsigned long pages = end_pfn - start_pfn;
6399 totalpages += pages;
6401 node_set_state(nid, N_MEMORY);
6407 * Find the PFN the Movable zone begins in each node. Kernel memory
6408 * is spread evenly between nodes as long as the nodes have enough
6409 * memory. When they don't, some nodes will have more kernelcore than
6412 static void __init find_zone_movable_pfns_for_nodes(void)
6415 unsigned long usable_startpfn;
6416 unsigned long kernelcore_node, kernelcore_remaining;
6417 /* save the state before borrow the nodemask */
6418 nodemask_t saved_node_state = node_states[N_MEMORY];
6419 unsigned long totalpages = early_calculate_totalpages();
6420 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6421 struct memblock_region *r;
6423 /* Need to find movable_zone earlier when movable_node is specified. */
6424 find_usable_zone_for_movable();
6427 * If movable_node is specified, ignore kernelcore and movablecore
6430 if (movable_node_is_enabled()) {
6431 for_each_memblock(memory, r) {
6432 if (!memblock_is_hotpluggable(r))
6437 usable_startpfn = PFN_DOWN(r->base);
6438 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6439 min(usable_startpfn, zone_movable_pfn[nid]) :
6447 * If kernelcore=mirror is specified, ignore movablecore option
6449 if (mirrored_kernelcore) {
6450 bool mem_below_4gb_not_mirrored = false;
6452 for_each_memblock(memory, r) {
6453 if (memblock_is_mirror(r))
6458 usable_startpfn = memblock_region_memory_base_pfn(r);
6460 if (usable_startpfn < 0x100000) {
6461 mem_below_4gb_not_mirrored = true;
6465 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6466 min(usable_startpfn, zone_movable_pfn[nid]) :
6470 if (mem_below_4gb_not_mirrored)
6471 pr_warn("This configuration results in unmirrored kernel memory.");
6477 * If movablecore=nn[KMG] was specified, calculate what size of
6478 * kernelcore that corresponds so that memory usable for
6479 * any allocation type is evenly spread. If both kernelcore
6480 * and movablecore are specified, then the value of kernelcore
6481 * will be used for required_kernelcore if it's greater than
6482 * what movablecore would have allowed.
6484 if (required_movablecore) {
6485 unsigned long corepages;
6488 * Round-up so that ZONE_MOVABLE is at least as large as what
6489 * was requested by the user
6491 required_movablecore =
6492 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6493 required_movablecore = min(totalpages, required_movablecore);
6494 corepages = totalpages - required_movablecore;
6496 required_kernelcore = max(required_kernelcore, corepages);
6500 * If kernelcore was not specified or kernelcore size is larger
6501 * than totalpages, there is no ZONE_MOVABLE.
6503 if (!required_kernelcore || required_kernelcore >= totalpages)
6506 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6507 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6510 /* Spread kernelcore memory as evenly as possible throughout nodes */
6511 kernelcore_node = required_kernelcore / usable_nodes;
6512 for_each_node_state(nid, N_MEMORY) {
6513 unsigned long start_pfn, end_pfn;
6516 * Recalculate kernelcore_node if the division per node
6517 * now exceeds what is necessary to satisfy the requested
6518 * amount of memory for the kernel
6520 if (required_kernelcore < kernelcore_node)
6521 kernelcore_node = required_kernelcore / usable_nodes;
6524 * As the map is walked, we track how much memory is usable
6525 * by the kernel using kernelcore_remaining. When it is
6526 * 0, the rest of the node is usable by ZONE_MOVABLE
6528 kernelcore_remaining = kernelcore_node;
6530 /* Go through each range of PFNs within this node */
6531 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6532 unsigned long size_pages;
6534 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6535 if (start_pfn >= end_pfn)
6538 /* Account for what is only usable for kernelcore */
6539 if (start_pfn < usable_startpfn) {
6540 unsigned long kernel_pages;
6541 kernel_pages = min(end_pfn, usable_startpfn)
6544 kernelcore_remaining -= min(kernel_pages,
6545 kernelcore_remaining);
6546 required_kernelcore -= min(kernel_pages,
6547 required_kernelcore);
6549 /* Continue if range is now fully accounted */
6550 if (end_pfn <= usable_startpfn) {
6553 * Push zone_movable_pfn to the end so
6554 * that if we have to rebalance
6555 * kernelcore across nodes, we will
6556 * not double account here
6558 zone_movable_pfn[nid] = end_pfn;
6561 start_pfn = usable_startpfn;
6565 * The usable PFN range for ZONE_MOVABLE is from
6566 * start_pfn->end_pfn. Calculate size_pages as the
6567 * number of pages used as kernelcore
6569 size_pages = end_pfn - start_pfn;
6570 if (size_pages > kernelcore_remaining)
6571 size_pages = kernelcore_remaining;
6572 zone_movable_pfn[nid] = start_pfn + size_pages;
6575 * Some kernelcore has been met, update counts and
6576 * break if the kernelcore for this node has been
6579 required_kernelcore -= min(required_kernelcore,
6581 kernelcore_remaining -= size_pages;
6582 if (!kernelcore_remaining)
6588 * If there is still required_kernelcore, we do another pass with one
6589 * less node in the count. This will push zone_movable_pfn[nid] further
6590 * along on the nodes that still have memory until kernelcore is
6594 if (usable_nodes && required_kernelcore > usable_nodes)
6598 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6599 for (nid = 0; nid < MAX_NUMNODES; nid++)
6600 zone_movable_pfn[nid] =
6601 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6604 /* restore the node_state */
6605 node_states[N_MEMORY] = saved_node_state;
6608 /* Any regular or high memory on that node ? */
6609 static void check_for_memory(pg_data_t *pgdat, int nid)
6611 enum zone_type zone_type;
6613 if (N_MEMORY == N_NORMAL_MEMORY)
6616 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6617 struct zone *zone = &pgdat->node_zones[zone_type];
6618 if (populated_zone(zone)) {
6619 node_set_state(nid, N_HIGH_MEMORY);
6620 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6621 zone_type <= ZONE_NORMAL)
6622 node_set_state(nid, N_NORMAL_MEMORY);
6629 * free_area_init_nodes - Initialise all pg_data_t and zone data
6630 * @max_zone_pfn: an array of max PFNs for each zone
6632 * This will call free_area_init_node() for each active node in the system.
6633 * Using the page ranges provided by memblock_set_node(), the size of each
6634 * zone in each node and their holes is calculated. If the maximum PFN
6635 * between two adjacent zones match, it is assumed that the zone is empty.
6636 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6637 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6638 * starts where the previous one ended. For example, ZONE_DMA32 starts
6639 * at arch_max_dma_pfn.
6641 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6643 unsigned long start_pfn, end_pfn;
6646 /* Record where the zone boundaries are */
6647 memset(arch_zone_lowest_possible_pfn, 0,
6648 sizeof(arch_zone_lowest_possible_pfn));
6649 memset(arch_zone_highest_possible_pfn, 0,
6650 sizeof(arch_zone_highest_possible_pfn));
6652 start_pfn = find_min_pfn_with_active_regions();
6654 for (i = 0; i < MAX_NR_ZONES; i++) {
6655 if (i == ZONE_MOVABLE)
6658 end_pfn = max(max_zone_pfn[i], start_pfn);
6659 arch_zone_lowest_possible_pfn[i] = start_pfn;
6660 arch_zone_highest_possible_pfn[i] = end_pfn;
6662 start_pfn = end_pfn;
6665 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6666 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6667 find_zone_movable_pfns_for_nodes();
6669 /* Print out the zone ranges */
6670 pr_info("Zone ranges:\n");
6671 for (i = 0; i < MAX_NR_ZONES; i++) {
6672 if (i == ZONE_MOVABLE)
6674 pr_info(" %-8s ", zone_names[i]);
6675 if (arch_zone_lowest_possible_pfn[i] ==
6676 arch_zone_highest_possible_pfn[i])
6679 pr_cont("[mem %#018Lx-%#018Lx]\n",
6680 (u64)arch_zone_lowest_possible_pfn[i]
6682 ((u64)arch_zone_highest_possible_pfn[i]
6683 << PAGE_SHIFT) - 1);
6686 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6687 pr_info("Movable zone start for each node\n");
6688 for (i = 0; i < MAX_NUMNODES; i++) {
6689 if (zone_movable_pfn[i])
6690 pr_info(" Node %d: %#018Lx\n", i,
6691 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6694 /* Print out the early node map */
6695 pr_info("Early memory node ranges\n");
6696 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6697 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6698 (u64)start_pfn << PAGE_SHIFT,
6699 ((u64)end_pfn << PAGE_SHIFT) - 1);
6701 /* Initialise every node */
6702 mminit_verify_pageflags_layout();
6703 setup_nr_node_ids();
6704 for_each_online_node(nid) {
6705 pg_data_t *pgdat = NODE_DATA(nid);
6706 free_area_init_node(nid, NULL,
6707 find_min_pfn_for_node(nid), NULL);
6709 /* Any memory on that node */
6710 if (pgdat->node_present_pages)
6711 node_set_state(nid, N_MEMORY);
6712 check_for_memory(pgdat, nid);
6714 zero_resv_unavail();
6717 static int __init cmdline_parse_core(char *p, unsigned long *core)
6719 unsigned long long coremem;
6723 coremem = memparse(p, &p);
6724 *core = coremem >> PAGE_SHIFT;
6726 /* Paranoid check that UL is enough for the coremem value */
6727 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6733 * kernelcore=size sets the amount of memory for use for allocations that
6734 * cannot be reclaimed or migrated.
6736 static int __init cmdline_parse_kernelcore(char *p)
6738 /* parse kernelcore=mirror */
6739 if (parse_option_str(p, "mirror")) {
6740 mirrored_kernelcore = true;
6744 return cmdline_parse_core(p, &required_kernelcore);
6748 * movablecore=size sets the amount of memory for use for allocations that
6749 * can be reclaimed or migrated.
6751 static int __init cmdline_parse_movablecore(char *p)
6753 return cmdline_parse_core(p, &required_movablecore);
6756 early_param("kernelcore", cmdline_parse_kernelcore);
6757 early_param("movablecore", cmdline_parse_movablecore);
6759 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6761 void adjust_managed_page_count(struct page *page, long count)
6763 spin_lock(&managed_page_count_lock);
6764 page_zone(page)->managed_pages += count;
6765 totalram_pages += count;
6766 #ifdef CONFIG_HIGHMEM
6767 if (PageHighMem(page))
6768 totalhigh_pages += count;
6770 spin_unlock(&managed_page_count_lock);
6772 EXPORT_SYMBOL(adjust_managed_page_count);
6774 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6777 unsigned long pages = 0;
6779 start = (void *)PAGE_ALIGN((unsigned long)start);
6780 end = (void *)((unsigned long)end & PAGE_MASK);
6781 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6782 if ((unsigned int)poison <= 0xFF)
6783 memset(pos, poison, PAGE_SIZE);
6784 free_reserved_page(virt_to_page(pos));
6788 pr_info("Freeing %s memory: %ldK\n",
6789 s, pages << (PAGE_SHIFT - 10));
6793 EXPORT_SYMBOL(free_reserved_area);
6795 #ifdef CONFIG_HIGHMEM
6796 void free_highmem_page(struct page *page)
6798 __free_reserved_page(page);
6800 page_zone(page)->managed_pages++;
6806 void __init mem_init_print_info(const char *str)
6808 unsigned long physpages, codesize, datasize, rosize, bss_size;
6809 unsigned long init_code_size, init_data_size;
6811 physpages = get_num_physpages();
6812 codesize = _etext - _stext;
6813 datasize = _edata - _sdata;
6814 rosize = __end_rodata - __start_rodata;
6815 bss_size = __bss_stop - __bss_start;
6816 init_data_size = __init_end - __init_begin;
6817 init_code_size = _einittext - _sinittext;
6820 * Detect special cases and adjust section sizes accordingly:
6821 * 1) .init.* may be embedded into .data sections
6822 * 2) .init.text.* may be out of [__init_begin, __init_end],
6823 * please refer to arch/tile/kernel/vmlinux.lds.S.
6824 * 3) .rodata.* may be embedded into .text or .data sections.
6826 #define adj_init_size(start, end, size, pos, adj) \
6828 if (start <= pos && pos < end && size > adj) \
6832 adj_init_size(__init_begin, __init_end, init_data_size,
6833 _sinittext, init_code_size);
6834 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6835 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6836 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6837 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6839 #undef adj_init_size
6841 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6842 #ifdef CONFIG_HIGHMEM
6846 nr_free_pages() << (PAGE_SHIFT - 10),
6847 physpages << (PAGE_SHIFT - 10),
6848 codesize >> 10, datasize >> 10, rosize >> 10,
6849 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6850 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6851 totalcma_pages << (PAGE_SHIFT - 10),
6852 #ifdef CONFIG_HIGHMEM
6853 totalhigh_pages << (PAGE_SHIFT - 10),
6855 str ? ", " : "", str ? str : "");
6859 * set_dma_reserve - set the specified number of pages reserved in the first zone
6860 * @new_dma_reserve: The number of pages to mark reserved
6862 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6863 * In the DMA zone, a significant percentage may be consumed by kernel image
6864 * and other unfreeable allocations which can skew the watermarks badly. This
6865 * function may optionally be used to account for unfreeable pages in the
6866 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6867 * smaller per-cpu batchsize.
6869 void __init set_dma_reserve(unsigned long new_dma_reserve)
6871 dma_reserve = new_dma_reserve;
6874 void __init free_area_init(unsigned long *zones_size)
6876 free_area_init_node(0, zones_size,
6877 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6878 zero_resv_unavail();
6881 static int page_alloc_cpu_dead(unsigned int cpu)
6884 lru_add_drain_cpu(cpu);
6888 * Spill the event counters of the dead processor
6889 * into the current processors event counters.
6890 * This artificially elevates the count of the current
6893 vm_events_fold_cpu(cpu);
6896 * Zero the differential counters of the dead processor
6897 * so that the vm statistics are consistent.
6899 * This is only okay since the processor is dead and cannot
6900 * race with what we are doing.
6902 cpu_vm_stats_fold(cpu);
6906 void __init page_alloc_init(void)
6910 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6911 "mm/page_alloc:dead", NULL,
6912 page_alloc_cpu_dead);
6917 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6918 * or min_free_kbytes changes.
6920 static void calculate_totalreserve_pages(void)
6922 struct pglist_data *pgdat;
6923 unsigned long reserve_pages = 0;
6924 enum zone_type i, j;
6926 for_each_online_pgdat(pgdat) {
6928 pgdat->totalreserve_pages = 0;
6930 for (i = 0; i < MAX_NR_ZONES; i++) {
6931 struct zone *zone = pgdat->node_zones + i;
6934 /* Find valid and maximum lowmem_reserve in the zone */
6935 for (j = i; j < MAX_NR_ZONES; j++) {
6936 if (zone->lowmem_reserve[j] > max)
6937 max = zone->lowmem_reserve[j];
6940 /* we treat the high watermark as reserved pages. */
6941 max += high_wmark_pages(zone);
6943 if (max > zone->managed_pages)
6944 max = zone->managed_pages;
6946 pgdat->totalreserve_pages += max;
6948 reserve_pages += max;
6951 totalreserve_pages = reserve_pages;
6955 * setup_per_zone_lowmem_reserve - called whenever
6956 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6957 * has a correct pages reserved value, so an adequate number of
6958 * pages are left in the zone after a successful __alloc_pages().
6960 static void setup_per_zone_lowmem_reserve(void)
6962 struct pglist_data *pgdat;
6963 enum zone_type j, idx;
6965 for_each_online_pgdat(pgdat) {
6966 for (j = 0; j < MAX_NR_ZONES; j++) {
6967 struct zone *zone = pgdat->node_zones + j;
6968 unsigned long managed_pages = zone->managed_pages;
6970 zone->lowmem_reserve[j] = 0;
6974 struct zone *lower_zone;
6978 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6979 sysctl_lowmem_reserve_ratio[idx] = 1;
6981 lower_zone = pgdat->node_zones + idx;
6982 lower_zone->lowmem_reserve[j] = managed_pages /
6983 sysctl_lowmem_reserve_ratio[idx];
6984 managed_pages += lower_zone->managed_pages;
6989 /* update totalreserve_pages */
6990 calculate_totalreserve_pages();
6993 static void __setup_per_zone_wmarks(void)
6995 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6996 unsigned long lowmem_pages = 0;
6998 unsigned long flags;
7000 /* Calculate total number of !ZONE_HIGHMEM pages */
7001 for_each_zone(zone) {
7002 if (!is_highmem(zone))
7003 lowmem_pages += zone->managed_pages;
7006 for_each_zone(zone) {
7009 spin_lock_irqsave(&zone->lock, flags);
7010 tmp = (u64)pages_min * zone->managed_pages;
7011 do_div(tmp, lowmem_pages);
7012 if (is_highmem(zone)) {
7014 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7015 * need highmem pages, so cap pages_min to a small
7018 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7019 * deltas control asynch page reclaim, and so should
7020 * not be capped for highmem.
7022 unsigned long min_pages;
7024 min_pages = zone->managed_pages / 1024;
7025 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7026 zone->watermark[WMARK_MIN] = min_pages;
7029 * If it's a lowmem zone, reserve a number of pages
7030 * proportionate to the zone's size.
7032 zone->watermark[WMARK_MIN] = tmp;
7036 * Set the kswapd watermarks distance according to the
7037 * scale factor in proportion to available memory, but
7038 * ensure a minimum size on small systems.
7040 tmp = max_t(u64, tmp >> 2,
7041 mult_frac(zone->managed_pages,
7042 watermark_scale_factor, 10000));
7044 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7045 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7047 spin_unlock_irqrestore(&zone->lock, flags);
7050 /* update totalreserve_pages */
7051 calculate_totalreserve_pages();
7055 * setup_per_zone_wmarks - called when min_free_kbytes changes
7056 * or when memory is hot-{added|removed}
7058 * Ensures that the watermark[min,low,high] values for each zone are set
7059 * correctly with respect to min_free_kbytes.
7061 void setup_per_zone_wmarks(void)
7063 static DEFINE_SPINLOCK(lock);
7066 __setup_per_zone_wmarks();
7071 * Initialise min_free_kbytes.
7073 * For small machines we want it small (128k min). For large machines
7074 * we want it large (64MB max). But it is not linear, because network
7075 * bandwidth does not increase linearly with machine size. We use
7077 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7078 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7094 int __meminit init_per_zone_wmark_min(void)
7096 unsigned long lowmem_kbytes;
7097 int new_min_free_kbytes;
7099 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7100 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7102 if (new_min_free_kbytes > user_min_free_kbytes) {
7103 min_free_kbytes = new_min_free_kbytes;
7104 if (min_free_kbytes < 128)
7105 min_free_kbytes = 128;
7106 if (min_free_kbytes > 65536)
7107 min_free_kbytes = 65536;
7109 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7110 new_min_free_kbytes, user_min_free_kbytes);
7112 setup_per_zone_wmarks();
7113 refresh_zone_stat_thresholds();
7114 setup_per_zone_lowmem_reserve();
7117 setup_min_unmapped_ratio();
7118 setup_min_slab_ratio();
7123 core_initcall(init_per_zone_wmark_min)
7126 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7127 * that we can call two helper functions whenever min_free_kbytes
7130 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7131 void __user *buffer, size_t *length, loff_t *ppos)
7135 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7140 user_min_free_kbytes = min_free_kbytes;
7141 setup_per_zone_wmarks();
7146 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7147 void __user *buffer, size_t *length, loff_t *ppos)
7151 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7156 setup_per_zone_wmarks();
7162 static void setup_min_unmapped_ratio(void)
7167 for_each_online_pgdat(pgdat)
7168 pgdat->min_unmapped_pages = 0;
7171 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7172 sysctl_min_unmapped_ratio) / 100;
7176 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7177 void __user *buffer, size_t *length, loff_t *ppos)
7181 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7185 setup_min_unmapped_ratio();
7190 static void setup_min_slab_ratio(void)
7195 for_each_online_pgdat(pgdat)
7196 pgdat->min_slab_pages = 0;
7199 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7200 sysctl_min_slab_ratio) / 100;
7203 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7204 void __user *buffer, size_t *length, loff_t *ppos)
7208 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7212 setup_min_slab_ratio();
7219 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7220 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7221 * whenever sysctl_lowmem_reserve_ratio changes.
7223 * The reserve ratio obviously has absolutely no relation with the
7224 * minimum watermarks. The lowmem reserve ratio can only make sense
7225 * if in function of the boot time zone sizes.
7227 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7228 void __user *buffer, size_t *length, loff_t *ppos)
7230 proc_dointvec_minmax(table, write, buffer, length, ppos);
7231 setup_per_zone_lowmem_reserve();
7236 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7237 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7238 * pagelist can have before it gets flushed back to buddy allocator.
7240 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7241 void __user *buffer, size_t *length, loff_t *ppos)
7244 int old_percpu_pagelist_fraction;
7247 mutex_lock(&pcp_batch_high_lock);
7248 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7250 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7251 if (!write || ret < 0)
7254 /* Sanity checking to avoid pcp imbalance */
7255 if (percpu_pagelist_fraction &&
7256 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7257 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7263 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7266 for_each_populated_zone(zone) {
7269 for_each_possible_cpu(cpu)
7270 pageset_set_high_and_batch(zone,
7271 per_cpu_ptr(zone->pageset, cpu));
7274 mutex_unlock(&pcp_batch_high_lock);
7279 int hashdist = HASHDIST_DEFAULT;
7281 static int __init set_hashdist(char *str)
7285 hashdist = simple_strtoul(str, &str, 0);
7288 __setup("hashdist=", set_hashdist);
7291 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7293 * Returns the number of pages that arch has reserved but
7294 * is not known to alloc_large_system_hash().
7296 static unsigned long __init arch_reserved_kernel_pages(void)
7303 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7304 * machines. As memory size is increased the scale is also increased but at
7305 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7306 * quadruples the scale is increased by one, which means the size of hash table
7307 * only doubles, instead of quadrupling as well.
7308 * Because 32-bit systems cannot have large physical memory, where this scaling
7309 * makes sense, it is disabled on such platforms.
7311 #if __BITS_PER_LONG > 32
7312 #define ADAPT_SCALE_BASE (64ul << 30)
7313 #define ADAPT_SCALE_SHIFT 2
7314 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7318 * allocate a large system hash table from bootmem
7319 * - it is assumed that the hash table must contain an exact power-of-2
7320 * quantity of entries
7321 * - limit is the number of hash buckets, not the total allocation size
7323 void *__init alloc_large_system_hash(const char *tablename,
7324 unsigned long bucketsize,
7325 unsigned long numentries,
7328 unsigned int *_hash_shift,
7329 unsigned int *_hash_mask,
7330 unsigned long low_limit,
7331 unsigned long high_limit)
7333 unsigned long long max = high_limit;
7334 unsigned long log2qty, size;
7338 /* allow the kernel cmdline to have a say */
7340 /* round applicable memory size up to nearest megabyte */
7341 numentries = nr_kernel_pages;
7342 numentries -= arch_reserved_kernel_pages();
7344 /* It isn't necessary when PAGE_SIZE >= 1MB */
7345 if (PAGE_SHIFT < 20)
7346 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7348 #if __BITS_PER_LONG > 32
7350 unsigned long adapt;
7352 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7353 adapt <<= ADAPT_SCALE_SHIFT)
7358 /* limit to 1 bucket per 2^scale bytes of low memory */
7359 if (scale > PAGE_SHIFT)
7360 numentries >>= (scale - PAGE_SHIFT);
7362 numentries <<= (PAGE_SHIFT - scale);
7364 /* Make sure we've got at least a 0-order allocation.. */
7365 if (unlikely(flags & HASH_SMALL)) {
7366 /* Makes no sense without HASH_EARLY */
7367 WARN_ON(!(flags & HASH_EARLY));
7368 if (!(numentries >> *_hash_shift)) {
7369 numentries = 1UL << *_hash_shift;
7370 BUG_ON(!numentries);
7372 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7373 numentries = PAGE_SIZE / bucketsize;
7375 numentries = roundup_pow_of_two(numentries);
7377 /* limit allocation size to 1/16 total memory by default */
7379 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7380 do_div(max, bucketsize);
7382 max = min(max, 0x80000000ULL);
7384 if (numentries < low_limit)
7385 numentries = low_limit;
7386 if (numentries > max)
7389 log2qty = ilog2(numentries);
7391 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7393 size = bucketsize << log2qty;
7394 if (flags & HASH_EARLY) {
7395 if (flags & HASH_ZERO)
7396 table = memblock_virt_alloc_nopanic(size, 0);
7398 table = memblock_virt_alloc_raw(size, 0);
7399 } else if (hashdist) {
7400 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7403 * If bucketsize is not a power-of-two, we may free
7404 * some pages at the end of hash table which
7405 * alloc_pages_exact() automatically does
7407 if (get_order(size) < MAX_ORDER) {
7408 table = alloc_pages_exact(size, gfp_flags);
7409 kmemleak_alloc(table, size, 1, gfp_flags);
7412 } while (!table && size > PAGE_SIZE && --log2qty);
7415 panic("Failed to allocate %s hash table\n", tablename);
7417 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7418 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7421 *_hash_shift = log2qty;
7423 *_hash_mask = (1 << log2qty) - 1;
7429 * This function checks whether pageblock includes unmovable pages or not.
7430 * If @count is not zero, it is okay to include less @count unmovable pages
7432 * PageLRU check without isolation or lru_lock could race so that
7433 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7434 * check without lock_page also may miss some movable non-lru pages at
7435 * race condition. So you can't expect this function should be exact.
7437 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7439 bool skip_hwpoisoned_pages)
7441 unsigned long pfn, iter, found;
7444 * For avoiding noise data, lru_add_drain_all() should be called
7445 * If ZONE_MOVABLE, the zone never contains unmovable pages
7447 if (zone_idx(zone) == ZONE_MOVABLE)
7451 * CMA allocations (alloc_contig_range) really need to mark isolate
7452 * CMA pageblocks even when they are not movable in fact so consider
7453 * them movable here.
7455 if (is_migrate_cma(migratetype) &&
7456 is_migrate_cma(get_pageblock_migratetype(page)))
7459 pfn = page_to_pfn(page);
7460 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7461 unsigned long check = pfn + iter;
7463 if (!pfn_valid_within(check))
7466 page = pfn_to_page(check);
7468 if (PageReserved(page))
7472 * Hugepages are not in LRU lists, but they're movable.
7473 * We need not scan over tail pages bacause we don't
7474 * handle each tail page individually in migration.
7476 if (PageHuge(page)) {
7477 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7482 * We can't use page_count without pin a page
7483 * because another CPU can free compound page.
7484 * This check already skips compound tails of THP
7485 * because their page->_refcount is zero at all time.
7487 if (!page_ref_count(page)) {
7488 if (PageBuddy(page))
7489 iter += (1 << page_order(page)) - 1;
7494 * The HWPoisoned page may be not in buddy system, and
7495 * page_count() is not 0.
7497 if (skip_hwpoisoned_pages && PageHWPoison(page))
7500 if (__PageMovable(page))
7506 * If there are RECLAIMABLE pages, we need to check
7507 * it. But now, memory offline itself doesn't call
7508 * shrink_node_slabs() and it still to be fixed.
7511 * If the page is not RAM, page_count()should be 0.
7512 * we don't need more check. This is an _used_ not-movable page.
7514 * The problematic thing here is PG_reserved pages. PG_reserved
7515 * is set to both of a memory hole page and a _used_ kernel
7524 bool is_pageblock_removable_nolock(struct page *page)
7530 * We have to be careful here because we are iterating over memory
7531 * sections which are not zone aware so we might end up outside of
7532 * the zone but still within the section.
7533 * We have to take care about the node as well. If the node is offline
7534 * its NODE_DATA will be NULL - see page_zone.
7536 if (!node_online(page_to_nid(page)))
7539 zone = page_zone(page);
7540 pfn = page_to_pfn(page);
7541 if (!zone_spans_pfn(zone, pfn))
7544 return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, true);
7547 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7549 static unsigned long pfn_max_align_down(unsigned long pfn)
7551 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7552 pageblock_nr_pages) - 1);
7555 static unsigned long pfn_max_align_up(unsigned long pfn)
7557 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7558 pageblock_nr_pages));
7561 /* [start, end) must belong to a single zone. */
7562 static int __alloc_contig_migrate_range(struct compact_control *cc,
7563 unsigned long start, unsigned long end)
7565 /* This function is based on compact_zone() from compaction.c. */
7566 unsigned long nr_reclaimed;
7567 unsigned long pfn = start;
7568 unsigned int tries = 0;
7573 while (pfn < end || !list_empty(&cc->migratepages)) {
7574 if (fatal_signal_pending(current)) {
7579 if (list_empty(&cc->migratepages)) {
7580 cc->nr_migratepages = 0;
7581 pfn = isolate_migratepages_range(cc, pfn, end);
7587 } else if (++tries == 5) {
7588 ret = ret < 0 ? ret : -EBUSY;
7592 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7594 cc->nr_migratepages -= nr_reclaimed;
7596 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7597 NULL, 0, cc->mode, MR_CMA);
7600 putback_movable_pages(&cc->migratepages);
7607 * alloc_contig_range() -- tries to allocate given range of pages
7608 * @start: start PFN to allocate
7609 * @end: one-past-the-last PFN to allocate
7610 * @migratetype: migratetype of the underlaying pageblocks (either
7611 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7612 * in range must have the same migratetype and it must
7613 * be either of the two.
7614 * @gfp_mask: GFP mask to use during compaction
7616 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7617 * aligned, however it's the caller's responsibility to guarantee that
7618 * we are the only thread that changes migrate type of pageblocks the
7621 * The PFN range must belong to a single zone.
7623 * Returns zero on success or negative error code. On success all
7624 * pages which PFN is in [start, end) are allocated for the caller and
7625 * need to be freed with free_contig_range().
7627 int alloc_contig_range(unsigned long start, unsigned long end,
7628 unsigned migratetype, gfp_t gfp_mask)
7630 unsigned long outer_start, outer_end;
7634 struct compact_control cc = {
7635 .nr_migratepages = 0,
7637 .zone = page_zone(pfn_to_page(start)),
7638 .mode = MIGRATE_SYNC,
7639 .ignore_skip_hint = true,
7640 .no_set_skip_hint = true,
7641 .gfp_mask = current_gfp_context(gfp_mask),
7643 INIT_LIST_HEAD(&cc.migratepages);
7646 * What we do here is we mark all pageblocks in range as
7647 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7648 * have different sizes, and due to the way page allocator
7649 * work, we align the range to biggest of the two pages so
7650 * that page allocator won't try to merge buddies from
7651 * different pageblocks and change MIGRATE_ISOLATE to some
7652 * other migration type.
7654 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7655 * migrate the pages from an unaligned range (ie. pages that
7656 * we are interested in). This will put all the pages in
7657 * range back to page allocator as MIGRATE_ISOLATE.
7659 * When this is done, we take the pages in range from page
7660 * allocator removing them from the buddy system. This way
7661 * page allocator will never consider using them.
7663 * This lets us mark the pageblocks back as
7664 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7665 * aligned range but not in the unaligned, original range are
7666 * put back to page allocator so that buddy can use them.
7669 ret = start_isolate_page_range(pfn_max_align_down(start),
7670 pfn_max_align_up(end), migratetype,
7676 * In case of -EBUSY, we'd like to know which page causes problem.
7677 * So, just fall through. test_pages_isolated() has a tracepoint
7678 * which will report the busy page.
7680 * It is possible that busy pages could become available before
7681 * the call to test_pages_isolated, and the range will actually be
7682 * allocated. So, if we fall through be sure to clear ret so that
7683 * -EBUSY is not accidentally used or returned to caller.
7685 ret = __alloc_contig_migrate_range(&cc, start, end);
7686 if (ret && ret != -EBUSY)
7691 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7692 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7693 * more, all pages in [start, end) are free in page allocator.
7694 * What we are going to do is to allocate all pages from
7695 * [start, end) (that is remove them from page allocator).
7697 * The only problem is that pages at the beginning and at the
7698 * end of interesting range may be not aligned with pages that
7699 * page allocator holds, ie. they can be part of higher order
7700 * pages. Because of this, we reserve the bigger range and
7701 * once this is done free the pages we are not interested in.
7703 * We don't have to hold zone->lock here because the pages are
7704 * isolated thus they won't get removed from buddy.
7707 lru_add_drain_all();
7708 drain_all_pages(cc.zone);
7711 outer_start = start;
7712 while (!PageBuddy(pfn_to_page(outer_start))) {
7713 if (++order >= MAX_ORDER) {
7714 outer_start = start;
7717 outer_start &= ~0UL << order;
7720 if (outer_start != start) {
7721 order = page_order(pfn_to_page(outer_start));
7724 * outer_start page could be small order buddy page and
7725 * it doesn't include start page. Adjust outer_start
7726 * in this case to report failed page properly
7727 * on tracepoint in test_pages_isolated()
7729 if (outer_start + (1UL << order) <= start)
7730 outer_start = start;
7733 /* Make sure the range is really isolated. */
7734 if (test_pages_isolated(outer_start, end, false)) {
7735 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7736 __func__, outer_start, end);
7741 /* Grab isolated pages from freelists. */
7742 outer_end = isolate_freepages_range(&cc, outer_start, end);
7748 /* Free head and tail (if any) */
7749 if (start != outer_start)
7750 free_contig_range(outer_start, start - outer_start);
7751 if (end != outer_end)
7752 free_contig_range(end, outer_end - end);
7755 undo_isolate_page_range(pfn_max_align_down(start),
7756 pfn_max_align_up(end), migratetype);
7760 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7762 unsigned int count = 0;
7764 for (; nr_pages--; pfn++) {
7765 struct page *page = pfn_to_page(pfn);
7767 count += page_count(page) != 1;
7770 WARN(count != 0, "%d pages are still in use!\n", count);
7774 #ifdef CONFIG_MEMORY_HOTPLUG
7776 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7777 * page high values need to be recalulated.
7779 void __meminit zone_pcp_update(struct zone *zone)
7782 mutex_lock(&pcp_batch_high_lock);
7783 for_each_possible_cpu(cpu)
7784 pageset_set_high_and_batch(zone,
7785 per_cpu_ptr(zone->pageset, cpu));
7786 mutex_unlock(&pcp_batch_high_lock);
7790 void zone_pcp_reset(struct zone *zone)
7792 unsigned long flags;
7794 struct per_cpu_pageset *pset;
7796 /* avoid races with drain_pages() */
7797 local_irq_save(flags);
7798 if (zone->pageset != &boot_pageset) {
7799 for_each_online_cpu(cpu) {
7800 pset = per_cpu_ptr(zone->pageset, cpu);
7801 drain_zonestat(zone, pset);
7803 free_percpu(zone->pageset);
7804 zone->pageset = &boot_pageset;
7806 local_irq_restore(flags);
7809 #ifdef CONFIG_MEMORY_HOTREMOVE
7811 * All pages in the range must be in a single zone and isolated
7812 * before calling this.
7815 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7819 unsigned int order, i;
7821 unsigned long flags;
7822 /* find the first valid pfn */
7823 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7828 offline_mem_sections(pfn, end_pfn);
7829 zone = page_zone(pfn_to_page(pfn));
7830 spin_lock_irqsave(&zone->lock, flags);
7832 while (pfn < end_pfn) {
7833 if (!pfn_valid(pfn)) {
7837 page = pfn_to_page(pfn);
7839 * The HWPoisoned page may be not in buddy system, and
7840 * page_count() is not 0.
7842 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7844 SetPageReserved(page);
7848 BUG_ON(page_count(page));
7849 BUG_ON(!PageBuddy(page));
7850 order = page_order(page);
7851 #ifdef CONFIG_DEBUG_VM
7852 pr_info("remove from free list %lx %d %lx\n",
7853 pfn, 1 << order, end_pfn);
7855 list_del(&page->lru);
7856 rmv_page_order(page);
7857 zone->free_area[order].nr_free--;
7858 for (i = 0; i < (1 << order); i++)
7859 SetPageReserved((page+i));
7860 pfn += (1 << order);
7862 spin_unlock_irqrestore(&zone->lock, flags);
7866 bool is_free_buddy_page(struct page *page)
7868 struct zone *zone = page_zone(page);
7869 unsigned long pfn = page_to_pfn(page);
7870 unsigned long flags;
7873 spin_lock_irqsave(&zone->lock, flags);
7874 for (order = 0; order < MAX_ORDER; order++) {
7875 struct page *page_head = page - (pfn & ((1 << order) - 1));
7877 if (PageBuddy(page_head) && page_order(page_head) >= order)
7880 spin_unlock_irqrestore(&zone->lock, flags);
7882 return order < MAX_ORDER;