2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kasan.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/memremap.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <trace/events/oom.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/sched/mm.h>
64 #include <linux/page_owner.h>
65 #include <linux/kthread.h>
66 #include <linux/memcontrol.h>
67 #include <linux/ftrace.h>
68 #include <linux/lockdep.h>
69 #include <linux/nmi.h>
70 #include <linux/khugepaged.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 - 1, pageblock_order);
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) {
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 (order < MAX_ORDER - 1) {
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)))
876 max_order = order + 1;
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);
1027 * Check tail pages before head page information is cleared to
1028 * avoid checking PageCompound for order-0 pages.
1030 if (unlikely(order)) {
1031 bool compound = PageCompound(page);
1034 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1037 ClearPageDoubleMap(page);
1038 for (i = 1; i < (1 << order); i++) {
1040 bad += free_tail_pages_check(page, page + i);
1041 if (unlikely(free_pages_check(page + i))) {
1045 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1048 if (PageMappingFlags(page))
1049 page->mapping = NULL;
1050 if (memcg_kmem_enabled() && PageKmemcg(page))
1051 memcg_kmem_uncharge(page, order);
1053 bad += free_pages_check(page);
1057 page_cpupid_reset_last(page);
1058 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1059 reset_page_owner(page, order);
1061 if (!PageHighMem(page)) {
1062 debug_check_no_locks_freed(page_address(page),
1063 PAGE_SIZE << order);
1064 debug_check_no_obj_freed(page_address(page),
1065 PAGE_SIZE << order);
1067 arch_free_page(page, order);
1068 kernel_poison_pages(page, 1 << order, 0);
1069 kernel_map_pages(page, 1 << order, 0);
1070 kasan_free_pages(page, order);
1075 #ifdef CONFIG_DEBUG_VM
1076 static inline bool free_pcp_prepare(struct page *page)
1078 return free_pages_prepare(page, 0, true);
1081 static inline bool bulkfree_pcp_prepare(struct page *page)
1086 static bool free_pcp_prepare(struct page *page)
1088 return free_pages_prepare(page, 0, false);
1091 static bool bulkfree_pcp_prepare(struct page *page)
1093 return free_pages_check(page);
1095 #endif /* CONFIG_DEBUG_VM */
1098 * Frees a number of pages from the PCP lists
1099 * Assumes all pages on list are in same zone, and of same order.
1100 * count is the number of pages to free.
1102 * If the zone was previously in an "all pages pinned" state then look to
1103 * see if this freeing clears that state.
1105 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1106 * pinned" detection logic.
1108 static void free_pcppages_bulk(struct zone *zone, int count,
1109 struct per_cpu_pages *pcp)
1111 int migratetype = 0;
1113 bool isolated_pageblocks;
1115 spin_lock(&zone->lock);
1116 isolated_pageblocks = has_isolate_pageblock(zone);
1119 * Ensure proper count is passed which otherwise would stuck in the
1120 * below while (list_empty(list)) loop.
1122 count = min(pcp->count, count);
1125 struct list_head *list;
1128 * Remove pages from lists in a round-robin fashion. A
1129 * batch_free count is maintained that is incremented when an
1130 * empty list is encountered. This is so more pages are freed
1131 * off fuller lists instead of spinning excessively around empty
1136 if (++migratetype == MIGRATE_PCPTYPES)
1138 list = &pcp->lists[migratetype];
1139 } while (list_empty(list));
1141 /* This is the only non-empty list. Free them all. */
1142 if (batch_free == MIGRATE_PCPTYPES)
1146 int mt; /* migratetype of the to-be-freed page */
1148 page = list_last_entry(list, struct page, lru);
1149 /* must delete as __free_one_page list manipulates */
1150 list_del(&page->lru);
1152 mt = get_pcppage_migratetype(page);
1153 /* MIGRATE_ISOLATE page should not go to pcplists */
1154 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1155 /* Pageblock could have been isolated meanwhile */
1156 if (unlikely(isolated_pageblocks))
1157 mt = get_pageblock_migratetype(page);
1159 if (bulkfree_pcp_prepare(page))
1162 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1163 trace_mm_page_pcpu_drain(page, 0, mt);
1164 } while (--count && --batch_free && !list_empty(list));
1166 spin_unlock(&zone->lock);
1169 static void free_one_page(struct zone *zone,
1170 struct page *page, unsigned long pfn,
1174 spin_lock(&zone->lock);
1175 if (unlikely(has_isolate_pageblock(zone) ||
1176 is_migrate_isolate(migratetype))) {
1177 migratetype = get_pfnblock_migratetype(page, pfn);
1179 __free_one_page(page, pfn, zone, order, migratetype);
1180 spin_unlock(&zone->lock);
1183 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1184 unsigned long zone, int nid)
1186 set_page_links(page, zone, nid, pfn);
1187 init_page_count(page);
1188 page_mapcount_reset(page);
1189 page_cpupid_reset_last(page);
1191 INIT_LIST_HEAD(&page->lru);
1192 #ifdef WANT_PAGE_VIRTUAL
1193 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1194 if (!is_highmem_idx(zone))
1195 set_page_address(page, __va(pfn << PAGE_SHIFT));
1199 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1202 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1205 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1206 static void __meminit init_reserved_page(unsigned long pfn)
1211 if (!early_page_uninitialised(pfn))
1214 nid = early_pfn_to_nid(pfn);
1215 pgdat = NODE_DATA(nid);
1217 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1218 struct zone *zone = &pgdat->node_zones[zid];
1220 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1223 __init_single_pfn(pfn, zid, nid);
1226 static inline void init_reserved_page(unsigned long pfn)
1229 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1232 * Initialised pages do not have PageReserved set. This function is
1233 * called for each range allocated by the bootmem allocator and
1234 * marks the pages PageReserved. The remaining valid pages are later
1235 * sent to the buddy page allocator.
1237 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1239 unsigned long start_pfn = PFN_DOWN(start);
1240 unsigned long end_pfn = PFN_UP(end);
1242 for (; start_pfn < end_pfn; start_pfn++) {
1243 if (pfn_valid(start_pfn)) {
1244 struct page *page = pfn_to_page(start_pfn);
1246 init_reserved_page(start_pfn);
1248 /* Avoid false-positive PageTail() */
1249 INIT_LIST_HEAD(&page->lru);
1251 SetPageReserved(page);
1256 static void __free_pages_ok(struct page *page, unsigned int order)
1258 unsigned long flags;
1260 unsigned long pfn = page_to_pfn(page);
1262 if (!free_pages_prepare(page, order, true))
1265 migratetype = get_pfnblock_migratetype(page, pfn);
1266 local_irq_save(flags);
1267 __count_vm_events(PGFREE, 1 << order);
1268 free_one_page(page_zone(page), page, pfn, order, migratetype);
1269 local_irq_restore(flags);
1272 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1274 unsigned int nr_pages = 1 << order;
1275 struct page *p = page;
1279 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1281 __ClearPageReserved(p);
1282 set_page_count(p, 0);
1284 __ClearPageReserved(p);
1285 set_page_count(p, 0);
1287 page_zone(page)->managed_pages += nr_pages;
1288 set_page_refcounted(page);
1289 __free_pages(page, order);
1292 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1293 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1295 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1297 int __meminit early_pfn_to_nid(unsigned long pfn)
1299 static DEFINE_SPINLOCK(early_pfn_lock);
1302 spin_lock(&early_pfn_lock);
1303 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1305 nid = first_online_node;
1306 spin_unlock(&early_pfn_lock);
1312 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1313 static inline bool __meminit __maybe_unused
1314 meminit_pfn_in_nid(unsigned long pfn, int node,
1315 struct mminit_pfnnid_cache *state)
1319 nid = __early_pfn_to_nid(pfn, state);
1320 if (nid >= 0 && nid != node)
1325 /* Only safe to use early in boot when initialisation is single-threaded */
1326 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1328 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1333 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1337 static inline bool __meminit __maybe_unused
1338 meminit_pfn_in_nid(unsigned long pfn, int node,
1339 struct mminit_pfnnid_cache *state)
1346 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1349 if (early_page_uninitialised(pfn))
1351 return __free_pages_boot_core(page, order);
1355 * Check that the whole (or subset of) a pageblock given by the interval of
1356 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1357 * with the migration of free compaction scanner. The scanners then need to
1358 * use only pfn_valid_within() check for arches that allow holes within
1361 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1363 * It's possible on some configurations to have a setup like node0 node1 node0
1364 * i.e. it's possible that all pages within a zones range of pages do not
1365 * belong to a single zone. We assume that a border between node0 and node1
1366 * can occur within a single pageblock, but not a node0 node1 node0
1367 * interleaving within a single pageblock. It is therefore sufficient to check
1368 * the first and last page of a pageblock and avoid checking each individual
1369 * page in a pageblock.
1371 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1372 unsigned long end_pfn, struct zone *zone)
1374 struct page *start_page;
1375 struct page *end_page;
1377 /* end_pfn is one past the range we are checking */
1380 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1383 start_page = pfn_to_online_page(start_pfn);
1387 if (page_zone(start_page) != zone)
1390 end_page = pfn_to_page(end_pfn);
1392 /* This gives a shorter code than deriving page_zone(end_page) */
1393 if (page_zone_id(start_page) != page_zone_id(end_page))
1399 void set_zone_contiguous(struct zone *zone)
1401 unsigned long block_start_pfn = zone->zone_start_pfn;
1402 unsigned long block_end_pfn;
1404 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1405 for (; block_start_pfn < zone_end_pfn(zone);
1406 block_start_pfn = block_end_pfn,
1407 block_end_pfn += pageblock_nr_pages) {
1409 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1411 if (!__pageblock_pfn_to_page(block_start_pfn,
1412 block_end_pfn, zone))
1417 /* We confirm that there is no hole */
1418 zone->contiguous = true;
1421 void clear_zone_contiguous(struct zone *zone)
1423 zone->contiguous = false;
1426 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1427 static void __init deferred_free_range(struct page *page,
1428 unsigned long pfn, int nr_pages)
1435 /* Free a large naturally-aligned chunk if possible */
1436 if (nr_pages == pageblock_nr_pages &&
1437 (pfn & (pageblock_nr_pages - 1)) == 0) {
1438 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1439 __free_pages_boot_core(page, pageblock_order);
1443 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1444 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1445 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1446 __free_pages_boot_core(page, 0);
1450 /* Completion tracking for deferred_init_memmap() threads */
1451 static atomic_t pgdat_init_n_undone __initdata;
1452 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1454 static inline void __init pgdat_init_report_one_done(void)
1456 if (atomic_dec_and_test(&pgdat_init_n_undone))
1457 complete(&pgdat_init_all_done_comp);
1460 /* Initialise remaining memory on a node */
1461 static int __init deferred_init_memmap(void *data)
1463 pg_data_t *pgdat = data;
1464 int nid = pgdat->node_id;
1465 struct mminit_pfnnid_cache nid_init_state = { };
1466 unsigned long start = jiffies;
1467 unsigned long nr_pages = 0;
1468 unsigned long walk_start, walk_end;
1471 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1472 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1474 if (first_init_pfn == ULONG_MAX) {
1475 pgdat_init_report_one_done();
1479 /* Bind memory initialisation thread to a local node if possible */
1480 if (!cpumask_empty(cpumask))
1481 set_cpus_allowed_ptr(current, cpumask);
1483 /* Sanity check boundaries */
1484 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1485 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1486 pgdat->first_deferred_pfn = ULONG_MAX;
1488 /* Only the highest zone is deferred so find it */
1489 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1490 zone = pgdat->node_zones + zid;
1491 if (first_init_pfn < zone_end_pfn(zone))
1495 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1496 unsigned long pfn, end_pfn;
1497 struct page *page = NULL;
1498 struct page *free_base_page = NULL;
1499 unsigned long free_base_pfn = 0;
1502 end_pfn = min(walk_end, zone_end_pfn(zone));
1503 pfn = first_init_pfn;
1504 if (pfn < walk_start)
1506 if (pfn < zone->zone_start_pfn)
1507 pfn = zone->zone_start_pfn;
1509 for (; pfn < end_pfn; pfn++) {
1510 if (!pfn_valid_within(pfn))
1514 * Ensure pfn_valid is checked every
1515 * pageblock_nr_pages for memory holes
1517 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1518 if (!pfn_valid(pfn)) {
1524 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1529 /* Minimise pfn page lookups and scheduler checks */
1530 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1533 nr_pages += nr_to_free;
1534 deferred_free_range(free_base_page,
1535 free_base_pfn, nr_to_free);
1536 free_base_page = NULL;
1537 free_base_pfn = nr_to_free = 0;
1539 page = pfn_to_page(pfn);
1544 VM_BUG_ON(page_zone(page) != zone);
1548 __init_single_page(page, pfn, zid, nid);
1549 if (!free_base_page) {
1550 free_base_page = page;
1551 free_base_pfn = pfn;
1556 /* Where possible, batch up pages for a single free */
1559 /* Free the current block of pages to allocator */
1560 nr_pages += nr_to_free;
1561 deferred_free_range(free_base_page, free_base_pfn,
1563 free_base_page = NULL;
1564 free_base_pfn = nr_to_free = 0;
1566 /* Free the last block of pages to allocator */
1567 nr_pages += nr_to_free;
1568 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1570 first_init_pfn = max(end_pfn, first_init_pfn);
1573 /* Sanity check that the next zone really is unpopulated */
1574 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1576 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1577 jiffies_to_msecs(jiffies - start));
1579 pgdat_init_report_one_done();
1582 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1584 void __init page_alloc_init_late(void)
1588 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1591 /* There will be num_node_state(N_MEMORY) threads */
1592 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1593 for_each_node_state(nid, N_MEMORY) {
1594 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1597 /* Block until all are initialised */
1598 wait_for_completion(&pgdat_init_all_done_comp);
1600 /* Reinit limits that are based on free pages after the kernel is up */
1601 files_maxfiles_init();
1603 #ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1604 /* Discard memblock private memory */
1608 for_each_populated_zone(zone)
1609 set_zone_contiguous(zone);
1613 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1614 void __init init_cma_reserved_pageblock(struct page *page)
1616 unsigned i = pageblock_nr_pages;
1617 struct page *p = page;
1620 __ClearPageReserved(p);
1621 set_page_count(p, 0);
1624 set_pageblock_migratetype(page, MIGRATE_CMA);
1626 if (pageblock_order >= MAX_ORDER) {
1627 i = pageblock_nr_pages;
1630 set_page_refcounted(p);
1631 __free_pages(p, MAX_ORDER - 1);
1632 p += MAX_ORDER_NR_PAGES;
1633 } while (i -= MAX_ORDER_NR_PAGES);
1635 set_page_refcounted(page);
1636 __free_pages(page, pageblock_order);
1639 adjust_managed_page_count(page, pageblock_nr_pages);
1644 * The order of subdivision here is critical for the IO subsystem.
1645 * Please do not alter this order without good reasons and regression
1646 * testing. Specifically, as large blocks of memory are subdivided,
1647 * the order in which smaller blocks are delivered depends on the order
1648 * they're subdivided in this function. This is the primary factor
1649 * influencing the order in which pages are delivered to the IO
1650 * subsystem according to empirical testing, and this is also justified
1651 * by considering the behavior of a buddy system containing a single
1652 * large block of memory acted on by a series of small allocations.
1653 * This behavior is a critical factor in sglist merging's success.
1657 static inline void expand(struct zone *zone, struct page *page,
1658 int low, int high, struct free_area *area,
1661 unsigned long size = 1 << high;
1663 while (high > low) {
1667 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1670 * Mark as guard pages (or page), that will allow to
1671 * merge back to allocator when buddy will be freed.
1672 * Corresponding page table entries will not be touched,
1673 * pages will stay not present in virtual address space
1675 if (set_page_guard(zone, &page[size], high, migratetype))
1678 list_add(&page[size].lru, &area->free_list[migratetype]);
1680 set_page_order(&page[size], high);
1684 static void check_new_page_bad(struct page *page)
1686 const char *bad_reason = NULL;
1687 unsigned long bad_flags = 0;
1689 if (unlikely(atomic_read(&page->_mapcount) != -1))
1690 bad_reason = "nonzero mapcount";
1691 if (unlikely(page->mapping != NULL))
1692 bad_reason = "non-NULL mapping";
1693 if (unlikely(page_ref_count(page) != 0))
1694 bad_reason = "nonzero _count";
1695 if (unlikely(page->flags & __PG_HWPOISON)) {
1696 bad_reason = "HWPoisoned (hardware-corrupted)";
1697 bad_flags = __PG_HWPOISON;
1698 /* Don't complain about hwpoisoned pages */
1699 page_mapcount_reset(page); /* remove PageBuddy */
1702 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1703 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1704 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1707 if (unlikely(page->mem_cgroup))
1708 bad_reason = "page still charged to cgroup";
1710 bad_page(page, bad_reason, bad_flags);
1714 * This page is about to be returned from the page allocator
1716 static inline int check_new_page(struct page *page)
1718 if (likely(page_expected_state(page,
1719 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1722 check_new_page_bad(page);
1726 static inline bool free_pages_prezeroed(void)
1728 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1729 page_poisoning_enabled();
1732 #ifdef CONFIG_DEBUG_VM
1733 static bool check_pcp_refill(struct page *page)
1738 static bool check_new_pcp(struct page *page)
1740 return check_new_page(page);
1743 static bool check_pcp_refill(struct page *page)
1745 return check_new_page(page);
1747 static bool check_new_pcp(struct page *page)
1751 #endif /* CONFIG_DEBUG_VM */
1753 static bool check_new_pages(struct page *page, unsigned int order)
1756 for (i = 0; i < (1 << order); i++) {
1757 struct page *p = page + i;
1759 if (unlikely(check_new_page(p)))
1766 inline void post_alloc_hook(struct page *page, unsigned int order,
1769 set_page_private(page, 0);
1770 set_page_refcounted(page);
1772 arch_alloc_page(page, order);
1773 kernel_map_pages(page, 1 << order, 1);
1774 kasan_alloc_pages(page, order);
1775 kernel_poison_pages(page, 1 << order, 1);
1776 set_page_owner(page, order, gfp_flags);
1779 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1780 unsigned int alloc_flags)
1784 post_alloc_hook(page, order, gfp_flags);
1786 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1787 for (i = 0; i < (1 << order); i++)
1788 clear_highpage(page + i);
1790 if (order && (gfp_flags & __GFP_COMP))
1791 prep_compound_page(page, order);
1794 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1795 * allocate the page. The expectation is that the caller is taking
1796 * steps that will free more memory. The caller should avoid the page
1797 * being used for !PFMEMALLOC purposes.
1799 if (alloc_flags & ALLOC_NO_WATERMARKS)
1800 set_page_pfmemalloc(page);
1802 clear_page_pfmemalloc(page);
1806 * Go through the free lists for the given migratetype and remove
1807 * the smallest available page from the freelists
1810 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1813 unsigned int current_order;
1814 struct free_area *area;
1817 /* Find a page of the appropriate size in the preferred list */
1818 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1819 area = &(zone->free_area[current_order]);
1820 page = list_first_entry_or_null(&area->free_list[migratetype],
1824 list_del(&page->lru);
1825 rmv_page_order(page);
1827 expand(zone, page, order, current_order, area, migratetype);
1828 set_pcppage_migratetype(page, migratetype);
1837 * This array describes the order lists are fallen back to when
1838 * the free lists for the desirable migrate type are depleted
1840 static int fallbacks[MIGRATE_TYPES][4] = {
1841 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1842 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1843 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1845 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1847 #ifdef CONFIG_MEMORY_ISOLATION
1848 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1853 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1856 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1859 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1860 unsigned int order) { return NULL; }
1864 * Move the free pages in a range to the free lists of the requested type.
1865 * Note that start_page and end_pages are not aligned on a pageblock
1866 * boundary. If alignment is required, use move_freepages_block()
1868 static int move_freepages(struct zone *zone,
1869 struct page *start_page, struct page *end_page,
1870 int migratetype, int *num_movable)
1874 int pages_moved = 0;
1876 #ifndef CONFIG_HOLES_IN_ZONE
1878 * page_zone is not safe to call in this context when
1879 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1880 * anyway as we check zone boundaries in move_freepages_block().
1881 * Remove at a later date when no bug reports exist related to
1882 * grouping pages by mobility
1884 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1890 for (page = start_page; page <= end_page;) {
1891 if (!pfn_valid_within(page_to_pfn(page))) {
1896 /* Make sure we are not inadvertently changing nodes */
1897 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1899 if (!PageBuddy(page)) {
1901 * We assume that pages that could be isolated for
1902 * migration are movable. But we don't actually try
1903 * isolating, as that would be expensive.
1906 (PageLRU(page) || __PageMovable(page)))
1913 order = page_order(page);
1914 list_move(&page->lru,
1915 &zone->free_area[order].free_list[migratetype]);
1917 pages_moved += 1 << order;
1923 int move_freepages_block(struct zone *zone, struct page *page,
1924 int migratetype, int *num_movable)
1926 unsigned long start_pfn, end_pfn;
1927 struct page *start_page, *end_page;
1929 start_pfn = page_to_pfn(page);
1930 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1931 start_page = pfn_to_page(start_pfn);
1932 end_page = start_page + pageblock_nr_pages - 1;
1933 end_pfn = start_pfn + pageblock_nr_pages - 1;
1935 /* Do not cross zone boundaries */
1936 if (!zone_spans_pfn(zone, start_pfn))
1938 if (!zone_spans_pfn(zone, end_pfn))
1941 return move_freepages(zone, start_page, end_page, migratetype,
1945 static void change_pageblock_range(struct page *pageblock_page,
1946 int start_order, int migratetype)
1948 int nr_pageblocks = 1 << (start_order - pageblock_order);
1950 while (nr_pageblocks--) {
1951 set_pageblock_migratetype(pageblock_page, migratetype);
1952 pageblock_page += pageblock_nr_pages;
1957 * When we are falling back to another migratetype during allocation, try to
1958 * steal extra free pages from the same pageblocks to satisfy further
1959 * allocations, instead of polluting multiple pageblocks.
1961 * If we are stealing a relatively large buddy page, it is likely there will
1962 * be more free pages in the pageblock, so try to steal them all. For
1963 * reclaimable and unmovable allocations, we steal regardless of page size,
1964 * as fragmentation caused by those allocations polluting movable pageblocks
1965 * is worse than movable allocations stealing from unmovable and reclaimable
1968 static bool can_steal_fallback(unsigned int order, int start_mt)
1971 * Leaving this order check is intended, although there is
1972 * relaxed order check in next check. The reason is that
1973 * we can actually steal whole pageblock if this condition met,
1974 * but, below check doesn't guarantee it and that is just heuristic
1975 * so could be changed anytime.
1977 if (order >= pageblock_order)
1980 if (order >= pageblock_order / 2 ||
1981 start_mt == MIGRATE_RECLAIMABLE ||
1982 start_mt == MIGRATE_UNMOVABLE ||
1983 page_group_by_mobility_disabled)
1990 * This function implements actual steal behaviour. If order is large enough,
1991 * we can steal whole pageblock. If not, we first move freepages in this
1992 * pageblock to our migratetype and determine how many already-allocated pages
1993 * are there in the pageblock with a compatible migratetype. If at least half
1994 * of pages are free or compatible, we can change migratetype of the pageblock
1995 * itself, so pages freed in the future will be put on the correct free list.
1997 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1998 int start_type, bool whole_block)
2000 unsigned int current_order = page_order(page);
2001 struct free_area *area;
2002 int free_pages, movable_pages, alike_pages;
2005 old_block_type = get_pageblock_migratetype(page);
2008 * This can happen due to races and we want to prevent broken
2009 * highatomic accounting.
2011 if (is_migrate_highatomic(old_block_type))
2014 /* Take ownership for orders >= pageblock_order */
2015 if (current_order >= pageblock_order) {
2016 change_pageblock_range(page, current_order, start_type);
2020 /* We are not allowed to try stealing from the whole block */
2024 free_pages = move_freepages_block(zone, page, start_type,
2027 * Determine how many pages are compatible with our allocation.
2028 * For movable allocation, it's the number of movable pages which
2029 * we just obtained. For other types it's a bit more tricky.
2031 if (start_type == MIGRATE_MOVABLE) {
2032 alike_pages = movable_pages;
2035 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2036 * to MOVABLE pageblock, consider all non-movable pages as
2037 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2038 * vice versa, be conservative since we can't distinguish the
2039 * exact migratetype of non-movable pages.
2041 if (old_block_type == MIGRATE_MOVABLE)
2042 alike_pages = pageblock_nr_pages
2043 - (free_pages + movable_pages);
2048 /* moving whole block can fail due to zone boundary conditions */
2053 * If a sufficient number of pages in the block are either free or of
2054 * comparable migratability as our allocation, claim the whole block.
2056 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2057 page_group_by_mobility_disabled)
2058 set_pageblock_migratetype(page, start_type);
2063 area = &zone->free_area[current_order];
2064 list_move(&page->lru, &area->free_list[start_type]);
2068 * Check whether there is a suitable fallback freepage with requested order.
2069 * If only_stealable is true, this function returns fallback_mt only if
2070 * we can steal other freepages all together. This would help to reduce
2071 * fragmentation due to mixed migratetype pages in one pageblock.
2073 int find_suitable_fallback(struct free_area *area, unsigned int order,
2074 int migratetype, bool only_stealable, bool *can_steal)
2079 if (area->nr_free == 0)
2084 fallback_mt = fallbacks[migratetype][i];
2085 if (fallback_mt == MIGRATE_TYPES)
2088 if (list_empty(&area->free_list[fallback_mt]))
2091 if (can_steal_fallback(order, migratetype))
2094 if (!only_stealable)
2105 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2106 * there are no empty page blocks that contain a page with a suitable order
2108 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2109 unsigned int alloc_order)
2112 unsigned long max_managed, flags;
2115 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2116 * Check is race-prone but harmless.
2118 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2119 if (zone->nr_reserved_highatomic >= max_managed)
2122 spin_lock_irqsave(&zone->lock, flags);
2124 /* Recheck the nr_reserved_highatomic limit under the lock */
2125 if (zone->nr_reserved_highatomic >= max_managed)
2129 mt = get_pageblock_migratetype(page);
2130 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2131 && !is_migrate_cma(mt)) {
2132 zone->nr_reserved_highatomic += pageblock_nr_pages;
2133 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2134 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2138 spin_unlock_irqrestore(&zone->lock, flags);
2142 * Used when an allocation is about to fail under memory pressure. This
2143 * potentially hurts the reliability of high-order allocations when under
2144 * intense memory pressure but failed atomic allocations should be easier
2145 * to recover from than an OOM.
2147 * If @force is true, try to unreserve a pageblock even though highatomic
2148 * pageblock is exhausted.
2150 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2153 struct zonelist *zonelist = ac->zonelist;
2154 unsigned long flags;
2161 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2164 * Preserve at least one pageblock unless memory pressure
2167 if (!force && zone->nr_reserved_highatomic <=
2171 spin_lock_irqsave(&zone->lock, flags);
2172 for (order = 0; order < MAX_ORDER; order++) {
2173 struct free_area *area = &(zone->free_area[order]);
2175 page = list_first_entry_or_null(
2176 &area->free_list[MIGRATE_HIGHATOMIC],
2182 * In page freeing path, migratetype change is racy so
2183 * we can counter several free pages in a pageblock
2184 * in this loop althoug we changed the pageblock type
2185 * from highatomic to ac->migratetype. So we should
2186 * adjust the count once.
2188 if (is_migrate_highatomic_page(page)) {
2190 * It should never happen but changes to
2191 * locking could inadvertently allow a per-cpu
2192 * drain to add pages to MIGRATE_HIGHATOMIC
2193 * while unreserving so be safe and watch for
2196 zone->nr_reserved_highatomic -= min(
2198 zone->nr_reserved_highatomic);
2202 * Convert to ac->migratetype and avoid the normal
2203 * pageblock stealing heuristics. Minimally, the caller
2204 * is doing the work and needs the pages. More
2205 * importantly, if the block was always converted to
2206 * MIGRATE_UNMOVABLE or another type then the number
2207 * of pageblocks that cannot be completely freed
2210 set_pageblock_migratetype(page, ac->migratetype);
2211 ret = move_freepages_block(zone, page, ac->migratetype,
2214 spin_unlock_irqrestore(&zone->lock, flags);
2218 spin_unlock_irqrestore(&zone->lock, flags);
2225 * Try finding a free buddy page on the fallback list and put it on the free
2226 * list of requested migratetype, possibly along with other pages from the same
2227 * block, depending on fragmentation avoidance heuristics. Returns true if
2228 * fallback was found so that __rmqueue_smallest() can grab it.
2230 * The use of signed ints for order and current_order is a deliberate
2231 * deviation from the rest of this file, to make the for loop
2232 * condition simpler.
2235 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2237 struct free_area *area;
2244 * Find the largest available free page in the other list. This roughly
2245 * approximates finding the pageblock with the most free pages, which
2246 * would be too costly to do exactly.
2248 for (current_order = MAX_ORDER - 1; current_order >= order;
2250 area = &(zone->free_area[current_order]);
2251 fallback_mt = find_suitable_fallback(area, current_order,
2252 start_migratetype, false, &can_steal);
2253 if (fallback_mt == -1)
2257 * We cannot steal all free pages from the pageblock and the
2258 * requested migratetype is movable. In that case it's better to
2259 * steal and split the smallest available page instead of the
2260 * largest available page, because even if the next movable
2261 * allocation falls back into a different pageblock than this
2262 * one, it won't cause permanent fragmentation.
2264 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2265 && current_order > order)
2274 for (current_order = order; current_order < MAX_ORDER;
2276 area = &(zone->free_area[current_order]);
2277 fallback_mt = find_suitable_fallback(area, current_order,
2278 start_migratetype, false, &can_steal);
2279 if (fallback_mt != -1)
2284 * This should not happen - we already found a suitable fallback
2285 * when looking for the largest page.
2287 VM_BUG_ON(current_order == MAX_ORDER);
2290 page = list_first_entry(&area->free_list[fallback_mt],
2293 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2295 trace_mm_page_alloc_extfrag(page, order, current_order,
2296 start_migratetype, fallback_mt);
2303 * Do the hard work of removing an element from the buddy allocator.
2304 * Call me with the zone->lock already held.
2306 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2312 page = __rmqueue_smallest(zone, order, migratetype);
2313 if (unlikely(!page)) {
2314 if (migratetype == MIGRATE_MOVABLE)
2315 page = __rmqueue_cma_fallback(zone, order);
2317 if (!page && __rmqueue_fallback(zone, order, migratetype))
2321 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2326 * Obtain a specified number of elements from the buddy allocator, all under
2327 * a single hold of the lock, for efficiency. Add them to the supplied list.
2328 * Returns the number of new pages which were placed at *list.
2330 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2331 unsigned long count, struct list_head *list,
2332 int migratetype, bool cold)
2336 spin_lock(&zone->lock);
2337 for (i = 0; i < count; ++i) {
2338 struct page *page = __rmqueue(zone, order, migratetype);
2339 if (unlikely(page == NULL))
2342 if (unlikely(check_pcp_refill(page)))
2346 * Split buddy pages returned by expand() are received here
2347 * in physical page order. The page is added to the callers and
2348 * list and the list head then moves forward. From the callers
2349 * perspective, the linked list is ordered by page number in
2350 * some conditions. This is useful for IO devices that can
2351 * merge IO requests if the physical pages are ordered
2355 list_add(&page->lru, list);
2357 list_add_tail(&page->lru, list);
2360 if (is_migrate_cma(get_pcppage_migratetype(page)))
2361 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2366 * i pages were removed from the buddy list even if some leak due
2367 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2368 * on i. Do not confuse with 'alloced' which is the number of
2369 * pages added to the pcp list.
2371 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2372 spin_unlock(&zone->lock);
2378 * Called from the vmstat counter updater to drain pagesets of this
2379 * currently executing processor on remote nodes after they have
2382 * Note that this function must be called with the thread pinned to
2383 * a single processor.
2385 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2387 unsigned long flags;
2388 int to_drain, batch;
2390 local_irq_save(flags);
2391 batch = READ_ONCE(pcp->batch);
2392 to_drain = min(pcp->count, batch);
2394 free_pcppages_bulk(zone, to_drain, pcp);
2395 pcp->count -= to_drain;
2397 local_irq_restore(flags);
2402 * Drain pcplists of the indicated processor and zone.
2404 * The processor must either be the current processor and the
2405 * thread pinned to the current processor or a processor that
2408 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2410 unsigned long flags;
2411 struct per_cpu_pageset *pset;
2412 struct per_cpu_pages *pcp;
2414 local_irq_save(flags);
2415 pset = per_cpu_ptr(zone->pageset, cpu);
2419 free_pcppages_bulk(zone, pcp->count, pcp);
2422 local_irq_restore(flags);
2426 * Drain pcplists of all zones on the indicated processor.
2428 * The processor must either be the current processor and the
2429 * thread pinned to the current processor or a processor that
2432 static void drain_pages(unsigned int cpu)
2436 for_each_populated_zone(zone) {
2437 drain_pages_zone(cpu, zone);
2442 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2444 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2445 * the single zone's pages.
2447 void drain_local_pages(struct zone *zone)
2449 int cpu = smp_processor_id();
2452 drain_pages_zone(cpu, zone);
2457 static void drain_local_pages_wq(struct work_struct *work)
2460 * drain_all_pages doesn't use proper cpu hotplug protection so
2461 * we can race with cpu offline when the WQ can move this from
2462 * a cpu pinned worker to an unbound one. We can operate on a different
2463 * cpu which is allright but we also have to make sure to not move to
2467 drain_local_pages(NULL);
2472 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2474 * When zone parameter is non-NULL, spill just the single zone's pages.
2476 * Note that this can be extremely slow as the draining happens in a workqueue.
2478 void drain_all_pages(struct zone *zone)
2483 * Allocate in the BSS so we wont require allocation in
2484 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2486 static cpumask_t cpus_with_pcps;
2489 * Make sure nobody triggers this path before mm_percpu_wq is fully
2492 if (WARN_ON_ONCE(!mm_percpu_wq))
2496 * Do not drain if one is already in progress unless it's specific to
2497 * a zone. Such callers are primarily CMA and memory hotplug and need
2498 * the drain to be complete when the call returns.
2500 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2503 mutex_lock(&pcpu_drain_mutex);
2507 * We don't care about racing with CPU hotplug event
2508 * as offline notification will cause the notified
2509 * cpu to drain that CPU pcps and on_each_cpu_mask
2510 * disables preemption as part of its processing
2512 for_each_online_cpu(cpu) {
2513 struct per_cpu_pageset *pcp;
2515 bool has_pcps = false;
2518 pcp = per_cpu_ptr(zone->pageset, cpu);
2522 for_each_populated_zone(z) {
2523 pcp = per_cpu_ptr(z->pageset, cpu);
2524 if (pcp->pcp.count) {
2532 cpumask_set_cpu(cpu, &cpus_with_pcps);
2534 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2537 for_each_cpu(cpu, &cpus_with_pcps) {
2538 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2539 INIT_WORK(work, drain_local_pages_wq);
2540 queue_work_on(cpu, mm_percpu_wq, work);
2542 for_each_cpu(cpu, &cpus_with_pcps)
2543 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2545 mutex_unlock(&pcpu_drain_mutex);
2548 #ifdef CONFIG_HIBERNATION
2551 * Touch the watchdog for every WD_PAGE_COUNT pages.
2553 #define WD_PAGE_COUNT (128*1024)
2555 void mark_free_pages(struct zone *zone)
2557 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2558 unsigned long flags;
2559 unsigned int order, t;
2562 if (zone_is_empty(zone))
2565 spin_lock_irqsave(&zone->lock, flags);
2567 max_zone_pfn = zone_end_pfn(zone);
2568 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2569 if (pfn_valid(pfn)) {
2570 page = pfn_to_page(pfn);
2572 if (!--page_count) {
2573 touch_nmi_watchdog();
2574 page_count = WD_PAGE_COUNT;
2577 if (page_zone(page) != zone)
2580 if (!swsusp_page_is_forbidden(page))
2581 swsusp_unset_page_free(page);
2584 for_each_migratetype_order(order, t) {
2585 list_for_each_entry(page,
2586 &zone->free_area[order].free_list[t], lru) {
2589 pfn = page_to_pfn(page);
2590 for (i = 0; i < (1UL << order); i++) {
2591 if (!--page_count) {
2592 touch_nmi_watchdog();
2593 page_count = WD_PAGE_COUNT;
2595 swsusp_set_page_free(pfn_to_page(pfn + i));
2599 spin_unlock_irqrestore(&zone->lock, flags);
2601 #endif /* CONFIG_PM */
2604 * Free a 0-order page
2605 * cold == true ? free a cold page : free a hot page
2607 void free_hot_cold_page(struct page *page, bool cold)
2609 struct zone *zone = page_zone(page);
2610 struct per_cpu_pages *pcp;
2611 unsigned long flags;
2612 unsigned long pfn = page_to_pfn(page);
2615 if (!free_pcp_prepare(page))
2618 migratetype = get_pfnblock_migratetype(page, pfn);
2619 set_pcppage_migratetype(page, migratetype);
2620 local_irq_save(flags);
2621 __count_vm_event(PGFREE);
2624 * We only track unmovable, reclaimable and movable on pcp lists.
2625 * Free ISOLATE pages back to the allocator because they are being
2626 * offlined but treat HIGHATOMIC as movable pages so we can get those
2627 * areas back if necessary. Otherwise, we may have to free
2628 * excessively into the page allocator
2630 if (migratetype >= MIGRATE_PCPTYPES) {
2631 if (unlikely(is_migrate_isolate(migratetype))) {
2632 free_one_page(zone, page, pfn, 0, migratetype);
2635 migratetype = MIGRATE_MOVABLE;
2638 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2640 list_add(&page->lru, &pcp->lists[migratetype]);
2642 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2644 if (pcp->count >= pcp->high) {
2645 unsigned long batch = READ_ONCE(pcp->batch);
2646 free_pcppages_bulk(zone, batch, pcp);
2647 pcp->count -= batch;
2651 local_irq_restore(flags);
2655 * Free a list of 0-order pages
2657 void free_hot_cold_page_list(struct list_head *list, bool cold)
2659 struct page *page, *next;
2661 list_for_each_entry_safe(page, next, list, lru) {
2662 trace_mm_page_free_batched(page, cold);
2663 free_hot_cold_page(page, cold);
2668 * split_page takes a non-compound higher-order page, and splits it into
2669 * n (1<<order) sub-pages: page[0..n]
2670 * Each sub-page must be freed individually.
2672 * Note: this is probably too low level an operation for use in drivers.
2673 * Please consult with lkml before using this in your driver.
2675 void split_page(struct page *page, unsigned int order)
2679 VM_BUG_ON_PAGE(PageCompound(page), page);
2680 VM_BUG_ON_PAGE(!page_count(page), page);
2682 for (i = 1; i < (1 << order); i++)
2683 set_page_refcounted(page + i);
2684 split_page_owner(page, order);
2686 EXPORT_SYMBOL_GPL(split_page);
2688 int __isolate_free_page(struct page *page, unsigned int order)
2690 unsigned long watermark;
2694 BUG_ON(!PageBuddy(page));
2696 zone = page_zone(page);
2697 mt = get_pageblock_migratetype(page);
2699 if (!is_migrate_isolate(mt)) {
2701 * Obey watermarks as if the page was being allocated. We can
2702 * emulate a high-order watermark check with a raised order-0
2703 * watermark, because we already know our high-order page
2706 watermark = min_wmark_pages(zone) + (1UL << order);
2707 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2710 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2713 /* Remove page from free list */
2714 list_del(&page->lru);
2715 zone->free_area[order].nr_free--;
2716 rmv_page_order(page);
2719 * Set the pageblock if the isolated page is at least half of a
2722 if (order >= pageblock_order - 1) {
2723 struct page *endpage = page + (1 << order) - 1;
2724 for (; page < endpage; page += pageblock_nr_pages) {
2725 int mt = get_pageblock_migratetype(page);
2726 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2727 && !is_migrate_highatomic(mt))
2728 set_pageblock_migratetype(page,
2734 return 1UL << order;
2738 * Update NUMA hit/miss statistics
2740 * Must be called with interrupts disabled.
2742 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2745 enum numa_stat_item local_stat = NUMA_LOCAL;
2747 if (z->node != numa_node_id())
2748 local_stat = NUMA_OTHER;
2750 if (z->node == preferred_zone->node)
2751 __inc_numa_state(z, NUMA_HIT);
2753 __inc_numa_state(z, NUMA_MISS);
2754 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2756 __inc_numa_state(z, local_stat);
2760 /* Remove page from the per-cpu list, caller must protect the list */
2761 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2762 bool cold, struct per_cpu_pages *pcp,
2763 struct list_head *list)
2768 if (list_empty(list)) {
2769 pcp->count += rmqueue_bulk(zone, 0,
2772 if (unlikely(list_empty(list)))
2777 page = list_last_entry(list, struct page, lru);
2779 page = list_first_entry(list, struct page, lru);
2781 list_del(&page->lru);
2783 } while (check_new_pcp(page));
2788 /* Lock and remove page from the per-cpu list */
2789 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2790 struct zone *zone, unsigned int order,
2791 gfp_t gfp_flags, int migratetype)
2793 struct per_cpu_pages *pcp;
2794 struct list_head *list;
2795 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2797 unsigned long flags;
2799 local_irq_save(flags);
2800 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2801 list = &pcp->lists[migratetype];
2802 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2804 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2805 zone_statistics(preferred_zone, zone);
2807 local_irq_restore(flags);
2812 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2815 struct page *rmqueue(struct zone *preferred_zone,
2816 struct zone *zone, unsigned int order,
2817 gfp_t gfp_flags, unsigned int alloc_flags,
2820 unsigned long flags;
2823 if (likely(order == 0)) {
2824 page = rmqueue_pcplist(preferred_zone, zone, order,
2825 gfp_flags, migratetype);
2830 * We most definitely don't want callers attempting to
2831 * allocate greater than order-1 page units with __GFP_NOFAIL.
2833 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2834 spin_lock_irqsave(&zone->lock, flags);
2838 if (alloc_flags & ALLOC_HARDER) {
2839 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2841 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2844 page = __rmqueue(zone, order, migratetype);
2845 } while (page && check_new_pages(page, order));
2846 spin_unlock(&zone->lock);
2849 __mod_zone_freepage_state(zone, -(1 << order),
2850 get_pcppage_migratetype(page));
2852 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2853 zone_statistics(preferred_zone, zone);
2854 local_irq_restore(flags);
2857 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2861 local_irq_restore(flags);
2865 #ifdef CONFIG_FAIL_PAGE_ALLOC
2868 struct fault_attr attr;
2870 bool ignore_gfp_highmem;
2871 bool ignore_gfp_reclaim;
2873 } fail_page_alloc = {
2874 .attr = FAULT_ATTR_INITIALIZER,
2875 .ignore_gfp_reclaim = true,
2876 .ignore_gfp_highmem = true,
2880 static int __init setup_fail_page_alloc(char *str)
2882 return setup_fault_attr(&fail_page_alloc.attr, str);
2884 __setup("fail_page_alloc=", setup_fail_page_alloc);
2886 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2888 if (order < fail_page_alloc.min_order)
2890 if (gfp_mask & __GFP_NOFAIL)
2892 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2894 if (fail_page_alloc.ignore_gfp_reclaim &&
2895 (gfp_mask & __GFP_DIRECT_RECLAIM))
2898 return should_fail(&fail_page_alloc.attr, 1 << order);
2901 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2903 static int __init fail_page_alloc_debugfs(void)
2905 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2908 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2909 &fail_page_alloc.attr);
2911 return PTR_ERR(dir);
2913 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2914 &fail_page_alloc.ignore_gfp_reclaim))
2916 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2917 &fail_page_alloc.ignore_gfp_highmem))
2919 if (!debugfs_create_u32("min-order", mode, dir,
2920 &fail_page_alloc.min_order))
2925 debugfs_remove_recursive(dir);
2930 late_initcall(fail_page_alloc_debugfs);
2932 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2934 #else /* CONFIG_FAIL_PAGE_ALLOC */
2936 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2941 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2944 * Return true if free base pages are above 'mark'. For high-order checks it
2945 * will return true of the order-0 watermark is reached and there is at least
2946 * one free page of a suitable size. Checking now avoids taking the zone lock
2947 * to check in the allocation paths if no pages are free.
2949 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2950 int classzone_idx, unsigned int alloc_flags,
2955 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
2957 /* free_pages may go negative - that's OK */
2958 free_pages -= (1 << order) - 1;
2960 if (alloc_flags & ALLOC_HIGH)
2964 * If the caller does not have rights to ALLOC_HARDER then subtract
2965 * the high-atomic reserves. This will over-estimate the size of the
2966 * atomic reserve but it avoids a search.
2968 if (likely(!alloc_harder)) {
2969 free_pages -= z->nr_reserved_highatomic;
2972 * OOM victims can try even harder than normal ALLOC_HARDER
2973 * users on the grounds that it's definitely going to be in
2974 * the exit path shortly and free memory. Any allocation it
2975 * makes during the free path will be small and short-lived.
2977 if (alloc_flags & ALLOC_OOM)
2985 /* If allocation can't use CMA areas don't use free CMA pages */
2986 if (!(alloc_flags & ALLOC_CMA))
2987 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2991 * Check watermarks for an order-0 allocation request. If these
2992 * are not met, then a high-order request also cannot go ahead
2993 * even if a suitable page happened to be free.
2995 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2998 /* If this is an order-0 request then the watermark is fine */
3002 /* For a high-order request, check at least one suitable page is free */
3003 for (o = order; o < MAX_ORDER; o++) {
3004 struct free_area *area = &z->free_area[o];
3010 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3011 if (!list_empty(&area->free_list[mt]))
3016 if ((alloc_flags & ALLOC_CMA) &&
3017 !list_empty(&area->free_list[MIGRATE_CMA])) {
3022 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3028 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3029 int classzone_idx, unsigned int alloc_flags)
3031 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3032 zone_page_state(z, NR_FREE_PAGES));
3035 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3036 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3038 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3042 /* If allocation can't use CMA areas don't use free CMA pages */
3043 if (!(alloc_flags & ALLOC_CMA))
3044 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3048 * Fast check for order-0 only. If this fails then the reserves
3049 * need to be calculated. There is a corner case where the check
3050 * passes but only the high-order atomic reserve are free. If
3051 * the caller is !atomic then it'll uselessly search the free
3052 * list. That corner case is then slower but it is harmless.
3054 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3057 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3061 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3062 unsigned long mark, int classzone_idx)
3064 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3066 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3067 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3069 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3074 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3076 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3079 #else /* CONFIG_NUMA */
3080 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3084 #endif /* CONFIG_NUMA */
3087 * get_page_from_freelist goes through the zonelist trying to allocate
3090 static struct page *
3091 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3092 const struct alloc_context *ac)
3094 struct zoneref *z = ac->preferred_zoneref;
3096 struct pglist_data *last_pgdat_dirty_limit = NULL;
3099 * Scan zonelist, looking for a zone with enough free.
3100 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3102 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3107 if (cpusets_enabled() &&
3108 (alloc_flags & ALLOC_CPUSET) &&
3109 !__cpuset_zone_allowed(zone, gfp_mask))
3112 * When allocating a page cache page for writing, we
3113 * want to get it from a node that is within its dirty
3114 * limit, such that no single node holds more than its
3115 * proportional share of globally allowed dirty pages.
3116 * The dirty limits take into account the node's
3117 * lowmem reserves and high watermark so that kswapd
3118 * should be able to balance it without having to
3119 * write pages from its LRU list.
3121 * XXX: For now, allow allocations to potentially
3122 * exceed the per-node dirty limit in the slowpath
3123 * (spread_dirty_pages unset) before going into reclaim,
3124 * which is important when on a NUMA setup the allowed
3125 * nodes are together not big enough to reach the
3126 * global limit. The proper fix for these situations
3127 * will require awareness of nodes in the
3128 * dirty-throttling and the flusher threads.
3130 if (ac->spread_dirty_pages) {
3131 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3134 if (!node_dirty_ok(zone->zone_pgdat)) {
3135 last_pgdat_dirty_limit = zone->zone_pgdat;
3140 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3141 if (!zone_watermark_fast(zone, order, mark,
3142 ac_classzone_idx(ac), alloc_flags)) {
3145 /* Checked here to keep the fast path fast */
3146 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3147 if (alloc_flags & ALLOC_NO_WATERMARKS)
3150 if (node_reclaim_mode == 0 ||
3151 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3154 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3156 case NODE_RECLAIM_NOSCAN:
3159 case NODE_RECLAIM_FULL:
3160 /* scanned but unreclaimable */
3163 /* did we reclaim enough */
3164 if (zone_watermark_ok(zone, order, mark,
3165 ac_classzone_idx(ac), alloc_flags))
3173 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3174 gfp_mask, alloc_flags, ac->migratetype);
3176 prep_new_page(page, order, gfp_mask, alloc_flags);
3179 * If this is a high-order atomic allocation then check
3180 * if the pageblock should be reserved for the future
3182 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3183 reserve_highatomic_pageblock(page, zone, order);
3193 * Large machines with many possible nodes should not always dump per-node
3194 * meminfo in irq context.
3196 static inline bool should_suppress_show_mem(void)
3201 ret = in_interrupt();
3206 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3208 unsigned int filter = SHOW_MEM_FILTER_NODES;
3209 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3211 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3215 * This documents exceptions given to allocations in certain
3216 * contexts that are allowed to allocate outside current's set
3219 if (!(gfp_mask & __GFP_NOMEMALLOC))
3220 if (tsk_is_oom_victim(current) ||
3221 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3222 filter &= ~SHOW_MEM_FILTER_NODES;
3223 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3224 filter &= ~SHOW_MEM_FILTER_NODES;
3226 show_mem(filter, nodemask);
3229 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3231 struct va_format vaf;
3233 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3234 DEFAULT_RATELIMIT_BURST);
3236 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3239 pr_warn("%s: ", current->comm);
3241 va_start(args, fmt);
3244 pr_cont("%pV", &vaf);
3247 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3249 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3251 pr_cont("(null)\n");
3253 cpuset_print_current_mems_allowed();
3256 warn_alloc_show_mem(gfp_mask, nodemask);
3259 static inline struct page *
3260 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3261 unsigned int alloc_flags,
3262 const struct alloc_context *ac)
3266 page = get_page_from_freelist(gfp_mask, order,
3267 alloc_flags|ALLOC_CPUSET, ac);
3269 * fallback to ignore cpuset restriction if our nodes
3273 page = get_page_from_freelist(gfp_mask, order,
3279 static inline struct page *
3280 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3281 const struct alloc_context *ac, unsigned long *did_some_progress)
3283 struct oom_control oc = {
3284 .zonelist = ac->zonelist,
3285 .nodemask = ac->nodemask,
3287 .gfp_mask = gfp_mask,
3292 *did_some_progress = 0;
3295 * Acquire the oom lock. If that fails, somebody else is
3296 * making progress for us.
3298 if (!mutex_trylock(&oom_lock)) {
3299 *did_some_progress = 1;
3300 schedule_timeout_uninterruptible(1);
3305 * Go through the zonelist yet one more time, keep very high watermark
3306 * here, this is only to catch a parallel oom killing, we must fail if
3307 * we're still under heavy pressure. But make sure that this reclaim
3308 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3309 * allocation which will never fail due to oom_lock already held.
3311 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3312 ~__GFP_DIRECT_RECLAIM, order,
3313 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3317 /* Coredumps can quickly deplete all memory reserves */
3318 if (current->flags & PF_DUMPCORE)
3320 /* The OOM killer will not help higher order allocs */
3321 if (order > PAGE_ALLOC_COSTLY_ORDER)
3324 * We have already exhausted all our reclaim opportunities without any
3325 * success so it is time to admit defeat. We will skip the OOM killer
3326 * because it is very likely that the caller has a more reasonable
3327 * fallback than shooting a random task.
3329 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3331 /* The OOM killer does not needlessly kill tasks for lowmem */
3332 if (ac->high_zoneidx < ZONE_NORMAL)
3334 if (pm_suspended_storage())
3337 * XXX: GFP_NOFS allocations should rather fail than rely on
3338 * other request to make a forward progress.
3339 * We are in an unfortunate situation where out_of_memory cannot
3340 * do much for this context but let's try it to at least get
3341 * access to memory reserved if the current task is killed (see
3342 * out_of_memory). Once filesystems are ready to handle allocation
3343 * failures more gracefully we should just bail out here.
3346 /* The OOM killer may not free memory on a specific node */
3347 if (gfp_mask & __GFP_THISNODE)
3350 /* Exhausted what can be done so it's blamo time */
3351 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3352 *did_some_progress = 1;
3355 * Help non-failing allocations by giving them access to memory
3358 if (gfp_mask & __GFP_NOFAIL)
3359 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3360 ALLOC_NO_WATERMARKS, ac);
3363 mutex_unlock(&oom_lock);
3368 * Maximum number of compaction retries wit a progress before OOM
3369 * killer is consider as the only way to move forward.
3371 #define MAX_COMPACT_RETRIES 16
3373 #ifdef CONFIG_COMPACTION
3374 /* Try memory compaction for high-order allocations before reclaim */
3375 static struct page *
3376 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3377 unsigned int alloc_flags, const struct alloc_context *ac,
3378 enum compact_priority prio, enum compact_result *compact_result)
3381 unsigned int noreclaim_flag;
3386 noreclaim_flag = memalloc_noreclaim_save();
3387 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3389 memalloc_noreclaim_restore(noreclaim_flag);
3391 if (*compact_result <= COMPACT_INACTIVE)
3395 * At least in one zone compaction wasn't deferred or skipped, so let's
3396 * count a compaction stall
3398 count_vm_event(COMPACTSTALL);
3400 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3403 struct zone *zone = page_zone(page);
3405 zone->compact_blockskip_flush = false;
3406 compaction_defer_reset(zone, order, true);
3407 count_vm_event(COMPACTSUCCESS);
3412 * It's bad if compaction run occurs and fails. The most likely reason
3413 * is that pages exist, but not enough to satisfy watermarks.
3415 count_vm_event(COMPACTFAIL);
3423 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3424 enum compact_result compact_result,
3425 enum compact_priority *compact_priority,
3426 int *compaction_retries)
3428 int max_retries = MAX_COMPACT_RETRIES;
3431 int retries = *compaction_retries;
3432 enum compact_priority priority = *compact_priority;
3437 if (compaction_made_progress(compact_result))
3438 (*compaction_retries)++;
3441 * compaction considers all the zone as desperately out of memory
3442 * so it doesn't really make much sense to retry except when the
3443 * failure could be caused by insufficient priority
3445 if (compaction_failed(compact_result))
3446 goto check_priority;
3449 * make sure the compaction wasn't deferred or didn't bail out early
3450 * due to locks contention before we declare that we should give up.
3451 * But do not retry if the given zonelist is not suitable for
3454 if (compaction_withdrawn(compact_result)) {
3455 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3460 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3461 * costly ones because they are de facto nofail and invoke OOM
3462 * killer to move on while costly can fail and users are ready
3463 * to cope with that. 1/4 retries is rather arbitrary but we
3464 * would need much more detailed feedback from compaction to
3465 * make a better decision.
3467 if (order > PAGE_ALLOC_COSTLY_ORDER)
3469 if (*compaction_retries <= max_retries) {
3475 * Make sure there are attempts at the highest priority if we exhausted
3476 * all retries or failed at the lower priorities.
3479 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3480 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3482 if (*compact_priority > min_priority) {
3483 (*compact_priority)--;
3484 *compaction_retries = 0;
3488 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3492 static inline struct page *
3493 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3494 unsigned int alloc_flags, const struct alloc_context *ac,
3495 enum compact_priority prio, enum compact_result *compact_result)
3497 *compact_result = COMPACT_SKIPPED;
3502 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3503 enum compact_result compact_result,
3504 enum compact_priority *compact_priority,
3505 int *compaction_retries)
3510 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3514 * There are setups with compaction disabled which would prefer to loop
3515 * inside the allocator rather than hit the oom killer prematurely.
3516 * Let's give them a good hope and keep retrying while the order-0
3517 * watermarks are OK.
3519 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3521 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3522 ac_classzone_idx(ac), alloc_flags))
3527 #endif /* CONFIG_COMPACTION */
3529 #ifdef CONFIG_LOCKDEP
3530 struct lockdep_map __fs_reclaim_map =
3531 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3533 static bool __need_fs_reclaim(gfp_t gfp_mask)
3535 gfp_mask = current_gfp_context(gfp_mask);
3537 /* no reclaim without waiting on it */
3538 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3541 /* this guy won't enter reclaim */
3542 if (current->flags & PF_MEMALLOC)
3545 /* We're only interested __GFP_FS allocations for now */
3546 if (!(gfp_mask & __GFP_FS))
3549 if (gfp_mask & __GFP_NOLOCKDEP)
3555 void fs_reclaim_acquire(gfp_t gfp_mask)
3557 if (__need_fs_reclaim(gfp_mask))
3558 lock_map_acquire(&__fs_reclaim_map);
3560 EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3562 void fs_reclaim_release(gfp_t gfp_mask)
3564 if (__need_fs_reclaim(gfp_mask))
3565 lock_map_release(&__fs_reclaim_map);
3567 EXPORT_SYMBOL_GPL(fs_reclaim_release);
3571 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3572 * have been rebuilt so allocation retries. Reader side does not lock and
3573 * retries the allocation if zonelist changes. Writer side is protected by the
3574 * embedded spin_lock.
3576 static DEFINE_SEQLOCK(zonelist_update_seq);
3578 static unsigned int zonelist_iter_begin(void)
3580 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3581 return read_seqbegin(&zonelist_update_seq);
3586 static unsigned int check_retry_zonelist(unsigned int seq)
3588 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3589 return read_seqretry(&zonelist_update_seq, seq);
3594 /* Perform direct synchronous page reclaim */
3596 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3597 const struct alloc_context *ac)
3599 struct reclaim_state reclaim_state;
3601 unsigned int noreclaim_flag;
3605 /* We now go into synchronous reclaim */
3606 cpuset_memory_pressure_bump();
3607 noreclaim_flag = memalloc_noreclaim_save();
3608 fs_reclaim_acquire(gfp_mask);
3609 reclaim_state.reclaimed_slab = 0;
3610 current->reclaim_state = &reclaim_state;
3612 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3615 current->reclaim_state = NULL;
3616 fs_reclaim_release(gfp_mask);
3617 memalloc_noreclaim_restore(noreclaim_flag);
3624 /* The really slow allocator path where we enter direct reclaim */
3625 static inline struct page *
3626 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3627 unsigned int alloc_flags, const struct alloc_context *ac,
3628 unsigned long *did_some_progress)
3630 struct page *page = NULL;
3631 bool drained = false;
3633 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3634 if (unlikely(!(*did_some_progress)))
3638 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3641 * If an allocation failed after direct reclaim, it could be because
3642 * pages are pinned on the per-cpu lists or in high alloc reserves.
3643 * Shrink them them and try again
3645 if (!page && !drained) {
3646 unreserve_highatomic_pageblock(ac, false);
3647 drain_all_pages(NULL);
3655 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3659 pg_data_t *last_pgdat = NULL;
3661 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3662 ac->high_zoneidx, ac->nodemask) {
3663 if (last_pgdat != zone->zone_pgdat)
3664 wakeup_kswapd(zone, order, ac->high_zoneidx);
3665 last_pgdat = zone->zone_pgdat;
3669 static inline unsigned int
3670 gfp_to_alloc_flags(gfp_t gfp_mask)
3672 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3674 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3675 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3678 * The caller may dip into page reserves a bit more if the caller
3679 * cannot run direct reclaim, or if the caller has realtime scheduling
3680 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3681 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3683 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3685 if (gfp_mask & __GFP_ATOMIC) {
3687 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3688 * if it can't schedule.
3690 if (!(gfp_mask & __GFP_NOMEMALLOC))
3691 alloc_flags |= ALLOC_HARDER;
3693 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3694 * comment for __cpuset_node_allowed().
3696 alloc_flags &= ~ALLOC_CPUSET;
3697 } else if (unlikely(rt_task(current)) && !in_interrupt())
3698 alloc_flags |= ALLOC_HARDER;
3701 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3702 alloc_flags |= ALLOC_CMA;
3707 static bool oom_reserves_allowed(struct task_struct *tsk)
3709 if (!tsk_is_oom_victim(tsk))
3713 * !MMU doesn't have oom reaper so give access to memory reserves
3714 * only to the thread with TIF_MEMDIE set
3716 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3723 * Distinguish requests which really need access to full memory
3724 * reserves from oom victims which can live with a portion of it
3726 static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3728 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3730 if (gfp_mask & __GFP_MEMALLOC)
3731 return ALLOC_NO_WATERMARKS;
3732 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3733 return ALLOC_NO_WATERMARKS;
3734 if (!in_interrupt()) {
3735 if (current->flags & PF_MEMALLOC)
3736 return ALLOC_NO_WATERMARKS;
3737 else if (oom_reserves_allowed(current))
3744 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3746 return !!__gfp_pfmemalloc_flags(gfp_mask);
3750 * Checks whether it makes sense to retry the reclaim to make a forward progress
3751 * for the given allocation request.
3753 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3754 * without success, or when we couldn't even meet the watermark if we
3755 * reclaimed all remaining pages on the LRU lists.
3757 * Returns true if a retry is viable or false to enter the oom path.
3760 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3761 struct alloc_context *ac, int alloc_flags,
3762 bool did_some_progress, int *no_progress_loops)
3768 * Costly allocations might have made a progress but this doesn't mean
3769 * their order will become available due to high fragmentation so
3770 * always increment the no progress counter for them
3772 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3773 *no_progress_loops = 0;
3775 (*no_progress_loops)++;
3778 * Make sure we converge to OOM if we cannot make any progress
3779 * several times in the row.
3781 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3782 /* Before OOM, exhaust highatomic_reserve */
3783 return unreserve_highatomic_pageblock(ac, true);
3787 * Keep reclaiming pages while there is a chance this will lead
3788 * somewhere. If none of the target zones can satisfy our allocation
3789 * request even if all reclaimable pages are considered then we are
3790 * screwed and have to go OOM.
3792 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3794 unsigned long available;
3795 unsigned long reclaimable;
3796 unsigned long min_wmark = min_wmark_pages(zone);
3799 available = reclaimable = zone_reclaimable_pages(zone);
3800 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3803 * Would the allocation succeed if we reclaimed all
3804 * reclaimable pages?
3806 wmark = __zone_watermark_ok(zone, order, min_wmark,
3807 ac_classzone_idx(ac), alloc_flags, available);
3808 trace_reclaim_retry_zone(z, order, reclaimable,
3809 available, min_wmark, *no_progress_loops, wmark);
3812 * If we didn't make any progress and have a lot of
3813 * dirty + writeback pages then we should wait for
3814 * an IO to complete to slow down the reclaim and
3815 * prevent from pre mature OOM
3817 if (!did_some_progress) {
3818 unsigned long write_pending;
3820 write_pending = zone_page_state_snapshot(zone,
3821 NR_ZONE_WRITE_PENDING);
3823 if (2 * write_pending > reclaimable) {
3824 congestion_wait(BLK_RW_ASYNC, HZ/10);
3830 * Memory allocation/reclaim might be called from a WQ
3831 * context and the current implementation of the WQ
3832 * concurrency control doesn't recognize that
3833 * a particular WQ is congested if the worker thread is
3834 * looping without ever sleeping. Therefore we have to
3835 * do a short sleep here rather than calling
3838 if (current->flags & PF_WQ_WORKER)
3839 schedule_timeout_uninterruptible(1);
3851 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3854 * It's possible that cpuset's mems_allowed and the nodemask from
3855 * mempolicy don't intersect. This should be normally dealt with by
3856 * policy_nodemask(), but it's possible to race with cpuset update in
3857 * such a way the check therein was true, and then it became false
3858 * before we got our cpuset_mems_cookie here.
3859 * This assumes that for all allocations, ac->nodemask can come only
3860 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3861 * when it does not intersect with the cpuset restrictions) or the
3862 * caller can deal with a violated nodemask.
3864 if (cpusets_enabled() && ac->nodemask &&
3865 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3866 ac->nodemask = NULL;
3871 * When updating a task's mems_allowed or mempolicy nodemask, it is
3872 * possible to race with parallel threads in such a way that our
3873 * allocation can fail while the mask is being updated. If we are about
3874 * to fail, check if the cpuset changed during allocation and if so,
3877 if (read_mems_allowed_retry(cpuset_mems_cookie))
3883 static inline struct page *
3884 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3885 struct alloc_context *ac)
3887 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3888 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3889 struct page *page = NULL;
3890 unsigned int alloc_flags;
3891 unsigned long did_some_progress;
3892 enum compact_priority compact_priority;
3893 enum compact_result compact_result;
3894 int compaction_retries;
3895 int no_progress_loops;
3896 unsigned int cpuset_mems_cookie;
3897 unsigned int zonelist_iter_cookie;
3901 * We also sanity check to catch abuse of atomic reserves being used by
3902 * callers that are not in atomic context.
3904 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3905 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3906 gfp_mask &= ~__GFP_ATOMIC;
3909 compaction_retries = 0;
3910 no_progress_loops = 0;
3911 compact_priority = DEF_COMPACT_PRIORITY;
3912 cpuset_mems_cookie = read_mems_allowed_begin();
3913 zonelist_iter_cookie = zonelist_iter_begin();
3916 * The fast path uses conservative alloc_flags to succeed only until
3917 * kswapd needs to be woken up, and to avoid the cost of setting up
3918 * alloc_flags precisely. So we do that now.
3920 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3923 * We need to recalculate the starting point for the zonelist iterator
3924 * because we might have used different nodemask in the fast path, or
3925 * there was a cpuset modification and we are retrying - otherwise we
3926 * could end up iterating over non-eligible zones endlessly.
3928 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3929 ac->high_zoneidx, ac->nodemask);
3930 if (!ac->preferred_zoneref->zone)
3933 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3934 wake_all_kswapds(order, ac);
3937 * The adjusted alloc_flags might result in immediate success, so try
3940 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3945 * For costly allocations, try direct compaction first, as it's likely
3946 * that we have enough base pages and don't need to reclaim. For non-
3947 * movable high-order allocations, do that as well, as compaction will
3948 * try prevent permanent fragmentation by migrating from blocks of the
3950 * Don't try this for allocations that are allowed to ignore
3951 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3953 if (can_direct_reclaim &&
3955 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3956 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3957 page = __alloc_pages_direct_compact(gfp_mask, order,
3959 INIT_COMPACT_PRIORITY,
3965 * Checks for costly allocations with __GFP_NORETRY, which
3966 * includes THP page fault allocations
3968 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
3970 * If compaction is deferred for high-order allocations,
3971 * it is because sync compaction recently failed. If
3972 * this is the case and the caller requested a THP
3973 * allocation, we do not want to heavily disrupt the
3974 * system, so we fail the allocation instead of entering
3977 if (compact_result == COMPACT_DEFERRED)
3981 * Looks like reclaim/compaction is worth trying, but
3982 * sync compaction could be very expensive, so keep
3983 * using async compaction.
3985 compact_priority = INIT_COMPACT_PRIORITY;
3990 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3991 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3992 wake_all_kswapds(order, ac);
3994 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
3996 alloc_flags = reserve_flags;
3999 * Reset the zonelist iterators if memory policies can be ignored.
4000 * These allocations are high priority and system rather than user
4003 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4004 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4005 ac->high_zoneidx, ac->nodemask);
4008 /* Attempt with potentially adjusted zonelist and alloc_flags */
4009 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4013 /* Caller is not willing to reclaim, we can't balance anything */
4014 if (!can_direct_reclaim)
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))
4062 * Deal with possible cpuset update races or zonelist updates to avoid
4063 * a unnecessary OOM kill.
4065 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4066 check_retry_zonelist(zonelist_iter_cookie))
4069 /* Reclaim has failed us, start killing things */
4070 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4074 /* Avoid allocations with no watermarks from looping endlessly */
4075 if (tsk_is_oom_victim(current) &&
4076 (alloc_flags == ALLOC_OOM ||
4077 (gfp_mask & __GFP_NOMEMALLOC)))
4080 /* Retry as long as the OOM killer is making progress */
4081 if (did_some_progress) {
4082 no_progress_loops = 0;
4088 * Deal with possible cpuset update races or zonelist updates to avoid
4089 * a unnecessary OOM kill.
4091 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4092 check_retry_zonelist(zonelist_iter_cookie))
4096 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4099 if (gfp_mask & __GFP_NOFAIL) {
4101 * All existing users of the __GFP_NOFAIL are blockable, so warn
4102 * of any new users that actually require GFP_NOWAIT
4104 if (WARN_ON_ONCE(!can_direct_reclaim))
4108 * PF_MEMALLOC request from this context is rather bizarre
4109 * because we cannot reclaim anything and only can loop waiting
4110 * for somebody to do a work for us
4112 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4115 * non failing costly orders are a hard requirement which we
4116 * are not prepared for much so let's warn about these users
4117 * so that we can identify them and convert them to something
4120 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4123 * Help non-failing allocations by giving them access to memory
4124 * reserves but do not use ALLOC_NO_WATERMARKS because this
4125 * could deplete whole memory reserves which would just make
4126 * the situation worse
4128 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4136 warn_alloc(gfp_mask, ac->nodemask,
4137 "page allocation failure: order:%u", order);
4142 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4143 int preferred_nid, nodemask_t *nodemask,
4144 struct alloc_context *ac, gfp_t *alloc_mask,
4145 unsigned int *alloc_flags)
4147 ac->high_zoneidx = gfp_zone(gfp_mask);
4148 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4149 ac->nodemask = nodemask;
4150 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4152 if (cpusets_enabled()) {
4153 *alloc_mask |= __GFP_HARDWALL;
4155 ac->nodemask = &cpuset_current_mems_allowed;
4157 *alloc_flags |= ALLOC_CPUSET;
4160 fs_reclaim_acquire(gfp_mask);
4161 fs_reclaim_release(gfp_mask);
4163 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4165 if (should_fail_alloc_page(gfp_mask, order))
4168 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4169 *alloc_flags |= ALLOC_CMA;
4174 /* Determine whether to spread dirty pages and what the first usable zone */
4175 static inline void finalise_ac(gfp_t gfp_mask,
4176 unsigned int order, struct alloc_context *ac)
4178 /* Dirty zone balancing only done in the fast path */
4179 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4182 * The preferred zone is used for statistics but crucially it is
4183 * also used as the starting point for the zonelist iterator. It
4184 * may get reset for allocations that ignore memory policies.
4186 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4187 ac->high_zoneidx, ac->nodemask);
4191 * This is the 'heart' of the zoned buddy allocator.
4194 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4195 nodemask_t *nodemask)
4198 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4199 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4200 struct alloc_context ac = { };
4203 * There are several places where we assume that the order value is sane
4204 * so bail out early if the request is out of bound.
4206 if (unlikely(order >= MAX_ORDER)) {
4207 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4211 gfp_mask &= gfp_allowed_mask;
4212 alloc_mask = gfp_mask;
4213 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4216 finalise_ac(gfp_mask, order, &ac);
4218 /* First allocation attempt */
4219 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4224 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4225 * resp. GFP_NOIO which has to be inherited for all allocation requests
4226 * from a particular context which has been marked by
4227 * memalloc_no{fs,io}_{save,restore}.
4229 alloc_mask = current_gfp_context(gfp_mask);
4230 ac.spread_dirty_pages = false;
4233 * Restore the original nodemask if it was potentially replaced with
4234 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4236 if (unlikely(ac.nodemask != nodemask))
4237 ac.nodemask = nodemask;
4239 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4242 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4243 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4244 __free_pages(page, order);
4248 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4252 EXPORT_SYMBOL(__alloc_pages_nodemask);
4255 * Common helper functions.
4257 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4262 * __get_free_pages() returns a 32-bit address, which cannot represent
4265 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4267 page = alloc_pages(gfp_mask, order);
4270 return (unsigned long) page_address(page);
4272 EXPORT_SYMBOL(__get_free_pages);
4274 unsigned long get_zeroed_page(gfp_t gfp_mask)
4276 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4278 EXPORT_SYMBOL(get_zeroed_page);
4280 void __free_pages(struct page *page, unsigned int order)
4282 if (put_page_testzero(page)) {
4284 free_hot_cold_page(page, false);
4286 __free_pages_ok(page, order);
4290 EXPORT_SYMBOL(__free_pages);
4292 void free_pages(unsigned long addr, unsigned int order)
4295 VM_BUG_ON(!virt_addr_valid((void *)addr));
4296 __free_pages(virt_to_page((void *)addr), order);
4300 EXPORT_SYMBOL(free_pages);
4304 * An arbitrary-length arbitrary-offset area of memory which resides
4305 * within a 0 or higher order page. Multiple fragments within that page
4306 * are individually refcounted, in the page's reference counter.
4308 * The page_frag functions below provide a simple allocation framework for
4309 * page fragments. This is used by the network stack and network device
4310 * drivers to provide a backing region of memory for use as either an
4311 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4313 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4316 struct page *page = NULL;
4317 gfp_t gfp = gfp_mask;
4319 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4320 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4322 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4323 PAGE_FRAG_CACHE_MAX_ORDER);
4324 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4326 if (unlikely(!page))
4327 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4329 nc->va = page ? page_address(page) : NULL;
4334 void __page_frag_cache_drain(struct page *page, unsigned int count)
4336 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4338 if (page_ref_sub_and_test(page, count)) {
4339 unsigned int order = compound_order(page);
4342 free_hot_cold_page(page, false);
4344 __free_pages_ok(page, order);
4347 EXPORT_SYMBOL(__page_frag_cache_drain);
4349 void *page_frag_alloc(struct page_frag_cache *nc,
4350 unsigned int fragsz, gfp_t gfp_mask)
4352 unsigned int size = PAGE_SIZE;
4356 if (unlikely(!nc->va)) {
4358 page = __page_frag_cache_refill(nc, gfp_mask);
4362 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4363 /* if size can vary use size else just use PAGE_SIZE */
4366 /* Even if we own the page, we do not use atomic_set().
4367 * This would break get_page_unless_zero() users.
4369 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4371 /* reset page count bias and offset to start of new frag */
4372 nc->pfmemalloc = page_is_pfmemalloc(page);
4373 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4377 offset = nc->offset - fragsz;
4378 if (unlikely(offset < 0)) {
4379 page = virt_to_page(nc->va);
4381 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4384 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4385 /* if size can vary use size else just use PAGE_SIZE */
4388 /* OK, page count is 0, we can safely set it */
4389 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4391 /* reset page count bias and offset to start of new frag */
4392 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4393 offset = size - fragsz;
4394 if (unlikely(offset < 0)) {
4396 * The caller is trying to allocate a fragment
4397 * with fragsz > PAGE_SIZE but the cache isn't big
4398 * enough to satisfy the request, this may
4399 * happen in low memory conditions.
4400 * We don't release the cache page because
4401 * it could make memory pressure worse
4402 * so we simply return NULL here.
4409 nc->offset = offset;
4411 return nc->va + offset;
4413 EXPORT_SYMBOL(page_frag_alloc);
4416 * Frees a page fragment allocated out of either a compound or order 0 page.
4418 void page_frag_free(void *addr)
4420 struct page *page = virt_to_head_page(addr);
4422 if (unlikely(put_page_testzero(page)))
4423 __free_pages_ok(page, compound_order(page));
4425 EXPORT_SYMBOL(page_frag_free);
4427 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4431 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4432 unsigned long used = addr + PAGE_ALIGN(size);
4434 split_page(virt_to_page((void *)addr), order);
4435 while (used < alloc_end) {
4440 return (void *)addr;
4444 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4445 * @size: the number of bytes to allocate
4446 * @gfp_mask: GFP flags for the allocation
4448 * This function is similar to alloc_pages(), except that it allocates the
4449 * minimum number of pages to satisfy the request. alloc_pages() can only
4450 * allocate memory in power-of-two pages.
4452 * This function is also limited by MAX_ORDER.
4454 * Memory allocated by this function must be released by free_pages_exact().
4456 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4458 unsigned int order = get_order(size);
4461 addr = __get_free_pages(gfp_mask, order);
4462 return make_alloc_exact(addr, order, size);
4464 EXPORT_SYMBOL(alloc_pages_exact);
4467 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4469 * @nid: the preferred node ID where memory should be allocated
4470 * @size: the number of bytes to allocate
4471 * @gfp_mask: GFP flags for the allocation
4473 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4476 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4478 unsigned int order = get_order(size);
4479 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4482 return make_alloc_exact((unsigned long)page_address(p), order, size);
4486 * free_pages_exact - release memory allocated via alloc_pages_exact()
4487 * @virt: the value returned by alloc_pages_exact.
4488 * @size: size of allocation, same value as passed to alloc_pages_exact().
4490 * Release the memory allocated by a previous call to alloc_pages_exact.
4492 void free_pages_exact(void *virt, size_t size)
4494 unsigned long addr = (unsigned long)virt;
4495 unsigned long end = addr + PAGE_ALIGN(size);
4497 while (addr < end) {
4502 EXPORT_SYMBOL(free_pages_exact);
4505 * nr_free_zone_pages - count number of pages beyond high watermark
4506 * @offset: The zone index of the highest zone
4508 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4509 * high watermark within all zones at or below a given zone index. For each
4510 * zone, the number of pages is calculated as:
4512 * nr_free_zone_pages = managed_pages - high_pages
4514 static unsigned long nr_free_zone_pages(int offset)
4519 /* Just pick one node, since fallback list is circular */
4520 unsigned long sum = 0;
4522 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4524 for_each_zone_zonelist(zone, z, zonelist, offset) {
4525 unsigned long size = zone->managed_pages;
4526 unsigned long high = high_wmark_pages(zone);
4535 * nr_free_buffer_pages - count number of pages beyond high watermark
4537 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4538 * watermark within ZONE_DMA and ZONE_NORMAL.
4540 unsigned long nr_free_buffer_pages(void)
4542 return nr_free_zone_pages(gfp_zone(GFP_USER));
4544 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4547 * nr_free_pagecache_pages - count number of pages beyond high watermark
4549 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4550 * high watermark within all zones.
4552 unsigned long nr_free_pagecache_pages(void)
4554 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4557 static inline void show_node(struct zone *zone)
4559 if (IS_ENABLED(CONFIG_NUMA))
4560 printk("Node %d ", zone_to_nid(zone));
4563 long si_mem_available(void)
4566 unsigned long pagecache;
4567 unsigned long wmark_low = 0;
4568 unsigned long pages[NR_LRU_LISTS];
4572 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4573 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4576 wmark_low += zone->watermark[WMARK_LOW];
4579 * Estimate the amount of memory available for userspace allocations,
4580 * without causing swapping.
4582 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4585 * Not all the page cache can be freed, otherwise the system will
4586 * start swapping. Assume at least half of the page cache, or the
4587 * low watermark worth of cache, needs to stay.
4589 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4590 pagecache -= min(pagecache / 2, wmark_low);
4591 available += pagecache;
4594 * Part of the reclaimable slab consists of items that are in use,
4595 * and cannot be freed. Cap this estimate at the low watermark.
4597 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4598 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4602 * Part of the kernel memory, which can be released under memory
4605 available += global_node_page_state(NR_INDIRECTLY_RECLAIMABLE_BYTES) >>
4612 EXPORT_SYMBOL_GPL(si_mem_available);
4614 void si_meminfo(struct sysinfo *val)
4616 val->totalram = totalram_pages;
4617 val->sharedram = global_node_page_state(NR_SHMEM);
4618 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4619 val->bufferram = nr_blockdev_pages();
4620 val->totalhigh = totalhigh_pages;
4621 val->freehigh = nr_free_highpages();
4622 val->mem_unit = PAGE_SIZE;
4625 EXPORT_SYMBOL(si_meminfo);
4628 void si_meminfo_node(struct sysinfo *val, int nid)
4630 int zone_type; /* needs to be signed */
4631 unsigned long managed_pages = 0;
4632 unsigned long managed_highpages = 0;
4633 unsigned long free_highpages = 0;
4634 pg_data_t *pgdat = NODE_DATA(nid);
4636 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4637 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4638 val->totalram = managed_pages;
4639 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4640 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4641 #ifdef CONFIG_HIGHMEM
4642 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4643 struct zone *zone = &pgdat->node_zones[zone_type];
4645 if (is_highmem(zone)) {
4646 managed_highpages += zone->managed_pages;
4647 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4650 val->totalhigh = managed_highpages;
4651 val->freehigh = free_highpages;
4653 val->totalhigh = managed_highpages;
4654 val->freehigh = free_highpages;
4656 val->mem_unit = PAGE_SIZE;
4661 * Determine whether the node should be displayed or not, depending on whether
4662 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4664 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4666 if (!(flags & SHOW_MEM_FILTER_NODES))
4670 * no node mask - aka implicit memory numa policy. Do not bother with
4671 * the synchronization - read_mems_allowed_begin - because we do not
4672 * have to be precise here.
4675 nodemask = &cpuset_current_mems_allowed;
4677 return !node_isset(nid, *nodemask);
4680 #define K(x) ((x) << (PAGE_SHIFT-10))
4682 static void show_migration_types(unsigned char type)
4684 static const char types[MIGRATE_TYPES] = {
4685 [MIGRATE_UNMOVABLE] = 'U',
4686 [MIGRATE_MOVABLE] = 'M',
4687 [MIGRATE_RECLAIMABLE] = 'E',
4688 [MIGRATE_HIGHATOMIC] = 'H',
4690 [MIGRATE_CMA] = 'C',
4692 #ifdef CONFIG_MEMORY_ISOLATION
4693 [MIGRATE_ISOLATE] = 'I',
4696 char tmp[MIGRATE_TYPES + 1];
4700 for (i = 0; i < MIGRATE_TYPES; i++) {
4701 if (type & (1 << i))
4706 printk(KERN_CONT "(%s) ", tmp);
4710 * Show free area list (used inside shift_scroll-lock stuff)
4711 * We also calculate the percentage fragmentation. We do this by counting the
4712 * memory on each free list with the exception of the first item on the list.
4715 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4718 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4720 unsigned long free_pcp = 0;
4725 for_each_populated_zone(zone) {
4726 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4729 for_each_online_cpu(cpu)
4730 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4733 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4734 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4735 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4736 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4737 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4738 " free:%lu free_pcp:%lu free_cma:%lu\n",
4739 global_node_page_state(NR_ACTIVE_ANON),
4740 global_node_page_state(NR_INACTIVE_ANON),
4741 global_node_page_state(NR_ISOLATED_ANON),
4742 global_node_page_state(NR_ACTIVE_FILE),
4743 global_node_page_state(NR_INACTIVE_FILE),
4744 global_node_page_state(NR_ISOLATED_FILE),
4745 global_node_page_state(NR_UNEVICTABLE),
4746 global_node_page_state(NR_FILE_DIRTY),
4747 global_node_page_state(NR_WRITEBACK),
4748 global_node_page_state(NR_UNSTABLE_NFS),
4749 global_node_page_state(NR_SLAB_RECLAIMABLE),
4750 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4751 global_node_page_state(NR_FILE_MAPPED),
4752 global_node_page_state(NR_SHMEM),
4753 global_zone_page_state(NR_PAGETABLE),
4754 global_zone_page_state(NR_BOUNCE),
4755 global_zone_page_state(NR_FREE_PAGES),
4757 global_zone_page_state(NR_FREE_CMA_PAGES));
4759 for_each_online_pgdat(pgdat) {
4760 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4764 " active_anon:%lukB"
4765 " inactive_anon:%lukB"
4766 " active_file:%lukB"
4767 " inactive_file:%lukB"
4768 " unevictable:%lukB"
4769 " isolated(anon):%lukB"
4770 " isolated(file):%lukB"
4775 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4777 " shmem_pmdmapped: %lukB"
4780 " writeback_tmp:%lukB"
4782 " all_unreclaimable? %s"
4785 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4786 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4787 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4788 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4789 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4790 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4791 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4792 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4793 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4794 K(node_page_state(pgdat, NR_WRITEBACK)),
4795 K(node_page_state(pgdat, NR_SHMEM)),
4796 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4797 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4798 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4800 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4802 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4803 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4804 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4808 for_each_populated_zone(zone) {
4811 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4815 for_each_online_cpu(cpu)
4816 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4825 " active_anon:%lukB"
4826 " inactive_anon:%lukB"
4827 " active_file:%lukB"
4828 " inactive_file:%lukB"
4829 " unevictable:%lukB"
4830 " writepending:%lukB"
4834 " kernel_stack:%lukB"
4842 K(zone_page_state(zone, NR_FREE_PAGES)),
4843 K(min_wmark_pages(zone)),
4844 K(low_wmark_pages(zone)),
4845 K(high_wmark_pages(zone)),
4846 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4847 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4848 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4849 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4850 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4851 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4852 K(zone->present_pages),
4853 K(zone->managed_pages),
4854 K(zone_page_state(zone, NR_MLOCK)),
4855 zone_page_state(zone, NR_KERNEL_STACK_KB),
4856 K(zone_page_state(zone, NR_PAGETABLE)),
4857 K(zone_page_state(zone, NR_BOUNCE)),
4859 K(this_cpu_read(zone->pageset->pcp.count)),
4860 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4861 printk("lowmem_reserve[]:");
4862 for (i = 0; i < MAX_NR_ZONES; i++)
4863 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4864 printk(KERN_CONT "\n");
4867 for_each_populated_zone(zone) {
4869 unsigned long nr[MAX_ORDER], flags, total = 0;
4870 unsigned char types[MAX_ORDER];
4872 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4875 printk(KERN_CONT "%s: ", zone->name);
4877 spin_lock_irqsave(&zone->lock, flags);
4878 for (order = 0; order < MAX_ORDER; order++) {
4879 struct free_area *area = &zone->free_area[order];
4882 nr[order] = area->nr_free;
4883 total += nr[order] << order;
4886 for (type = 0; type < MIGRATE_TYPES; type++) {
4887 if (!list_empty(&area->free_list[type]))
4888 types[order] |= 1 << type;
4891 spin_unlock_irqrestore(&zone->lock, flags);
4892 for (order = 0; order < MAX_ORDER; order++) {
4893 printk(KERN_CONT "%lu*%lukB ",
4894 nr[order], K(1UL) << order);
4896 show_migration_types(types[order]);
4898 printk(KERN_CONT "= %lukB\n", K(total));
4901 hugetlb_show_meminfo();
4903 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4905 show_swap_cache_info();
4908 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4910 zoneref->zone = zone;
4911 zoneref->zone_idx = zone_idx(zone);
4915 * Builds allocation fallback zone lists.
4917 * Add all populated zones of a node to the zonelist.
4919 static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
4922 enum zone_type zone_type = MAX_NR_ZONES;
4927 zone = pgdat->node_zones + zone_type;
4928 if (populated_zone(zone)) {
4929 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
4930 check_highest_zone(zone_type);
4932 } while (zone_type);
4939 static int __parse_numa_zonelist_order(char *s)
4942 * We used to support different zonlists modes but they turned
4943 * out to be just not useful. Let's keep the warning in place
4944 * if somebody still use the cmd line parameter so that we do
4945 * not fail it silently
4947 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
4948 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
4954 static __init int setup_numa_zonelist_order(char *s)
4959 return __parse_numa_zonelist_order(s);
4961 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4963 char numa_zonelist_order[] = "Node";
4966 * sysctl handler for numa_zonelist_order
4968 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4969 void __user *buffer, size_t *length,
4976 return proc_dostring(table, write, buffer, length, ppos);
4977 str = memdup_user_nul(buffer, 16);
4979 return PTR_ERR(str);
4981 ret = __parse_numa_zonelist_order(str);
4987 #define MAX_NODE_LOAD (nr_online_nodes)
4988 static int node_load[MAX_NUMNODES];
4991 * find_next_best_node - find the next node that should appear in a given node's fallback list
4992 * @node: node whose fallback list we're appending
4993 * @used_node_mask: nodemask_t of already used nodes
4995 * We use a number of factors to determine which is the next node that should
4996 * appear on a given node's fallback list. The node should not have appeared
4997 * already in @node's fallback list, and it should be the next closest node
4998 * according to the distance array (which contains arbitrary distance values
4999 * from each node to each node in the system), and should also prefer nodes
5000 * with no CPUs, since presumably they'll have very little allocation pressure
5001 * on them otherwise.
5002 * It returns -1 if no node is found.
5004 static int find_next_best_node(int node, nodemask_t *used_node_mask)
5007 int min_val = INT_MAX;
5008 int best_node = NUMA_NO_NODE;
5009 const struct cpumask *tmp = cpumask_of_node(0);
5011 /* Use the local node if we haven't already */
5012 if (!node_isset(node, *used_node_mask)) {
5013 node_set(node, *used_node_mask);
5017 for_each_node_state(n, N_MEMORY) {
5019 /* Don't want a node to appear more than once */
5020 if (node_isset(n, *used_node_mask))
5023 /* Use the distance array to find the distance */
5024 val = node_distance(node, n);
5026 /* Penalize nodes under us ("prefer the next node") */
5029 /* Give preference to headless and unused nodes */
5030 tmp = cpumask_of_node(n);
5031 if (!cpumask_empty(tmp))
5032 val += PENALTY_FOR_NODE_WITH_CPUS;
5034 /* Slight preference for less loaded node */
5035 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5036 val += node_load[n];
5038 if (val < min_val) {
5045 node_set(best_node, *used_node_mask);
5052 * Build zonelists ordered by node and zones within node.
5053 * This results in maximum locality--normal zone overflows into local
5054 * DMA zone, if any--but risks exhausting DMA zone.
5056 static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5059 struct zoneref *zonerefs;
5062 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5064 for (i = 0; i < nr_nodes; i++) {
5067 pg_data_t *node = NODE_DATA(node_order[i]);
5069 nr_zones = build_zonerefs_node(node, zonerefs);
5070 zonerefs += nr_zones;
5072 zonerefs->zone = NULL;
5073 zonerefs->zone_idx = 0;
5077 * Build gfp_thisnode zonelists
5079 static void build_thisnode_zonelists(pg_data_t *pgdat)
5081 struct zoneref *zonerefs;
5084 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5085 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5086 zonerefs += nr_zones;
5087 zonerefs->zone = NULL;
5088 zonerefs->zone_idx = 0;
5092 * Build zonelists ordered by zone and nodes within zones.
5093 * This results in conserving DMA zone[s] until all Normal memory is
5094 * exhausted, but results in overflowing to remote node while memory
5095 * may still exist in local DMA zone.
5098 static void build_zonelists(pg_data_t *pgdat)
5100 static int node_order[MAX_NUMNODES];
5101 int node, load, nr_nodes = 0;
5102 nodemask_t used_mask;
5103 int local_node, prev_node;
5105 /* NUMA-aware ordering of nodes */
5106 local_node = pgdat->node_id;
5107 load = nr_online_nodes;
5108 prev_node = local_node;
5109 nodes_clear(used_mask);
5111 memset(node_order, 0, sizeof(node_order));
5112 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5114 * We don't want to pressure a particular node.
5115 * So adding penalty to the first node in same
5116 * distance group to make it round-robin.
5118 if (node_distance(local_node, node) !=
5119 node_distance(local_node, prev_node))
5120 node_load[node] = load;
5122 node_order[nr_nodes++] = node;
5127 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5128 build_thisnode_zonelists(pgdat);
5131 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5133 * Return node id of node used for "local" allocations.
5134 * I.e., first node id of first zone in arg node's generic zonelist.
5135 * Used for initializing percpu 'numa_mem', which is used primarily
5136 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5138 int local_memory_node(int node)
5142 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5143 gfp_zone(GFP_KERNEL),
5145 return z->zone->node;
5149 static void setup_min_unmapped_ratio(void);
5150 static void setup_min_slab_ratio(void);
5151 #else /* CONFIG_NUMA */
5153 static void build_zonelists(pg_data_t *pgdat)
5155 int node, local_node;
5156 struct zoneref *zonerefs;
5159 local_node = pgdat->node_id;
5161 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5162 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5163 zonerefs += nr_zones;
5166 * Now we build the zonelist so that it contains the zones
5167 * of all the other nodes.
5168 * We don't want to pressure a particular node, so when
5169 * building the zones for node N, we make sure that the
5170 * zones coming right after the local ones are those from
5171 * node N+1 (modulo N)
5173 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5174 if (!node_online(node))
5176 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5177 zonerefs += nr_zones;
5179 for (node = 0; node < local_node; node++) {
5180 if (!node_online(node))
5182 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5183 zonerefs += nr_zones;
5186 zonerefs->zone = NULL;
5187 zonerefs->zone_idx = 0;
5190 #endif /* CONFIG_NUMA */
5193 * Boot pageset table. One per cpu which is going to be used for all
5194 * zones and all nodes. The parameters will be set in such a way
5195 * that an item put on a list will immediately be handed over to
5196 * the buddy list. This is safe since pageset manipulation is done
5197 * with interrupts disabled.
5199 * The boot_pagesets must be kept even after bootup is complete for
5200 * unused processors and/or zones. They do play a role for bootstrapping
5201 * hotplugged processors.
5203 * zoneinfo_show() and maybe other functions do
5204 * not check if the processor is online before following the pageset pointer.
5205 * Other parts of the kernel may not check if the zone is available.
5207 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5208 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5209 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5211 static void __build_all_zonelists(void *data)
5214 int __maybe_unused cpu;
5215 pg_data_t *self = data;
5217 write_seqlock(&zonelist_update_seq);
5220 memset(node_load, 0, sizeof(node_load));
5224 * This node is hotadded and no memory is yet present. So just
5225 * building zonelists is fine - no need to touch other nodes.
5227 if (self && !node_online(self->node_id)) {
5228 build_zonelists(self);
5230 for_each_online_node(nid) {
5231 pg_data_t *pgdat = NODE_DATA(nid);
5233 build_zonelists(pgdat);
5236 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5238 * We now know the "local memory node" for each node--
5239 * i.e., the node of the first zone in the generic zonelist.
5240 * Set up numa_mem percpu variable for on-line cpus. During
5241 * boot, only the boot cpu should be on-line; we'll init the
5242 * secondary cpus' numa_mem as they come on-line. During
5243 * node/memory hotplug, we'll fixup all on-line cpus.
5245 for_each_online_cpu(cpu)
5246 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5250 write_sequnlock(&zonelist_update_seq);
5253 static noinline void __init
5254 build_all_zonelists_init(void)
5258 __build_all_zonelists(NULL);
5261 * Initialize the boot_pagesets that are going to be used
5262 * for bootstrapping processors. The real pagesets for
5263 * each zone will be allocated later when the per cpu
5264 * allocator is available.
5266 * boot_pagesets are used also for bootstrapping offline
5267 * cpus if the system is already booted because the pagesets
5268 * are needed to initialize allocators on a specific cpu too.
5269 * F.e. the percpu allocator needs the page allocator which
5270 * needs the percpu allocator in order to allocate its pagesets
5271 * (a chicken-egg dilemma).
5273 for_each_possible_cpu(cpu)
5274 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5276 mminit_verify_zonelist();
5277 cpuset_init_current_mems_allowed();
5281 * unless system_state == SYSTEM_BOOTING.
5283 * __ref due to call of __init annotated helper build_all_zonelists_init
5284 * [protected by SYSTEM_BOOTING].
5286 void __ref build_all_zonelists(pg_data_t *pgdat)
5288 if (system_state == SYSTEM_BOOTING) {
5289 build_all_zonelists_init();
5291 __build_all_zonelists(pgdat);
5292 /* cpuset refresh routine should be here */
5294 vm_total_pages = nr_free_pagecache_pages();
5296 * Disable grouping by mobility if the number of pages in the
5297 * system is too low to allow the mechanism to work. It would be
5298 * more accurate, but expensive to check per-zone. This check is
5299 * made on memory-hotadd so a system can start with mobility
5300 * disabled and enable it later
5302 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5303 page_group_by_mobility_disabled = 1;
5305 page_group_by_mobility_disabled = 0;
5307 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5309 page_group_by_mobility_disabled ? "off" : "on",
5312 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5317 * Initially all pages are reserved - free ones are freed
5318 * up by free_all_bootmem() once the early boot process is
5319 * done. Non-atomic initialization, single-pass.
5321 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5322 unsigned long start_pfn, enum memmap_context context)
5324 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5325 unsigned long end_pfn = start_pfn + size;
5326 pg_data_t *pgdat = NODE_DATA(nid);
5328 unsigned long nr_initialised = 0;
5329 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5330 struct memblock_region *r = NULL, *tmp;
5333 if (highest_memmap_pfn < end_pfn - 1)
5334 highest_memmap_pfn = end_pfn - 1;
5337 * Honor reservation requested by the driver for this ZONE_DEVICE
5340 if (altmap && start_pfn == altmap->base_pfn)
5341 start_pfn += altmap->reserve;
5343 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5345 * There can be holes in boot-time mem_map[]s handed to this
5346 * function. They do not exist on hotplugged memory.
5348 if (context != MEMMAP_EARLY)
5351 if (!early_pfn_valid(pfn))
5353 if (!early_pfn_in_nid(pfn, nid))
5355 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5358 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5360 * Check given memblock attribute by firmware which can affect
5361 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5362 * mirrored, it's an overlapped memmap init. skip it.
5364 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5365 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5366 for_each_memblock(memory, tmp)
5367 if (pfn < memblock_region_memory_end_pfn(tmp))
5371 if (pfn >= memblock_region_memory_base_pfn(r) &&
5372 memblock_is_mirror(r)) {
5373 /* already initialized as NORMAL */
5374 pfn = memblock_region_memory_end_pfn(r);
5382 * Mark the block movable so that blocks are reserved for
5383 * movable at startup. This will force kernel allocations
5384 * to reserve their blocks rather than leaking throughout
5385 * the address space during boot when many long-lived
5386 * kernel allocations are made.
5388 * bitmap is created for zone's valid pfn range. but memmap
5389 * can be created for invalid pages (for alignment)
5390 * check here not to call set_pageblock_migratetype() against
5393 if (!(pfn & (pageblock_nr_pages - 1))) {
5394 struct page *page = pfn_to_page(pfn);
5396 __init_single_page(page, pfn, zone, nid);
5397 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5400 __init_single_pfn(pfn, zone, nid);
5405 static void __meminit zone_init_free_lists(struct zone *zone)
5407 unsigned int order, t;
5408 for_each_migratetype_order(order, t) {
5409 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5410 zone->free_area[order].nr_free = 0;
5414 #ifndef __HAVE_ARCH_MEMMAP_INIT
5415 #define memmap_init(size, nid, zone, start_pfn) \
5416 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5419 static int zone_batchsize(struct zone *zone)
5425 * The per-cpu-pages pools are set to around 1000th of the
5426 * size of the zone. But no more than 1/2 of a meg.
5428 * OK, so we don't know how big the cache is. So guess.
5430 batch = zone->managed_pages / 1024;
5431 if (batch * PAGE_SIZE > 512 * 1024)
5432 batch = (512 * 1024) / PAGE_SIZE;
5433 batch /= 4; /* We effectively *= 4 below */
5438 * Clamp the batch to a 2^n - 1 value. Having a power
5439 * of 2 value was found to be more likely to have
5440 * suboptimal cache aliasing properties in some cases.
5442 * For example if 2 tasks are alternately allocating
5443 * batches of pages, one task can end up with a lot
5444 * of pages of one half of the possible page colors
5445 * and the other with pages of the other colors.
5447 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5452 /* The deferral and batching of frees should be suppressed under NOMMU
5455 * The problem is that NOMMU needs to be able to allocate large chunks
5456 * of contiguous memory as there's no hardware page translation to
5457 * assemble apparent contiguous memory from discontiguous pages.
5459 * Queueing large contiguous runs of pages for batching, however,
5460 * causes the pages to actually be freed in smaller chunks. As there
5461 * can be a significant delay between the individual batches being
5462 * recycled, this leads to the once large chunks of space being
5463 * fragmented and becoming unavailable for high-order allocations.
5470 * pcp->high and pcp->batch values are related and dependent on one another:
5471 * ->batch must never be higher then ->high.
5472 * The following function updates them in a safe manner without read side
5475 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5476 * those fields changing asynchronously (acording the the above rule).
5478 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5479 * outside of boot time (or some other assurance that no concurrent updaters
5482 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5483 unsigned long batch)
5485 /* start with a fail safe value for batch */
5489 /* Update high, then batch, in order */
5496 /* a companion to pageset_set_high() */
5497 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5499 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5502 static void pageset_init(struct per_cpu_pageset *p)
5504 struct per_cpu_pages *pcp;
5507 memset(p, 0, sizeof(*p));
5511 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5512 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5515 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5518 pageset_set_batch(p, batch);
5522 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5523 * to the value high for the pageset p.
5525 static void pageset_set_high(struct per_cpu_pageset *p,
5528 unsigned long batch = max(1UL, high / 4);
5529 if ((high / 4) > (PAGE_SHIFT * 8))
5530 batch = PAGE_SHIFT * 8;
5532 pageset_update(&p->pcp, high, batch);
5535 static void pageset_set_high_and_batch(struct zone *zone,
5536 struct per_cpu_pageset *pcp)
5538 if (percpu_pagelist_fraction)
5539 pageset_set_high(pcp,
5540 (zone->managed_pages /
5541 percpu_pagelist_fraction));
5543 pageset_set_batch(pcp, zone_batchsize(zone));
5546 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5548 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5551 pageset_set_high_and_batch(zone, pcp);
5554 void __meminit setup_zone_pageset(struct zone *zone)
5557 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5558 for_each_possible_cpu(cpu)
5559 zone_pageset_init(zone, cpu);
5563 * Allocate per cpu pagesets and initialize them.
5564 * Before this call only boot pagesets were available.
5566 void __init setup_per_cpu_pageset(void)
5568 struct pglist_data *pgdat;
5571 for_each_populated_zone(zone)
5572 setup_zone_pageset(zone);
5574 for_each_online_pgdat(pgdat)
5575 pgdat->per_cpu_nodestats =
5576 alloc_percpu(struct per_cpu_nodestat);
5579 static __meminit void zone_pcp_init(struct zone *zone)
5582 * per cpu subsystem is not up at this point. The following code
5583 * relies on the ability of the linker to provide the
5584 * offset of a (static) per cpu variable into the per cpu area.
5586 zone->pageset = &boot_pageset;
5588 if (populated_zone(zone))
5589 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5590 zone->name, zone->present_pages,
5591 zone_batchsize(zone));
5594 void __meminit init_currently_empty_zone(struct zone *zone,
5595 unsigned long zone_start_pfn,
5598 struct pglist_data *pgdat = zone->zone_pgdat;
5599 int zone_idx = zone_idx(zone) + 1;
5601 if (zone_idx > pgdat->nr_zones)
5602 pgdat->nr_zones = zone_idx;
5604 zone->zone_start_pfn = zone_start_pfn;
5606 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5607 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5609 (unsigned long)zone_idx(zone),
5610 zone_start_pfn, (zone_start_pfn + size));
5612 zone_init_free_lists(zone);
5613 zone->initialized = 1;
5616 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5617 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5620 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5622 int __meminit __early_pfn_to_nid(unsigned long pfn,
5623 struct mminit_pfnnid_cache *state)
5625 unsigned long start_pfn, end_pfn;
5628 if (state->last_start <= pfn && pfn < state->last_end)
5629 return state->last_nid;
5631 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5633 state->last_start = start_pfn;
5634 state->last_end = end_pfn;
5635 state->last_nid = nid;
5640 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5643 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5644 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5645 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5647 * If an architecture guarantees that all ranges registered contain no holes
5648 * and may be freed, this this function may be used instead of calling
5649 * memblock_free_early_nid() manually.
5651 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5653 unsigned long start_pfn, end_pfn;
5656 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5657 start_pfn = min(start_pfn, max_low_pfn);
5658 end_pfn = min(end_pfn, max_low_pfn);
5660 if (start_pfn < end_pfn)
5661 memblock_free_early_nid(PFN_PHYS(start_pfn),
5662 (end_pfn - start_pfn) << PAGE_SHIFT,
5668 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5669 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5671 * If an architecture guarantees that all ranges registered contain no holes and may
5672 * be freed, this function may be used instead of calling memory_present() manually.
5674 void __init sparse_memory_present_with_active_regions(int nid)
5676 unsigned long start_pfn, end_pfn;
5679 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5680 memory_present(this_nid, start_pfn, end_pfn);
5684 * get_pfn_range_for_nid - Return the start and end page frames for a node
5685 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5686 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5687 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5689 * It returns the start and end page frame of a node based on information
5690 * provided by memblock_set_node(). If called for a node
5691 * with no available memory, a warning is printed and the start and end
5694 void __meminit get_pfn_range_for_nid(unsigned int nid,
5695 unsigned long *start_pfn, unsigned long *end_pfn)
5697 unsigned long this_start_pfn, this_end_pfn;
5703 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5704 *start_pfn = min(*start_pfn, this_start_pfn);
5705 *end_pfn = max(*end_pfn, this_end_pfn);
5708 if (*start_pfn == -1UL)
5713 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5714 * assumption is made that zones within a node are ordered in monotonic
5715 * increasing memory addresses so that the "highest" populated zone is used
5717 static void __init find_usable_zone_for_movable(void)
5720 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5721 if (zone_index == ZONE_MOVABLE)
5724 if (arch_zone_highest_possible_pfn[zone_index] >
5725 arch_zone_lowest_possible_pfn[zone_index])
5729 VM_BUG_ON(zone_index == -1);
5730 movable_zone = zone_index;
5734 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5735 * because it is sized independent of architecture. Unlike the other zones,
5736 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5737 * in each node depending on the size of each node and how evenly kernelcore
5738 * is distributed. This helper function adjusts the zone ranges
5739 * provided by the architecture for a given node by using the end of the
5740 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5741 * zones within a node are in order of monotonic increases memory addresses
5743 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5744 unsigned long zone_type,
5745 unsigned long node_start_pfn,
5746 unsigned long node_end_pfn,
5747 unsigned long *zone_start_pfn,
5748 unsigned long *zone_end_pfn)
5750 /* Only adjust if ZONE_MOVABLE is on this node */
5751 if (zone_movable_pfn[nid]) {
5752 /* Size ZONE_MOVABLE */
5753 if (zone_type == ZONE_MOVABLE) {
5754 *zone_start_pfn = zone_movable_pfn[nid];
5755 *zone_end_pfn = min(node_end_pfn,
5756 arch_zone_highest_possible_pfn[movable_zone]);
5758 /* Adjust for ZONE_MOVABLE starting within this range */
5759 } else if (!mirrored_kernelcore &&
5760 *zone_start_pfn < zone_movable_pfn[nid] &&
5761 *zone_end_pfn > zone_movable_pfn[nid]) {
5762 *zone_end_pfn = zone_movable_pfn[nid];
5764 /* Check if this whole range is within ZONE_MOVABLE */
5765 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5766 *zone_start_pfn = *zone_end_pfn;
5771 * Return the number of pages a zone spans in a node, including holes
5772 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5774 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5775 unsigned long zone_type,
5776 unsigned long node_start_pfn,
5777 unsigned long node_end_pfn,
5778 unsigned long *zone_start_pfn,
5779 unsigned long *zone_end_pfn,
5780 unsigned long *ignored)
5782 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5783 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5784 /* When hotadd a new node from cpu_up(), the node should be empty */
5785 if (!node_start_pfn && !node_end_pfn)
5788 /* Get the start and end of the zone */
5789 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5790 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5791 adjust_zone_range_for_zone_movable(nid, zone_type,
5792 node_start_pfn, node_end_pfn,
5793 zone_start_pfn, zone_end_pfn);
5795 /* Check that this node has pages within the zone's required range */
5796 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5799 /* Move the zone boundaries inside the node if necessary */
5800 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5801 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5803 /* Return the spanned pages */
5804 return *zone_end_pfn - *zone_start_pfn;
5808 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5809 * then all holes in the requested range will be accounted for.
5811 unsigned long __meminit __absent_pages_in_range(int nid,
5812 unsigned long range_start_pfn,
5813 unsigned long range_end_pfn)
5815 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5816 unsigned long start_pfn, end_pfn;
5819 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5820 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5821 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5822 nr_absent -= end_pfn - start_pfn;
5828 * absent_pages_in_range - Return number of page frames in holes within a range
5829 * @start_pfn: The start PFN to start searching for holes
5830 * @end_pfn: The end PFN to stop searching for holes
5832 * It returns the number of pages frames in memory holes within a range.
5834 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5835 unsigned long end_pfn)
5837 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5840 /* Return the number of page frames in holes in a zone on a node */
5841 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5842 unsigned long zone_type,
5843 unsigned long node_start_pfn,
5844 unsigned long node_end_pfn,
5845 unsigned long *ignored)
5847 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5848 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5849 unsigned long zone_start_pfn, zone_end_pfn;
5850 unsigned long nr_absent;
5852 /* When hotadd a new node from cpu_up(), the node should be empty */
5853 if (!node_start_pfn && !node_end_pfn)
5856 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5857 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5859 adjust_zone_range_for_zone_movable(nid, zone_type,
5860 node_start_pfn, node_end_pfn,
5861 &zone_start_pfn, &zone_end_pfn);
5862 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5865 * ZONE_MOVABLE handling.
5866 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5869 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5870 unsigned long start_pfn, end_pfn;
5871 struct memblock_region *r;
5873 for_each_memblock(memory, r) {
5874 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5875 zone_start_pfn, zone_end_pfn);
5876 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5877 zone_start_pfn, zone_end_pfn);
5879 if (zone_type == ZONE_MOVABLE &&
5880 memblock_is_mirror(r))
5881 nr_absent += end_pfn - start_pfn;
5883 if (zone_type == ZONE_NORMAL &&
5884 !memblock_is_mirror(r))
5885 nr_absent += end_pfn - start_pfn;
5892 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5893 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5894 unsigned long zone_type,
5895 unsigned long node_start_pfn,
5896 unsigned long node_end_pfn,
5897 unsigned long *zone_start_pfn,
5898 unsigned long *zone_end_pfn,
5899 unsigned long *zones_size)
5903 *zone_start_pfn = node_start_pfn;
5904 for (zone = 0; zone < zone_type; zone++)
5905 *zone_start_pfn += zones_size[zone];
5907 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5909 return zones_size[zone_type];
5912 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5913 unsigned long zone_type,
5914 unsigned long node_start_pfn,
5915 unsigned long node_end_pfn,
5916 unsigned long *zholes_size)
5921 return zholes_size[zone_type];
5924 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5926 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5927 unsigned long node_start_pfn,
5928 unsigned long node_end_pfn,
5929 unsigned long *zones_size,
5930 unsigned long *zholes_size)
5932 unsigned long realtotalpages = 0, totalpages = 0;
5935 for (i = 0; i < MAX_NR_ZONES; i++) {
5936 struct zone *zone = pgdat->node_zones + i;
5937 unsigned long zone_start_pfn, zone_end_pfn;
5938 unsigned long size, real_size;
5940 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5946 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5947 node_start_pfn, node_end_pfn,
5950 zone->zone_start_pfn = zone_start_pfn;
5952 zone->zone_start_pfn = 0;
5953 zone->spanned_pages = size;
5954 zone->present_pages = real_size;
5957 realtotalpages += real_size;
5960 pgdat->node_spanned_pages = totalpages;
5961 pgdat->node_present_pages = realtotalpages;
5962 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5966 #ifndef CONFIG_SPARSEMEM
5968 * Calculate the size of the zone->blockflags rounded to an unsigned long
5969 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5970 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5971 * round what is now in bits to nearest long in bits, then return it in
5974 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5976 unsigned long usemapsize;
5978 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5979 usemapsize = roundup(zonesize, pageblock_nr_pages);
5980 usemapsize = usemapsize >> pageblock_order;
5981 usemapsize *= NR_PAGEBLOCK_BITS;
5982 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5984 return usemapsize / 8;
5987 static void __init setup_usemap(struct pglist_data *pgdat,
5989 unsigned long zone_start_pfn,
5990 unsigned long zonesize)
5992 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5993 zone->pageblock_flags = NULL;
5995 zone->pageblock_flags =
5996 memblock_virt_alloc_node_nopanic(usemapsize,
6000 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6001 unsigned long zone_start_pfn, unsigned long zonesize) {}
6002 #endif /* CONFIG_SPARSEMEM */
6004 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6006 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6007 void __paginginit set_pageblock_order(void)
6011 /* Check that pageblock_nr_pages has not already been setup */
6012 if (pageblock_order)
6015 if (HPAGE_SHIFT > PAGE_SHIFT)
6016 order = HUGETLB_PAGE_ORDER;
6018 order = MAX_ORDER - 1;
6021 * Assume the largest contiguous order of interest is a huge page.
6022 * This value may be variable depending on boot parameters on IA64 and
6025 pageblock_order = order;
6027 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6030 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6031 * is unused as pageblock_order is set at compile-time. See
6032 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6035 void __paginginit set_pageblock_order(void)
6039 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6041 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6042 unsigned long present_pages)
6044 unsigned long pages = spanned_pages;
6047 * Provide a more accurate estimation if there are holes within
6048 * the zone and SPARSEMEM is in use. If there are holes within the
6049 * zone, each populated memory region may cost us one or two extra
6050 * memmap pages due to alignment because memmap pages for each
6051 * populated regions may not be naturally aligned on page boundary.
6052 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6054 if (spanned_pages > present_pages + (present_pages >> 4) &&
6055 IS_ENABLED(CONFIG_SPARSEMEM))
6056 pages = present_pages;
6058 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6062 * Set up the zone data structures:
6063 * - mark all pages reserved
6064 * - mark all memory queues empty
6065 * - clear the memory bitmaps
6067 * NOTE: pgdat should get zeroed by caller.
6069 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6072 int nid = pgdat->node_id;
6074 pgdat_resize_init(pgdat);
6075 #ifdef CONFIG_NUMA_BALANCING
6076 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6077 pgdat->numabalancing_migrate_nr_pages = 0;
6078 pgdat->numabalancing_migrate_next_window = jiffies;
6080 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6081 spin_lock_init(&pgdat->split_queue_lock);
6082 INIT_LIST_HEAD(&pgdat->split_queue);
6083 pgdat->split_queue_len = 0;
6085 init_waitqueue_head(&pgdat->kswapd_wait);
6086 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6087 #ifdef CONFIG_COMPACTION
6088 init_waitqueue_head(&pgdat->kcompactd_wait);
6090 pgdat_page_ext_init(pgdat);
6091 spin_lock_init(&pgdat->lru_lock);
6092 lruvec_init(node_lruvec(pgdat));
6094 pgdat->per_cpu_nodestats = &boot_nodestats;
6096 for (j = 0; j < MAX_NR_ZONES; j++) {
6097 struct zone *zone = pgdat->node_zones + j;
6098 unsigned long size, realsize, freesize, memmap_pages;
6099 unsigned long zone_start_pfn = zone->zone_start_pfn;
6101 size = zone->spanned_pages;
6102 realsize = freesize = zone->present_pages;
6105 * Adjust freesize so that it accounts for how much memory
6106 * is used by this zone for memmap. This affects the watermark
6107 * and per-cpu initialisations
6109 memmap_pages = calc_memmap_size(size, realsize);
6110 if (!is_highmem_idx(j)) {
6111 if (freesize >= memmap_pages) {
6112 freesize -= memmap_pages;
6115 " %s zone: %lu pages used for memmap\n",
6116 zone_names[j], memmap_pages);
6118 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6119 zone_names[j], memmap_pages, freesize);
6122 /* Account for reserved pages */
6123 if (j == 0 && freesize > dma_reserve) {
6124 freesize -= dma_reserve;
6125 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6126 zone_names[0], dma_reserve);
6129 if (!is_highmem_idx(j))
6130 nr_kernel_pages += freesize;
6131 /* Charge for highmem memmap if there are enough kernel pages */
6132 else if (nr_kernel_pages > memmap_pages * 2)
6133 nr_kernel_pages -= memmap_pages;
6134 nr_all_pages += freesize;
6137 * Set an approximate value for lowmem here, it will be adjusted
6138 * when the bootmem allocator frees pages into the buddy system.
6139 * And all highmem pages will be managed by the buddy system.
6141 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6145 zone->name = zone_names[j];
6146 zone->zone_pgdat = pgdat;
6147 spin_lock_init(&zone->lock);
6148 zone_seqlock_init(zone);
6149 zone_pcp_init(zone);
6154 set_pageblock_order();
6155 setup_usemap(pgdat, zone, zone_start_pfn, size);
6156 init_currently_empty_zone(zone, zone_start_pfn, size);
6157 memmap_init(size, nid, j, zone_start_pfn);
6161 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6163 unsigned long __maybe_unused start = 0;
6164 unsigned long __maybe_unused offset = 0;
6166 /* Skip empty nodes */
6167 if (!pgdat->node_spanned_pages)
6170 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6171 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6172 offset = pgdat->node_start_pfn - start;
6173 /* ia64 gets its own node_mem_map, before this, without bootmem */
6174 if (!pgdat->node_mem_map) {
6175 unsigned long size, end;
6179 * The zone's endpoints aren't required to be MAX_ORDER
6180 * aligned but the node_mem_map endpoints must be in order
6181 * for the buddy allocator to function correctly.
6183 end = pgdat_end_pfn(pgdat);
6184 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6185 size = (end - start) * sizeof(struct page);
6186 map = alloc_remap(pgdat->node_id, size);
6188 map = memblock_virt_alloc_node_nopanic(size,
6190 pgdat->node_mem_map = map + offset;
6192 #ifndef CONFIG_NEED_MULTIPLE_NODES
6194 * With no DISCONTIG, the global mem_map is just set as node 0's
6196 if (pgdat == NODE_DATA(0)) {
6197 mem_map = NODE_DATA(0)->node_mem_map;
6198 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6199 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6201 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6204 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6207 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6208 unsigned long node_start_pfn, unsigned long *zholes_size)
6210 pg_data_t *pgdat = NODE_DATA(nid);
6211 unsigned long start_pfn = 0;
6212 unsigned long end_pfn = 0;
6214 /* pg_data_t should be reset to zero when it's allocated */
6215 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6217 pgdat->node_id = nid;
6218 pgdat->node_start_pfn = node_start_pfn;
6219 pgdat->per_cpu_nodestats = NULL;
6220 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6221 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6222 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6223 (u64)start_pfn << PAGE_SHIFT,
6224 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6226 start_pfn = node_start_pfn;
6228 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6229 zones_size, zholes_size);
6231 alloc_node_mem_map(pgdat);
6232 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6233 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6234 nid, (unsigned long)pgdat,
6235 (unsigned long)pgdat->node_mem_map);
6238 reset_deferred_meminit(pgdat);
6239 free_area_init_core(pgdat);
6242 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6244 #if MAX_NUMNODES > 1
6246 * Figure out the number of possible node ids.
6248 void __init setup_nr_node_ids(void)
6250 unsigned int highest;
6252 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6253 nr_node_ids = highest + 1;
6258 * node_map_pfn_alignment - determine the maximum internode alignment
6260 * This function should be called after node map is populated and sorted.
6261 * It calculates the maximum power of two alignment which can distinguish
6264 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6265 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6266 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6267 * shifted, 1GiB is enough and this function will indicate so.
6269 * This is used to test whether pfn -> nid mapping of the chosen memory
6270 * model has fine enough granularity to avoid incorrect mapping for the
6271 * populated node map.
6273 * Returns the determined alignment in pfn's. 0 if there is no alignment
6274 * requirement (single node).
6276 unsigned long __init node_map_pfn_alignment(void)
6278 unsigned long accl_mask = 0, last_end = 0;
6279 unsigned long start, end, mask;
6283 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6284 if (!start || last_nid < 0 || last_nid == nid) {
6291 * Start with a mask granular enough to pin-point to the
6292 * start pfn and tick off bits one-by-one until it becomes
6293 * too coarse to separate the current node from the last.
6295 mask = ~((1 << __ffs(start)) - 1);
6296 while (mask && last_end <= (start & (mask << 1)))
6299 /* accumulate all internode masks */
6303 /* convert mask to number of pages */
6304 return ~accl_mask + 1;
6307 /* Find the lowest pfn for a node */
6308 static unsigned long __init find_min_pfn_for_node(int nid)
6310 unsigned long min_pfn = ULONG_MAX;
6311 unsigned long start_pfn;
6314 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6315 min_pfn = min(min_pfn, start_pfn);
6317 if (min_pfn == ULONG_MAX) {
6318 pr_warn("Could not find start_pfn for node %d\n", nid);
6326 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6328 * It returns the minimum PFN based on information provided via
6329 * memblock_set_node().
6331 unsigned long __init find_min_pfn_with_active_regions(void)
6333 return find_min_pfn_for_node(MAX_NUMNODES);
6337 * early_calculate_totalpages()
6338 * Sum pages in active regions for movable zone.
6339 * Populate N_MEMORY for calculating usable_nodes.
6341 static unsigned long __init early_calculate_totalpages(void)
6343 unsigned long totalpages = 0;
6344 unsigned long start_pfn, end_pfn;
6347 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6348 unsigned long pages = end_pfn - start_pfn;
6350 totalpages += pages;
6352 node_set_state(nid, N_MEMORY);
6358 * Find the PFN the Movable zone begins in each node. Kernel memory
6359 * is spread evenly between nodes as long as the nodes have enough
6360 * memory. When they don't, some nodes will have more kernelcore than
6363 static void __init find_zone_movable_pfns_for_nodes(void)
6366 unsigned long usable_startpfn;
6367 unsigned long kernelcore_node, kernelcore_remaining;
6368 /* save the state before borrow the nodemask */
6369 nodemask_t saved_node_state = node_states[N_MEMORY];
6370 unsigned long totalpages = early_calculate_totalpages();
6371 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6372 struct memblock_region *r;
6374 /* Need to find movable_zone earlier when movable_node is specified. */
6375 find_usable_zone_for_movable();
6378 * If movable_node is specified, ignore kernelcore and movablecore
6381 if (movable_node_is_enabled()) {
6382 for_each_memblock(memory, r) {
6383 if (!memblock_is_hotpluggable(r))
6388 usable_startpfn = PFN_DOWN(r->base);
6389 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6390 min(usable_startpfn, zone_movable_pfn[nid]) :
6398 * If kernelcore=mirror is specified, ignore movablecore option
6400 if (mirrored_kernelcore) {
6401 bool mem_below_4gb_not_mirrored = false;
6403 for_each_memblock(memory, r) {
6404 if (memblock_is_mirror(r))
6409 usable_startpfn = memblock_region_memory_base_pfn(r);
6411 if (usable_startpfn < 0x100000) {
6412 mem_below_4gb_not_mirrored = true;
6416 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6417 min(usable_startpfn, zone_movable_pfn[nid]) :
6421 if (mem_below_4gb_not_mirrored)
6422 pr_warn("This configuration results in unmirrored kernel memory.");
6428 * If movablecore=nn[KMG] was specified, calculate what size of
6429 * kernelcore that corresponds so that memory usable for
6430 * any allocation type is evenly spread. If both kernelcore
6431 * and movablecore are specified, then the value of kernelcore
6432 * will be used for required_kernelcore if it's greater than
6433 * what movablecore would have allowed.
6435 if (required_movablecore) {
6436 unsigned long corepages;
6439 * Round-up so that ZONE_MOVABLE is at least as large as what
6440 * was requested by the user
6442 required_movablecore =
6443 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6444 required_movablecore = min(totalpages, required_movablecore);
6445 corepages = totalpages - required_movablecore;
6447 required_kernelcore = max(required_kernelcore, corepages);
6451 * If kernelcore was not specified or kernelcore size is larger
6452 * than totalpages, there is no ZONE_MOVABLE.
6454 if (!required_kernelcore || required_kernelcore >= totalpages)
6457 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6458 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6461 /* Spread kernelcore memory as evenly as possible throughout nodes */
6462 kernelcore_node = required_kernelcore / usable_nodes;
6463 for_each_node_state(nid, N_MEMORY) {
6464 unsigned long start_pfn, end_pfn;
6467 * Recalculate kernelcore_node if the division per node
6468 * now exceeds what is necessary to satisfy the requested
6469 * amount of memory for the kernel
6471 if (required_kernelcore < kernelcore_node)
6472 kernelcore_node = required_kernelcore / usable_nodes;
6475 * As the map is walked, we track how much memory is usable
6476 * by the kernel using kernelcore_remaining. When it is
6477 * 0, the rest of the node is usable by ZONE_MOVABLE
6479 kernelcore_remaining = kernelcore_node;
6481 /* Go through each range of PFNs within this node */
6482 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6483 unsigned long size_pages;
6485 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6486 if (start_pfn >= end_pfn)
6489 /* Account for what is only usable for kernelcore */
6490 if (start_pfn < usable_startpfn) {
6491 unsigned long kernel_pages;
6492 kernel_pages = min(end_pfn, usable_startpfn)
6495 kernelcore_remaining -= min(kernel_pages,
6496 kernelcore_remaining);
6497 required_kernelcore -= min(kernel_pages,
6498 required_kernelcore);
6500 /* Continue if range is now fully accounted */
6501 if (end_pfn <= usable_startpfn) {
6504 * Push zone_movable_pfn to the end so
6505 * that if we have to rebalance
6506 * kernelcore across nodes, we will
6507 * not double account here
6509 zone_movable_pfn[nid] = end_pfn;
6512 start_pfn = usable_startpfn;
6516 * The usable PFN range for ZONE_MOVABLE is from
6517 * start_pfn->end_pfn. Calculate size_pages as the
6518 * number of pages used as kernelcore
6520 size_pages = end_pfn - start_pfn;
6521 if (size_pages > kernelcore_remaining)
6522 size_pages = kernelcore_remaining;
6523 zone_movable_pfn[nid] = start_pfn + size_pages;
6526 * Some kernelcore has been met, update counts and
6527 * break if the kernelcore for this node has been
6530 required_kernelcore -= min(required_kernelcore,
6532 kernelcore_remaining -= size_pages;
6533 if (!kernelcore_remaining)
6539 * If there is still required_kernelcore, we do another pass with one
6540 * less node in the count. This will push zone_movable_pfn[nid] further
6541 * along on the nodes that still have memory until kernelcore is
6545 if (usable_nodes && required_kernelcore > usable_nodes)
6549 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6550 for (nid = 0; nid < MAX_NUMNODES; nid++) {
6551 unsigned long start_pfn, end_pfn;
6553 zone_movable_pfn[nid] =
6554 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6556 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6557 if (zone_movable_pfn[nid] >= end_pfn)
6558 zone_movable_pfn[nid] = 0;
6562 /* restore the node_state */
6563 node_states[N_MEMORY] = saved_node_state;
6566 /* Any regular or high memory on that node ? */
6567 static void check_for_memory(pg_data_t *pgdat, int nid)
6569 enum zone_type zone_type;
6571 if (N_MEMORY == N_NORMAL_MEMORY)
6574 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6575 struct zone *zone = &pgdat->node_zones[zone_type];
6576 if (populated_zone(zone)) {
6577 node_set_state(nid, N_HIGH_MEMORY);
6578 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6579 zone_type <= ZONE_NORMAL)
6580 node_set_state(nid, N_NORMAL_MEMORY);
6587 * free_area_init_nodes - Initialise all pg_data_t and zone data
6588 * @max_zone_pfn: an array of max PFNs for each zone
6590 * This will call free_area_init_node() for each active node in the system.
6591 * Using the page ranges provided by memblock_set_node(), the size of each
6592 * zone in each node and their holes is calculated. If the maximum PFN
6593 * between two adjacent zones match, it is assumed that the zone is empty.
6594 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6595 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6596 * starts where the previous one ended. For example, ZONE_DMA32 starts
6597 * at arch_max_dma_pfn.
6599 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6601 unsigned long start_pfn, end_pfn;
6604 /* Record where the zone boundaries are */
6605 memset(arch_zone_lowest_possible_pfn, 0,
6606 sizeof(arch_zone_lowest_possible_pfn));
6607 memset(arch_zone_highest_possible_pfn, 0,
6608 sizeof(arch_zone_highest_possible_pfn));
6610 start_pfn = find_min_pfn_with_active_regions();
6612 for (i = 0; i < MAX_NR_ZONES; i++) {
6613 if (i == ZONE_MOVABLE)
6616 end_pfn = max(max_zone_pfn[i], start_pfn);
6617 arch_zone_lowest_possible_pfn[i] = start_pfn;
6618 arch_zone_highest_possible_pfn[i] = end_pfn;
6620 start_pfn = end_pfn;
6623 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6624 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6625 find_zone_movable_pfns_for_nodes();
6627 /* Print out the zone ranges */
6628 pr_info("Zone ranges:\n");
6629 for (i = 0; i < MAX_NR_ZONES; i++) {
6630 if (i == ZONE_MOVABLE)
6632 pr_info(" %-8s ", zone_names[i]);
6633 if (arch_zone_lowest_possible_pfn[i] ==
6634 arch_zone_highest_possible_pfn[i])
6637 pr_cont("[mem %#018Lx-%#018Lx]\n",
6638 (u64)arch_zone_lowest_possible_pfn[i]
6640 ((u64)arch_zone_highest_possible_pfn[i]
6641 << PAGE_SHIFT) - 1);
6644 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6645 pr_info("Movable zone start for each node\n");
6646 for (i = 0; i < MAX_NUMNODES; i++) {
6647 if (zone_movable_pfn[i])
6648 pr_info(" Node %d: %#018Lx\n", i,
6649 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6652 /* Print out the early node map */
6653 pr_info("Early memory node ranges\n");
6654 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6655 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6656 (u64)start_pfn << PAGE_SHIFT,
6657 ((u64)end_pfn << PAGE_SHIFT) - 1);
6659 /* Initialise every node */
6660 mminit_verify_pageflags_layout();
6661 setup_nr_node_ids();
6662 for_each_online_node(nid) {
6663 pg_data_t *pgdat = NODE_DATA(nid);
6664 free_area_init_node(nid, NULL,
6665 find_min_pfn_for_node(nid), NULL);
6667 /* Any memory on that node */
6668 if (pgdat->node_present_pages)
6669 node_set_state(nid, N_MEMORY);
6670 check_for_memory(pgdat, nid);
6674 static int __init cmdline_parse_core(char *p, unsigned long *core)
6676 unsigned long long coremem;
6680 coremem = memparse(p, &p);
6681 *core = coremem >> PAGE_SHIFT;
6683 /* Paranoid check that UL is enough for the coremem value */
6684 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6690 * kernelcore=size sets the amount of memory for use for allocations that
6691 * cannot be reclaimed or migrated.
6693 static int __init cmdline_parse_kernelcore(char *p)
6695 /* parse kernelcore=mirror */
6696 if (parse_option_str(p, "mirror")) {
6697 mirrored_kernelcore = true;
6701 return cmdline_parse_core(p, &required_kernelcore);
6705 * movablecore=size sets the amount of memory for use for allocations that
6706 * can be reclaimed or migrated.
6708 static int __init cmdline_parse_movablecore(char *p)
6710 return cmdline_parse_core(p, &required_movablecore);
6713 early_param("kernelcore", cmdline_parse_kernelcore);
6714 early_param("movablecore", cmdline_parse_movablecore);
6716 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6718 void adjust_managed_page_count(struct page *page, long count)
6720 spin_lock(&managed_page_count_lock);
6721 page_zone(page)->managed_pages += count;
6722 totalram_pages += count;
6723 #ifdef CONFIG_HIGHMEM
6724 if (PageHighMem(page))
6725 totalhigh_pages += count;
6727 spin_unlock(&managed_page_count_lock);
6729 EXPORT_SYMBOL(adjust_managed_page_count);
6731 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6734 unsigned long pages = 0;
6736 start = (void *)PAGE_ALIGN((unsigned long)start);
6737 end = (void *)((unsigned long)end & PAGE_MASK);
6738 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6739 if ((unsigned int)poison <= 0xFF)
6740 memset(pos, poison, PAGE_SIZE);
6741 free_reserved_page(virt_to_page(pos));
6745 pr_info("Freeing %s memory: %ldK\n",
6746 s, pages << (PAGE_SHIFT - 10));
6750 EXPORT_SYMBOL(free_reserved_area);
6752 #ifdef CONFIG_HIGHMEM
6753 void free_highmem_page(struct page *page)
6755 __free_reserved_page(page);
6757 page_zone(page)->managed_pages++;
6763 void __init mem_init_print_info(const char *str)
6765 unsigned long physpages, codesize, datasize, rosize, bss_size;
6766 unsigned long init_code_size, init_data_size;
6768 physpages = get_num_physpages();
6769 codesize = _etext - _stext;
6770 datasize = _edata - _sdata;
6771 rosize = __end_rodata - __start_rodata;
6772 bss_size = __bss_stop - __bss_start;
6773 init_data_size = __init_end - __init_begin;
6774 init_code_size = _einittext - _sinittext;
6777 * Detect special cases and adjust section sizes accordingly:
6778 * 1) .init.* may be embedded into .data sections
6779 * 2) .init.text.* may be out of [__init_begin, __init_end],
6780 * please refer to arch/tile/kernel/vmlinux.lds.S.
6781 * 3) .rodata.* may be embedded into .text or .data sections.
6783 #define adj_init_size(start, end, size, pos, adj) \
6785 if (&start[0] <= &pos[0] && &pos[0] < &end[0] && size > adj) \
6789 adj_init_size(__init_begin, __init_end, init_data_size,
6790 _sinittext, init_code_size);
6791 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6792 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6793 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6794 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6796 #undef adj_init_size
6798 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6799 #ifdef CONFIG_HIGHMEM
6803 nr_free_pages() << (PAGE_SHIFT - 10),
6804 physpages << (PAGE_SHIFT - 10),
6805 codesize >> 10, datasize >> 10, rosize >> 10,
6806 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6807 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6808 totalcma_pages << (PAGE_SHIFT - 10),
6809 #ifdef CONFIG_HIGHMEM
6810 totalhigh_pages << (PAGE_SHIFT - 10),
6812 str ? ", " : "", str ? str : "");
6816 * set_dma_reserve - set the specified number of pages reserved in the first zone
6817 * @new_dma_reserve: The number of pages to mark reserved
6819 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6820 * In the DMA zone, a significant percentage may be consumed by kernel image
6821 * and other unfreeable allocations which can skew the watermarks badly. This
6822 * function may optionally be used to account for unfreeable pages in the
6823 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6824 * smaller per-cpu batchsize.
6826 void __init set_dma_reserve(unsigned long new_dma_reserve)
6828 dma_reserve = new_dma_reserve;
6831 void __init free_area_init(unsigned long *zones_size)
6833 free_area_init_node(0, zones_size,
6834 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6837 static int page_alloc_cpu_dead(unsigned int cpu)
6840 lru_add_drain_cpu(cpu);
6844 * Spill the event counters of the dead processor
6845 * into the current processors event counters.
6846 * This artificially elevates the count of the current
6849 vm_events_fold_cpu(cpu);
6852 * Zero the differential counters of the dead processor
6853 * so that the vm statistics are consistent.
6855 * This is only okay since the processor is dead and cannot
6856 * race with what we are doing.
6858 cpu_vm_stats_fold(cpu);
6862 void __init page_alloc_init(void)
6866 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6867 "mm/page_alloc:dead", NULL,
6868 page_alloc_cpu_dead);
6873 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6874 * or min_free_kbytes changes.
6876 static void calculate_totalreserve_pages(void)
6878 struct pglist_data *pgdat;
6879 unsigned long reserve_pages = 0;
6880 enum zone_type i, j;
6882 for_each_online_pgdat(pgdat) {
6884 pgdat->totalreserve_pages = 0;
6886 for (i = 0; i < MAX_NR_ZONES; i++) {
6887 struct zone *zone = pgdat->node_zones + i;
6890 /* Find valid and maximum lowmem_reserve in the zone */
6891 for (j = i; j < MAX_NR_ZONES; j++) {
6892 if (zone->lowmem_reserve[j] > max)
6893 max = zone->lowmem_reserve[j];
6896 /* we treat the high watermark as reserved pages. */
6897 max += high_wmark_pages(zone);
6899 if (max > zone->managed_pages)
6900 max = zone->managed_pages;
6902 pgdat->totalreserve_pages += max;
6904 reserve_pages += max;
6907 totalreserve_pages = reserve_pages;
6911 * setup_per_zone_lowmem_reserve - called whenever
6912 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6913 * has a correct pages reserved value, so an adequate number of
6914 * pages are left in the zone after a successful __alloc_pages().
6916 static void setup_per_zone_lowmem_reserve(void)
6918 struct pglist_data *pgdat;
6919 enum zone_type j, idx;
6921 for_each_online_pgdat(pgdat) {
6922 for (j = 0; j < MAX_NR_ZONES; j++) {
6923 struct zone *zone = pgdat->node_zones + j;
6924 unsigned long managed_pages = zone->managed_pages;
6926 zone->lowmem_reserve[j] = 0;
6930 struct zone *lower_zone;
6934 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6935 sysctl_lowmem_reserve_ratio[idx] = 1;
6937 lower_zone = pgdat->node_zones + idx;
6938 lower_zone->lowmem_reserve[j] = managed_pages /
6939 sysctl_lowmem_reserve_ratio[idx];
6940 managed_pages += lower_zone->managed_pages;
6945 /* update totalreserve_pages */
6946 calculate_totalreserve_pages();
6949 static void __setup_per_zone_wmarks(void)
6951 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6952 unsigned long lowmem_pages = 0;
6954 unsigned long flags;
6956 /* Calculate total number of !ZONE_HIGHMEM pages */
6957 for_each_zone(zone) {
6958 if (!is_highmem(zone))
6959 lowmem_pages += zone->managed_pages;
6962 for_each_zone(zone) {
6965 spin_lock_irqsave(&zone->lock, flags);
6966 tmp = (u64)pages_min * zone->managed_pages;
6967 do_div(tmp, lowmem_pages);
6968 if (is_highmem(zone)) {
6970 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6971 * need highmem pages, so cap pages_min to a small
6974 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6975 * deltas control asynch page reclaim, and so should
6976 * not be capped for highmem.
6978 unsigned long min_pages;
6980 min_pages = zone->managed_pages / 1024;
6981 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6982 zone->watermark[WMARK_MIN] = min_pages;
6985 * If it's a lowmem zone, reserve a number of pages
6986 * proportionate to the zone's size.
6988 zone->watermark[WMARK_MIN] = tmp;
6992 * Set the kswapd watermarks distance according to the
6993 * scale factor in proportion to available memory, but
6994 * ensure a minimum size on small systems.
6996 tmp = max_t(u64, tmp >> 2,
6997 mult_frac(zone->managed_pages,
6998 watermark_scale_factor, 10000));
7000 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7001 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7003 spin_unlock_irqrestore(&zone->lock, flags);
7006 /* update totalreserve_pages */
7007 calculate_totalreserve_pages();
7011 * setup_per_zone_wmarks - called when min_free_kbytes changes
7012 * or when memory is hot-{added|removed}
7014 * Ensures that the watermark[min,low,high] values for each zone are set
7015 * correctly with respect to min_free_kbytes.
7017 void setup_per_zone_wmarks(void)
7019 static DEFINE_SPINLOCK(lock);
7022 __setup_per_zone_wmarks();
7027 * Initialise min_free_kbytes.
7029 * For small machines we want it small (128k min). For large machines
7030 * we want it large (64MB max). But it is not linear, because network
7031 * bandwidth does not increase linearly with machine size. We use
7033 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7034 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7050 int __meminit init_per_zone_wmark_min(void)
7052 unsigned long lowmem_kbytes;
7053 int new_min_free_kbytes;
7055 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7056 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7058 if (new_min_free_kbytes > user_min_free_kbytes) {
7059 min_free_kbytes = new_min_free_kbytes;
7060 if (min_free_kbytes < 128)
7061 min_free_kbytes = 128;
7062 if (min_free_kbytes > 65536)
7063 min_free_kbytes = 65536;
7065 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7066 new_min_free_kbytes, user_min_free_kbytes);
7068 setup_per_zone_wmarks();
7069 refresh_zone_stat_thresholds();
7070 setup_per_zone_lowmem_reserve();
7073 setup_min_unmapped_ratio();
7074 setup_min_slab_ratio();
7077 khugepaged_min_free_kbytes_update();
7081 postcore_initcall(init_per_zone_wmark_min)
7084 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7085 * that we can call two helper functions whenever min_free_kbytes
7088 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7089 void __user *buffer, size_t *length, loff_t *ppos)
7093 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7098 user_min_free_kbytes = min_free_kbytes;
7099 setup_per_zone_wmarks();
7104 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7105 void __user *buffer, size_t *length, loff_t *ppos)
7109 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7114 setup_per_zone_wmarks();
7120 static void setup_min_unmapped_ratio(void)
7125 for_each_online_pgdat(pgdat)
7126 pgdat->min_unmapped_pages = 0;
7129 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7130 sysctl_min_unmapped_ratio) / 100;
7134 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7135 void __user *buffer, size_t *length, loff_t *ppos)
7139 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7143 setup_min_unmapped_ratio();
7148 static void setup_min_slab_ratio(void)
7153 for_each_online_pgdat(pgdat)
7154 pgdat->min_slab_pages = 0;
7157 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7158 sysctl_min_slab_ratio) / 100;
7161 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7162 void __user *buffer, size_t *length, loff_t *ppos)
7166 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7170 setup_min_slab_ratio();
7177 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7178 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7179 * whenever sysctl_lowmem_reserve_ratio changes.
7181 * The reserve ratio obviously has absolutely no relation with the
7182 * minimum watermarks. The lowmem reserve ratio can only make sense
7183 * if in function of the boot time zone sizes.
7185 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7186 void __user *buffer, size_t *length, loff_t *ppos)
7188 proc_dointvec_minmax(table, write, buffer, length, ppos);
7189 setup_per_zone_lowmem_reserve();
7194 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7195 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7196 * pagelist can have before it gets flushed back to buddy allocator.
7198 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7199 void __user *buffer, size_t *length, loff_t *ppos)
7202 int old_percpu_pagelist_fraction;
7205 mutex_lock(&pcp_batch_high_lock);
7206 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7208 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7209 if (!write || ret < 0)
7212 /* Sanity checking to avoid pcp imbalance */
7213 if (percpu_pagelist_fraction &&
7214 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7215 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7221 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7224 for_each_populated_zone(zone) {
7227 for_each_possible_cpu(cpu)
7228 pageset_set_high_and_batch(zone,
7229 per_cpu_ptr(zone->pageset, cpu));
7232 mutex_unlock(&pcp_batch_high_lock);
7237 int hashdist = HASHDIST_DEFAULT;
7239 static int __init set_hashdist(char *str)
7243 hashdist = simple_strtoul(str, &str, 0);
7246 __setup("hashdist=", set_hashdist);
7249 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7251 * Returns the number of pages that arch has reserved but
7252 * is not known to alloc_large_system_hash().
7254 static unsigned long __init arch_reserved_kernel_pages(void)
7261 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7262 * machines. As memory size is increased the scale is also increased but at
7263 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7264 * quadruples the scale is increased by one, which means the size of hash table
7265 * only doubles, instead of quadrupling as well.
7266 * Because 32-bit systems cannot have large physical memory, where this scaling
7267 * makes sense, it is disabled on such platforms.
7269 #if __BITS_PER_LONG > 32
7270 #define ADAPT_SCALE_BASE (64ul << 30)
7271 #define ADAPT_SCALE_SHIFT 2
7272 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7276 * allocate a large system hash table from bootmem
7277 * - it is assumed that the hash table must contain an exact power-of-2
7278 * quantity of entries
7279 * - limit is the number of hash buckets, not the total allocation size
7281 void *__init alloc_large_system_hash(const char *tablename,
7282 unsigned long bucketsize,
7283 unsigned long numentries,
7286 unsigned int *_hash_shift,
7287 unsigned int *_hash_mask,
7288 unsigned long low_limit,
7289 unsigned long high_limit)
7291 unsigned long long max = high_limit;
7292 unsigned long log2qty, size;
7296 /* allow the kernel cmdline to have a say */
7298 /* round applicable memory size up to nearest megabyte */
7299 numentries = nr_kernel_pages;
7300 numentries -= arch_reserved_kernel_pages();
7302 /* It isn't necessary when PAGE_SIZE >= 1MB */
7303 if (PAGE_SHIFT < 20)
7304 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7306 #if __BITS_PER_LONG > 32
7308 unsigned long adapt;
7310 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7311 adapt <<= ADAPT_SCALE_SHIFT)
7316 /* limit to 1 bucket per 2^scale bytes of low memory */
7317 if (scale > PAGE_SHIFT)
7318 numentries >>= (scale - PAGE_SHIFT);
7320 numentries <<= (PAGE_SHIFT - scale);
7322 /* Make sure we've got at least a 0-order allocation.. */
7323 if (unlikely(flags & HASH_SMALL)) {
7324 /* Makes no sense without HASH_EARLY */
7325 WARN_ON(!(flags & HASH_EARLY));
7326 if (!(numentries >> *_hash_shift)) {
7327 numentries = 1UL << *_hash_shift;
7328 BUG_ON(!numentries);
7330 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7331 numentries = PAGE_SIZE / bucketsize;
7333 numentries = roundup_pow_of_two(numentries);
7335 /* limit allocation size to 1/16 total memory by default */
7337 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7338 do_div(max, bucketsize);
7340 max = min(max, 0x80000000ULL);
7342 if (numentries < low_limit)
7343 numentries = low_limit;
7344 if (numentries > max)
7347 log2qty = ilog2(numentries);
7350 * memblock allocator returns zeroed memory already, so HASH_ZERO is
7351 * currently not used when HASH_EARLY is specified.
7353 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7355 size = bucketsize << log2qty;
7356 if (flags & HASH_EARLY)
7357 table = memblock_virt_alloc_nopanic(size, 0);
7359 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7362 * If bucketsize is not a power-of-two, we may free
7363 * some pages at the end of hash table which
7364 * alloc_pages_exact() automatically does
7366 if (get_order(size) < MAX_ORDER) {
7367 table = alloc_pages_exact(size, gfp_flags);
7368 kmemleak_alloc(table, size, 1, gfp_flags);
7371 } while (!table && size > PAGE_SIZE && --log2qty);
7374 panic("Failed to allocate %s hash table\n", tablename);
7376 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7377 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7380 *_hash_shift = log2qty;
7382 *_hash_mask = (1 << log2qty) - 1;
7388 * This function checks whether pageblock includes unmovable pages or not.
7389 * If @count is not zero, it is okay to include less @count unmovable pages
7391 * PageLRU check without isolation or lru_lock could race so that
7392 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7393 * check without lock_page also may miss some movable non-lru pages at
7394 * race condition. So you can't expect this function should be exact.
7396 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7397 bool skip_hwpoisoned_pages)
7399 unsigned long pfn, iter, found;
7403 * For avoiding noise data, lru_add_drain_all() should be called
7404 * If ZONE_MOVABLE, the zone never contains unmovable pages
7406 if (zone_idx(zone) == ZONE_MOVABLE)
7408 mt = get_pageblock_migratetype(page);
7409 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7412 pfn = page_to_pfn(page);
7413 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7414 unsigned long check = pfn + iter;
7416 if (!pfn_valid_within(check))
7419 page = pfn_to_page(check);
7422 * Hugepages are not in LRU lists, but they're movable.
7423 * We need not scan over tail pages bacause we don't
7424 * handle each tail page individually in migration.
7426 if (PageHuge(page)) {
7427 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7432 * We can't use page_count without pin a page
7433 * because another CPU can free compound page.
7434 * This check already skips compound tails of THP
7435 * because their page->_refcount is zero at all time.
7437 if (!page_ref_count(page)) {
7438 if (PageBuddy(page))
7439 iter += (1 << page_order(page)) - 1;
7444 * The HWPoisoned page may be not in buddy system, and
7445 * page_count() is not 0.
7447 if (skip_hwpoisoned_pages && PageHWPoison(page))
7450 if (__PageMovable(page))
7456 * If there are RECLAIMABLE pages, we need to check
7457 * it. But now, memory offline itself doesn't call
7458 * shrink_node_slabs() and it still to be fixed.
7461 * If the page is not RAM, page_count()should be 0.
7462 * we don't need more check. This is an _used_ not-movable page.
7464 * The problematic thing here is PG_reserved pages. PG_reserved
7465 * is set to both of a memory hole page and a _used_ kernel
7474 bool is_pageblock_removable_nolock(struct page *page)
7480 * We have to be careful here because we are iterating over memory
7481 * sections which are not zone aware so we might end up outside of
7482 * the zone but still within the section.
7483 * We have to take care about the node as well. If the node is offline
7484 * its NODE_DATA will be NULL - see page_zone.
7486 if (!node_online(page_to_nid(page)))
7489 zone = page_zone(page);
7490 pfn = page_to_pfn(page);
7491 if (!zone_spans_pfn(zone, pfn))
7494 return !has_unmovable_pages(zone, page, 0, true);
7497 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7499 static unsigned long pfn_max_align_down(unsigned long pfn)
7501 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7502 pageblock_nr_pages) - 1);
7505 static unsigned long pfn_max_align_up(unsigned long pfn)
7507 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7508 pageblock_nr_pages));
7511 /* [start, end) must belong to a single zone. */
7512 static int __alloc_contig_migrate_range(struct compact_control *cc,
7513 unsigned long start, unsigned long end)
7515 /* This function is based on compact_zone() from compaction.c. */
7516 unsigned long nr_reclaimed;
7517 unsigned long pfn = start;
7518 unsigned int tries = 0;
7523 while (pfn < end || !list_empty(&cc->migratepages)) {
7524 if (fatal_signal_pending(current)) {
7529 if (list_empty(&cc->migratepages)) {
7530 cc->nr_migratepages = 0;
7531 pfn = isolate_migratepages_range(cc, pfn, end);
7537 } else if (++tries == 5) {
7538 ret = ret < 0 ? ret : -EBUSY;
7542 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7544 cc->nr_migratepages -= nr_reclaimed;
7546 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7547 NULL, 0, cc->mode, MR_CMA);
7550 putback_movable_pages(&cc->migratepages);
7557 * alloc_contig_range() -- tries to allocate given range of pages
7558 * @start: start PFN to allocate
7559 * @end: one-past-the-last PFN to allocate
7560 * @migratetype: migratetype of the underlaying pageblocks (either
7561 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7562 * in range must have the same migratetype and it must
7563 * be either of the two.
7564 * @gfp_mask: GFP mask to use during compaction
7566 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7567 * aligned, however it's the caller's responsibility to guarantee that
7568 * we are the only thread that changes migrate type of pageblocks the
7571 * The PFN range must belong to a single zone.
7573 * Returns zero on success or negative error code. On success all
7574 * pages which PFN is in [start, end) are allocated for the caller and
7575 * need to be freed with free_contig_range().
7577 int alloc_contig_range(unsigned long start, unsigned long end,
7578 unsigned migratetype, gfp_t gfp_mask)
7580 unsigned long outer_start, outer_end;
7584 struct compact_control cc = {
7585 .nr_migratepages = 0,
7587 .zone = page_zone(pfn_to_page(start)),
7588 .mode = MIGRATE_SYNC,
7589 .ignore_skip_hint = true,
7590 .gfp_mask = current_gfp_context(gfp_mask),
7592 INIT_LIST_HEAD(&cc.migratepages);
7595 * What we do here is we mark all pageblocks in range as
7596 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7597 * have different sizes, and due to the way page allocator
7598 * work, we align the range to biggest of the two pages so
7599 * that page allocator won't try to merge buddies from
7600 * different pageblocks and change MIGRATE_ISOLATE to some
7601 * other migration type.
7603 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7604 * migrate the pages from an unaligned range (ie. pages that
7605 * we are interested in). This will put all the pages in
7606 * range back to page allocator as MIGRATE_ISOLATE.
7608 * When this is done, we take the pages in range from page
7609 * allocator removing them from the buddy system. This way
7610 * page allocator will never consider using them.
7612 * This lets us mark the pageblocks back as
7613 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7614 * aligned range but not in the unaligned, original range are
7615 * put back to page allocator so that buddy can use them.
7618 ret = start_isolate_page_range(pfn_max_align_down(start),
7619 pfn_max_align_up(end), migratetype,
7625 * In case of -EBUSY, we'd like to know which page causes problem.
7626 * So, just fall through. test_pages_isolated() has a tracepoint
7627 * which will report the busy page.
7629 * It is possible that busy pages could become available before
7630 * the call to test_pages_isolated, and the range will actually be
7631 * allocated. So, if we fall through be sure to clear ret so that
7632 * -EBUSY is not accidentally used or returned to caller.
7634 ret = __alloc_contig_migrate_range(&cc, start, end);
7635 if (ret && ret != -EBUSY)
7640 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7641 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7642 * more, all pages in [start, end) are free in page allocator.
7643 * What we are going to do is to allocate all pages from
7644 * [start, end) (that is remove them from page allocator).
7646 * The only problem is that pages at the beginning and at the
7647 * end of interesting range may be not aligned with pages that
7648 * page allocator holds, ie. they can be part of higher order
7649 * pages. Because of this, we reserve the bigger range and
7650 * once this is done free the pages we are not interested in.
7652 * We don't have to hold zone->lock here because the pages are
7653 * isolated thus they won't get removed from buddy.
7656 lru_add_drain_all();
7657 drain_all_pages(cc.zone);
7660 outer_start = start;
7661 while (!PageBuddy(pfn_to_page(outer_start))) {
7662 if (++order >= MAX_ORDER) {
7663 outer_start = start;
7666 outer_start &= ~0UL << order;
7669 if (outer_start != start) {
7670 order = page_order(pfn_to_page(outer_start));
7673 * outer_start page could be small order buddy page and
7674 * it doesn't include start page. Adjust outer_start
7675 * in this case to report failed page properly
7676 * on tracepoint in test_pages_isolated()
7678 if (outer_start + (1UL << order) <= start)
7679 outer_start = start;
7682 /* Make sure the range is really isolated. */
7683 if (test_pages_isolated(outer_start, end, false)) {
7684 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7685 __func__, outer_start, end);
7690 /* Grab isolated pages from freelists. */
7691 outer_end = isolate_freepages_range(&cc, outer_start, end);
7697 /* Free head and tail (if any) */
7698 if (start != outer_start)
7699 free_contig_range(outer_start, start - outer_start);
7700 if (end != outer_end)
7701 free_contig_range(end, outer_end - end);
7704 undo_isolate_page_range(pfn_max_align_down(start),
7705 pfn_max_align_up(end), migratetype);
7709 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7711 unsigned int count = 0;
7713 for (; nr_pages--; pfn++) {
7714 struct page *page = pfn_to_page(pfn);
7716 count += page_count(page) != 1;
7719 WARN(count != 0, "%d pages are still in use!\n", count);
7723 #ifdef CONFIG_MEMORY_HOTPLUG
7725 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7726 * page high values need to be recalulated.
7728 void __meminit zone_pcp_update(struct zone *zone)
7731 mutex_lock(&pcp_batch_high_lock);
7732 for_each_possible_cpu(cpu)
7733 pageset_set_high_and_batch(zone,
7734 per_cpu_ptr(zone->pageset, cpu));
7735 mutex_unlock(&pcp_batch_high_lock);
7739 void zone_pcp_reset(struct zone *zone)
7741 unsigned long flags;
7743 struct per_cpu_pageset *pset;
7745 /* avoid races with drain_pages() */
7746 local_irq_save(flags);
7747 if (zone->pageset != &boot_pageset) {
7748 for_each_online_cpu(cpu) {
7749 pset = per_cpu_ptr(zone->pageset, cpu);
7750 drain_zonestat(zone, pset);
7752 free_percpu(zone->pageset);
7753 zone->pageset = &boot_pageset;
7755 local_irq_restore(flags);
7758 #ifdef CONFIG_MEMORY_HOTREMOVE
7760 * All pages in the range must be in a single zone and isolated
7761 * before calling this.
7764 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7768 unsigned int order, i;
7770 unsigned long flags;
7771 /* find the first valid pfn */
7772 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7777 offline_mem_sections(pfn, end_pfn);
7778 zone = page_zone(pfn_to_page(pfn));
7779 spin_lock_irqsave(&zone->lock, flags);
7781 while (pfn < end_pfn) {
7782 if (!pfn_valid(pfn)) {
7786 page = pfn_to_page(pfn);
7788 * The HWPoisoned page may be not in buddy system, and
7789 * page_count() is not 0.
7791 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7793 SetPageReserved(page);
7797 BUG_ON(page_count(page));
7798 BUG_ON(!PageBuddy(page));
7799 order = page_order(page);
7800 #ifdef CONFIG_DEBUG_VM
7801 pr_info("remove from free list %lx %d %lx\n",
7802 pfn, 1 << order, end_pfn);
7804 list_del(&page->lru);
7805 rmv_page_order(page);
7806 zone->free_area[order].nr_free--;
7807 for (i = 0; i < (1 << order); i++)
7808 SetPageReserved((page+i));
7809 pfn += (1 << order);
7811 spin_unlock_irqrestore(&zone->lock, flags);
7815 bool is_free_buddy_page(struct page *page)
7817 struct zone *zone = page_zone(page);
7818 unsigned long pfn = page_to_pfn(page);
7819 unsigned long flags;
7822 spin_lock_irqsave(&zone->lock, flags);
7823 for (order = 0; order < MAX_ORDER; order++) {
7824 struct page *page_head = page - (pfn & ((1 << order) - 1));
7826 if (PageBuddy(page_head) && page_order(page_head) >= order)
7829 spin_unlock_irqrestore(&zone->lock, flags);
7831 return order < MAX_ORDER;