2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/memremap.h>
47 #include <linux/stop_machine.h>
48 #include <linux/sort.h>
49 #include <linux/pfn.h>
50 #include <linux/backing-dev.h>
51 #include <linux/fault-inject.h>
52 #include <linux/page-isolation.h>
53 #include <linux/page_ext.h>
54 #include <linux/debugobjects.h>
55 #include <linux/kmemleak.h>
56 #include <linux/compaction.h>
57 #include <trace/events/kmem.h>
58 #include <trace/events/oom.h>
59 #include <linux/prefetch.h>
60 #include <linux/mm_inline.h>
61 #include <linux/migrate.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
64 #include <linux/sched/mm.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
68 #include <linux/ftrace.h>
70 #include <asm/sections.h>
71 #include <asm/tlbflush.h>
72 #include <asm/div64.h>
75 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
76 static DEFINE_MUTEX(pcp_batch_high_lock);
77 #define MIN_PERCPU_PAGELIST_FRACTION (8)
79 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
80 DEFINE_PER_CPU(int, numa_node);
81 EXPORT_PER_CPU_SYMBOL(numa_node);
84 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
86 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
87 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
88 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
89 * defined in <linux/topology.h>.
91 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
92 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
93 int _node_numa_mem_[MAX_NUMNODES];
96 /* work_structs for global per-cpu drains */
97 DEFINE_MUTEX(pcpu_drain_mutex);
98 DEFINE_PER_CPU(struct work_struct, pcpu_drain);
100 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
101 volatile unsigned long latent_entropy __latent_entropy;
102 EXPORT_SYMBOL(latent_entropy);
106 * Array of node states.
108 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
109 [N_POSSIBLE] = NODE_MASK_ALL,
110 [N_ONLINE] = { { [0] = 1UL } },
112 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
113 #ifdef CONFIG_HIGHMEM
114 [N_HIGH_MEMORY] = { { [0] = 1UL } },
116 [N_MEMORY] = { { [0] = 1UL } },
117 [N_CPU] = { { [0] = 1UL } },
120 EXPORT_SYMBOL(node_states);
122 /* Protect totalram_pages and zone->managed_pages */
123 static DEFINE_SPINLOCK(managed_page_count_lock);
125 unsigned long totalram_pages __read_mostly;
126 unsigned long totalreserve_pages __read_mostly;
127 unsigned long totalcma_pages __read_mostly;
129 int percpu_pagelist_fraction;
130 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
133 * A cached value of the page's pageblock's migratetype, used when the page is
134 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
135 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
136 * Also the migratetype set in the page does not necessarily match the pcplist
137 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
138 * other index - this ensures that it will be put on the correct CMA freelist.
140 static inline int get_pcppage_migratetype(struct page *page)
145 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
147 page->index = migratetype;
150 #ifdef CONFIG_PM_SLEEP
152 * The following functions are used by the suspend/hibernate code to temporarily
153 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
154 * while devices are suspended. To avoid races with the suspend/hibernate code,
155 * they should always be called with pm_mutex held (gfp_allowed_mask also should
156 * only be modified with pm_mutex held, unless the suspend/hibernate code is
157 * guaranteed not to run in parallel with that modification).
160 static gfp_t saved_gfp_mask;
162 void pm_restore_gfp_mask(void)
164 WARN_ON(!mutex_is_locked(&pm_mutex));
165 if (saved_gfp_mask) {
166 gfp_allowed_mask = saved_gfp_mask;
171 void pm_restrict_gfp_mask(void)
173 WARN_ON(!mutex_is_locked(&pm_mutex));
174 WARN_ON(saved_gfp_mask);
175 saved_gfp_mask = gfp_allowed_mask;
176 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
179 bool pm_suspended_storage(void)
181 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
185 #endif /* CONFIG_PM_SLEEP */
187 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
188 unsigned int pageblock_order __read_mostly;
191 static void __free_pages_ok(struct page *page, unsigned int order);
194 * results with 256, 32 in the lowmem_reserve sysctl:
195 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
196 * 1G machine -> (16M dma, 784M normal, 224M high)
197 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
198 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
199 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
201 * TBD: should special case ZONE_DMA32 machines here - in those we normally
202 * don't need any ZONE_NORMAL reservation
204 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
205 #ifdef CONFIG_ZONE_DMA
208 #ifdef CONFIG_ZONE_DMA32
211 #ifdef CONFIG_HIGHMEM
217 EXPORT_SYMBOL(totalram_pages);
219 static char * const zone_names[MAX_NR_ZONES] = {
220 #ifdef CONFIG_ZONE_DMA
223 #ifdef CONFIG_ZONE_DMA32
227 #ifdef CONFIG_HIGHMEM
231 #ifdef CONFIG_ZONE_DEVICE
236 char * const migratetype_names[MIGRATE_TYPES] = {
244 #ifdef CONFIG_MEMORY_ISOLATION
249 compound_page_dtor * const compound_page_dtors[] = {
252 #ifdef CONFIG_HUGETLB_PAGE
255 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
260 int min_free_kbytes = 1024;
261 int user_min_free_kbytes = -1;
262 int watermark_scale_factor = 10;
264 static unsigned long __meminitdata nr_kernel_pages;
265 static unsigned long __meminitdata nr_all_pages;
266 static unsigned long __meminitdata dma_reserve;
268 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
269 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
270 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
271 static unsigned long __initdata required_kernelcore;
272 static unsigned long __initdata required_movablecore;
273 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
274 static bool mirrored_kernelcore;
276 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
278 EXPORT_SYMBOL(movable_zone);
279 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
282 int nr_node_ids __read_mostly = MAX_NUMNODES;
283 int nr_online_nodes __read_mostly = 1;
284 EXPORT_SYMBOL(nr_node_ids);
285 EXPORT_SYMBOL(nr_online_nodes);
288 int page_group_by_mobility_disabled __read_mostly;
290 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
291 static inline void reset_deferred_meminit(pg_data_t *pgdat)
293 unsigned long max_initialise;
294 unsigned long reserved_lowmem;
297 * Initialise at least 2G of a node but also take into account that
298 * two large system hashes that can take up 1GB for 0.25TB/node.
300 max_initialise = max(2UL << (30 - PAGE_SHIFT),
301 (pgdat->node_spanned_pages >> 8));
304 * Compensate the all the memblock reservations (e.g. crash kernel)
305 * from the initial estimation to make sure we will initialize enough
308 reserved_lowmem = memblock_reserved_memory_within(pgdat->node_start_pfn,
309 pgdat->node_start_pfn + max_initialise);
310 max_initialise += reserved_lowmem;
312 pgdat->static_init_size = min(max_initialise, pgdat->node_spanned_pages);
313 pgdat->first_deferred_pfn = ULONG_MAX;
316 /* Returns true if the struct page for the pfn is uninitialised */
317 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
319 int nid = early_pfn_to_nid(pfn);
321 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
328 * Returns false when the remaining initialisation should be deferred until
329 * later in the boot cycle when it can be parallelised.
331 static inline bool update_defer_init(pg_data_t *pgdat,
332 unsigned long pfn, unsigned long zone_end,
333 unsigned long *nr_initialised)
335 /* Always populate low zones for address-contrained allocations */
336 if (zone_end < pgdat_end_pfn(pgdat))
339 if ((*nr_initialised > pgdat->static_init_size) &&
340 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
341 pgdat->first_deferred_pfn = pfn;
348 static inline void reset_deferred_meminit(pg_data_t *pgdat)
352 static inline bool early_page_uninitialised(unsigned long pfn)
357 static inline bool update_defer_init(pg_data_t *pgdat,
358 unsigned long pfn, unsigned long zone_end,
359 unsigned long *nr_initialised)
365 /* Return a pointer to the bitmap storing bits affecting a block of pages */
366 static inline unsigned long *get_pageblock_bitmap(struct page *page,
369 #ifdef CONFIG_SPARSEMEM
370 return __pfn_to_section(pfn)->pageblock_flags;
372 return page_zone(page)->pageblock_flags;
373 #endif /* CONFIG_SPARSEMEM */
376 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
378 #ifdef CONFIG_SPARSEMEM
379 pfn &= (PAGES_PER_SECTION-1);
380 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
382 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
383 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
384 #endif /* CONFIG_SPARSEMEM */
388 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
389 * @page: The page within the block of interest
390 * @pfn: The target page frame number
391 * @end_bitidx: The last bit of interest to retrieve
392 * @mask: mask of bits that the caller is interested in
394 * Return: pageblock_bits flags
396 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
398 unsigned long end_bitidx,
401 unsigned long *bitmap;
402 unsigned long bitidx, word_bitidx;
405 bitmap = get_pageblock_bitmap(page, pfn);
406 bitidx = pfn_to_bitidx(page, pfn);
407 word_bitidx = bitidx / BITS_PER_LONG;
408 bitidx &= (BITS_PER_LONG-1);
410 word = bitmap[word_bitidx];
411 bitidx += end_bitidx;
412 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
415 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
416 unsigned long end_bitidx,
419 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
422 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
424 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
428 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
429 * @page: The page within the block of interest
430 * @flags: The flags to set
431 * @pfn: The target page frame number
432 * @end_bitidx: The last bit of interest
433 * @mask: mask of bits that the caller is interested in
435 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
437 unsigned long end_bitidx,
440 unsigned long *bitmap;
441 unsigned long bitidx, word_bitidx;
442 unsigned long old_word, word;
444 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
446 bitmap = get_pageblock_bitmap(page, pfn);
447 bitidx = pfn_to_bitidx(page, pfn);
448 word_bitidx = bitidx / BITS_PER_LONG;
449 bitidx &= (BITS_PER_LONG-1);
451 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
453 bitidx += end_bitidx;
454 mask <<= (BITS_PER_LONG - bitidx - 1);
455 flags <<= (BITS_PER_LONG - bitidx - 1);
457 word = READ_ONCE(bitmap[word_bitidx]);
459 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
460 if (word == old_word)
466 void set_pageblock_migratetype(struct page *page, int migratetype)
468 if (unlikely(page_group_by_mobility_disabled &&
469 migratetype < MIGRATE_PCPTYPES))
470 migratetype = MIGRATE_UNMOVABLE;
472 set_pageblock_flags_group(page, (unsigned long)migratetype,
473 PB_migrate, PB_migrate_end);
476 #ifdef CONFIG_DEBUG_VM
477 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
481 unsigned long pfn = page_to_pfn(page);
482 unsigned long sp, start_pfn;
485 seq = zone_span_seqbegin(zone);
486 start_pfn = zone->zone_start_pfn;
487 sp = zone->spanned_pages;
488 if (!zone_spans_pfn(zone, pfn))
490 } while (zone_span_seqretry(zone, seq));
493 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
494 pfn, zone_to_nid(zone), zone->name,
495 start_pfn, start_pfn + sp);
500 static int page_is_consistent(struct zone *zone, struct page *page)
502 if (!pfn_valid_within(page_to_pfn(page)))
504 if (zone != page_zone(page))
510 * Temporary debugging check for pages not lying within a given zone.
512 static int __maybe_unused bad_range(struct zone *zone, struct page *page)
514 if (page_outside_zone_boundaries(zone, page))
516 if (!page_is_consistent(zone, page))
522 static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
528 static void bad_page(struct page *page, const char *reason,
529 unsigned long bad_flags)
531 static unsigned long resume;
532 static unsigned long nr_shown;
533 static unsigned long nr_unshown;
536 * Allow a burst of 60 reports, then keep quiet for that minute;
537 * or allow a steady drip of one report per second.
539 if (nr_shown == 60) {
540 if (time_before(jiffies, resume)) {
546 "BUG: Bad page state: %lu messages suppressed\n",
553 resume = jiffies + 60 * HZ;
555 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
556 current->comm, page_to_pfn(page));
557 __dump_page(page, reason);
558 bad_flags &= page->flags;
560 pr_alert("bad because of flags: %#lx(%pGp)\n",
561 bad_flags, &bad_flags);
562 dump_page_owner(page);
567 /* Leave bad fields for debug, except PageBuddy could make trouble */
568 page_mapcount_reset(page); /* remove PageBuddy */
569 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
573 * Higher-order pages are called "compound pages". They are structured thusly:
575 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
577 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
578 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
580 * The first tail page's ->compound_dtor holds the offset in array of compound
581 * page destructors. See compound_page_dtors.
583 * The first tail page's ->compound_order holds the order of allocation.
584 * This usage means that zero-order pages may not be compound.
587 void free_compound_page(struct page *page)
589 __free_pages_ok(page, compound_order(page));
592 void prep_compound_page(struct page *page, unsigned int order)
595 int nr_pages = 1 << order;
597 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
598 set_compound_order(page, order);
600 for (i = 1; i < nr_pages; i++) {
601 struct page *p = page + i;
602 set_page_count(p, 0);
603 p->mapping = TAIL_MAPPING;
604 set_compound_head(p, page);
606 atomic_set(compound_mapcount_ptr(page), -1);
609 #ifdef CONFIG_DEBUG_PAGEALLOC
610 unsigned int _debug_guardpage_minorder;
611 bool _debug_pagealloc_enabled __read_mostly
612 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
613 EXPORT_SYMBOL(_debug_pagealloc_enabled);
614 bool _debug_guardpage_enabled __read_mostly;
616 static int __init early_debug_pagealloc(char *buf)
620 return kstrtobool(buf, &_debug_pagealloc_enabled);
622 early_param("debug_pagealloc", early_debug_pagealloc);
624 static bool need_debug_guardpage(void)
626 /* If we don't use debug_pagealloc, we don't need guard page */
627 if (!debug_pagealloc_enabled())
630 if (!debug_guardpage_minorder())
636 static void init_debug_guardpage(void)
638 if (!debug_pagealloc_enabled())
641 if (!debug_guardpage_minorder())
644 _debug_guardpage_enabled = true;
647 struct page_ext_operations debug_guardpage_ops = {
648 .need = need_debug_guardpage,
649 .init = init_debug_guardpage,
652 static int __init debug_guardpage_minorder_setup(char *buf)
656 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
657 pr_err("Bad debug_guardpage_minorder value\n");
660 _debug_guardpage_minorder = res;
661 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
664 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
666 static inline bool set_page_guard(struct zone *zone, struct page *page,
667 unsigned int order, int migratetype)
669 struct page_ext *page_ext;
671 if (!debug_guardpage_enabled())
674 if (order >= debug_guardpage_minorder())
677 page_ext = lookup_page_ext(page);
678 if (unlikely(!page_ext))
681 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
683 INIT_LIST_HEAD(&page->lru);
684 set_page_private(page, order);
685 /* Guard pages are not available for any usage */
686 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
691 static inline void clear_page_guard(struct zone *zone, struct page *page,
692 unsigned int order, int migratetype)
694 struct page_ext *page_ext;
696 if (!debug_guardpage_enabled())
699 page_ext = lookup_page_ext(page);
700 if (unlikely(!page_ext))
703 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
705 set_page_private(page, 0);
706 if (!is_migrate_isolate(migratetype))
707 __mod_zone_freepage_state(zone, (1 << order), migratetype);
710 struct page_ext_operations debug_guardpage_ops;
711 static inline bool set_page_guard(struct zone *zone, struct page *page,
712 unsigned int order, int migratetype) { return false; }
713 static inline void clear_page_guard(struct zone *zone, struct page *page,
714 unsigned int order, int migratetype) {}
717 static inline void set_page_order(struct page *page, unsigned int order)
719 set_page_private(page, order);
720 __SetPageBuddy(page);
723 static inline void rmv_page_order(struct page *page)
725 __ClearPageBuddy(page);
726 set_page_private(page, 0);
730 * This function checks whether a page is free && is the buddy
731 * we can do coalesce a page and its buddy if
732 * (a) the buddy is not in a hole (check before calling!) &&
733 * (b) the buddy is in the buddy system &&
734 * (c) a page and its buddy have the same order &&
735 * (d) a page and its buddy are in the same zone.
737 * For recording whether a page is in the buddy system, we set ->_mapcount
738 * PAGE_BUDDY_MAPCOUNT_VALUE.
739 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
740 * serialized by zone->lock.
742 * For recording page's order, we use page_private(page).
744 static inline int page_is_buddy(struct page *page, struct page *buddy,
747 if (page_is_guard(buddy) && page_order(buddy) == order) {
748 if (page_zone_id(page) != page_zone_id(buddy))
751 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
756 if (PageBuddy(buddy) && page_order(buddy) == order) {
758 * zone check is done late to avoid uselessly
759 * calculating zone/node ids for pages that could
762 if (page_zone_id(page) != page_zone_id(buddy))
765 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
773 * Freeing function for a buddy system allocator.
775 * The concept of a buddy system is to maintain direct-mapped table
776 * (containing bit values) for memory blocks of various "orders".
777 * The bottom level table contains the map for the smallest allocatable
778 * units of memory (here, pages), and each level above it describes
779 * pairs of units from the levels below, hence, "buddies".
780 * At a high level, all that happens here is marking the table entry
781 * at the bottom level available, and propagating the changes upward
782 * as necessary, plus some accounting needed to play nicely with other
783 * parts of the VM system.
784 * At each level, we keep a list of pages, which are heads of continuous
785 * free pages of length of (1 << order) and marked with _mapcount
786 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
788 * So when we are allocating or freeing one, we can derive the state of the
789 * other. That is, if we allocate a small block, and both were
790 * free, the remainder of the region must be split into blocks.
791 * If a block is freed, and its buddy is also free, then this
792 * triggers coalescing into a block of larger size.
797 static inline void __free_one_page(struct page *page,
799 struct zone *zone, unsigned int order,
802 unsigned long combined_pfn;
803 unsigned long uninitialized_var(buddy_pfn);
805 unsigned int max_order;
807 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
809 VM_BUG_ON(!zone_is_initialized(zone));
810 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
812 VM_BUG_ON(migratetype == -1);
813 if (likely(!is_migrate_isolate(migratetype)))
814 __mod_zone_freepage_state(zone, 1 << order, migratetype);
816 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
817 VM_BUG_ON_PAGE(bad_range(zone, page), page);
820 while (order < max_order - 1) {
821 buddy_pfn = __find_buddy_pfn(pfn, order);
822 buddy = page + (buddy_pfn - pfn);
824 if (!pfn_valid_within(buddy_pfn))
826 if (!page_is_buddy(page, buddy, order))
829 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
830 * merge with it and move up one order.
832 if (page_is_guard(buddy)) {
833 clear_page_guard(zone, buddy, order, migratetype);
835 list_del(&buddy->lru);
836 zone->free_area[order].nr_free--;
837 rmv_page_order(buddy);
839 combined_pfn = buddy_pfn & pfn;
840 page = page + (combined_pfn - pfn);
844 if (max_order < MAX_ORDER) {
845 /* If we are here, it means order is >= pageblock_order.
846 * We want to prevent merge between freepages on isolate
847 * pageblock and normal pageblock. Without this, pageblock
848 * isolation could cause incorrect freepage or CMA accounting.
850 * We don't want to hit this code for the more frequent
853 if (unlikely(has_isolate_pageblock(zone))) {
856 buddy_pfn = __find_buddy_pfn(pfn, order);
857 buddy = page + (buddy_pfn - pfn);
858 buddy_mt = get_pageblock_migratetype(buddy);
860 if (migratetype != buddy_mt
861 && (is_migrate_isolate(migratetype) ||
862 is_migrate_isolate(buddy_mt)))
866 goto continue_merging;
870 set_page_order(page, order);
873 * If this is not the largest possible page, check if the buddy
874 * of the next-highest order is free. If it is, it's possible
875 * that pages are being freed that will coalesce soon. In case,
876 * that is happening, add the free page to the tail of the list
877 * so it's less likely to be used soon and more likely to be merged
878 * as a higher order page
880 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
881 struct page *higher_page, *higher_buddy;
882 combined_pfn = buddy_pfn & pfn;
883 higher_page = page + (combined_pfn - pfn);
884 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
885 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
886 if (pfn_valid_within(buddy_pfn) &&
887 page_is_buddy(higher_page, higher_buddy, order + 1)) {
888 list_add_tail(&page->lru,
889 &zone->free_area[order].free_list[migratetype]);
894 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
896 zone->free_area[order].nr_free++;
900 * A bad page could be due to a number of fields. Instead of multiple branches,
901 * try and check multiple fields with one check. The caller must do a detailed
902 * check if necessary.
904 static inline bool page_expected_state(struct page *page,
905 unsigned long check_flags)
907 if (unlikely(atomic_read(&page->_mapcount) != -1))
910 if (unlikely((unsigned long)page->mapping |
911 page_ref_count(page) |
913 (unsigned long)page->mem_cgroup |
915 (page->flags & check_flags)))
921 static void free_pages_check_bad(struct page *page)
923 const char *bad_reason;
924 unsigned long bad_flags;
929 if (unlikely(atomic_read(&page->_mapcount) != -1))
930 bad_reason = "nonzero mapcount";
931 if (unlikely(page->mapping != NULL))
932 bad_reason = "non-NULL mapping";
933 if (unlikely(page_ref_count(page) != 0))
934 bad_reason = "nonzero _refcount";
935 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
936 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
937 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
940 if (unlikely(page->mem_cgroup))
941 bad_reason = "page still charged to cgroup";
943 bad_page(page, bad_reason, bad_flags);
946 static inline int free_pages_check(struct page *page)
948 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
951 /* Something has gone sideways, find it */
952 free_pages_check_bad(page);
956 static int free_tail_pages_check(struct page *head_page, struct page *page)
961 * We rely page->lru.next never has bit 0 set, unless the page
962 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
964 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
966 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
970 switch (page - head_page) {
972 /* the first tail page: ->mapping is compound_mapcount() */
973 if (unlikely(compound_mapcount(page))) {
974 bad_page(page, "nonzero compound_mapcount", 0);
980 * the second tail page: ->mapping is
981 * page_deferred_list().next -- ignore value.
985 if (page->mapping != TAIL_MAPPING) {
986 bad_page(page, "corrupted mapping in tail page", 0);
991 if (unlikely(!PageTail(page))) {
992 bad_page(page, "PageTail not set", 0);
995 if (unlikely(compound_head(page) != head_page)) {
996 bad_page(page, "compound_head not consistent", 0);
1001 page->mapping = NULL;
1002 clear_compound_head(page);
1006 static __always_inline bool free_pages_prepare(struct page *page,
1007 unsigned int order, bool check_free)
1011 VM_BUG_ON_PAGE(PageTail(page), page);
1013 trace_mm_page_free(page, order);
1014 kmemcheck_free_shadow(page, order);
1017 * Check tail pages before head page information is cleared to
1018 * avoid checking PageCompound for order-0 pages.
1020 if (unlikely(order)) {
1021 bool compound = PageCompound(page);
1024 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1027 ClearPageDoubleMap(page);
1028 for (i = 1; i < (1 << order); i++) {
1030 bad += free_tail_pages_check(page, page + i);
1031 if (unlikely(free_pages_check(page + i))) {
1035 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1038 if (PageMappingFlags(page))
1039 page->mapping = NULL;
1040 if (memcg_kmem_enabled() && PageKmemcg(page))
1041 memcg_kmem_uncharge(page, order);
1043 bad += free_pages_check(page);
1047 page_cpupid_reset_last(page);
1048 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1049 reset_page_owner(page, order);
1051 if (!PageHighMem(page)) {
1052 debug_check_no_locks_freed(page_address(page),
1053 PAGE_SIZE << order);
1054 debug_check_no_obj_freed(page_address(page),
1055 PAGE_SIZE << order);
1057 arch_free_page(page, order);
1058 kernel_poison_pages(page, 1 << order, 0);
1059 kernel_map_pages(page, 1 << order, 0);
1060 kasan_free_pages(page, order);
1065 #ifdef CONFIG_DEBUG_VM
1066 static inline bool free_pcp_prepare(struct page *page)
1068 return free_pages_prepare(page, 0, true);
1071 static inline bool bulkfree_pcp_prepare(struct page *page)
1076 static bool free_pcp_prepare(struct page *page)
1078 return free_pages_prepare(page, 0, false);
1081 static bool bulkfree_pcp_prepare(struct page *page)
1083 return free_pages_check(page);
1085 #endif /* CONFIG_DEBUG_VM */
1088 * Frees a number of pages from the PCP lists
1089 * Assumes all pages on list are in same zone, and of same order.
1090 * count is the number of pages to free.
1092 * If the zone was previously in an "all pages pinned" state then look to
1093 * see if this freeing clears that state.
1095 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1096 * pinned" detection logic.
1098 static void free_pcppages_bulk(struct zone *zone, int count,
1099 struct per_cpu_pages *pcp)
1101 int migratetype = 0;
1103 bool isolated_pageblocks;
1105 spin_lock(&zone->lock);
1106 isolated_pageblocks = has_isolate_pageblock(zone);
1110 struct list_head *list;
1113 * Remove pages from lists in a round-robin fashion. A
1114 * batch_free count is maintained that is incremented when an
1115 * empty list is encountered. This is so more pages are freed
1116 * off fuller lists instead of spinning excessively around empty
1121 if (++migratetype == MIGRATE_PCPTYPES)
1123 list = &pcp->lists[migratetype];
1124 } while (list_empty(list));
1126 /* This is the only non-empty list. Free them all. */
1127 if (batch_free == MIGRATE_PCPTYPES)
1131 int mt; /* migratetype of the to-be-freed page */
1133 page = list_last_entry(list, struct page, lru);
1134 /* must delete as __free_one_page list manipulates */
1135 list_del(&page->lru);
1137 mt = get_pcppage_migratetype(page);
1138 /* MIGRATE_ISOLATE page should not go to pcplists */
1139 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1140 /* Pageblock could have been isolated meanwhile */
1141 if (unlikely(isolated_pageblocks))
1142 mt = get_pageblock_migratetype(page);
1144 if (bulkfree_pcp_prepare(page))
1147 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1148 trace_mm_page_pcpu_drain(page, 0, mt);
1149 } while (--count && --batch_free && !list_empty(list));
1151 spin_unlock(&zone->lock);
1154 static void free_one_page(struct zone *zone,
1155 struct page *page, unsigned long pfn,
1159 spin_lock(&zone->lock);
1160 if (unlikely(has_isolate_pageblock(zone) ||
1161 is_migrate_isolate(migratetype))) {
1162 migratetype = get_pfnblock_migratetype(page, pfn);
1164 __free_one_page(page, pfn, zone, order, migratetype);
1165 spin_unlock(&zone->lock);
1168 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1169 unsigned long zone, int nid)
1171 set_page_links(page, zone, nid, pfn);
1172 init_page_count(page);
1173 page_mapcount_reset(page);
1174 page_cpupid_reset_last(page);
1176 INIT_LIST_HEAD(&page->lru);
1177 #ifdef WANT_PAGE_VIRTUAL
1178 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1179 if (!is_highmem_idx(zone))
1180 set_page_address(page, __va(pfn << PAGE_SHIFT));
1184 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1187 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1190 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1191 static void init_reserved_page(unsigned long pfn)
1196 if (!early_page_uninitialised(pfn))
1199 nid = early_pfn_to_nid(pfn);
1200 pgdat = NODE_DATA(nid);
1202 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1203 struct zone *zone = &pgdat->node_zones[zid];
1205 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1208 __init_single_pfn(pfn, zid, nid);
1211 static inline void init_reserved_page(unsigned long pfn)
1214 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1217 * Initialised pages do not have PageReserved set. This function is
1218 * called for each range allocated by the bootmem allocator and
1219 * marks the pages PageReserved. The remaining valid pages are later
1220 * sent to the buddy page allocator.
1222 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1224 unsigned long start_pfn = PFN_DOWN(start);
1225 unsigned long end_pfn = PFN_UP(end);
1227 for (; start_pfn < end_pfn; start_pfn++) {
1228 if (pfn_valid(start_pfn)) {
1229 struct page *page = pfn_to_page(start_pfn);
1231 init_reserved_page(start_pfn);
1233 /* Avoid false-positive PageTail() */
1234 INIT_LIST_HEAD(&page->lru);
1236 SetPageReserved(page);
1241 static void __free_pages_ok(struct page *page, unsigned int order)
1243 unsigned long flags;
1245 unsigned long pfn = page_to_pfn(page);
1247 if (!free_pages_prepare(page, order, true))
1250 migratetype = get_pfnblock_migratetype(page, pfn);
1251 local_irq_save(flags);
1252 __count_vm_events(PGFREE, 1 << order);
1253 free_one_page(page_zone(page), page, pfn, order, migratetype);
1254 local_irq_restore(flags);
1257 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1259 unsigned int nr_pages = 1 << order;
1260 struct page *p = page;
1264 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1266 __ClearPageReserved(p);
1267 set_page_count(p, 0);
1269 __ClearPageReserved(p);
1270 set_page_count(p, 0);
1272 page_zone(page)->managed_pages += nr_pages;
1273 set_page_refcounted(page);
1274 __free_pages(page, order);
1277 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1278 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1280 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1282 int __meminit early_pfn_to_nid(unsigned long pfn)
1284 static DEFINE_SPINLOCK(early_pfn_lock);
1287 spin_lock(&early_pfn_lock);
1288 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1290 nid = first_online_node;
1291 spin_unlock(&early_pfn_lock);
1297 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1298 static inline bool __meminit __maybe_unused
1299 meminit_pfn_in_nid(unsigned long pfn, int node,
1300 struct mminit_pfnnid_cache *state)
1304 nid = __early_pfn_to_nid(pfn, state);
1305 if (nid >= 0 && nid != node)
1310 /* Only safe to use early in boot when initialisation is single-threaded */
1311 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1313 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1318 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1322 static inline bool __meminit __maybe_unused
1323 meminit_pfn_in_nid(unsigned long pfn, int node,
1324 struct mminit_pfnnid_cache *state)
1331 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1334 if (early_page_uninitialised(pfn))
1336 return __free_pages_boot_core(page, order);
1340 * Check that the whole (or subset of) a pageblock given by the interval of
1341 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1342 * with the migration of free compaction scanner. The scanners then need to
1343 * use only pfn_valid_within() check for arches that allow holes within
1346 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1348 * It's possible on some configurations to have a setup like node0 node1 node0
1349 * i.e. it's possible that all pages within a zones range of pages do not
1350 * belong to a single zone. We assume that a border between node0 and node1
1351 * can occur within a single pageblock, but not a node0 node1 node0
1352 * interleaving within a single pageblock. It is therefore sufficient to check
1353 * the first and last page of a pageblock and avoid checking each individual
1354 * page in a pageblock.
1356 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1357 unsigned long end_pfn, struct zone *zone)
1359 struct page *start_page;
1360 struct page *end_page;
1362 /* end_pfn is one past the range we are checking */
1365 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1368 start_page = pfn_to_online_page(start_pfn);
1372 if (page_zone(start_page) != zone)
1375 end_page = pfn_to_page(end_pfn);
1377 /* This gives a shorter code than deriving page_zone(end_page) */
1378 if (page_zone_id(start_page) != page_zone_id(end_page))
1384 void set_zone_contiguous(struct zone *zone)
1386 unsigned long block_start_pfn = zone->zone_start_pfn;
1387 unsigned long block_end_pfn;
1389 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1390 for (; block_start_pfn < zone_end_pfn(zone);
1391 block_start_pfn = block_end_pfn,
1392 block_end_pfn += pageblock_nr_pages) {
1394 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1396 if (!__pageblock_pfn_to_page(block_start_pfn,
1397 block_end_pfn, zone))
1401 /* We confirm that there is no hole */
1402 zone->contiguous = true;
1405 void clear_zone_contiguous(struct zone *zone)
1407 zone->contiguous = false;
1410 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1411 static void __init deferred_free_range(struct page *page,
1412 unsigned long pfn, int nr_pages)
1419 /* Free a large naturally-aligned chunk if possible */
1420 if (nr_pages == pageblock_nr_pages &&
1421 (pfn & (pageblock_nr_pages - 1)) == 0) {
1422 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1423 __free_pages_boot_core(page, pageblock_order);
1427 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1428 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1429 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1430 __free_pages_boot_core(page, 0);
1434 /* Completion tracking for deferred_init_memmap() threads */
1435 static atomic_t pgdat_init_n_undone __initdata;
1436 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1438 static inline void __init pgdat_init_report_one_done(void)
1440 if (atomic_dec_and_test(&pgdat_init_n_undone))
1441 complete(&pgdat_init_all_done_comp);
1444 /* Initialise remaining memory on a node */
1445 static int __init deferred_init_memmap(void *data)
1447 pg_data_t *pgdat = data;
1448 int nid = pgdat->node_id;
1449 struct mminit_pfnnid_cache nid_init_state = { };
1450 unsigned long start = jiffies;
1451 unsigned long nr_pages = 0;
1452 unsigned long walk_start, walk_end;
1455 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1456 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1458 if (first_init_pfn == ULONG_MAX) {
1459 pgdat_init_report_one_done();
1463 /* Bind memory initialisation thread to a local node if possible */
1464 if (!cpumask_empty(cpumask))
1465 set_cpus_allowed_ptr(current, cpumask);
1467 /* Sanity check boundaries */
1468 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1469 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1470 pgdat->first_deferred_pfn = ULONG_MAX;
1472 /* Only the highest zone is deferred so find it */
1473 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1474 zone = pgdat->node_zones + zid;
1475 if (first_init_pfn < zone_end_pfn(zone))
1479 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1480 unsigned long pfn, end_pfn;
1481 struct page *page = NULL;
1482 struct page *free_base_page = NULL;
1483 unsigned long free_base_pfn = 0;
1486 end_pfn = min(walk_end, zone_end_pfn(zone));
1487 pfn = first_init_pfn;
1488 if (pfn < walk_start)
1490 if (pfn < zone->zone_start_pfn)
1491 pfn = zone->zone_start_pfn;
1493 for (; pfn < end_pfn; pfn++) {
1494 if (!pfn_valid_within(pfn))
1498 * Ensure pfn_valid is checked every
1499 * pageblock_nr_pages for memory holes
1501 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1502 if (!pfn_valid(pfn)) {
1508 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1513 /* Minimise pfn page lookups and scheduler checks */
1514 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1517 nr_pages += nr_to_free;
1518 deferred_free_range(free_base_page,
1519 free_base_pfn, nr_to_free);
1520 free_base_page = NULL;
1521 free_base_pfn = nr_to_free = 0;
1523 page = pfn_to_page(pfn);
1528 VM_BUG_ON(page_zone(page) != zone);
1532 __init_single_page(page, pfn, zid, nid);
1533 if (!free_base_page) {
1534 free_base_page = page;
1535 free_base_pfn = pfn;
1540 /* Where possible, batch up pages for a single free */
1543 /* Free the current block of pages to allocator */
1544 nr_pages += nr_to_free;
1545 deferred_free_range(free_base_page, free_base_pfn,
1547 free_base_page = NULL;
1548 free_base_pfn = nr_to_free = 0;
1550 /* Free the last block of pages to allocator */
1551 nr_pages += nr_to_free;
1552 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1554 first_init_pfn = max(end_pfn, first_init_pfn);
1557 /* Sanity check that the next zone really is unpopulated */
1558 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1560 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1561 jiffies_to_msecs(jiffies - start));
1563 pgdat_init_report_one_done();
1566 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1568 void __init page_alloc_init_late(void)
1572 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1575 /* There will be num_node_state(N_MEMORY) threads */
1576 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1577 for_each_node_state(nid, N_MEMORY) {
1578 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1581 /* Block until all are initialised */
1582 wait_for_completion(&pgdat_init_all_done_comp);
1584 /* Reinit limits that are based on free pages after the kernel is up */
1585 files_maxfiles_init();
1588 for_each_populated_zone(zone)
1589 set_zone_contiguous(zone);
1593 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1594 void __init init_cma_reserved_pageblock(struct page *page)
1596 unsigned i = pageblock_nr_pages;
1597 struct page *p = page;
1600 __ClearPageReserved(p);
1601 set_page_count(p, 0);
1604 set_pageblock_migratetype(page, MIGRATE_CMA);
1606 if (pageblock_order >= MAX_ORDER) {
1607 i = pageblock_nr_pages;
1610 set_page_refcounted(p);
1611 __free_pages(p, MAX_ORDER - 1);
1612 p += MAX_ORDER_NR_PAGES;
1613 } while (i -= MAX_ORDER_NR_PAGES);
1615 set_page_refcounted(page);
1616 __free_pages(page, pageblock_order);
1619 adjust_managed_page_count(page, pageblock_nr_pages);
1624 * The order of subdivision here is critical for the IO subsystem.
1625 * Please do not alter this order without good reasons and regression
1626 * testing. Specifically, as large blocks of memory are subdivided,
1627 * the order in which smaller blocks are delivered depends on the order
1628 * they're subdivided in this function. This is the primary factor
1629 * influencing the order in which pages are delivered to the IO
1630 * subsystem according to empirical testing, and this is also justified
1631 * by considering the behavior of a buddy system containing a single
1632 * large block of memory acted on by a series of small allocations.
1633 * This behavior is a critical factor in sglist merging's success.
1637 static inline void expand(struct zone *zone, struct page *page,
1638 int low, int high, struct free_area *area,
1641 unsigned long size = 1 << high;
1643 while (high > low) {
1647 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1650 * Mark as guard pages (or page), that will allow to
1651 * merge back to allocator when buddy will be freed.
1652 * Corresponding page table entries will not be touched,
1653 * pages will stay not present in virtual address space
1655 if (set_page_guard(zone, &page[size], high, migratetype))
1658 list_add(&page[size].lru, &area->free_list[migratetype]);
1660 set_page_order(&page[size], high);
1664 static void check_new_page_bad(struct page *page)
1666 const char *bad_reason = NULL;
1667 unsigned long bad_flags = 0;
1669 if (unlikely(atomic_read(&page->_mapcount) != -1))
1670 bad_reason = "nonzero mapcount";
1671 if (unlikely(page->mapping != NULL))
1672 bad_reason = "non-NULL mapping";
1673 if (unlikely(page_ref_count(page) != 0))
1674 bad_reason = "nonzero _count";
1675 if (unlikely(page->flags & __PG_HWPOISON)) {
1676 bad_reason = "HWPoisoned (hardware-corrupted)";
1677 bad_flags = __PG_HWPOISON;
1678 /* Don't complain about hwpoisoned pages */
1679 page_mapcount_reset(page); /* remove PageBuddy */
1682 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1683 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1684 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1687 if (unlikely(page->mem_cgroup))
1688 bad_reason = "page still charged to cgroup";
1690 bad_page(page, bad_reason, bad_flags);
1694 * This page is about to be returned from the page allocator
1696 static inline int check_new_page(struct page *page)
1698 if (likely(page_expected_state(page,
1699 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1702 check_new_page_bad(page);
1706 static inline bool free_pages_prezeroed(void)
1708 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1709 page_poisoning_enabled();
1712 #ifdef CONFIG_DEBUG_VM
1713 static bool check_pcp_refill(struct page *page)
1718 static bool check_new_pcp(struct page *page)
1720 return check_new_page(page);
1723 static bool check_pcp_refill(struct page *page)
1725 return check_new_page(page);
1727 static bool check_new_pcp(struct page *page)
1731 #endif /* CONFIG_DEBUG_VM */
1733 static bool check_new_pages(struct page *page, unsigned int order)
1736 for (i = 0; i < (1 << order); i++) {
1737 struct page *p = page + i;
1739 if (unlikely(check_new_page(p)))
1746 inline void post_alloc_hook(struct page *page, unsigned int order,
1749 set_page_private(page, 0);
1750 set_page_refcounted(page);
1752 arch_alloc_page(page, order);
1753 kernel_map_pages(page, 1 << order, 1);
1754 kernel_poison_pages(page, 1 << order, 1);
1755 kasan_alloc_pages(page, order);
1756 set_page_owner(page, order, gfp_flags);
1759 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1760 unsigned int alloc_flags)
1764 post_alloc_hook(page, order, gfp_flags);
1766 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1767 for (i = 0; i < (1 << order); i++)
1768 clear_highpage(page + i);
1770 if (order && (gfp_flags & __GFP_COMP))
1771 prep_compound_page(page, order);
1774 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1775 * allocate the page. The expectation is that the caller is taking
1776 * steps that will free more memory. The caller should avoid the page
1777 * being used for !PFMEMALLOC purposes.
1779 if (alloc_flags & ALLOC_NO_WATERMARKS)
1780 set_page_pfmemalloc(page);
1782 clear_page_pfmemalloc(page);
1786 * Go through the free lists for the given migratetype and remove
1787 * the smallest available page from the freelists
1790 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1793 unsigned int current_order;
1794 struct free_area *area;
1797 /* Find a page of the appropriate size in the preferred list */
1798 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1799 area = &(zone->free_area[current_order]);
1800 page = list_first_entry_or_null(&area->free_list[migratetype],
1804 list_del(&page->lru);
1805 rmv_page_order(page);
1807 expand(zone, page, order, current_order, area, migratetype);
1808 set_pcppage_migratetype(page, migratetype);
1817 * This array describes the order lists are fallen back to when
1818 * the free lists for the desirable migrate type are depleted
1820 static int fallbacks[MIGRATE_TYPES][4] = {
1821 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1822 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1823 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1825 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1827 #ifdef CONFIG_MEMORY_ISOLATION
1828 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1833 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1836 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1839 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1840 unsigned int order) { return NULL; }
1844 * Move the free pages in a range to the free lists of the requested type.
1845 * Note that start_page and end_pages are not aligned on a pageblock
1846 * boundary. If alignment is required, use move_freepages_block()
1848 static int move_freepages(struct zone *zone,
1849 struct page *start_page, struct page *end_page,
1850 int migratetype, int *num_movable)
1854 int pages_moved = 0;
1856 #ifndef CONFIG_HOLES_IN_ZONE
1858 * page_zone is not safe to call in this context when
1859 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1860 * anyway as we check zone boundaries in move_freepages_block().
1861 * Remove at a later date when no bug reports exist related to
1862 * grouping pages by mobility
1864 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1870 for (page = start_page; page <= end_page;) {
1871 if (!pfn_valid_within(page_to_pfn(page))) {
1876 /* Make sure we are not inadvertently changing nodes */
1877 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1879 if (!PageBuddy(page)) {
1881 * We assume that pages that could be isolated for
1882 * migration are movable. But we don't actually try
1883 * isolating, as that would be expensive.
1886 (PageLRU(page) || __PageMovable(page)))
1893 order = page_order(page);
1894 list_move(&page->lru,
1895 &zone->free_area[order].free_list[migratetype]);
1897 pages_moved += 1 << order;
1903 int move_freepages_block(struct zone *zone, struct page *page,
1904 int migratetype, int *num_movable)
1906 unsigned long start_pfn, end_pfn;
1907 struct page *start_page, *end_page;
1909 start_pfn = page_to_pfn(page);
1910 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1911 start_page = pfn_to_page(start_pfn);
1912 end_page = start_page + pageblock_nr_pages - 1;
1913 end_pfn = start_pfn + pageblock_nr_pages - 1;
1915 /* Do not cross zone boundaries */
1916 if (!zone_spans_pfn(zone, start_pfn))
1918 if (!zone_spans_pfn(zone, end_pfn))
1921 return move_freepages(zone, start_page, end_page, migratetype,
1925 static void change_pageblock_range(struct page *pageblock_page,
1926 int start_order, int migratetype)
1928 int nr_pageblocks = 1 << (start_order - pageblock_order);
1930 while (nr_pageblocks--) {
1931 set_pageblock_migratetype(pageblock_page, migratetype);
1932 pageblock_page += pageblock_nr_pages;
1937 * When we are falling back to another migratetype during allocation, try to
1938 * steal extra free pages from the same pageblocks to satisfy further
1939 * allocations, instead of polluting multiple pageblocks.
1941 * If we are stealing a relatively large buddy page, it is likely there will
1942 * be more free pages in the pageblock, so try to steal them all. For
1943 * reclaimable and unmovable allocations, we steal regardless of page size,
1944 * as fragmentation caused by those allocations polluting movable pageblocks
1945 * is worse than movable allocations stealing from unmovable and reclaimable
1948 static bool can_steal_fallback(unsigned int order, int start_mt)
1951 * Leaving this order check is intended, although there is
1952 * relaxed order check in next check. The reason is that
1953 * we can actually steal whole pageblock if this condition met,
1954 * but, below check doesn't guarantee it and that is just heuristic
1955 * so could be changed anytime.
1957 if (order >= pageblock_order)
1960 if (order >= pageblock_order / 2 ||
1961 start_mt == MIGRATE_RECLAIMABLE ||
1962 start_mt == MIGRATE_UNMOVABLE ||
1963 page_group_by_mobility_disabled)
1970 * This function implements actual steal behaviour. If order is large enough,
1971 * we can steal whole pageblock. If not, we first move freepages in this
1972 * pageblock to our migratetype and determine how many already-allocated pages
1973 * are there in the pageblock with a compatible migratetype. If at least half
1974 * of pages are free or compatible, we can change migratetype of the pageblock
1975 * itself, so pages freed in the future will be put on the correct free list.
1977 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1978 int start_type, bool whole_block)
1980 unsigned int current_order = page_order(page);
1981 struct free_area *area;
1982 int free_pages, movable_pages, alike_pages;
1985 old_block_type = get_pageblock_migratetype(page);
1988 * This can happen due to races and we want to prevent broken
1989 * highatomic accounting.
1991 if (is_migrate_highatomic(old_block_type))
1994 /* Take ownership for orders >= pageblock_order */
1995 if (current_order >= pageblock_order) {
1996 change_pageblock_range(page, current_order, start_type);
2000 /* We are not allowed to try stealing from the whole block */
2004 free_pages = move_freepages_block(zone, page, start_type,
2007 * Determine how many pages are compatible with our allocation.
2008 * For movable allocation, it's the number of movable pages which
2009 * we just obtained. For other types it's a bit more tricky.
2011 if (start_type == MIGRATE_MOVABLE) {
2012 alike_pages = movable_pages;
2015 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2016 * to MOVABLE pageblock, consider all non-movable pages as
2017 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2018 * vice versa, be conservative since we can't distinguish the
2019 * exact migratetype of non-movable pages.
2021 if (old_block_type == MIGRATE_MOVABLE)
2022 alike_pages = pageblock_nr_pages
2023 - (free_pages + movable_pages);
2028 /* moving whole block can fail due to zone boundary conditions */
2033 * If a sufficient number of pages in the block are either free or of
2034 * comparable migratability as our allocation, claim the whole block.
2036 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2037 page_group_by_mobility_disabled)
2038 set_pageblock_migratetype(page, start_type);
2043 area = &zone->free_area[current_order];
2044 list_move(&page->lru, &area->free_list[start_type]);
2048 * Check whether there is a suitable fallback freepage with requested order.
2049 * If only_stealable is true, this function returns fallback_mt only if
2050 * we can steal other freepages all together. This would help to reduce
2051 * fragmentation due to mixed migratetype pages in one pageblock.
2053 int find_suitable_fallback(struct free_area *area, unsigned int order,
2054 int migratetype, bool only_stealable, bool *can_steal)
2059 if (area->nr_free == 0)
2064 fallback_mt = fallbacks[migratetype][i];
2065 if (fallback_mt == MIGRATE_TYPES)
2068 if (list_empty(&area->free_list[fallback_mt]))
2071 if (can_steal_fallback(order, migratetype))
2074 if (!only_stealable)
2085 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2086 * there are no empty page blocks that contain a page with a suitable order
2088 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2089 unsigned int alloc_order)
2092 unsigned long max_managed, flags;
2095 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2096 * Check is race-prone but harmless.
2098 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2099 if (zone->nr_reserved_highatomic >= max_managed)
2102 spin_lock_irqsave(&zone->lock, flags);
2104 /* Recheck the nr_reserved_highatomic limit under the lock */
2105 if (zone->nr_reserved_highatomic >= max_managed)
2109 mt = get_pageblock_migratetype(page);
2110 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2111 && !is_migrate_cma(mt)) {
2112 zone->nr_reserved_highatomic += pageblock_nr_pages;
2113 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2114 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2118 spin_unlock_irqrestore(&zone->lock, flags);
2122 * Used when an allocation is about to fail under memory pressure. This
2123 * potentially hurts the reliability of high-order allocations when under
2124 * intense memory pressure but failed atomic allocations should be easier
2125 * to recover from than an OOM.
2127 * If @force is true, try to unreserve a pageblock even though highatomic
2128 * pageblock is exhausted.
2130 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2133 struct zonelist *zonelist = ac->zonelist;
2134 unsigned long flags;
2141 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2144 * Preserve at least one pageblock unless memory pressure
2147 if (!force && zone->nr_reserved_highatomic <=
2151 spin_lock_irqsave(&zone->lock, flags);
2152 for (order = 0; order < MAX_ORDER; order++) {
2153 struct free_area *area = &(zone->free_area[order]);
2155 page = list_first_entry_or_null(
2156 &area->free_list[MIGRATE_HIGHATOMIC],
2162 * In page freeing path, migratetype change is racy so
2163 * we can counter several free pages in a pageblock
2164 * in this loop althoug we changed the pageblock type
2165 * from highatomic to ac->migratetype. So we should
2166 * adjust the count once.
2168 if (is_migrate_highatomic_page(page)) {
2170 * It should never happen but changes to
2171 * locking could inadvertently allow a per-cpu
2172 * drain to add pages to MIGRATE_HIGHATOMIC
2173 * while unreserving so be safe and watch for
2176 zone->nr_reserved_highatomic -= min(
2178 zone->nr_reserved_highatomic);
2182 * Convert to ac->migratetype and avoid the normal
2183 * pageblock stealing heuristics. Minimally, the caller
2184 * is doing the work and needs the pages. More
2185 * importantly, if the block was always converted to
2186 * MIGRATE_UNMOVABLE or another type then the number
2187 * of pageblocks that cannot be completely freed
2190 set_pageblock_migratetype(page, ac->migratetype);
2191 ret = move_freepages_block(zone, page, ac->migratetype,
2194 spin_unlock_irqrestore(&zone->lock, flags);
2198 spin_unlock_irqrestore(&zone->lock, flags);
2205 * Try finding a free buddy page on the fallback list and put it on the free
2206 * list of requested migratetype, possibly along with other pages from the same
2207 * block, depending on fragmentation avoidance heuristics. Returns true if
2208 * fallback was found so that __rmqueue_smallest() can grab it.
2211 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2213 struct free_area *area;
2214 unsigned int current_order;
2220 * Find the largest available free page in the other list. This roughly
2221 * approximates finding the pageblock with the most free pages, which
2222 * would be too costly to do exactly.
2224 for (current_order = MAX_ORDER-1;
2225 current_order >= order && current_order <= MAX_ORDER-1;
2227 area = &(zone->free_area[current_order]);
2228 fallback_mt = find_suitable_fallback(area, current_order,
2229 start_migratetype, false, &can_steal);
2230 if (fallback_mt == -1)
2234 * We cannot steal all free pages from the pageblock and the
2235 * requested migratetype is movable. In that case it's better to
2236 * steal and split the smallest available page instead of the
2237 * largest available page, because even if the next movable
2238 * allocation falls back into a different pageblock than this
2239 * one, it won't cause permanent fragmentation.
2241 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2242 && current_order > order)
2251 for (current_order = order; current_order < MAX_ORDER;
2253 area = &(zone->free_area[current_order]);
2254 fallback_mt = find_suitable_fallback(area, current_order,
2255 start_migratetype, false, &can_steal);
2256 if (fallback_mt != -1)
2261 * This should not happen - we already found a suitable fallback
2262 * when looking for the largest page.
2264 VM_BUG_ON(current_order == MAX_ORDER);
2267 page = list_first_entry(&area->free_list[fallback_mt],
2270 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2272 trace_mm_page_alloc_extfrag(page, order, current_order,
2273 start_migratetype, fallback_mt);
2280 * Do the hard work of removing an element from the buddy allocator.
2281 * Call me with the zone->lock already held.
2283 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2289 page = __rmqueue_smallest(zone, order, migratetype);
2290 if (unlikely(!page)) {
2291 if (migratetype == MIGRATE_MOVABLE)
2292 page = __rmqueue_cma_fallback(zone, order);
2294 if (!page && __rmqueue_fallback(zone, order, migratetype))
2298 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2303 * Obtain a specified number of elements from the buddy allocator, all under
2304 * a single hold of the lock, for efficiency. Add them to the supplied list.
2305 * Returns the number of new pages which were placed at *list.
2307 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2308 unsigned long count, struct list_head *list,
2309 int migratetype, bool cold)
2313 spin_lock(&zone->lock);
2314 for (i = 0; i < count; ++i) {
2315 struct page *page = __rmqueue(zone, order, migratetype);
2316 if (unlikely(page == NULL))
2319 if (unlikely(check_pcp_refill(page)))
2323 * Split buddy pages returned by expand() are received here
2324 * in physical page order. The page is added to the callers and
2325 * list and the list head then moves forward. From the callers
2326 * perspective, the linked list is ordered by page number in
2327 * some conditions. This is useful for IO devices that can
2328 * merge IO requests if the physical pages are ordered
2332 list_add(&page->lru, list);
2334 list_add_tail(&page->lru, list);
2337 if (is_migrate_cma(get_pcppage_migratetype(page)))
2338 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2343 * i pages were removed from the buddy list even if some leak due
2344 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2345 * on i. Do not confuse with 'alloced' which is the number of
2346 * pages added to the pcp list.
2348 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2349 spin_unlock(&zone->lock);
2355 * Called from the vmstat counter updater to drain pagesets of this
2356 * currently executing processor on remote nodes after they have
2359 * Note that this function must be called with the thread pinned to
2360 * a single processor.
2362 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2364 unsigned long flags;
2365 int to_drain, batch;
2367 local_irq_save(flags);
2368 batch = READ_ONCE(pcp->batch);
2369 to_drain = min(pcp->count, batch);
2371 free_pcppages_bulk(zone, to_drain, pcp);
2372 pcp->count -= to_drain;
2374 local_irq_restore(flags);
2379 * Drain pcplists of the indicated processor and zone.
2381 * The processor must either be the current processor and the
2382 * thread pinned to the current processor or a processor that
2385 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2387 unsigned long flags;
2388 struct per_cpu_pageset *pset;
2389 struct per_cpu_pages *pcp;
2391 local_irq_save(flags);
2392 pset = per_cpu_ptr(zone->pageset, cpu);
2396 free_pcppages_bulk(zone, pcp->count, pcp);
2399 local_irq_restore(flags);
2403 * Drain pcplists of all zones on the indicated processor.
2405 * The processor must either be the current processor and the
2406 * thread pinned to the current processor or a processor that
2409 static void drain_pages(unsigned int cpu)
2413 for_each_populated_zone(zone) {
2414 drain_pages_zone(cpu, zone);
2419 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2421 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2422 * the single zone's pages.
2424 void drain_local_pages(struct zone *zone)
2426 int cpu = smp_processor_id();
2429 drain_pages_zone(cpu, zone);
2434 static void drain_local_pages_wq(struct work_struct *work)
2437 * drain_all_pages doesn't use proper cpu hotplug protection so
2438 * we can race with cpu offline when the WQ can move this from
2439 * a cpu pinned worker to an unbound one. We can operate on a different
2440 * cpu which is allright but we also have to make sure to not move to
2444 drain_local_pages(NULL);
2449 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2451 * When zone parameter is non-NULL, spill just the single zone's pages.
2453 * Note that this can be extremely slow as the draining happens in a workqueue.
2455 void drain_all_pages(struct zone *zone)
2460 * Allocate in the BSS so we wont require allocation in
2461 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2463 static cpumask_t cpus_with_pcps;
2466 * Make sure nobody triggers this path before mm_percpu_wq is fully
2469 if (WARN_ON_ONCE(!mm_percpu_wq))
2472 /* Workqueues cannot recurse */
2473 if (current->flags & PF_WQ_WORKER)
2477 * Do not drain if one is already in progress unless it's specific to
2478 * a zone. Such callers are primarily CMA and memory hotplug and need
2479 * the drain to be complete when the call returns.
2481 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2484 mutex_lock(&pcpu_drain_mutex);
2488 * We don't care about racing with CPU hotplug event
2489 * as offline notification will cause the notified
2490 * cpu to drain that CPU pcps and on_each_cpu_mask
2491 * disables preemption as part of its processing
2493 for_each_online_cpu(cpu) {
2494 struct per_cpu_pageset *pcp;
2496 bool has_pcps = false;
2499 pcp = per_cpu_ptr(zone->pageset, cpu);
2503 for_each_populated_zone(z) {
2504 pcp = per_cpu_ptr(z->pageset, cpu);
2505 if (pcp->pcp.count) {
2513 cpumask_set_cpu(cpu, &cpus_with_pcps);
2515 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2518 for_each_cpu(cpu, &cpus_with_pcps) {
2519 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2520 INIT_WORK(work, drain_local_pages_wq);
2521 queue_work_on(cpu, mm_percpu_wq, work);
2523 for_each_cpu(cpu, &cpus_with_pcps)
2524 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2526 mutex_unlock(&pcpu_drain_mutex);
2529 #ifdef CONFIG_HIBERNATION
2531 void mark_free_pages(struct zone *zone)
2533 unsigned long pfn, max_zone_pfn;
2534 unsigned long flags;
2535 unsigned int order, t;
2538 if (zone_is_empty(zone))
2541 spin_lock_irqsave(&zone->lock, flags);
2543 max_zone_pfn = zone_end_pfn(zone);
2544 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2545 if (pfn_valid(pfn)) {
2546 page = pfn_to_page(pfn);
2548 if (page_zone(page) != zone)
2551 if (!swsusp_page_is_forbidden(page))
2552 swsusp_unset_page_free(page);
2555 for_each_migratetype_order(order, t) {
2556 list_for_each_entry(page,
2557 &zone->free_area[order].free_list[t], lru) {
2560 pfn = page_to_pfn(page);
2561 for (i = 0; i < (1UL << order); i++)
2562 swsusp_set_page_free(pfn_to_page(pfn + i));
2565 spin_unlock_irqrestore(&zone->lock, flags);
2567 #endif /* CONFIG_PM */
2570 * Free a 0-order page
2571 * cold == true ? free a cold page : free a hot page
2573 void free_hot_cold_page(struct page *page, bool cold)
2575 struct zone *zone = page_zone(page);
2576 struct per_cpu_pages *pcp;
2577 unsigned long flags;
2578 unsigned long pfn = page_to_pfn(page);
2581 if (!free_pcp_prepare(page))
2584 migratetype = get_pfnblock_migratetype(page, pfn);
2585 set_pcppage_migratetype(page, migratetype);
2586 local_irq_save(flags);
2587 __count_vm_event(PGFREE);
2590 * We only track unmovable, reclaimable and movable on pcp lists.
2591 * Free ISOLATE pages back to the allocator because they are being
2592 * offlined but treat HIGHATOMIC as movable pages so we can get those
2593 * areas back if necessary. Otherwise, we may have to free
2594 * excessively into the page allocator
2596 if (migratetype >= MIGRATE_PCPTYPES) {
2597 if (unlikely(is_migrate_isolate(migratetype))) {
2598 free_one_page(zone, page, pfn, 0, migratetype);
2601 migratetype = MIGRATE_MOVABLE;
2604 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2606 list_add(&page->lru, &pcp->lists[migratetype]);
2608 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2610 if (pcp->count >= pcp->high) {
2611 unsigned long batch = READ_ONCE(pcp->batch);
2612 free_pcppages_bulk(zone, batch, pcp);
2613 pcp->count -= batch;
2617 local_irq_restore(flags);
2621 * Free a list of 0-order pages
2623 void free_hot_cold_page_list(struct list_head *list, bool cold)
2625 struct page *page, *next;
2627 list_for_each_entry_safe(page, next, list, lru) {
2628 trace_mm_page_free_batched(page, cold);
2629 free_hot_cold_page(page, cold);
2634 * split_page takes a non-compound higher-order page, and splits it into
2635 * n (1<<order) sub-pages: page[0..n]
2636 * Each sub-page must be freed individually.
2638 * Note: this is probably too low level an operation for use in drivers.
2639 * Please consult with lkml before using this in your driver.
2641 void split_page(struct page *page, unsigned int order)
2645 VM_BUG_ON_PAGE(PageCompound(page), page);
2646 VM_BUG_ON_PAGE(!page_count(page), page);
2648 #ifdef CONFIG_KMEMCHECK
2650 * Split shadow pages too, because free(page[0]) would
2651 * otherwise free the whole shadow.
2653 if (kmemcheck_page_is_tracked(page))
2654 split_page(virt_to_page(page[0].shadow), order);
2657 for (i = 1; i < (1 << order); i++)
2658 set_page_refcounted(page + i);
2659 split_page_owner(page, order);
2661 EXPORT_SYMBOL_GPL(split_page);
2663 int __isolate_free_page(struct page *page, unsigned int order)
2665 unsigned long watermark;
2669 BUG_ON(!PageBuddy(page));
2671 zone = page_zone(page);
2672 mt = get_pageblock_migratetype(page);
2674 if (!is_migrate_isolate(mt)) {
2676 * Obey watermarks as if the page was being allocated. We can
2677 * emulate a high-order watermark check with a raised order-0
2678 * watermark, because we already know our high-order page
2681 watermark = min_wmark_pages(zone) + (1UL << order);
2682 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2685 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2688 /* Remove page from free list */
2689 list_del(&page->lru);
2690 zone->free_area[order].nr_free--;
2691 rmv_page_order(page);
2694 * Set the pageblock if the isolated page is at least half of a
2697 if (order >= pageblock_order - 1) {
2698 struct page *endpage = page + (1 << order) - 1;
2699 for (; page < endpage; page += pageblock_nr_pages) {
2700 int mt = get_pageblock_migratetype(page);
2701 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2702 && !is_migrate_highatomic(mt))
2703 set_pageblock_migratetype(page,
2709 return 1UL << order;
2713 * Update NUMA hit/miss statistics
2715 * Must be called with interrupts disabled.
2717 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2720 enum zone_stat_item local_stat = NUMA_LOCAL;
2722 if (z->node != numa_node_id())
2723 local_stat = NUMA_OTHER;
2725 if (z->node == preferred_zone->node)
2726 __inc_zone_state(z, NUMA_HIT);
2728 __inc_zone_state(z, NUMA_MISS);
2729 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2731 __inc_zone_state(z, local_stat);
2735 /* Remove page from the per-cpu list, caller must protect the list */
2736 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2737 bool cold, struct per_cpu_pages *pcp,
2738 struct list_head *list)
2743 if (list_empty(list)) {
2744 pcp->count += rmqueue_bulk(zone, 0,
2747 if (unlikely(list_empty(list)))
2752 page = list_last_entry(list, struct page, lru);
2754 page = list_first_entry(list, struct page, lru);
2756 list_del(&page->lru);
2758 } while (check_new_pcp(page));
2763 /* Lock and remove page from the per-cpu list */
2764 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2765 struct zone *zone, unsigned int order,
2766 gfp_t gfp_flags, int migratetype)
2768 struct per_cpu_pages *pcp;
2769 struct list_head *list;
2770 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2772 unsigned long flags;
2774 local_irq_save(flags);
2775 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2776 list = &pcp->lists[migratetype];
2777 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2779 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2780 zone_statistics(preferred_zone, zone);
2782 local_irq_restore(flags);
2787 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2790 struct page *rmqueue(struct zone *preferred_zone,
2791 struct zone *zone, unsigned int order,
2792 gfp_t gfp_flags, unsigned int alloc_flags,
2795 unsigned long flags;
2798 if (likely(order == 0)) {
2799 page = rmqueue_pcplist(preferred_zone, zone, order,
2800 gfp_flags, migratetype);
2805 * We most definitely don't want callers attempting to
2806 * allocate greater than order-1 page units with __GFP_NOFAIL.
2808 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2809 spin_lock_irqsave(&zone->lock, flags);
2813 if (alloc_flags & ALLOC_HARDER) {
2814 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2816 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2819 page = __rmqueue(zone, order, migratetype);
2820 } while (page && check_new_pages(page, order));
2821 spin_unlock(&zone->lock);
2824 __mod_zone_freepage_state(zone, -(1 << order),
2825 get_pcppage_migratetype(page));
2827 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2828 zone_statistics(preferred_zone, zone);
2829 local_irq_restore(flags);
2832 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2836 local_irq_restore(flags);
2840 #ifdef CONFIG_FAIL_PAGE_ALLOC
2843 struct fault_attr attr;
2845 bool ignore_gfp_highmem;
2846 bool ignore_gfp_reclaim;
2848 } fail_page_alloc = {
2849 .attr = FAULT_ATTR_INITIALIZER,
2850 .ignore_gfp_reclaim = true,
2851 .ignore_gfp_highmem = true,
2855 static int __init setup_fail_page_alloc(char *str)
2857 return setup_fault_attr(&fail_page_alloc.attr, str);
2859 __setup("fail_page_alloc=", setup_fail_page_alloc);
2861 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2863 if (order < fail_page_alloc.min_order)
2865 if (gfp_mask & __GFP_NOFAIL)
2867 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2869 if (fail_page_alloc.ignore_gfp_reclaim &&
2870 (gfp_mask & __GFP_DIRECT_RECLAIM))
2873 return should_fail(&fail_page_alloc.attr, 1 << order);
2876 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2878 static int __init fail_page_alloc_debugfs(void)
2880 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2883 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2884 &fail_page_alloc.attr);
2886 return PTR_ERR(dir);
2888 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2889 &fail_page_alloc.ignore_gfp_reclaim))
2891 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2892 &fail_page_alloc.ignore_gfp_highmem))
2894 if (!debugfs_create_u32("min-order", mode, dir,
2895 &fail_page_alloc.min_order))
2900 debugfs_remove_recursive(dir);
2905 late_initcall(fail_page_alloc_debugfs);
2907 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2909 #else /* CONFIG_FAIL_PAGE_ALLOC */
2911 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2916 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2919 * Return true if free base pages are above 'mark'. For high-order checks it
2920 * will return true of the order-0 watermark is reached and there is at least
2921 * one free page of a suitable size. Checking now avoids taking the zone lock
2922 * to check in the allocation paths if no pages are free.
2924 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2925 int classzone_idx, unsigned int alloc_flags,
2930 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2932 /* free_pages may go negative - that's OK */
2933 free_pages -= (1 << order) - 1;
2935 if (alloc_flags & ALLOC_HIGH)
2939 * If the caller does not have rights to ALLOC_HARDER then subtract
2940 * the high-atomic reserves. This will over-estimate the size of the
2941 * atomic reserve but it avoids a search.
2943 if (likely(!alloc_harder))
2944 free_pages -= z->nr_reserved_highatomic;
2949 /* If allocation can't use CMA areas don't use free CMA pages */
2950 if (!(alloc_flags & ALLOC_CMA))
2951 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2955 * Check watermarks for an order-0 allocation request. If these
2956 * are not met, then a high-order request also cannot go ahead
2957 * even if a suitable page happened to be free.
2959 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2962 /* If this is an order-0 request then the watermark is fine */
2966 /* For a high-order request, check at least one suitable page is free */
2967 for (o = order; o < MAX_ORDER; o++) {
2968 struct free_area *area = &z->free_area[o];
2977 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2978 if (!list_empty(&area->free_list[mt]))
2983 if ((alloc_flags & ALLOC_CMA) &&
2984 !list_empty(&area->free_list[MIGRATE_CMA])) {
2992 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2993 int classzone_idx, unsigned int alloc_flags)
2995 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2996 zone_page_state(z, NR_FREE_PAGES));
2999 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3000 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3002 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3006 /* If allocation can't use CMA areas don't use free CMA pages */
3007 if (!(alloc_flags & ALLOC_CMA))
3008 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3012 * Fast check for order-0 only. If this fails then the reserves
3013 * need to be calculated. There is a corner case where the check
3014 * passes but only the high-order atomic reserve are free. If
3015 * the caller is !atomic then it'll uselessly search the free
3016 * list. That corner case is then slower but it is harmless.
3018 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3021 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3025 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3026 unsigned long mark, int classzone_idx)
3028 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3030 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3031 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3033 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3038 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3040 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3043 #else /* CONFIG_NUMA */
3044 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3048 #endif /* CONFIG_NUMA */
3051 * get_page_from_freelist goes through the zonelist trying to allocate
3054 static struct page *
3055 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3056 const struct alloc_context *ac)
3058 struct zoneref *z = ac->preferred_zoneref;
3060 struct pglist_data *last_pgdat_dirty_limit = NULL;
3063 * Scan zonelist, looking for a zone with enough free.
3064 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3066 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3071 if (cpusets_enabled() &&
3072 (alloc_flags & ALLOC_CPUSET) &&
3073 !__cpuset_zone_allowed(zone, gfp_mask))
3076 * When allocating a page cache page for writing, we
3077 * want to get it from a node that is within its dirty
3078 * limit, such that no single node holds more than its
3079 * proportional share of globally allowed dirty pages.
3080 * The dirty limits take into account the node's
3081 * lowmem reserves and high watermark so that kswapd
3082 * should be able to balance it without having to
3083 * write pages from its LRU list.
3085 * XXX: For now, allow allocations to potentially
3086 * exceed the per-node dirty limit in the slowpath
3087 * (spread_dirty_pages unset) before going into reclaim,
3088 * which is important when on a NUMA setup the allowed
3089 * nodes are together not big enough to reach the
3090 * global limit. The proper fix for these situations
3091 * will require awareness of nodes in the
3092 * dirty-throttling and the flusher threads.
3094 if (ac->spread_dirty_pages) {
3095 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3098 if (!node_dirty_ok(zone->zone_pgdat)) {
3099 last_pgdat_dirty_limit = zone->zone_pgdat;
3104 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3105 if (!zone_watermark_fast(zone, order, mark,
3106 ac_classzone_idx(ac), alloc_flags)) {
3109 /* Checked here to keep the fast path fast */
3110 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3111 if (alloc_flags & ALLOC_NO_WATERMARKS)
3114 if (node_reclaim_mode == 0 ||
3115 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3118 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3120 case NODE_RECLAIM_NOSCAN:
3123 case NODE_RECLAIM_FULL:
3124 /* scanned but unreclaimable */
3127 /* did we reclaim enough */
3128 if (zone_watermark_ok(zone, order, mark,
3129 ac_classzone_idx(ac), alloc_flags))
3137 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3138 gfp_mask, alloc_flags, ac->migratetype);
3140 prep_new_page(page, order, gfp_mask, alloc_flags);
3143 * If this is a high-order atomic allocation then check
3144 * if the pageblock should be reserved for the future
3146 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3147 reserve_highatomic_pageblock(page, zone, order);
3157 * Large machines with many possible nodes should not always dump per-node
3158 * meminfo in irq context.
3160 static inline bool should_suppress_show_mem(void)
3165 ret = in_interrupt();
3170 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3172 unsigned int filter = SHOW_MEM_FILTER_NODES;
3173 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3175 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3179 * This documents exceptions given to allocations in certain
3180 * contexts that are allowed to allocate outside current's set
3183 if (!(gfp_mask & __GFP_NOMEMALLOC))
3184 if (test_thread_flag(TIF_MEMDIE) ||
3185 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3186 filter &= ~SHOW_MEM_FILTER_NODES;
3187 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3188 filter &= ~SHOW_MEM_FILTER_NODES;
3190 show_mem(filter, nodemask);
3193 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3195 struct va_format vaf;
3197 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3198 DEFAULT_RATELIMIT_BURST);
3200 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3203 pr_warn("%s: ", current->comm);
3205 va_start(args, fmt);
3208 pr_cont("%pV", &vaf);
3211 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3213 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3215 pr_cont("(null)\n");
3217 cpuset_print_current_mems_allowed();
3220 warn_alloc_show_mem(gfp_mask, nodemask);
3223 static inline struct page *
3224 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3225 unsigned int alloc_flags,
3226 const struct alloc_context *ac)
3230 page = get_page_from_freelist(gfp_mask, order,
3231 alloc_flags|ALLOC_CPUSET, ac);
3233 * fallback to ignore cpuset restriction if our nodes
3237 page = get_page_from_freelist(gfp_mask, order,
3243 static inline struct page *
3244 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3245 const struct alloc_context *ac, unsigned long *did_some_progress)
3247 struct oom_control oc = {
3248 .zonelist = ac->zonelist,
3249 .nodemask = ac->nodemask,
3251 .gfp_mask = gfp_mask,
3256 *did_some_progress = 0;
3259 * Acquire the oom lock. If that fails, somebody else is
3260 * making progress for us.
3262 if (!mutex_trylock(&oom_lock)) {
3263 *did_some_progress = 1;
3264 schedule_timeout_uninterruptible(1);
3269 * Go through the zonelist yet one more time, keep very high watermark
3270 * here, this is only to catch a parallel oom killing, we must fail if
3271 * we're still under heavy pressure.
3273 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3274 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3278 /* Coredumps can quickly deplete all memory reserves */
3279 if (current->flags & PF_DUMPCORE)
3281 /* The OOM killer will not help higher order allocs */
3282 if (order > PAGE_ALLOC_COSTLY_ORDER)
3284 /* The OOM killer does not needlessly kill tasks for lowmem */
3285 if (ac->high_zoneidx < ZONE_NORMAL)
3287 if (pm_suspended_storage())
3290 * XXX: GFP_NOFS allocations should rather fail than rely on
3291 * other request to make a forward progress.
3292 * We are in an unfortunate situation where out_of_memory cannot
3293 * do much for this context but let's try it to at least get
3294 * access to memory reserved if the current task is killed (see
3295 * out_of_memory). Once filesystems are ready to handle allocation
3296 * failures more gracefully we should just bail out here.
3299 /* The OOM killer may not free memory on a specific node */
3300 if (gfp_mask & __GFP_THISNODE)
3303 /* Exhausted what can be done so it's blamo time */
3304 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3305 *did_some_progress = 1;
3308 * Help non-failing allocations by giving them access to memory
3311 if (gfp_mask & __GFP_NOFAIL)
3312 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3313 ALLOC_NO_WATERMARKS, ac);
3316 mutex_unlock(&oom_lock);
3321 * Maximum number of compaction retries wit a progress before OOM
3322 * killer is consider as the only way to move forward.
3324 #define MAX_COMPACT_RETRIES 16
3326 #ifdef CONFIG_COMPACTION
3327 /* Try memory compaction for high-order allocations before reclaim */
3328 static struct page *
3329 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3330 unsigned int alloc_flags, const struct alloc_context *ac,
3331 enum compact_priority prio, enum compact_result *compact_result)
3334 unsigned int noreclaim_flag;
3339 noreclaim_flag = memalloc_noreclaim_save();
3340 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3342 memalloc_noreclaim_restore(noreclaim_flag);
3344 if (*compact_result <= COMPACT_INACTIVE)
3348 * At least in one zone compaction wasn't deferred or skipped, so let's
3349 * count a compaction stall
3351 count_vm_event(COMPACTSTALL);
3353 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3356 struct zone *zone = page_zone(page);
3358 zone->compact_blockskip_flush = false;
3359 compaction_defer_reset(zone, order, true);
3360 count_vm_event(COMPACTSUCCESS);
3365 * It's bad if compaction run occurs and fails. The most likely reason
3366 * is that pages exist, but not enough to satisfy watermarks.
3368 count_vm_event(COMPACTFAIL);
3376 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3377 enum compact_result compact_result,
3378 enum compact_priority *compact_priority,
3379 int *compaction_retries)
3381 int max_retries = MAX_COMPACT_RETRIES;
3384 int retries = *compaction_retries;
3385 enum compact_priority priority = *compact_priority;
3390 if (compaction_made_progress(compact_result))
3391 (*compaction_retries)++;
3394 * compaction considers all the zone as desperately out of memory
3395 * so it doesn't really make much sense to retry except when the
3396 * failure could be caused by insufficient priority
3398 if (compaction_failed(compact_result))
3399 goto check_priority;
3402 * make sure the compaction wasn't deferred or didn't bail out early
3403 * due to locks contention before we declare that we should give up.
3404 * But do not retry if the given zonelist is not suitable for
3407 if (compaction_withdrawn(compact_result)) {
3408 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3413 * !costly requests are much more important than __GFP_REPEAT
3414 * costly ones because they are de facto nofail and invoke OOM
3415 * killer to move on while costly can fail and users are ready
3416 * to cope with that. 1/4 retries is rather arbitrary but we
3417 * would need much more detailed feedback from compaction to
3418 * make a better decision.
3420 if (order > PAGE_ALLOC_COSTLY_ORDER)
3422 if (*compaction_retries <= max_retries) {
3428 * Make sure there are attempts at the highest priority if we exhausted
3429 * all retries or failed at the lower priorities.
3432 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3433 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3435 if (*compact_priority > min_priority) {
3436 (*compact_priority)--;
3437 *compaction_retries = 0;
3441 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3445 static inline struct page *
3446 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3447 unsigned int alloc_flags, const struct alloc_context *ac,
3448 enum compact_priority prio, enum compact_result *compact_result)
3450 *compact_result = COMPACT_SKIPPED;
3455 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3456 enum compact_result compact_result,
3457 enum compact_priority *compact_priority,
3458 int *compaction_retries)
3463 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3467 * There are setups with compaction disabled which would prefer to loop
3468 * inside the allocator rather than hit the oom killer prematurely.
3469 * Let's give them a good hope and keep retrying while the order-0
3470 * watermarks are OK.
3472 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3474 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3475 ac_classzone_idx(ac), alloc_flags))
3480 #endif /* CONFIG_COMPACTION */
3482 /* Perform direct synchronous page reclaim */
3484 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3485 const struct alloc_context *ac)
3487 struct reclaim_state reclaim_state;
3489 unsigned int noreclaim_flag;
3493 /* We now go into synchronous reclaim */
3494 cpuset_memory_pressure_bump();
3495 noreclaim_flag = memalloc_noreclaim_save();
3496 lockdep_set_current_reclaim_state(gfp_mask);
3497 reclaim_state.reclaimed_slab = 0;
3498 current->reclaim_state = &reclaim_state;
3500 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3503 current->reclaim_state = NULL;
3504 lockdep_clear_current_reclaim_state();
3505 memalloc_noreclaim_restore(noreclaim_flag);
3512 /* The really slow allocator path where we enter direct reclaim */
3513 static inline struct page *
3514 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3515 unsigned int alloc_flags, const struct alloc_context *ac,
3516 unsigned long *did_some_progress)
3518 struct page *page = NULL;
3519 bool drained = false;
3521 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3522 if (unlikely(!(*did_some_progress)))
3526 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3529 * If an allocation failed after direct reclaim, it could be because
3530 * pages are pinned on the per-cpu lists or in high alloc reserves.
3531 * Shrink them them and try again
3533 if (!page && !drained) {
3534 unreserve_highatomic_pageblock(ac, false);
3535 drain_all_pages(NULL);
3543 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3547 pg_data_t *last_pgdat = NULL;
3549 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3550 ac->high_zoneidx, ac->nodemask) {
3551 if (last_pgdat != zone->zone_pgdat)
3552 wakeup_kswapd(zone, order, ac->high_zoneidx);
3553 last_pgdat = zone->zone_pgdat;
3557 static inline unsigned int
3558 gfp_to_alloc_flags(gfp_t gfp_mask)
3560 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3562 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3563 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3566 * The caller may dip into page reserves a bit more if the caller
3567 * cannot run direct reclaim, or if the caller has realtime scheduling
3568 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3569 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3571 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3573 if (gfp_mask & __GFP_ATOMIC) {
3575 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3576 * if it can't schedule.
3578 if (!(gfp_mask & __GFP_NOMEMALLOC))
3579 alloc_flags |= ALLOC_HARDER;
3581 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3582 * comment for __cpuset_node_allowed().
3584 alloc_flags &= ~ALLOC_CPUSET;
3585 } else if (unlikely(rt_task(current)) && !in_interrupt())
3586 alloc_flags |= ALLOC_HARDER;
3589 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3590 alloc_flags |= ALLOC_CMA;
3595 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3597 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3600 if (gfp_mask & __GFP_MEMALLOC)
3602 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3604 if (!in_interrupt() &&
3605 ((current->flags & PF_MEMALLOC) ||
3606 unlikely(test_thread_flag(TIF_MEMDIE))))
3613 * Checks whether it makes sense to retry the reclaim to make a forward progress
3614 * for the given allocation request.
3616 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3617 * without success, or when we couldn't even meet the watermark if we
3618 * reclaimed all remaining pages on the LRU lists.
3620 * Returns true if a retry is viable or false to enter the oom path.
3623 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3624 struct alloc_context *ac, int alloc_flags,
3625 bool did_some_progress, int *no_progress_loops)
3631 * Costly allocations might have made a progress but this doesn't mean
3632 * their order will become available due to high fragmentation so
3633 * always increment the no progress counter for them
3635 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3636 *no_progress_loops = 0;
3638 (*no_progress_loops)++;
3641 * Make sure we converge to OOM if we cannot make any progress
3642 * several times in the row.
3644 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3645 /* Before OOM, exhaust highatomic_reserve */
3646 return unreserve_highatomic_pageblock(ac, true);
3650 * Keep reclaiming pages while there is a chance this will lead
3651 * somewhere. If none of the target zones can satisfy our allocation
3652 * request even if all reclaimable pages are considered then we are
3653 * screwed and have to go OOM.
3655 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3657 unsigned long available;
3658 unsigned long reclaimable;
3659 unsigned long min_wmark = min_wmark_pages(zone);
3662 available = reclaimable = zone_reclaimable_pages(zone);
3663 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3666 * Would the allocation succeed if we reclaimed all
3667 * reclaimable pages?
3669 wmark = __zone_watermark_ok(zone, order, min_wmark,
3670 ac_classzone_idx(ac), alloc_flags, available);
3671 trace_reclaim_retry_zone(z, order, reclaimable,
3672 available, min_wmark, *no_progress_loops, wmark);
3675 * If we didn't make any progress and have a lot of
3676 * dirty + writeback pages then we should wait for
3677 * an IO to complete to slow down the reclaim and
3678 * prevent from pre mature OOM
3680 if (!did_some_progress) {
3681 unsigned long write_pending;
3683 write_pending = zone_page_state_snapshot(zone,
3684 NR_ZONE_WRITE_PENDING);
3686 if (2 * write_pending > reclaimable) {
3687 congestion_wait(BLK_RW_ASYNC, HZ/10);
3693 * Memory allocation/reclaim might be called from a WQ
3694 * context and the current implementation of the WQ
3695 * concurrency control doesn't recognize that
3696 * a particular WQ is congested if the worker thread is
3697 * looping without ever sleeping. Therefore we have to
3698 * do a short sleep here rather than calling
3701 if (current->flags & PF_WQ_WORKER)
3702 schedule_timeout_uninterruptible(1);
3714 check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
3717 * It's possible that cpuset's mems_allowed and the nodemask from
3718 * mempolicy don't intersect. This should be normally dealt with by
3719 * policy_nodemask(), but it's possible to race with cpuset update in
3720 * such a way the check therein was true, and then it became false
3721 * before we got our cpuset_mems_cookie here.
3722 * This assumes that for all allocations, ac->nodemask can come only
3723 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
3724 * when it does not intersect with the cpuset restrictions) or the
3725 * caller can deal with a violated nodemask.
3727 if (cpusets_enabled() && ac->nodemask &&
3728 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
3729 ac->nodemask = NULL;
3734 * When updating a task's mems_allowed or mempolicy nodemask, it is
3735 * possible to race with parallel threads in such a way that our
3736 * allocation can fail while the mask is being updated. If we are about
3737 * to fail, check if the cpuset changed during allocation and if so,
3740 if (read_mems_allowed_retry(cpuset_mems_cookie))
3746 static inline struct page *
3747 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3748 struct alloc_context *ac)
3750 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3751 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
3752 struct page *page = NULL;
3753 unsigned int alloc_flags;
3754 unsigned long did_some_progress;
3755 enum compact_priority compact_priority;
3756 enum compact_result compact_result;
3757 int compaction_retries;
3758 int no_progress_loops;
3759 unsigned long alloc_start = jiffies;
3760 unsigned int stall_timeout = 10 * HZ;
3761 unsigned int cpuset_mems_cookie;
3764 * In the slowpath, we sanity check order to avoid ever trying to
3765 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3766 * be using allocators in order of preference for an area that is
3769 if (order >= MAX_ORDER) {
3770 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3775 * We also sanity check to catch abuse of atomic reserves being used by
3776 * callers that are not in atomic context.
3778 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3779 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3780 gfp_mask &= ~__GFP_ATOMIC;
3783 compaction_retries = 0;
3784 no_progress_loops = 0;
3785 compact_priority = DEF_COMPACT_PRIORITY;
3786 cpuset_mems_cookie = read_mems_allowed_begin();
3789 * The fast path uses conservative alloc_flags to succeed only until
3790 * kswapd needs to be woken up, and to avoid the cost of setting up
3791 * alloc_flags precisely. So we do that now.
3793 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3796 * We need to recalculate the starting point for the zonelist iterator
3797 * because we might have used different nodemask in the fast path, or
3798 * there was a cpuset modification and we are retrying - otherwise we
3799 * could end up iterating over non-eligible zones endlessly.
3801 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3802 ac->high_zoneidx, ac->nodemask);
3803 if (!ac->preferred_zoneref->zone)
3806 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3807 wake_all_kswapds(order, ac);
3810 * The adjusted alloc_flags might result in immediate success, so try
3813 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3818 * For costly allocations, try direct compaction first, as it's likely
3819 * that we have enough base pages and don't need to reclaim. For non-
3820 * movable high-order allocations, do that as well, as compaction will
3821 * try prevent permanent fragmentation by migrating from blocks of the
3823 * Don't try this for allocations that are allowed to ignore
3824 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
3826 if (can_direct_reclaim &&
3828 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
3829 && !gfp_pfmemalloc_allowed(gfp_mask)) {
3830 page = __alloc_pages_direct_compact(gfp_mask, order,
3832 INIT_COMPACT_PRIORITY,
3838 * Checks for costly allocations with __GFP_NORETRY, which
3839 * includes THP page fault allocations
3841 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
3843 * If compaction is deferred for high-order allocations,
3844 * it is because sync compaction recently failed. If
3845 * this is the case and the caller requested a THP
3846 * allocation, we do not want to heavily disrupt the
3847 * system, so we fail the allocation instead of entering
3850 if (compact_result == COMPACT_DEFERRED)
3854 * Looks like reclaim/compaction is worth trying, but
3855 * sync compaction could be very expensive, so keep
3856 * using async compaction.
3858 compact_priority = INIT_COMPACT_PRIORITY;
3863 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3864 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3865 wake_all_kswapds(order, ac);
3867 if (gfp_pfmemalloc_allowed(gfp_mask))
3868 alloc_flags = ALLOC_NO_WATERMARKS;
3871 * Reset the zonelist iterators if memory policies can be ignored.
3872 * These allocations are high priority and system rather than user
3875 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3876 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3877 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3878 ac->high_zoneidx, ac->nodemask);
3881 /* Attempt with potentially adjusted zonelist and alloc_flags */
3882 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3886 /* Caller is not willing to reclaim, we can't balance anything */
3887 if (!can_direct_reclaim)
3890 /* Make sure we know about allocations which stall for too long */
3891 if (time_after(jiffies, alloc_start + stall_timeout)) {
3892 warn_alloc(gfp_mask & ~__GFP_NOWARN, ac->nodemask,
3893 "page allocation stalls for %ums, order:%u",
3894 jiffies_to_msecs(jiffies-alloc_start), order);
3895 stall_timeout += 10 * HZ;
3898 /* Avoid recursion of direct reclaim */
3899 if (current->flags & PF_MEMALLOC)
3902 /* Try direct reclaim and then allocating */
3903 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3904 &did_some_progress);
3908 /* Try direct compaction and then allocating */
3909 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3910 compact_priority, &compact_result);
3914 /* Do not loop if specifically requested */
3915 if (gfp_mask & __GFP_NORETRY)
3919 * Do not retry costly high order allocations unless they are
3922 if (costly_order && !(gfp_mask & __GFP_REPEAT))
3925 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3926 did_some_progress > 0, &no_progress_loops))
3930 * It doesn't make any sense to retry for the compaction if the order-0
3931 * reclaim is not able to make any progress because the current
3932 * implementation of the compaction depends on the sufficient amount
3933 * of free memory (see __compaction_suitable)
3935 if (did_some_progress > 0 &&
3936 should_compact_retry(ac, order, alloc_flags,
3937 compact_result, &compact_priority,
3938 &compaction_retries))
3942 /* Deal with possible cpuset update races before we start OOM killing */
3943 if (check_retry_cpuset(cpuset_mems_cookie, ac))
3946 /* Reclaim has failed us, start killing things */
3947 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3951 /* Avoid allocations with no watermarks from looping endlessly */
3952 if (test_thread_flag(TIF_MEMDIE) &&
3953 (alloc_flags == ALLOC_NO_WATERMARKS ||
3954 (gfp_mask & __GFP_NOMEMALLOC)))
3957 /* Retry as long as the OOM killer is making progress */
3958 if (did_some_progress) {
3959 no_progress_loops = 0;
3964 /* Deal with possible cpuset update races before we fail */
3965 if (check_retry_cpuset(cpuset_mems_cookie, ac))
3969 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
3972 if (gfp_mask & __GFP_NOFAIL) {
3974 * All existing users of the __GFP_NOFAIL are blockable, so warn
3975 * of any new users that actually require GFP_NOWAIT
3977 if (WARN_ON_ONCE(!can_direct_reclaim))
3981 * PF_MEMALLOC request from this context is rather bizarre
3982 * because we cannot reclaim anything and only can loop waiting
3983 * for somebody to do a work for us
3985 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
3988 * non failing costly orders are a hard requirement which we
3989 * are not prepared for much so let's warn about these users
3990 * so that we can identify them and convert them to something
3993 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
3996 * Help non-failing allocations by giving them access to memory
3997 * reserves but do not use ALLOC_NO_WATERMARKS because this
3998 * could deplete whole memory reserves which would just make
3999 * the situation worse
4001 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4009 warn_alloc(gfp_mask, ac->nodemask,
4010 "page allocation failure: order:%u", order);
4015 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4016 int preferred_nid, nodemask_t *nodemask,
4017 struct alloc_context *ac, gfp_t *alloc_mask,
4018 unsigned int *alloc_flags)
4020 ac->high_zoneidx = gfp_zone(gfp_mask);
4021 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4022 ac->nodemask = nodemask;
4023 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4025 if (cpusets_enabled()) {
4026 *alloc_mask |= __GFP_HARDWALL;
4028 ac->nodemask = &cpuset_current_mems_allowed;
4030 *alloc_flags |= ALLOC_CPUSET;
4033 lockdep_trace_alloc(gfp_mask);
4035 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4037 if (should_fail_alloc_page(gfp_mask, order))
4040 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4041 *alloc_flags |= ALLOC_CMA;
4046 /* Determine whether to spread dirty pages and what the first usable zone */
4047 static inline void finalise_ac(gfp_t gfp_mask,
4048 unsigned int order, struct alloc_context *ac)
4050 /* Dirty zone balancing only done in the fast path */
4051 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4054 * The preferred zone is used for statistics but crucially it is
4055 * also used as the starting point for the zonelist iterator. It
4056 * may get reset for allocations that ignore memory policies.
4058 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4059 ac->high_zoneidx, ac->nodemask);
4063 * This is the 'heart' of the zoned buddy allocator.
4066 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4067 nodemask_t *nodemask)
4070 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4071 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
4072 struct alloc_context ac = { };
4074 gfp_mask &= gfp_allowed_mask;
4075 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4078 finalise_ac(gfp_mask, order, &ac);
4080 /* First allocation attempt */
4081 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4086 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4087 * resp. GFP_NOIO which has to be inherited for all allocation requests
4088 * from a particular context which has been marked by
4089 * memalloc_no{fs,io}_{save,restore}.
4091 alloc_mask = current_gfp_context(gfp_mask);
4092 ac.spread_dirty_pages = false;
4095 * Restore the original nodemask if it was potentially replaced with
4096 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4098 if (unlikely(ac.nodemask != nodemask))
4099 ac.nodemask = nodemask;
4101 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4104 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4105 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4106 __free_pages(page, order);
4110 if (kmemcheck_enabled && page)
4111 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
4113 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4117 EXPORT_SYMBOL(__alloc_pages_nodemask);
4120 * Common helper functions.
4122 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4127 * __get_free_pages() returns a 32-bit address, which cannot represent
4130 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4132 page = alloc_pages(gfp_mask, order);
4135 return (unsigned long) page_address(page);
4137 EXPORT_SYMBOL(__get_free_pages);
4139 unsigned long get_zeroed_page(gfp_t gfp_mask)
4141 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4143 EXPORT_SYMBOL(get_zeroed_page);
4145 void __free_pages(struct page *page, unsigned int order)
4147 if (put_page_testzero(page)) {
4149 free_hot_cold_page(page, false);
4151 __free_pages_ok(page, order);
4155 EXPORT_SYMBOL(__free_pages);
4157 void free_pages(unsigned long addr, unsigned int order)
4160 VM_BUG_ON(!virt_addr_valid((void *)addr));
4161 __free_pages(virt_to_page((void *)addr), order);
4165 EXPORT_SYMBOL(free_pages);
4169 * An arbitrary-length arbitrary-offset area of memory which resides
4170 * within a 0 or higher order page. Multiple fragments within that page
4171 * are individually refcounted, in the page's reference counter.
4173 * The page_frag functions below provide a simple allocation framework for
4174 * page fragments. This is used by the network stack and network device
4175 * drivers to provide a backing region of memory for use as either an
4176 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4178 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4181 struct page *page = NULL;
4182 gfp_t gfp = gfp_mask;
4184 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4185 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4187 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4188 PAGE_FRAG_CACHE_MAX_ORDER);
4189 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4191 if (unlikely(!page))
4192 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4194 nc->va = page ? page_address(page) : NULL;
4199 void __page_frag_cache_drain(struct page *page, unsigned int count)
4201 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4203 if (page_ref_sub_and_test(page, count)) {
4204 unsigned int order = compound_order(page);
4207 free_hot_cold_page(page, false);
4209 __free_pages_ok(page, order);
4212 EXPORT_SYMBOL(__page_frag_cache_drain);
4214 void *page_frag_alloc(struct page_frag_cache *nc,
4215 unsigned int fragsz, gfp_t gfp_mask)
4217 unsigned int size = PAGE_SIZE;
4221 if (unlikely(!nc->va)) {
4223 page = __page_frag_cache_refill(nc, gfp_mask);
4227 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4228 /* if size can vary use size else just use PAGE_SIZE */
4231 /* Even if we own the page, we do not use atomic_set().
4232 * This would break get_page_unless_zero() users.
4234 page_ref_add(page, size - 1);
4236 /* reset page count bias and offset to start of new frag */
4237 nc->pfmemalloc = page_is_pfmemalloc(page);
4238 nc->pagecnt_bias = size;
4242 offset = nc->offset - fragsz;
4243 if (unlikely(offset < 0)) {
4244 page = virt_to_page(nc->va);
4246 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4249 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4250 /* if size can vary use size else just use PAGE_SIZE */
4253 /* OK, page count is 0, we can safely set it */
4254 set_page_count(page, size);
4256 /* reset page count bias and offset to start of new frag */
4257 nc->pagecnt_bias = size;
4258 offset = size - fragsz;
4262 nc->offset = offset;
4264 return nc->va + offset;
4266 EXPORT_SYMBOL(page_frag_alloc);
4269 * Frees a page fragment allocated out of either a compound or order 0 page.
4271 void page_frag_free(void *addr)
4273 struct page *page = virt_to_head_page(addr);
4275 if (unlikely(put_page_testzero(page)))
4276 __free_pages_ok(page, compound_order(page));
4278 EXPORT_SYMBOL(page_frag_free);
4280 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4284 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4285 unsigned long used = addr + PAGE_ALIGN(size);
4287 split_page(virt_to_page((void *)addr), order);
4288 while (used < alloc_end) {
4293 return (void *)addr;
4297 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4298 * @size: the number of bytes to allocate
4299 * @gfp_mask: GFP flags for the allocation
4301 * This function is similar to alloc_pages(), except that it allocates the
4302 * minimum number of pages to satisfy the request. alloc_pages() can only
4303 * allocate memory in power-of-two pages.
4305 * This function is also limited by MAX_ORDER.
4307 * Memory allocated by this function must be released by free_pages_exact().
4309 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4311 unsigned int order = get_order(size);
4314 addr = __get_free_pages(gfp_mask, order);
4315 return make_alloc_exact(addr, order, size);
4317 EXPORT_SYMBOL(alloc_pages_exact);
4320 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4322 * @nid: the preferred node ID where memory should be allocated
4323 * @size: the number of bytes to allocate
4324 * @gfp_mask: GFP flags for the allocation
4326 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4329 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4331 unsigned int order = get_order(size);
4332 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4335 return make_alloc_exact((unsigned long)page_address(p), order, size);
4339 * free_pages_exact - release memory allocated via alloc_pages_exact()
4340 * @virt: the value returned by alloc_pages_exact.
4341 * @size: size of allocation, same value as passed to alloc_pages_exact().
4343 * Release the memory allocated by a previous call to alloc_pages_exact.
4345 void free_pages_exact(void *virt, size_t size)
4347 unsigned long addr = (unsigned long)virt;
4348 unsigned long end = addr + PAGE_ALIGN(size);
4350 while (addr < end) {
4355 EXPORT_SYMBOL(free_pages_exact);
4358 * nr_free_zone_pages - count number of pages beyond high watermark
4359 * @offset: The zone index of the highest zone
4361 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4362 * high watermark within all zones at or below a given zone index. For each
4363 * zone, the number of pages is calculated as:
4365 * nr_free_zone_pages = managed_pages - high_pages
4367 static unsigned long nr_free_zone_pages(int offset)
4372 /* Just pick one node, since fallback list is circular */
4373 unsigned long sum = 0;
4375 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4377 for_each_zone_zonelist(zone, z, zonelist, offset) {
4378 unsigned long size = zone->managed_pages;
4379 unsigned long high = high_wmark_pages(zone);
4388 * nr_free_buffer_pages - count number of pages beyond high watermark
4390 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4391 * watermark within ZONE_DMA and ZONE_NORMAL.
4393 unsigned long nr_free_buffer_pages(void)
4395 return nr_free_zone_pages(gfp_zone(GFP_USER));
4397 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4400 * nr_free_pagecache_pages - count number of pages beyond high watermark
4402 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4403 * high watermark within all zones.
4405 unsigned long nr_free_pagecache_pages(void)
4407 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4410 static inline void show_node(struct zone *zone)
4412 if (IS_ENABLED(CONFIG_NUMA))
4413 printk("Node %d ", zone_to_nid(zone));
4416 long si_mem_available(void)
4419 unsigned long pagecache;
4420 unsigned long wmark_low = 0;
4421 unsigned long pages[NR_LRU_LISTS];
4425 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4426 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4429 wmark_low += zone->watermark[WMARK_LOW];
4432 * Estimate the amount of memory available for userspace allocations,
4433 * without causing swapping.
4435 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4438 * Not all the page cache can be freed, otherwise the system will
4439 * start swapping. Assume at least half of the page cache, or the
4440 * low watermark worth of cache, needs to stay.
4442 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4443 pagecache -= min(pagecache / 2, wmark_low);
4444 available += pagecache;
4447 * Part of the reclaimable slab consists of items that are in use,
4448 * and cannot be freed. Cap this estimate at the low watermark.
4450 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4451 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4457 EXPORT_SYMBOL_GPL(si_mem_available);
4459 void si_meminfo(struct sysinfo *val)
4461 val->totalram = totalram_pages;
4462 val->sharedram = global_node_page_state(NR_SHMEM);
4463 val->freeram = global_page_state(NR_FREE_PAGES);
4464 val->bufferram = nr_blockdev_pages();
4465 val->totalhigh = totalhigh_pages;
4466 val->freehigh = nr_free_highpages();
4467 val->mem_unit = PAGE_SIZE;
4470 EXPORT_SYMBOL(si_meminfo);
4473 void si_meminfo_node(struct sysinfo *val, int nid)
4475 int zone_type; /* needs to be signed */
4476 unsigned long managed_pages = 0;
4477 unsigned long managed_highpages = 0;
4478 unsigned long free_highpages = 0;
4479 pg_data_t *pgdat = NODE_DATA(nid);
4481 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4482 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4483 val->totalram = managed_pages;
4484 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4485 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4486 #ifdef CONFIG_HIGHMEM
4487 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4488 struct zone *zone = &pgdat->node_zones[zone_type];
4490 if (is_highmem(zone)) {
4491 managed_highpages += zone->managed_pages;
4492 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4495 val->totalhigh = managed_highpages;
4496 val->freehigh = free_highpages;
4498 val->totalhigh = managed_highpages;
4499 val->freehigh = free_highpages;
4501 val->mem_unit = PAGE_SIZE;
4506 * Determine whether the node should be displayed or not, depending on whether
4507 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4509 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4511 if (!(flags & SHOW_MEM_FILTER_NODES))
4515 * no node mask - aka implicit memory numa policy. Do not bother with
4516 * the synchronization - read_mems_allowed_begin - because we do not
4517 * have to be precise here.
4520 nodemask = &cpuset_current_mems_allowed;
4522 return !node_isset(nid, *nodemask);
4525 #define K(x) ((x) << (PAGE_SHIFT-10))
4527 static void show_migration_types(unsigned char type)
4529 static const char types[MIGRATE_TYPES] = {
4530 [MIGRATE_UNMOVABLE] = 'U',
4531 [MIGRATE_MOVABLE] = 'M',
4532 [MIGRATE_RECLAIMABLE] = 'E',
4533 [MIGRATE_HIGHATOMIC] = 'H',
4535 [MIGRATE_CMA] = 'C',
4537 #ifdef CONFIG_MEMORY_ISOLATION
4538 [MIGRATE_ISOLATE] = 'I',
4541 char tmp[MIGRATE_TYPES + 1];
4545 for (i = 0; i < MIGRATE_TYPES; i++) {
4546 if (type & (1 << i))
4551 printk(KERN_CONT "(%s) ", tmp);
4555 * Show free area list (used inside shift_scroll-lock stuff)
4556 * We also calculate the percentage fragmentation. We do this by counting the
4557 * memory on each free list with the exception of the first item on the list.
4560 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4563 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4565 unsigned long free_pcp = 0;
4570 for_each_populated_zone(zone) {
4571 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4574 for_each_online_cpu(cpu)
4575 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4578 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4579 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4580 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4581 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4582 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4583 " free:%lu free_pcp:%lu free_cma:%lu\n",
4584 global_node_page_state(NR_ACTIVE_ANON),
4585 global_node_page_state(NR_INACTIVE_ANON),
4586 global_node_page_state(NR_ISOLATED_ANON),
4587 global_node_page_state(NR_ACTIVE_FILE),
4588 global_node_page_state(NR_INACTIVE_FILE),
4589 global_node_page_state(NR_ISOLATED_FILE),
4590 global_node_page_state(NR_UNEVICTABLE),
4591 global_node_page_state(NR_FILE_DIRTY),
4592 global_node_page_state(NR_WRITEBACK),
4593 global_node_page_state(NR_UNSTABLE_NFS),
4594 global_page_state(NR_SLAB_RECLAIMABLE),
4595 global_page_state(NR_SLAB_UNRECLAIMABLE),
4596 global_node_page_state(NR_FILE_MAPPED),
4597 global_node_page_state(NR_SHMEM),
4598 global_page_state(NR_PAGETABLE),
4599 global_page_state(NR_BOUNCE),
4600 global_page_state(NR_FREE_PAGES),
4602 global_page_state(NR_FREE_CMA_PAGES));
4604 for_each_online_pgdat(pgdat) {
4605 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4609 " active_anon:%lukB"
4610 " inactive_anon:%lukB"
4611 " active_file:%lukB"
4612 " inactive_file:%lukB"
4613 " unevictable:%lukB"
4614 " isolated(anon):%lukB"
4615 " isolated(file):%lukB"
4620 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4622 " shmem_pmdmapped: %lukB"
4625 " writeback_tmp:%lukB"
4627 " all_unreclaimable? %s"
4630 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4631 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4632 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4633 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4634 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4635 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4636 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4637 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4638 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4639 K(node_page_state(pgdat, NR_WRITEBACK)),
4640 K(node_page_state(pgdat, NR_SHMEM)),
4641 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4642 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4643 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4645 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4647 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4648 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4649 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4653 for_each_populated_zone(zone) {
4656 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4660 for_each_online_cpu(cpu)
4661 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4670 " active_anon:%lukB"
4671 " inactive_anon:%lukB"
4672 " active_file:%lukB"
4673 " inactive_file:%lukB"
4674 " unevictable:%lukB"
4675 " writepending:%lukB"
4679 " kernel_stack:%lukB"
4687 K(zone_page_state(zone, NR_FREE_PAGES)),
4688 K(min_wmark_pages(zone)),
4689 K(low_wmark_pages(zone)),
4690 K(high_wmark_pages(zone)),
4691 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4692 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4693 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4694 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4695 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4696 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4697 K(zone->present_pages),
4698 K(zone->managed_pages),
4699 K(zone_page_state(zone, NR_MLOCK)),
4700 zone_page_state(zone, NR_KERNEL_STACK_KB),
4701 K(zone_page_state(zone, NR_PAGETABLE)),
4702 K(zone_page_state(zone, NR_BOUNCE)),
4704 K(this_cpu_read(zone->pageset->pcp.count)),
4705 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4706 printk("lowmem_reserve[]:");
4707 for (i = 0; i < MAX_NR_ZONES; i++)
4708 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4709 printk(KERN_CONT "\n");
4712 for_each_populated_zone(zone) {
4714 unsigned long nr[MAX_ORDER], flags, total = 0;
4715 unsigned char types[MAX_ORDER];
4717 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4720 printk(KERN_CONT "%s: ", zone->name);
4722 spin_lock_irqsave(&zone->lock, flags);
4723 for (order = 0; order < MAX_ORDER; order++) {
4724 struct free_area *area = &zone->free_area[order];
4727 nr[order] = area->nr_free;
4728 total += nr[order] << order;
4731 for (type = 0; type < MIGRATE_TYPES; type++) {
4732 if (!list_empty(&area->free_list[type]))
4733 types[order] |= 1 << type;
4736 spin_unlock_irqrestore(&zone->lock, flags);
4737 for (order = 0; order < MAX_ORDER; order++) {
4738 printk(KERN_CONT "%lu*%lukB ",
4739 nr[order], K(1UL) << order);
4741 show_migration_types(types[order]);
4743 printk(KERN_CONT "= %lukB\n", K(total));
4746 hugetlb_show_meminfo();
4748 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4750 show_swap_cache_info();
4753 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4755 zoneref->zone = zone;
4756 zoneref->zone_idx = zone_idx(zone);
4760 * Builds allocation fallback zone lists.
4762 * Add all populated zones of a node to the zonelist.
4764 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4768 enum zone_type zone_type = MAX_NR_ZONES;
4772 zone = pgdat->node_zones + zone_type;
4773 if (managed_zone(zone)) {
4774 zoneref_set_zone(zone,
4775 &zonelist->_zonerefs[nr_zones++]);
4776 check_highest_zone(zone_type);
4778 } while (zone_type);
4786 * 0 = automatic detection of better ordering.
4787 * 1 = order by ([node] distance, -zonetype)
4788 * 2 = order by (-zonetype, [node] distance)
4790 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4791 * the same zonelist. So only NUMA can configure this param.
4793 #define ZONELIST_ORDER_DEFAULT 0
4794 #define ZONELIST_ORDER_NODE 1
4795 #define ZONELIST_ORDER_ZONE 2
4797 /* zonelist order in the kernel.
4798 * set_zonelist_order() will set this to NODE or ZONE.
4800 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4801 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4805 /* The value user specified ....changed by config */
4806 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4807 /* string for sysctl */
4808 #define NUMA_ZONELIST_ORDER_LEN 16
4809 char numa_zonelist_order[16] = "default";
4812 * interface for configure zonelist ordering.
4813 * command line option "numa_zonelist_order"
4814 * = "[dD]efault - default, automatic configuration.
4815 * = "[nN]ode - order by node locality, then by zone within node
4816 * = "[zZ]one - order by zone, then by locality within zone
4819 static int __parse_numa_zonelist_order(char *s)
4821 if (*s == 'd' || *s == 'D') {
4822 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4823 } else if (*s == 'n' || *s == 'N') {
4824 user_zonelist_order = ZONELIST_ORDER_NODE;
4825 } else if (*s == 'z' || *s == 'Z') {
4826 user_zonelist_order = ZONELIST_ORDER_ZONE;
4828 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4834 static __init int setup_numa_zonelist_order(char *s)
4841 ret = __parse_numa_zonelist_order(s);
4843 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4847 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4850 * sysctl handler for numa_zonelist_order
4852 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4853 void __user *buffer, size_t *length,
4856 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4858 static DEFINE_MUTEX(zl_order_mutex);
4860 mutex_lock(&zl_order_mutex);
4862 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4866 strcpy(saved_string, (char *)table->data);
4868 ret = proc_dostring(table, write, buffer, length, ppos);
4872 int oldval = user_zonelist_order;
4874 ret = __parse_numa_zonelist_order((char *)table->data);
4877 * bogus value. restore saved string
4879 strncpy((char *)table->data, saved_string,
4880 NUMA_ZONELIST_ORDER_LEN);
4881 user_zonelist_order = oldval;
4882 } else if (oldval != user_zonelist_order) {
4883 mutex_lock(&zonelists_mutex);
4884 build_all_zonelists(NULL, NULL);
4885 mutex_unlock(&zonelists_mutex);
4889 mutex_unlock(&zl_order_mutex);
4894 #define MAX_NODE_LOAD (nr_online_nodes)
4895 static int node_load[MAX_NUMNODES];
4898 * find_next_best_node - find the next node that should appear in a given node's fallback list
4899 * @node: node whose fallback list we're appending
4900 * @used_node_mask: nodemask_t of already used nodes
4902 * We use a number of factors to determine which is the next node that should
4903 * appear on a given node's fallback list. The node should not have appeared
4904 * already in @node's fallback list, and it should be the next closest node
4905 * according to the distance array (which contains arbitrary distance values
4906 * from each node to each node in the system), and should also prefer nodes
4907 * with no CPUs, since presumably they'll have very little allocation pressure
4908 * on them otherwise.
4909 * It returns -1 if no node is found.
4911 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4914 int min_val = INT_MAX;
4915 int best_node = NUMA_NO_NODE;
4916 const struct cpumask *tmp = cpumask_of_node(0);
4918 /* Use the local node if we haven't already */
4919 if (!node_isset(node, *used_node_mask)) {
4920 node_set(node, *used_node_mask);
4924 for_each_node_state(n, N_MEMORY) {
4926 /* Don't want a node to appear more than once */
4927 if (node_isset(n, *used_node_mask))
4930 /* Use the distance array to find the distance */
4931 val = node_distance(node, n);
4933 /* Penalize nodes under us ("prefer the next node") */
4936 /* Give preference to headless and unused nodes */
4937 tmp = cpumask_of_node(n);
4938 if (!cpumask_empty(tmp))
4939 val += PENALTY_FOR_NODE_WITH_CPUS;
4941 /* Slight preference for less loaded node */
4942 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4943 val += node_load[n];
4945 if (val < min_val) {
4952 node_set(best_node, *used_node_mask);
4959 * Build zonelists ordered by node and zones within node.
4960 * This results in maximum locality--normal zone overflows into local
4961 * DMA zone, if any--but risks exhausting DMA zone.
4963 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4966 struct zonelist *zonelist;
4968 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4969 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4971 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4972 zonelist->_zonerefs[j].zone = NULL;
4973 zonelist->_zonerefs[j].zone_idx = 0;
4977 * Build gfp_thisnode zonelists
4979 static void build_thisnode_zonelists(pg_data_t *pgdat)
4982 struct zonelist *zonelist;
4984 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
4985 j = build_zonelists_node(pgdat, zonelist, 0);
4986 zonelist->_zonerefs[j].zone = NULL;
4987 zonelist->_zonerefs[j].zone_idx = 0;
4991 * Build zonelists ordered by zone and nodes within zones.
4992 * This results in conserving DMA zone[s] until all Normal memory is
4993 * exhausted, but results in overflowing to remote node while memory
4994 * may still exist in local DMA zone.
4996 static int node_order[MAX_NUMNODES];
4998 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
5001 int zone_type; /* needs to be signed */
5003 struct zonelist *zonelist;
5005 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5007 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
5008 for (j = 0; j < nr_nodes; j++) {
5009 node = node_order[j];
5010 z = &NODE_DATA(node)->node_zones[zone_type];
5011 if (managed_zone(z)) {
5013 &zonelist->_zonerefs[pos++]);
5014 check_highest_zone(zone_type);
5018 zonelist->_zonerefs[pos].zone = NULL;
5019 zonelist->_zonerefs[pos].zone_idx = 0;
5022 #if defined(CONFIG_64BIT)
5024 * Devices that require DMA32/DMA are relatively rare and do not justify a
5025 * penalty to every machine in case the specialised case applies. Default
5026 * to Node-ordering on 64-bit NUMA machines
5028 static int default_zonelist_order(void)
5030 return ZONELIST_ORDER_NODE;
5034 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
5035 * by the kernel. If processes running on node 0 deplete the low memory zone
5036 * then reclaim will occur more frequency increasing stalls and potentially
5037 * be easier to OOM if a large percentage of the zone is under writeback or
5038 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
5039 * Hence, default to zone ordering on 32-bit.
5041 static int default_zonelist_order(void)
5043 return ZONELIST_ORDER_ZONE;
5045 #endif /* CONFIG_64BIT */
5047 static void set_zonelist_order(void)
5049 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
5050 current_zonelist_order = default_zonelist_order();
5052 current_zonelist_order = user_zonelist_order;
5055 static void build_zonelists(pg_data_t *pgdat)
5058 nodemask_t used_mask;
5059 int local_node, prev_node;
5060 struct zonelist *zonelist;
5061 unsigned int order = current_zonelist_order;
5063 /* initialize zonelists */
5064 for (i = 0; i < MAX_ZONELISTS; i++) {
5065 zonelist = pgdat->node_zonelists + i;
5066 zonelist->_zonerefs[0].zone = NULL;
5067 zonelist->_zonerefs[0].zone_idx = 0;
5070 /* NUMA-aware ordering of nodes */
5071 local_node = pgdat->node_id;
5072 load = nr_online_nodes;
5073 prev_node = local_node;
5074 nodes_clear(used_mask);
5076 memset(node_order, 0, sizeof(node_order));
5079 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5081 * We don't want to pressure a particular node.
5082 * So adding penalty to the first node in same
5083 * distance group to make it round-robin.
5085 if (node_distance(local_node, node) !=
5086 node_distance(local_node, prev_node))
5087 node_load[node] = load;
5091 if (order == ZONELIST_ORDER_NODE)
5092 build_zonelists_in_node_order(pgdat, node);
5094 node_order[i++] = node; /* remember order */
5097 if (order == ZONELIST_ORDER_ZONE) {
5098 /* calculate node order -- i.e., DMA last! */
5099 build_zonelists_in_zone_order(pgdat, i);
5102 build_thisnode_zonelists(pgdat);
5105 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5107 * Return node id of node used for "local" allocations.
5108 * I.e., first node id of first zone in arg node's generic zonelist.
5109 * Used for initializing percpu 'numa_mem', which is used primarily
5110 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5112 int local_memory_node(int node)
5116 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5117 gfp_zone(GFP_KERNEL),
5119 return z->zone->node;
5123 static void setup_min_unmapped_ratio(void);
5124 static void setup_min_slab_ratio(void);
5125 #else /* CONFIG_NUMA */
5127 static void set_zonelist_order(void)
5129 current_zonelist_order = ZONELIST_ORDER_ZONE;
5132 static void build_zonelists(pg_data_t *pgdat)
5134 int node, local_node;
5136 struct zonelist *zonelist;
5138 local_node = pgdat->node_id;
5140 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5141 j = build_zonelists_node(pgdat, zonelist, 0);
5144 * Now we build the zonelist so that it contains the zones
5145 * of all the other nodes.
5146 * We don't want to pressure a particular node, so when
5147 * building the zones for node N, we make sure that the
5148 * zones coming right after the local ones are those from
5149 * node N+1 (modulo N)
5151 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5152 if (!node_online(node))
5154 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5156 for (node = 0; node < local_node; node++) {
5157 if (!node_online(node))
5159 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5162 zonelist->_zonerefs[j].zone = NULL;
5163 zonelist->_zonerefs[j].zone_idx = 0;
5166 #endif /* CONFIG_NUMA */
5169 * Boot pageset table. One per cpu which is going to be used for all
5170 * zones and all nodes. The parameters will be set in such a way
5171 * that an item put on a list will immediately be handed over to
5172 * the buddy list. This is safe since pageset manipulation is done
5173 * with interrupts disabled.
5175 * The boot_pagesets must be kept even after bootup is complete for
5176 * unused processors and/or zones. They do play a role for bootstrapping
5177 * hotplugged processors.
5179 * zoneinfo_show() and maybe other functions do
5180 * not check if the processor is online before following the pageset pointer.
5181 * Other parts of the kernel may not check if the zone is available.
5183 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5184 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5185 static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5186 static void setup_zone_pageset(struct zone *zone);
5189 * Global mutex to protect against size modification of zonelists
5190 * as well as to serialize pageset setup for the new populated zone.
5192 DEFINE_MUTEX(zonelists_mutex);
5194 /* return values int ....just for stop_machine() */
5195 static int __build_all_zonelists(void *data)
5199 pg_data_t *self = data;
5202 memset(node_load, 0, sizeof(node_load));
5205 if (self && !node_online(self->node_id)) {
5206 build_zonelists(self);
5209 for_each_online_node(nid) {
5210 pg_data_t *pgdat = NODE_DATA(nid);
5212 build_zonelists(pgdat);
5216 * Initialize the boot_pagesets that are going to be used
5217 * for bootstrapping processors. The real pagesets for
5218 * each zone will be allocated later when the per cpu
5219 * allocator is available.
5221 * boot_pagesets are used also for bootstrapping offline
5222 * cpus if the system is already booted because the pagesets
5223 * are needed to initialize allocators on a specific cpu too.
5224 * F.e. the percpu allocator needs the page allocator which
5225 * needs the percpu allocator in order to allocate its pagesets
5226 * (a chicken-egg dilemma).
5228 for_each_possible_cpu(cpu) {
5229 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5231 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5233 * We now know the "local memory node" for each node--
5234 * i.e., the node of the first zone in the generic zonelist.
5235 * Set up numa_mem percpu variable for on-line cpus. During
5236 * boot, only the boot cpu should be on-line; we'll init the
5237 * secondary cpus' numa_mem as they come on-line. During
5238 * node/memory hotplug, we'll fixup all on-line cpus.
5240 if (cpu_online(cpu))
5241 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5248 static noinline void __init
5249 build_all_zonelists_init(void)
5251 __build_all_zonelists(NULL);
5252 mminit_verify_zonelist();
5253 cpuset_init_current_mems_allowed();
5257 * Called with zonelists_mutex held always
5258 * unless system_state == SYSTEM_BOOTING.
5260 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5261 * [we're only called with non-NULL zone through __meminit paths] and
5262 * (2) call of __init annotated helper build_all_zonelists_init
5263 * [protected by SYSTEM_BOOTING].
5265 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5267 set_zonelist_order();
5269 if (system_state == SYSTEM_BOOTING) {
5270 build_all_zonelists_init();
5272 #ifdef CONFIG_MEMORY_HOTPLUG
5274 setup_zone_pageset(zone);
5276 /* we have to stop all cpus to guarantee there is no user
5278 stop_machine(__build_all_zonelists, pgdat, NULL);
5279 /* cpuset refresh routine should be here */
5281 vm_total_pages = nr_free_pagecache_pages();
5283 * Disable grouping by mobility if the number of pages in the
5284 * system is too low to allow the mechanism to work. It would be
5285 * more accurate, but expensive to check per-zone. This check is
5286 * made on memory-hotadd so a system can start with mobility
5287 * disabled and enable it later
5289 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5290 page_group_by_mobility_disabled = 1;
5292 page_group_by_mobility_disabled = 0;
5294 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5296 zonelist_order_name[current_zonelist_order],
5297 page_group_by_mobility_disabled ? "off" : "on",
5300 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5305 * Initially all pages are reserved - free ones are freed
5306 * up by free_all_bootmem() once the early boot process is
5307 * done. Non-atomic initialization, single-pass.
5309 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5310 unsigned long start_pfn, enum memmap_context context)
5312 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5313 unsigned long end_pfn = start_pfn + size;
5314 pg_data_t *pgdat = NODE_DATA(nid);
5316 unsigned long nr_initialised = 0;
5317 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5318 struct memblock_region *r = NULL, *tmp;
5321 if (highest_memmap_pfn < end_pfn - 1)
5322 highest_memmap_pfn = end_pfn - 1;
5325 * Honor reservation requested by the driver for this ZONE_DEVICE
5328 if (altmap && start_pfn == altmap->base_pfn)
5329 start_pfn += altmap->reserve;
5331 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5333 * There can be holes in boot-time mem_map[]s handed to this
5334 * function. They do not exist on hotplugged memory.
5336 if (context != MEMMAP_EARLY)
5339 if (!early_pfn_valid(pfn)) {
5340 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5342 * Skip to the pfn preceding the next valid one (or
5343 * end_pfn), such that we hit a valid pfn (or end_pfn)
5344 * on our next iteration of the loop.
5346 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5350 if (!early_pfn_in_nid(pfn, nid))
5352 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5355 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5357 * Check given memblock attribute by firmware which can affect
5358 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5359 * mirrored, it's an overlapped memmap init. skip it.
5361 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5362 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5363 for_each_memblock(memory, tmp)
5364 if (pfn < memblock_region_memory_end_pfn(tmp))
5368 if (pfn >= memblock_region_memory_base_pfn(r) &&
5369 memblock_is_mirror(r)) {
5370 /* already initialized as NORMAL */
5371 pfn = memblock_region_memory_end_pfn(r);
5379 * Mark the block movable so that blocks are reserved for
5380 * movable at startup. This will force kernel allocations
5381 * to reserve their blocks rather than leaking throughout
5382 * the address space during boot when many long-lived
5383 * kernel allocations are made.
5385 * bitmap is created for zone's valid pfn range. but memmap
5386 * can be created for invalid pages (for alignment)
5387 * check here not to call set_pageblock_migratetype() against
5390 if (!(pfn & (pageblock_nr_pages - 1))) {
5391 struct page *page = pfn_to_page(pfn);
5393 __init_single_page(page, pfn, zone, nid);
5394 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5396 __init_single_pfn(pfn, zone, nid);
5401 static void __meminit zone_init_free_lists(struct zone *zone)
5403 unsigned int order, t;
5404 for_each_migratetype_order(order, t) {
5405 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5406 zone->free_area[order].nr_free = 0;
5410 #ifndef __HAVE_ARCH_MEMMAP_INIT
5411 #define memmap_init(size, nid, zone, start_pfn) \
5412 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5415 static int zone_batchsize(struct zone *zone)
5421 * The per-cpu-pages pools are set to around 1000th of the
5422 * size of the zone. But no more than 1/2 of a meg.
5424 * OK, so we don't know how big the cache is. So guess.
5426 batch = zone->managed_pages / 1024;
5427 if (batch * PAGE_SIZE > 512 * 1024)
5428 batch = (512 * 1024) / PAGE_SIZE;
5429 batch /= 4; /* We effectively *= 4 below */
5434 * Clamp the batch to a 2^n - 1 value. Having a power
5435 * of 2 value was found to be more likely to have
5436 * suboptimal cache aliasing properties in some cases.
5438 * For example if 2 tasks are alternately allocating
5439 * batches of pages, one task can end up with a lot
5440 * of pages of one half of the possible page colors
5441 * and the other with pages of the other colors.
5443 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5448 /* The deferral and batching of frees should be suppressed under NOMMU
5451 * The problem is that NOMMU needs to be able to allocate large chunks
5452 * of contiguous memory as there's no hardware page translation to
5453 * assemble apparent contiguous memory from discontiguous pages.
5455 * Queueing large contiguous runs of pages for batching, however,
5456 * causes the pages to actually be freed in smaller chunks. As there
5457 * can be a significant delay between the individual batches being
5458 * recycled, this leads to the once large chunks of space being
5459 * fragmented and becoming unavailable for high-order allocations.
5466 * pcp->high and pcp->batch values are related and dependent on one another:
5467 * ->batch must never be higher then ->high.
5468 * The following function updates them in a safe manner without read side
5471 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5472 * those fields changing asynchronously (acording the the above rule).
5474 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5475 * outside of boot time (or some other assurance that no concurrent updaters
5478 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5479 unsigned long batch)
5481 /* start with a fail safe value for batch */
5485 /* Update high, then batch, in order */
5492 /* a companion to pageset_set_high() */
5493 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5495 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5498 static void pageset_init(struct per_cpu_pageset *p)
5500 struct per_cpu_pages *pcp;
5503 memset(p, 0, sizeof(*p));
5507 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5508 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5511 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5514 pageset_set_batch(p, batch);
5518 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5519 * to the value high for the pageset p.
5521 static void pageset_set_high(struct per_cpu_pageset *p,
5524 unsigned long batch = max(1UL, high / 4);
5525 if ((high / 4) > (PAGE_SHIFT * 8))
5526 batch = PAGE_SHIFT * 8;
5528 pageset_update(&p->pcp, high, batch);
5531 static void pageset_set_high_and_batch(struct zone *zone,
5532 struct per_cpu_pageset *pcp)
5534 if (percpu_pagelist_fraction)
5535 pageset_set_high(pcp,
5536 (zone->managed_pages /
5537 percpu_pagelist_fraction));
5539 pageset_set_batch(pcp, zone_batchsize(zone));
5542 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5544 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5547 pageset_set_high_and_batch(zone, pcp);
5550 static void __meminit setup_zone_pageset(struct zone *zone)
5553 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5554 for_each_possible_cpu(cpu)
5555 zone_pageset_init(zone, cpu);
5559 * Allocate per cpu pagesets and initialize them.
5560 * Before this call only boot pagesets were available.
5562 void __init setup_per_cpu_pageset(void)
5564 struct pglist_data *pgdat;
5567 for_each_populated_zone(zone)
5568 setup_zone_pageset(zone);
5570 for_each_online_pgdat(pgdat)
5571 pgdat->per_cpu_nodestats =
5572 alloc_percpu(struct per_cpu_nodestat);
5575 static __meminit void zone_pcp_init(struct zone *zone)
5578 * per cpu subsystem is not up at this point. The following code
5579 * relies on the ability of the linker to provide the
5580 * offset of a (static) per cpu variable into the per cpu area.
5582 zone->pageset = &boot_pageset;
5584 if (populated_zone(zone))
5585 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5586 zone->name, zone->present_pages,
5587 zone_batchsize(zone));
5590 void __meminit init_currently_empty_zone(struct zone *zone,
5591 unsigned long zone_start_pfn,
5594 struct pglist_data *pgdat = zone->zone_pgdat;
5596 pgdat->nr_zones = zone_idx(zone) + 1;
5598 zone->zone_start_pfn = zone_start_pfn;
5600 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5601 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5603 (unsigned long)zone_idx(zone),
5604 zone_start_pfn, (zone_start_pfn + size));
5606 zone_init_free_lists(zone);
5607 zone->initialized = 1;
5610 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5611 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5614 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5616 int __meminit __early_pfn_to_nid(unsigned long pfn,
5617 struct mminit_pfnnid_cache *state)
5619 unsigned long start_pfn, end_pfn;
5622 if (state->last_start <= pfn && pfn < state->last_end)
5623 return state->last_nid;
5625 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5627 state->last_start = start_pfn;
5628 state->last_end = end_pfn;
5629 state->last_nid = nid;
5634 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5637 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5638 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5639 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5641 * If an architecture guarantees that all ranges registered contain no holes
5642 * and may be freed, this this function may be used instead of calling
5643 * memblock_free_early_nid() manually.
5645 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5647 unsigned long start_pfn, end_pfn;
5650 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5651 start_pfn = min(start_pfn, max_low_pfn);
5652 end_pfn = min(end_pfn, max_low_pfn);
5654 if (start_pfn < end_pfn)
5655 memblock_free_early_nid(PFN_PHYS(start_pfn),
5656 (end_pfn - start_pfn) << PAGE_SHIFT,
5662 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5663 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5665 * If an architecture guarantees that all ranges registered contain no holes and may
5666 * be freed, this function may be used instead of calling memory_present() manually.
5668 void __init sparse_memory_present_with_active_regions(int nid)
5670 unsigned long start_pfn, end_pfn;
5673 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5674 memory_present(this_nid, start_pfn, end_pfn);
5678 * get_pfn_range_for_nid - Return the start and end page frames for a node
5679 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5680 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5681 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5683 * It returns the start and end page frame of a node based on information
5684 * provided by memblock_set_node(). If called for a node
5685 * with no available memory, a warning is printed and the start and end
5688 void __meminit get_pfn_range_for_nid(unsigned int nid,
5689 unsigned long *start_pfn, unsigned long *end_pfn)
5691 unsigned long this_start_pfn, this_end_pfn;
5697 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5698 *start_pfn = min(*start_pfn, this_start_pfn);
5699 *end_pfn = max(*end_pfn, this_end_pfn);
5702 if (*start_pfn == -1UL)
5707 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5708 * assumption is made that zones within a node are ordered in monotonic
5709 * increasing memory addresses so that the "highest" populated zone is used
5711 static void __init find_usable_zone_for_movable(void)
5714 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5715 if (zone_index == ZONE_MOVABLE)
5718 if (arch_zone_highest_possible_pfn[zone_index] >
5719 arch_zone_lowest_possible_pfn[zone_index])
5723 VM_BUG_ON(zone_index == -1);
5724 movable_zone = zone_index;
5728 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5729 * because it is sized independent of architecture. Unlike the other zones,
5730 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5731 * in each node depending on the size of each node and how evenly kernelcore
5732 * is distributed. This helper function adjusts the zone ranges
5733 * provided by the architecture for a given node by using the end of the
5734 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5735 * zones within a node are in order of monotonic increases memory addresses
5737 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5738 unsigned long zone_type,
5739 unsigned long node_start_pfn,
5740 unsigned long node_end_pfn,
5741 unsigned long *zone_start_pfn,
5742 unsigned long *zone_end_pfn)
5744 /* Only adjust if ZONE_MOVABLE is on this node */
5745 if (zone_movable_pfn[nid]) {
5746 /* Size ZONE_MOVABLE */
5747 if (zone_type == ZONE_MOVABLE) {
5748 *zone_start_pfn = zone_movable_pfn[nid];
5749 *zone_end_pfn = min(node_end_pfn,
5750 arch_zone_highest_possible_pfn[movable_zone]);
5752 /* Adjust for ZONE_MOVABLE starting within this range */
5753 } else if (!mirrored_kernelcore &&
5754 *zone_start_pfn < zone_movable_pfn[nid] &&
5755 *zone_end_pfn > zone_movable_pfn[nid]) {
5756 *zone_end_pfn = zone_movable_pfn[nid];
5758 /* Check if this whole range is within ZONE_MOVABLE */
5759 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5760 *zone_start_pfn = *zone_end_pfn;
5765 * Return the number of pages a zone spans in a node, including holes
5766 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5768 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5769 unsigned long zone_type,
5770 unsigned long node_start_pfn,
5771 unsigned long node_end_pfn,
5772 unsigned long *zone_start_pfn,
5773 unsigned long *zone_end_pfn,
5774 unsigned long *ignored)
5776 /* When hotadd a new node from cpu_up(), the node should be empty */
5777 if (!node_start_pfn && !node_end_pfn)
5780 /* Get the start and end of the zone */
5781 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5782 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5783 adjust_zone_range_for_zone_movable(nid, zone_type,
5784 node_start_pfn, node_end_pfn,
5785 zone_start_pfn, zone_end_pfn);
5787 /* Check that this node has pages within the zone's required range */
5788 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5791 /* Move the zone boundaries inside the node if necessary */
5792 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5793 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5795 /* Return the spanned pages */
5796 return *zone_end_pfn - *zone_start_pfn;
5800 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5801 * then all holes in the requested range will be accounted for.
5803 unsigned long __meminit __absent_pages_in_range(int nid,
5804 unsigned long range_start_pfn,
5805 unsigned long range_end_pfn)
5807 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5808 unsigned long start_pfn, end_pfn;
5811 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5812 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5813 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5814 nr_absent -= end_pfn - start_pfn;
5820 * absent_pages_in_range - Return number of page frames in holes within a range
5821 * @start_pfn: The start PFN to start searching for holes
5822 * @end_pfn: The end PFN to stop searching for holes
5824 * It returns the number of pages frames in memory holes within a range.
5826 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5827 unsigned long end_pfn)
5829 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5832 /* Return the number of page frames in holes in a zone on a node */
5833 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5834 unsigned long zone_type,
5835 unsigned long node_start_pfn,
5836 unsigned long node_end_pfn,
5837 unsigned long *ignored)
5839 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5840 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5841 unsigned long zone_start_pfn, zone_end_pfn;
5842 unsigned long nr_absent;
5844 /* When hotadd a new node from cpu_up(), the node should be empty */
5845 if (!node_start_pfn && !node_end_pfn)
5848 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5849 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5851 adjust_zone_range_for_zone_movable(nid, zone_type,
5852 node_start_pfn, node_end_pfn,
5853 &zone_start_pfn, &zone_end_pfn);
5854 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5857 * ZONE_MOVABLE handling.
5858 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5861 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5862 unsigned long start_pfn, end_pfn;
5863 struct memblock_region *r;
5865 for_each_memblock(memory, r) {
5866 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5867 zone_start_pfn, zone_end_pfn);
5868 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5869 zone_start_pfn, zone_end_pfn);
5871 if (zone_type == ZONE_MOVABLE &&
5872 memblock_is_mirror(r))
5873 nr_absent += end_pfn - start_pfn;
5875 if (zone_type == ZONE_NORMAL &&
5876 !memblock_is_mirror(r))
5877 nr_absent += end_pfn - start_pfn;
5884 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5885 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5886 unsigned long zone_type,
5887 unsigned long node_start_pfn,
5888 unsigned long node_end_pfn,
5889 unsigned long *zone_start_pfn,
5890 unsigned long *zone_end_pfn,
5891 unsigned long *zones_size)
5895 *zone_start_pfn = node_start_pfn;
5896 for (zone = 0; zone < zone_type; zone++)
5897 *zone_start_pfn += zones_size[zone];
5899 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5901 return zones_size[zone_type];
5904 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5905 unsigned long zone_type,
5906 unsigned long node_start_pfn,
5907 unsigned long node_end_pfn,
5908 unsigned long *zholes_size)
5913 return zholes_size[zone_type];
5916 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5918 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5919 unsigned long node_start_pfn,
5920 unsigned long node_end_pfn,
5921 unsigned long *zones_size,
5922 unsigned long *zholes_size)
5924 unsigned long realtotalpages = 0, totalpages = 0;
5927 for (i = 0; i < MAX_NR_ZONES; i++) {
5928 struct zone *zone = pgdat->node_zones + i;
5929 unsigned long zone_start_pfn, zone_end_pfn;
5930 unsigned long size, real_size;
5932 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5938 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5939 node_start_pfn, node_end_pfn,
5942 zone->zone_start_pfn = zone_start_pfn;
5944 zone->zone_start_pfn = 0;
5945 zone->spanned_pages = size;
5946 zone->present_pages = real_size;
5949 realtotalpages += real_size;
5952 pgdat->node_spanned_pages = totalpages;
5953 pgdat->node_present_pages = realtotalpages;
5954 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5958 #ifndef CONFIG_SPARSEMEM
5960 * Calculate the size of the zone->blockflags rounded to an unsigned long
5961 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5962 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5963 * round what is now in bits to nearest long in bits, then return it in
5966 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5968 unsigned long usemapsize;
5970 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5971 usemapsize = roundup(zonesize, pageblock_nr_pages);
5972 usemapsize = usemapsize >> pageblock_order;
5973 usemapsize *= NR_PAGEBLOCK_BITS;
5974 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5976 return usemapsize / 8;
5979 static void __init setup_usemap(struct pglist_data *pgdat,
5981 unsigned long zone_start_pfn,
5982 unsigned long zonesize)
5984 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5985 zone->pageblock_flags = NULL;
5987 zone->pageblock_flags =
5988 memblock_virt_alloc_node_nopanic(usemapsize,
5992 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5993 unsigned long zone_start_pfn, unsigned long zonesize) {}
5994 #endif /* CONFIG_SPARSEMEM */
5996 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5998 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5999 void __paginginit set_pageblock_order(void)
6003 /* Check that pageblock_nr_pages has not already been setup */
6004 if (pageblock_order)
6007 if (HPAGE_SHIFT > PAGE_SHIFT)
6008 order = HUGETLB_PAGE_ORDER;
6010 order = MAX_ORDER - 1;
6013 * Assume the largest contiguous order of interest is a huge page.
6014 * This value may be variable depending on boot parameters on IA64 and
6017 pageblock_order = order;
6019 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6022 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6023 * is unused as pageblock_order is set at compile-time. See
6024 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6027 void __paginginit set_pageblock_order(void)
6031 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6033 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6034 unsigned long present_pages)
6036 unsigned long pages = spanned_pages;
6039 * Provide a more accurate estimation if there are holes within
6040 * the zone and SPARSEMEM is in use. If there are holes within the
6041 * zone, each populated memory region may cost us one or two extra
6042 * memmap pages due to alignment because memmap pages for each
6043 * populated regions may not be naturally aligned on page boundary.
6044 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6046 if (spanned_pages > present_pages + (present_pages >> 4) &&
6047 IS_ENABLED(CONFIG_SPARSEMEM))
6048 pages = present_pages;
6050 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6054 * Set up the zone data structures:
6055 * - mark all pages reserved
6056 * - mark all memory queues empty
6057 * - clear the memory bitmaps
6059 * NOTE: pgdat should get zeroed by caller.
6061 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6064 int nid = pgdat->node_id;
6066 pgdat_resize_init(pgdat);
6067 #ifdef CONFIG_NUMA_BALANCING
6068 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6069 pgdat->numabalancing_migrate_nr_pages = 0;
6070 pgdat->numabalancing_migrate_next_window = jiffies;
6072 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6073 spin_lock_init(&pgdat->split_queue_lock);
6074 INIT_LIST_HEAD(&pgdat->split_queue);
6075 pgdat->split_queue_len = 0;
6077 init_waitqueue_head(&pgdat->kswapd_wait);
6078 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6079 #ifdef CONFIG_COMPACTION
6080 init_waitqueue_head(&pgdat->kcompactd_wait);
6082 pgdat_page_ext_init(pgdat);
6083 spin_lock_init(&pgdat->lru_lock);
6084 lruvec_init(node_lruvec(pgdat));
6086 pgdat->per_cpu_nodestats = &boot_nodestats;
6088 for (j = 0; j < MAX_NR_ZONES; j++) {
6089 struct zone *zone = pgdat->node_zones + j;
6090 unsigned long size, realsize, freesize, memmap_pages;
6091 unsigned long zone_start_pfn = zone->zone_start_pfn;
6093 size = zone->spanned_pages;
6094 realsize = freesize = zone->present_pages;
6097 * Adjust freesize so that it accounts for how much memory
6098 * is used by this zone for memmap. This affects the watermark
6099 * and per-cpu initialisations
6101 memmap_pages = calc_memmap_size(size, realsize);
6102 if (!is_highmem_idx(j)) {
6103 if (freesize >= memmap_pages) {
6104 freesize -= memmap_pages;
6107 " %s zone: %lu pages used for memmap\n",
6108 zone_names[j], memmap_pages);
6110 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6111 zone_names[j], memmap_pages, freesize);
6114 /* Account for reserved pages */
6115 if (j == 0 && freesize > dma_reserve) {
6116 freesize -= dma_reserve;
6117 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6118 zone_names[0], dma_reserve);
6121 if (!is_highmem_idx(j))
6122 nr_kernel_pages += freesize;
6123 /* Charge for highmem memmap if there are enough kernel pages */
6124 else if (nr_kernel_pages > memmap_pages * 2)
6125 nr_kernel_pages -= memmap_pages;
6126 nr_all_pages += freesize;
6129 * Set an approximate value for lowmem here, it will be adjusted
6130 * when the bootmem allocator frees pages into the buddy system.
6131 * And all highmem pages will be managed by the buddy system.
6133 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6137 zone->name = zone_names[j];
6138 zone->zone_pgdat = pgdat;
6139 spin_lock_init(&zone->lock);
6140 zone_seqlock_init(zone);
6141 zone_pcp_init(zone);
6146 set_pageblock_order();
6147 setup_usemap(pgdat, zone, zone_start_pfn, size);
6148 init_currently_empty_zone(zone, zone_start_pfn, size);
6149 memmap_init(size, nid, j, zone_start_pfn);
6153 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6155 unsigned long __maybe_unused start = 0;
6156 unsigned long __maybe_unused offset = 0;
6158 /* Skip empty nodes */
6159 if (!pgdat->node_spanned_pages)
6162 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6163 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6164 offset = pgdat->node_start_pfn - start;
6165 /* ia64 gets its own node_mem_map, before this, without bootmem */
6166 if (!pgdat->node_mem_map) {
6167 unsigned long size, end;
6171 * The zone's endpoints aren't required to be MAX_ORDER
6172 * aligned but the node_mem_map endpoints must be in order
6173 * for the buddy allocator to function correctly.
6175 end = pgdat_end_pfn(pgdat);
6176 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6177 size = (end - start) * sizeof(struct page);
6178 map = alloc_remap(pgdat->node_id, size);
6180 map = memblock_virt_alloc_node_nopanic(size,
6182 pgdat->node_mem_map = map + offset;
6184 #ifndef CONFIG_NEED_MULTIPLE_NODES
6186 * With no DISCONTIG, the global mem_map is just set as node 0's
6188 if (pgdat == NODE_DATA(0)) {
6189 mem_map = NODE_DATA(0)->node_mem_map;
6190 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6191 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6193 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6196 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6199 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6200 unsigned long node_start_pfn, unsigned long *zholes_size)
6202 pg_data_t *pgdat = NODE_DATA(nid);
6203 unsigned long start_pfn = 0;
6204 unsigned long end_pfn = 0;
6206 /* pg_data_t should be reset to zero when it's allocated */
6207 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6209 pgdat->node_id = nid;
6210 pgdat->node_start_pfn = node_start_pfn;
6211 pgdat->per_cpu_nodestats = NULL;
6212 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6213 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6214 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6215 (u64)start_pfn << PAGE_SHIFT,
6216 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6218 start_pfn = node_start_pfn;
6220 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6221 zones_size, zholes_size);
6223 alloc_node_mem_map(pgdat);
6224 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6225 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6226 nid, (unsigned long)pgdat,
6227 (unsigned long)pgdat->node_mem_map);
6230 reset_deferred_meminit(pgdat);
6231 free_area_init_core(pgdat);
6234 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6236 #if MAX_NUMNODES > 1
6238 * Figure out the number of possible node ids.
6240 void __init setup_nr_node_ids(void)
6242 unsigned int highest;
6244 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6245 nr_node_ids = highest + 1;
6250 * node_map_pfn_alignment - determine the maximum internode alignment
6252 * This function should be called after node map is populated and sorted.
6253 * It calculates the maximum power of two alignment which can distinguish
6256 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6257 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6258 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6259 * shifted, 1GiB is enough and this function will indicate so.
6261 * This is used to test whether pfn -> nid mapping of the chosen memory
6262 * model has fine enough granularity to avoid incorrect mapping for the
6263 * populated node map.
6265 * Returns the determined alignment in pfn's. 0 if there is no alignment
6266 * requirement (single node).
6268 unsigned long __init node_map_pfn_alignment(void)
6270 unsigned long accl_mask = 0, last_end = 0;
6271 unsigned long start, end, mask;
6275 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6276 if (!start || last_nid < 0 || last_nid == nid) {
6283 * Start with a mask granular enough to pin-point to the
6284 * start pfn and tick off bits one-by-one until it becomes
6285 * too coarse to separate the current node from the last.
6287 mask = ~((1 << __ffs(start)) - 1);
6288 while (mask && last_end <= (start & (mask << 1)))
6291 /* accumulate all internode masks */
6295 /* convert mask to number of pages */
6296 return ~accl_mask + 1;
6299 /* Find the lowest pfn for a node */
6300 static unsigned long __init find_min_pfn_for_node(int nid)
6302 unsigned long min_pfn = ULONG_MAX;
6303 unsigned long start_pfn;
6306 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6307 min_pfn = min(min_pfn, start_pfn);
6309 if (min_pfn == ULONG_MAX) {
6310 pr_warn("Could not find start_pfn for node %d\n", nid);
6318 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6320 * It returns the minimum PFN based on information provided via
6321 * memblock_set_node().
6323 unsigned long __init find_min_pfn_with_active_regions(void)
6325 return find_min_pfn_for_node(MAX_NUMNODES);
6329 * early_calculate_totalpages()
6330 * Sum pages in active regions for movable zone.
6331 * Populate N_MEMORY for calculating usable_nodes.
6333 static unsigned long __init early_calculate_totalpages(void)
6335 unsigned long totalpages = 0;
6336 unsigned long start_pfn, end_pfn;
6339 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6340 unsigned long pages = end_pfn - start_pfn;
6342 totalpages += pages;
6344 node_set_state(nid, N_MEMORY);
6350 * Find the PFN the Movable zone begins in each node. Kernel memory
6351 * is spread evenly between nodes as long as the nodes have enough
6352 * memory. When they don't, some nodes will have more kernelcore than
6355 static void __init find_zone_movable_pfns_for_nodes(void)
6358 unsigned long usable_startpfn;
6359 unsigned long kernelcore_node, kernelcore_remaining;
6360 /* save the state before borrow the nodemask */
6361 nodemask_t saved_node_state = node_states[N_MEMORY];
6362 unsigned long totalpages = early_calculate_totalpages();
6363 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6364 struct memblock_region *r;
6366 /* Need to find movable_zone earlier when movable_node is specified. */
6367 find_usable_zone_for_movable();
6370 * If movable_node is specified, ignore kernelcore and movablecore
6373 if (movable_node_is_enabled()) {
6374 for_each_memblock(memory, r) {
6375 if (!memblock_is_hotpluggable(r))
6380 usable_startpfn = PFN_DOWN(r->base);
6381 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6382 min(usable_startpfn, zone_movable_pfn[nid]) :
6390 * If kernelcore=mirror is specified, ignore movablecore option
6392 if (mirrored_kernelcore) {
6393 bool mem_below_4gb_not_mirrored = false;
6395 for_each_memblock(memory, r) {
6396 if (memblock_is_mirror(r))
6401 usable_startpfn = memblock_region_memory_base_pfn(r);
6403 if (usable_startpfn < 0x100000) {
6404 mem_below_4gb_not_mirrored = true;
6408 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6409 min(usable_startpfn, zone_movable_pfn[nid]) :
6413 if (mem_below_4gb_not_mirrored)
6414 pr_warn("This configuration results in unmirrored kernel memory.");
6420 * If movablecore=nn[KMG] was specified, calculate what size of
6421 * kernelcore that corresponds so that memory usable for
6422 * any allocation type is evenly spread. If both kernelcore
6423 * and movablecore are specified, then the value of kernelcore
6424 * will be used for required_kernelcore if it's greater than
6425 * what movablecore would have allowed.
6427 if (required_movablecore) {
6428 unsigned long corepages;
6431 * Round-up so that ZONE_MOVABLE is at least as large as what
6432 * was requested by the user
6434 required_movablecore =
6435 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6436 required_movablecore = min(totalpages, required_movablecore);
6437 corepages = totalpages - required_movablecore;
6439 required_kernelcore = max(required_kernelcore, corepages);
6443 * If kernelcore was not specified or kernelcore size is larger
6444 * than totalpages, there is no ZONE_MOVABLE.
6446 if (!required_kernelcore || required_kernelcore >= totalpages)
6449 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6450 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6453 /* Spread kernelcore memory as evenly as possible throughout nodes */
6454 kernelcore_node = required_kernelcore / usable_nodes;
6455 for_each_node_state(nid, N_MEMORY) {
6456 unsigned long start_pfn, end_pfn;
6459 * Recalculate kernelcore_node if the division per node
6460 * now exceeds what is necessary to satisfy the requested
6461 * amount of memory for the kernel
6463 if (required_kernelcore < kernelcore_node)
6464 kernelcore_node = required_kernelcore / usable_nodes;
6467 * As the map is walked, we track how much memory is usable
6468 * by the kernel using kernelcore_remaining. When it is
6469 * 0, the rest of the node is usable by ZONE_MOVABLE
6471 kernelcore_remaining = kernelcore_node;
6473 /* Go through each range of PFNs within this node */
6474 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6475 unsigned long size_pages;
6477 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6478 if (start_pfn >= end_pfn)
6481 /* Account for what is only usable for kernelcore */
6482 if (start_pfn < usable_startpfn) {
6483 unsigned long kernel_pages;
6484 kernel_pages = min(end_pfn, usable_startpfn)
6487 kernelcore_remaining -= min(kernel_pages,
6488 kernelcore_remaining);
6489 required_kernelcore -= min(kernel_pages,
6490 required_kernelcore);
6492 /* Continue if range is now fully accounted */
6493 if (end_pfn <= usable_startpfn) {
6496 * Push zone_movable_pfn to the end so
6497 * that if we have to rebalance
6498 * kernelcore across nodes, we will
6499 * not double account here
6501 zone_movable_pfn[nid] = end_pfn;
6504 start_pfn = usable_startpfn;
6508 * The usable PFN range for ZONE_MOVABLE is from
6509 * start_pfn->end_pfn. Calculate size_pages as the
6510 * number of pages used as kernelcore
6512 size_pages = end_pfn - start_pfn;
6513 if (size_pages > kernelcore_remaining)
6514 size_pages = kernelcore_remaining;
6515 zone_movable_pfn[nid] = start_pfn + size_pages;
6518 * Some kernelcore has been met, update counts and
6519 * break if the kernelcore for this node has been
6522 required_kernelcore -= min(required_kernelcore,
6524 kernelcore_remaining -= size_pages;
6525 if (!kernelcore_remaining)
6531 * If there is still required_kernelcore, we do another pass with one
6532 * less node in the count. This will push zone_movable_pfn[nid] further
6533 * along on the nodes that still have memory until kernelcore is
6537 if (usable_nodes && required_kernelcore > usable_nodes)
6541 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6542 for (nid = 0; nid < MAX_NUMNODES; nid++)
6543 zone_movable_pfn[nid] =
6544 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6547 /* restore the node_state */
6548 node_states[N_MEMORY] = saved_node_state;
6551 /* Any regular or high memory on that node ? */
6552 static void check_for_memory(pg_data_t *pgdat, int nid)
6554 enum zone_type zone_type;
6556 if (N_MEMORY == N_NORMAL_MEMORY)
6559 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6560 struct zone *zone = &pgdat->node_zones[zone_type];
6561 if (populated_zone(zone)) {
6562 node_set_state(nid, N_HIGH_MEMORY);
6563 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6564 zone_type <= ZONE_NORMAL)
6565 node_set_state(nid, N_NORMAL_MEMORY);
6572 * free_area_init_nodes - Initialise all pg_data_t and zone data
6573 * @max_zone_pfn: an array of max PFNs for each zone
6575 * This will call free_area_init_node() for each active node in the system.
6576 * Using the page ranges provided by memblock_set_node(), the size of each
6577 * zone in each node and their holes is calculated. If the maximum PFN
6578 * between two adjacent zones match, it is assumed that the zone is empty.
6579 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6580 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6581 * starts where the previous one ended. For example, ZONE_DMA32 starts
6582 * at arch_max_dma_pfn.
6584 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6586 unsigned long start_pfn, end_pfn;
6589 /* Record where the zone boundaries are */
6590 memset(arch_zone_lowest_possible_pfn, 0,
6591 sizeof(arch_zone_lowest_possible_pfn));
6592 memset(arch_zone_highest_possible_pfn, 0,
6593 sizeof(arch_zone_highest_possible_pfn));
6595 start_pfn = find_min_pfn_with_active_regions();
6597 for (i = 0; i < MAX_NR_ZONES; i++) {
6598 if (i == ZONE_MOVABLE)
6601 end_pfn = max(max_zone_pfn[i], start_pfn);
6602 arch_zone_lowest_possible_pfn[i] = start_pfn;
6603 arch_zone_highest_possible_pfn[i] = end_pfn;
6605 start_pfn = end_pfn;
6608 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6609 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6610 find_zone_movable_pfns_for_nodes();
6612 /* Print out the zone ranges */
6613 pr_info("Zone ranges:\n");
6614 for (i = 0; i < MAX_NR_ZONES; i++) {
6615 if (i == ZONE_MOVABLE)
6617 pr_info(" %-8s ", zone_names[i]);
6618 if (arch_zone_lowest_possible_pfn[i] ==
6619 arch_zone_highest_possible_pfn[i])
6622 pr_cont("[mem %#018Lx-%#018Lx]\n",
6623 (u64)arch_zone_lowest_possible_pfn[i]
6625 ((u64)arch_zone_highest_possible_pfn[i]
6626 << PAGE_SHIFT) - 1);
6629 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6630 pr_info("Movable zone start for each node\n");
6631 for (i = 0; i < MAX_NUMNODES; i++) {
6632 if (zone_movable_pfn[i])
6633 pr_info(" Node %d: %#018Lx\n", i,
6634 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6637 /* Print out the early node map */
6638 pr_info("Early memory node ranges\n");
6639 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6640 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6641 (u64)start_pfn << PAGE_SHIFT,
6642 ((u64)end_pfn << PAGE_SHIFT) - 1);
6644 /* Initialise every node */
6645 mminit_verify_pageflags_layout();
6646 setup_nr_node_ids();
6647 for_each_online_node(nid) {
6648 pg_data_t *pgdat = NODE_DATA(nid);
6649 free_area_init_node(nid, NULL,
6650 find_min_pfn_for_node(nid), NULL);
6652 /* Any memory on that node */
6653 if (pgdat->node_present_pages)
6654 node_set_state(nid, N_MEMORY);
6655 check_for_memory(pgdat, nid);
6659 static int __init cmdline_parse_core(char *p, unsigned long *core)
6661 unsigned long long coremem;
6665 coremem = memparse(p, &p);
6666 *core = coremem >> PAGE_SHIFT;
6668 /* Paranoid check that UL is enough for the coremem value */
6669 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6675 * kernelcore=size sets the amount of memory for use for allocations that
6676 * cannot be reclaimed or migrated.
6678 static int __init cmdline_parse_kernelcore(char *p)
6680 /* parse kernelcore=mirror */
6681 if (parse_option_str(p, "mirror")) {
6682 mirrored_kernelcore = true;
6686 return cmdline_parse_core(p, &required_kernelcore);
6690 * movablecore=size sets the amount of memory for use for allocations that
6691 * can be reclaimed or migrated.
6693 static int __init cmdline_parse_movablecore(char *p)
6695 return cmdline_parse_core(p, &required_movablecore);
6698 early_param("kernelcore", cmdline_parse_kernelcore);
6699 early_param("movablecore", cmdline_parse_movablecore);
6701 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6703 void adjust_managed_page_count(struct page *page, long count)
6705 spin_lock(&managed_page_count_lock);
6706 page_zone(page)->managed_pages += count;
6707 totalram_pages += count;
6708 #ifdef CONFIG_HIGHMEM
6709 if (PageHighMem(page))
6710 totalhigh_pages += count;
6712 spin_unlock(&managed_page_count_lock);
6714 EXPORT_SYMBOL(adjust_managed_page_count);
6716 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6719 unsigned long pages = 0;
6721 start = (void *)PAGE_ALIGN((unsigned long)start);
6722 end = (void *)((unsigned long)end & PAGE_MASK);
6723 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6724 if ((unsigned int)poison <= 0xFF)
6725 memset(pos, poison, PAGE_SIZE);
6726 free_reserved_page(virt_to_page(pos));
6730 pr_info("Freeing %s memory: %ldK\n",
6731 s, pages << (PAGE_SHIFT - 10));
6735 EXPORT_SYMBOL(free_reserved_area);
6737 #ifdef CONFIG_HIGHMEM
6738 void free_highmem_page(struct page *page)
6740 __free_reserved_page(page);
6742 page_zone(page)->managed_pages++;
6748 void __init mem_init_print_info(const char *str)
6750 unsigned long physpages, codesize, datasize, rosize, bss_size;
6751 unsigned long init_code_size, init_data_size;
6753 physpages = get_num_physpages();
6754 codesize = _etext - _stext;
6755 datasize = _edata - _sdata;
6756 rosize = __end_rodata - __start_rodata;
6757 bss_size = __bss_stop - __bss_start;
6758 init_data_size = __init_end - __init_begin;
6759 init_code_size = _einittext - _sinittext;
6762 * Detect special cases and adjust section sizes accordingly:
6763 * 1) .init.* may be embedded into .data sections
6764 * 2) .init.text.* may be out of [__init_begin, __init_end],
6765 * please refer to arch/tile/kernel/vmlinux.lds.S.
6766 * 3) .rodata.* may be embedded into .text or .data sections.
6768 #define adj_init_size(start, end, size, pos, adj) \
6770 if (start <= pos && pos < end && size > adj) \
6774 adj_init_size(__init_begin, __init_end, init_data_size,
6775 _sinittext, init_code_size);
6776 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6777 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6778 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6779 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6781 #undef adj_init_size
6783 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6784 #ifdef CONFIG_HIGHMEM
6788 nr_free_pages() << (PAGE_SHIFT - 10),
6789 physpages << (PAGE_SHIFT - 10),
6790 codesize >> 10, datasize >> 10, rosize >> 10,
6791 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6792 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6793 totalcma_pages << (PAGE_SHIFT - 10),
6794 #ifdef CONFIG_HIGHMEM
6795 totalhigh_pages << (PAGE_SHIFT - 10),
6797 str ? ", " : "", str ? str : "");
6801 * set_dma_reserve - set the specified number of pages reserved in the first zone
6802 * @new_dma_reserve: The number of pages to mark reserved
6804 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6805 * In the DMA zone, a significant percentage may be consumed by kernel image
6806 * and other unfreeable allocations which can skew the watermarks badly. This
6807 * function may optionally be used to account for unfreeable pages in the
6808 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6809 * smaller per-cpu batchsize.
6811 void __init set_dma_reserve(unsigned long new_dma_reserve)
6813 dma_reserve = new_dma_reserve;
6816 void __init free_area_init(unsigned long *zones_size)
6818 free_area_init_node(0, zones_size,
6819 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6822 static int page_alloc_cpu_dead(unsigned int cpu)
6825 lru_add_drain_cpu(cpu);
6829 * Spill the event counters of the dead processor
6830 * into the current processors event counters.
6831 * This artificially elevates the count of the current
6834 vm_events_fold_cpu(cpu);
6837 * Zero the differential counters of the dead processor
6838 * so that the vm statistics are consistent.
6840 * This is only okay since the processor is dead and cannot
6841 * race with what we are doing.
6843 cpu_vm_stats_fold(cpu);
6847 void __init page_alloc_init(void)
6851 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6852 "mm/page_alloc:dead", NULL,
6853 page_alloc_cpu_dead);
6858 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6859 * or min_free_kbytes changes.
6861 static void calculate_totalreserve_pages(void)
6863 struct pglist_data *pgdat;
6864 unsigned long reserve_pages = 0;
6865 enum zone_type i, j;
6867 for_each_online_pgdat(pgdat) {
6869 pgdat->totalreserve_pages = 0;
6871 for (i = 0; i < MAX_NR_ZONES; i++) {
6872 struct zone *zone = pgdat->node_zones + i;
6875 /* Find valid and maximum lowmem_reserve in the zone */
6876 for (j = i; j < MAX_NR_ZONES; j++) {
6877 if (zone->lowmem_reserve[j] > max)
6878 max = zone->lowmem_reserve[j];
6881 /* we treat the high watermark as reserved pages. */
6882 max += high_wmark_pages(zone);
6884 if (max > zone->managed_pages)
6885 max = zone->managed_pages;
6887 pgdat->totalreserve_pages += max;
6889 reserve_pages += max;
6892 totalreserve_pages = reserve_pages;
6896 * setup_per_zone_lowmem_reserve - called whenever
6897 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6898 * has a correct pages reserved value, so an adequate number of
6899 * pages are left in the zone after a successful __alloc_pages().
6901 static void setup_per_zone_lowmem_reserve(void)
6903 struct pglist_data *pgdat;
6904 enum zone_type j, idx;
6906 for_each_online_pgdat(pgdat) {
6907 for (j = 0; j < MAX_NR_ZONES; j++) {
6908 struct zone *zone = pgdat->node_zones + j;
6909 unsigned long managed_pages = zone->managed_pages;
6911 zone->lowmem_reserve[j] = 0;
6915 struct zone *lower_zone;
6919 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6920 sysctl_lowmem_reserve_ratio[idx] = 1;
6922 lower_zone = pgdat->node_zones + idx;
6923 lower_zone->lowmem_reserve[j] = managed_pages /
6924 sysctl_lowmem_reserve_ratio[idx];
6925 managed_pages += lower_zone->managed_pages;
6930 /* update totalreserve_pages */
6931 calculate_totalreserve_pages();
6934 static void __setup_per_zone_wmarks(void)
6936 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6937 unsigned long lowmem_pages = 0;
6939 unsigned long flags;
6941 /* Calculate total number of !ZONE_HIGHMEM pages */
6942 for_each_zone(zone) {
6943 if (!is_highmem(zone))
6944 lowmem_pages += zone->managed_pages;
6947 for_each_zone(zone) {
6950 spin_lock_irqsave(&zone->lock, flags);
6951 tmp = (u64)pages_min * zone->managed_pages;
6952 do_div(tmp, lowmem_pages);
6953 if (is_highmem(zone)) {
6955 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6956 * need highmem pages, so cap pages_min to a small
6959 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6960 * deltas control asynch page reclaim, and so should
6961 * not be capped for highmem.
6963 unsigned long min_pages;
6965 min_pages = zone->managed_pages / 1024;
6966 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6967 zone->watermark[WMARK_MIN] = min_pages;
6970 * If it's a lowmem zone, reserve a number of pages
6971 * proportionate to the zone's size.
6973 zone->watermark[WMARK_MIN] = tmp;
6977 * Set the kswapd watermarks distance according to the
6978 * scale factor in proportion to available memory, but
6979 * ensure a minimum size on small systems.
6981 tmp = max_t(u64, tmp >> 2,
6982 mult_frac(zone->managed_pages,
6983 watermark_scale_factor, 10000));
6985 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6986 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6988 spin_unlock_irqrestore(&zone->lock, flags);
6991 /* update totalreserve_pages */
6992 calculate_totalreserve_pages();
6996 * setup_per_zone_wmarks - called when min_free_kbytes changes
6997 * or when memory is hot-{added|removed}
6999 * Ensures that the watermark[min,low,high] values for each zone are set
7000 * correctly with respect to min_free_kbytes.
7002 void setup_per_zone_wmarks(void)
7004 mutex_lock(&zonelists_mutex);
7005 __setup_per_zone_wmarks();
7006 mutex_unlock(&zonelists_mutex);
7010 * Initialise min_free_kbytes.
7012 * For small machines we want it small (128k min). For large machines
7013 * we want it large (64MB max). But it is not linear, because network
7014 * bandwidth does not increase linearly with machine size. We use
7016 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7017 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7033 int __meminit init_per_zone_wmark_min(void)
7035 unsigned long lowmem_kbytes;
7036 int new_min_free_kbytes;
7038 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7039 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7041 if (new_min_free_kbytes > user_min_free_kbytes) {
7042 min_free_kbytes = new_min_free_kbytes;
7043 if (min_free_kbytes < 128)
7044 min_free_kbytes = 128;
7045 if (min_free_kbytes > 65536)
7046 min_free_kbytes = 65536;
7048 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7049 new_min_free_kbytes, user_min_free_kbytes);
7051 setup_per_zone_wmarks();
7052 refresh_zone_stat_thresholds();
7053 setup_per_zone_lowmem_reserve();
7056 setup_min_unmapped_ratio();
7057 setup_min_slab_ratio();
7062 core_initcall(init_per_zone_wmark_min)
7065 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7066 * that we can call two helper functions whenever min_free_kbytes
7069 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7070 void __user *buffer, size_t *length, loff_t *ppos)
7074 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7079 user_min_free_kbytes = min_free_kbytes;
7080 setup_per_zone_wmarks();
7085 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7086 void __user *buffer, size_t *length, loff_t *ppos)
7090 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7095 setup_per_zone_wmarks();
7101 static void setup_min_unmapped_ratio(void)
7106 for_each_online_pgdat(pgdat)
7107 pgdat->min_unmapped_pages = 0;
7110 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7111 sysctl_min_unmapped_ratio) / 100;
7115 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7116 void __user *buffer, size_t *length, loff_t *ppos)
7120 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7124 setup_min_unmapped_ratio();
7129 static void setup_min_slab_ratio(void)
7134 for_each_online_pgdat(pgdat)
7135 pgdat->min_slab_pages = 0;
7138 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7139 sysctl_min_slab_ratio) / 100;
7142 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7143 void __user *buffer, size_t *length, loff_t *ppos)
7147 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7151 setup_min_slab_ratio();
7158 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7159 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7160 * whenever sysctl_lowmem_reserve_ratio changes.
7162 * The reserve ratio obviously has absolutely no relation with the
7163 * minimum watermarks. The lowmem reserve ratio can only make sense
7164 * if in function of the boot time zone sizes.
7166 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7167 void __user *buffer, size_t *length, loff_t *ppos)
7169 proc_dointvec_minmax(table, write, buffer, length, ppos);
7170 setup_per_zone_lowmem_reserve();
7175 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7176 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7177 * pagelist can have before it gets flushed back to buddy allocator.
7179 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7180 void __user *buffer, size_t *length, loff_t *ppos)
7183 int old_percpu_pagelist_fraction;
7186 mutex_lock(&pcp_batch_high_lock);
7187 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7189 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7190 if (!write || ret < 0)
7193 /* Sanity checking to avoid pcp imbalance */
7194 if (percpu_pagelist_fraction &&
7195 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7196 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7202 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7205 for_each_populated_zone(zone) {
7208 for_each_possible_cpu(cpu)
7209 pageset_set_high_and_batch(zone,
7210 per_cpu_ptr(zone->pageset, cpu));
7213 mutex_unlock(&pcp_batch_high_lock);
7218 int hashdist = HASHDIST_DEFAULT;
7220 static int __init set_hashdist(char *str)
7224 hashdist = simple_strtoul(str, &str, 0);
7227 __setup("hashdist=", set_hashdist);
7230 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7232 * Returns the number of pages that arch has reserved but
7233 * is not known to alloc_large_system_hash().
7235 static unsigned long __init arch_reserved_kernel_pages(void)
7242 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7243 * machines. As memory size is increased the scale is also increased but at
7244 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7245 * quadruples the scale is increased by one, which means the size of hash table
7246 * only doubles, instead of quadrupling as well.
7247 * Because 32-bit systems cannot have large physical memory, where this scaling
7248 * makes sense, it is disabled on such platforms.
7250 #if __BITS_PER_LONG > 32
7251 #define ADAPT_SCALE_BASE (64ul << 30)
7252 #define ADAPT_SCALE_SHIFT 2
7253 #define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7257 * allocate a large system hash table from bootmem
7258 * - it is assumed that the hash table must contain an exact power-of-2
7259 * quantity of entries
7260 * - limit is the number of hash buckets, not the total allocation size
7262 void *__init alloc_large_system_hash(const char *tablename,
7263 unsigned long bucketsize,
7264 unsigned long numentries,
7267 unsigned int *_hash_shift,
7268 unsigned int *_hash_mask,
7269 unsigned long low_limit,
7270 unsigned long high_limit)
7272 unsigned long long max = high_limit;
7273 unsigned long log2qty, size;
7277 /* allow the kernel cmdline to have a say */
7279 /* round applicable memory size up to nearest megabyte */
7280 numentries = nr_kernel_pages;
7281 numentries -= arch_reserved_kernel_pages();
7283 /* It isn't necessary when PAGE_SIZE >= 1MB */
7284 if (PAGE_SHIFT < 20)
7285 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7287 #if __BITS_PER_LONG > 32
7289 unsigned long adapt;
7291 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7292 adapt <<= ADAPT_SCALE_SHIFT)
7297 /* limit to 1 bucket per 2^scale bytes of low memory */
7298 if (scale > PAGE_SHIFT)
7299 numentries >>= (scale - PAGE_SHIFT);
7301 numentries <<= (PAGE_SHIFT - scale);
7303 /* Make sure we've got at least a 0-order allocation.. */
7304 if (unlikely(flags & HASH_SMALL)) {
7305 /* Makes no sense without HASH_EARLY */
7306 WARN_ON(!(flags & HASH_EARLY));
7307 if (!(numentries >> *_hash_shift)) {
7308 numentries = 1UL << *_hash_shift;
7309 BUG_ON(!numentries);
7311 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7312 numentries = PAGE_SIZE / bucketsize;
7314 numentries = roundup_pow_of_two(numentries);
7316 /* limit allocation size to 1/16 total memory by default */
7318 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7319 do_div(max, bucketsize);
7321 max = min(max, 0x80000000ULL);
7323 if (numentries < low_limit)
7324 numentries = low_limit;
7325 if (numentries > max)
7328 log2qty = ilog2(numentries);
7331 * memblock allocator returns zeroed memory already, so HASH_ZERO is
7332 * currently not used when HASH_EARLY is specified.
7334 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7336 size = bucketsize << log2qty;
7337 if (flags & HASH_EARLY)
7338 table = memblock_virt_alloc_nopanic(size, 0);
7340 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7343 * If bucketsize is not a power-of-two, we may free
7344 * some pages at the end of hash table which
7345 * alloc_pages_exact() automatically does
7347 if (get_order(size) < MAX_ORDER) {
7348 table = alloc_pages_exact(size, gfp_flags);
7349 kmemleak_alloc(table, size, 1, gfp_flags);
7352 } while (!table && size > PAGE_SIZE && --log2qty);
7355 panic("Failed to allocate %s hash table\n", tablename);
7357 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7358 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7361 *_hash_shift = log2qty;
7363 *_hash_mask = (1 << log2qty) - 1;
7369 * This function checks whether pageblock includes unmovable pages or not.
7370 * If @count is not zero, it is okay to include less @count unmovable pages
7372 * PageLRU check without isolation or lru_lock could race so that
7373 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7374 * check without lock_page also may miss some movable non-lru pages at
7375 * race condition. So you can't expect this function should be exact.
7377 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7378 bool skip_hwpoisoned_pages)
7380 unsigned long pfn, iter, found;
7384 * For avoiding noise data, lru_add_drain_all() should be called
7385 * If ZONE_MOVABLE, the zone never contains unmovable pages
7387 if (zone_idx(zone) == ZONE_MOVABLE)
7389 mt = get_pageblock_migratetype(page);
7390 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7393 pfn = page_to_pfn(page);
7394 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7395 unsigned long check = pfn + iter;
7397 if (!pfn_valid_within(check))
7400 page = pfn_to_page(check);
7403 * Hugepages are not in LRU lists, but they're movable.
7404 * We need not scan over tail pages bacause we don't
7405 * handle each tail page individually in migration.
7407 if (PageHuge(page)) {
7408 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7413 * We can't use page_count without pin a page
7414 * because another CPU can free compound page.
7415 * This check already skips compound tails of THP
7416 * because their page->_refcount is zero at all time.
7418 if (!page_ref_count(page)) {
7419 if (PageBuddy(page))
7420 iter += (1 << page_order(page)) - 1;
7425 * The HWPoisoned page may be not in buddy system, and
7426 * page_count() is not 0.
7428 if (skip_hwpoisoned_pages && PageHWPoison(page))
7431 if (__PageMovable(page))
7437 * If there are RECLAIMABLE pages, we need to check
7438 * it. But now, memory offline itself doesn't call
7439 * shrink_node_slabs() and it still to be fixed.
7442 * If the page is not RAM, page_count()should be 0.
7443 * we don't need more check. This is an _used_ not-movable page.
7445 * The problematic thing here is PG_reserved pages. PG_reserved
7446 * is set to both of a memory hole page and a _used_ kernel
7455 bool is_pageblock_removable_nolock(struct page *page)
7461 * We have to be careful here because we are iterating over memory
7462 * sections which are not zone aware so we might end up outside of
7463 * the zone but still within the section.
7464 * We have to take care about the node as well. If the node is offline
7465 * its NODE_DATA will be NULL - see page_zone.
7467 if (!node_online(page_to_nid(page)))
7470 zone = page_zone(page);
7471 pfn = page_to_pfn(page);
7472 if (!zone_spans_pfn(zone, pfn))
7475 return !has_unmovable_pages(zone, page, 0, true);
7478 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7480 static unsigned long pfn_max_align_down(unsigned long pfn)
7482 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7483 pageblock_nr_pages) - 1);
7486 static unsigned long pfn_max_align_up(unsigned long pfn)
7488 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7489 pageblock_nr_pages));
7492 /* [start, end) must belong to a single zone. */
7493 static int __alloc_contig_migrate_range(struct compact_control *cc,
7494 unsigned long start, unsigned long end)
7496 /* This function is based on compact_zone() from compaction.c. */
7497 unsigned long nr_reclaimed;
7498 unsigned long pfn = start;
7499 unsigned int tries = 0;
7504 while (pfn < end || !list_empty(&cc->migratepages)) {
7505 if (fatal_signal_pending(current)) {
7510 if (list_empty(&cc->migratepages)) {
7511 cc->nr_migratepages = 0;
7512 pfn = isolate_migratepages_range(cc, pfn, end);
7518 } else if (++tries == 5) {
7519 ret = ret < 0 ? ret : -EBUSY;
7523 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7525 cc->nr_migratepages -= nr_reclaimed;
7527 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7528 NULL, 0, cc->mode, MR_CMA);
7531 putback_movable_pages(&cc->migratepages);
7538 * alloc_contig_range() -- tries to allocate given range of pages
7539 * @start: start PFN to allocate
7540 * @end: one-past-the-last PFN to allocate
7541 * @migratetype: migratetype of the underlaying pageblocks (either
7542 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7543 * in range must have the same migratetype and it must
7544 * be either of the two.
7545 * @gfp_mask: GFP mask to use during compaction
7547 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7548 * aligned, however it's the caller's responsibility to guarantee that
7549 * we are the only thread that changes migrate type of pageblocks the
7552 * The PFN range must belong to a single zone.
7554 * Returns zero on success or negative error code. On success all
7555 * pages which PFN is in [start, end) are allocated for the caller and
7556 * need to be freed with free_contig_range().
7558 int alloc_contig_range(unsigned long start, unsigned long end,
7559 unsigned migratetype, gfp_t gfp_mask)
7561 unsigned long outer_start, outer_end;
7565 struct compact_control cc = {
7566 .nr_migratepages = 0,
7568 .zone = page_zone(pfn_to_page(start)),
7569 .mode = MIGRATE_SYNC,
7570 .ignore_skip_hint = true,
7571 .gfp_mask = current_gfp_context(gfp_mask),
7573 INIT_LIST_HEAD(&cc.migratepages);
7576 * What we do here is we mark all pageblocks in range as
7577 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7578 * have different sizes, and due to the way page allocator
7579 * work, we align the range to biggest of the two pages so
7580 * that page allocator won't try to merge buddies from
7581 * different pageblocks and change MIGRATE_ISOLATE to some
7582 * other migration type.
7584 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7585 * migrate the pages from an unaligned range (ie. pages that
7586 * we are interested in). This will put all the pages in
7587 * range back to page allocator as MIGRATE_ISOLATE.
7589 * When this is done, we take the pages in range from page
7590 * allocator removing them from the buddy system. This way
7591 * page allocator will never consider using them.
7593 * This lets us mark the pageblocks back as
7594 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7595 * aligned range but not in the unaligned, original range are
7596 * put back to page allocator so that buddy can use them.
7599 ret = start_isolate_page_range(pfn_max_align_down(start),
7600 pfn_max_align_up(end), migratetype,
7606 * In case of -EBUSY, we'd like to know which page causes problem.
7607 * So, just fall through. We will check it in test_pages_isolated().
7609 ret = __alloc_contig_migrate_range(&cc, start, end);
7610 if (ret && ret != -EBUSY)
7614 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7615 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7616 * more, all pages in [start, end) are free in page allocator.
7617 * What we are going to do is to allocate all pages from
7618 * [start, end) (that is remove them from page allocator).
7620 * The only problem is that pages at the beginning and at the
7621 * end of interesting range may be not aligned with pages that
7622 * page allocator holds, ie. they can be part of higher order
7623 * pages. Because of this, we reserve the bigger range and
7624 * once this is done free the pages we are not interested in.
7626 * We don't have to hold zone->lock here because the pages are
7627 * isolated thus they won't get removed from buddy.
7630 lru_add_drain_all();
7631 drain_all_pages(cc.zone);
7634 outer_start = start;
7635 while (!PageBuddy(pfn_to_page(outer_start))) {
7636 if (++order >= MAX_ORDER) {
7637 outer_start = start;
7640 outer_start &= ~0UL << order;
7643 if (outer_start != start) {
7644 order = page_order(pfn_to_page(outer_start));
7647 * outer_start page could be small order buddy page and
7648 * it doesn't include start page. Adjust outer_start
7649 * in this case to report failed page properly
7650 * on tracepoint in test_pages_isolated()
7652 if (outer_start + (1UL << order) <= start)
7653 outer_start = start;
7656 /* Make sure the range is really isolated. */
7657 if (test_pages_isolated(outer_start, end, false)) {
7658 pr_info("%s: [%lx, %lx) PFNs busy\n",
7659 __func__, outer_start, end);
7664 /* Grab isolated pages from freelists. */
7665 outer_end = isolate_freepages_range(&cc, outer_start, end);
7671 /* Free head and tail (if any) */
7672 if (start != outer_start)
7673 free_contig_range(outer_start, start - outer_start);
7674 if (end != outer_end)
7675 free_contig_range(end, outer_end - end);
7678 undo_isolate_page_range(pfn_max_align_down(start),
7679 pfn_max_align_up(end), migratetype);
7683 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7685 unsigned int count = 0;
7687 for (; nr_pages--; pfn++) {
7688 struct page *page = pfn_to_page(pfn);
7690 count += page_count(page) != 1;
7693 WARN(count != 0, "%d pages are still in use!\n", count);
7697 #ifdef CONFIG_MEMORY_HOTPLUG
7699 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7700 * page high values need to be recalulated.
7702 void __meminit zone_pcp_update(struct zone *zone)
7705 mutex_lock(&pcp_batch_high_lock);
7706 for_each_possible_cpu(cpu)
7707 pageset_set_high_and_batch(zone,
7708 per_cpu_ptr(zone->pageset, cpu));
7709 mutex_unlock(&pcp_batch_high_lock);
7713 void zone_pcp_reset(struct zone *zone)
7715 unsigned long flags;
7717 struct per_cpu_pageset *pset;
7719 /* avoid races with drain_pages() */
7720 local_irq_save(flags);
7721 if (zone->pageset != &boot_pageset) {
7722 for_each_online_cpu(cpu) {
7723 pset = per_cpu_ptr(zone->pageset, cpu);
7724 drain_zonestat(zone, pset);
7726 free_percpu(zone->pageset);
7727 zone->pageset = &boot_pageset;
7729 local_irq_restore(flags);
7732 #ifdef CONFIG_MEMORY_HOTREMOVE
7734 * All pages in the range must be in a single zone and isolated
7735 * before calling this.
7738 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7742 unsigned int order, i;
7744 unsigned long flags;
7745 /* find the first valid pfn */
7746 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7751 offline_mem_sections(pfn, end_pfn);
7752 zone = page_zone(pfn_to_page(pfn));
7753 spin_lock_irqsave(&zone->lock, flags);
7755 while (pfn < end_pfn) {
7756 if (!pfn_valid(pfn)) {
7760 page = pfn_to_page(pfn);
7762 * The HWPoisoned page may be not in buddy system, and
7763 * page_count() is not 0.
7765 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7767 SetPageReserved(page);
7771 BUG_ON(page_count(page));
7772 BUG_ON(!PageBuddy(page));
7773 order = page_order(page);
7774 #ifdef CONFIG_DEBUG_VM
7775 pr_info("remove from free list %lx %d %lx\n",
7776 pfn, 1 << order, end_pfn);
7778 list_del(&page->lru);
7779 rmv_page_order(page);
7780 zone->free_area[order].nr_free--;
7781 for (i = 0; i < (1 << order); i++)
7782 SetPageReserved((page+i));
7783 pfn += (1 << order);
7785 spin_unlock_irqrestore(&zone->lock, flags);
7789 bool is_free_buddy_page(struct page *page)
7791 struct zone *zone = page_zone(page);
7792 unsigned long pfn = page_to_pfn(page);
7793 unsigned long flags;
7796 spin_lock_irqsave(&zone->lock, flags);
7797 for (order = 0; order < MAX_ORDER; order++) {
7798 struct page *page_head = page - (pfn & ((1 << order) - 1));
7800 if (PageBuddy(page_head) && page_order(page_head) >= order)
7803 spin_unlock_irqrestore(&zone->lock, flags);
7805 return order < MAX_ORDER;