2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
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 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
88 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
89 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
90 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
91 * defined in <linux/topology.h>.
93 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
94 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
95 int _node_numa_mem_[MAX_NUMNODES];
98 /* work_structs for global per-cpu drains */
99 DEFINE_MUTEX(pcpu_drain_mutex);
100 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
102 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
103 volatile unsigned long latent_entropy __latent_entropy;
104 EXPORT_SYMBOL(latent_entropy);
108 * Array of node states.
110 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
111 [N_POSSIBLE] = NODE_MASK_ALL,
112 [N_ONLINE] = { { [0] = 1UL } },
114 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
115 #ifdef CONFIG_HIGHMEM
116 [N_HIGH_MEMORY] = { { [0] = 1UL } },
118 [N_MEMORY] = { { [0] = 1UL } },
119 [N_CPU] = { { [0] = 1UL } },
122 EXPORT_SYMBOL(node_states);
124 /* Protect totalram_pages and zone->managed_pages */
125 static DEFINE_SPINLOCK(managed_page_count_lock);
127 unsigned long totalram_pages __read_mostly;
128 unsigned long totalreserve_pages __read_mostly;
129 unsigned long totalcma_pages __read_mostly;
131 int percpu_pagelist_fraction;
132 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
135 * A cached value of the page's pageblock's migratetype, used when the page is
136 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
137 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
138 * Also the migratetype set in the page does not necessarily match the pcplist
139 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
140 * other index - this ensures that it will be put on the correct CMA freelist.
142 static inline int get_pcppage_migratetype(struct page *page)
147 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
149 page->index = migratetype;
152 #ifdef CONFIG_PM_SLEEP
154 * The following functions are used by the suspend/hibernate code to temporarily
155 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
156 * while devices are suspended. To avoid races with the suspend/hibernate code,
157 * they should always be called with pm_mutex held (gfp_allowed_mask also should
158 * only be modified with pm_mutex held, unless the suspend/hibernate code is
159 * guaranteed not to run in parallel with that modification).
162 static gfp_t saved_gfp_mask;
164 void pm_restore_gfp_mask(void)
166 WARN_ON(!mutex_is_locked(&pm_mutex));
167 if (saved_gfp_mask) {
168 gfp_allowed_mask = saved_gfp_mask;
173 void pm_restrict_gfp_mask(void)
175 WARN_ON(!mutex_is_locked(&pm_mutex));
176 WARN_ON(saved_gfp_mask);
177 saved_gfp_mask = gfp_allowed_mask;
178 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
181 bool pm_suspended_storage(void)
183 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
187 #endif /* CONFIG_PM_SLEEP */
189 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
190 unsigned int pageblock_order __read_mostly;
193 static void __free_pages_ok(struct page *page, unsigned int order);
196 * results with 256, 32 in the lowmem_reserve sysctl:
197 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
198 * 1G machine -> (16M dma, 784M normal, 224M high)
199 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
200 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
201 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 * TBD: should special case ZONE_DMA32 machines here - in those we normally
204 * don't need any ZONE_NORMAL reservation
206 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
207 #ifdef CONFIG_ZONE_DMA
210 #ifdef CONFIG_ZONE_DMA32
213 #ifdef CONFIG_HIGHMEM
219 EXPORT_SYMBOL(totalram_pages);
221 static char * const zone_names[MAX_NR_ZONES] = {
222 #ifdef CONFIG_ZONE_DMA
225 #ifdef CONFIG_ZONE_DMA32
229 #ifdef CONFIG_HIGHMEM
233 #ifdef CONFIG_ZONE_DEVICE
238 char * const migratetype_names[MIGRATE_TYPES] = {
246 #ifdef CONFIG_MEMORY_ISOLATION
251 compound_page_dtor * const compound_page_dtors[] = {
254 #ifdef CONFIG_HUGETLB_PAGE
257 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
262 int min_free_kbytes = 1024;
263 int user_min_free_kbytes = -1;
264 int watermark_scale_factor = 10;
266 static unsigned long __meminitdata nr_kernel_pages;
267 static unsigned long __meminitdata nr_all_pages;
268 static unsigned long __meminitdata dma_reserve;
270 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
271 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
272 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
273 static unsigned long __initdata required_kernelcore;
274 static unsigned long __initdata required_movablecore;
275 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
276 static bool mirrored_kernelcore;
278 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
280 EXPORT_SYMBOL(movable_zone);
281 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
284 int nr_node_ids __read_mostly = MAX_NUMNODES;
285 int nr_online_nodes __read_mostly = 1;
286 EXPORT_SYMBOL(nr_node_ids);
287 EXPORT_SYMBOL(nr_online_nodes);
290 int page_group_by_mobility_disabled __read_mostly;
292 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
295 * Determine how many pages need to be initialized durig early boot
296 * (non-deferred initialization).
297 * The value of first_deferred_pfn will be set later, once non-deferred pages
298 * are initialized, but for now set it ULONG_MAX.
300 static inline void reset_deferred_meminit(pg_data_t *pgdat)
302 phys_addr_t start_addr, end_addr;
303 unsigned long max_pgcnt;
304 unsigned long reserved;
307 * Initialise at least 2G of a node but also take into account that
308 * two large system hashes that can take up 1GB for 0.25TB/node.
310 max_pgcnt = max(2UL << (30 - PAGE_SHIFT),
311 (pgdat->node_spanned_pages >> 8));
314 * Compensate the all the memblock reservations (e.g. crash kernel)
315 * from the initial estimation to make sure we will initialize enough
318 start_addr = PFN_PHYS(pgdat->node_start_pfn);
319 end_addr = PFN_PHYS(pgdat->node_start_pfn + max_pgcnt);
320 reserved = memblock_reserved_memory_within(start_addr, end_addr);
321 max_pgcnt += PHYS_PFN(reserved);
323 pgdat->static_init_pgcnt = min(max_pgcnt, pgdat->node_spanned_pages);
324 pgdat->first_deferred_pfn = ULONG_MAX;
327 /* Returns true if the struct page for the pfn is uninitialised */
328 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
330 int nid = early_pfn_to_nid(pfn);
332 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
339 * Returns false when the remaining initialisation should be deferred until
340 * later in the boot cycle when it can be parallelised.
342 static inline bool update_defer_init(pg_data_t *pgdat,
343 unsigned long pfn, unsigned long zone_end,
344 unsigned long *nr_initialised)
346 /* Always populate low zones for address-contrained allocations */
347 if (zone_end < pgdat_end_pfn(pgdat))
350 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
351 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
352 pgdat->first_deferred_pfn = pfn;
359 static inline void reset_deferred_meminit(pg_data_t *pgdat)
363 static inline bool early_page_uninitialised(unsigned long pfn)
368 static inline bool update_defer_init(pg_data_t *pgdat,
369 unsigned long pfn, unsigned long zone_end,
370 unsigned long *nr_initialised)
376 /* Return a pointer to the bitmap storing bits affecting a block of pages */
377 static inline unsigned long *get_pageblock_bitmap(struct page *page,
380 #ifdef CONFIG_SPARSEMEM
381 return __pfn_to_section(pfn)->pageblock_flags;
383 return page_zone(page)->pageblock_flags;
384 #endif /* CONFIG_SPARSEMEM */
387 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
389 #ifdef CONFIG_SPARSEMEM
390 pfn &= (PAGES_PER_SECTION-1);
391 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
393 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
394 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
395 #endif /* CONFIG_SPARSEMEM */
399 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
400 * @page: The page within the block of interest
401 * @pfn: The target page frame number
402 * @end_bitidx: The last bit of interest to retrieve
403 * @mask: mask of bits that the caller is interested in
405 * Return: pageblock_bits flags
407 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
409 unsigned long end_bitidx,
412 unsigned long *bitmap;
413 unsigned long bitidx, word_bitidx;
416 bitmap = get_pageblock_bitmap(page, pfn);
417 bitidx = pfn_to_bitidx(page, pfn);
418 word_bitidx = bitidx / BITS_PER_LONG;
419 bitidx &= (BITS_PER_LONG-1);
421 word = bitmap[word_bitidx];
422 bitidx += end_bitidx;
423 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
426 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
427 unsigned long end_bitidx,
430 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
433 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
435 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
439 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
440 * @page: The page within the block of interest
441 * @flags: The flags to set
442 * @pfn: The target page frame number
443 * @end_bitidx: The last bit of interest
444 * @mask: mask of bits that the caller is interested in
446 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
448 unsigned long end_bitidx,
451 unsigned long *bitmap;
452 unsigned long bitidx, word_bitidx;
453 unsigned long old_word, word;
455 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
457 bitmap = get_pageblock_bitmap(page, pfn);
458 bitidx = pfn_to_bitidx(page, pfn);
459 word_bitidx = bitidx / BITS_PER_LONG;
460 bitidx &= (BITS_PER_LONG-1);
462 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
464 bitidx += end_bitidx;
465 mask <<= (BITS_PER_LONG - bitidx - 1);
466 flags <<= (BITS_PER_LONG - bitidx - 1);
468 word = READ_ONCE(bitmap[word_bitidx]);
470 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
471 if (word == old_word)
477 void set_pageblock_migratetype(struct page *page, int migratetype)
479 if (unlikely(page_group_by_mobility_disabled &&
480 migratetype < MIGRATE_PCPTYPES))
481 migratetype = MIGRATE_UNMOVABLE;
483 set_pageblock_flags_group(page, (unsigned long)migratetype,
484 PB_migrate, PB_migrate_end);
487 #ifdef CONFIG_DEBUG_VM
488 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
492 unsigned long pfn = page_to_pfn(page);
493 unsigned long sp, start_pfn;
496 seq = zone_span_seqbegin(zone);
497 start_pfn = zone->zone_start_pfn;
498 sp = zone->spanned_pages;
499 if (!zone_spans_pfn(zone, pfn))
501 } while (zone_span_seqretry(zone, seq));
504 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
505 pfn, zone_to_nid(zone), zone->name,
506 start_pfn, start_pfn + sp);
511 static int page_is_consistent(struct zone *zone, struct page *page)
513 if (!pfn_valid_within(page_to_pfn(page)))
515 if (zone != page_zone(page))
521 * Temporary debugging check for pages not lying within a given zone.
523 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
525 if (page_outside_zone_boundaries(zone, page))
527 if (!page_is_consistent(zone, page))
533 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
539 static void bad_page(struct page *page, const char *reason,
540 unsigned long bad_flags)
542 static unsigned long resume;
543 static unsigned long nr_shown;
544 static unsigned long nr_unshown;
547 * Allow a burst of 60 reports, then keep quiet for that minute;
548 * or allow a steady drip of one report per second.
550 if (nr_shown == 60) {
551 if (time_before(jiffies, resume)) {
557 "BUG: Bad page state: %lu messages suppressed\n",
564 resume = jiffies + 60 * HZ;
566 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
567 current->comm, page_to_pfn(page));
568 __dump_page(page, reason);
569 bad_flags &= page->flags;
571 pr_alert("bad because of flags: %#lx(%pGp)\n",
572 bad_flags, &bad_flags);
573 dump_page_owner(page);
578 /* Leave bad fields for debug, except PageBuddy could make trouble */
579 page_mapcount_reset(page); /* remove PageBuddy */
580 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
584 * Higher-order pages are called "compound pages". They are structured thusly:
586 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
588 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
589 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
591 * The first tail page's ->compound_dtor holds the offset in array of compound
592 * page destructors. See compound_page_dtors.
594 * The first tail page's ->compound_order holds the order of allocation.
595 * This usage means that zero-order pages may not be compound.
598 void free_compound_page(struct page *page)
600 __free_pages_ok(page, compound_order(page));
603 void prep_compound_page(struct page *page, unsigned int order)
606 int nr_pages = 1 << order;
608 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
609 set_compound_order(page, order);
611 for (i = 1; i < nr_pages; i++) {
612 struct page *p = page + i;
613 set_page_count(p, 0);
614 p->mapping = TAIL_MAPPING;
615 set_compound_head(p, page);
617 atomic_set(compound_mapcount_ptr(page), -1);
620 #ifdef CONFIG_DEBUG_PAGEALLOC
621 unsigned int _debug_guardpage_minorder;
622 bool _debug_pagealloc_enabled __read_mostly
623 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
624 EXPORT_SYMBOL(_debug_pagealloc_enabled);
625 bool _debug_guardpage_enabled __read_mostly;
627 static int __init early_debug_pagealloc(char *buf)
631 return kstrtobool(buf, &_debug_pagealloc_enabled);
633 early_param("debug_pagealloc", early_debug_pagealloc);
635 static bool need_debug_guardpage(void)
637 /* If we don't use debug_pagealloc, we don't need guard page */
638 if (!debug_pagealloc_enabled())
641 if (!debug_guardpage_minorder())
647 static void init_debug_guardpage(void)
649 if (!debug_pagealloc_enabled())
652 if (!debug_guardpage_minorder())
655 _debug_guardpage_enabled = true;
658 struct page_ext_operations debug_guardpage_ops = {
659 .need = need_debug_guardpage,
660 .init = init_debug_guardpage,
663 static int __init debug_guardpage_minorder_setup(char *buf)
667 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
668 pr_err("Bad debug_guardpage_minorder value\n");
671 _debug_guardpage_minorder = res;
672 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
675 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
677 static inline bool set_page_guard(struct zone *zone, struct page *page,
678 unsigned int order, int migratetype)
680 struct page_ext *page_ext;
682 if (!debug_guardpage_enabled())
685 if (order >= debug_guardpage_minorder())
688 page_ext = lookup_page_ext(page);
689 if (unlikely(!page_ext))
692 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
694 INIT_LIST_HEAD(&page->lru);
695 set_page_private(page, order);
696 /* Guard pages are not available for any usage */
697 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
702 static inline void clear_page_guard(struct zone *zone, struct page *page,
703 unsigned int order, int migratetype)
705 struct page_ext *page_ext;
707 if (!debug_guardpage_enabled())
710 page_ext = lookup_page_ext(page);
711 if (unlikely(!page_ext))
714 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
716 set_page_private(page, 0);
717 if (!is_migrate_isolate(migratetype))
718 __mod_zone_freepage_state(zone, (1 << order), migratetype);
721 struct page_ext_operations debug_guardpage_ops;
722 static inline bool set_page_guard(struct zone *zone, struct page *page,
723 unsigned int order, int migratetype) { return false; }
724 static inline void clear_page_guard(struct zone *zone, struct page *page,
725 unsigned int order, int migratetype) {}
728 static inline void set_page_order(struct page *page, unsigned int order)
730 set_page_private(page, order);
731 __SetPageBuddy(page);
734 static inline void rmv_page_order(struct page *page)
736 __ClearPageBuddy(page);
737 set_page_private(page, 0);
741 * This function checks whether a page is free && is the buddy
742 * we can do coalesce a page and its buddy if
743 * (a) the buddy is not in a hole (check before calling!) &&
744 * (b) the buddy is in the buddy system &&
745 * (c) a page and its buddy have the same order &&
746 * (d) a page and its buddy are in the same zone.
748 * For recording whether a page is in the buddy system, we set ->_mapcount
749 * PAGE_BUDDY_MAPCOUNT_VALUE.
750 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
751 * serialized by zone->lock.
753 * For recording page's order, we use page_private(page).
755 static inline int page_is_buddy(struct page *page, struct page *buddy,
758 if (page_is_guard(buddy) && page_order(buddy) == order) {
759 if (page_zone_id(page) != page_zone_id(buddy))
762 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
767 if (PageBuddy(buddy) && page_order(buddy) == order) {
769 * zone check is done late to avoid uselessly
770 * calculating zone/node ids for pages that could
773 if (page_zone_id(page) != page_zone_id(buddy))
776 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
784 * Freeing function for a buddy system allocator.
786 * The concept of a buddy system is to maintain direct-mapped table
787 * (containing bit values) for memory blocks of various "orders".
788 * The bottom level table contains the map for the smallest allocatable
789 * units of memory (here, pages), and each level above it describes
790 * pairs of units from the levels below, hence, "buddies".
791 * At a high level, all that happens here is marking the table entry
792 * at the bottom level available, and propagating the changes upward
793 * as necessary, plus some accounting needed to play nicely with other
794 * parts of the VM system.
795 * At each level, we keep a list of pages, which are heads of continuous
796 * free pages of length of (1 << order) and marked with _mapcount
797 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
799 * So when we are allocating or freeing one, we can derive the state of the
800 * other. That is, if we allocate a small block, and both were
801 * free, the remainder of the region must be split into blocks.
802 * If a block is freed, and its buddy is also free, then this
803 * triggers coalescing into a block of larger size.
808 static inline void __free_one_page(struct page *page,
810 struct zone *zone, unsigned int order,
813 unsigned long combined_pfn;
814 unsigned long uninitialized_var(buddy_pfn);
816 unsigned int max_order;
818 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
820 VM_BUG_ON(!zone_is_initialized(zone));
821 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
823 VM_BUG_ON(migratetype == -1);
824 if (likely(!is_migrate_isolate(migratetype)))
825 __mod_zone_freepage_state(zone, 1 << order, migratetype);
827 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
828 VM_BUG_ON_PAGE(bad_range(zone, page), page);
831 while (order < max_order - 1) {
832 buddy_pfn = __find_buddy_pfn(pfn, order);
833 buddy = page + (buddy_pfn - pfn);
835 if (!pfn_valid_within(buddy_pfn))
837 if (!page_is_buddy(page, buddy, order))
840 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
841 * merge with it and move up one order.
843 if (page_is_guard(buddy)) {
844 clear_page_guard(zone, buddy, order, migratetype);
846 list_del(&buddy->lru);
847 zone->free_area[order].nr_free--;
848 rmv_page_order(buddy);
850 combined_pfn = buddy_pfn & pfn;
851 page = page + (combined_pfn - pfn);
855 if (max_order < MAX_ORDER) {
856 /* If we are here, it means order is >= pageblock_order.
857 * We want to prevent merge between freepages on isolate
858 * pageblock and normal pageblock. Without this, pageblock
859 * isolation could cause incorrect freepage or CMA accounting.
861 * We don't want to hit this code for the more frequent
864 if (unlikely(has_isolate_pageblock(zone))) {
867 buddy_pfn = __find_buddy_pfn(pfn, order);
868 buddy = page + (buddy_pfn - pfn);
869 buddy_mt = get_pageblock_migratetype(buddy);
871 if (migratetype != buddy_mt
872 && (is_migrate_isolate(migratetype) ||
873 is_migrate_isolate(buddy_mt)))
877 goto continue_merging;
881 set_page_order(page, order);
884 * If this is not the largest possible page, check if the buddy
885 * of the next-highest order is free. If it is, it's possible
886 * that pages are being freed that will coalesce soon. In case,
887 * that is happening, add the free page to the tail of the list
888 * so it's less likely to be used soon and more likely to be merged
889 * as a higher order page
891 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
892 struct page *higher_page, *higher_buddy;
893 combined_pfn = buddy_pfn & pfn;
894 higher_page = page + (combined_pfn - pfn);
895 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
896 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
897 if (pfn_valid_within(buddy_pfn) &&
898 page_is_buddy(higher_page, higher_buddy, order + 1)) {
899 list_add_tail(&page->lru,
900 &zone->free_area[order].free_list[migratetype]);
905 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
907 zone->free_area[order].nr_free++;
911 * A bad page could be due to a number of fields. Instead of multiple branches,
912 * try and check multiple fields with one check. The caller must do a detailed
913 * check if necessary.
915 static inline bool page_expected_state(struct page *page,
916 unsigned long check_flags)
918 if (unlikely(atomic_read(&page->_mapcount) != -1))
921 if (unlikely((unsigned long)page->mapping |
922 page_ref_count(page) |
924 (unsigned long)page->mem_cgroup |
926 (page->flags & check_flags)))
932 static void free_pages_check_bad(struct page *page)
934 const char *bad_reason;
935 unsigned long bad_flags;
940 if (unlikely(atomic_read(&page->_mapcount) != -1))
941 bad_reason = "nonzero mapcount";
942 if (unlikely(page->mapping != NULL))
943 bad_reason = "non-NULL mapping";
944 if (unlikely(page_ref_count(page) != 0))
945 bad_reason = "nonzero _refcount";
946 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
947 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
948 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
951 if (unlikely(page->mem_cgroup))
952 bad_reason = "page still charged to cgroup";
954 bad_page(page, bad_reason, bad_flags);
957 static inline int free_pages_check(struct page *page)
959 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
962 /* Something has gone sideways, find it */
963 free_pages_check_bad(page);
967 static int free_tail_pages_check(struct page *head_page, struct page *page)
972 * We rely page->lru.next never has bit 0 set, unless the page
973 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
975 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
977 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
981 switch (page - head_page) {
983 /* the first tail page: ->mapping is compound_mapcount() */
984 if (unlikely(compound_mapcount(page))) {
985 bad_page(page, "nonzero compound_mapcount", 0);
991 * the second tail page: ->mapping is
992 * page_deferred_list().next -- ignore value.
996 if (page->mapping != TAIL_MAPPING) {
997 bad_page(page, "corrupted mapping in tail page", 0);
1002 if (unlikely(!PageTail(page))) {
1003 bad_page(page, "PageTail not set", 0);
1006 if (unlikely(compound_head(page) != head_page)) {
1007 bad_page(page, "compound_head not consistent", 0);
1012 page->mapping = NULL;
1013 clear_compound_head(page);
1017 static __always_inline bool free_pages_prepare(struct page *page,
1018 unsigned int order, bool check_free)
1022 VM_BUG_ON_PAGE(PageTail(page), page);
1024 trace_mm_page_free(page, order);
1025 kmemcheck_free_shadow(page, order);
1028 * Check tail pages before head page information is cleared to
1029 * avoid checking PageCompound for order-0 pages.
1031 if (unlikely(order)) {
1032 bool compound = PageCompound(page);
1035 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1038 ClearPageDoubleMap(page);
1039 for (i = 1; i < (1 << order); i++) {
1041 bad += free_tail_pages_check(page, page + i);
1042 if (unlikely(free_pages_check(page + i))) {
1046 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1049 if (PageMappingFlags(page))
1050 page->mapping = NULL;
1051 if (memcg_kmem_enabled() && PageKmemcg(page))
1052 memcg_kmem_uncharge(page, order);
1054 bad += free_pages_check(page);
1058 page_cpupid_reset_last(page);
1059 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1060 reset_page_owner(page, order);
1062 if (!PageHighMem(page)) {
1063 debug_check_no_locks_freed(page_address(page),
1064 PAGE_SIZE << order);
1065 debug_check_no_obj_freed(page_address(page),
1066 PAGE_SIZE << order);
1068 arch_free_page(page, order);
1069 kernel_poison_pages(page, 1 << order, 0);
1070 kernel_map_pages(page, 1 << order, 0);
1071 kasan_free_pages(page, order);
1076 #ifdef CONFIG_DEBUG_VM
1077 static inline bool free_pcp_prepare(struct page *page)
1079 return free_pages_prepare(page, 0, true);
1082 static inline bool bulkfree_pcp_prepare(struct page *page)
1087 static bool free_pcp_prepare(struct page *page)
1089 return free_pages_prepare(page, 0, false);
1092 static bool bulkfree_pcp_prepare(struct page *page)
1094 return free_pages_check(page);
1096 #endif /* CONFIG_DEBUG_VM */
1099 * Frees a number of pages from the PCP lists
1100 * Assumes all pages on list are in same zone, and of same order.
1101 * count is the number of pages to free.
1103 * If the zone was previously in an "all pages pinned" state then look to
1104 * see if this freeing clears that state.
1106 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1107 * pinned" detection logic.
1109 static void free_pcppages_bulk(struct zone *zone, int count,
1110 struct per_cpu_pages *pcp)
1112 int migratetype = 0;
1114 bool isolated_pageblocks;
1116 spin_lock(&zone->lock);
1117 isolated_pageblocks = has_isolate_pageblock(zone);
1121 struct list_head *list;
1124 * Remove pages from lists in a round-robin fashion. A
1125 * batch_free count is maintained that is incremented when an
1126 * empty list is encountered. This is so more pages are freed
1127 * off fuller lists instead of spinning excessively around empty
1132 if (++migratetype == MIGRATE_PCPTYPES)
1134 list = &pcp->lists[migratetype];
1135 } while (list_empty(list));
1137 /* This is the only non-empty list. Free them all. */
1138 if (batch_free == MIGRATE_PCPTYPES)
1142 int mt; /* migratetype of the to-be-freed page */
1144 page = list_last_entry(list, struct page, lru);
1145 /* must delete as __free_one_page list manipulates */
1146 list_del(&page->lru);
1148 mt = get_pcppage_migratetype(page);
1149 /* MIGRATE_ISOLATE page should not go to pcplists */
1150 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1151 /* Pageblock could have been isolated meanwhile */
1152 if (unlikely(isolated_pageblocks))
1153 mt = get_pageblock_migratetype(page);
1155 if (bulkfree_pcp_prepare(page))
1158 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1159 trace_mm_page_pcpu_drain(page, 0, mt);
1160 } while (--count && --batch_free && !list_empty(list));
1162 spin_unlock(&zone->lock);
1165 static void free_one_page(struct zone *zone,
1166 struct page *page, unsigned long pfn,
1170 spin_lock(&zone->lock);
1171 if (unlikely(has_isolate_pageblock(zone) ||
1172 is_migrate_isolate(migratetype))) {
1173 migratetype = get_pfnblock_migratetype(page, pfn);
1175 __free_one_page(page, pfn, zone, order, migratetype);
1176 spin_unlock(&zone->lock);
1179 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1180 unsigned long zone, int nid)
1182 set_page_links(page, zone, nid, pfn);
1183 init_page_count(page);
1184 page_mapcount_reset(page);
1185 page_cpupid_reset_last(page);
1187 INIT_LIST_HEAD(&page->lru);
1188 #ifdef WANT_PAGE_VIRTUAL
1189 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1190 if (!is_highmem_idx(zone))
1191 set_page_address(page, __va(pfn << PAGE_SHIFT));
1195 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1198 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1201 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1202 static void __meminit init_reserved_page(unsigned long pfn)
1207 if (!early_page_uninitialised(pfn))
1210 nid = early_pfn_to_nid(pfn);
1211 pgdat = NODE_DATA(nid);
1213 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1214 struct zone *zone = &pgdat->node_zones[zid];
1216 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1219 __init_single_pfn(pfn, zid, nid);
1222 static inline void init_reserved_page(unsigned long pfn)
1225 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1228 * Initialised pages do not have PageReserved set. This function is
1229 * called for each range allocated by the bootmem allocator and
1230 * marks the pages PageReserved. The remaining valid pages are later
1231 * sent to the buddy page allocator.
1233 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1235 unsigned long start_pfn = PFN_DOWN(start);
1236 unsigned long end_pfn = PFN_UP(end);
1238 for (; start_pfn < end_pfn; start_pfn++) {
1239 if (pfn_valid(start_pfn)) {
1240 struct page *page = pfn_to_page(start_pfn);
1242 init_reserved_page(start_pfn);
1244 /* Avoid false-positive PageTail() */
1245 INIT_LIST_HEAD(&page->lru);
1247 SetPageReserved(page);
1252 static void __free_pages_ok(struct page *page, unsigned int order)
1254 unsigned long flags;
1256 unsigned long pfn = page_to_pfn(page);
1258 if (!free_pages_prepare(page, order, true))
1261 migratetype = get_pfnblock_migratetype(page, pfn);
1262 local_irq_save(flags);
1263 __count_vm_events(PGFREE, 1 << order);
1264 free_one_page(page_zone(page), page, pfn, order, migratetype);
1265 local_irq_restore(flags);
1268 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1270 unsigned int nr_pages = 1 << order;
1271 struct page *p = page;
1275 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1277 __ClearPageReserved(p);
1278 set_page_count(p, 0);
1280 __ClearPageReserved(p);
1281 set_page_count(p, 0);
1283 page_zone(page)->managed_pages += nr_pages;
1284 set_page_refcounted(page);
1285 __free_pages(page, order);
1288 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1289 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1291 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1293 int __meminit early_pfn_to_nid(unsigned long pfn)
1295 static DEFINE_SPINLOCK(early_pfn_lock);
1298 spin_lock(&early_pfn_lock);
1299 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1301 nid = first_online_node;
1302 spin_unlock(&early_pfn_lock);
1308 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1309 static inline bool __meminit __maybe_unused
1310 meminit_pfn_in_nid(unsigned long pfn, int node,
1311 struct mminit_pfnnid_cache *state)
1315 nid = __early_pfn_to_nid(pfn, state);
1316 if (nid >= 0 && nid != node)
1321 /* Only safe to use early in boot when initialisation is single-threaded */
1322 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1324 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1329 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1333 static inline bool __meminit __maybe_unused
1334 meminit_pfn_in_nid(unsigned long pfn, int node,
1335 struct mminit_pfnnid_cache *state)
1342 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1345 if (early_page_uninitialised(pfn))
1347 return __free_pages_boot_core(page, order);
1351 * Check that the whole (or subset of) a pageblock given by the interval of
1352 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1353 * with the migration of free compaction scanner. The scanners then need to
1354 * use only pfn_valid_within() check for arches that allow holes within
1357 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1359 * It's possible on some configurations to have a setup like node0 node1 node0
1360 * i.e. it's possible that all pages within a zones range of pages do not
1361 * belong to a single zone. We assume that a border between node0 and node1
1362 * can occur within a single pageblock, but not a node0 node1 node0
1363 * interleaving within a single pageblock. It is therefore sufficient to check
1364 * the first and last page of a pageblock and avoid checking each individual
1365 * page in a pageblock.
1367 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1368 unsigned long end_pfn, struct zone *zone)
1370 struct page *start_page;
1371 struct page *end_page;
1373 /* end_pfn is one past the range we are checking */
1376 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1379 start_page = pfn_to_online_page(start_pfn);
1383 if (page_zone(start_page) != zone)
1386 end_page = pfn_to_page(end_pfn);
1388 /* This gives a shorter code than deriving page_zone(end_page) */
1389 if (page_zone_id(start_page) != page_zone_id(end_page))
1395 void set_zone_contiguous(struct zone *zone)
1397 unsigned long block_start_pfn = zone->zone_start_pfn;
1398 unsigned long block_end_pfn;
1400 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1401 for (; block_start_pfn < zone_end_pfn(zone);
1402 block_start_pfn = block_end_pfn,
1403 block_end_pfn += pageblock_nr_pages) {
1405 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1407 if (!__pageblock_pfn_to_page(block_start_pfn,
1408 block_end_pfn, zone))
1412 /* We confirm that there is no hole */
1413 zone->contiguous = true;
1416 void clear_zone_contiguous(struct zone *zone)
1418 zone->contiguous = false;
1421 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1422 static void __init deferred_free_range(struct page *page,
1423 unsigned long pfn, int nr_pages)
1430 /* Free a large naturally-aligned chunk if possible */
1431 if (nr_pages == pageblock_nr_pages &&
1432 (pfn & (pageblock_nr_pages - 1)) == 0) {
1433 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1434 __free_pages_boot_core(page, pageblock_order);
1438 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1439 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1440 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1441 __free_pages_boot_core(page, 0);
1445 /* Completion tracking for deferred_init_memmap() threads */
1446 static atomic_t pgdat_init_n_undone __initdata;
1447 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1449 static inline void __init pgdat_init_report_one_done(void)
1451 if (atomic_dec_and_test(&pgdat_init_n_undone))
1452 complete(&pgdat_init_all_done_comp);
1455 /* Initialise remaining memory on a node */
1456 static int __init deferred_init_memmap(void *data)
1458 pg_data_t *pgdat = data;
1459 int nid = pgdat->node_id;
1460 struct mminit_pfnnid_cache nid_init_state = { };
1461 unsigned long start = jiffies;
1462 unsigned long nr_pages = 0;
1463 unsigned long walk_start, walk_end;
1466 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1467 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1469 if (first_init_pfn == ULONG_MAX) {
1470 pgdat_init_report_one_done();
1474 /* Bind memory initialisation thread to a local node if possible */
1475 if (!cpumask_empty(cpumask))
1476 set_cpus_allowed_ptr(current, cpumask);
1478 /* Sanity check boundaries */
1479 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1480 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1481 pgdat->first_deferred_pfn = ULONG_MAX;
1483 /* Only the highest zone is deferred so find it */
1484 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1485 zone = pgdat->node_zones + zid;
1486 if (first_init_pfn < zone_end_pfn(zone))
1490 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1491 unsigned long pfn, end_pfn;
1492 struct page *page = NULL;
1493 struct page *free_base_page = NULL;
1494 unsigned long free_base_pfn = 0;
1497 end_pfn = min(walk_end, zone_end_pfn(zone));
1498 pfn = first_init_pfn;
1499 if (pfn < walk_start)
1501 if (pfn < zone->zone_start_pfn)
1502 pfn = zone->zone_start_pfn;
1504 for (; pfn < end_pfn; pfn++) {
1505 if (!pfn_valid_within(pfn))
1509 * Ensure pfn_valid is checked every
1510 * pageblock_nr_pages for memory holes
1512 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1513 if (!pfn_valid(pfn)) {
1519 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1524 /* Minimise pfn page lookups and scheduler checks */
1525 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1528 nr_pages += nr_to_free;
1529 deferred_free_range(free_base_page,
1530 free_base_pfn, nr_to_free);
1531 free_base_page = NULL;
1532 free_base_pfn = nr_to_free = 0;
1534 page = pfn_to_page(pfn);
1539 VM_BUG_ON(page_zone(page) != zone);
1543 __init_single_page(page, pfn, zid, nid);
1544 if (!free_base_page) {
1545 free_base_page = page;
1546 free_base_pfn = pfn;
1551 /* Where possible, batch up pages for a single free */
1554 /* Free the current block of pages to allocator */
1555 nr_pages += nr_to_free;
1556 deferred_free_range(free_base_page, free_base_pfn,
1558 free_base_page = NULL;
1559 free_base_pfn = nr_to_free = 0;
1561 /* Free the last block of pages to allocator */
1562 nr_pages += nr_to_free;
1563 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1565 first_init_pfn = max(end_pfn, first_init_pfn);
1568 /* Sanity check that the next zone really is unpopulated */
1569 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1571 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1572 jiffies_to_msecs(jiffies - start));
1574 pgdat_init_report_one_done();
1577 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1579 void __init page_alloc_init_late(void)
1583 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1586 /* There will be num_node_state(N_MEMORY) threads */
1587 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1588 for_each_node_state(nid, N_MEMORY) {
1589 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1592 /* Block until all are initialised */
1593 wait_for_completion(&pgdat_init_all_done_comp);
1595 /* Reinit limits that are based on free pages after the kernel is up */
1596 files_maxfiles_init();
1598 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1599 /* Discard memblock private memory */
1603 for_each_populated_zone(zone)
1604 set_zone_contiguous(zone);
1608 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1609 void __init init_cma_reserved_pageblock(struct page *page)
1611 unsigned i = pageblock_nr_pages;
1612 struct page *p = page;
1615 __ClearPageReserved(p);
1616 set_page_count(p, 0);
1619 set_pageblock_migratetype(page, MIGRATE_CMA);
1621 if (pageblock_order >= MAX_ORDER) {
1622 i = pageblock_nr_pages;
1625 set_page_refcounted(p);
1626 __free_pages(p, MAX_ORDER - 1);
1627 p += MAX_ORDER_NR_PAGES;
1628 } while (i -= MAX_ORDER_NR_PAGES);
1630 set_page_refcounted(page);
1631 __free_pages(page, pageblock_order);
1634 adjust_managed_page_count(page, pageblock_nr_pages);
1639 * The order of subdivision here is critical for the IO subsystem.
1640 * Please do not alter this order without good reasons and regression
1641 * testing. Specifically, as large blocks of memory are subdivided,
1642 * the order in which smaller blocks are delivered depends on the order
1643 * they're subdivided in this function. This is the primary factor
1644 * influencing the order in which pages are delivered to the IO
1645 * subsystem according to empirical testing, and this is also justified
1646 * by considering the behavior of a buddy system containing a single
1647 * large block of memory acted on by a series of small allocations.
1648 * This behavior is a critical factor in sglist merging's success.
1652 static inline void expand(struct zone *zone, struct page *page,
1653 int low, int high, struct free_area *area,
1656 unsigned long size = 1 << high;
1658 while (high > low) {
1662 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1665 * Mark as guard pages (or page), that will allow to
1666 * merge back to allocator when buddy will be freed.
1667 * Corresponding page table entries will not be touched,
1668 * pages will stay not present in virtual address space
1670 if (set_page_guard(zone, &page[size], high, migratetype))
1673 list_add(&page[size].lru, &area->free_list[migratetype]);
1675 set_page_order(&page[size], high);
1679 static void check_new_page_bad(struct page *page)
1681 const char *bad_reason = NULL;
1682 unsigned long bad_flags = 0;
1684 if (unlikely(atomic_read(&page->_mapcount) != -1))
1685 bad_reason = "nonzero mapcount";
1686 if (unlikely(page->mapping != NULL))
1687 bad_reason = "non-NULL mapping";
1688 if (unlikely(page_ref_count(page) != 0))
1689 bad_reason = "nonzero _count";
1690 if (unlikely(page->flags & __PG_HWPOISON)) {
1691 bad_reason = "HWPoisoned (hardware-corrupted)";
1692 bad_flags = __PG_HWPOISON;
1693 /* Don't complain about hwpoisoned pages */
1694 page_mapcount_reset(page); /* remove PageBuddy */
1697 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1698 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1699 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1702 if (unlikely(page->mem_cgroup))
1703 bad_reason = "page still charged to cgroup";
1705 bad_page(page, bad_reason, bad_flags);
1709 * This page is about to be returned from the page allocator
1711 static inline int check_new_page(struct page *page)
1713 if (likely(page_expected_state(page,
1714 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1717 check_new_page_bad(page);
1721 static inline bool free_pages_prezeroed(void)
1723 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1724 page_poisoning_enabled();
1727 #ifdef CONFIG_DEBUG_VM
1728 static bool check_pcp_refill(struct page *page)
1733 static bool check_new_pcp(struct page *page)
1735 return check_new_page(page);
1738 static bool check_pcp_refill(struct page *page)
1740 return check_new_page(page);
1742 static bool check_new_pcp(struct page *page)
1746 #endif /* CONFIG_DEBUG_VM */
1748 static bool check_new_pages(struct page *page, unsigned int order)
1751 for (i = 0; i < (1 << order); i++) {
1752 struct page *p = page + i;
1754 if (unlikely(check_new_page(p)))
1761 inline void post_alloc_hook(struct page *page, unsigned int order,
1764 set_page_private(page, 0);
1765 set_page_refcounted(page);
1767 arch_alloc_page(page, order);
1768 kernel_map_pages(page, 1 << order, 1);
1769 kernel_poison_pages(page, 1 << order, 1);
1770 kasan_alloc_pages(page, order);
1771 set_page_owner(page, order, gfp_flags);
1774 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1775 unsigned int alloc_flags)
1779 post_alloc_hook(page, order, gfp_flags);
1781 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1782 for (i = 0; i < (1 << order); i++)
1783 clear_highpage(page + i);
1785 if (order && (gfp_flags & __GFP_COMP))
1786 prep_compound_page(page, order);
1789 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1790 * allocate the page. The expectation is that the caller is taking
1791 * steps that will free more memory. The caller should avoid the page
1792 * being used for !PFMEMALLOC purposes.
1794 if (alloc_flags & ALLOC_NO_WATERMARKS)
1795 set_page_pfmemalloc(page);
1797 clear_page_pfmemalloc(page);
1801 * Go through the free lists for the given migratetype and remove
1802 * the smallest available page from the freelists
1805 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1808 unsigned int current_order;
1809 struct free_area *area;
1812 /* Find a page of the appropriate size in the preferred list */
1813 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1814 area = &(zone->free_area[current_order]);
1815 page = list_first_entry_or_null(&area->free_list[migratetype],
1819 list_del(&page->lru);
1820 rmv_page_order(page);
1822 expand(zone, page, order, current_order, area, migratetype);
1823 set_pcppage_migratetype(page, migratetype);
1832 * This array describes the order lists are fallen back to when
1833 * the free lists for the desirable migrate type are depleted
1835 static int fallbacks[MIGRATE_TYPES][4] = {
1836 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1837 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1838 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1840 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1842 #ifdef CONFIG_MEMORY_ISOLATION
1843 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1848 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1851 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1854 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1855 unsigned int order) { return NULL; }
1859 * Move the free pages in a range to the free lists of the requested type.
1860 * Note that start_page and end_pages are not aligned on a pageblock
1861 * boundary. If alignment is required, use move_freepages_block()
1863 static int move_freepages(struct zone *zone,
1864 struct page *start_page, struct page *end_page,
1865 int migratetype, int *num_movable)
1869 int pages_moved = 0;
1871 #ifndef CONFIG_HOLES_IN_ZONE
1873 * page_zone is not safe to call in this context when
1874 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1875 * anyway as we check zone boundaries in move_freepages_block().
1876 * Remove at a later date when no bug reports exist related to
1877 * grouping pages by mobility
1879 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1885 for (page = start_page; page <= end_page;) {
1886 if (!pfn_valid_within(page_to_pfn(page))) {
1891 /* Make sure we are not inadvertently changing nodes */
1892 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1894 if (!PageBuddy(page)) {
1896 * We assume that pages that could be isolated for
1897 * migration are movable. But we don't actually try
1898 * isolating, as that would be expensive.
1901 (PageLRU(page) || __PageMovable(page)))
1908 order = page_order(page);
1909 list_move(&page->lru,
1910 &zone->free_area[order].free_list[migratetype]);
1912 pages_moved += 1 << order;
1918 int move_freepages_block(struct zone *zone, struct page *page,
1919 int migratetype, int *num_movable)
1921 unsigned long start_pfn, end_pfn;
1922 struct page *start_page, *end_page;
1924 start_pfn = page_to_pfn(page);
1925 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1926 start_page = pfn_to_page(start_pfn);
1927 end_page = start_page + pageblock_nr_pages - 1;
1928 end_pfn = start_pfn + pageblock_nr_pages - 1;
1930 /* Do not cross zone boundaries */
1931 if (!zone_spans_pfn(zone, start_pfn))
1933 if (!zone_spans_pfn(zone, end_pfn))
1936 return move_freepages(zone, start_page, end_page, migratetype,
1940 static void change_pageblock_range(struct page *pageblock_page,
1941 int start_order, int migratetype)
1943 int nr_pageblocks = 1 << (start_order - pageblock_order);
1945 while (nr_pageblocks--) {
1946 set_pageblock_migratetype(pageblock_page, migratetype);
1947 pageblock_page += pageblock_nr_pages;
1952 * When we are falling back to another migratetype during allocation, try to
1953 * steal extra free pages from the same pageblocks to satisfy further
1954 * allocations, instead of polluting multiple pageblocks.
1956 * If we are stealing a relatively large buddy page, it is likely there will
1957 * be more free pages in the pageblock, so try to steal them all. For
1958 * reclaimable and unmovable allocations, we steal regardless of page size,
1959 * as fragmentation caused by those allocations polluting movable pageblocks
1960 * is worse than movable allocations stealing from unmovable and reclaimable
1963 static bool can_steal_fallback(unsigned int order, int start_mt)
1966 * Leaving this order check is intended, although there is
1967 * relaxed order check in next check. The reason is that
1968 * we can actually steal whole pageblock if this condition met,
1969 * but, below check doesn't guarantee it and that is just heuristic
1970 * so could be changed anytime.
1972 if (order >= pageblock_order)
1975 if (order >= pageblock_order / 2 ||
1976 start_mt == MIGRATE_RECLAIMABLE ||
1977 start_mt == MIGRATE_UNMOVABLE ||
1978 page_group_by_mobility_disabled)
1985 * This function implements actual steal behaviour. If order is large enough,
1986 * we can steal whole pageblock. If not, we first move freepages in this
1987 * pageblock to our migratetype and determine how many already-allocated pages
1988 * are there in the pageblock with a compatible migratetype. If at least half
1989 * of pages are free or compatible, we can change migratetype of the pageblock
1990 * itself, so pages freed in the future will be put on the correct free list.
1992 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1993 int start_type, bool whole_block)
1995 unsigned int current_order = page_order(page);
1996 struct free_area *area;
1997 int free_pages, movable_pages, alike_pages;
2000 old_block_type = get_pageblock_migratetype(page);
2003 * This can happen due to races and we want to prevent broken
2004 * highatomic accounting.
2006 if (is_migrate_highatomic(old_block_type))
2009 /* Take ownership for orders >= pageblock_order */
2010 if (current_order >= pageblock_order) {
2011 change_pageblock_range(page, current_order, start_type);
2015 /* We are not allowed to try stealing from the whole block */
2019 free_pages = move_freepages_block(zone, page, start_type,
2022 * Determine how many pages are compatible with our allocation.
2023 * For movable allocation, it's the number of movable pages which
2024 * we just obtained. For other types it's a bit more tricky.
2026 if (start_type == MIGRATE_MOVABLE) {
2027 alike_pages = movable_pages;
2030 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2031 * to MOVABLE pageblock, consider all non-movable pages as
2032 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2033 * vice versa, be conservative since we can't distinguish the
2034 * exact migratetype of non-movable pages.
2036 if (old_block_type == MIGRATE_MOVABLE)
2037 alike_pages = pageblock_nr_pages
2038 - (free_pages + movable_pages);
2043 /* moving whole block can fail due to zone boundary conditions */
2048 * If a sufficient number of pages in the block are either free or of
2049 * comparable migratability as our allocation, claim the whole block.
2051 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2052 page_group_by_mobility_disabled)
2053 set_pageblock_migratetype(page, start_type);
2058 area = &zone->free_area[current_order];
2059 list_move(&page->lru, &area->free_list[start_type]);
2063 * Check whether there is a suitable fallback freepage with requested order.
2064 * If only_stealable is true, this function returns fallback_mt only if
2065 * we can steal other freepages all together. This would help to reduce
2066 * fragmentation due to mixed migratetype pages in one pageblock.
2068 int find_suitable_fallback(struct free_area *area, unsigned int order,
2069 int migratetype, bool only_stealable, bool *can_steal)
2074 if (area->nr_free == 0)
2079 fallback_mt = fallbacks[migratetype][i];
2080 if (fallback_mt == MIGRATE_TYPES)
2083 if (list_empty(&area->free_list[fallback_mt]))
2086 if (can_steal_fallback(order, migratetype))
2089 if (!only_stealable)
2100 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2101 * there are no empty page blocks that contain a page with a suitable order
2103 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2104 unsigned int alloc_order)
2107 unsigned long max_managed, flags;
2110 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2111 * Check is race-prone but harmless.
2113 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2114 if (zone->nr_reserved_highatomic >= max_managed)
2117 spin_lock_irqsave(&zone->lock, flags);
2119 /* Recheck the nr_reserved_highatomic limit under the lock */
2120 if (zone->nr_reserved_highatomic >= max_managed)
2124 mt = get_pageblock_migratetype(page);
2125 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2126 && !is_migrate_cma(mt)) {
2127 zone->nr_reserved_highatomic += pageblock_nr_pages;
2128 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2129 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2133 spin_unlock_irqrestore(&zone->lock, flags);
2137 * Used when an allocation is about to fail under memory pressure. This
2138 * potentially hurts the reliability of high-order allocations when under
2139 * intense memory pressure but failed atomic allocations should be easier
2140 * to recover from than an OOM.
2142 * If @force is true, try to unreserve a pageblock even though highatomic
2143 * pageblock is exhausted.
2145 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2148 struct zonelist *zonelist = ac->zonelist;
2149 unsigned long flags;
2156 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2159 * Preserve at least one pageblock unless memory pressure
2162 if (!force && zone->nr_reserved_highatomic <=
2166 spin_lock_irqsave(&zone->lock, flags);
2167 for (order = 0; order < MAX_ORDER; order++) {
2168 struct free_area *area = &(zone->free_area[order]);
2170 page = list_first_entry_or_null(
2171 &area->free_list[MIGRATE_HIGHATOMIC],
2177 * In page freeing path, migratetype change is racy so
2178 * we can counter several free pages in a pageblock
2179 * in this loop althoug we changed the pageblock type
2180 * from highatomic to ac->migratetype. So we should
2181 * adjust the count once.
2183 if (is_migrate_highatomic_page(page)) {
2185 * It should never happen but changes to
2186 * locking could inadvertently allow a per-cpu
2187 * drain to add pages to MIGRATE_HIGHATOMIC
2188 * while unreserving so be safe and watch for
2191 zone->nr_reserved_highatomic -= min(
2193 zone->nr_reserved_highatomic);
2197 * Convert to ac->migratetype and avoid the normal
2198 * pageblock stealing heuristics. Minimally, the caller
2199 * is doing the work and needs the pages. More
2200 * importantly, if the block was always converted to
2201 * MIGRATE_UNMOVABLE or another type then the number
2202 * of pageblocks that cannot be completely freed
2205 set_pageblock_migratetype(page, ac->migratetype);
2206 ret = move_freepages_block(zone, page, ac->migratetype,
2209 spin_unlock_irqrestore(&zone->lock, flags);
2213 spin_unlock_irqrestore(&zone->lock, flags);
2220 * Try finding a free buddy page on the fallback list and put it on the free
2221 * list of requested migratetype, possibly along with other pages from the same
2222 * block, depending on fragmentation avoidance heuristics. Returns true if
2223 * fallback was found so that __rmqueue_smallest() can grab it.
2225 * The use of signed ints for order and current_order is a deliberate
2226 * deviation from the rest of this file, to make the for loop
2227 * condition simpler.
2230 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2232 struct free_area *area;
2239 * Find the largest available free page in the other list. This roughly
2240 * approximates finding the pageblock with the most free pages, which
2241 * would be too costly to do exactly.
2243 for (current_order = MAX_ORDER - 1; current_order >= order;
2245 area = &(zone->free_area[current_order]);
2246 fallback_mt = find_suitable_fallback(area, current_order,
2247 start_migratetype, false, &can_steal);
2248 if (fallback_mt == -1)
2252 * We cannot steal all free pages from the pageblock and the
2253 * requested migratetype is movable. In that case it's better to
2254 * steal and split the smallest available page instead of the
2255 * largest available page, because even if the next movable
2256 * allocation falls back into a different pageblock than this
2257 * one, it won't cause permanent fragmentation.
2259 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2260 && current_order > order)
2269 for (current_order = order; current_order < MAX_ORDER;
2271 area = &(zone->free_area[current_order]);
2272 fallback_mt = find_suitable_fallback(area, current_order,
2273 start_migratetype, false, &can_steal);
2274 if (fallback_mt != -1)
2279 * This should not happen - we already found a suitable fallback
2280 * when looking for the largest page.
2282 VM_BUG_ON(current_order == MAX_ORDER);
2285 page = list_first_entry(&area->free_list[fallback_mt],
2288 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2290 trace_mm_page_alloc_extfrag(page, order, current_order,
2291 start_migratetype, fallback_mt);
2298 * Do the hard work of removing an element from the buddy allocator.
2299 * Call me with the zone->lock already held.
2301 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2307 page = __rmqueue_smallest(zone, order, migratetype);
2308 if (unlikely(!page)) {
2309 if (migratetype == MIGRATE_MOVABLE)
2310 page = __rmqueue_cma_fallback(zone, order);
2312 if (!page && __rmqueue_fallback(zone, order, migratetype))
2316 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2321 * Obtain a specified number of elements from the buddy allocator, all under
2322 * a single hold of the lock, for efficiency. Add them to the supplied list.
2323 * Returns the number of new pages which were placed at *list.
2325 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2326 unsigned long count, struct list_head *list,
2327 int migratetype, bool cold)
2331 spin_lock(&zone->lock);
2332 for (i = 0; i < count; ++i) {
2333 struct page *page = __rmqueue(zone, order, migratetype);
2334 if (unlikely(page == NULL))
2337 if (unlikely(check_pcp_refill(page)))
2341 * Split buddy pages returned by expand() are received here
2342 * in physical page order. The page is added to the callers and
2343 * list and the list head then moves forward. From the callers
2344 * perspective, the linked list is ordered by page number in
2345 * some conditions. This is useful for IO devices that can
2346 * merge IO requests if the physical pages are ordered
2350 list_add(&page->lru, list);
2352 list_add_tail(&page->lru, list);
2355 if (is_migrate_cma(get_pcppage_migratetype(page)))
2356 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2361 * i pages were removed from the buddy list even if some leak due
2362 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2363 * on i. Do not confuse with 'alloced' which is the number of
2364 * pages added to the pcp list.
2366 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2367 spin_unlock(&zone->lock);
2373 * Called from the vmstat counter updater to drain pagesets of this
2374 * currently executing processor on remote nodes after they have
2377 * Note that this function must be called with the thread pinned to
2378 * a single processor.
2380 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2382 unsigned long flags;
2383 int to_drain, batch;
2385 local_irq_save(flags);
2386 batch = READ_ONCE(pcp->batch);
2387 to_drain = min(pcp->count, batch);
2389 free_pcppages_bulk(zone, to_drain, pcp);
2390 pcp->count -= to_drain;
2392 local_irq_restore(flags);
2397 * Drain pcplists of the indicated processor and zone.
2399 * The processor must either be the current processor and the
2400 * thread pinned to the current processor or a processor that
2403 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2405 unsigned long flags;
2406 struct per_cpu_pageset *pset;
2407 struct per_cpu_pages *pcp;
2409 local_irq_save(flags);
2410 pset = per_cpu_ptr(zone->pageset, cpu);
2414 free_pcppages_bulk(zone, pcp->count, pcp);
2417 local_irq_restore(flags);
2421 * Drain pcplists of all zones on the indicated processor.
2423 * The processor must either be the current processor and the
2424 * thread pinned to the current processor or a processor that
2427 static void drain_pages(unsigned int cpu)
2431 for_each_populated_zone(zone) {
2432 drain_pages_zone(cpu, zone);
2437 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2439 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2440 * the single zone's pages.
2442 void drain_local_pages(struct zone *zone)
2444 int cpu = smp_processor_id();
2447 drain_pages_zone(cpu, zone);
2452 static void drain_local_pages_wq(struct work_struct *work)
2455 * drain_all_pages doesn't use proper cpu hotplug protection so
2456 * we can race with cpu offline when the WQ can move this from
2457 * a cpu pinned worker to an unbound one. We can operate on a different
2458 * cpu which is allright but we also have to make sure to not move to
2462 drain_local_pages(NULL);
2467 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2469 * When zone parameter is non-NULL, spill just the single zone's pages.
2471 * Note that this can be extremely slow as the draining happens in a workqueue.
2473 void drain_all_pages(struct zone *zone)
2478 * Allocate in the BSS so we wont require allocation in
2479 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2481 static cpumask_t cpus_with_pcps;
2484 * Make sure nobody triggers this path before mm_percpu_wq is fully
2487 if (WARN_ON_ONCE(!mm_percpu_wq))
2491 * Do not drain if one is already in progress unless it's specific to
2492 * a zone. Such callers are primarily CMA and memory hotplug and need
2493 * the drain to be complete when the call returns.
2495 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2498 mutex_lock(&pcpu_drain_mutex);
2502 * We don't care about racing with CPU hotplug event
2503 * as offline notification will cause the notified
2504 * cpu to drain that CPU pcps and on_each_cpu_mask
2505 * disables preemption as part of its processing
2507 for_each_online_cpu(cpu) {
2508 struct per_cpu_pageset *pcp;
2510 bool has_pcps = false;
2513 pcp = per_cpu_ptr(zone->pageset, cpu);
2517 for_each_populated_zone(z) {
2518 pcp = per_cpu_ptr(z->pageset, cpu);
2519 if (pcp->pcp.count) {
2527 cpumask_set_cpu(cpu, &cpus_with_pcps);
2529 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2532 for_each_cpu(cpu, &cpus_with_pcps) {
2533 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2534 INIT_WORK(work, drain_local_pages_wq);
2535 queue_work_on(cpu, mm_percpu_wq, work);
2537 for_each_cpu(cpu, &cpus_with_pcps)
2538 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2540 mutex_unlock(&pcpu_drain_mutex);
2543 #ifdef CONFIG_HIBERNATION
2546 * Touch the watchdog for every WD_PAGE_COUNT pages.
2548 #define WD_PAGE_COUNT (128*1024)
2550 void mark_free_pages(struct zone *zone)
2552 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2553 unsigned long flags;
2554 unsigned int order, t;
2557 if (zone_is_empty(zone))
2560 spin_lock_irqsave(&zone->lock, flags);
2562 max_zone_pfn = zone_end_pfn(zone);
2563 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2564 if (pfn_valid(pfn)) {
2565 page = pfn_to_page(pfn);
2567 if (!--page_count) {
2568 touch_nmi_watchdog();
2569 page_count = WD_PAGE_COUNT;
2572 if (page_zone(page) != zone)
2575 if (!swsusp_page_is_forbidden(page))
2576 swsusp_unset_page_free(page);
2579 for_each_migratetype_order(order, t) {
2580 list_for_each_entry(page,
2581 &zone->free_area[order].free_list[t], lru) {
2584 pfn = page_to_pfn(page);
2585 for (i = 0; i < (1UL << order); i++) {
2586 if (!--page_count) {
2587 touch_nmi_watchdog();
2588 page_count = WD_PAGE_COUNT;
2590 swsusp_set_page_free(pfn_to_page(pfn + i));
2594 spin_unlock_irqrestore(&zone->lock, flags);
2596 #endif /* CONFIG_PM */
2599 * Free a 0-order page
2600 * cold == true ? free a cold page : free a hot page
2602 void free_hot_cold_page(struct page *page, bool cold)
2604 struct zone *zone = page_zone(page);
2605 struct per_cpu_pages *pcp;
2606 unsigned long flags;
2607 unsigned long pfn = page_to_pfn(page);
2610 if (!free_pcp_prepare(page))
2613 migratetype = get_pfnblock_migratetype(page, pfn);
2614 set_pcppage_migratetype(page, migratetype);
2615 local_irq_save(flags);
2616 __count_vm_event(PGFREE);
2619 * We only track unmovable, reclaimable and movable on pcp lists.
2620 * Free ISOLATE pages back to the allocator because they are being
2621 * offlined but treat HIGHATOMIC as movable pages so we can get those
2622 * areas back if necessary. Otherwise, we may have to free
2623 * excessively into the page allocator
2625 if (migratetype >= MIGRATE_PCPTYPES) {
2626 if (unlikely(is_migrate_isolate(migratetype))) {
2627 free_one_page(zone, page, pfn, 0, migratetype);
2630 migratetype = MIGRATE_MOVABLE;
2633 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2635 list_add(&page->lru, &pcp->lists[migratetype]);
2637 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2639 if (pcp->count >= pcp->high) {
2640 unsigned long batch = READ_ONCE(pcp->batch);
2641 free_pcppages_bulk(zone, batch, pcp);
2642 pcp->count -= batch;
2646 local_irq_restore(flags);
2650 * Free a list of 0-order pages
2652 void free_hot_cold_page_list(struct list_head *list, bool cold)
2654 struct page *page, *next;
2656 list_for_each_entry_safe(page, next, list, lru) {
2657 trace_mm_page_free_batched(page, cold);
2658 free_hot_cold_page(page, cold);
2663 * split_page takes a non-compound higher-order page, and splits it into
2664 * n (1<<order) sub-pages: page[0..n]
2665 * Each sub-page must be freed individually.
2667 * Note: this is probably too low level an operation for use in drivers.
2668 * Please consult with lkml before using this in your driver.
2670 void split_page(struct page *page, unsigned int order)
2674 VM_BUG_ON_PAGE(PageCompound(page), page);
2675 VM_BUG_ON_PAGE(!page_count(page), page);
2677 #ifdef CONFIG_KMEMCHECK
2679 * Split shadow pages too, because free(page[0]) would
2680 * otherwise free the whole shadow.
2682 if (kmemcheck_page_is_tracked(page))
2683 split_page(virt_to_page(page[0].shadow), order);
2686 for (i = 1; i < (1 << order); i++)
2687 set_page_refcounted(page + i);
2688 split_page_owner(page, order);
2690 EXPORT_SYMBOL_GPL(split_page);
2692 int __isolate_free_page(struct page *page, unsigned int order)
2694 unsigned long watermark;
2698 BUG_ON(!PageBuddy(page));
2700 zone = page_zone(page);
2701 mt = get_pageblock_migratetype(page);
2703 if (!is_migrate_isolate(mt)) {
2705 * Obey watermarks as if the page was being allocated. We can
2706 * emulate a high-order watermark check with a raised order-0
2707 * watermark, because we already know our high-order page
2710 watermark = min_wmark_pages(zone) + (1UL << order);
2711 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2714 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2717 /* Remove page from free list */
2718 list_del(&page->lru);
2719 zone->free_area[order].nr_free--;
2720 rmv_page_order(page);
2723 * Set the pageblock if the isolated page is at least half of a
2726 if (order >= pageblock_order - 1) {
2727 struct page *endpage = page + (1 << order) - 1;
2728 for (; page < endpage; page += pageblock_nr_pages) {
2729 int mt = get_pageblock_migratetype(page);
2730 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2731 && !is_migrate_highatomic(mt))
2732 set_pageblock_migratetype(page,
2738 return 1UL << order;
2742 * Update NUMA hit/miss statistics
2744 * Must be called with interrupts disabled.
2746 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2749 enum numa_stat_item local_stat = NUMA_LOCAL;
2751 if (z->node != numa_node_id())
2752 local_stat = NUMA_OTHER;
2754 if (z->node == preferred_zone->node)
2755 __inc_numa_state(z, NUMA_HIT);
2757 __inc_numa_state(z, NUMA_MISS);
2758 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2760 __inc_numa_state(z, local_stat);
2764 /* Remove page from the per-cpu list, caller must protect the list */
2765 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2766 bool cold, struct per_cpu_pages *pcp,
2767 struct list_head *list)
2772 if (list_empty(list)) {
2773 pcp->count += rmqueue_bulk(zone, 0,
2776 if (unlikely(list_empty(list)))
2781 page = list_last_entry(list, struct page, lru);
2783 page = list_first_entry(list, struct page, lru);
2785 list_del(&page->lru);
2787 } while (check_new_pcp(page));
2792 /* Lock and remove page from the per-cpu list */
2793 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2794 struct zone *zone, unsigned int order,
2795 gfp_t gfp_flags, int migratetype)
2797 struct per_cpu_pages *pcp;
2798 struct list_head *list;
2799 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2801 unsigned long flags;
2803 local_irq_save(flags);
2804 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2805 list = &pcp->lists[migratetype];
2806 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2808 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2809 zone_statistics(preferred_zone, zone);
2811 local_irq_restore(flags);
2816 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2819 struct page *rmqueue(struct zone *preferred_zone,
2820 struct zone *zone, unsigned int order,
2821 gfp_t gfp_flags, unsigned int alloc_flags,
2824 unsigned long flags;
2827 if (likely(order == 0)) {
2828 page = rmqueue_pcplist(preferred_zone, zone, order,
2829 gfp_flags, migratetype);
2834 * We most definitely don't want callers attempting to
2835 * allocate greater than order-1 page units with __GFP_NOFAIL.
2837 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2838 spin_lock_irqsave(&zone->lock, flags);
2842 if (alloc_flags & ALLOC_HARDER) {
2843 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2845 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2848 page = __rmqueue(zone, order, migratetype);
2849 } while (page && check_new_pages(page, order));
2850 spin_unlock(&zone->lock);
2853 __mod_zone_freepage_state(zone, -(1 << order),
2854 get_pcppage_migratetype(page));
2856 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2857 zone_statistics(preferred_zone, zone);
2858 local_irq_restore(flags);
2861 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2865 local_irq_restore(flags);
2869 #ifdef CONFIG_FAIL_PAGE_ALLOC
2872 struct fault_attr attr;
2874 bool ignore_gfp_highmem;
2875 bool ignore_gfp_reclaim;
2877 } fail_page_alloc = {
2878 .attr = FAULT_ATTR_INITIALIZER,
2879 .ignore_gfp_reclaim = true,
2880 .ignore_gfp_highmem = true,
2884 static int __init setup_fail_page_alloc(char *str)
2886 return setup_fault_attr(&fail_page_alloc.attr, str);
2888 __setup("fail_page_alloc=", setup_fail_page_alloc);
2890 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2892 if (order < fail_page_alloc.min_order)
2894 if (gfp_mask & __GFP_NOFAIL)
2896 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2898 if (fail_page_alloc.ignore_gfp_reclaim &&
2899 (gfp_mask & __GFP_DIRECT_RECLAIM))
2902 return should_fail(&fail_page_alloc.attr, 1 << order);
2905 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2907 static int __init fail_page_alloc_debugfs(void)
2909 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2912 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2913 &fail_page_alloc.attr);
2915 return PTR_ERR(dir);
2917 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2918 &fail_page_alloc.ignore_gfp_reclaim))
2920 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2921 &fail_page_alloc.ignore_gfp_highmem))
2923 if (!debugfs_create_u32("min-order", mode, dir,
2924 &fail_page_alloc.min_order))
2929 debugfs_remove_recursive(dir);
2934 late_initcall(fail_page_alloc_debugfs);
2936 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2938 #else /* CONFIG_FAIL_PAGE_ALLOC */
2940 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2945 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2948 * Return true if free base pages are above 'mark'. For high-order checks it
2949 * will return true of the order-0 watermark is reached and there is at least
2950 * one free page of a suitable size. Checking now avoids taking the zone lock
2951 * to check in the allocation paths if no pages are free.
2953 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2954 int classzone_idx, unsigned int alloc_flags,
2959 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
2961 /* free_pages may go negative - that's OK */
2962 free_pages -= (1 << order) - 1;
2964 if (alloc_flags & ALLOC_HIGH)
2968 * If the caller does not have rights to ALLOC_HARDER then subtract
2969 * the high-atomic reserves. This will over-estimate the size of the
2970 * atomic reserve but it avoids a search.
2972 if (likely(!alloc_harder)) {
2973 free_pages -= z->nr_reserved_highatomic;
2976 * OOM victims can try even harder than normal ALLOC_HARDER
2977 * users on the grounds that it's definitely going to be in
2978 * the exit path shortly and free memory. Any allocation it
2979 * makes during the free path will be small and short-lived.
2981 if (alloc_flags & ALLOC_OOM)
2989 /* If allocation can't use CMA areas don't use free CMA pages */
2990 if (!(alloc_flags & ALLOC_CMA))
2991 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2995 * Check watermarks for an order-0 allocation request. If these
2996 * are not met, then a high-order request also cannot go ahead
2997 * even if a suitable page happened to be free.
2999 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3002 /* If this is an order-0 request then the watermark is fine */
3006 /* For a high-order request, check at least one suitable page is free */
3007 for (o = order; o < MAX_ORDER; o++) {
3008 struct free_area *area = &z->free_area[o];
3017 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3018 if (!list_empty(&area->free_list[mt]))
3023 if ((alloc_flags & ALLOC_CMA) &&
3024 !list_empty(&area->free_list[MIGRATE_CMA])) {
3032 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3033 int classzone_idx, unsigned int alloc_flags)
3035 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3036 zone_page_state(z, NR_FREE_PAGES));
3039 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3040 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3042 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3046 /* If allocation can't use CMA areas don't use free CMA pages */
3047 if (!(alloc_flags & ALLOC_CMA))
3048 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3052 * Fast check for order-0 only. If this fails then the reserves
3053 * need to be calculated. There is a corner case where the check
3054 * passes but only the high-order atomic reserve are free. If
3055 * the caller is !atomic then it'll uselessly search the free
3056 * list. That corner case is then slower but it is harmless.
3058 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3061 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3065 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3066 unsigned long mark, int classzone_idx)
3068 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3070 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3071 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3073 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3078 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3080 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3083 #else /* CONFIG_NUMA */
3084 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3088 #endif /* CONFIG_NUMA */
3091 * get_page_from_freelist goes through the zonelist trying to allocate
3094 static struct page *
3095 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3096 const struct alloc_context *ac)
3098 struct zoneref *z = ac->preferred_zoneref;
3100 struct pglist_data *last_pgdat_dirty_limit = NULL;
3103 * Scan zonelist, looking for a zone with enough free.
3104 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3106 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3111 if (cpusets_enabled() &&
3112 (alloc_flags & ALLOC_CPUSET) &&
3113 !__cpuset_zone_allowed(zone, gfp_mask))
3116 * When allocating a page cache page for writing, we
3117 * want to get it from a node that is within its dirty
3118 * limit, such that no single node holds more than its
3119 * proportional share of globally allowed dirty pages.
3120 * The dirty limits take into account the node's
3121 * lowmem reserves and high watermark so that kswapd
3122 * should be able to balance it without having to
3123 * write pages from its LRU list.
3125 * XXX: For now, allow allocations to potentially
3126 * exceed the per-node dirty limit in the slowpath
3127 * (spread_dirty_pages unset) before going into reclaim,
3128 * which is important when on a NUMA setup the allowed
3129 * nodes are together not big enough to reach the
3130 * global limit. The proper fix for these situations
3131 * will require awareness of nodes in the
3132 * dirty-throttling and the flusher threads.
3134 if (ac->spread_dirty_pages) {
3135 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3138 if (!node_dirty_ok(zone->zone_pgdat)) {
3139 last_pgdat_dirty_limit = zone->zone_pgdat;
3144 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3145 if (!zone_watermark_fast(zone, order, mark,
3146 ac_classzone_idx(ac), alloc_flags)) {
3149 /* Checked here to keep the fast path fast */
3150 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3151 if (alloc_flags & ALLOC_NO_WATERMARKS)
3154 if (node_reclaim_mode == 0 ||
3155 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3158 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3160 case NODE_RECLAIM_NOSCAN:
3163 case NODE_RECLAIM_FULL:
3164 /* scanned but unreclaimable */
3167 /* did we reclaim enough */
3168 if (zone_watermark_ok(zone, order, mark,
3169 ac_classzone_idx(ac), alloc_flags))
3177 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3178 gfp_mask, alloc_flags, ac->migratetype);
3180 prep_new_page(page, order, gfp_mask, alloc_flags);
3183 * If this is a high-order atomic allocation then check
3184 * if the pageblock should be reserved for the future
3186 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3187 reserve_highatomic_pageblock(page, zone, order);
3197 * Large machines with many possible nodes should not always dump per-node
3198 * meminfo in irq context.
3200 static inline bool should_suppress_show_mem(void)
3205 ret = in_interrupt();
3210 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3212 unsigned int filter = SHOW_MEM_FILTER_NODES;
3213 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3215 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3219 * This documents exceptions given to allocations in certain
3220 * contexts that are allowed to allocate outside current's set
3223 if (!(gfp_mask & __GFP_NOMEMALLOC))
3224 if (tsk_is_oom_victim(current) ||
3225 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3226 filter &= ~SHOW_MEM_FILTER_NODES;
3227 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3228 filter &= ~SHOW_MEM_FILTER_NODES;
3230 show_mem(filter, nodemask);
3233 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3235 struct va_format vaf;
3237 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3238 DEFAULT_RATELIMIT_BURST);
3240 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3243 pr_warn("%s: ", current->comm);
3245 va_start(args, fmt);
3248 pr_cont("%pV", &vaf);
3251 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3253 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3255 pr_cont("(null)\n");
3257 cpuset_print_current_mems_allowed();
3260 warn_alloc_show_mem(gfp_mask, nodemask);
3263 static inline struct page *
3264 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3265 unsigned int alloc_flags,
3266 const struct alloc_context *ac)
3270 page = get_page_from_freelist(gfp_mask, order,
3271 alloc_flags|ALLOC_CPUSET, ac);
3273 * fallback to ignore cpuset restriction if our nodes
3277 page = get_page_from_freelist(gfp_mask, order,
3283 static inline struct page *
3284 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3285 const struct alloc_context *ac, unsigned long *did_some_progress)
3287 struct oom_control oc = {
3288 .zonelist = ac->zonelist,
3289 .nodemask = ac->nodemask,
3291 .gfp_mask = gfp_mask,
3296 *did_some_progress = 0;
3299 * Acquire the oom lock. If that fails, somebody else is
3300 * making progress for us.
3302 if (!mutex_trylock(&oom_lock)) {
3303 *did_some_progress = 1;
3304 schedule_timeout_uninterruptible(1);
3309 * Go through the zonelist yet one more time, keep very high watermark
3310 * here, this is only to catch a parallel oom killing, we must fail if
3311 * we're still under heavy pressure. But make sure that this reclaim
3312 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3313 * allocation which will never fail due to oom_lock already held.
3315 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3316 ~__GFP_DIRECT_RECLAIM, order,
3317 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3321 /* Coredumps can quickly deplete all memory reserves */
3322 if (current->flags & PF_DUMPCORE)
3324 /* The OOM killer will not help higher order allocs */
3325 if (order > PAGE_ALLOC_COSTLY_ORDER)
3328 * We have already exhausted all our reclaim opportunities without any
3329 * success so it is time to admit defeat. We will skip the OOM killer
3330 * because it is very likely that the caller has a more reasonable
3331 * fallback than shooting a random task.
3333 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3335 /* The OOM killer does not needlessly kill tasks for lowmem */
3336 if (ac->high_zoneidx < ZONE_NORMAL)
3338 if (pm_suspended_storage())
3341 * XXX: GFP_NOFS allocations should rather fail than rely on
3342 * other request to make a forward progress.
3343 * We are in an unfortunate situation where out_of_memory cannot
3344 * do much for this context but let's try it to at least get
3345 * access to memory reserved if the current task is killed (see
3346 * out_of_memory). Once filesystems are ready to handle allocation
3347 * failures more gracefully we should just bail out here.
3350 /* The OOM killer may not free memory on a specific node */
3351 if (gfp_mask & __GFP_THISNODE)
3354 /* Exhausted what can be done so it's blamo time */
3355 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3356 *did_some_progress = 1;
3359 * Help non-failing allocations by giving them access to memory
3362 if (gfp_mask & __GFP_NOFAIL)
3363 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3364 ALLOC_NO_WATERMARKS, ac);
3367 mutex_unlock(&oom_lock);
3372 * Maximum number of compaction retries wit a progress before OOM
3373 * killer is consider as the only way to move forward.
3375 #define MAX_COMPACT_RETRIES 16
3377 #ifdef CONFIG_COMPACTION
3378 /* Try memory compaction for high-order allocations before reclaim */
3379 static struct page *
3380 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3381 unsigned int alloc_flags, const struct alloc_context *ac,
3382 enum compact_priority prio, enum compact_result *compact_result)
3385 unsigned int noreclaim_flag;
3390 noreclaim_flag = memalloc_noreclaim_save();
3391 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3393 memalloc_noreclaim_restore(noreclaim_flag);
3395 if (*compact_result <= COMPACT_INACTIVE)
3399 * At least in one zone compaction wasn't deferred or skipped, so let's
3400 * count a compaction stall
3402 count_vm_event(COMPACTSTALL);
3404 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3407 struct zone *zone = page_zone(page);
3409 zone->compact_blockskip_flush = false;
3410 compaction_defer_reset(zone, order, true);
3411 count_vm_event(COMPACTSUCCESS);
3416 * It's bad if compaction run occurs and fails. The most likely reason
3417 * is that pages exist, but not enough to satisfy watermarks.
3419 count_vm_event(COMPACTFAIL);
3427 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3428 enum compact_result compact_result,
3429 enum compact_priority *compact_priority,
3430 int *compaction_retries)
3432 int max_retries = MAX_COMPACT_RETRIES;
3435 int retries = *compaction_retries;
3436 enum compact_priority priority = *compact_priority;
3441 if (compaction_made_progress(compact_result))
3442 (*compaction_retries)++;
3445 * compaction considers all the zone as desperately out of memory
3446 * so it doesn't really make much sense to retry except when the
3447 * failure could be caused by insufficient priority
3449 if (compaction_failed(compact_result))
3450 goto check_priority;
3453 * make sure the compaction wasn't deferred or didn't bail out early
3454 * due to locks contention before we declare that we should give up.
3455 * But do not retry if the given zonelist is not suitable for
3458 if (compaction_withdrawn(compact_result)) {
3459 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3464 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3465 * costly ones because they are de facto nofail and invoke OOM
3466 * killer to move on while costly can fail and users are ready
3467 * to cope with that. 1/4 retries is rather arbitrary but we
3468 * would need much more detailed feedback from compaction to
3469 * make a better decision.
3471 if (order > PAGE_ALLOC_COSTLY_ORDER)
3473 if (*compaction_retries <= max_retries) {
3479 * Make sure there are attempts at the highest priority if we exhausted
3480 * all retries or failed at the lower priorities.
3483 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3484 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3486 if (*compact_priority > min_priority) {
3487 (*compact_priority)--;
3488 *compaction_retries = 0;
3492 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3496 static inline struct page *
3497 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3498 unsigned int alloc_flags, const struct alloc_context *ac,
3499 enum compact_priority prio, enum compact_result *compact_result)
3501 *compact_result = COMPACT_SKIPPED;
3506 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3507 enum compact_result compact_result,
3508 enum compact_priority *compact_priority,
3509 int *compaction_retries)
3514 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3518 * There are setups with compaction disabled which would prefer to loop
3519 * inside the allocator rather than hit the oom killer prematurely.
3520 * Let's give them a good hope and keep retrying while the order-0
3521 * watermarks are OK.
3523 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3525 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3526 ac_classzone_idx(ac), alloc_flags))
3531 #endif /* CONFIG_COMPACTION */
3533 #ifdef CONFIG_LOCKDEP
3534 struct lockdep_map __fs_reclaim_map =
3535 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3537 static bool __need_fs_reclaim(gfp_t gfp_mask)
3539 gfp_mask = current_gfp_context(gfp_mask);
3541 /* no reclaim without waiting on it */
3542 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3545 /* this guy won't enter reclaim */
3546 if ((current->flags & PF_MEMALLOC) && !(gfp_mask & __GFP_NOMEMALLOC))
3549 /* We're only interested __GFP_FS allocations for now */
3550 if (!(gfp_mask & __GFP_FS))
3553 if (gfp_mask & __GFP_NOLOCKDEP)
3559 void fs_reclaim_acquire(gfp_t gfp_mask)
3561 if (__need_fs_reclaim(gfp_mask))
3562 lock_map_acquire(&__fs_reclaim_map);
3564 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3566 void fs_reclaim_release(gfp_t gfp_mask)
3568 if (__need_fs_reclaim(gfp_mask))
3569 lock_map_release(&__fs_reclaim_map);
3571 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3574 /* Perform direct synchronous page reclaim */
3576 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3577 const struct alloc_context *ac)
3579 struct reclaim_state reclaim_state;
3581 unsigned int noreclaim_flag;
3585 /* We now go into synchronous reclaim */
3586 cpuset_memory_pressure_bump();
3587 noreclaim_flag = memalloc_noreclaim_save();
3588 fs_reclaim_acquire(gfp_mask);
3589 reclaim_state.reclaimed_slab = 0;
3590 current->reclaim_state = &reclaim_state;
3592 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3595 current->reclaim_state = NULL;
3596 fs_reclaim_release(gfp_mask);
3597 memalloc_noreclaim_restore(noreclaim_flag);
3604 /* The really slow allocator path where we enter direct reclaim */
3605 static inline struct page *
3606 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3607 unsigned int alloc_flags, const struct alloc_context *ac,
3608 unsigned long *did_some_progress)
3610 struct page *page = NULL;
3611 bool drained = false;
3613 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3614 if (unlikely(!(*did_some_progress)))
3618 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3621 * If an allocation failed after direct reclaim, it could be because
3622 * pages are pinned on the per-cpu lists or in high alloc reserves.
3623 * Shrink them them and try again
3625 if (!page && !drained) {
3626 unreserve_highatomic_pageblock(ac, false);
3627 drain_all_pages(NULL);
3635 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3639 pg_data_t *last_pgdat = NULL;
3641 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3642 ac->high_zoneidx, ac->nodemask) {
3643 if (last_pgdat != zone->zone_pgdat)
3644 wakeup_kswapd(zone, order, ac->high_zoneidx);
3645 last_pgdat = zone->zone_pgdat;
3649 static inline unsigned int
3650 gfp_to_alloc_flags(gfp_t gfp_mask)
3652 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3654 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3655 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3658 * The caller may dip into page reserves a bit more if the caller
3659 * cannot run direct reclaim, or if the caller has realtime scheduling
3660 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3661 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3663 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3665 if (gfp_mask & __GFP_ATOMIC) {
3667 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3668 * if it can't schedule.
3670 if (!(gfp_mask & __GFP_NOMEMALLOC))
3671 alloc_flags |= ALLOC_HARDER;
3673 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3674 * comment for __cpuset_node_allowed().
3676 alloc_flags &= ~ALLOC_CPUSET;
3677 } else if (unlikely(rt_task(current)) && !in_interrupt())
3678 alloc_flags |= ALLOC_HARDER;
3681 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3682 alloc_flags |= ALLOC_CMA;
3687 static bool oom_reserves_allowed(struct task_struct *tsk)
3689 if (!tsk_is_oom_victim(tsk))
3693 * !MMU doesn't have oom reaper so give access to memory reserves
3694 * only to the thread with TIF_MEMDIE set
3696 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3703 * Distinguish requests which really need access to full memory
3704 * reserves from oom victims which can live with a portion of it
3706 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3708 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3710 if (gfp_mask & __GFP_MEMALLOC)
3711 return ALLOC_NO_WATERMARKS;
3712 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3713 return ALLOC_NO_WATERMARKS;
3714 if (!in_interrupt()) {
3715 if (current->flags & PF_MEMALLOC)
3716 return ALLOC_NO_WATERMARKS;
3717 else if (oom_reserves_allowed(current))
3724 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3726 return !!__gfp_pfmemalloc_flags(gfp_mask);
3730 * Checks whether it makes sense to retry the reclaim to make a forward progress
3731 * for the given allocation request.
3733 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3734 * without success, or when we couldn't even meet the watermark if we
3735 * reclaimed all remaining pages on the LRU lists.
3737 * Returns true if a retry is viable or false to enter the oom path.
3740 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3741 struct alloc_context *ac, int alloc_flags,
3742 bool did_some_progress, int *no_progress_loops)
3748 * Costly allocations might have made a progress but this doesn't mean
3749 * their order will become available due to high fragmentation so
3750 * always increment the no progress counter for them
3752 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3753 *no_progress_loops = 0;
3755 (*no_progress_loops)++;
3758 * Make sure we converge to OOM if we cannot make any progress
3759 * several times in the row.
3761 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3762 /* Before OOM, exhaust highatomic_reserve */
3763 return unreserve_highatomic_pageblock(ac, true);
3767 * Keep reclaiming pages while there is a chance this will lead
3768 * somewhere. If none of the target zones can satisfy our allocation
3769 * request even if all reclaimable pages are considered then we are
3770 * screwed and have to go OOM.
3772 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3774 unsigned long available;
3775 unsigned long reclaimable;
3776 unsigned long min_wmark = min_wmark_pages(zone);
3779 available = reclaimable = zone_reclaimable_pages(zone);
3780 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3783 * Would the allocation succeed if we reclaimed all
3784 * reclaimable pages?
3786 wmark = __zone_watermark_ok(zone, order, min_wmark,
3787 ac_classzone_idx(ac), alloc_flags, available);
3788 trace_reclaim_retry_zone(z, order, reclaimable,
3789 available, min_wmark, *no_progress_loops, wmark);
3792 * If we didn't make any progress and have a lot of
3793 * dirty + writeback pages then we should wait for
3794 * an IO to complete to slow down the reclaim and
3795 * prevent from pre mature OOM
3797 if (!did_some_progress) {
3798 unsigned long write_pending;
3800 write_pending = zone_page_state_snapshot(zone,
3801 NR_ZONE_WRITE_PENDING);
3803 if (2 * write_pending > reclaimable) {
3804 congestion_wait(BLK_RW_ASYNC, HZ/10);
3810 * Memory allocation/reclaim might be called from a WQ
3811 * context and the current implementation of the WQ
3812 * concurrency control doesn't recognize that
3813 * a particular WQ is congested if the worker thread is
3814 * looping without ever sleeping. Therefore we have to
3815 * do a short sleep here rather than calling
3818 if (current->flags & PF_WQ_WORKER)
3819 schedule_timeout_uninterruptible(1);
3831 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3834 * It's possible that cpuset's mems_allowed and the nodemask from
3835 * mempolicy don't intersect. This should be normally dealt with by
3836 * policy_nodemask(), but it's possible to race with cpuset update in
3837 * such a way the check therein was true, and then it became false
3838 * before we got our cpuset_mems_cookie here.
3839 * This assumes that for all allocations, ac->nodemask can come only
3840 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3841 * when it does not intersect with the cpuset restrictions) or the
3842 * caller can deal with a violated nodemask.
3844 if (cpusets_enabled() && ac->nodemask &&
3845 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3846 ac->nodemask = NULL;
3851 * When updating a task's mems_allowed or mempolicy nodemask, it is
3852 * possible to race with parallel threads in such a way that our
3853 * allocation can fail while the mask is being updated. If we are about
3854 * to fail, check if the cpuset changed during allocation and if so,
3857 if (read_mems_allowed_retry(cpuset_mems_cookie))
3863 static inline struct page *
3864 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3865 struct alloc_context *ac)
3867 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3868 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3869 struct page *page = NULL;
3870 unsigned int alloc_flags;
3871 unsigned long did_some_progress;
3872 enum compact_priority compact_priority;
3873 enum compact_result compact_result;
3874 int compaction_retries;
3875 int no_progress_loops;
3876 unsigned long alloc_start = jiffies;
3877 unsigned int stall_timeout = 10 * HZ;
3878 unsigned int cpuset_mems_cookie;
3882 * In the slowpath, we sanity check order to avoid ever trying to
3883 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3884 * be using allocators in order of preference for an area that is
3887 if (order >= MAX_ORDER) {
3888 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3893 * We also sanity check to catch abuse of atomic reserves being used by
3894 * callers that are not in atomic context.
3896 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3897 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3898 gfp_mask &= ~__GFP_ATOMIC;
3901 compaction_retries = 0;
3902 no_progress_loops = 0;
3903 compact_priority = DEF_COMPACT_PRIORITY;
3904 cpuset_mems_cookie = read_mems_allowed_begin();
3907 * The fast path uses conservative alloc_flags to succeed only until
3908 * kswapd needs to be woken up, and to avoid the cost of setting up
3909 * alloc_flags precisely. So we do that now.
3911 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3914 * We need to recalculate the starting point for the zonelist iterator
3915 * because we might have used different nodemask in the fast path, or
3916 * there was a cpuset modification and we are retrying - otherwise we
3917 * could end up iterating over non-eligible zones endlessly.
3919 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3920 ac->high_zoneidx, ac->nodemask);
3921 if (!ac->preferred_zoneref->zone)
3924 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3925 wake_all_kswapds(order, ac);
3928 * The adjusted alloc_flags might result in immediate success, so try
3931 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3936 * For costly allocations, try direct compaction first, as it's likely
3937 * that we have enough base pages and don't need to reclaim. For non-
3938 * movable high-order allocations, do that as well, as compaction will
3939 * try prevent permanent fragmentation by migrating from blocks of the
3941 * Don't try this for allocations that are allowed to ignore
3942 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3944 if (can_direct_reclaim &&
3946 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3947 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3948 page = __alloc_pages_direct_compact(gfp_mask, order,
3950 INIT_COMPACT_PRIORITY,
3956 * Checks for costly allocations with __GFP_NORETRY, which
3957 * includes THP page fault allocations
3959 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
3961 * If compaction is deferred for high-order allocations,
3962 * it is because sync compaction recently failed. If
3963 * this is the case and the caller requested a THP
3964 * allocation, we do not want to heavily disrupt the
3965 * system, so we fail the allocation instead of entering
3968 if (compact_result == COMPACT_DEFERRED)
3972 * Looks like reclaim/compaction is worth trying, but
3973 * sync compaction could be very expensive, so keep
3974 * using async compaction.
3976 compact_priority = INIT_COMPACT_PRIORITY;
3981 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3982 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3983 wake_all_kswapds(order, ac);
3985 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
3987 alloc_flags = reserve_flags;
3990 * Reset the zonelist iterators if memory policies can be ignored.
3991 * These allocations are high priority and system rather than user
3994 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
3995 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3996 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3997 ac->high_zoneidx, ac->nodemask);
4000 /* Attempt with potentially adjusted zonelist and alloc_flags */
4001 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4005 /* Caller is not willing to reclaim, we can't balance anything */
4006 if (!can_direct_reclaim)
4009 /* Make sure we know about allocations which stall for too long */
4010 if (time_after(jiffies, alloc_start + stall_timeout)) {
4011 warn_alloc(gfp_mask & ~__GFP_NOWARN, ac->nodemask,
4012 "page allocation stalls for %ums, order:%u",
4013 jiffies_to_msecs(jiffies-alloc_start), order);
4014 stall_timeout += 10 * HZ;
4017 /* Avoid recursion of direct reclaim */
4018 if (current->flags & PF_MEMALLOC)
4021 /* Try direct reclaim and then allocating */
4022 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4023 &did_some_progress);
4027 /* Try direct compaction and then allocating */
4028 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4029 compact_priority, &compact_result);
4033 /* Do not loop if specifically requested */
4034 if (gfp_mask & __GFP_NORETRY)
4038 * Do not retry costly high order allocations unless they are
4039 * __GFP_RETRY_MAYFAIL
4041 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4044 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4045 did_some_progress > 0, &no_progress_loops))
4049 * It doesn't make any sense to retry for the compaction if the order-0
4050 * reclaim is not able to make any progress because the current
4051 * implementation of the compaction depends on the sufficient amount
4052 * of free memory (see __compaction_suitable)
4054 if (did_some_progress > 0 &&
4055 should_compact_retry(ac, order, alloc_flags,
4056 compact_result, &compact_priority,
4057 &compaction_retries))
4061 /* Deal with possible cpuset update races before we start OOM killing */
4062 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4065 /* Reclaim has failed us, start killing things */
4066 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4070 /* Avoid allocations with no watermarks from looping endlessly */
4071 if (tsk_is_oom_victim(current) &&
4072 (alloc_flags == ALLOC_OOM ||
4073 (gfp_mask & __GFP_NOMEMALLOC)))
4076 /* Retry as long as the OOM killer is making progress */
4077 if (did_some_progress) {
4078 no_progress_loops = 0;
4083 /* Deal with possible cpuset update races before we fail */
4084 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4088 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4091 if (gfp_mask & __GFP_NOFAIL) {
4093 * All existing users of the __GFP_NOFAIL are blockable, so warn
4094 * of any new users that actually require GFP_NOWAIT
4096 if (WARN_ON_ONCE(!can_direct_reclaim))
4100 * PF_MEMALLOC request from this context is rather bizarre
4101 * because we cannot reclaim anything and only can loop waiting
4102 * for somebody to do a work for us
4104 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4107 * non failing costly orders are a hard requirement which we
4108 * are not prepared for much so let's warn about these users
4109 * so that we can identify them and convert them to something
4112 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4115 * Help non-failing allocations by giving them access to memory
4116 * reserves but do not use ALLOC_NO_WATERMARKS because this
4117 * could deplete whole memory reserves which would just make
4118 * the situation worse
4120 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4128 warn_alloc(gfp_mask, ac->nodemask,
4129 "page allocation failure: order:%u", order);
4134 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4135 int preferred_nid, nodemask_t *nodemask,
4136 struct alloc_context *ac, gfp_t *alloc_mask,
4137 unsigned int *alloc_flags)
4139 ac->high_zoneidx = gfp_zone(gfp_mask);
4140 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4141 ac->nodemask = nodemask;
4142 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4144 if (cpusets_enabled()) {
4145 *alloc_mask |= __GFP_HARDWALL;
4147 ac->nodemask = &cpuset_current_mems_allowed;
4149 *alloc_flags |= ALLOC_CPUSET;
4152 fs_reclaim_acquire(gfp_mask);
4153 fs_reclaim_release(gfp_mask);
4155 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4157 if (should_fail_alloc_page(gfp_mask, order))
4160 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4161 *alloc_flags |= ALLOC_CMA;
4166 /* Determine whether to spread dirty pages and what the first usable zone */
4167 static inline void finalise_ac(gfp_t gfp_mask,
4168 unsigned int order, struct alloc_context *ac)
4170 /* Dirty zone balancing only done in the fast path */
4171 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4174 * The preferred zone is used for statistics but crucially it is
4175 * also used as the starting point for the zonelist iterator. It
4176 * may get reset for allocations that ignore memory policies.
4178 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4179 ac->high_zoneidx, ac->nodemask);
4183 * This is the 'heart' of the zoned buddy allocator.
4186 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4187 nodemask_t *nodemask)
4190 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4191 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4192 struct alloc_context ac = { };
4194 gfp_mask &= gfp_allowed_mask;
4195 alloc_mask = gfp_mask;
4196 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4199 finalise_ac(gfp_mask, order, &ac);
4201 /* First allocation attempt */
4202 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4207 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4208 * resp. GFP_NOIO which has to be inherited for all allocation requests
4209 * from a particular context which has been marked by
4210 * memalloc_no{fs,io}_{save,restore}.
4212 alloc_mask = current_gfp_context(gfp_mask);
4213 ac.spread_dirty_pages = false;
4216 * Restore the original nodemask if it was potentially replaced with
4217 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4219 if (unlikely(ac.nodemask != nodemask))
4220 ac.nodemask = nodemask;
4222 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4225 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4226 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4227 __free_pages(page, order);
4231 if (kmemcheck_enabled && page)
4232 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
4234 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4238 EXPORT_SYMBOL(__alloc_pages_nodemask);
4241 * Common helper functions.
4243 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4248 * __get_free_pages() returns a 32-bit address, which cannot represent
4251 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4253 page = alloc_pages(gfp_mask, order);
4256 return (unsigned long) page_address(page);
4258 EXPORT_SYMBOL(__get_free_pages);
4260 unsigned long get_zeroed_page(gfp_t gfp_mask)
4262 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4264 EXPORT_SYMBOL(get_zeroed_page);
4266 void __free_pages(struct page *page, unsigned int order)
4268 if (put_page_testzero(page)) {
4270 free_hot_cold_page(page, false);
4272 __free_pages_ok(page, order);
4276 EXPORT_SYMBOL(__free_pages);
4278 void free_pages(unsigned long addr, unsigned int order)
4281 VM_BUG_ON(!virt_addr_valid((void *)addr));
4282 __free_pages(virt_to_page((void *)addr), order);
4286 EXPORT_SYMBOL(free_pages);
4290 * An arbitrary-length arbitrary-offset area of memory which resides
4291 * within a 0 or higher order page. Multiple fragments within that page
4292 * are individually refcounted, in the page's reference counter.
4294 * The page_frag functions below provide a simple allocation framework for
4295 * page fragments. This is used by the network stack and network device
4296 * drivers to provide a backing region of memory for use as either an
4297 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4299 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4302 struct page *page = NULL;
4303 gfp_t gfp = gfp_mask;
4305 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4306 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4308 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4309 PAGE_FRAG_CACHE_MAX_ORDER);
4310 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4312 if (unlikely(!page))
4313 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4315 nc->va = page ? page_address(page) : NULL;
4320 void __page_frag_cache_drain(struct page *page, unsigned int count)
4322 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4324 if (page_ref_sub_and_test(page, count)) {
4325 unsigned int order = compound_order(page);
4328 free_hot_cold_page(page, false);
4330 __free_pages_ok(page, order);
4333 EXPORT_SYMBOL(__page_frag_cache_drain);
4335 void *page_frag_alloc(struct page_frag_cache *nc,
4336 unsigned int fragsz, gfp_t gfp_mask)
4338 unsigned int size = PAGE_SIZE;
4342 if (unlikely(!nc->va)) {
4344 page = __page_frag_cache_refill(nc, gfp_mask);
4348 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4349 /* if size can vary use size else just use PAGE_SIZE */
4352 /* Even if we own the page, we do not use atomic_set().
4353 * This would break get_page_unless_zero() users.
4355 page_ref_add(page, size - 1);
4357 /* reset page count bias and offset to start of new frag */
4358 nc->pfmemalloc = page_is_pfmemalloc(page);
4359 nc->pagecnt_bias = size;
4363 offset = nc->offset - fragsz;
4364 if (unlikely(offset < 0)) {
4365 page = virt_to_page(nc->va);
4367 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4370 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4371 /* if size can vary use size else just use PAGE_SIZE */
4374 /* OK, page count is 0, we can safely set it */
4375 set_page_count(page, size);
4377 /* reset page count bias and offset to start of new frag */
4378 nc->pagecnt_bias = size;
4379 offset = size - fragsz;
4383 nc->offset = offset;
4385 return nc->va + offset;
4387 EXPORT_SYMBOL(page_frag_alloc);
4390 * Frees a page fragment allocated out of either a compound or order 0 page.
4392 void page_frag_free(void *addr)
4394 struct page *page = virt_to_head_page(addr);
4396 if (unlikely(put_page_testzero(page)))
4397 __free_pages_ok(page, compound_order(page));
4399 EXPORT_SYMBOL(page_frag_free);
4401 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4405 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4406 unsigned long used = addr + PAGE_ALIGN(size);
4408 split_page(virt_to_page((void *)addr), order);
4409 while (used < alloc_end) {
4414 return (void *)addr;
4418 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4419 * @size: the number of bytes to allocate
4420 * @gfp_mask: GFP flags for the allocation
4422 * This function is similar to alloc_pages(), except that it allocates the
4423 * minimum number of pages to satisfy the request. alloc_pages() can only
4424 * allocate memory in power-of-two pages.
4426 * This function is also limited by MAX_ORDER.
4428 * Memory allocated by this function must be released by free_pages_exact().
4430 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4432 unsigned int order = get_order(size);
4435 addr = __get_free_pages(gfp_mask, order);
4436 return make_alloc_exact(addr, order, size);
4438 EXPORT_SYMBOL(alloc_pages_exact);
4441 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4443 * @nid: the preferred node ID where memory should be allocated
4444 * @size: the number of bytes to allocate
4445 * @gfp_mask: GFP flags for the allocation
4447 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4450 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4452 unsigned int order = get_order(size);
4453 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4456 return make_alloc_exact((unsigned long)page_address(p), order, size);
4460 * free_pages_exact - release memory allocated via alloc_pages_exact()
4461 * @virt: the value returned by alloc_pages_exact.
4462 * @size: size of allocation, same value as passed to alloc_pages_exact().
4464 * Release the memory allocated by a previous call to alloc_pages_exact.
4466 void free_pages_exact(void *virt, size_t size)
4468 unsigned long addr = (unsigned long)virt;
4469 unsigned long end = addr + PAGE_ALIGN(size);
4471 while (addr < end) {
4476 EXPORT_SYMBOL(free_pages_exact);
4479 * nr_free_zone_pages - count number of pages beyond high watermark
4480 * @offset: The zone index of the highest zone
4482 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4483 * high watermark within all zones at or below a given zone index. For each
4484 * zone, the number of pages is calculated as:
4486 * nr_free_zone_pages = managed_pages - high_pages
4488 static unsigned long nr_free_zone_pages(int offset)
4493 /* Just pick one node, since fallback list is circular */
4494 unsigned long sum = 0;
4496 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4498 for_each_zone_zonelist(zone, z, zonelist, offset) {
4499 unsigned long size = zone->managed_pages;
4500 unsigned long high = high_wmark_pages(zone);
4509 * nr_free_buffer_pages - count number of pages beyond high watermark
4511 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4512 * watermark within ZONE_DMA and ZONE_NORMAL.
4514 unsigned long nr_free_buffer_pages(void)
4516 return nr_free_zone_pages(gfp_zone(GFP_USER));
4518 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4521 * nr_free_pagecache_pages - count number of pages beyond high watermark
4523 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4524 * high watermark within all zones.
4526 unsigned long nr_free_pagecache_pages(void)
4528 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4531 static inline void show_node(struct zone *zone)
4533 if (IS_ENABLED(CONFIG_NUMA))
4534 printk("Node %d ", zone_to_nid(zone));
4537 long si_mem_available(void)
4540 unsigned long pagecache;
4541 unsigned long wmark_low = 0;
4542 unsigned long pages[NR_LRU_LISTS];
4546 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4547 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4550 wmark_low += zone->watermark[WMARK_LOW];
4553 * Estimate the amount of memory available for userspace allocations,
4554 * without causing swapping.
4556 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4559 * Not all the page cache can be freed, otherwise the system will
4560 * start swapping. Assume at least half of the page cache, or the
4561 * low watermark worth of cache, needs to stay.
4563 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4564 pagecache -= min(pagecache / 2, wmark_low);
4565 available += pagecache;
4568 * Part of the reclaimable slab consists of items that are in use,
4569 * and cannot be freed. Cap this estimate at the low watermark.
4571 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4572 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4579 EXPORT_SYMBOL_GPL(si_mem_available);
4581 void si_meminfo(struct sysinfo *val)
4583 val->totalram = totalram_pages;
4584 val->sharedram = global_node_page_state(NR_SHMEM);
4585 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4586 val->bufferram = nr_blockdev_pages();
4587 val->totalhigh = totalhigh_pages;
4588 val->freehigh = nr_free_highpages();
4589 val->mem_unit = PAGE_SIZE;
4592 EXPORT_SYMBOL(si_meminfo);
4595 void si_meminfo_node(struct sysinfo *val, int nid)
4597 int zone_type; /* needs to be signed */
4598 unsigned long managed_pages = 0;
4599 unsigned long managed_highpages = 0;
4600 unsigned long free_highpages = 0;
4601 pg_data_t *pgdat = NODE_DATA(nid);
4603 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4604 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4605 val->totalram = managed_pages;
4606 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4607 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4608 #ifdef CONFIG_HIGHMEM
4609 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4610 struct zone *zone = &pgdat->node_zones[zone_type];
4612 if (is_highmem(zone)) {
4613 managed_highpages += zone->managed_pages;
4614 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4617 val->totalhigh = managed_highpages;
4618 val->freehigh = free_highpages;
4620 val->totalhigh = managed_highpages;
4621 val->freehigh = free_highpages;
4623 val->mem_unit = PAGE_SIZE;
4628 * Determine whether the node should be displayed or not, depending on whether
4629 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4631 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4633 if (!(flags & SHOW_MEM_FILTER_NODES))
4637 * no node mask - aka implicit memory numa policy. Do not bother with
4638 * the synchronization - read_mems_allowed_begin - because we do not
4639 * have to be precise here.
4642 nodemask = &cpuset_current_mems_allowed;
4644 return !node_isset(nid, *nodemask);
4647 #define K(x) ((x) << (PAGE_SHIFT-10))
4649 static void show_migration_types(unsigned char type)
4651 static const char types[MIGRATE_TYPES] = {
4652 [MIGRATE_UNMOVABLE] = 'U',
4653 [MIGRATE_MOVABLE] = 'M',
4654 [MIGRATE_RECLAIMABLE] = 'E',
4655 [MIGRATE_HIGHATOMIC] = 'H',
4657 [MIGRATE_CMA] = 'C',
4659 #ifdef CONFIG_MEMORY_ISOLATION
4660 [MIGRATE_ISOLATE] = 'I',
4663 char tmp[MIGRATE_TYPES + 1];
4667 for (i = 0; i < MIGRATE_TYPES; i++) {
4668 if (type & (1 << i))
4673 printk(KERN_CONT "(%s) ", tmp);
4677 * Show free area list (used inside shift_scroll-lock stuff)
4678 * We also calculate the percentage fragmentation. We do this by counting the
4679 * memory on each free list with the exception of the first item on the list.
4682 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4685 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4687 unsigned long free_pcp = 0;
4692 for_each_populated_zone(zone) {
4693 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4696 for_each_online_cpu(cpu)
4697 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4700 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4701 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4702 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4703 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4704 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4705 " free:%lu free_pcp:%lu free_cma:%lu\n",
4706 global_node_page_state(NR_ACTIVE_ANON),
4707 global_node_page_state(NR_INACTIVE_ANON),
4708 global_node_page_state(NR_ISOLATED_ANON),
4709 global_node_page_state(NR_ACTIVE_FILE),
4710 global_node_page_state(NR_INACTIVE_FILE),
4711 global_node_page_state(NR_ISOLATED_FILE),
4712 global_node_page_state(NR_UNEVICTABLE),
4713 global_node_page_state(NR_FILE_DIRTY),
4714 global_node_page_state(NR_WRITEBACK),
4715 global_node_page_state(NR_UNSTABLE_NFS),
4716 global_node_page_state(NR_SLAB_RECLAIMABLE),
4717 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4718 global_node_page_state(NR_FILE_MAPPED),
4719 global_node_page_state(NR_SHMEM),
4720 global_zone_page_state(NR_PAGETABLE),
4721 global_zone_page_state(NR_BOUNCE),
4722 global_zone_page_state(NR_FREE_PAGES),
4724 global_zone_page_state(NR_FREE_CMA_PAGES));
4726 for_each_online_pgdat(pgdat) {
4727 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4731 " active_anon:%lukB"
4732 " inactive_anon:%lukB"
4733 " active_file:%lukB"
4734 " inactive_file:%lukB"
4735 " unevictable:%lukB"
4736 " isolated(anon):%lukB"
4737 " isolated(file):%lukB"
4742 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4744 " shmem_pmdmapped: %lukB"
4747 " writeback_tmp:%lukB"
4749 " all_unreclaimable? %s"
4752 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4753 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4754 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4755 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4756 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4757 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4758 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4759 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4760 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4761 K(node_page_state(pgdat, NR_WRITEBACK)),
4762 K(node_page_state(pgdat, NR_SHMEM)),
4763 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4764 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4765 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4767 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4769 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4770 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4771 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4775 for_each_populated_zone(zone) {
4778 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4782 for_each_online_cpu(cpu)
4783 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4792 " active_anon:%lukB"
4793 " inactive_anon:%lukB"
4794 " active_file:%lukB"
4795 " inactive_file:%lukB"
4796 " unevictable:%lukB"
4797 " writepending:%lukB"
4801 " kernel_stack:%lukB"
4809 K(zone_page_state(zone, NR_FREE_PAGES)),
4810 K(min_wmark_pages(zone)),
4811 K(low_wmark_pages(zone)),
4812 K(high_wmark_pages(zone)),
4813 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4814 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4815 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4816 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4817 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4818 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4819 K(zone->present_pages),
4820 K(zone->managed_pages),
4821 K(zone_page_state(zone, NR_MLOCK)),
4822 zone_page_state(zone, NR_KERNEL_STACK_KB),
4823 K(zone_page_state(zone, NR_PAGETABLE)),
4824 K(zone_page_state(zone, NR_BOUNCE)),
4826 K(this_cpu_read(zone->pageset->pcp.count)),
4827 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4828 printk("lowmem_reserve[]:");
4829 for (i = 0; i < MAX_NR_ZONES; i++)
4830 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4831 printk(KERN_CONT "\n");
4834 for_each_populated_zone(zone) {
4836 unsigned long nr[MAX_ORDER], flags, total = 0;
4837 unsigned char types[MAX_ORDER];
4839 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4842 printk(KERN_CONT "%s: ", zone->name);
4844 spin_lock_irqsave(&zone->lock, flags);
4845 for (order = 0; order < MAX_ORDER; order++) {
4846 struct free_area *area = &zone->free_area[order];
4849 nr[order] = area->nr_free;
4850 total += nr[order] << order;
4853 for (type = 0; type < MIGRATE_TYPES; type++) {
4854 if (!list_empty(&area->free_list[type]))
4855 types[order] |= 1 << type;
4858 spin_unlock_irqrestore(&zone->lock, flags);
4859 for (order = 0; order < MAX_ORDER; order++) {
4860 printk(KERN_CONT "%lu*%lukB ",
4861 nr[order], K(1UL) << order);
4863 show_migration_types(types[order]);
4865 printk(KERN_CONT "= %lukB\n", K(total));
4868 hugetlb_show_meminfo();
4870 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4872 show_swap_cache_info();
4875 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4877 zoneref->zone = zone;
4878 zoneref->zone_idx = zone_idx(zone);
4882 * Builds allocation fallback zone lists.
4884 * Add all populated zones of a node to the zonelist.
4886 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4889 enum zone_type zone_type = MAX_NR_ZONES;
4894 zone = pgdat->node_zones + zone_type;
4895 if (managed_zone(zone)) {
4896 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4897 check_highest_zone(zone_type);
4899 } while (zone_type);
4906 static int __parse_numa_zonelist_order(char *s)
4909 * We used to support different zonlists modes but they turned
4910 * out to be just not useful. Let's keep the warning in place
4911 * if somebody still use the cmd line parameter so that we do
4912 * not fail it silently
4914 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4915 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4921 static __init int setup_numa_zonelist_order(char *s)
4926 return __parse_numa_zonelist_order(s);
4928 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4930 char numa_zonelist_order[] = "Node";
4933 * sysctl handler for numa_zonelist_order
4935 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4936 void __user *buffer, size_t *length,
4943 return proc_dostring(table, write, buffer, length, ppos);
4944 str = memdup_user_nul(buffer, 16);
4946 return PTR_ERR(str);
4948 ret = __parse_numa_zonelist_order(str);
4954 #define MAX_NODE_LOAD (nr_online_nodes)
4955 static int node_load[MAX_NUMNODES];
4958 * find_next_best_node - find the next node that should appear in a given node's fallback list
4959 * @node: node whose fallback list we're appending
4960 * @used_node_mask: nodemask_t of already used nodes
4962 * We use a number of factors to determine which is the next node that should
4963 * appear on a given node's fallback list. The node should not have appeared
4964 * already in @node's fallback list, and it should be the next closest node
4965 * according to the distance array (which contains arbitrary distance values
4966 * from each node to each node in the system), and should also prefer nodes
4967 * with no CPUs, since presumably they'll have very little allocation pressure
4968 * on them otherwise.
4969 * It returns -1 if no node is found.
4971 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4974 int min_val = INT_MAX;
4975 int best_node = NUMA_NO_NODE;
4976 const struct cpumask *tmp = cpumask_of_node(0);
4978 /* Use the local node if we haven't already */
4979 if (!node_isset(node, *used_node_mask)) {
4980 node_set(node, *used_node_mask);
4984 for_each_node_state(n, N_MEMORY) {
4986 /* Don't want a node to appear more than once */
4987 if (node_isset(n, *used_node_mask))
4990 /* Use the distance array to find the distance */
4991 val = node_distance(node, n);
4993 /* Penalize nodes under us ("prefer the next node") */
4996 /* Give preference to headless and unused nodes */
4997 tmp = cpumask_of_node(n);
4998 if (!cpumask_empty(tmp))
4999 val += PENALTY_FOR_NODE_WITH_CPUS;
5001 /* Slight preference for less loaded node */
5002 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5003 val += node_load[n];
5005 if (val < min_val) {
5012 node_set(best_node, *used_node_mask);
5019 * Build zonelists ordered by node and zones within node.
5020 * This results in maximum locality--normal zone overflows into local
5021 * DMA zone, if any--but risks exhausting DMA zone.
5023 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5026 struct zoneref *zonerefs;
5029 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5031 for (i = 0; i < nr_nodes; i++) {
5034 pg_data_t *node = NODE_DATA(node_order[i]);
5036 nr_zones = build_zonerefs_node(node, zonerefs);
5037 zonerefs += nr_zones;
5039 zonerefs->zone = NULL;
5040 zonerefs->zone_idx = 0;
5044 * Build gfp_thisnode zonelists
5046 static void build_thisnode_zonelists(pg_data_t *pgdat)
5048 struct zoneref *zonerefs;
5051 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5052 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5053 zonerefs += nr_zones;
5054 zonerefs->zone = NULL;
5055 zonerefs->zone_idx = 0;
5059 * Build zonelists ordered by zone and nodes within zones.
5060 * This results in conserving DMA zone[s] until all Normal memory is
5061 * exhausted, but results in overflowing to remote node while memory
5062 * may still exist in local DMA zone.
5065 static void build_zonelists(pg_data_t *pgdat)
5067 static int node_order[MAX_NUMNODES];
5068 int node, load, nr_nodes = 0;
5069 nodemask_t used_mask;
5070 int local_node, prev_node;
5072 /* NUMA-aware ordering of nodes */
5073 local_node = pgdat->node_id;
5074 load = nr_online_nodes;
5075 prev_node = local_node;
5076 nodes_clear(used_mask);
5078 memset(node_order, 0, sizeof(node_order));
5079 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5081 * We don't want to pressure a particular node.
5082 * So adding penalty to the first node in same
5083 * distance group to make it round-robin.
5085 if (node_distance(local_node, node) !=
5086 node_distance(local_node, prev_node))
5087 node_load[node] = load;
5089 node_order[nr_nodes++] = node;
5094 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5095 build_thisnode_zonelists(pgdat);
5098 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5100 * Return node id of node used for "local" allocations.
5101 * I.e., first node id of first zone in arg node's generic zonelist.
5102 * Used for initializing percpu 'numa_mem', which is used primarily
5103 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5105 int local_memory_node(int node)
5109 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5110 gfp_zone(GFP_KERNEL),
5112 return z->zone->node;
5116 static void setup_min_unmapped_ratio(void);
5117 static void setup_min_slab_ratio(void);
5118 #else /* CONFIG_NUMA */
5120 static void build_zonelists(pg_data_t *pgdat)
5122 int node, local_node;
5123 struct zoneref *zonerefs;
5126 local_node = pgdat->node_id;
5128 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5129 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5130 zonerefs += nr_zones;
5133 * Now we build the zonelist so that it contains the zones
5134 * of all the other nodes.
5135 * We don't want to pressure a particular node, so when
5136 * building the zones for node N, we make sure that the
5137 * zones coming right after the local ones are those from
5138 * node N+1 (modulo N)
5140 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5141 if (!node_online(node))
5143 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5144 zonerefs += nr_zones;
5146 for (node = 0; node < local_node; node++) {
5147 if (!node_online(node))
5149 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5150 zonerefs += nr_zones;
5153 zonerefs->zone = NULL;
5154 zonerefs->zone_idx = 0;
5157 #endif /* CONFIG_NUMA */
5160 * Boot pageset table. One per cpu which is going to be used for all
5161 * zones and all nodes. The parameters will be set in such a way
5162 * that an item put on a list will immediately be handed over to
5163 * the buddy list. This is safe since pageset manipulation is done
5164 * with interrupts disabled.
5166 * The boot_pagesets must be kept even after bootup is complete for
5167 * unused processors and/or zones. They do play a role for bootstrapping
5168 * hotplugged processors.
5170 * zoneinfo_show() and maybe other functions do
5171 * not check if the processor is online before following the pageset pointer.
5172 * Other parts of the kernel may not check if the zone is available.
5174 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5175 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5176 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5178 static void __build_all_zonelists(void *data)
5181 int __maybe_unused cpu;
5182 pg_data_t *self = data;
5183 static DEFINE_SPINLOCK(lock);
5188 memset(node_load, 0, sizeof(node_load));
5192 * This node is hotadded and no memory is yet present. So just
5193 * building zonelists is fine - no need to touch other nodes.
5195 if (self && !node_online(self->node_id)) {
5196 build_zonelists(self);
5198 for_each_online_node(nid) {
5199 pg_data_t *pgdat = NODE_DATA(nid);
5201 build_zonelists(pgdat);
5204 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5206 * We now know the "local memory node" for each node--
5207 * i.e., the node of the first zone in the generic zonelist.
5208 * Set up numa_mem percpu variable for on-line cpus. During
5209 * boot, only the boot cpu should be on-line; we'll init the
5210 * secondary cpus' numa_mem as they come on-line. During
5211 * node/memory hotplug, we'll fixup all on-line cpus.
5213 for_each_online_cpu(cpu)
5214 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5221 static noinline void __init
5222 build_all_zonelists_init(void)
5226 __build_all_zonelists(NULL);
5229 * Initialize the boot_pagesets that are going to be used
5230 * for bootstrapping processors. The real pagesets for
5231 * each zone will be allocated later when the per cpu
5232 * allocator is available.
5234 * boot_pagesets are used also for bootstrapping offline
5235 * cpus if the system is already booted because the pagesets
5236 * are needed to initialize allocators on a specific cpu too.
5237 * F.e. the percpu allocator needs the page allocator which
5238 * needs the percpu allocator in order to allocate its pagesets
5239 * (a chicken-egg dilemma).
5241 for_each_possible_cpu(cpu)
5242 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5244 mminit_verify_zonelist();
5245 cpuset_init_current_mems_allowed();
5249 * unless system_state == SYSTEM_BOOTING.
5251 * __ref due to call of __init annotated helper build_all_zonelists_init
5252 * [protected by SYSTEM_BOOTING].
5254 void __ref build_all_zonelists(pg_data_t *pgdat)
5256 if (system_state == SYSTEM_BOOTING) {
5257 build_all_zonelists_init();
5259 __build_all_zonelists(pgdat);
5260 /* cpuset refresh routine should be here */
5262 vm_total_pages = nr_free_pagecache_pages();
5264 * Disable grouping by mobility if the number of pages in the
5265 * system is too low to allow the mechanism to work. It would be
5266 * more accurate, but expensive to check per-zone. This check is
5267 * made on memory-hotadd so a system can start with mobility
5268 * disabled and enable it later
5270 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5271 page_group_by_mobility_disabled = 1;
5273 page_group_by_mobility_disabled = 0;
5275 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5277 page_group_by_mobility_disabled ? "off" : "on",
5280 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5285 * Initially all pages are reserved - free ones are freed
5286 * up by free_all_bootmem() once the early boot process is
5287 * done. Non-atomic initialization, single-pass.
5289 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5290 unsigned long start_pfn, enum memmap_context context)
5292 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5293 unsigned long end_pfn = start_pfn + size;
5294 pg_data_t *pgdat = NODE_DATA(nid);
5296 unsigned long nr_initialised = 0;
5297 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5298 struct memblock_region *r = NULL, *tmp;
5301 if (highest_memmap_pfn < end_pfn - 1)
5302 highest_memmap_pfn = end_pfn - 1;
5305 * Honor reservation requested by the driver for this ZONE_DEVICE
5308 if (altmap && start_pfn == altmap->base_pfn)
5309 start_pfn += altmap->reserve;
5311 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5313 * There can be holes in boot-time mem_map[]s handed to this
5314 * function. They do not exist on hotplugged memory.
5316 if (context != MEMMAP_EARLY)
5319 if (!early_pfn_valid(pfn)) {
5320 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5322 * Skip to the pfn preceding the next valid one (or
5323 * end_pfn), such that we hit a valid pfn (or end_pfn)
5324 * on our next iteration of the loop.
5326 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5330 if (!early_pfn_in_nid(pfn, nid))
5332 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5335 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5337 * Check given memblock attribute by firmware which can affect
5338 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5339 * mirrored, it's an overlapped memmap init. skip it.
5341 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5342 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5343 for_each_memblock(memory, tmp)
5344 if (pfn < memblock_region_memory_end_pfn(tmp))
5348 if (pfn >= memblock_region_memory_base_pfn(r) &&
5349 memblock_is_mirror(r)) {
5350 /* already initialized as NORMAL */
5351 pfn = memblock_region_memory_end_pfn(r);
5359 * Mark the block movable so that blocks are reserved for
5360 * movable at startup. This will force kernel allocations
5361 * to reserve their blocks rather than leaking throughout
5362 * the address space during boot when many long-lived
5363 * kernel allocations are made.
5365 * bitmap is created for zone's valid pfn range. but memmap
5366 * can be created for invalid pages (for alignment)
5367 * check here not to call set_pageblock_migratetype() against
5370 if (!(pfn & (pageblock_nr_pages - 1))) {
5371 struct page *page = pfn_to_page(pfn);
5373 __init_single_page(page, pfn, zone, nid);
5374 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5377 __init_single_pfn(pfn, zone, nid);
5382 static void __meminit zone_init_free_lists(struct zone *zone)
5384 unsigned int order, t;
5385 for_each_migratetype_order(order, t) {
5386 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5387 zone->free_area[order].nr_free = 0;
5391 #ifndef __HAVE_ARCH_MEMMAP_INIT
5392 #define memmap_init(size, nid, zone, start_pfn) \
5393 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5396 static int zone_batchsize(struct zone *zone)
5402 * The per-cpu-pages pools are set to around 1000th of the
5403 * size of the zone. But no more than 1/2 of a meg.
5405 * OK, so we don't know how big the cache is. So guess.
5407 batch = zone->managed_pages / 1024;
5408 if (batch * PAGE_SIZE > 512 * 1024)
5409 batch = (512 * 1024) / PAGE_SIZE;
5410 batch /= 4; /* We effectively *= 4 below */
5415 * Clamp the batch to a 2^n - 1 value. Having a power
5416 * of 2 value was found to be more likely to have
5417 * suboptimal cache aliasing properties in some cases.
5419 * For example if 2 tasks are alternately allocating
5420 * batches of pages, one task can end up with a lot
5421 * of pages of one half of the possible page colors
5422 * and the other with pages of the other colors.
5424 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5429 /* The deferral and batching of frees should be suppressed under NOMMU
5432 * The problem is that NOMMU needs to be able to allocate large chunks
5433 * of contiguous memory as there's no hardware page translation to
5434 * assemble apparent contiguous memory from discontiguous pages.
5436 * Queueing large contiguous runs of pages for batching, however,
5437 * causes the pages to actually be freed in smaller chunks. As there
5438 * can be a significant delay between the individual batches being
5439 * recycled, this leads to the once large chunks of space being
5440 * fragmented and becoming unavailable for high-order allocations.
5447 * pcp->high and pcp->batch values are related and dependent on one another:
5448 * ->batch must never be higher then ->high.
5449 * The following function updates them in a safe manner without read side
5452 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5453 * those fields changing asynchronously (acording the the above rule).
5455 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5456 * outside of boot time (or some other assurance that no concurrent updaters
5459 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5460 unsigned long batch)
5462 /* start with a fail safe value for batch */
5466 /* Update high, then batch, in order */
5473 /* a companion to pageset_set_high() */
5474 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5476 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5479 static void pageset_init(struct per_cpu_pageset *p)
5481 struct per_cpu_pages *pcp;
5484 memset(p, 0, sizeof(*p));
5488 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5489 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5492 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5495 pageset_set_batch(p, batch);
5499 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5500 * to the value high for the pageset p.
5502 static void pageset_set_high(struct per_cpu_pageset *p,
5505 unsigned long batch = max(1UL, high / 4);
5506 if ((high / 4) > (PAGE_SHIFT * 8))
5507 batch = PAGE_SHIFT * 8;
5509 pageset_update(&p->pcp, high, batch);
5512 static void pageset_set_high_and_batch(struct zone *zone,
5513 struct per_cpu_pageset *pcp)
5515 if (percpu_pagelist_fraction)
5516 pageset_set_high(pcp,
5517 (zone->managed_pages /
5518 percpu_pagelist_fraction));
5520 pageset_set_batch(pcp, zone_batchsize(zone));
5523 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5525 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5528 pageset_set_high_and_batch(zone, pcp);
5531 void __meminit setup_zone_pageset(struct zone *zone)
5534 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5535 for_each_possible_cpu(cpu)
5536 zone_pageset_init(zone, cpu);
5540 * Allocate per cpu pagesets and initialize them.
5541 * Before this call only boot pagesets were available.
5543 void __init setup_per_cpu_pageset(void)
5545 struct pglist_data *pgdat;
5548 for_each_populated_zone(zone)
5549 setup_zone_pageset(zone);
5551 for_each_online_pgdat(pgdat)
5552 pgdat->per_cpu_nodestats =
5553 alloc_percpu(struct per_cpu_nodestat);
5556 static __meminit void zone_pcp_init(struct zone *zone)
5559 * per cpu subsystem is not up at this point. The following code
5560 * relies on the ability of the linker to provide the
5561 * offset of a (static) per cpu variable into the per cpu area.
5563 zone->pageset = &boot_pageset;
5565 if (populated_zone(zone))
5566 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5567 zone->name, zone->present_pages,
5568 zone_batchsize(zone));
5571 void __meminit init_currently_empty_zone(struct zone *zone,
5572 unsigned long zone_start_pfn,
5575 struct pglist_data *pgdat = zone->zone_pgdat;
5577 pgdat->nr_zones = zone_idx(zone) + 1;
5579 zone->zone_start_pfn = zone_start_pfn;
5581 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5582 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5584 (unsigned long)zone_idx(zone),
5585 zone_start_pfn, (zone_start_pfn + size));
5587 zone_init_free_lists(zone);
5588 zone->initialized = 1;
5591 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5592 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5595 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5597 int __meminit __early_pfn_to_nid(unsigned long pfn,
5598 struct mminit_pfnnid_cache *state)
5600 unsigned long start_pfn, end_pfn;
5603 if (state->last_start <= pfn && pfn < state->last_end)
5604 return state->last_nid;
5606 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5608 state->last_start = start_pfn;
5609 state->last_end = end_pfn;
5610 state->last_nid = nid;
5615 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5618 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5619 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5620 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5622 * If an architecture guarantees that all ranges registered contain no holes
5623 * and may be freed, this this function may be used instead of calling
5624 * memblock_free_early_nid() manually.
5626 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5628 unsigned long start_pfn, end_pfn;
5631 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5632 start_pfn = min(start_pfn, max_low_pfn);
5633 end_pfn = min(end_pfn, max_low_pfn);
5635 if (start_pfn < end_pfn)
5636 memblock_free_early_nid(PFN_PHYS(start_pfn),
5637 (end_pfn - start_pfn) << PAGE_SHIFT,
5643 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5644 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5646 * If an architecture guarantees that all ranges registered contain no holes and may
5647 * be freed, this function may be used instead of calling memory_present() manually.
5649 void __init sparse_memory_present_with_active_regions(int nid)
5651 unsigned long start_pfn, end_pfn;
5654 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5655 memory_present(this_nid, start_pfn, end_pfn);
5659 * get_pfn_range_for_nid - Return the start and end page frames for a node
5660 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5661 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5662 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5664 * It returns the start and end page frame of a node based on information
5665 * provided by memblock_set_node(). If called for a node
5666 * with no available memory, a warning is printed and the start and end
5669 void __meminit get_pfn_range_for_nid(unsigned int nid,
5670 unsigned long *start_pfn, unsigned long *end_pfn)
5672 unsigned long this_start_pfn, this_end_pfn;
5678 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5679 *start_pfn = min(*start_pfn, this_start_pfn);
5680 *end_pfn = max(*end_pfn, this_end_pfn);
5683 if (*start_pfn == -1UL)
5688 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5689 * assumption is made that zones within a node are ordered in monotonic
5690 * increasing memory addresses so that the "highest" populated zone is used
5692 static void __init find_usable_zone_for_movable(void)
5695 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5696 if (zone_index == ZONE_MOVABLE)
5699 if (arch_zone_highest_possible_pfn[zone_index] >
5700 arch_zone_lowest_possible_pfn[zone_index])
5704 VM_BUG_ON(zone_index == -1);
5705 movable_zone = zone_index;
5709 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5710 * because it is sized independent of architecture. Unlike the other zones,
5711 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5712 * in each node depending on the size of each node and how evenly kernelcore
5713 * is distributed. This helper function adjusts the zone ranges
5714 * provided by the architecture for a given node by using the end of the
5715 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5716 * zones within a node are in order of monotonic increases memory addresses
5718 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5719 unsigned long zone_type,
5720 unsigned long node_start_pfn,
5721 unsigned long node_end_pfn,
5722 unsigned long *zone_start_pfn,
5723 unsigned long *zone_end_pfn)
5725 /* Only adjust if ZONE_MOVABLE is on this node */
5726 if (zone_movable_pfn[nid]) {
5727 /* Size ZONE_MOVABLE */
5728 if (zone_type == ZONE_MOVABLE) {
5729 *zone_start_pfn = zone_movable_pfn[nid];
5730 *zone_end_pfn = min(node_end_pfn,
5731 arch_zone_highest_possible_pfn[movable_zone]);
5733 /* Adjust for ZONE_MOVABLE starting within this range */
5734 } else if (!mirrored_kernelcore &&
5735 *zone_start_pfn < zone_movable_pfn[nid] &&
5736 *zone_end_pfn > zone_movable_pfn[nid]) {
5737 *zone_end_pfn = zone_movable_pfn[nid];
5739 /* Check if this whole range is within ZONE_MOVABLE */
5740 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5741 *zone_start_pfn = *zone_end_pfn;
5746 * Return the number of pages a zone spans in a node, including holes
5747 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5749 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5750 unsigned long zone_type,
5751 unsigned long node_start_pfn,
5752 unsigned long node_end_pfn,
5753 unsigned long *zone_start_pfn,
5754 unsigned long *zone_end_pfn,
5755 unsigned long *ignored)
5757 /* When hotadd a new node from cpu_up(), the node should be empty */
5758 if (!node_start_pfn && !node_end_pfn)
5761 /* Get the start and end of the zone */
5762 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5763 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5764 adjust_zone_range_for_zone_movable(nid, zone_type,
5765 node_start_pfn, node_end_pfn,
5766 zone_start_pfn, zone_end_pfn);
5768 /* Check that this node has pages within the zone's required range */
5769 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5772 /* Move the zone boundaries inside the node if necessary */
5773 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5774 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5776 /* Return the spanned pages */
5777 return *zone_end_pfn - *zone_start_pfn;
5781 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5782 * then all holes in the requested range will be accounted for.
5784 unsigned long __meminit __absent_pages_in_range(int nid,
5785 unsigned long range_start_pfn,
5786 unsigned long range_end_pfn)
5788 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5789 unsigned long start_pfn, end_pfn;
5792 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5793 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5794 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5795 nr_absent -= end_pfn - start_pfn;
5801 * absent_pages_in_range - Return number of page frames in holes within a range
5802 * @start_pfn: The start PFN to start searching for holes
5803 * @end_pfn: The end PFN to stop searching for holes
5805 * It returns the number of pages frames in memory holes within a range.
5807 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5808 unsigned long end_pfn)
5810 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5813 /* Return the number of page frames in holes in a zone on a node */
5814 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5815 unsigned long zone_type,
5816 unsigned long node_start_pfn,
5817 unsigned long node_end_pfn,
5818 unsigned long *ignored)
5820 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5821 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5822 unsigned long zone_start_pfn, zone_end_pfn;
5823 unsigned long nr_absent;
5825 /* When hotadd a new node from cpu_up(), the node should be empty */
5826 if (!node_start_pfn && !node_end_pfn)
5829 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5830 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5832 adjust_zone_range_for_zone_movable(nid, zone_type,
5833 node_start_pfn, node_end_pfn,
5834 &zone_start_pfn, &zone_end_pfn);
5835 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5838 * ZONE_MOVABLE handling.
5839 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5842 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5843 unsigned long start_pfn, end_pfn;
5844 struct memblock_region *r;
5846 for_each_memblock(memory, r) {
5847 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5848 zone_start_pfn, zone_end_pfn);
5849 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5850 zone_start_pfn, zone_end_pfn);
5852 if (zone_type == ZONE_MOVABLE &&
5853 memblock_is_mirror(r))
5854 nr_absent += end_pfn - start_pfn;
5856 if (zone_type == ZONE_NORMAL &&
5857 !memblock_is_mirror(r))
5858 nr_absent += end_pfn - start_pfn;
5865 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5866 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5867 unsigned long zone_type,
5868 unsigned long node_start_pfn,
5869 unsigned long node_end_pfn,
5870 unsigned long *zone_start_pfn,
5871 unsigned long *zone_end_pfn,
5872 unsigned long *zones_size)
5876 *zone_start_pfn = node_start_pfn;
5877 for (zone = 0; zone < zone_type; zone++)
5878 *zone_start_pfn += zones_size[zone];
5880 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5882 return zones_size[zone_type];
5885 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5886 unsigned long zone_type,
5887 unsigned long node_start_pfn,
5888 unsigned long node_end_pfn,
5889 unsigned long *zholes_size)
5894 return zholes_size[zone_type];
5897 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5899 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5900 unsigned long node_start_pfn,
5901 unsigned long node_end_pfn,
5902 unsigned long *zones_size,
5903 unsigned long *zholes_size)
5905 unsigned long realtotalpages = 0, totalpages = 0;
5908 for (i = 0; i < MAX_NR_ZONES; i++) {
5909 struct zone *zone = pgdat->node_zones + i;
5910 unsigned long zone_start_pfn, zone_end_pfn;
5911 unsigned long size, real_size;
5913 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5919 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5920 node_start_pfn, node_end_pfn,
5923 zone->zone_start_pfn = zone_start_pfn;
5925 zone->zone_start_pfn = 0;
5926 zone->spanned_pages = size;
5927 zone->present_pages = real_size;
5930 realtotalpages += real_size;
5933 pgdat->node_spanned_pages = totalpages;
5934 pgdat->node_present_pages = realtotalpages;
5935 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5939 #ifndef CONFIG_SPARSEMEM
5941 * Calculate the size of the zone->blockflags rounded to an unsigned long
5942 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5943 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5944 * round what is now in bits to nearest long in bits, then return it in
5947 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5949 unsigned long usemapsize;
5951 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5952 usemapsize = roundup(zonesize, pageblock_nr_pages);
5953 usemapsize = usemapsize >> pageblock_order;
5954 usemapsize *= NR_PAGEBLOCK_BITS;
5955 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5957 return usemapsize / 8;
5960 static void __init setup_usemap(struct pglist_data *pgdat,
5962 unsigned long zone_start_pfn,
5963 unsigned long zonesize)
5965 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5966 zone->pageblock_flags = NULL;
5968 zone->pageblock_flags =
5969 memblock_virt_alloc_node_nopanic(usemapsize,
5973 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5974 unsigned long zone_start_pfn, unsigned long zonesize) {}
5975 #endif /* CONFIG_SPARSEMEM */
5977 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5979 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5980 void __paginginit set_pageblock_order(void)
5984 /* Check that pageblock_nr_pages has not already been setup */
5985 if (pageblock_order)
5988 if (HPAGE_SHIFT > PAGE_SHIFT)
5989 order = HUGETLB_PAGE_ORDER;
5991 order = MAX_ORDER - 1;
5994 * Assume the largest contiguous order of interest is a huge page.
5995 * This value may be variable depending on boot parameters on IA64 and
5998 pageblock_order = order;
6000 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6003 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6004 * is unused as pageblock_order is set at compile-time. See
6005 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6008 void __paginginit set_pageblock_order(void)
6012 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6014 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6015 unsigned long present_pages)
6017 unsigned long pages = spanned_pages;
6020 * Provide a more accurate estimation if there are holes within
6021 * the zone and SPARSEMEM is in use. If there are holes within the
6022 * zone, each populated memory region may cost us one or two extra
6023 * memmap pages due to alignment because memmap pages for each
6024 * populated regions may not be naturally aligned on page boundary.
6025 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6027 if (spanned_pages > present_pages + (present_pages >> 4) &&
6028 IS_ENABLED(CONFIG_SPARSEMEM))
6029 pages = present_pages;
6031 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6035 * Set up the zone data structures:
6036 * - mark all pages reserved
6037 * - mark all memory queues empty
6038 * - clear the memory bitmaps
6040 * NOTE: pgdat should get zeroed by caller.
6042 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6045 int nid = pgdat->node_id;
6047 pgdat_resize_init(pgdat);
6048 #ifdef CONFIG_NUMA_BALANCING
6049 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6050 pgdat->numabalancing_migrate_nr_pages = 0;
6051 pgdat->numabalancing_migrate_next_window = jiffies;
6053 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6054 spin_lock_init(&pgdat->split_queue_lock);
6055 INIT_LIST_HEAD(&pgdat->split_queue);
6056 pgdat->split_queue_len = 0;
6058 init_waitqueue_head(&pgdat->kswapd_wait);
6059 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6060 #ifdef CONFIG_COMPACTION
6061 init_waitqueue_head(&pgdat->kcompactd_wait);
6063 pgdat_page_ext_init(pgdat);
6064 spin_lock_init(&pgdat->lru_lock);
6065 lruvec_init(node_lruvec(pgdat));
6067 pgdat->per_cpu_nodestats = &boot_nodestats;
6069 for (j = 0; j < MAX_NR_ZONES; j++) {
6070 struct zone *zone = pgdat->node_zones + j;
6071 unsigned long size, realsize, freesize, memmap_pages;
6072 unsigned long zone_start_pfn = zone->zone_start_pfn;
6074 size = zone->spanned_pages;
6075 realsize = freesize = zone->present_pages;
6078 * Adjust freesize so that it accounts for how much memory
6079 * is used by this zone for memmap. This affects the watermark
6080 * and per-cpu initialisations
6082 memmap_pages = calc_memmap_size(size, realsize);
6083 if (!is_highmem_idx(j)) {
6084 if (freesize >= memmap_pages) {
6085 freesize -= memmap_pages;
6088 " %s zone: %lu pages used for memmap\n",
6089 zone_names[j], memmap_pages);
6091 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6092 zone_names[j], memmap_pages, freesize);
6095 /* Account for reserved pages */
6096 if (j == 0 && freesize > dma_reserve) {
6097 freesize -= dma_reserve;
6098 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6099 zone_names[0], dma_reserve);
6102 if (!is_highmem_idx(j))
6103 nr_kernel_pages += freesize;
6104 /* Charge for highmem memmap if there are enough kernel pages */
6105 else if (nr_kernel_pages > memmap_pages * 2)
6106 nr_kernel_pages -= memmap_pages;
6107 nr_all_pages += freesize;
6110 * Set an approximate value for lowmem here, it will be adjusted
6111 * when the bootmem allocator frees pages into the buddy system.
6112 * And all highmem pages will be managed by the buddy system.
6114 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6118 zone->name = zone_names[j];
6119 zone->zone_pgdat = pgdat;
6120 spin_lock_init(&zone->lock);
6121 zone_seqlock_init(zone);
6122 zone_pcp_init(zone);
6127 set_pageblock_order();
6128 setup_usemap(pgdat, zone, zone_start_pfn, size);
6129 init_currently_empty_zone(zone, zone_start_pfn, size);
6130 memmap_init(size, nid, j, zone_start_pfn);
6134 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6136 unsigned long __maybe_unused start = 0;
6137 unsigned long __maybe_unused offset = 0;
6139 /* Skip empty nodes */
6140 if (!pgdat->node_spanned_pages)
6143 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6144 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6145 offset = pgdat->node_start_pfn - start;
6146 /* ia64 gets its own node_mem_map, before this, without bootmem */
6147 if (!pgdat->node_mem_map) {
6148 unsigned long size, end;
6152 * The zone's endpoints aren't required to be MAX_ORDER
6153 * aligned but the node_mem_map endpoints must be in order
6154 * for the buddy allocator to function correctly.
6156 end = pgdat_end_pfn(pgdat);
6157 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6158 size = (end - start) * sizeof(struct page);
6159 map = alloc_remap(pgdat->node_id, size);
6161 map = memblock_virt_alloc_node_nopanic(size,
6163 pgdat->node_mem_map = map + offset;
6165 #ifndef CONFIG_NEED_MULTIPLE_NODES
6167 * With no DISCONTIG, the global mem_map is just set as node 0's
6169 if (pgdat == NODE_DATA(0)) {
6170 mem_map = NODE_DATA(0)->node_mem_map;
6171 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6172 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6174 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6177 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6180 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6181 unsigned long node_start_pfn, unsigned long *zholes_size)
6183 pg_data_t *pgdat = NODE_DATA(nid);
6184 unsigned long start_pfn = 0;
6185 unsigned long end_pfn = 0;
6187 /* pg_data_t should be reset to zero when it's allocated */
6188 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6190 pgdat->node_id = nid;
6191 pgdat->node_start_pfn = node_start_pfn;
6192 pgdat->per_cpu_nodestats = NULL;
6193 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6194 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6195 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6196 (u64)start_pfn << PAGE_SHIFT,
6197 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6199 start_pfn = node_start_pfn;
6201 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6202 zones_size, zholes_size);
6204 alloc_node_mem_map(pgdat);
6205 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6206 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6207 nid, (unsigned long)pgdat,
6208 (unsigned long)pgdat->node_mem_map);
6211 reset_deferred_meminit(pgdat);
6212 free_area_init_core(pgdat);
6215 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6217 #if MAX_NUMNODES > 1
6219 * Figure out the number of possible node ids.
6221 void __init setup_nr_node_ids(void)
6223 unsigned int highest;
6225 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6226 nr_node_ids = highest + 1;
6231 * node_map_pfn_alignment - determine the maximum internode alignment
6233 * This function should be called after node map is populated and sorted.
6234 * It calculates the maximum power of two alignment which can distinguish
6237 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6238 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6239 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6240 * shifted, 1GiB is enough and this function will indicate so.
6242 * This is used to test whether pfn -> nid mapping of the chosen memory
6243 * model has fine enough granularity to avoid incorrect mapping for the
6244 * populated node map.
6246 * Returns the determined alignment in pfn's. 0 if there is no alignment
6247 * requirement (single node).
6249 unsigned long __init node_map_pfn_alignment(void)
6251 unsigned long accl_mask = 0, last_end = 0;
6252 unsigned long start, end, mask;
6256 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6257 if (!start || last_nid < 0 || last_nid == nid) {
6264 * Start with a mask granular enough to pin-point to the
6265 * start pfn and tick off bits one-by-one until it becomes
6266 * too coarse to separate the current node from the last.
6268 mask = ~((1 << __ffs(start)) - 1);
6269 while (mask && last_end <= (start & (mask << 1)))
6272 /* accumulate all internode masks */
6276 /* convert mask to number of pages */
6277 return ~accl_mask + 1;
6280 /* Find the lowest pfn for a node */
6281 static unsigned long __init find_min_pfn_for_node(int nid)
6283 unsigned long min_pfn = ULONG_MAX;
6284 unsigned long start_pfn;
6287 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6288 min_pfn = min(min_pfn, start_pfn);
6290 if (min_pfn == ULONG_MAX) {
6291 pr_warn("Could not find start_pfn for node %d\n", nid);
6299 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6301 * It returns the minimum PFN based on information provided via
6302 * memblock_set_node().
6304 unsigned long __init find_min_pfn_with_active_regions(void)
6306 return find_min_pfn_for_node(MAX_NUMNODES);
6310 * early_calculate_totalpages()
6311 * Sum pages in active regions for movable zone.
6312 * Populate N_MEMORY for calculating usable_nodes.
6314 static unsigned long __init early_calculate_totalpages(void)
6316 unsigned long totalpages = 0;
6317 unsigned long start_pfn, end_pfn;
6320 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6321 unsigned long pages = end_pfn - start_pfn;
6323 totalpages += pages;
6325 node_set_state(nid, N_MEMORY);
6331 * Find the PFN the Movable zone begins in each node. Kernel memory
6332 * is spread evenly between nodes as long as the nodes have enough
6333 * memory. When they don't, some nodes will have more kernelcore than
6336 static void __init find_zone_movable_pfns_for_nodes(void)
6339 unsigned long usable_startpfn;
6340 unsigned long kernelcore_node, kernelcore_remaining;
6341 /* save the state before borrow the nodemask */
6342 nodemask_t saved_node_state = node_states[N_MEMORY];
6343 unsigned long totalpages = early_calculate_totalpages();
6344 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6345 struct memblock_region *r;
6347 /* Need to find movable_zone earlier when movable_node is specified. */
6348 find_usable_zone_for_movable();
6351 * If movable_node is specified, ignore kernelcore and movablecore
6354 if (movable_node_is_enabled()) {
6355 for_each_memblock(memory, r) {
6356 if (!memblock_is_hotpluggable(r))
6361 usable_startpfn = PFN_DOWN(r->base);
6362 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6363 min(usable_startpfn, zone_movable_pfn[nid]) :
6371 * If kernelcore=mirror is specified, ignore movablecore option
6373 if (mirrored_kernelcore) {
6374 bool mem_below_4gb_not_mirrored = false;
6376 for_each_memblock(memory, r) {
6377 if (memblock_is_mirror(r))
6382 usable_startpfn = memblock_region_memory_base_pfn(r);
6384 if (usable_startpfn < 0x100000) {
6385 mem_below_4gb_not_mirrored = true;
6389 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6390 min(usable_startpfn, zone_movable_pfn[nid]) :
6394 if (mem_below_4gb_not_mirrored)
6395 pr_warn("This configuration results in unmirrored kernel memory.");
6401 * If movablecore=nn[KMG] was specified, calculate what size of
6402 * kernelcore that corresponds so that memory usable for
6403 * any allocation type is evenly spread. If both kernelcore
6404 * and movablecore are specified, then the value of kernelcore
6405 * will be used for required_kernelcore if it's greater than
6406 * what movablecore would have allowed.
6408 if (required_movablecore) {
6409 unsigned long corepages;
6412 * Round-up so that ZONE_MOVABLE is at least as large as what
6413 * was requested by the user
6415 required_movablecore =
6416 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6417 required_movablecore = min(totalpages, required_movablecore);
6418 corepages = totalpages - required_movablecore;
6420 required_kernelcore = max(required_kernelcore, corepages);
6424 * If kernelcore was not specified or kernelcore size is larger
6425 * than totalpages, there is no ZONE_MOVABLE.
6427 if (!required_kernelcore || required_kernelcore >= totalpages)
6430 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6431 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6434 /* Spread kernelcore memory as evenly as possible throughout nodes */
6435 kernelcore_node = required_kernelcore / usable_nodes;
6436 for_each_node_state(nid, N_MEMORY) {
6437 unsigned long start_pfn, end_pfn;
6440 * Recalculate kernelcore_node if the division per node
6441 * now exceeds what is necessary to satisfy the requested
6442 * amount of memory for the kernel
6444 if (required_kernelcore < kernelcore_node)
6445 kernelcore_node = required_kernelcore / usable_nodes;
6448 * As the map is walked, we track how much memory is usable
6449 * by the kernel using kernelcore_remaining. When it is
6450 * 0, the rest of the node is usable by ZONE_MOVABLE
6452 kernelcore_remaining = kernelcore_node;
6454 /* Go through each range of PFNs within this node */
6455 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6456 unsigned long size_pages;
6458 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6459 if (start_pfn >= end_pfn)
6462 /* Account for what is only usable for kernelcore */
6463 if (start_pfn < usable_startpfn) {
6464 unsigned long kernel_pages;
6465 kernel_pages = min(end_pfn, usable_startpfn)
6468 kernelcore_remaining -= min(kernel_pages,
6469 kernelcore_remaining);
6470 required_kernelcore -= min(kernel_pages,
6471 required_kernelcore);
6473 /* Continue if range is now fully accounted */
6474 if (end_pfn <= usable_startpfn) {
6477 * Push zone_movable_pfn to the end so
6478 * that if we have to rebalance
6479 * kernelcore across nodes, we will
6480 * not double account here
6482 zone_movable_pfn[nid] = end_pfn;
6485 start_pfn = usable_startpfn;
6489 * The usable PFN range for ZONE_MOVABLE is from
6490 * start_pfn->end_pfn. Calculate size_pages as the
6491 * number of pages used as kernelcore
6493 size_pages = end_pfn - start_pfn;
6494 if (size_pages > kernelcore_remaining)
6495 size_pages = kernelcore_remaining;
6496 zone_movable_pfn[nid] = start_pfn + size_pages;
6499 * Some kernelcore has been met, update counts and
6500 * break if the kernelcore for this node has been
6503 required_kernelcore -= min(required_kernelcore,
6505 kernelcore_remaining -= size_pages;
6506 if (!kernelcore_remaining)
6512 * If there is still required_kernelcore, we do another pass with one
6513 * less node in the count. This will push zone_movable_pfn[nid] further
6514 * along on the nodes that still have memory until kernelcore is
6518 if (usable_nodes && required_kernelcore > usable_nodes)
6522 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6523 for (nid = 0; nid < MAX_NUMNODES; nid++)
6524 zone_movable_pfn[nid] =
6525 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6528 /* restore the node_state */
6529 node_states[N_MEMORY] = saved_node_state;
6532 /* Any regular or high memory on that node ? */
6533 static void check_for_memory(pg_data_t *pgdat, int nid)
6535 enum zone_type zone_type;
6537 if (N_MEMORY == N_NORMAL_MEMORY)
6540 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6541 struct zone *zone = &pgdat->node_zones[zone_type];
6542 if (populated_zone(zone)) {
6543 node_set_state(nid, N_HIGH_MEMORY);
6544 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6545 zone_type <= ZONE_NORMAL)
6546 node_set_state(nid, N_NORMAL_MEMORY);
6553 * free_area_init_nodes - Initialise all pg_data_t and zone data
6554 * @max_zone_pfn: an array of max PFNs for each zone
6556 * This will call free_area_init_node() for each active node in the system.
6557 * Using the page ranges provided by memblock_set_node(), the size of each
6558 * zone in each node and their holes is calculated. If the maximum PFN
6559 * between two adjacent zones match, it is assumed that the zone is empty.
6560 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6561 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6562 * starts where the previous one ended. For example, ZONE_DMA32 starts
6563 * at arch_max_dma_pfn.
6565 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6567 unsigned long start_pfn, end_pfn;
6570 /* Record where the zone boundaries are */
6571 memset(arch_zone_lowest_possible_pfn, 0,
6572 sizeof(arch_zone_lowest_possible_pfn));
6573 memset(arch_zone_highest_possible_pfn, 0,
6574 sizeof(arch_zone_highest_possible_pfn));
6576 start_pfn = find_min_pfn_with_active_regions();
6578 for (i = 0; i < MAX_NR_ZONES; i++) {
6579 if (i == ZONE_MOVABLE)
6582 end_pfn = max(max_zone_pfn[i], start_pfn);
6583 arch_zone_lowest_possible_pfn[i] = start_pfn;
6584 arch_zone_highest_possible_pfn[i] = end_pfn;
6586 start_pfn = end_pfn;
6589 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6590 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6591 find_zone_movable_pfns_for_nodes();
6593 /* Print out the zone ranges */
6594 pr_info("Zone ranges:\n");
6595 for (i = 0; i < MAX_NR_ZONES; i++) {
6596 if (i == ZONE_MOVABLE)
6598 pr_info(" %-8s ", zone_names[i]);
6599 if (arch_zone_lowest_possible_pfn[i] ==
6600 arch_zone_highest_possible_pfn[i])
6603 pr_cont("[mem %#018Lx-%#018Lx]\n",
6604 (u64)arch_zone_lowest_possible_pfn[i]
6606 ((u64)arch_zone_highest_possible_pfn[i]
6607 << PAGE_SHIFT) - 1);
6610 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6611 pr_info("Movable zone start for each node\n");
6612 for (i = 0; i < MAX_NUMNODES; i++) {
6613 if (zone_movable_pfn[i])
6614 pr_info(" Node %d: %#018Lx\n", i,
6615 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6618 /* Print out the early node map */
6619 pr_info("Early memory node ranges\n");
6620 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6621 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6622 (u64)start_pfn << PAGE_SHIFT,
6623 ((u64)end_pfn << PAGE_SHIFT) - 1);
6625 /* Initialise every node */
6626 mminit_verify_pageflags_layout();
6627 setup_nr_node_ids();
6628 for_each_online_node(nid) {
6629 pg_data_t *pgdat = NODE_DATA(nid);
6630 free_area_init_node(nid, NULL,
6631 find_min_pfn_for_node(nid), NULL);
6633 /* Any memory on that node */
6634 if (pgdat->node_present_pages)
6635 node_set_state(nid, N_MEMORY);
6636 check_for_memory(pgdat, nid);
6640 static int __init cmdline_parse_core(char *p, unsigned long *core)
6642 unsigned long long coremem;
6646 coremem = memparse(p, &p);
6647 *core = coremem >> PAGE_SHIFT;
6649 /* Paranoid check that UL is enough for the coremem value */
6650 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6656 * kernelcore=size sets the amount of memory for use for allocations that
6657 * cannot be reclaimed or migrated.
6659 static int __init cmdline_parse_kernelcore(char *p)
6661 /* parse kernelcore=mirror */
6662 if (parse_option_str(p, "mirror")) {
6663 mirrored_kernelcore = true;
6667 return cmdline_parse_core(p, &required_kernelcore);
6671 * movablecore=size sets the amount of memory for use for allocations that
6672 * can be reclaimed or migrated.
6674 static int __init cmdline_parse_movablecore(char *p)
6676 return cmdline_parse_core(p, &required_movablecore);
6679 early_param("kernelcore", cmdline_parse_kernelcore);
6680 early_param("movablecore", cmdline_parse_movablecore);
6682 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6684 void adjust_managed_page_count(struct page *page, long count)
6686 spin_lock(&managed_page_count_lock);
6687 page_zone(page)->managed_pages += count;
6688 totalram_pages += count;
6689 #ifdef CONFIG_HIGHMEM
6690 if (PageHighMem(page))
6691 totalhigh_pages += count;
6693 spin_unlock(&managed_page_count_lock);
6695 EXPORT_SYMBOL(adjust_managed_page_count);
6697 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6700 unsigned long pages = 0;
6702 start = (void *)PAGE_ALIGN((unsigned long)start);
6703 end = (void *)((unsigned long)end & PAGE_MASK);
6704 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6705 if ((unsigned int)poison <= 0xFF)
6706 memset(pos, poison, PAGE_SIZE);
6707 free_reserved_page(virt_to_page(pos));
6711 pr_info("Freeing %s memory: %ldK\n",
6712 s, pages << (PAGE_SHIFT - 10));
6716 EXPORT_SYMBOL(free_reserved_area);
6718 #ifdef CONFIG_HIGHMEM
6719 void free_highmem_page(struct page *page)
6721 __free_reserved_page(page);
6723 page_zone(page)->managed_pages++;
6729 void __init mem_init_print_info(const char *str)
6731 unsigned long physpages, codesize, datasize, rosize, bss_size;
6732 unsigned long init_code_size, init_data_size;
6734 physpages = get_num_physpages();
6735 codesize = _etext - _stext;
6736 datasize = _edata - _sdata;
6737 rosize = __end_rodata - __start_rodata;
6738 bss_size = __bss_stop - __bss_start;
6739 init_data_size = __init_end - __init_begin;
6740 init_code_size = _einittext - _sinittext;
6743 * Detect special cases and adjust section sizes accordingly:
6744 * 1) .init.* may be embedded into .data sections
6745 * 2) .init.text.* may be out of [__init_begin, __init_end],
6746 * please refer to arch/tile/kernel/vmlinux.lds.S.
6747 * 3) .rodata.* may be embedded into .text or .data sections.
6749 #define adj_init_size(start, end, size, pos, adj) \
6751 if (start <= pos && pos < end && size > adj) \
6755 adj_init_size(__init_begin, __init_end, init_data_size,
6756 _sinittext, init_code_size);
6757 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6758 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6759 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6760 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6762 #undef adj_init_size
6764 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6765 #ifdef CONFIG_HIGHMEM
6769 nr_free_pages() << (PAGE_SHIFT - 10),
6770 physpages << (PAGE_SHIFT - 10),
6771 codesize >> 10, datasize >> 10, rosize >> 10,
6772 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6773 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6774 totalcma_pages << (PAGE_SHIFT - 10),
6775 #ifdef CONFIG_HIGHMEM
6776 totalhigh_pages << (PAGE_SHIFT - 10),
6778 str ? ", " : "", str ? str : "");
6782 * set_dma_reserve - set the specified number of pages reserved in the first zone
6783 * @new_dma_reserve: The number of pages to mark reserved
6785 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6786 * In the DMA zone, a significant percentage may be consumed by kernel image
6787 * and other unfreeable allocations which can skew the watermarks badly. This
6788 * function may optionally be used to account for unfreeable pages in the
6789 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6790 * smaller per-cpu batchsize.
6792 void __init set_dma_reserve(unsigned long new_dma_reserve)
6794 dma_reserve = new_dma_reserve;
6797 void __init free_area_init(unsigned long *zones_size)
6799 free_area_init_node(0, zones_size,
6800 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6803 static int page_alloc_cpu_dead(unsigned int cpu)
6806 lru_add_drain_cpu(cpu);
6810 * Spill the event counters of the dead processor
6811 * into the current processors event counters.
6812 * This artificially elevates the count of the current
6815 vm_events_fold_cpu(cpu);
6818 * Zero the differential counters of the dead processor
6819 * so that the vm statistics are consistent.
6821 * This is only okay since the processor is dead and cannot
6822 * race with what we are doing.
6824 cpu_vm_stats_fold(cpu);
6828 void __init page_alloc_init(void)
6832 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6833 "mm/page_alloc:dead", NULL,
6834 page_alloc_cpu_dead);
6839 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6840 * or min_free_kbytes changes.
6842 static void calculate_totalreserve_pages(void)
6844 struct pglist_data *pgdat;
6845 unsigned long reserve_pages = 0;
6846 enum zone_type i, j;
6848 for_each_online_pgdat(pgdat) {
6850 pgdat->totalreserve_pages = 0;
6852 for (i = 0; i < MAX_NR_ZONES; i++) {
6853 struct zone *zone = pgdat->node_zones + i;
6856 /* Find valid and maximum lowmem_reserve in the zone */
6857 for (j = i; j < MAX_NR_ZONES; j++) {
6858 if (zone->lowmem_reserve[j] > max)
6859 max = zone->lowmem_reserve[j];
6862 /* we treat the high watermark as reserved pages. */
6863 max += high_wmark_pages(zone);
6865 if (max > zone->managed_pages)
6866 max = zone->managed_pages;
6868 pgdat->totalreserve_pages += max;
6870 reserve_pages += max;
6873 totalreserve_pages = reserve_pages;
6877 * setup_per_zone_lowmem_reserve - called whenever
6878 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6879 * has a correct pages reserved value, so an adequate number of
6880 * pages are left in the zone after a successful __alloc_pages().
6882 static void setup_per_zone_lowmem_reserve(void)
6884 struct pglist_data *pgdat;
6885 enum zone_type j, idx;
6887 for_each_online_pgdat(pgdat) {
6888 for (j = 0; j < MAX_NR_ZONES; j++) {
6889 struct zone *zone = pgdat->node_zones + j;
6890 unsigned long managed_pages = zone->managed_pages;
6892 zone->lowmem_reserve[j] = 0;
6896 struct zone *lower_zone;
6900 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6901 sysctl_lowmem_reserve_ratio[idx] = 1;
6903 lower_zone = pgdat->node_zones + idx;
6904 lower_zone->lowmem_reserve[j] = managed_pages /
6905 sysctl_lowmem_reserve_ratio[idx];
6906 managed_pages += lower_zone->managed_pages;
6911 /* update totalreserve_pages */
6912 calculate_totalreserve_pages();
6915 static void __setup_per_zone_wmarks(void)
6917 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6918 unsigned long lowmem_pages = 0;
6920 unsigned long flags;
6922 /* Calculate total number of !ZONE_HIGHMEM pages */
6923 for_each_zone(zone) {
6924 if (!is_highmem(zone))
6925 lowmem_pages += zone->managed_pages;
6928 for_each_zone(zone) {
6931 spin_lock_irqsave(&zone->lock, flags);
6932 tmp = (u64)pages_min * zone->managed_pages;
6933 do_div(tmp, lowmem_pages);
6934 if (is_highmem(zone)) {
6936 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6937 * need highmem pages, so cap pages_min to a small
6940 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6941 * deltas control asynch page reclaim, and so should
6942 * not be capped for highmem.
6944 unsigned long min_pages;
6946 min_pages = zone->managed_pages / 1024;
6947 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6948 zone->watermark[WMARK_MIN] = min_pages;
6951 * If it's a lowmem zone, reserve a number of pages
6952 * proportionate to the zone's size.
6954 zone->watermark[WMARK_MIN] = tmp;
6958 * Set the kswapd watermarks distance according to the
6959 * scale factor in proportion to available memory, but
6960 * ensure a minimum size on small systems.
6962 tmp = max_t(u64, tmp >> 2,
6963 mult_frac(zone->managed_pages,
6964 watermark_scale_factor, 10000));
6966 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6967 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6969 spin_unlock_irqrestore(&zone->lock, flags);
6972 /* update totalreserve_pages */
6973 calculate_totalreserve_pages();
6977 * setup_per_zone_wmarks - called when min_free_kbytes changes
6978 * or when memory is hot-{added|removed}
6980 * Ensures that the watermark[min,low,high] values for each zone are set
6981 * correctly with respect to min_free_kbytes.
6983 void setup_per_zone_wmarks(void)
6985 static DEFINE_SPINLOCK(lock);
6988 __setup_per_zone_wmarks();
6993 * Initialise min_free_kbytes.
6995 * For small machines we want it small (128k min). For large machines
6996 * we want it large (64MB max). But it is not linear, because network
6997 * bandwidth does not increase linearly with machine size. We use
6999 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7000 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7016 int __meminit init_per_zone_wmark_min(void)
7018 unsigned long lowmem_kbytes;
7019 int new_min_free_kbytes;
7021 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7022 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7024 if (new_min_free_kbytes > user_min_free_kbytes) {
7025 min_free_kbytes = new_min_free_kbytes;
7026 if (min_free_kbytes < 128)
7027 min_free_kbytes = 128;
7028 if (min_free_kbytes > 65536)
7029 min_free_kbytes = 65536;
7031 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7032 new_min_free_kbytes, user_min_free_kbytes);
7034 setup_per_zone_wmarks();
7035 refresh_zone_stat_thresholds();
7036 setup_per_zone_lowmem_reserve();
7039 setup_min_unmapped_ratio();
7040 setup_min_slab_ratio();
7045 core_initcall(init_per_zone_wmark_min)
7048 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7049 * that we can call two helper functions whenever min_free_kbytes
7052 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7053 void __user *buffer, size_t *length, loff_t *ppos)
7057 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7062 user_min_free_kbytes = min_free_kbytes;
7063 setup_per_zone_wmarks();
7068 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7069 void __user *buffer, size_t *length, loff_t *ppos)
7073 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7078 setup_per_zone_wmarks();
7084 static void setup_min_unmapped_ratio(void)
7089 for_each_online_pgdat(pgdat)
7090 pgdat->min_unmapped_pages = 0;
7093 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7094 sysctl_min_unmapped_ratio) / 100;
7098 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7099 void __user *buffer, size_t *length, loff_t *ppos)
7103 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7107 setup_min_unmapped_ratio();
7112 static void setup_min_slab_ratio(void)
7117 for_each_online_pgdat(pgdat)
7118 pgdat->min_slab_pages = 0;
7121 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7122 sysctl_min_slab_ratio) / 100;
7125 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7126 void __user *buffer, size_t *length, loff_t *ppos)
7130 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7134 setup_min_slab_ratio();
7141 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7142 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7143 * whenever sysctl_lowmem_reserve_ratio changes.
7145 * The reserve ratio obviously has absolutely no relation with the
7146 * minimum watermarks. The lowmem reserve ratio can only make sense
7147 * if in function of the boot time zone sizes.
7149 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7150 void __user *buffer, size_t *length, loff_t *ppos)
7152 proc_dointvec_minmax(table, write, buffer, length, ppos);
7153 setup_per_zone_lowmem_reserve();
7158 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7159 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7160 * pagelist can have before it gets flushed back to buddy allocator.
7162 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7163 void __user *buffer, size_t *length, loff_t *ppos)
7166 int old_percpu_pagelist_fraction;
7169 mutex_lock(&pcp_batch_high_lock);
7170 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7172 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7173 if (!write || ret < 0)
7176 /* Sanity checking to avoid pcp imbalance */
7177 if (percpu_pagelist_fraction &&
7178 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7179 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7185 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7188 for_each_populated_zone(zone) {
7191 for_each_possible_cpu(cpu)
7192 pageset_set_high_and_batch(zone,
7193 per_cpu_ptr(zone->pageset, cpu));
7196 mutex_unlock(&pcp_batch_high_lock);
7201 int hashdist = HASHDIST_DEFAULT;
7203 static int __init set_hashdist(char *str)
7207 hashdist = simple_strtoul(str, &str, 0);
7210 __setup("hashdist=", set_hashdist);
7213 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7215 * Returns the number of pages that arch has reserved but
7216 * is not known to alloc_large_system_hash().
7218 static unsigned long __init arch_reserved_kernel_pages(void)
7225 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7226 * machines. As memory size is increased the scale is also increased but at
7227 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7228 * quadruples the scale is increased by one, which means the size of hash table
7229 * only doubles, instead of quadrupling as well.
7230 * Because 32-bit systems cannot have large physical memory, where this scaling
7231 * makes sense, it is disabled on such platforms.
7233 #if __BITS_PER_LONG > 32
7234 #define ADAPT_SCALE_BASE (64ul << 30)
7235 #define ADAPT_SCALE_SHIFT 2
7236 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7240 * allocate a large system hash table from bootmem
7241 * - it is assumed that the hash table must contain an exact power-of-2
7242 * quantity of entries
7243 * - limit is the number of hash buckets, not the total allocation size
7245 void *__init alloc_large_system_hash(const char *tablename,
7246 unsigned long bucketsize,
7247 unsigned long numentries,
7250 unsigned int *_hash_shift,
7251 unsigned int *_hash_mask,
7252 unsigned long low_limit,
7253 unsigned long high_limit)
7255 unsigned long long max = high_limit;
7256 unsigned long log2qty, size;
7260 /* allow the kernel cmdline to have a say */
7262 /* round applicable memory size up to nearest megabyte */
7263 numentries = nr_kernel_pages;
7264 numentries -= arch_reserved_kernel_pages();
7266 /* It isn't necessary when PAGE_SIZE >= 1MB */
7267 if (PAGE_SHIFT < 20)
7268 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7270 #if __BITS_PER_LONG > 32
7272 unsigned long adapt;
7274 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7275 adapt <<= ADAPT_SCALE_SHIFT)
7280 /* limit to 1 bucket per 2^scale bytes of low memory */
7281 if (scale > PAGE_SHIFT)
7282 numentries >>= (scale - PAGE_SHIFT);
7284 numentries <<= (PAGE_SHIFT - scale);
7286 /* Make sure we've got at least a 0-order allocation.. */
7287 if (unlikely(flags & HASH_SMALL)) {
7288 /* Makes no sense without HASH_EARLY */
7289 WARN_ON(!(flags & HASH_EARLY));
7290 if (!(numentries >> *_hash_shift)) {
7291 numentries = 1UL << *_hash_shift;
7292 BUG_ON(!numentries);
7294 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7295 numentries = PAGE_SIZE / bucketsize;
7297 numentries = roundup_pow_of_two(numentries);
7299 /* limit allocation size to 1/16 total memory by default */
7301 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7302 do_div(max, bucketsize);
7304 max = min(max, 0x80000000ULL);
7306 if (numentries < low_limit)
7307 numentries = low_limit;
7308 if (numentries > max)
7311 log2qty = ilog2(numentries);
7314 * memblock allocator returns zeroed memory already, so HASH_ZERO is
7315 * currently not used when HASH_EARLY is specified.
7317 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7319 size = bucketsize << log2qty;
7320 if (flags & HASH_EARLY)
7321 table = memblock_virt_alloc_nopanic(size, 0);
7323 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7326 * If bucketsize is not a power-of-two, we may free
7327 * some pages at the end of hash table which
7328 * alloc_pages_exact() automatically does
7330 if (get_order(size) < MAX_ORDER) {
7331 table = alloc_pages_exact(size, gfp_flags);
7332 kmemleak_alloc(table, size, 1, gfp_flags);
7335 } while (!table && size > PAGE_SIZE && --log2qty);
7338 panic("Failed to allocate %s hash table\n", tablename);
7340 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7341 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7344 *_hash_shift = log2qty;
7346 *_hash_mask = (1 << log2qty) - 1;
7352 * This function checks whether pageblock includes unmovable pages or not.
7353 * If @count is not zero, it is okay to include less @count unmovable pages
7355 * PageLRU check without isolation or lru_lock could race so that
7356 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7357 * check without lock_page also may miss some movable non-lru pages at
7358 * race condition. So you can't expect this function should be exact.
7360 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7361 bool skip_hwpoisoned_pages)
7363 unsigned long pfn, iter, found;
7367 * For avoiding noise data, lru_add_drain_all() should be called
7368 * If ZONE_MOVABLE, the zone never contains unmovable pages
7370 if (zone_idx(zone) == ZONE_MOVABLE)
7372 mt = get_pageblock_migratetype(page);
7373 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7376 pfn = page_to_pfn(page);
7377 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7378 unsigned long check = pfn + iter;
7380 if (!pfn_valid_within(check))
7383 page = pfn_to_page(check);
7386 * Hugepages are not in LRU lists, but they're movable.
7387 * We need not scan over tail pages bacause we don't
7388 * handle each tail page individually in migration.
7390 if (PageHuge(page)) {
7391 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7396 * We can't use page_count without pin a page
7397 * because another CPU can free compound page.
7398 * This check already skips compound tails of THP
7399 * because their page->_refcount is zero at all time.
7401 if (!page_ref_count(page)) {
7402 if (PageBuddy(page))
7403 iter += (1 << page_order(page)) - 1;
7408 * The HWPoisoned page may be not in buddy system, and
7409 * page_count() is not 0.
7411 if (skip_hwpoisoned_pages && PageHWPoison(page))
7414 if (__PageMovable(page))
7420 * If there are RECLAIMABLE pages, we need to check
7421 * it. But now, memory offline itself doesn't call
7422 * shrink_node_slabs() and it still to be fixed.
7425 * If the page is not RAM, page_count()should be 0.
7426 * we don't need more check. This is an _used_ not-movable page.
7428 * The problematic thing here is PG_reserved pages. PG_reserved
7429 * is set to both of a memory hole page and a _used_ kernel
7438 bool is_pageblock_removable_nolock(struct page *page)
7444 * We have to be careful here because we are iterating over memory
7445 * sections which are not zone aware so we might end up outside of
7446 * the zone but still within the section.
7447 * We have to take care about the node as well. If the node is offline
7448 * its NODE_DATA will be NULL - see page_zone.
7450 if (!node_online(page_to_nid(page)))
7453 zone = page_zone(page);
7454 pfn = page_to_pfn(page);
7455 if (!zone_spans_pfn(zone, pfn))
7458 return !has_unmovable_pages(zone, page, 0, true);
7461 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7463 static unsigned long pfn_max_align_down(unsigned long pfn)
7465 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7466 pageblock_nr_pages) - 1);
7469 static unsigned long pfn_max_align_up(unsigned long pfn)
7471 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7472 pageblock_nr_pages));
7475 /* [start, end) must belong to a single zone. */
7476 static int __alloc_contig_migrate_range(struct compact_control *cc,
7477 unsigned long start, unsigned long end)
7479 /* This function is based on compact_zone() from compaction.c. */
7480 unsigned long nr_reclaimed;
7481 unsigned long pfn = start;
7482 unsigned int tries = 0;
7487 while (pfn < end || !list_empty(&cc->migratepages)) {
7488 if (fatal_signal_pending(current)) {
7493 if (list_empty(&cc->migratepages)) {
7494 cc->nr_migratepages = 0;
7495 pfn = isolate_migratepages_range(cc, pfn, end);
7501 } else if (++tries == 5) {
7502 ret = ret < 0 ? ret : -EBUSY;
7506 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7508 cc->nr_migratepages -= nr_reclaimed;
7510 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7511 NULL, 0, cc->mode, MR_CMA);
7514 putback_movable_pages(&cc->migratepages);
7521 * alloc_contig_range() -- tries to allocate given range of pages
7522 * @start: start PFN to allocate
7523 * @end: one-past-the-last PFN to allocate
7524 * @migratetype: migratetype of the underlaying pageblocks (either
7525 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7526 * in range must have the same migratetype and it must
7527 * be either of the two.
7528 * @gfp_mask: GFP mask to use during compaction
7530 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7531 * aligned, however it's the caller's responsibility to guarantee that
7532 * we are the only thread that changes migrate type of pageblocks the
7535 * The PFN range must belong to a single zone.
7537 * Returns zero on success or negative error code. On success all
7538 * pages which PFN is in [start, end) are allocated for the caller and
7539 * need to be freed with free_contig_range().
7541 int alloc_contig_range(unsigned long start, unsigned long end,
7542 unsigned migratetype, gfp_t gfp_mask)
7544 unsigned long outer_start, outer_end;
7548 struct compact_control cc = {
7549 .nr_migratepages = 0,
7551 .zone = page_zone(pfn_to_page(start)),
7552 .mode = MIGRATE_SYNC,
7553 .ignore_skip_hint = true,
7554 .gfp_mask = current_gfp_context(gfp_mask),
7556 INIT_LIST_HEAD(&cc.migratepages);
7559 * What we do here is we mark all pageblocks in range as
7560 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7561 * have different sizes, and due to the way page allocator
7562 * work, we align the range to biggest of the two pages so
7563 * that page allocator won't try to merge buddies from
7564 * different pageblocks and change MIGRATE_ISOLATE to some
7565 * other migration type.
7567 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7568 * migrate the pages from an unaligned range (ie. pages that
7569 * we are interested in). This will put all the pages in
7570 * range back to page allocator as MIGRATE_ISOLATE.
7572 * When this is done, we take the pages in range from page
7573 * allocator removing them from the buddy system. This way
7574 * page allocator will never consider using them.
7576 * This lets us mark the pageblocks back as
7577 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7578 * aligned range but not in the unaligned, original range are
7579 * put back to page allocator so that buddy can use them.
7582 ret = start_isolate_page_range(pfn_max_align_down(start),
7583 pfn_max_align_up(end), migratetype,
7589 * In case of -EBUSY, we'd like to know which page causes problem.
7590 * So, just fall through. test_pages_isolated() has a tracepoint
7591 * which will report the busy page.
7593 * It is possible that busy pages could become available before
7594 * the call to test_pages_isolated, and the range will actually be
7595 * allocated. So, if we fall through be sure to clear ret so that
7596 * -EBUSY is not accidentally used or returned to caller.
7598 ret = __alloc_contig_migrate_range(&cc, start, end);
7599 if (ret && ret != -EBUSY)
7604 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7605 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7606 * more, all pages in [start, end) are free in page allocator.
7607 * What we are going to do is to allocate all pages from
7608 * [start, end) (that is remove them from page allocator).
7610 * The only problem is that pages at the beginning and at the
7611 * end of interesting range may be not aligned with pages that
7612 * page allocator holds, ie. they can be part of higher order
7613 * pages. Because of this, we reserve the bigger range and
7614 * once this is done free the pages we are not interested in.
7616 * We don't have to hold zone->lock here because the pages are
7617 * isolated thus they won't get removed from buddy.
7620 lru_add_drain_all();
7621 drain_all_pages(cc.zone);
7624 outer_start = start;
7625 while (!PageBuddy(pfn_to_page(outer_start))) {
7626 if (++order >= MAX_ORDER) {
7627 outer_start = start;
7630 outer_start &= ~0UL << order;
7633 if (outer_start != start) {
7634 order = page_order(pfn_to_page(outer_start));
7637 * outer_start page could be small order buddy page and
7638 * it doesn't include start page. Adjust outer_start
7639 * in this case to report failed page properly
7640 * on tracepoint in test_pages_isolated()
7642 if (outer_start + (1UL << order) <= start)
7643 outer_start = start;
7646 /* Make sure the range is really isolated. */
7647 if (test_pages_isolated(outer_start, end, false)) {
7648 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7649 __func__, outer_start, end);
7654 /* Grab isolated pages from freelists. */
7655 outer_end = isolate_freepages_range(&cc, outer_start, end);
7661 /* Free head and tail (if any) */
7662 if (start != outer_start)
7663 free_contig_range(outer_start, start - outer_start);
7664 if (end != outer_end)
7665 free_contig_range(end, outer_end - end);
7668 undo_isolate_page_range(pfn_max_align_down(start),
7669 pfn_max_align_up(end), migratetype);
7673 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7675 unsigned int count = 0;
7677 for (; nr_pages--; pfn++) {
7678 struct page *page = pfn_to_page(pfn);
7680 count += page_count(page) != 1;
7683 WARN(count != 0, "%d pages are still in use!\n", count);
7687 #ifdef CONFIG_MEMORY_HOTPLUG
7689 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7690 * page high values need to be recalulated.
7692 void __meminit zone_pcp_update(struct zone *zone)
7695 mutex_lock(&pcp_batch_high_lock);
7696 for_each_possible_cpu(cpu)
7697 pageset_set_high_and_batch(zone,
7698 per_cpu_ptr(zone->pageset, cpu));
7699 mutex_unlock(&pcp_batch_high_lock);
7703 void zone_pcp_reset(struct zone *zone)
7705 unsigned long flags;
7707 struct per_cpu_pageset *pset;
7709 /* avoid races with drain_pages() */
7710 local_irq_save(flags);
7711 if (zone->pageset != &boot_pageset) {
7712 for_each_online_cpu(cpu) {
7713 pset = per_cpu_ptr(zone->pageset, cpu);
7714 drain_zonestat(zone, pset);
7716 free_percpu(zone->pageset);
7717 zone->pageset = &boot_pageset;
7719 local_irq_restore(flags);
7722 #ifdef CONFIG_MEMORY_HOTREMOVE
7724 * All pages in the range must be in a single zone and isolated
7725 * before calling this.
7728 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7732 unsigned int order, i;
7734 unsigned long flags;
7735 /* find the first valid pfn */
7736 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7741 offline_mem_sections(pfn, end_pfn);
7742 zone = page_zone(pfn_to_page(pfn));
7743 spin_lock_irqsave(&zone->lock, flags);
7745 while (pfn < end_pfn) {
7746 if (!pfn_valid(pfn)) {
7750 page = pfn_to_page(pfn);
7752 * The HWPoisoned page may be not in buddy system, and
7753 * page_count() is not 0.
7755 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7757 SetPageReserved(page);
7761 BUG_ON(page_count(page));
7762 BUG_ON(!PageBuddy(page));
7763 order = page_order(page);
7764 #ifdef CONFIG_DEBUG_VM
7765 pr_info("remove from free list %lx %d %lx\n",
7766 pfn, 1 << order, end_pfn);
7768 list_del(&page->lru);
7769 rmv_page_order(page);
7770 zone->free_area[order].nr_free--;
7771 for (i = 0; i < (1 << order); i++)
7772 SetPageReserved((page+i));
7773 pfn += (1 << order);
7775 spin_unlock_irqrestore(&zone->lock, flags);
7779 bool is_free_buddy_page(struct page *page)
7781 struct zone *zone = page_zone(page);
7782 unsigned long pfn = page_to_pfn(page);
7783 unsigned long flags;
7786 spin_lock_irqsave(&zone->lock, flags);
7787 for (order = 0; order < MAX_ORDER; order++) {
7788 struct page *page_head = page - (pfn & ((1 << order) - 1));
7790 if (PageBuddy(page_head) && page_order(page_head) >= order)
7793 spin_unlock_irqrestore(&zone->lock, flags);
7795 return order < MAX_ORDER;