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/page_ext.h>
63 #include <linux/hugetlb.h>
64 #include <linux/sched/rt.h>
65 #include <linux/page_owner.h>
66 #include <linux/kthread.h>
67 #include <linux/memcontrol.h>
69 #include <asm/sections.h>
70 #include <asm/tlbflush.h>
71 #include <asm/div64.h>
74 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
75 static DEFINE_MUTEX(pcp_batch_high_lock);
76 #define MIN_PERCPU_PAGELIST_FRACTION (8)
78 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
79 DEFINE_PER_CPU(int, numa_node);
80 EXPORT_PER_CPU_SYMBOL(numa_node);
83 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
85 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
86 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
87 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
88 * defined in <linux/topology.h>.
90 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
91 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
92 int _node_numa_mem_[MAX_NUMNODES];
95 #ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
96 volatile unsigned long latent_entropy __latent_entropy;
97 EXPORT_SYMBOL(latent_entropy);
101 * Array of node states.
103 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
104 [N_POSSIBLE] = NODE_MASK_ALL,
105 [N_ONLINE] = { { [0] = 1UL } },
107 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
108 #ifdef CONFIG_HIGHMEM
109 [N_HIGH_MEMORY] = { { [0] = 1UL } },
111 #ifdef CONFIG_MOVABLE_NODE
112 [N_MEMORY] = { { [0] = 1UL } },
114 [N_CPU] = { { [0] = 1UL } },
117 EXPORT_SYMBOL(node_states);
119 /* Protect totalram_pages and zone->managed_pages */
120 static DEFINE_SPINLOCK(managed_page_count_lock);
122 unsigned long totalram_pages __read_mostly;
123 unsigned long totalreserve_pages __read_mostly;
124 unsigned long totalcma_pages __read_mostly;
126 int percpu_pagelist_fraction;
127 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
130 * A cached value of the page's pageblock's migratetype, used when the page is
131 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
132 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
133 * Also the migratetype set in the page does not necessarily match the pcplist
134 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
135 * other index - this ensures that it will be put on the correct CMA freelist.
137 static inline int get_pcppage_migratetype(struct page *page)
142 static inline void set_pcppage_migratetype(struct page *page, int migratetype)
144 page->index = migratetype;
147 #ifdef CONFIG_PM_SLEEP
149 * The following functions are used by the suspend/hibernate code to temporarily
150 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
151 * while devices are suspended. To avoid races with the suspend/hibernate code,
152 * they should always be called with pm_mutex held (gfp_allowed_mask also should
153 * only be modified with pm_mutex held, unless the suspend/hibernate code is
154 * guaranteed not to run in parallel with that modification).
157 static gfp_t saved_gfp_mask;
159 void pm_restore_gfp_mask(void)
161 WARN_ON(!mutex_is_locked(&pm_mutex));
162 if (saved_gfp_mask) {
163 gfp_allowed_mask = saved_gfp_mask;
168 void pm_restrict_gfp_mask(void)
170 WARN_ON(!mutex_is_locked(&pm_mutex));
171 WARN_ON(saved_gfp_mask);
172 saved_gfp_mask = gfp_allowed_mask;
173 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
176 bool pm_suspended_storage(void)
178 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
182 #endif /* CONFIG_PM_SLEEP */
184 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
185 unsigned int pageblock_order __read_mostly;
188 static void __free_pages_ok(struct page *page, unsigned int order);
191 * results with 256, 32 in the lowmem_reserve sysctl:
192 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
193 * 1G machine -> (16M dma, 784M normal, 224M high)
194 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
195 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
196 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
198 * TBD: should special case ZONE_DMA32 machines here - in those we normally
199 * don't need any ZONE_NORMAL reservation
201 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
202 #ifdef CONFIG_ZONE_DMA
205 #ifdef CONFIG_ZONE_DMA32
208 #ifdef CONFIG_HIGHMEM
214 EXPORT_SYMBOL(totalram_pages);
216 static char * const zone_names[MAX_NR_ZONES] = {
217 #ifdef CONFIG_ZONE_DMA
220 #ifdef CONFIG_ZONE_DMA32
224 #ifdef CONFIG_HIGHMEM
228 #ifdef CONFIG_ZONE_DEVICE
233 char * const migratetype_names[MIGRATE_TYPES] = {
241 #ifdef CONFIG_MEMORY_ISOLATION
246 compound_page_dtor * const compound_page_dtors[] = {
249 #ifdef CONFIG_HUGETLB_PAGE
252 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
257 int min_free_kbytes = 1024;
258 int user_min_free_kbytes = -1;
259 int watermark_scale_factor = 10;
261 static unsigned long __meminitdata nr_kernel_pages;
262 static unsigned long __meminitdata nr_all_pages;
263 static unsigned long __meminitdata dma_reserve;
265 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
266 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
267 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
268 static unsigned long __initdata required_kernelcore;
269 static unsigned long __initdata required_movablecore;
270 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
271 static bool mirrored_kernelcore;
273 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
275 EXPORT_SYMBOL(movable_zone);
276 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
279 int nr_node_ids __read_mostly = MAX_NUMNODES;
280 int nr_online_nodes __read_mostly = 1;
281 EXPORT_SYMBOL(nr_node_ids);
282 EXPORT_SYMBOL(nr_online_nodes);
285 int page_group_by_mobility_disabled __read_mostly;
287 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
288 static inline void reset_deferred_meminit(pg_data_t *pgdat)
290 pgdat->first_deferred_pfn = ULONG_MAX;
293 /* Returns true if the struct page for the pfn is uninitialised */
294 static inline bool __meminit early_page_uninitialised(unsigned long pfn)
296 int nid = early_pfn_to_nid(pfn);
298 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
305 * Returns false when the remaining initialisation should be deferred until
306 * later in the boot cycle when it can be parallelised.
308 static inline bool update_defer_init(pg_data_t *pgdat,
309 unsigned long pfn, unsigned long zone_end,
310 unsigned long *nr_initialised)
312 unsigned long max_initialise;
314 /* Always populate low zones for address-contrained allocations */
315 if (zone_end < pgdat_end_pfn(pgdat))
318 * Initialise at least 2G of a node but also take into account that
319 * two large system hashes that can take up 1GB for 0.25TB/node.
321 max_initialise = max(2UL << (30 - PAGE_SHIFT),
322 (pgdat->node_spanned_pages >> 8));
325 if ((*nr_initialised > max_initialise) &&
326 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
327 pgdat->first_deferred_pfn = pfn;
334 static inline void reset_deferred_meminit(pg_data_t *pgdat)
338 static inline bool early_page_uninitialised(unsigned long pfn)
343 static inline bool update_defer_init(pg_data_t *pgdat,
344 unsigned long pfn, unsigned long zone_end,
345 unsigned long *nr_initialised)
351 /* Return a pointer to the bitmap storing bits affecting a block of pages */
352 static inline unsigned long *get_pageblock_bitmap(struct page *page,
355 #ifdef CONFIG_SPARSEMEM
356 return __pfn_to_section(pfn)->pageblock_flags;
358 return page_zone(page)->pageblock_flags;
359 #endif /* CONFIG_SPARSEMEM */
362 static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
364 #ifdef CONFIG_SPARSEMEM
365 pfn &= (PAGES_PER_SECTION-1);
366 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
368 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
369 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
370 #endif /* CONFIG_SPARSEMEM */
374 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
375 * @page: The page within the block of interest
376 * @pfn: The target page frame number
377 * @end_bitidx: The last bit of interest to retrieve
378 * @mask: mask of bits that the caller is interested in
380 * Return: pageblock_bits flags
382 static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
384 unsigned long end_bitidx,
387 unsigned long *bitmap;
388 unsigned long bitidx, word_bitidx;
391 bitmap = get_pageblock_bitmap(page, pfn);
392 bitidx = pfn_to_bitidx(page, pfn);
393 word_bitidx = bitidx / BITS_PER_LONG;
394 bitidx &= (BITS_PER_LONG-1);
396 word = bitmap[word_bitidx];
397 bitidx += end_bitidx;
398 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
401 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
402 unsigned long end_bitidx,
405 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
408 static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
410 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
414 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
415 * @page: The page within the block of interest
416 * @flags: The flags to set
417 * @pfn: The target page frame number
418 * @end_bitidx: The last bit of interest
419 * @mask: mask of bits that the caller is interested in
421 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
423 unsigned long end_bitidx,
426 unsigned long *bitmap;
427 unsigned long bitidx, word_bitidx;
428 unsigned long old_word, word;
430 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
432 bitmap = get_pageblock_bitmap(page, pfn);
433 bitidx = pfn_to_bitidx(page, pfn);
434 word_bitidx = bitidx / BITS_PER_LONG;
435 bitidx &= (BITS_PER_LONG-1);
437 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
439 bitidx += end_bitidx;
440 mask <<= (BITS_PER_LONG - bitidx - 1);
441 flags <<= (BITS_PER_LONG - bitidx - 1);
443 word = READ_ONCE(bitmap[word_bitidx]);
445 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
446 if (word == old_word)
452 void set_pageblock_migratetype(struct page *page, int migratetype)
454 if (unlikely(page_group_by_mobility_disabled &&
455 migratetype < MIGRATE_PCPTYPES))
456 migratetype = MIGRATE_UNMOVABLE;
458 set_pageblock_flags_group(page, (unsigned long)migratetype,
459 PB_migrate, PB_migrate_end);
462 #ifdef CONFIG_DEBUG_VM
463 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
467 unsigned long pfn = page_to_pfn(page);
468 unsigned long sp, start_pfn;
471 seq = zone_span_seqbegin(zone);
472 start_pfn = zone->zone_start_pfn;
473 sp = zone->spanned_pages;
474 if (!zone_spans_pfn(zone, pfn))
476 } while (zone_span_seqretry(zone, seq));
479 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
480 pfn, zone_to_nid(zone), zone->name,
481 start_pfn, start_pfn + sp);
486 static int page_is_consistent(struct zone *zone, struct page *page)
488 if (!pfn_valid_within(page_to_pfn(page)))
490 if (zone != page_zone(page))
496 * Temporary debugging check for pages not lying within a given zone.
498 static int bad_range(struct zone *zone, struct page *page)
500 if (page_outside_zone_boundaries(zone, page))
502 if (!page_is_consistent(zone, page))
508 static inline int bad_range(struct zone *zone, struct page *page)
514 static void bad_page(struct page *page, const char *reason,
515 unsigned long bad_flags)
517 static unsigned long resume;
518 static unsigned long nr_shown;
519 static unsigned long nr_unshown;
522 * Allow a burst of 60 reports, then keep quiet for that minute;
523 * or allow a steady drip of one report per second.
525 if (nr_shown == 60) {
526 if (time_before(jiffies, resume)) {
532 "BUG: Bad page state: %lu messages suppressed\n",
539 resume = jiffies + 60 * HZ;
541 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
542 current->comm, page_to_pfn(page));
543 __dump_page(page, reason);
544 bad_flags &= page->flags;
546 pr_alert("bad because of flags: %#lx(%pGp)\n",
547 bad_flags, &bad_flags);
548 dump_page_owner(page);
553 /* Leave bad fields for debug, except PageBuddy could make trouble */
554 page_mapcount_reset(page); /* remove PageBuddy */
555 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
559 * Higher-order pages are called "compound pages". They are structured thusly:
561 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
563 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
564 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
566 * The first tail page's ->compound_dtor holds the offset in array of compound
567 * page destructors. See compound_page_dtors.
569 * The first tail page's ->compound_order holds the order of allocation.
570 * This usage means that zero-order pages may not be compound.
573 void free_compound_page(struct page *page)
575 __free_pages_ok(page, compound_order(page));
578 void prep_compound_page(struct page *page, unsigned int order)
581 int nr_pages = 1 << order;
583 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
584 set_compound_order(page, order);
586 for (i = 1; i < nr_pages; i++) {
587 struct page *p = page + i;
588 set_page_count(p, 0);
589 p->mapping = TAIL_MAPPING;
590 set_compound_head(p, page);
592 atomic_set(compound_mapcount_ptr(page), -1);
595 #ifdef CONFIG_DEBUG_PAGEALLOC
596 unsigned int _debug_guardpage_minorder;
597 bool _debug_pagealloc_enabled __read_mostly
598 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
599 EXPORT_SYMBOL(_debug_pagealloc_enabled);
600 bool _debug_guardpage_enabled __read_mostly;
602 static int __init early_debug_pagealloc(char *buf)
606 return kstrtobool(buf, &_debug_pagealloc_enabled);
608 early_param("debug_pagealloc", early_debug_pagealloc);
610 static bool need_debug_guardpage(void)
612 /* If we don't use debug_pagealloc, we don't need guard page */
613 if (!debug_pagealloc_enabled())
616 if (!debug_guardpage_minorder())
622 static void init_debug_guardpage(void)
624 if (!debug_pagealloc_enabled())
627 if (!debug_guardpage_minorder())
630 _debug_guardpage_enabled = true;
633 struct page_ext_operations debug_guardpage_ops = {
634 .need = need_debug_guardpage,
635 .init = init_debug_guardpage,
638 static int __init debug_guardpage_minorder_setup(char *buf)
642 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
643 pr_err("Bad debug_guardpage_minorder value\n");
646 _debug_guardpage_minorder = res;
647 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
650 early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
652 static inline bool set_page_guard(struct zone *zone, struct page *page,
653 unsigned int order, int migratetype)
655 struct page_ext *page_ext;
657 if (!debug_guardpage_enabled())
660 if (order >= debug_guardpage_minorder())
663 page_ext = lookup_page_ext(page);
664 if (unlikely(!page_ext))
667 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
669 INIT_LIST_HEAD(&page->lru);
670 set_page_private(page, order);
671 /* Guard pages are not available for any usage */
672 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
677 static inline void clear_page_guard(struct zone *zone, struct page *page,
678 unsigned int order, int migratetype)
680 struct page_ext *page_ext;
682 if (!debug_guardpage_enabled())
685 page_ext = lookup_page_ext(page);
686 if (unlikely(!page_ext))
689 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
691 set_page_private(page, 0);
692 if (!is_migrate_isolate(migratetype))
693 __mod_zone_freepage_state(zone, (1 << order), migratetype);
696 struct page_ext_operations debug_guardpage_ops;
697 static inline bool set_page_guard(struct zone *zone, struct page *page,
698 unsigned int order, int migratetype) { return false; }
699 static inline void clear_page_guard(struct zone *zone, struct page *page,
700 unsigned int order, int migratetype) {}
703 static inline void set_page_order(struct page *page, unsigned int order)
705 set_page_private(page, order);
706 __SetPageBuddy(page);
709 static inline void rmv_page_order(struct page *page)
711 __ClearPageBuddy(page);
712 set_page_private(page, 0);
716 * This function checks whether a page is free && is the buddy
717 * we can do coalesce a page and its buddy if
718 * (a) the buddy is not in a hole (check before calling!) &&
719 * (b) the buddy is in the buddy system &&
720 * (c) a page and its buddy have the same order &&
721 * (d) a page and its buddy are in the same zone.
723 * For recording whether a page is in the buddy system, we set ->_mapcount
724 * PAGE_BUDDY_MAPCOUNT_VALUE.
725 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
726 * serialized by zone->lock.
728 * For recording page's order, we use page_private(page).
730 static inline int page_is_buddy(struct page *page, struct page *buddy,
733 if (page_is_guard(buddy) && page_order(buddy) == order) {
734 if (page_zone_id(page) != page_zone_id(buddy))
737 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
742 if (PageBuddy(buddy) && page_order(buddy) == order) {
744 * zone check is done late to avoid uselessly
745 * calculating zone/node ids for pages that could
748 if (page_zone_id(page) != page_zone_id(buddy))
751 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
759 * Freeing function for a buddy system allocator.
761 * The concept of a buddy system is to maintain direct-mapped table
762 * (containing bit values) for memory blocks of various "orders".
763 * The bottom level table contains the map for the smallest allocatable
764 * units of memory (here, pages), and each level above it describes
765 * pairs of units from the levels below, hence, "buddies".
766 * At a high level, all that happens here is marking the table entry
767 * at the bottom level available, and propagating the changes upward
768 * as necessary, plus some accounting needed to play nicely with other
769 * parts of the VM system.
770 * At each level, we keep a list of pages, which are heads of continuous
771 * free pages of length of (1 << order) and marked with _mapcount
772 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
774 * So when we are allocating or freeing one, we can derive the state of the
775 * other. That is, if we allocate a small block, and both were
776 * free, the remainder of the region must be split into blocks.
777 * If a block is freed, and its buddy is also free, then this
778 * triggers coalescing into a block of larger size.
783 static inline void __free_one_page(struct page *page,
785 struct zone *zone, unsigned int order,
788 unsigned long combined_pfn;
789 unsigned long uninitialized_var(buddy_pfn);
791 unsigned int max_order;
793 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
795 VM_BUG_ON(!zone_is_initialized(zone));
796 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
798 VM_BUG_ON(migratetype == -1);
799 if (likely(!is_migrate_isolate(migratetype)))
800 __mod_zone_freepage_state(zone, 1 << order, migratetype);
802 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
803 VM_BUG_ON_PAGE(bad_range(zone, page), page);
806 while (order < max_order - 1) {
807 buddy_pfn = __find_buddy_pfn(pfn, order);
808 buddy = page + (buddy_pfn - pfn);
810 if (!pfn_valid_within(buddy_pfn))
812 if (!page_is_buddy(page, buddy, order))
815 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
816 * merge with it and move up one order.
818 if (page_is_guard(buddy)) {
819 clear_page_guard(zone, buddy, order, migratetype);
821 list_del(&buddy->lru);
822 zone->free_area[order].nr_free--;
823 rmv_page_order(buddy);
825 combined_pfn = buddy_pfn & pfn;
826 page = page + (combined_pfn - pfn);
830 if (max_order < MAX_ORDER) {
831 /* If we are here, it means order is >= pageblock_order.
832 * We want to prevent merge between freepages on isolate
833 * pageblock and normal pageblock. Without this, pageblock
834 * isolation could cause incorrect freepage or CMA accounting.
836 * We don't want to hit this code for the more frequent
839 if (unlikely(has_isolate_pageblock(zone))) {
842 buddy_pfn = __find_buddy_pfn(pfn, order);
843 buddy = page + (buddy_pfn - pfn);
844 buddy_mt = get_pageblock_migratetype(buddy);
846 if (migratetype != buddy_mt
847 && (is_migrate_isolate(migratetype) ||
848 is_migrate_isolate(buddy_mt)))
852 goto continue_merging;
856 set_page_order(page, order);
859 * If this is not the largest possible page, check if the buddy
860 * of the next-highest order is free. If it is, it's possible
861 * that pages are being freed that will coalesce soon. In case,
862 * that is happening, add the free page to the tail of the list
863 * so it's less likely to be used soon and more likely to be merged
864 * as a higher order page
866 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
867 struct page *higher_page, *higher_buddy;
868 combined_pfn = buddy_pfn & pfn;
869 higher_page = page + (combined_pfn - pfn);
870 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
871 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
872 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
873 list_add_tail(&page->lru,
874 &zone->free_area[order].free_list[migratetype]);
879 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
881 zone->free_area[order].nr_free++;
885 * A bad page could be due to a number of fields. Instead of multiple branches,
886 * try and check multiple fields with one check. The caller must do a detailed
887 * check if necessary.
889 static inline bool page_expected_state(struct page *page,
890 unsigned long check_flags)
892 if (unlikely(atomic_read(&page->_mapcount) != -1))
895 if (unlikely((unsigned long)page->mapping |
896 page_ref_count(page) |
898 (unsigned long)page->mem_cgroup |
900 (page->flags & check_flags)))
906 static void free_pages_check_bad(struct page *page)
908 const char *bad_reason;
909 unsigned long bad_flags;
914 if (unlikely(atomic_read(&page->_mapcount) != -1))
915 bad_reason = "nonzero mapcount";
916 if (unlikely(page->mapping != NULL))
917 bad_reason = "non-NULL mapping";
918 if (unlikely(page_ref_count(page) != 0))
919 bad_reason = "nonzero _refcount";
920 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
921 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
922 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
925 if (unlikely(page->mem_cgroup))
926 bad_reason = "page still charged to cgroup";
928 bad_page(page, bad_reason, bad_flags);
931 static inline int free_pages_check(struct page *page)
933 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
936 /* Something has gone sideways, find it */
937 free_pages_check_bad(page);
941 static int free_tail_pages_check(struct page *head_page, struct page *page)
946 * We rely page->lru.next never has bit 0 set, unless the page
947 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
949 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
951 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
955 switch (page - head_page) {
957 /* the first tail page: ->mapping is compound_mapcount() */
958 if (unlikely(compound_mapcount(page))) {
959 bad_page(page, "nonzero compound_mapcount", 0);
965 * the second tail page: ->mapping is
966 * page_deferred_list().next -- ignore value.
970 if (page->mapping != TAIL_MAPPING) {
971 bad_page(page, "corrupted mapping in tail page", 0);
976 if (unlikely(!PageTail(page))) {
977 bad_page(page, "PageTail not set", 0);
980 if (unlikely(compound_head(page) != head_page)) {
981 bad_page(page, "compound_head not consistent", 0);
986 page->mapping = NULL;
987 clear_compound_head(page);
991 static __always_inline bool free_pages_prepare(struct page *page,
992 unsigned int order, bool check_free)
996 VM_BUG_ON_PAGE(PageTail(page), page);
998 trace_mm_page_free(page, order);
999 kmemcheck_free_shadow(page, order);
1002 * Check tail pages before head page information is cleared to
1003 * avoid checking PageCompound for order-0 pages.
1005 if (unlikely(order)) {
1006 bool compound = PageCompound(page);
1009 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1012 ClearPageDoubleMap(page);
1013 for (i = 1; i < (1 << order); i++) {
1015 bad += free_tail_pages_check(page, page + i);
1016 if (unlikely(free_pages_check(page + i))) {
1020 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1023 if (PageMappingFlags(page))
1024 page->mapping = NULL;
1025 if (memcg_kmem_enabled() && PageKmemcg(page))
1026 memcg_kmem_uncharge(page, order);
1028 bad += free_pages_check(page);
1032 page_cpupid_reset_last(page);
1033 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1034 reset_page_owner(page, order);
1036 if (!PageHighMem(page)) {
1037 debug_check_no_locks_freed(page_address(page),
1038 PAGE_SIZE << order);
1039 debug_check_no_obj_freed(page_address(page),
1040 PAGE_SIZE << order);
1042 arch_free_page(page, order);
1043 kernel_poison_pages(page, 1 << order, 0);
1044 kernel_map_pages(page, 1 << order, 0);
1045 kasan_free_pages(page, order);
1050 #ifdef CONFIG_DEBUG_VM
1051 static inline bool free_pcp_prepare(struct page *page)
1053 return free_pages_prepare(page, 0, true);
1056 static inline bool bulkfree_pcp_prepare(struct page *page)
1061 static bool free_pcp_prepare(struct page *page)
1063 return free_pages_prepare(page, 0, false);
1066 static bool bulkfree_pcp_prepare(struct page *page)
1068 return free_pages_check(page);
1070 #endif /* CONFIG_DEBUG_VM */
1073 * Frees a number of pages from the PCP lists
1074 * Assumes all pages on list are in same zone, and of same order.
1075 * count is the number of pages to free.
1077 * If the zone was previously in an "all pages pinned" state then look to
1078 * see if this freeing clears that state.
1080 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1081 * pinned" detection logic.
1083 static void free_pcppages_bulk(struct zone *zone, int count,
1084 struct per_cpu_pages *pcp)
1086 int migratetype = 0;
1088 unsigned long nr_scanned;
1089 bool isolated_pageblocks;
1091 spin_lock(&zone->lock);
1092 isolated_pageblocks = has_isolate_pageblock(zone);
1093 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1095 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1099 struct list_head *list;
1102 * Remove pages from lists in a round-robin fashion. A
1103 * batch_free count is maintained that is incremented when an
1104 * empty list is encountered. This is so more pages are freed
1105 * off fuller lists instead of spinning excessively around empty
1110 if (++migratetype == MIGRATE_PCPTYPES)
1112 list = &pcp->lists[migratetype];
1113 } while (list_empty(list));
1115 /* This is the only non-empty list. Free them all. */
1116 if (batch_free == MIGRATE_PCPTYPES)
1120 int mt; /* migratetype of the to-be-freed page */
1122 page = list_last_entry(list, struct page, lru);
1123 /* must delete as __free_one_page list manipulates */
1124 list_del(&page->lru);
1126 mt = get_pcppage_migratetype(page);
1127 /* MIGRATE_ISOLATE page should not go to pcplists */
1128 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1129 /* Pageblock could have been isolated meanwhile */
1130 if (unlikely(isolated_pageblocks))
1131 mt = get_pageblock_migratetype(page);
1133 if (bulkfree_pcp_prepare(page))
1136 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1137 trace_mm_page_pcpu_drain(page, 0, mt);
1138 } while (--count && --batch_free && !list_empty(list));
1140 spin_unlock(&zone->lock);
1143 static void free_one_page(struct zone *zone,
1144 struct page *page, unsigned long pfn,
1148 unsigned long nr_scanned;
1149 spin_lock(&zone->lock);
1150 nr_scanned = node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED);
1152 __mod_node_page_state(zone->zone_pgdat, NR_PAGES_SCANNED, -nr_scanned);
1154 if (unlikely(has_isolate_pageblock(zone) ||
1155 is_migrate_isolate(migratetype))) {
1156 migratetype = get_pfnblock_migratetype(page, pfn);
1158 __free_one_page(page, pfn, zone, order, migratetype);
1159 spin_unlock(&zone->lock);
1162 static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1163 unsigned long zone, int nid)
1165 set_page_links(page, zone, nid, pfn);
1166 init_page_count(page);
1167 page_mapcount_reset(page);
1168 page_cpupid_reset_last(page);
1170 INIT_LIST_HEAD(&page->lru);
1171 #ifdef WANT_PAGE_VIRTUAL
1172 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1173 if (!is_highmem_idx(zone))
1174 set_page_address(page, __va(pfn << PAGE_SHIFT));
1178 static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
1181 return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
1184 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1185 static void init_reserved_page(unsigned long pfn)
1190 if (!early_page_uninitialised(pfn))
1193 nid = early_pfn_to_nid(pfn);
1194 pgdat = NODE_DATA(nid);
1196 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1197 struct zone *zone = &pgdat->node_zones[zid];
1199 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1202 __init_single_pfn(pfn, zid, nid);
1205 static inline void init_reserved_page(unsigned long pfn)
1208 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1211 * Initialised pages do not have PageReserved set. This function is
1212 * called for each range allocated by the bootmem allocator and
1213 * marks the pages PageReserved. The remaining valid pages are later
1214 * sent to the buddy page allocator.
1216 void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1218 unsigned long start_pfn = PFN_DOWN(start);
1219 unsigned long end_pfn = PFN_UP(end);
1221 for (; start_pfn < end_pfn; start_pfn++) {
1222 if (pfn_valid(start_pfn)) {
1223 struct page *page = pfn_to_page(start_pfn);
1225 init_reserved_page(start_pfn);
1227 /* Avoid false-positive PageTail() */
1228 INIT_LIST_HEAD(&page->lru);
1230 SetPageReserved(page);
1235 static void __free_pages_ok(struct page *page, unsigned int order)
1237 unsigned long flags;
1239 unsigned long pfn = page_to_pfn(page);
1241 if (!free_pages_prepare(page, order, true))
1244 migratetype = get_pfnblock_migratetype(page, pfn);
1245 local_irq_save(flags);
1246 __count_vm_events(PGFREE, 1 << order);
1247 free_one_page(page_zone(page), page, pfn, order, migratetype);
1248 local_irq_restore(flags);
1251 static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1253 unsigned int nr_pages = 1 << order;
1254 struct page *p = page;
1258 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1260 __ClearPageReserved(p);
1261 set_page_count(p, 0);
1263 __ClearPageReserved(p);
1264 set_page_count(p, 0);
1266 page_zone(page)->managed_pages += nr_pages;
1267 set_page_refcounted(page);
1268 __free_pages(page, order);
1271 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1272 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1274 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1276 int __meminit early_pfn_to_nid(unsigned long pfn)
1278 static DEFINE_SPINLOCK(early_pfn_lock);
1281 spin_lock(&early_pfn_lock);
1282 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1284 nid = first_online_node;
1285 spin_unlock(&early_pfn_lock);
1291 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1292 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1293 struct mminit_pfnnid_cache *state)
1297 nid = __early_pfn_to_nid(pfn, state);
1298 if (nid >= 0 && nid != node)
1303 /* Only safe to use early in boot when initialisation is single-threaded */
1304 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1306 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1311 static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1315 static inline bool __meminit meminit_pfn_in_nid(unsigned long pfn, int node,
1316 struct mminit_pfnnid_cache *state)
1323 void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1326 if (early_page_uninitialised(pfn))
1328 return __free_pages_boot_core(page, order);
1332 * Check that the whole (or subset of) a pageblock given by the interval of
1333 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1334 * with the migration of free compaction scanner. The scanners then need to
1335 * use only pfn_valid_within() check for arches that allow holes within
1338 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1340 * It's possible on some configurations to have a setup like node0 node1 node0
1341 * i.e. it's possible that all pages within a zones range of pages do not
1342 * belong to a single zone. We assume that a border between node0 and node1
1343 * can occur within a single pageblock, but not a node0 node1 node0
1344 * interleaving within a single pageblock. It is therefore sufficient to check
1345 * the first and last page of a pageblock and avoid checking each individual
1346 * page in a pageblock.
1348 struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1349 unsigned long end_pfn, struct zone *zone)
1351 struct page *start_page;
1352 struct page *end_page;
1354 /* end_pfn is one past the range we are checking */
1357 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1360 start_page = pfn_to_page(start_pfn);
1362 if (page_zone(start_page) != zone)
1365 end_page = pfn_to_page(end_pfn);
1367 /* This gives a shorter code than deriving page_zone(end_page) */
1368 if (page_zone_id(start_page) != page_zone_id(end_page))
1374 void set_zone_contiguous(struct zone *zone)
1376 unsigned long block_start_pfn = zone->zone_start_pfn;
1377 unsigned long block_end_pfn;
1379 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1380 for (; block_start_pfn < zone_end_pfn(zone);
1381 block_start_pfn = block_end_pfn,
1382 block_end_pfn += pageblock_nr_pages) {
1384 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1386 if (!__pageblock_pfn_to_page(block_start_pfn,
1387 block_end_pfn, zone))
1391 /* We confirm that there is no hole */
1392 zone->contiguous = true;
1395 void clear_zone_contiguous(struct zone *zone)
1397 zone->contiguous = false;
1400 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1401 static void __init deferred_free_range(struct page *page,
1402 unsigned long pfn, int nr_pages)
1409 /* Free a large naturally-aligned chunk if possible */
1410 if (nr_pages == pageblock_nr_pages &&
1411 (pfn & (pageblock_nr_pages - 1)) == 0) {
1412 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1413 __free_pages_boot_core(page, pageblock_order);
1417 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1418 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1419 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1420 __free_pages_boot_core(page, 0);
1424 /* Completion tracking for deferred_init_memmap() threads */
1425 static atomic_t pgdat_init_n_undone __initdata;
1426 static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1428 static inline void __init pgdat_init_report_one_done(void)
1430 if (atomic_dec_and_test(&pgdat_init_n_undone))
1431 complete(&pgdat_init_all_done_comp);
1434 /* Initialise remaining memory on a node */
1435 static int __init deferred_init_memmap(void *data)
1437 pg_data_t *pgdat = data;
1438 int nid = pgdat->node_id;
1439 struct mminit_pfnnid_cache nid_init_state = { };
1440 unsigned long start = jiffies;
1441 unsigned long nr_pages = 0;
1442 unsigned long walk_start, walk_end;
1445 unsigned long first_init_pfn = pgdat->first_deferred_pfn;
1446 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1448 if (first_init_pfn == ULONG_MAX) {
1449 pgdat_init_report_one_done();
1453 /* Bind memory initialisation thread to a local node if possible */
1454 if (!cpumask_empty(cpumask))
1455 set_cpus_allowed_ptr(current, cpumask);
1457 /* Sanity check boundaries */
1458 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1459 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1460 pgdat->first_deferred_pfn = ULONG_MAX;
1462 /* Only the highest zone is deferred so find it */
1463 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1464 zone = pgdat->node_zones + zid;
1465 if (first_init_pfn < zone_end_pfn(zone))
1469 for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
1470 unsigned long pfn, end_pfn;
1471 struct page *page = NULL;
1472 struct page *free_base_page = NULL;
1473 unsigned long free_base_pfn = 0;
1476 end_pfn = min(walk_end, zone_end_pfn(zone));
1477 pfn = first_init_pfn;
1478 if (pfn < walk_start)
1480 if (pfn < zone->zone_start_pfn)
1481 pfn = zone->zone_start_pfn;
1483 for (; pfn < end_pfn; pfn++) {
1484 if (!pfn_valid_within(pfn))
1488 * Ensure pfn_valid is checked every
1489 * pageblock_nr_pages for memory holes
1491 if ((pfn & (pageblock_nr_pages - 1)) == 0) {
1492 if (!pfn_valid(pfn)) {
1498 if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
1503 /* Minimise pfn page lookups and scheduler checks */
1504 if (page && (pfn & (pageblock_nr_pages - 1)) != 0) {
1507 nr_pages += nr_to_free;
1508 deferred_free_range(free_base_page,
1509 free_base_pfn, nr_to_free);
1510 free_base_page = NULL;
1511 free_base_pfn = nr_to_free = 0;
1513 page = pfn_to_page(pfn);
1518 VM_BUG_ON(page_zone(page) != zone);
1522 __init_single_page(page, pfn, zid, nid);
1523 if (!free_base_page) {
1524 free_base_page = page;
1525 free_base_pfn = pfn;
1530 /* Where possible, batch up pages for a single free */
1533 /* Free the current block of pages to allocator */
1534 nr_pages += nr_to_free;
1535 deferred_free_range(free_base_page, free_base_pfn,
1537 free_base_page = NULL;
1538 free_base_pfn = nr_to_free = 0;
1540 /* Free the last block of pages to allocator */
1541 nr_pages += nr_to_free;
1542 deferred_free_range(free_base_page, free_base_pfn, nr_to_free);
1544 first_init_pfn = max(end_pfn, first_init_pfn);
1547 /* Sanity check that the next zone really is unpopulated */
1548 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1550 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1551 jiffies_to_msecs(jiffies - start));
1553 pgdat_init_report_one_done();
1556 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1558 void __init page_alloc_init_late(void)
1562 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1565 /* There will be num_node_state(N_MEMORY) threads */
1566 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1567 for_each_node_state(nid, N_MEMORY) {
1568 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1571 /* Block until all are initialised */
1572 wait_for_completion(&pgdat_init_all_done_comp);
1574 /* Reinit limits that are based on free pages after the kernel is up */
1575 files_maxfiles_init();
1578 for_each_populated_zone(zone)
1579 set_zone_contiguous(zone);
1583 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1584 void __init init_cma_reserved_pageblock(struct page *page)
1586 unsigned i = pageblock_nr_pages;
1587 struct page *p = page;
1590 __ClearPageReserved(p);
1591 set_page_count(p, 0);
1594 set_pageblock_migratetype(page, MIGRATE_CMA);
1596 if (pageblock_order >= MAX_ORDER) {
1597 i = pageblock_nr_pages;
1600 set_page_refcounted(p);
1601 __free_pages(p, MAX_ORDER - 1);
1602 p += MAX_ORDER_NR_PAGES;
1603 } while (i -= MAX_ORDER_NR_PAGES);
1605 set_page_refcounted(page);
1606 __free_pages(page, pageblock_order);
1609 adjust_managed_page_count(page, pageblock_nr_pages);
1614 * The order of subdivision here is critical for the IO subsystem.
1615 * Please do not alter this order without good reasons and regression
1616 * testing. Specifically, as large blocks of memory are subdivided,
1617 * the order in which smaller blocks are delivered depends on the order
1618 * they're subdivided in this function. This is the primary factor
1619 * influencing the order in which pages are delivered to the IO
1620 * subsystem according to empirical testing, and this is also justified
1621 * by considering the behavior of a buddy system containing a single
1622 * large block of memory acted on by a series of small allocations.
1623 * This behavior is a critical factor in sglist merging's success.
1627 static inline void expand(struct zone *zone, struct page *page,
1628 int low, int high, struct free_area *area,
1631 unsigned long size = 1 << high;
1633 while (high > low) {
1637 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1640 * Mark as guard pages (or page), that will allow to
1641 * merge back to allocator when buddy will be freed.
1642 * Corresponding page table entries will not be touched,
1643 * pages will stay not present in virtual address space
1645 if (set_page_guard(zone, &page[size], high, migratetype))
1648 list_add(&page[size].lru, &area->free_list[migratetype]);
1650 set_page_order(&page[size], high);
1654 static void check_new_page_bad(struct page *page)
1656 const char *bad_reason = NULL;
1657 unsigned long bad_flags = 0;
1659 if (unlikely(atomic_read(&page->_mapcount) != -1))
1660 bad_reason = "nonzero mapcount";
1661 if (unlikely(page->mapping != NULL))
1662 bad_reason = "non-NULL mapping";
1663 if (unlikely(page_ref_count(page) != 0))
1664 bad_reason = "nonzero _count";
1665 if (unlikely(page->flags & __PG_HWPOISON)) {
1666 bad_reason = "HWPoisoned (hardware-corrupted)";
1667 bad_flags = __PG_HWPOISON;
1668 /* Don't complain about hwpoisoned pages */
1669 page_mapcount_reset(page); /* remove PageBuddy */
1672 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1673 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1674 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1677 if (unlikely(page->mem_cgroup))
1678 bad_reason = "page still charged to cgroup";
1680 bad_page(page, bad_reason, bad_flags);
1684 * This page is about to be returned from the page allocator
1686 static inline int check_new_page(struct page *page)
1688 if (likely(page_expected_state(page,
1689 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1692 check_new_page_bad(page);
1696 static inline bool free_pages_prezeroed(bool poisoned)
1698 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1699 page_poisoning_enabled() && poisoned;
1702 #ifdef CONFIG_DEBUG_VM
1703 static bool check_pcp_refill(struct page *page)
1708 static bool check_new_pcp(struct page *page)
1710 return check_new_page(page);
1713 static bool check_pcp_refill(struct page *page)
1715 return check_new_page(page);
1717 static bool check_new_pcp(struct page *page)
1721 #endif /* CONFIG_DEBUG_VM */
1723 static bool check_new_pages(struct page *page, unsigned int order)
1726 for (i = 0; i < (1 << order); i++) {
1727 struct page *p = page + i;
1729 if (unlikely(check_new_page(p)))
1736 inline void post_alloc_hook(struct page *page, unsigned int order,
1739 set_page_private(page, 0);
1740 set_page_refcounted(page);
1742 arch_alloc_page(page, order);
1743 kernel_map_pages(page, 1 << order, 1);
1744 kernel_poison_pages(page, 1 << order, 1);
1745 kasan_alloc_pages(page, order);
1746 set_page_owner(page, order, gfp_flags);
1749 static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1750 unsigned int alloc_flags)
1753 bool poisoned = true;
1755 for (i = 0; i < (1 << order); i++) {
1756 struct page *p = page + i;
1758 poisoned &= page_is_poisoned(p);
1761 post_alloc_hook(page, order, gfp_flags);
1763 if (!free_pages_prezeroed(poisoned) && (gfp_flags & __GFP_ZERO))
1764 for (i = 0; i < (1 << order); i++)
1765 clear_highpage(page + i);
1767 if (order && (gfp_flags & __GFP_COMP))
1768 prep_compound_page(page, order);
1771 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1772 * allocate the page. The expectation is that the caller is taking
1773 * steps that will free more memory. The caller should avoid the page
1774 * being used for !PFMEMALLOC purposes.
1776 if (alloc_flags & ALLOC_NO_WATERMARKS)
1777 set_page_pfmemalloc(page);
1779 clear_page_pfmemalloc(page);
1783 * Go through the free lists for the given migratetype and remove
1784 * the smallest available page from the freelists
1787 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1790 unsigned int current_order;
1791 struct free_area *area;
1794 /* Find a page of the appropriate size in the preferred list */
1795 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1796 area = &(zone->free_area[current_order]);
1797 page = list_first_entry_or_null(&area->free_list[migratetype],
1801 list_del(&page->lru);
1802 rmv_page_order(page);
1804 expand(zone, page, order, current_order, area, migratetype);
1805 set_pcppage_migratetype(page, migratetype);
1814 * This array describes the order lists are fallen back to when
1815 * the free lists for the desirable migrate type are depleted
1817 static int fallbacks[MIGRATE_TYPES][4] = {
1818 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1819 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1820 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1822 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1824 #ifdef CONFIG_MEMORY_ISOLATION
1825 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1830 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1833 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1836 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1837 unsigned int order) { return NULL; }
1841 * Move the free pages in a range to the free lists of the requested type.
1842 * Note that start_page and end_pages are not aligned on a pageblock
1843 * boundary. If alignment is required, use move_freepages_block()
1845 int move_freepages(struct zone *zone,
1846 struct page *start_page, struct page *end_page,
1851 int pages_moved = 0;
1853 #ifndef CONFIG_HOLES_IN_ZONE
1855 * page_zone is not safe to call in this context when
1856 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1857 * anyway as we check zone boundaries in move_freepages_block().
1858 * Remove at a later date when no bug reports exist related to
1859 * grouping pages by mobility
1861 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1864 for (page = start_page; page <= end_page;) {
1865 if (!pfn_valid_within(page_to_pfn(page))) {
1870 /* Make sure we are not inadvertently changing nodes */
1871 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1873 if (!PageBuddy(page)) {
1878 order = page_order(page);
1879 list_move(&page->lru,
1880 &zone->free_area[order].free_list[migratetype]);
1882 pages_moved += 1 << order;
1888 int move_freepages_block(struct zone *zone, struct page *page,
1891 unsigned long start_pfn, end_pfn;
1892 struct page *start_page, *end_page;
1894 start_pfn = page_to_pfn(page);
1895 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1896 start_page = pfn_to_page(start_pfn);
1897 end_page = start_page + pageblock_nr_pages - 1;
1898 end_pfn = start_pfn + pageblock_nr_pages - 1;
1900 /* Do not cross zone boundaries */
1901 if (!zone_spans_pfn(zone, start_pfn))
1903 if (!zone_spans_pfn(zone, end_pfn))
1906 return move_freepages(zone, start_page, end_page, migratetype);
1909 static void change_pageblock_range(struct page *pageblock_page,
1910 int start_order, int migratetype)
1912 int nr_pageblocks = 1 << (start_order - pageblock_order);
1914 while (nr_pageblocks--) {
1915 set_pageblock_migratetype(pageblock_page, migratetype);
1916 pageblock_page += pageblock_nr_pages;
1921 * When we are falling back to another migratetype during allocation, try to
1922 * steal extra free pages from the same pageblocks to satisfy further
1923 * allocations, instead of polluting multiple pageblocks.
1925 * If we are stealing a relatively large buddy page, it is likely there will
1926 * be more free pages in the pageblock, so try to steal them all. For
1927 * reclaimable and unmovable allocations, we steal regardless of page size,
1928 * as fragmentation caused by those allocations polluting movable pageblocks
1929 * is worse than movable allocations stealing from unmovable and reclaimable
1932 static bool can_steal_fallback(unsigned int order, int start_mt)
1935 * Leaving this order check is intended, although there is
1936 * relaxed order check in next check. The reason is that
1937 * we can actually steal whole pageblock if this condition met,
1938 * but, below check doesn't guarantee it and that is just heuristic
1939 * so could be changed anytime.
1941 if (order >= pageblock_order)
1944 if (order >= pageblock_order / 2 ||
1945 start_mt == MIGRATE_RECLAIMABLE ||
1946 start_mt == MIGRATE_UNMOVABLE ||
1947 page_group_by_mobility_disabled)
1954 * This function implements actual steal behaviour. If order is large enough,
1955 * we can steal whole pageblock. If not, we first move freepages in this
1956 * pageblock and check whether half of pages are moved or not. If half of
1957 * pages are moved, we can change migratetype of pageblock and permanently
1958 * use it's pages as requested migratetype in the future.
1960 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1963 unsigned int current_order = page_order(page);
1966 /* Take ownership for orders >= pageblock_order */
1967 if (current_order >= pageblock_order) {
1968 change_pageblock_range(page, current_order, start_type);
1972 pages = move_freepages_block(zone, page, start_type);
1974 /* Claim the whole block if over half of it is free */
1975 if (pages >= (1 << (pageblock_order-1)) ||
1976 page_group_by_mobility_disabled)
1977 set_pageblock_migratetype(page, start_type);
1981 * Check whether there is a suitable fallback freepage with requested order.
1982 * If only_stealable is true, this function returns fallback_mt only if
1983 * we can steal other freepages all together. This would help to reduce
1984 * fragmentation due to mixed migratetype pages in one pageblock.
1986 int find_suitable_fallback(struct free_area *area, unsigned int order,
1987 int migratetype, bool only_stealable, bool *can_steal)
1992 if (area->nr_free == 0)
1997 fallback_mt = fallbacks[migratetype][i];
1998 if (fallback_mt == MIGRATE_TYPES)
2001 if (list_empty(&area->free_list[fallback_mt]))
2004 if (can_steal_fallback(order, migratetype))
2007 if (!only_stealable)
2018 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2019 * there are no empty page blocks that contain a page with a suitable order
2021 static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2022 unsigned int alloc_order)
2025 unsigned long max_managed, flags;
2028 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2029 * Check is race-prone but harmless.
2031 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2032 if (zone->nr_reserved_highatomic >= max_managed)
2035 spin_lock_irqsave(&zone->lock, flags);
2037 /* Recheck the nr_reserved_highatomic limit under the lock */
2038 if (zone->nr_reserved_highatomic >= max_managed)
2042 mt = get_pageblock_migratetype(page);
2043 if (mt != MIGRATE_HIGHATOMIC &&
2044 !is_migrate_isolate(mt) && !is_migrate_cma(mt)) {
2045 zone->nr_reserved_highatomic += pageblock_nr_pages;
2046 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2047 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC);
2051 spin_unlock_irqrestore(&zone->lock, flags);
2055 * Used when an allocation is about to fail under memory pressure. This
2056 * potentially hurts the reliability of high-order allocations when under
2057 * intense memory pressure but failed atomic allocations should be easier
2058 * to recover from than an OOM.
2060 * If @force is true, try to unreserve a pageblock even though highatomic
2061 * pageblock is exhausted.
2063 static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2066 struct zonelist *zonelist = ac->zonelist;
2067 unsigned long flags;
2074 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2077 * Preserve at least one pageblock unless memory pressure
2080 if (!force && zone->nr_reserved_highatomic <=
2084 spin_lock_irqsave(&zone->lock, flags);
2085 for (order = 0; order < MAX_ORDER; order++) {
2086 struct free_area *area = &(zone->free_area[order]);
2088 page = list_first_entry_or_null(
2089 &area->free_list[MIGRATE_HIGHATOMIC],
2095 * In page freeing path, migratetype change is racy so
2096 * we can counter several free pages in a pageblock
2097 * in this loop althoug we changed the pageblock type
2098 * from highatomic to ac->migratetype. So we should
2099 * adjust the count once.
2101 if (get_pageblock_migratetype(page) ==
2102 MIGRATE_HIGHATOMIC) {
2104 * It should never happen but changes to
2105 * locking could inadvertently allow a per-cpu
2106 * drain to add pages to MIGRATE_HIGHATOMIC
2107 * while unreserving so be safe and watch for
2110 zone->nr_reserved_highatomic -= min(
2112 zone->nr_reserved_highatomic);
2116 * Convert to ac->migratetype and avoid the normal
2117 * pageblock stealing heuristics. Minimally, the caller
2118 * is doing the work and needs the pages. More
2119 * importantly, if the block was always converted to
2120 * MIGRATE_UNMOVABLE or another type then the number
2121 * of pageblocks that cannot be completely freed
2124 set_pageblock_migratetype(page, ac->migratetype);
2125 ret = move_freepages_block(zone, page, ac->migratetype);
2127 spin_unlock_irqrestore(&zone->lock, flags);
2131 spin_unlock_irqrestore(&zone->lock, flags);
2137 /* Remove an element from the buddy allocator from the fallback list */
2138 static inline struct page *
2139 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
2141 struct free_area *area;
2142 unsigned int current_order;
2147 /* Find the largest possible block of pages in the other list */
2148 for (current_order = MAX_ORDER-1;
2149 current_order >= order && current_order <= MAX_ORDER-1;
2151 area = &(zone->free_area[current_order]);
2152 fallback_mt = find_suitable_fallback(area, current_order,
2153 start_migratetype, false, &can_steal);
2154 if (fallback_mt == -1)
2157 page = list_first_entry(&area->free_list[fallback_mt],
2160 get_pageblock_migratetype(page) != MIGRATE_HIGHATOMIC)
2161 steal_suitable_fallback(zone, page, start_migratetype);
2163 /* Remove the page from the freelists */
2165 list_del(&page->lru);
2166 rmv_page_order(page);
2168 expand(zone, page, order, current_order, area,
2171 * The pcppage_migratetype may differ from pageblock's
2172 * migratetype depending on the decisions in
2173 * find_suitable_fallback(). This is OK as long as it does not
2174 * differ for MIGRATE_CMA pageblocks. Those can be used as
2175 * fallback only via special __rmqueue_cma_fallback() function
2177 set_pcppage_migratetype(page, start_migratetype);
2179 trace_mm_page_alloc_extfrag(page, order, current_order,
2180 start_migratetype, fallback_mt);
2189 * Do the hard work of removing an element from the buddy allocator.
2190 * Call me with the zone->lock already held.
2192 static struct page *__rmqueue(struct zone *zone, unsigned int order,
2197 page = __rmqueue_smallest(zone, order, migratetype);
2198 if (unlikely(!page)) {
2199 if (migratetype == MIGRATE_MOVABLE)
2200 page = __rmqueue_cma_fallback(zone, order);
2203 page = __rmqueue_fallback(zone, order, migratetype);
2206 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2211 * Obtain a specified number of elements from the buddy allocator, all under
2212 * a single hold of the lock, for efficiency. Add them to the supplied list.
2213 * Returns the number of new pages which were placed at *list.
2215 static int rmqueue_bulk(struct zone *zone, unsigned int order,
2216 unsigned long count, struct list_head *list,
2217 int migratetype, bool cold)
2221 spin_lock(&zone->lock);
2222 for (i = 0; i < count; ++i) {
2223 struct page *page = __rmqueue(zone, order, migratetype);
2224 if (unlikely(page == NULL))
2227 if (unlikely(check_pcp_refill(page)))
2231 * Split buddy pages returned by expand() are received here
2232 * in physical page order. The page is added to the callers and
2233 * list and the list head then moves forward. From the callers
2234 * perspective, the linked list is ordered by page number in
2235 * some conditions. This is useful for IO devices that can
2236 * merge IO requests if the physical pages are ordered
2240 list_add(&page->lru, list);
2242 list_add_tail(&page->lru, list);
2245 if (is_migrate_cma(get_pcppage_migratetype(page)))
2246 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2251 * i pages were removed from the buddy list even if some leak due
2252 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2253 * on i. Do not confuse with 'alloced' which is the number of
2254 * pages added to the pcp list.
2256 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2257 spin_unlock(&zone->lock);
2263 * Called from the vmstat counter updater to drain pagesets of this
2264 * currently executing processor on remote nodes after they have
2267 * Note that this function must be called with the thread pinned to
2268 * a single processor.
2270 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2272 unsigned long flags;
2273 int to_drain, batch;
2275 local_irq_save(flags);
2276 batch = READ_ONCE(pcp->batch);
2277 to_drain = min(pcp->count, batch);
2279 free_pcppages_bulk(zone, to_drain, pcp);
2280 pcp->count -= to_drain;
2282 local_irq_restore(flags);
2287 * Drain pcplists of the indicated processor and zone.
2289 * The processor must either be the current processor and the
2290 * thread pinned to the current processor or a processor that
2293 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2295 unsigned long flags;
2296 struct per_cpu_pageset *pset;
2297 struct per_cpu_pages *pcp;
2299 local_irq_save(flags);
2300 pset = per_cpu_ptr(zone->pageset, cpu);
2304 free_pcppages_bulk(zone, pcp->count, pcp);
2307 local_irq_restore(flags);
2311 * Drain pcplists of all zones on the indicated processor.
2313 * The processor must either be the current processor and the
2314 * thread pinned to the current processor or a processor that
2317 static void drain_pages(unsigned int cpu)
2321 for_each_populated_zone(zone) {
2322 drain_pages_zone(cpu, zone);
2327 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2329 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2330 * the single zone's pages.
2332 void drain_local_pages(struct zone *zone)
2334 int cpu = smp_processor_id();
2337 drain_pages_zone(cpu, zone);
2342 static void drain_local_pages_wq(struct work_struct *work)
2344 drain_local_pages(NULL);
2348 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2350 * When zone parameter is non-NULL, spill just the single zone's pages.
2352 * Note that this can be extremely slow as the draining happens in a workqueue.
2354 void drain_all_pages(struct zone *zone)
2356 struct work_struct __percpu *works;
2360 * Allocate in the BSS so we wont require allocation in
2361 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2363 static cpumask_t cpus_with_pcps;
2365 /* Workqueues cannot recurse */
2366 if (current->flags & PF_WQ_WORKER)
2370 * As this can be called from reclaim context, do not reenter reclaim.
2371 * An allocation failure can be handled, it's simply slower
2374 works = alloc_percpu_gfp(struct work_struct, GFP_ATOMIC);
2377 * We don't care about racing with CPU hotplug event
2378 * as offline notification will cause the notified
2379 * cpu to drain that CPU pcps and on_each_cpu_mask
2380 * disables preemption as part of its processing
2382 for_each_online_cpu(cpu) {
2383 struct per_cpu_pageset *pcp;
2385 bool has_pcps = false;
2388 pcp = per_cpu_ptr(zone->pageset, cpu);
2392 for_each_populated_zone(z) {
2393 pcp = per_cpu_ptr(z->pageset, cpu);
2394 if (pcp->pcp.count) {
2402 cpumask_set_cpu(cpu, &cpus_with_pcps);
2404 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2408 for_each_cpu(cpu, &cpus_with_pcps) {
2409 struct work_struct *work = per_cpu_ptr(works, cpu);
2410 INIT_WORK(work, drain_local_pages_wq);
2411 schedule_work_on(cpu, work);
2413 for_each_cpu(cpu, &cpus_with_pcps)
2414 flush_work(per_cpu_ptr(works, cpu));
2416 for_each_cpu(cpu, &cpus_with_pcps) {
2417 struct work_struct work;
2419 INIT_WORK(&work, drain_local_pages_wq);
2420 schedule_work_on(cpu, &work);
2427 #ifdef CONFIG_HIBERNATION
2429 void mark_free_pages(struct zone *zone)
2431 unsigned long pfn, max_zone_pfn;
2432 unsigned long flags;
2433 unsigned int order, t;
2436 if (zone_is_empty(zone))
2439 spin_lock_irqsave(&zone->lock, flags);
2441 max_zone_pfn = zone_end_pfn(zone);
2442 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2443 if (pfn_valid(pfn)) {
2444 page = pfn_to_page(pfn);
2446 if (page_zone(page) != zone)
2449 if (!swsusp_page_is_forbidden(page))
2450 swsusp_unset_page_free(page);
2453 for_each_migratetype_order(order, t) {
2454 list_for_each_entry(page,
2455 &zone->free_area[order].free_list[t], lru) {
2458 pfn = page_to_pfn(page);
2459 for (i = 0; i < (1UL << order); i++)
2460 swsusp_set_page_free(pfn_to_page(pfn + i));
2463 spin_unlock_irqrestore(&zone->lock, flags);
2465 #endif /* CONFIG_PM */
2468 * Free a 0-order page
2469 * cold == true ? free a cold page : free a hot page
2471 void free_hot_cold_page(struct page *page, bool cold)
2473 struct zone *zone = page_zone(page);
2474 struct per_cpu_pages *pcp;
2475 unsigned long flags;
2476 unsigned long pfn = page_to_pfn(page);
2479 if (!free_pcp_prepare(page))
2482 migratetype = get_pfnblock_migratetype(page, pfn);
2483 set_pcppage_migratetype(page, migratetype);
2484 local_irq_save(flags);
2485 __count_vm_event(PGFREE);
2488 * We only track unmovable, reclaimable and movable on pcp lists.
2489 * Free ISOLATE pages back to the allocator because they are being
2490 * offlined but treat RESERVE as movable pages so we can get those
2491 * areas back if necessary. Otherwise, we may have to free
2492 * excessively into the page allocator
2494 if (migratetype >= MIGRATE_PCPTYPES) {
2495 if (unlikely(is_migrate_isolate(migratetype))) {
2496 free_one_page(zone, page, pfn, 0, migratetype);
2499 migratetype = MIGRATE_MOVABLE;
2502 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2504 list_add(&page->lru, &pcp->lists[migratetype]);
2506 list_add_tail(&page->lru, &pcp->lists[migratetype]);
2508 if (pcp->count >= pcp->high) {
2509 unsigned long batch = READ_ONCE(pcp->batch);
2510 free_pcppages_bulk(zone, batch, pcp);
2511 pcp->count -= batch;
2515 local_irq_restore(flags);
2519 * Free a list of 0-order pages
2521 void free_hot_cold_page_list(struct list_head *list, bool cold)
2523 struct page *page, *next;
2525 list_for_each_entry_safe(page, next, list, lru) {
2526 trace_mm_page_free_batched(page, cold);
2527 free_hot_cold_page(page, cold);
2532 * split_page takes a non-compound higher-order page, and splits it into
2533 * n (1<<order) sub-pages: page[0..n]
2534 * Each sub-page must be freed individually.
2536 * Note: this is probably too low level an operation for use in drivers.
2537 * Please consult with lkml before using this in your driver.
2539 void split_page(struct page *page, unsigned int order)
2543 VM_BUG_ON_PAGE(PageCompound(page), page);
2544 VM_BUG_ON_PAGE(!page_count(page), page);
2546 #ifdef CONFIG_KMEMCHECK
2548 * Split shadow pages too, because free(page[0]) would
2549 * otherwise free the whole shadow.
2551 if (kmemcheck_page_is_tracked(page))
2552 split_page(virt_to_page(page[0].shadow), order);
2555 for (i = 1; i < (1 << order); i++)
2556 set_page_refcounted(page + i);
2557 split_page_owner(page, order);
2559 EXPORT_SYMBOL_GPL(split_page);
2561 int __isolate_free_page(struct page *page, unsigned int order)
2563 unsigned long watermark;
2567 BUG_ON(!PageBuddy(page));
2569 zone = page_zone(page);
2570 mt = get_pageblock_migratetype(page);
2572 if (!is_migrate_isolate(mt)) {
2574 * Obey watermarks as if the page was being allocated. We can
2575 * emulate a high-order watermark check with a raised order-0
2576 * watermark, because we already know our high-order page
2579 watermark = min_wmark_pages(zone) + (1UL << order);
2580 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2583 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2586 /* Remove page from free list */
2587 list_del(&page->lru);
2588 zone->free_area[order].nr_free--;
2589 rmv_page_order(page);
2592 * Set the pageblock if the isolated page is at least half of a
2595 if (order >= pageblock_order - 1) {
2596 struct page *endpage = page + (1 << order) - 1;
2597 for (; page < endpage; page += pageblock_nr_pages) {
2598 int mt = get_pageblock_migratetype(page);
2599 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2600 && mt != MIGRATE_HIGHATOMIC)
2601 set_pageblock_migratetype(page,
2607 return 1UL << order;
2611 * Update NUMA hit/miss statistics
2613 * Must be called with interrupts disabled.
2615 static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2618 enum zone_stat_item local_stat = NUMA_LOCAL;
2620 if (z->node != numa_node_id())
2621 local_stat = NUMA_OTHER;
2623 if (z->node == preferred_zone->node)
2624 __inc_zone_state(z, NUMA_HIT);
2626 __inc_zone_state(z, NUMA_MISS);
2627 __inc_zone_state(preferred_zone, NUMA_FOREIGN);
2629 __inc_zone_state(z, local_stat);
2633 /* Remove page from the per-cpu list, caller must protect the list */
2634 static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2635 bool cold, struct per_cpu_pages *pcp,
2636 struct list_head *list)
2641 if (list_empty(list)) {
2642 pcp->count += rmqueue_bulk(zone, 0,
2645 if (unlikely(list_empty(list)))
2650 page = list_last_entry(list, struct page, lru);
2652 page = list_first_entry(list, struct page, lru);
2654 list_del(&page->lru);
2656 } while (check_new_pcp(page));
2661 /* Lock and remove page from the per-cpu list */
2662 static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2663 struct zone *zone, unsigned int order,
2664 gfp_t gfp_flags, int migratetype)
2666 struct per_cpu_pages *pcp;
2667 struct list_head *list;
2668 bool cold = ((gfp_flags & __GFP_COLD) != 0);
2670 unsigned long flags;
2672 local_irq_save(flags);
2673 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2674 list = &pcp->lists[migratetype];
2675 page = __rmqueue_pcplist(zone, migratetype, cold, pcp, list);
2677 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2678 zone_statistics(preferred_zone, zone);
2680 local_irq_restore(flags);
2685 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2688 struct page *rmqueue(struct zone *preferred_zone,
2689 struct zone *zone, unsigned int order,
2690 gfp_t gfp_flags, unsigned int alloc_flags,
2693 unsigned long flags;
2696 if (likely(order == 0)) {
2697 page = rmqueue_pcplist(preferred_zone, zone, order,
2698 gfp_flags, migratetype);
2703 * We most definitely don't want callers attempting to
2704 * allocate greater than order-1 page units with __GFP_NOFAIL.
2706 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2707 spin_lock_irqsave(&zone->lock, flags);
2711 if (alloc_flags & ALLOC_HARDER) {
2712 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2714 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2717 page = __rmqueue(zone, order, migratetype);
2718 } while (page && check_new_pages(page, order));
2719 spin_unlock(&zone->lock);
2722 __mod_zone_freepage_state(zone, -(1 << order),
2723 get_pcppage_migratetype(page));
2725 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2726 zone_statistics(preferred_zone, zone);
2727 local_irq_restore(flags);
2730 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
2734 local_irq_restore(flags);
2738 #ifdef CONFIG_FAIL_PAGE_ALLOC
2741 struct fault_attr attr;
2743 bool ignore_gfp_highmem;
2744 bool ignore_gfp_reclaim;
2746 } fail_page_alloc = {
2747 .attr = FAULT_ATTR_INITIALIZER,
2748 .ignore_gfp_reclaim = true,
2749 .ignore_gfp_highmem = true,
2753 static int __init setup_fail_page_alloc(char *str)
2755 return setup_fault_attr(&fail_page_alloc.attr, str);
2757 __setup("fail_page_alloc=", setup_fail_page_alloc);
2759 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2761 if (order < fail_page_alloc.min_order)
2763 if (gfp_mask & __GFP_NOFAIL)
2765 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
2767 if (fail_page_alloc.ignore_gfp_reclaim &&
2768 (gfp_mask & __GFP_DIRECT_RECLAIM))
2771 return should_fail(&fail_page_alloc.attr, 1 << order);
2774 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2776 static int __init fail_page_alloc_debugfs(void)
2778 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
2781 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
2782 &fail_page_alloc.attr);
2784 return PTR_ERR(dir);
2786 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
2787 &fail_page_alloc.ignore_gfp_reclaim))
2789 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
2790 &fail_page_alloc.ignore_gfp_highmem))
2792 if (!debugfs_create_u32("min-order", mode, dir,
2793 &fail_page_alloc.min_order))
2798 debugfs_remove_recursive(dir);
2803 late_initcall(fail_page_alloc_debugfs);
2805 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2807 #else /* CONFIG_FAIL_PAGE_ALLOC */
2809 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
2814 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2817 * Return true if free base pages are above 'mark'. For high-order checks it
2818 * will return true of the order-0 watermark is reached and there is at least
2819 * one free page of a suitable size. Checking now avoids taking the zone lock
2820 * to check in the allocation paths if no pages are free.
2822 bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2823 int classzone_idx, unsigned int alloc_flags,
2828 const bool alloc_harder = (alloc_flags & ALLOC_HARDER);
2830 /* free_pages may go negative - that's OK */
2831 free_pages -= (1 << order) - 1;
2833 if (alloc_flags & ALLOC_HIGH)
2837 * If the caller does not have rights to ALLOC_HARDER then subtract
2838 * the high-atomic reserves. This will over-estimate the size of the
2839 * atomic reserve but it avoids a search.
2841 if (likely(!alloc_harder))
2842 free_pages -= z->nr_reserved_highatomic;
2847 /* If allocation can't use CMA areas don't use free CMA pages */
2848 if (!(alloc_flags & ALLOC_CMA))
2849 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
2853 * Check watermarks for an order-0 allocation request. If these
2854 * are not met, then a high-order request also cannot go ahead
2855 * even if a suitable page happened to be free.
2857 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
2860 /* If this is an order-0 request then the watermark is fine */
2864 /* For a high-order request, check at least one suitable page is free */
2865 for (o = order; o < MAX_ORDER; o++) {
2866 struct free_area *area = &z->free_area[o];
2875 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
2876 if (!list_empty(&area->free_list[mt]))
2881 if ((alloc_flags & ALLOC_CMA) &&
2882 !list_empty(&area->free_list[MIGRATE_CMA])) {
2890 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
2891 int classzone_idx, unsigned int alloc_flags)
2893 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2894 zone_page_state(z, NR_FREE_PAGES));
2897 static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
2898 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
2900 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2904 /* If allocation can't use CMA areas don't use free CMA pages */
2905 if (!(alloc_flags & ALLOC_CMA))
2906 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
2910 * Fast check for order-0 only. If this fails then the reserves
2911 * need to be calculated. There is a corner case where the check
2912 * passes but only the high-order atomic reserve are free. If
2913 * the caller is !atomic then it'll uselessly search the free
2914 * list. That corner case is then slower but it is harmless.
2916 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
2919 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
2923 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
2924 unsigned long mark, int classzone_idx)
2926 long free_pages = zone_page_state(z, NR_FREE_PAGES);
2928 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
2929 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
2931 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
2936 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2938 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2941 #else /* CONFIG_NUMA */
2942 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2946 #endif /* CONFIG_NUMA */
2949 * get_page_from_freelist goes through the zonelist trying to allocate
2952 static struct page *
2953 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2954 const struct alloc_context *ac)
2956 struct zoneref *z = ac->preferred_zoneref;
2958 struct pglist_data *last_pgdat_dirty_limit = NULL;
2961 * Scan zonelist, looking for a zone with enough free.
2962 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2964 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
2969 if (cpusets_enabled() &&
2970 (alloc_flags & ALLOC_CPUSET) &&
2971 !__cpuset_zone_allowed(zone, gfp_mask))
2974 * When allocating a page cache page for writing, we
2975 * want to get it from a node that is within its dirty
2976 * limit, such that no single node holds more than its
2977 * proportional share of globally allowed dirty pages.
2978 * The dirty limits take into account the node's
2979 * lowmem reserves and high watermark so that kswapd
2980 * should be able to balance it without having to
2981 * write pages from its LRU list.
2983 * XXX: For now, allow allocations to potentially
2984 * exceed the per-node dirty limit in the slowpath
2985 * (spread_dirty_pages unset) before going into reclaim,
2986 * which is important when on a NUMA setup the allowed
2987 * nodes are together not big enough to reach the
2988 * global limit. The proper fix for these situations
2989 * will require awareness of nodes in the
2990 * dirty-throttling and the flusher threads.
2992 if (ac->spread_dirty_pages) {
2993 if (last_pgdat_dirty_limit == zone->zone_pgdat)
2996 if (!node_dirty_ok(zone->zone_pgdat)) {
2997 last_pgdat_dirty_limit = zone->zone_pgdat;
3002 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3003 if (!zone_watermark_fast(zone, order, mark,
3004 ac_classzone_idx(ac), alloc_flags)) {
3007 /* Checked here to keep the fast path fast */
3008 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3009 if (alloc_flags & ALLOC_NO_WATERMARKS)
3012 if (node_reclaim_mode == 0 ||
3013 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3016 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3018 case NODE_RECLAIM_NOSCAN:
3021 case NODE_RECLAIM_FULL:
3022 /* scanned but unreclaimable */
3025 /* did we reclaim enough */
3026 if (zone_watermark_ok(zone, order, mark,
3027 ac_classzone_idx(ac), alloc_flags))
3035 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3036 gfp_mask, alloc_flags, ac->migratetype);
3038 prep_new_page(page, order, gfp_mask, alloc_flags);
3041 * If this is a high-order atomic allocation then check
3042 * if the pageblock should be reserved for the future
3044 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3045 reserve_highatomic_pageblock(page, zone, order);
3055 * Large machines with many possible nodes should not always dump per-node
3056 * meminfo in irq context.
3058 static inline bool should_suppress_show_mem(void)
3063 ret = in_interrupt();
3068 static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3070 unsigned int filter = SHOW_MEM_FILTER_NODES;
3071 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3073 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3077 * This documents exceptions given to allocations in certain
3078 * contexts that are allowed to allocate outside current's set
3081 if (!(gfp_mask & __GFP_NOMEMALLOC))
3082 if (test_thread_flag(TIF_MEMDIE) ||
3083 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3084 filter &= ~SHOW_MEM_FILTER_NODES;
3085 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3086 filter &= ~SHOW_MEM_FILTER_NODES;
3088 show_mem(filter, nodemask);
3091 void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3093 struct va_format vaf;
3095 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3096 DEFAULT_RATELIMIT_BURST);
3098 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
3099 debug_guardpage_minorder() > 0)
3102 pr_warn("%s: ", current->comm);
3104 va_start(args, fmt);
3107 pr_cont("%pV", &vaf);
3110 pr_cont(", mode:%#x(%pGg), nodemask=", gfp_mask, &gfp_mask);
3112 pr_cont("%*pbl\n", nodemask_pr_args(nodemask));
3114 pr_cont("(null)\n");
3116 cpuset_print_current_mems_allowed();
3119 warn_alloc_show_mem(gfp_mask, nodemask);
3122 static inline struct page *
3123 __alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3124 unsigned int alloc_flags,
3125 const struct alloc_context *ac)
3129 page = get_page_from_freelist(gfp_mask, order,
3130 alloc_flags|ALLOC_CPUSET, ac);
3132 * fallback to ignore cpuset restriction if our nodes
3136 page = get_page_from_freelist(gfp_mask, order,
3142 static inline struct page *
3143 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3144 const struct alloc_context *ac, unsigned long *did_some_progress)
3146 struct oom_control oc = {
3147 .zonelist = ac->zonelist,
3148 .nodemask = ac->nodemask,
3150 .gfp_mask = gfp_mask,
3155 *did_some_progress = 0;
3158 * Acquire the oom lock. If that fails, somebody else is
3159 * making progress for us.
3161 if (!mutex_trylock(&oom_lock)) {
3162 *did_some_progress = 1;
3163 schedule_timeout_uninterruptible(1);
3168 * Go through the zonelist yet one more time, keep very high watermark
3169 * here, this is only to catch a parallel oom killing, we must fail if
3170 * we're still under heavy pressure.
3172 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
3173 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3177 /* Coredumps can quickly deplete all memory reserves */
3178 if (current->flags & PF_DUMPCORE)
3180 /* The OOM killer will not help higher order allocs */
3181 if (order > PAGE_ALLOC_COSTLY_ORDER)
3183 /* The OOM killer does not needlessly kill tasks for lowmem */
3184 if (ac->high_zoneidx < ZONE_NORMAL)
3186 if (pm_suspended_storage())
3189 * XXX: GFP_NOFS allocations should rather fail than rely on
3190 * other request to make a forward progress.
3191 * We are in an unfortunate situation where out_of_memory cannot
3192 * do much for this context but let's try it to at least get
3193 * access to memory reserved if the current task is killed (see
3194 * out_of_memory). Once filesystems are ready to handle allocation
3195 * failures more gracefully we should just bail out here.
3198 /* The OOM killer may not free memory on a specific node */
3199 if (gfp_mask & __GFP_THISNODE)
3202 /* Exhausted what can be done so it's blamo time */
3203 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3204 *did_some_progress = 1;
3207 * Help non-failing allocations by giving them access to memory
3210 if (gfp_mask & __GFP_NOFAIL)
3211 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3212 ALLOC_NO_WATERMARKS, ac);
3215 mutex_unlock(&oom_lock);
3220 * Maximum number of compaction retries wit a progress before OOM
3221 * killer is consider as the only way to move forward.
3223 #define MAX_COMPACT_RETRIES 16
3225 #ifdef CONFIG_COMPACTION
3226 /* Try memory compaction for high-order allocations before reclaim */
3227 static struct page *
3228 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3229 unsigned int alloc_flags, const struct alloc_context *ac,
3230 enum compact_priority prio, enum compact_result *compact_result)
3237 current->flags |= PF_MEMALLOC;
3238 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3240 current->flags &= ~PF_MEMALLOC;
3242 if (*compact_result <= COMPACT_INACTIVE)
3246 * At least in one zone compaction wasn't deferred or skipped, so let's
3247 * count a compaction stall
3249 count_vm_event(COMPACTSTALL);
3251 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3254 struct zone *zone = page_zone(page);
3256 zone->compact_blockskip_flush = false;
3257 compaction_defer_reset(zone, order, true);
3258 count_vm_event(COMPACTSUCCESS);
3263 * It's bad if compaction run occurs and fails. The most likely reason
3264 * is that pages exist, but not enough to satisfy watermarks.
3266 count_vm_event(COMPACTFAIL);
3274 should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3275 enum compact_result compact_result,
3276 enum compact_priority *compact_priority,
3277 int *compaction_retries)
3279 int max_retries = MAX_COMPACT_RETRIES;
3282 int retries = *compaction_retries;
3283 enum compact_priority priority = *compact_priority;
3288 if (compaction_made_progress(compact_result))
3289 (*compaction_retries)++;
3292 * compaction considers all the zone as desperately out of memory
3293 * so it doesn't really make much sense to retry except when the
3294 * failure could be caused by insufficient priority
3296 if (compaction_failed(compact_result))
3297 goto check_priority;
3300 * make sure the compaction wasn't deferred or didn't bail out early
3301 * due to locks contention before we declare that we should give up.
3302 * But do not retry if the given zonelist is not suitable for
3305 if (compaction_withdrawn(compact_result)) {
3306 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3311 * !costly requests are much more important than __GFP_REPEAT
3312 * costly ones because they are de facto nofail and invoke OOM
3313 * killer to move on while costly can fail and users are ready
3314 * to cope with that. 1/4 retries is rather arbitrary but we
3315 * would need much more detailed feedback from compaction to
3316 * make a better decision.
3318 if (order > PAGE_ALLOC_COSTLY_ORDER)
3320 if (*compaction_retries <= max_retries) {
3326 * Make sure there are attempts at the highest priority if we exhausted
3327 * all retries or failed at the lower priorities.
3330 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3331 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3333 if (*compact_priority > min_priority) {
3334 (*compact_priority)--;
3335 *compaction_retries = 0;
3339 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3343 static inline struct page *
3344 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3345 unsigned int alloc_flags, const struct alloc_context *ac,
3346 enum compact_priority prio, enum compact_result *compact_result)
3348 *compact_result = COMPACT_SKIPPED;
3353 should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3354 enum compact_result compact_result,
3355 enum compact_priority *compact_priority,
3356 int *compaction_retries)
3361 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3365 * There are setups with compaction disabled which would prefer to loop
3366 * inside the allocator rather than hit the oom killer prematurely.
3367 * Let's give them a good hope and keep retrying while the order-0
3368 * watermarks are OK.
3370 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3372 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3373 ac_classzone_idx(ac), alloc_flags))
3378 #endif /* CONFIG_COMPACTION */
3380 /* Perform direct synchronous page reclaim */
3382 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
3383 const struct alloc_context *ac)
3385 struct reclaim_state reclaim_state;
3390 /* We now go into synchronous reclaim */
3391 cpuset_memory_pressure_bump();
3392 current->flags |= PF_MEMALLOC;
3393 lockdep_set_current_reclaim_state(gfp_mask);
3394 reclaim_state.reclaimed_slab = 0;
3395 current->reclaim_state = &reclaim_state;
3397 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3400 current->reclaim_state = NULL;
3401 lockdep_clear_current_reclaim_state();
3402 current->flags &= ~PF_MEMALLOC;
3409 /* The really slow allocator path where we enter direct reclaim */
3410 static inline struct page *
3411 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3412 unsigned int alloc_flags, const struct alloc_context *ac,
3413 unsigned long *did_some_progress)
3415 struct page *page = NULL;
3416 bool drained = false;
3418 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3419 if (unlikely(!(*did_some_progress)))
3423 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3426 * If an allocation failed after direct reclaim, it could be because
3427 * pages are pinned on the per-cpu lists or in high alloc reserves.
3428 * Shrink them them and try again
3430 if (!page && !drained) {
3431 unreserve_highatomic_pageblock(ac, false);
3432 drain_all_pages(NULL);
3440 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
3444 pg_data_t *last_pgdat = NULL;
3446 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3447 ac->high_zoneidx, ac->nodemask) {
3448 if (last_pgdat != zone->zone_pgdat)
3449 wakeup_kswapd(zone, order, ac->high_zoneidx);
3450 last_pgdat = zone->zone_pgdat;
3454 static inline unsigned int
3455 gfp_to_alloc_flags(gfp_t gfp_mask)
3457 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3459 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3460 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3463 * The caller may dip into page reserves a bit more if the caller
3464 * cannot run direct reclaim, or if the caller has realtime scheduling
3465 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3466 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3468 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3470 if (gfp_mask & __GFP_ATOMIC) {
3472 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3473 * if it can't schedule.
3475 if (!(gfp_mask & __GFP_NOMEMALLOC))
3476 alloc_flags |= ALLOC_HARDER;
3478 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3479 * comment for __cpuset_node_allowed().
3481 alloc_flags &= ~ALLOC_CPUSET;
3482 } else if (unlikely(rt_task(current)) && !in_interrupt())
3483 alloc_flags |= ALLOC_HARDER;
3486 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3487 alloc_flags |= ALLOC_CMA;
3492 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3494 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3497 if (gfp_mask & __GFP_MEMALLOC)
3499 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3501 if (!in_interrupt() &&
3502 ((current->flags & PF_MEMALLOC) ||
3503 unlikely(test_thread_flag(TIF_MEMDIE))))
3510 * Maximum number of reclaim retries without any progress before OOM killer
3511 * is consider as the only way to move forward.
3513 #define MAX_RECLAIM_RETRIES 16
3516 * Checks whether it makes sense to retry the reclaim to make a forward progress
3517 * for the given allocation request.
3518 * The reclaim feedback represented by did_some_progress (any progress during
3519 * the last reclaim round) and no_progress_loops (number of reclaim rounds without
3520 * any progress in a row) is considered as well as the reclaimable pages on the
3521 * applicable zone list (with a backoff mechanism which is a function of
3522 * no_progress_loops).
3524 * Returns true if a retry is viable or false to enter the oom path.
3527 should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3528 struct alloc_context *ac, int alloc_flags,
3529 bool did_some_progress, int *no_progress_loops)
3535 * Costly allocations might have made a progress but this doesn't mean
3536 * their order will become available due to high fragmentation so
3537 * always increment the no progress counter for them
3539 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3540 *no_progress_loops = 0;
3542 (*no_progress_loops)++;
3545 * Make sure we converge to OOM if we cannot make any progress
3546 * several times in the row.
3548 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3549 /* Before OOM, exhaust highatomic_reserve */
3550 return unreserve_highatomic_pageblock(ac, true);
3554 * Keep reclaiming pages while there is a chance this will lead
3555 * somewhere. If none of the target zones can satisfy our allocation
3556 * request even if all reclaimable pages are considered then we are
3557 * screwed and have to go OOM.
3559 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3561 unsigned long available;
3562 unsigned long reclaimable;
3563 unsigned long min_wmark = min_wmark_pages(zone);
3566 available = reclaimable = zone_reclaimable_pages(zone);
3567 available -= DIV_ROUND_UP((*no_progress_loops) * available,
3568 MAX_RECLAIM_RETRIES);
3569 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3572 * Would the allocation succeed if we reclaimed the whole
3575 wmark = __zone_watermark_ok(zone, order, min_wmark,
3576 ac_classzone_idx(ac), alloc_flags, available);
3577 trace_reclaim_retry_zone(z, order, reclaimable,
3578 available, min_wmark, *no_progress_loops, wmark);
3581 * If we didn't make any progress and have a lot of
3582 * dirty + writeback pages then we should wait for
3583 * an IO to complete to slow down the reclaim and
3584 * prevent from pre mature OOM
3586 if (!did_some_progress) {
3587 unsigned long write_pending;
3589 write_pending = zone_page_state_snapshot(zone,
3590 NR_ZONE_WRITE_PENDING);
3592 if (2 * write_pending > reclaimable) {
3593 congestion_wait(BLK_RW_ASYNC, HZ/10);
3599 * Memory allocation/reclaim might be called from a WQ
3600 * context and the current implementation of the WQ
3601 * concurrency control doesn't recognize that
3602 * a particular WQ is congested if the worker thread is
3603 * looping without ever sleeping. Therefore we have to
3604 * do a short sleep here rather than calling
3607 if (current->flags & PF_WQ_WORKER)
3608 schedule_timeout_uninterruptible(1);
3619 static inline struct page *
3620 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
3621 struct alloc_context *ac)
3623 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
3624 struct page *page = NULL;
3625 unsigned int alloc_flags;
3626 unsigned long did_some_progress;
3627 enum compact_priority compact_priority;
3628 enum compact_result compact_result;
3629 int compaction_retries;
3630 int no_progress_loops;
3631 unsigned long alloc_start = jiffies;
3632 unsigned int stall_timeout = 10 * HZ;
3633 unsigned int cpuset_mems_cookie;
3636 * In the slowpath, we sanity check order to avoid ever trying to
3637 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3638 * be using allocators in order of preference for an area that is
3641 if (order >= MAX_ORDER) {
3642 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
3647 * We also sanity check to catch abuse of atomic reserves being used by
3648 * callers that are not in atomic context.
3650 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
3651 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
3652 gfp_mask &= ~__GFP_ATOMIC;
3655 compaction_retries = 0;
3656 no_progress_loops = 0;
3657 compact_priority = DEF_COMPACT_PRIORITY;
3658 cpuset_mems_cookie = read_mems_allowed_begin();
3661 * The fast path uses conservative alloc_flags to succeed only until
3662 * kswapd needs to be woken up, and to avoid the cost of setting up
3663 * alloc_flags precisely. So we do that now.
3665 alloc_flags = gfp_to_alloc_flags(gfp_mask);
3668 * We need to recalculate the starting point for the zonelist iterator
3669 * because we might have used different nodemask in the fast path, or
3670 * there was a cpuset modification and we are retrying - otherwise we
3671 * could end up iterating over non-eligible zones endlessly.
3673 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3674 ac->high_zoneidx, ac->nodemask);
3675 if (!ac->preferred_zoneref->zone)
3678 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3679 wake_all_kswapds(order, ac);
3682 * The adjusted alloc_flags might result in immediate success, so try
3685 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3690 * For costly allocations, try direct compaction first, as it's likely
3691 * that we have enough base pages and don't need to reclaim. Don't try
3692 * that for allocations that are allowed to ignore watermarks, as the
3693 * ALLOC_NO_WATERMARKS attempt didn't yet happen.
3695 if (can_direct_reclaim && order > PAGE_ALLOC_COSTLY_ORDER &&
3696 !gfp_pfmemalloc_allowed(gfp_mask)) {
3697 page = __alloc_pages_direct_compact(gfp_mask, order,
3699 INIT_COMPACT_PRIORITY,
3705 * Checks for costly allocations with __GFP_NORETRY, which
3706 * includes THP page fault allocations
3708 if (gfp_mask & __GFP_NORETRY) {
3710 * If compaction is deferred for high-order allocations,
3711 * it is because sync compaction recently failed. If
3712 * this is the case and the caller requested a THP
3713 * allocation, we do not want to heavily disrupt the
3714 * system, so we fail the allocation instead of entering
3717 if (compact_result == COMPACT_DEFERRED)
3721 * Looks like reclaim/compaction is worth trying, but
3722 * sync compaction could be very expensive, so keep
3723 * using async compaction.
3725 compact_priority = INIT_COMPACT_PRIORITY;
3730 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
3731 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3732 wake_all_kswapds(order, ac);
3734 if (gfp_pfmemalloc_allowed(gfp_mask))
3735 alloc_flags = ALLOC_NO_WATERMARKS;
3738 * Reset the zonelist iterators if memory policies can be ignored.
3739 * These allocations are high priority and system rather than user
3742 if (!(alloc_flags & ALLOC_CPUSET) || (alloc_flags & ALLOC_NO_WATERMARKS)) {
3743 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
3744 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3745 ac->high_zoneidx, ac->nodemask);
3748 /* Attempt with potentially adjusted zonelist and alloc_flags */
3749 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3753 /* Caller is not willing to reclaim, we can't balance anything */
3754 if (!can_direct_reclaim)
3757 /* Make sure we know about allocations which stall for too long */
3758 if (time_after(jiffies, alloc_start + stall_timeout)) {
3759 warn_alloc(gfp_mask, ac->nodemask,
3760 "page allocation stalls for %ums, order:%u",
3761 jiffies_to_msecs(jiffies-alloc_start), order);
3762 stall_timeout += 10 * HZ;
3765 /* Avoid recursion of direct reclaim */
3766 if (current->flags & PF_MEMALLOC)
3769 /* Try direct reclaim and then allocating */
3770 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
3771 &did_some_progress);
3775 /* Try direct compaction and then allocating */
3776 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
3777 compact_priority, &compact_result);
3781 /* Do not loop if specifically requested */
3782 if (gfp_mask & __GFP_NORETRY)
3786 * Do not retry costly high order allocations unless they are
3789 if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
3792 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
3793 did_some_progress > 0, &no_progress_loops))
3797 * It doesn't make any sense to retry for the compaction if the order-0
3798 * reclaim is not able to make any progress because the current
3799 * implementation of the compaction depends on the sufficient amount
3800 * of free memory (see __compaction_suitable)
3802 if (did_some_progress > 0 &&
3803 should_compact_retry(ac, order, alloc_flags,
3804 compact_result, &compact_priority,
3805 &compaction_retries))
3809 * It's possible we raced with cpuset update so the OOM would be
3810 * premature (see below the nopage: label for full explanation).
3812 if (read_mems_allowed_retry(cpuset_mems_cookie))
3815 /* Reclaim has failed us, start killing things */
3816 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
3820 /* Avoid allocations with no watermarks from looping endlessly */
3821 if (test_thread_flag(TIF_MEMDIE))
3824 /* Retry as long as the OOM killer is making progress */
3825 if (did_some_progress) {
3826 no_progress_loops = 0;
3832 * When updating a task's mems_allowed or mempolicy nodemask, it is
3833 * possible to race with parallel threads in such a way that our
3834 * allocation can fail while the mask is being updated. If we are about
3835 * to fail, check if the cpuset changed during allocation and if so,
3838 if (read_mems_allowed_retry(cpuset_mems_cookie))
3842 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
3845 if (gfp_mask & __GFP_NOFAIL) {
3847 * All existing users of the __GFP_NOFAIL are blockable, so warn
3848 * of any new users that actually require GFP_NOWAIT
3850 if (WARN_ON_ONCE(!can_direct_reclaim))
3854 * PF_MEMALLOC request from this context is rather bizarre
3855 * because we cannot reclaim anything and only can loop waiting
3856 * for somebody to do a work for us
3858 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
3861 * non failing costly orders are a hard requirement which we
3862 * are not prepared for much so let's warn about these users
3863 * so that we can identify them and convert them to something
3866 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
3869 * Help non-failing allocations by giving them access to memory
3870 * reserves but do not use ALLOC_NO_WATERMARKS because this
3871 * could deplete whole memory reserves which would just make
3872 * the situation worse
3874 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
3882 warn_alloc(gfp_mask, ac->nodemask,
3883 "page allocation failure: order:%u", order);
3888 static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
3889 struct zonelist *zonelist, nodemask_t *nodemask,
3890 struct alloc_context *ac, gfp_t *alloc_mask,
3891 unsigned int *alloc_flags)
3893 ac->high_zoneidx = gfp_zone(gfp_mask);
3894 ac->zonelist = zonelist;
3895 ac->nodemask = nodemask;
3896 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
3898 if (cpusets_enabled()) {
3899 *alloc_mask |= __GFP_HARDWALL;
3900 *alloc_flags |= ALLOC_CPUSET;
3902 ac->nodemask = &cpuset_current_mems_allowed;
3905 lockdep_trace_alloc(gfp_mask);
3907 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
3909 if (should_fail_alloc_page(gfp_mask, order))
3913 * Check the zones suitable for the gfp_mask contain at least one
3914 * valid zone. It's possible to have an empty zonelist as a result
3915 * of __GFP_THISNODE and a memoryless node
3917 if (unlikely(!ac->zonelist->_zonerefs->zone))
3920 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
3921 *alloc_flags |= ALLOC_CMA;
3926 /* Determine whether to spread dirty pages and what the first usable zone */
3927 static inline void finalise_ac(gfp_t gfp_mask,
3928 unsigned int order, struct alloc_context *ac)
3930 /* Dirty zone balancing only done in the fast path */
3931 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
3934 * The preferred zone is used for statistics but crucially it is
3935 * also used as the starting point for the zonelist iterator. It
3936 * may get reset for allocations that ignore memory policies.
3938 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
3939 ac->high_zoneidx, ac->nodemask);
3943 * This is the 'heart' of the zoned buddy allocator.
3946 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
3947 struct zonelist *zonelist, nodemask_t *nodemask)
3950 unsigned int alloc_flags = ALLOC_WMARK_LOW;
3951 gfp_t alloc_mask = gfp_mask; /* The gfp_t that was actually used for allocation */
3952 struct alloc_context ac = { };
3954 gfp_mask &= gfp_allowed_mask;
3955 if (!prepare_alloc_pages(gfp_mask, order, zonelist, nodemask, &ac, &alloc_mask, &alloc_flags))
3958 finalise_ac(gfp_mask, order, &ac);
3959 if (!ac.preferred_zoneref->zone) {
3962 * This might be due to race with cpuset_current_mems_allowed
3963 * update, so make sure we retry with original nodemask in the
3969 /* First allocation attempt */
3970 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
3976 * Runtime PM, block IO and its error handling path can deadlock
3977 * because I/O on the device might not complete.
3979 alloc_mask = memalloc_noio_flags(gfp_mask);
3980 ac.spread_dirty_pages = false;
3983 * Restore the original nodemask if it was potentially replaced with
3984 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
3986 if (unlikely(ac.nodemask != nodemask))
3987 ac.nodemask = nodemask;
3989 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
3992 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
3993 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
3994 __free_pages(page, order);
3998 if (kmemcheck_enabled && page)
3999 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
4001 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4005 EXPORT_SYMBOL(__alloc_pages_nodemask);
4008 * Common helper functions.
4010 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4015 * __get_free_pages() returns a 32-bit address, which cannot represent
4018 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4020 page = alloc_pages(gfp_mask, order);
4023 return (unsigned long) page_address(page);
4025 EXPORT_SYMBOL(__get_free_pages);
4027 unsigned long get_zeroed_page(gfp_t gfp_mask)
4029 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4031 EXPORT_SYMBOL(get_zeroed_page);
4033 void __free_pages(struct page *page, unsigned int order)
4035 if (put_page_testzero(page)) {
4037 free_hot_cold_page(page, false);
4039 __free_pages_ok(page, order);
4043 EXPORT_SYMBOL(__free_pages);
4045 void free_pages(unsigned long addr, unsigned int order)
4048 VM_BUG_ON(!virt_addr_valid((void *)addr));
4049 __free_pages(virt_to_page((void *)addr), order);
4053 EXPORT_SYMBOL(free_pages);
4057 * An arbitrary-length arbitrary-offset area of memory which resides
4058 * within a 0 or higher order page. Multiple fragments within that page
4059 * are individually refcounted, in the page's reference counter.
4061 * The page_frag functions below provide a simple allocation framework for
4062 * page fragments. This is used by the network stack and network device
4063 * drivers to provide a backing region of memory for use as either an
4064 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4066 static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4069 struct page *page = NULL;
4070 gfp_t gfp = gfp_mask;
4072 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4073 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4075 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4076 PAGE_FRAG_CACHE_MAX_ORDER);
4077 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4079 if (unlikely(!page))
4080 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4082 nc->va = page ? page_address(page) : NULL;
4087 void __page_frag_cache_drain(struct page *page, unsigned int count)
4089 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4091 if (page_ref_sub_and_test(page, count)) {
4092 unsigned int order = compound_order(page);
4095 free_hot_cold_page(page, false);
4097 __free_pages_ok(page, order);
4100 EXPORT_SYMBOL(__page_frag_cache_drain);
4102 void *page_frag_alloc(struct page_frag_cache *nc,
4103 unsigned int fragsz, gfp_t gfp_mask)
4105 unsigned int size = PAGE_SIZE;
4109 if (unlikely(!nc->va)) {
4111 page = __page_frag_cache_refill(nc, gfp_mask);
4115 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4116 /* if size can vary use size else just use PAGE_SIZE */
4119 /* Even if we own the page, we do not use atomic_set().
4120 * This would break get_page_unless_zero() users.
4122 page_ref_add(page, size - 1);
4124 /* reset page count bias and offset to start of new frag */
4125 nc->pfmemalloc = page_is_pfmemalloc(page);
4126 nc->pagecnt_bias = size;
4130 offset = nc->offset - fragsz;
4131 if (unlikely(offset < 0)) {
4132 page = virt_to_page(nc->va);
4134 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4137 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4138 /* if size can vary use size else just use PAGE_SIZE */
4141 /* OK, page count is 0, we can safely set it */
4142 set_page_count(page, size);
4144 /* reset page count bias and offset to start of new frag */
4145 nc->pagecnt_bias = size;
4146 offset = size - fragsz;
4150 nc->offset = offset;
4152 return nc->va + offset;
4154 EXPORT_SYMBOL(page_frag_alloc);
4157 * Frees a page fragment allocated out of either a compound or order 0 page.
4159 void page_frag_free(void *addr)
4161 struct page *page = virt_to_head_page(addr);
4163 if (unlikely(put_page_testzero(page)))
4164 __free_pages_ok(page, compound_order(page));
4166 EXPORT_SYMBOL(page_frag_free);
4168 static void *make_alloc_exact(unsigned long addr, unsigned int order,
4172 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4173 unsigned long used = addr + PAGE_ALIGN(size);
4175 split_page(virt_to_page((void *)addr), order);
4176 while (used < alloc_end) {
4181 return (void *)addr;
4185 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4186 * @size: the number of bytes to allocate
4187 * @gfp_mask: GFP flags for the allocation
4189 * This function is similar to alloc_pages(), except that it allocates the
4190 * minimum number of pages to satisfy the request. alloc_pages() can only
4191 * allocate memory in power-of-two pages.
4193 * This function is also limited by MAX_ORDER.
4195 * Memory allocated by this function must be released by free_pages_exact().
4197 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4199 unsigned int order = get_order(size);
4202 addr = __get_free_pages(gfp_mask, order);
4203 return make_alloc_exact(addr, order, size);
4205 EXPORT_SYMBOL(alloc_pages_exact);
4208 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4210 * @nid: the preferred node ID where memory should be allocated
4211 * @size: the number of bytes to allocate
4212 * @gfp_mask: GFP flags for the allocation
4214 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4217 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4219 unsigned int order = get_order(size);
4220 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4223 return make_alloc_exact((unsigned long)page_address(p), order, size);
4227 * free_pages_exact - release memory allocated via alloc_pages_exact()
4228 * @virt: the value returned by alloc_pages_exact.
4229 * @size: size of allocation, same value as passed to alloc_pages_exact().
4231 * Release the memory allocated by a previous call to alloc_pages_exact.
4233 void free_pages_exact(void *virt, size_t size)
4235 unsigned long addr = (unsigned long)virt;
4236 unsigned long end = addr + PAGE_ALIGN(size);
4238 while (addr < end) {
4243 EXPORT_SYMBOL(free_pages_exact);
4246 * nr_free_zone_pages - count number of pages beyond high watermark
4247 * @offset: The zone index of the highest zone
4249 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4250 * high watermark within all zones at or below a given zone index. For each
4251 * zone, the number of pages is calculated as:
4252 * managed_pages - high_pages
4254 static unsigned long nr_free_zone_pages(int offset)
4259 /* Just pick one node, since fallback list is circular */
4260 unsigned long sum = 0;
4262 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4264 for_each_zone_zonelist(zone, z, zonelist, offset) {
4265 unsigned long size = zone->managed_pages;
4266 unsigned long high = high_wmark_pages(zone);
4275 * nr_free_buffer_pages - count number of pages beyond high watermark
4277 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4278 * watermark within ZONE_DMA and ZONE_NORMAL.
4280 unsigned long nr_free_buffer_pages(void)
4282 return nr_free_zone_pages(gfp_zone(GFP_USER));
4284 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4287 * nr_free_pagecache_pages - count number of pages beyond high watermark
4289 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4290 * high watermark within all zones.
4292 unsigned long nr_free_pagecache_pages(void)
4294 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4297 static inline void show_node(struct zone *zone)
4299 if (IS_ENABLED(CONFIG_NUMA))
4300 printk("Node %d ", zone_to_nid(zone));
4303 long si_mem_available(void)
4306 unsigned long pagecache;
4307 unsigned long wmark_low = 0;
4308 unsigned long pages[NR_LRU_LISTS];
4312 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4313 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4316 wmark_low += zone->watermark[WMARK_LOW];
4319 * Estimate the amount of memory available for userspace allocations,
4320 * without causing swapping.
4322 available = global_page_state(NR_FREE_PAGES) - totalreserve_pages;
4325 * Not all the page cache can be freed, otherwise the system will
4326 * start swapping. Assume at least half of the page cache, or the
4327 * low watermark worth of cache, needs to stay.
4329 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4330 pagecache -= min(pagecache / 2, wmark_low);
4331 available += pagecache;
4334 * Part of the reclaimable slab consists of items that are in use,
4335 * and cannot be freed. Cap this estimate at the low watermark.
4337 available += global_page_state(NR_SLAB_RECLAIMABLE) -
4338 min(global_page_state(NR_SLAB_RECLAIMABLE) / 2, wmark_low);
4344 EXPORT_SYMBOL_GPL(si_mem_available);
4346 void si_meminfo(struct sysinfo *val)
4348 val->totalram = totalram_pages;
4349 val->sharedram = global_node_page_state(NR_SHMEM);
4350 val->freeram = global_page_state(NR_FREE_PAGES);
4351 val->bufferram = nr_blockdev_pages();
4352 val->totalhigh = totalhigh_pages;
4353 val->freehigh = nr_free_highpages();
4354 val->mem_unit = PAGE_SIZE;
4357 EXPORT_SYMBOL(si_meminfo);
4360 void si_meminfo_node(struct sysinfo *val, int nid)
4362 int zone_type; /* needs to be signed */
4363 unsigned long managed_pages = 0;
4364 unsigned long managed_highpages = 0;
4365 unsigned long free_highpages = 0;
4366 pg_data_t *pgdat = NODE_DATA(nid);
4368 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4369 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4370 val->totalram = managed_pages;
4371 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4372 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4373 #ifdef CONFIG_HIGHMEM
4374 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4375 struct zone *zone = &pgdat->node_zones[zone_type];
4377 if (is_highmem(zone)) {
4378 managed_highpages += zone->managed_pages;
4379 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4382 val->totalhigh = managed_highpages;
4383 val->freehigh = free_highpages;
4385 val->totalhigh = managed_highpages;
4386 val->freehigh = free_highpages;
4388 val->mem_unit = PAGE_SIZE;
4393 * Determine whether the node should be displayed or not, depending on whether
4394 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4396 static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4398 if (!(flags & SHOW_MEM_FILTER_NODES))
4402 * no node mask - aka implicit memory numa policy. Do not bother with
4403 * the synchronization - read_mems_allowed_begin - because we do not
4404 * have to be precise here.
4407 nodemask = &cpuset_current_mems_allowed;
4409 return !node_isset(nid, *nodemask);
4412 #define K(x) ((x) << (PAGE_SHIFT-10))
4414 static void show_migration_types(unsigned char type)
4416 static const char types[MIGRATE_TYPES] = {
4417 [MIGRATE_UNMOVABLE] = 'U',
4418 [MIGRATE_MOVABLE] = 'M',
4419 [MIGRATE_RECLAIMABLE] = 'E',
4420 [MIGRATE_HIGHATOMIC] = 'H',
4422 [MIGRATE_CMA] = 'C',
4424 #ifdef CONFIG_MEMORY_ISOLATION
4425 [MIGRATE_ISOLATE] = 'I',
4428 char tmp[MIGRATE_TYPES + 1];
4432 for (i = 0; i < MIGRATE_TYPES; i++) {
4433 if (type & (1 << i))
4438 printk(KERN_CONT "(%s) ", tmp);
4442 * Show free area list (used inside shift_scroll-lock stuff)
4443 * We also calculate the percentage fragmentation. We do this by counting the
4444 * memory on each free list with the exception of the first item on the list.
4447 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4450 void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4452 unsigned long free_pcp = 0;
4457 for_each_populated_zone(zone) {
4458 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4461 for_each_online_cpu(cpu)
4462 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4465 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4466 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4467 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4468 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4469 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4470 " free:%lu free_pcp:%lu free_cma:%lu\n",
4471 global_node_page_state(NR_ACTIVE_ANON),
4472 global_node_page_state(NR_INACTIVE_ANON),
4473 global_node_page_state(NR_ISOLATED_ANON),
4474 global_node_page_state(NR_ACTIVE_FILE),
4475 global_node_page_state(NR_INACTIVE_FILE),
4476 global_node_page_state(NR_ISOLATED_FILE),
4477 global_node_page_state(NR_UNEVICTABLE),
4478 global_node_page_state(NR_FILE_DIRTY),
4479 global_node_page_state(NR_WRITEBACK),
4480 global_node_page_state(NR_UNSTABLE_NFS),
4481 global_page_state(NR_SLAB_RECLAIMABLE),
4482 global_page_state(NR_SLAB_UNRECLAIMABLE),
4483 global_node_page_state(NR_FILE_MAPPED),
4484 global_node_page_state(NR_SHMEM),
4485 global_page_state(NR_PAGETABLE),
4486 global_page_state(NR_BOUNCE),
4487 global_page_state(NR_FREE_PAGES),
4489 global_page_state(NR_FREE_CMA_PAGES));
4491 for_each_online_pgdat(pgdat) {
4492 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4496 " active_anon:%lukB"
4497 " inactive_anon:%lukB"
4498 " active_file:%lukB"
4499 " inactive_file:%lukB"
4500 " unevictable:%lukB"
4501 " isolated(anon):%lukB"
4502 " isolated(file):%lukB"
4507 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4509 " shmem_pmdmapped: %lukB"
4512 " writeback_tmp:%lukB"
4514 " pages_scanned:%lu"
4515 " all_unreclaimable? %s"
4518 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4519 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4520 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4521 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4522 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4523 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4524 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4525 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4526 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4527 K(node_page_state(pgdat, NR_WRITEBACK)),
4528 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4529 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4530 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4532 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4534 K(node_page_state(pgdat, NR_SHMEM)),
4535 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4536 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4537 node_page_state(pgdat, NR_PAGES_SCANNED),
4538 !pgdat_reclaimable(pgdat) ? "yes" : "no");
4541 for_each_populated_zone(zone) {
4544 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4548 for_each_online_cpu(cpu)
4549 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4558 " active_anon:%lukB"
4559 " inactive_anon:%lukB"
4560 " active_file:%lukB"
4561 " inactive_file:%lukB"
4562 " unevictable:%lukB"
4563 " writepending:%lukB"
4567 " slab_reclaimable:%lukB"
4568 " slab_unreclaimable:%lukB"
4569 " kernel_stack:%lukB"
4577 K(zone_page_state(zone, NR_FREE_PAGES)),
4578 K(min_wmark_pages(zone)),
4579 K(low_wmark_pages(zone)),
4580 K(high_wmark_pages(zone)),
4581 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4582 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4583 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4584 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4585 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4586 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4587 K(zone->present_pages),
4588 K(zone->managed_pages),
4589 K(zone_page_state(zone, NR_MLOCK)),
4590 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
4591 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
4592 zone_page_state(zone, NR_KERNEL_STACK_KB),
4593 K(zone_page_state(zone, NR_PAGETABLE)),
4594 K(zone_page_state(zone, NR_BOUNCE)),
4596 K(this_cpu_read(zone->pageset->pcp.count)),
4597 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4598 printk("lowmem_reserve[]:");
4599 for (i = 0; i < MAX_NR_ZONES; i++)
4600 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4601 printk(KERN_CONT "\n");
4604 for_each_populated_zone(zone) {
4606 unsigned long nr[MAX_ORDER], flags, total = 0;
4607 unsigned char types[MAX_ORDER];
4609 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4612 printk(KERN_CONT "%s: ", zone->name);
4614 spin_lock_irqsave(&zone->lock, flags);
4615 for (order = 0; order < MAX_ORDER; order++) {
4616 struct free_area *area = &zone->free_area[order];
4619 nr[order] = area->nr_free;
4620 total += nr[order] << order;
4623 for (type = 0; type < MIGRATE_TYPES; type++) {
4624 if (!list_empty(&area->free_list[type]))
4625 types[order] |= 1 << type;
4628 spin_unlock_irqrestore(&zone->lock, flags);
4629 for (order = 0; order < MAX_ORDER; order++) {
4630 printk(KERN_CONT "%lu*%lukB ",
4631 nr[order], K(1UL) << order);
4633 show_migration_types(types[order]);
4635 printk(KERN_CONT "= %lukB\n", K(total));
4638 hugetlb_show_meminfo();
4640 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
4642 show_swap_cache_info();
4645 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
4647 zoneref->zone = zone;
4648 zoneref->zone_idx = zone_idx(zone);
4652 * Builds allocation fallback zone lists.
4654 * Add all populated zones of a node to the zonelist.
4656 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
4660 enum zone_type zone_type = MAX_NR_ZONES;
4664 zone = pgdat->node_zones + zone_type;
4665 if (managed_zone(zone)) {
4666 zoneref_set_zone(zone,
4667 &zonelist->_zonerefs[nr_zones++]);
4668 check_highest_zone(zone_type);
4670 } while (zone_type);
4678 * 0 = automatic detection of better ordering.
4679 * 1 = order by ([node] distance, -zonetype)
4680 * 2 = order by (-zonetype, [node] distance)
4682 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
4683 * the same zonelist. So only NUMA can configure this param.
4685 #define ZONELIST_ORDER_DEFAULT 0
4686 #define ZONELIST_ORDER_NODE 1
4687 #define ZONELIST_ORDER_ZONE 2
4689 /* zonelist order in the kernel.
4690 * set_zonelist_order() will set this to NODE or ZONE.
4692 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
4693 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
4697 /* The value user specified ....changed by config */
4698 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4699 /* string for sysctl */
4700 #define NUMA_ZONELIST_ORDER_LEN 16
4701 char numa_zonelist_order[16] = "default";
4704 * interface for configure zonelist ordering.
4705 * command line option "numa_zonelist_order"
4706 * = "[dD]efault - default, automatic configuration.
4707 * = "[nN]ode - order by node locality, then by zone within node
4708 * = "[zZ]one - order by zone, then by locality within zone
4711 static int __parse_numa_zonelist_order(char *s)
4713 if (*s == 'd' || *s == 'D') {
4714 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
4715 } else if (*s == 'n' || *s == 'N') {
4716 user_zonelist_order = ZONELIST_ORDER_NODE;
4717 } else if (*s == 'z' || *s == 'Z') {
4718 user_zonelist_order = ZONELIST_ORDER_ZONE;
4720 pr_warn("Ignoring invalid numa_zonelist_order value: %s\n", s);
4726 static __init int setup_numa_zonelist_order(char *s)
4733 ret = __parse_numa_zonelist_order(s);
4735 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
4739 early_param("numa_zonelist_order", setup_numa_zonelist_order);
4742 * sysctl handler for numa_zonelist_order
4744 int numa_zonelist_order_handler(struct ctl_table *table, int write,
4745 void __user *buffer, size_t *length,
4748 char saved_string[NUMA_ZONELIST_ORDER_LEN];
4750 static DEFINE_MUTEX(zl_order_mutex);
4752 mutex_lock(&zl_order_mutex);
4754 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
4758 strcpy(saved_string, (char *)table->data);
4760 ret = proc_dostring(table, write, buffer, length, ppos);
4764 int oldval = user_zonelist_order;
4766 ret = __parse_numa_zonelist_order((char *)table->data);
4769 * bogus value. restore saved string
4771 strncpy((char *)table->data, saved_string,
4772 NUMA_ZONELIST_ORDER_LEN);
4773 user_zonelist_order = oldval;
4774 } else if (oldval != user_zonelist_order) {
4775 mutex_lock(&zonelists_mutex);
4776 build_all_zonelists(NULL, NULL);
4777 mutex_unlock(&zonelists_mutex);
4781 mutex_unlock(&zl_order_mutex);
4786 #define MAX_NODE_LOAD (nr_online_nodes)
4787 static int node_load[MAX_NUMNODES];
4790 * find_next_best_node - find the next node that should appear in a given node's fallback list
4791 * @node: node whose fallback list we're appending
4792 * @used_node_mask: nodemask_t of already used nodes
4794 * We use a number of factors to determine which is the next node that should
4795 * appear on a given node's fallback list. The node should not have appeared
4796 * already in @node's fallback list, and it should be the next closest node
4797 * according to the distance array (which contains arbitrary distance values
4798 * from each node to each node in the system), and should also prefer nodes
4799 * with no CPUs, since presumably they'll have very little allocation pressure
4800 * on them otherwise.
4801 * It returns -1 if no node is found.
4803 static int find_next_best_node(int node, nodemask_t *used_node_mask)
4806 int min_val = INT_MAX;
4807 int best_node = NUMA_NO_NODE;
4808 const struct cpumask *tmp = cpumask_of_node(0);
4810 /* Use the local node if we haven't already */
4811 if (!node_isset(node, *used_node_mask)) {
4812 node_set(node, *used_node_mask);
4816 for_each_node_state(n, N_MEMORY) {
4818 /* Don't want a node to appear more than once */
4819 if (node_isset(n, *used_node_mask))
4822 /* Use the distance array to find the distance */
4823 val = node_distance(node, n);
4825 /* Penalize nodes under us ("prefer the next node") */
4828 /* Give preference to headless and unused nodes */
4829 tmp = cpumask_of_node(n);
4830 if (!cpumask_empty(tmp))
4831 val += PENALTY_FOR_NODE_WITH_CPUS;
4833 /* Slight preference for less loaded node */
4834 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
4835 val += node_load[n];
4837 if (val < min_val) {
4844 node_set(best_node, *used_node_mask);
4851 * Build zonelists ordered by node and zones within node.
4852 * This results in maximum locality--normal zone overflows into local
4853 * DMA zone, if any--but risks exhausting DMA zone.
4855 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
4858 struct zonelist *zonelist;
4860 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4861 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
4863 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
4864 zonelist->_zonerefs[j].zone = NULL;
4865 zonelist->_zonerefs[j].zone_idx = 0;
4869 * Build gfp_thisnode zonelists
4871 static void build_thisnode_zonelists(pg_data_t *pgdat)
4874 struct zonelist *zonelist;
4876 zonelist = &pgdat->node_zonelists[ZONELIST_NOFALLBACK];
4877 j = build_zonelists_node(pgdat, zonelist, 0);
4878 zonelist->_zonerefs[j].zone = NULL;
4879 zonelist->_zonerefs[j].zone_idx = 0;
4883 * Build zonelists ordered by zone and nodes within zones.
4884 * This results in conserving DMA zone[s] until all Normal memory is
4885 * exhausted, but results in overflowing to remote node while memory
4886 * may still exist in local DMA zone.
4888 static int node_order[MAX_NUMNODES];
4890 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
4893 int zone_type; /* needs to be signed */
4895 struct zonelist *zonelist;
4897 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
4899 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
4900 for (j = 0; j < nr_nodes; j++) {
4901 node = node_order[j];
4902 z = &NODE_DATA(node)->node_zones[zone_type];
4903 if (managed_zone(z)) {
4905 &zonelist->_zonerefs[pos++]);
4906 check_highest_zone(zone_type);
4910 zonelist->_zonerefs[pos].zone = NULL;
4911 zonelist->_zonerefs[pos].zone_idx = 0;
4914 #if defined(CONFIG_64BIT)
4916 * Devices that require DMA32/DMA are relatively rare and do not justify a
4917 * penalty to every machine in case the specialised case applies. Default
4918 * to Node-ordering on 64-bit NUMA machines
4920 static int default_zonelist_order(void)
4922 return ZONELIST_ORDER_NODE;
4926 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4927 * by the kernel. If processes running on node 0 deplete the low memory zone
4928 * then reclaim will occur more frequency increasing stalls and potentially
4929 * be easier to OOM if a large percentage of the zone is under writeback or
4930 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4931 * Hence, default to zone ordering on 32-bit.
4933 static int default_zonelist_order(void)
4935 return ZONELIST_ORDER_ZONE;
4937 #endif /* CONFIG_64BIT */
4939 static void set_zonelist_order(void)
4941 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
4942 current_zonelist_order = default_zonelist_order();
4944 current_zonelist_order = user_zonelist_order;
4947 static void build_zonelists(pg_data_t *pgdat)
4950 nodemask_t used_mask;
4951 int local_node, prev_node;
4952 struct zonelist *zonelist;
4953 unsigned int order = current_zonelist_order;
4955 /* initialize zonelists */
4956 for (i = 0; i < MAX_ZONELISTS; i++) {
4957 zonelist = pgdat->node_zonelists + i;
4958 zonelist->_zonerefs[0].zone = NULL;
4959 zonelist->_zonerefs[0].zone_idx = 0;
4962 /* NUMA-aware ordering of nodes */
4963 local_node = pgdat->node_id;
4964 load = nr_online_nodes;
4965 prev_node = local_node;
4966 nodes_clear(used_mask);
4968 memset(node_order, 0, sizeof(node_order));
4971 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
4973 * We don't want to pressure a particular node.
4974 * So adding penalty to the first node in same
4975 * distance group to make it round-robin.
4977 if (node_distance(local_node, node) !=
4978 node_distance(local_node, prev_node))
4979 node_load[node] = load;
4983 if (order == ZONELIST_ORDER_NODE)
4984 build_zonelists_in_node_order(pgdat, node);
4986 node_order[i++] = node; /* remember order */
4989 if (order == ZONELIST_ORDER_ZONE) {
4990 /* calculate node order -- i.e., DMA last! */
4991 build_zonelists_in_zone_order(pgdat, i);
4994 build_thisnode_zonelists(pgdat);
4997 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4999 * Return node id of node used for "local" allocations.
5000 * I.e., first node id of first zone in arg node's generic zonelist.
5001 * Used for initializing percpu 'numa_mem', which is used primarily
5002 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5004 int local_memory_node(int node)
5008 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5009 gfp_zone(GFP_KERNEL),
5011 return z->zone->node;
5015 static void setup_min_unmapped_ratio(void);
5016 static void setup_min_slab_ratio(void);
5017 #else /* CONFIG_NUMA */
5019 static void set_zonelist_order(void)
5021 current_zonelist_order = ZONELIST_ORDER_ZONE;
5024 static void build_zonelists(pg_data_t *pgdat)
5026 int node, local_node;
5028 struct zonelist *zonelist;
5030 local_node = pgdat->node_id;
5032 zonelist = &pgdat->node_zonelists[ZONELIST_FALLBACK];
5033 j = build_zonelists_node(pgdat, zonelist, 0);
5036 * Now we build the zonelist so that it contains the zones
5037 * of all the other nodes.
5038 * We don't want to pressure a particular node, so when
5039 * building the zones for node N, we make sure that the
5040 * zones coming right after the local ones are those from
5041 * node N+1 (modulo N)
5043 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5044 if (!node_online(node))
5046 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5048 for (node = 0; node < local_node; node++) {
5049 if (!node_online(node))
5051 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
5054 zonelist->_zonerefs[j].zone = NULL;
5055 zonelist->_zonerefs[j].zone_idx = 0;
5058 #endif /* CONFIG_NUMA */
5061 * Boot pageset table. One per cpu which is going to be used for all
5062 * zones and all nodes. The parameters will be set in such a way
5063 * that an item put on a list will immediately be handed over to
5064 * the buddy list. This is safe since pageset manipulation is done
5065 * with interrupts disabled.
5067 * The boot_pagesets must be kept even after bootup is complete for
5068 * unused processors and/or zones. They do play a role for bootstrapping
5069 * hotplugged processors.
5071 * zoneinfo_show() and maybe other functions do
5072 * not check if the processor is online before following the pageset pointer.
5073 * Other parts of the kernel may not check if the zone is available.
5075 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5076 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5077 static void setup_zone_pageset(struct zone *zone);
5080 * Global mutex to protect against size modification of zonelists
5081 * as well as to serialize pageset setup for the new populated zone.
5083 DEFINE_MUTEX(zonelists_mutex);
5085 /* return values int ....just for stop_machine() */
5086 static int __build_all_zonelists(void *data)
5090 pg_data_t *self = data;
5093 memset(node_load, 0, sizeof(node_load));
5096 if (self && !node_online(self->node_id)) {
5097 build_zonelists(self);
5100 for_each_online_node(nid) {
5101 pg_data_t *pgdat = NODE_DATA(nid);
5103 build_zonelists(pgdat);
5107 * Initialize the boot_pagesets that are going to be used
5108 * for bootstrapping processors. The real pagesets for
5109 * each zone will be allocated later when the per cpu
5110 * allocator is available.
5112 * boot_pagesets are used also for bootstrapping offline
5113 * cpus if the system is already booted because the pagesets
5114 * are needed to initialize allocators on a specific cpu too.
5115 * F.e. the percpu allocator needs the page allocator which
5116 * needs the percpu allocator in order to allocate its pagesets
5117 * (a chicken-egg dilemma).
5119 for_each_possible_cpu(cpu) {
5120 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5122 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
5124 * We now know the "local memory node" for each node--
5125 * i.e., the node of the first zone in the generic zonelist.
5126 * Set up numa_mem percpu variable for on-line cpus. During
5127 * boot, only the boot cpu should be on-line; we'll init the
5128 * secondary cpus' numa_mem as they come on-line. During
5129 * node/memory hotplug, we'll fixup all on-line cpus.
5131 if (cpu_online(cpu))
5132 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5139 static noinline void __init
5140 build_all_zonelists_init(void)
5142 __build_all_zonelists(NULL);
5143 mminit_verify_zonelist();
5144 cpuset_init_current_mems_allowed();
5148 * Called with zonelists_mutex held always
5149 * unless system_state == SYSTEM_BOOTING.
5151 * __ref due to (1) call of __meminit annotated setup_zone_pageset
5152 * [we're only called with non-NULL zone through __meminit paths] and
5153 * (2) call of __init annotated helper build_all_zonelists_init
5154 * [protected by SYSTEM_BOOTING].
5156 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
5158 set_zonelist_order();
5160 if (system_state == SYSTEM_BOOTING) {
5161 build_all_zonelists_init();
5163 #ifdef CONFIG_MEMORY_HOTPLUG
5165 setup_zone_pageset(zone);
5167 /* we have to stop all cpus to guarantee there is no user
5169 stop_machine(__build_all_zonelists, pgdat, NULL);
5170 /* cpuset refresh routine should be here */
5172 vm_total_pages = nr_free_pagecache_pages();
5174 * Disable grouping by mobility if the number of pages in the
5175 * system is too low to allow the mechanism to work. It would be
5176 * more accurate, but expensive to check per-zone. This check is
5177 * made on memory-hotadd so a system can start with mobility
5178 * disabled and enable it later
5180 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5181 page_group_by_mobility_disabled = 1;
5183 page_group_by_mobility_disabled = 0;
5185 pr_info("Built %i zonelists in %s order, mobility grouping %s. Total pages: %ld\n",
5187 zonelist_order_name[current_zonelist_order],
5188 page_group_by_mobility_disabled ? "off" : "on",
5191 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5196 * Initially all pages are reserved - free ones are freed
5197 * up by free_all_bootmem() once the early boot process is
5198 * done. Non-atomic initialization, single-pass.
5200 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5201 unsigned long start_pfn, enum memmap_context context)
5203 struct vmem_altmap *altmap = to_vmem_altmap(__pfn_to_phys(start_pfn));
5204 unsigned long end_pfn = start_pfn + size;
5205 pg_data_t *pgdat = NODE_DATA(nid);
5207 unsigned long nr_initialised = 0;
5208 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5209 struct memblock_region *r = NULL, *tmp;
5212 if (highest_memmap_pfn < end_pfn - 1)
5213 highest_memmap_pfn = end_pfn - 1;
5216 * Honor reservation requested by the driver for this ZONE_DEVICE
5219 if (altmap && start_pfn == altmap->base_pfn)
5220 start_pfn += altmap->reserve;
5222 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5224 * There can be holes in boot-time mem_map[]s handed to this
5225 * function. They do not exist on hotplugged memory.
5227 if (context != MEMMAP_EARLY)
5230 if (!early_pfn_valid(pfn)) {
5231 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5233 * Skip to the pfn preceding the next valid one (or
5234 * end_pfn), such that we hit a valid pfn (or end_pfn)
5235 * on our next iteration of the loop.
5237 pfn = memblock_next_valid_pfn(pfn, end_pfn) - 1;
5241 if (!early_pfn_in_nid(pfn, nid))
5243 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5246 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5248 * Check given memblock attribute by firmware which can affect
5249 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5250 * mirrored, it's an overlapped memmap init. skip it.
5252 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5253 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5254 for_each_memblock(memory, tmp)
5255 if (pfn < memblock_region_memory_end_pfn(tmp))
5259 if (pfn >= memblock_region_memory_base_pfn(r) &&
5260 memblock_is_mirror(r)) {
5261 /* already initialized as NORMAL */
5262 pfn = memblock_region_memory_end_pfn(r);
5270 * Mark the block movable so that blocks are reserved for
5271 * movable at startup. This will force kernel allocations
5272 * to reserve their blocks rather than leaking throughout
5273 * the address space during boot when many long-lived
5274 * kernel allocations are made.
5276 * bitmap is created for zone's valid pfn range. but memmap
5277 * can be created for invalid pages (for alignment)
5278 * check here not to call set_pageblock_migratetype() against
5281 if (!(pfn & (pageblock_nr_pages - 1))) {
5282 struct page *page = pfn_to_page(pfn);
5284 __init_single_page(page, pfn, zone, nid);
5285 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5287 __init_single_pfn(pfn, zone, nid);
5292 static void __meminit zone_init_free_lists(struct zone *zone)
5294 unsigned int order, t;
5295 for_each_migratetype_order(order, t) {
5296 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5297 zone->free_area[order].nr_free = 0;
5301 #ifndef __HAVE_ARCH_MEMMAP_INIT
5302 #define memmap_init(size, nid, zone, start_pfn) \
5303 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
5306 static int zone_batchsize(struct zone *zone)
5312 * The per-cpu-pages pools are set to around 1000th of the
5313 * size of the zone. But no more than 1/2 of a meg.
5315 * OK, so we don't know how big the cache is. So guess.
5317 batch = zone->managed_pages / 1024;
5318 if (batch * PAGE_SIZE > 512 * 1024)
5319 batch = (512 * 1024) / PAGE_SIZE;
5320 batch /= 4; /* We effectively *= 4 below */
5325 * Clamp the batch to a 2^n - 1 value. Having a power
5326 * of 2 value was found to be more likely to have
5327 * suboptimal cache aliasing properties in some cases.
5329 * For example if 2 tasks are alternately allocating
5330 * batches of pages, one task can end up with a lot
5331 * of pages of one half of the possible page colors
5332 * and the other with pages of the other colors.
5334 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5339 /* The deferral and batching of frees should be suppressed under NOMMU
5342 * The problem is that NOMMU needs to be able to allocate large chunks
5343 * of contiguous memory as there's no hardware page translation to
5344 * assemble apparent contiguous memory from discontiguous pages.
5346 * Queueing large contiguous runs of pages for batching, however,
5347 * causes the pages to actually be freed in smaller chunks. As there
5348 * can be a significant delay between the individual batches being
5349 * recycled, this leads to the once large chunks of space being
5350 * fragmented and becoming unavailable for high-order allocations.
5357 * pcp->high and pcp->batch values are related and dependent on one another:
5358 * ->batch must never be higher then ->high.
5359 * The following function updates them in a safe manner without read side
5362 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5363 * those fields changing asynchronously (acording the the above rule).
5365 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5366 * outside of boot time (or some other assurance that no concurrent updaters
5369 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5370 unsigned long batch)
5372 /* start with a fail safe value for batch */
5376 /* Update high, then batch, in order */
5383 /* a companion to pageset_set_high() */
5384 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5386 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5389 static void pageset_init(struct per_cpu_pageset *p)
5391 struct per_cpu_pages *pcp;
5394 memset(p, 0, sizeof(*p));
5398 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5399 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5402 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5405 pageset_set_batch(p, batch);
5409 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5410 * to the value high for the pageset p.
5412 static void pageset_set_high(struct per_cpu_pageset *p,
5415 unsigned long batch = max(1UL, high / 4);
5416 if ((high / 4) > (PAGE_SHIFT * 8))
5417 batch = PAGE_SHIFT * 8;
5419 pageset_update(&p->pcp, high, batch);
5422 static void pageset_set_high_and_batch(struct zone *zone,
5423 struct per_cpu_pageset *pcp)
5425 if (percpu_pagelist_fraction)
5426 pageset_set_high(pcp,
5427 (zone->managed_pages /
5428 percpu_pagelist_fraction));
5430 pageset_set_batch(pcp, zone_batchsize(zone));
5433 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5435 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5438 pageset_set_high_and_batch(zone, pcp);
5441 static void __meminit setup_zone_pageset(struct zone *zone)
5444 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5445 for_each_possible_cpu(cpu)
5446 zone_pageset_init(zone, cpu);
5450 * Allocate per cpu pagesets and initialize them.
5451 * Before this call only boot pagesets were available.
5453 void __init setup_per_cpu_pageset(void)
5455 struct pglist_data *pgdat;
5458 for_each_populated_zone(zone)
5459 setup_zone_pageset(zone);
5461 for_each_online_pgdat(pgdat)
5462 pgdat->per_cpu_nodestats =
5463 alloc_percpu(struct per_cpu_nodestat);
5466 static __meminit void zone_pcp_init(struct zone *zone)
5469 * per cpu subsystem is not up at this point. The following code
5470 * relies on the ability of the linker to provide the
5471 * offset of a (static) per cpu variable into the per cpu area.
5473 zone->pageset = &boot_pageset;
5475 if (populated_zone(zone))
5476 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5477 zone->name, zone->present_pages,
5478 zone_batchsize(zone));
5481 int __meminit init_currently_empty_zone(struct zone *zone,
5482 unsigned long zone_start_pfn,
5485 struct pglist_data *pgdat = zone->zone_pgdat;
5487 pgdat->nr_zones = zone_idx(zone) + 1;
5489 zone->zone_start_pfn = zone_start_pfn;
5491 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5492 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5494 (unsigned long)zone_idx(zone),
5495 zone_start_pfn, (zone_start_pfn + size));
5497 zone_init_free_lists(zone);
5498 zone->initialized = 1;
5503 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5504 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5507 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5509 int __meminit __early_pfn_to_nid(unsigned long pfn,
5510 struct mminit_pfnnid_cache *state)
5512 unsigned long start_pfn, end_pfn;
5515 if (state->last_start <= pfn && pfn < state->last_end)
5516 return state->last_nid;
5518 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5520 state->last_start = start_pfn;
5521 state->last_end = end_pfn;
5522 state->last_nid = nid;
5527 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5530 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5531 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5532 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5534 * If an architecture guarantees that all ranges registered contain no holes
5535 * and may be freed, this this function may be used instead of calling
5536 * memblock_free_early_nid() manually.
5538 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5540 unsigned long start_pfn, end_pfn;
5543 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5544 start_pfn = min(start_pfn, max_low_pfn);
5545 end_pfn = min(end_pfn, max_low_pfn);
5547 if (start_pfn < end_pfn)
5548 memblock_free_early_nid(PFN_PHYS(start_pfn),
5549 (end_pfn - start_pfn) << PAGE_SHIFT,
5555 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5556 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5558 * If an architecture guarantees that all ranges registered contain no holes and may
5559 * be freed, this function may be used instead of calling memory_present() manually.
5561 void __init sparse_memory_present_with_active_regions(int nid)
5563 unsigned long start_pfn, end_pfn;
5566 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5567 memory_present(this_nid, start_pfn, end_pfn);
5571 * get_pfn_range_for_nid - Return the start and end page frames for a node
5572 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5573 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5574 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5576 * It returns the start and end page frame of a node based on information
5577 * provided by memblock_set_node(). If called for a node
5578 * with no available memory, a warning is printed and the start and end
5581 void __meminit get_pfn_range_for_nid(unsigned int nid,
5582 unsigned long *start_pfn, unsigned long *end_pfn)
5584 unsigned long this_start_pfn, this_end_pfn;
5590 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5591 *start_pfn = min(*start_pfn, this_start_pfn);
5592 *end_pfn = max(*end_pfn, this_end_pfn);
5595 if (*start_pfn == -1UL)
5600 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5601 * assumption is made that zones within a node are ordered in monotonic
5602 * increasing memory addresses so that the "highest" populated zone is used
5604 static void __init find_usable_zone_for_movable(void)
5607 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5608 if (zone_index == ZONE_MOVABLE)
5611 if (arch_zone_highest_possible_pfn[zone_index] >
5612 arch_zone_lowest_possible_pfn[zone_index])
5616 VM_BUG_ON(zone_index == -1);
5617 movable_zone = zone_index;
5621 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5622 * because it is sized independent of architecture. Unlike the other zones,
5623 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5624 * in each node depending on the size of each node and how evenly kernelcore
5625 * is distributed. This helper function adjusts the zone ranges
5626 * provided by the architecture for a given node by using the end of the
5627 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5628 * zones within a node are in order of monotonic increases memory addresses
5630 static void __meminit adjust_zone_range_for_zone_movable(int nid,
5631 unsigned long zone_type,
5632 unsigned long node_start_pfn,
5633 unsigned long node_end_pfn,
5634 unsigned long *zone_start_pfn,
5635 unsigned long *zone_end_pfn)
5637 /* Only adjust if ZONE_MOVABLE is on this node */
5638 if (zone_movable_pfn[nid]) {
5639 /* Size ZONE_MOVABLE */
5640 if (zone_type == ZONE_MOVABLE) {
5641 *zone_start_pfn = zone_movable_pfn[nid];
5642 *zone_end_pfn = min(node_end_pfn,
5643 arch_zone_highest_possible_pfn[movable_zone]);
5645 /* Adjust for ZONE_MOVABLE starting within this range */
5646 } else if (!mirrored_kernelcore &&
5647 *zone_start_pfn < zone_movable_pfn[nid] &&
5648 *zone_end_pfn > zone_movable_pfn[nid]) {
5649 *zone_end_pfn = zone_movable_pfn[nid];
5651 /* Check if this whole range is within ZONE_MOVABLE */
5652 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5653 *zone_start_pfn = *zone_end_pfn;
5658 * Return the number of pages a zone spans in a node, including holes
5659 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5661 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5662 unsigned long zone_type,
5663 unsigned long node_start_pfn,
5664 unsigned long node_end_pfn,
5665 unsigned long *zone_start_pfn,
5666 unsigned long *zone_end_pfn,
5667 unsigned long *ignored)
5669 /* When hotadd a new node from cpu_up(), the node should be empty */
5670 if (!node_start_pfn && !node_end_pfn)
5673 /* Get the start and end of the zone */
5674 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5675 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5676 adjust_zone_range_for_zone_movable(nid, zone_type,
5677 node_start_pfn, node_end_pfn,
5678 zone_start_pfn, zone_end_pfn);
5680 /* Check that this node has pages within the zone's required range */
5681 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5684 /* Move the zone boundaries inside the node if necessary */
5685 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5686 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5688 /* Return the spanned pages */
5689 return *zone_end_pfn - *zone_start_pfn;
5693 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5694 * then all holes in the requested range will be accounted for.
5696 unsigned long __meminit __absent_pages_in_range(int nid,
5697 unsigned long range_start_pfn,
5698 unsigned long range_end_pfn)
5700 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5701 unsigned long start_pfn, end_pfn;
5704 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5705 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5706 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5707 nr_absent -= end_pfn - start_pfn;
5713 * absent_pages_in_range - Return number of page frames in holes within a range
5714 * @start_pfn: The start PFN to start searching for holes
5715 * @end_pfn: The end PFN to stop searching for holes
5717 * It returns the number of pages frames in memory holes within a range.
5719 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5720 unsigned long end_pfn)
5722 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5725 /* Return the number of page frames in holes in a zone on a node */
5726 static unsigned long __meminit zone_absent_pages_in_node(int nid,
5727 unsigned long zone_type,
5728 unsigned long node_start_pfn,
5729 unsigned long node_end_pfn,
5730 unsigned long *ignored)
5732 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5733 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5734 unsigned long zone_start_pfn, zone_end_pfn;
5735 unsigned long nr_absent;
5737 /* When hotadd a new node from cpu_up(), the node should be empty */
5738 if (!node_start_pfn && !node_end_pfn)
5741 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5742 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5744 adjust_zone_range_for_zone_movable(nid, zone_type,
5745 node_start_pfn, node_end_pfn,
5746 &zone_start_pfn, &zone_end_pfn);
5747 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5750 * ZONE_MOVABLE handling.
5751 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
5754 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
5755 unsigned long start_pfn, end_pfn;
5756 struct memblock_region *r;
5758 for_each_memblock(memory, r) {
5759 start_pfn = clamp(memblock_region_memory_base_pfn(r),
5760 zone_start_pfn, zone_end_pfn);
5761 end_pfn = clamp(memblock_region_memory_end_pfn(r),
5762 zone_start_pfn, zone_end_pfn);
5764 if (zone_type == ZONE_MOVABLE &&
5765 memblock_is_mirror(r))
5766 nr_absent += end_pfn - start_pfn;
5768 if (zone_type == ZONE_NORMAL &&
5769 !memblock_is_mirror(r))
5770 nr_absent += end_pfn - start_pfn;
5777 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5778 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
5779 unsigned long zone_type,
5780 unsigned long node_start_pfn,
5781 unsigned long node_end_pfn,
5782 unsigned long *zone_start_pfn,
5783 unsigned long *zone_end_pfn,
5784 unsigned long *zones_size)
5788 *zone_start_pfn = node_start_pfn;
5789 for (zone = 0; zone < zone_type; zone++)
5790 *zone_start_pfn += zones_size[zone];
5792 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
5794 return zones_size[zone_type];
5797 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
5798 unsigned long zone_type,
5799 unsigned long node_start_pfn,
5800 unsigned long node_end_pfn,
5801 unsigned long *zholes_size)
5806 return zholes_size[zone_type];
5809 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5811 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
5812 unsigned long node_start_pfn,
5813 unsigned long node_end_pfn,
5814 unsigned long *zones_size,
5815 unsigned long *zholes_size)
5817 unsigned long realtotalpages = 0, totalpages = 0;
5820 for (i = 0; i < MAX_NR_ZONES; i++) {
5821 struct zone *zone = pgdat->node_zones + i;
5822 unsigned long zone_start_pfn, zone_end_pfn;
5823 unsigned long size, real_size;
5825 size = zone_spanned_pages_in_node(pgdat->node_id, i,
5831 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
5832 node_start_pfn, node_end_pfn,
5835 zone->zone_start_pfn = zone_start_pfn;
5837 zone->zone_start_pfn = 0;
5838 zone->spanned_pages = size;
5839 zone->present_pages = real_size;
5842 realtotalpages += real_size;
5845 pgdat->node_spanned_pages = totalpages;
5846 pgdat->node_present_pages = realtotalpages;
5847 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
5851 #ifndef CONFIG_SPARSEMEM
5853 * Calculate the size of the zone->blockflags rounded to an unsigned long
5854 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5855 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5856 * round what is now in bits to nearest long in bits, then return it in
5859 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
5861 unsigned long usemapsize;
5863 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
5864 usemapsize = roundup(zonesize, pageblock_nr_pages);
5865 usemapsize = usemapsize >> pageblock_order;
5866 usemapsize *= NR_PAGEBLOCK_BITS;
5867 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
5869 return usemapsize / 8;
5872 static void __init setup_usemap(struct pglist_data *pgdat,
5874 unsigned long zone_start_pfn,
5875 unsigned long zonesize)
5877 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
5878 zone->pageblock_flags = NULL;
5880 zone->pageblock_flags =
5881 memblock_virt_alloc_node_nopanic(usemapsize,
5885 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
5886 unsigned long zone_start_pfn, unsigned long zonesize) {}
5887 #endif /* CONFIG_SPARSEMEM */
5889 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5891 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5892 void __paginginit set_pageblock_order(void)
5896 /* Check that pageblock_nr_pages has not already been setup */
5897 if (pageblock_order)
5900 if (HPAGE_SHIFT > PAGE_SHIFT)
5901 order = HUGETLB_PAGE_ORDER;
5903 order = MAX_ORDER - 1;
5906 * Assume the largest contiguous order of interest is a huge page.
5907 * This value may be variable depending on boot parameters on IA64 and
5910 pageblock_order = order;
5912 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5915 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5916 * is unused as pageblock_order is set at compile-time. See
5917 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5920 void __paginginit set_pageblock_order(void)
5924 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5926 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
5927 unsigned long present_pages)
5929 unsigned long pages = spanned_pages;
5932 * Provide a more accurate estimation if there are holes within
5933 * the zone and SPARSEMEM is in use. If there are holes within the
5934 * zone, each populated memory region may cost us one or two extra
5935 * memmap pages due to alignment because memmap pages for each
5936 * populated regions may not naturally algined on page boundary.
5937 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5939 if (spanned_pages > present_pages + (present_pages >> 4) &&
5940 IS_ENABLED(CONFIG_SPARSEMEM))
5941 pages = present_pages;
5943 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
5947 * Set up the zone data structures:
5948 * - mark all pages reserved
5949 * - mark all memory queues empty
5950 * - clear the memory bitmaps
5952 * NOTE: pgdat should get zeroed by caller.
5954 static void __paginginit free_area_init_core(struct pglist_data *pgdat)
5957 int nid = pgdat->node_id;
5960 pgdat_resize_init(pgdat);
5961 #ifdef CONFIG_NUMA_BALANCING
5962 spin_lock_init(&pgdat->numabalancing_migrate_lock);
5963 pgdat->numabalancing_migrate_nr_pages = 0;
5964 pgdat->numabalancing_migrate_next_window = jiffies;
5966 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5967 spin_lock_init(&pgdat->split_queue_lock);
5968 INIT_LIST_HEAD(&pgdat->split_queue);
5969 pgdat->split_queue_len = 0;
5971 init_waitqueue_head(&pgdat->kswapd_wait);
5972 init_waitqueue_head(&pgdat->pfmemalloc_wait);
5973 #ifdef CONFIG_COMPACTION
5974 init_waitqueue_head(&pgdat->kcompactd_wait);
5976 pgdat_page_ext_init(pgdat);
5977 spin_lock_init(&pgdat->lru_lock);
5978 lruvec_init(node_lruvec(pgdat));
5980 for (j = 0; j < MAX_NR_ZONES; j++) {
5981 struct zone *zone = pgdat->node_zones + j;
5982 unsigned long size, realsize, freesize, memmap_pages;
5983 unsigned long zone_start_pfn = zone->zone_start_pfn;
5985 size = zone->spanned_pages;
5986 realsize = freesize = zone->present_pages;
5989 * Adjust freesize so that it accounts for how much memory
5990 * is used by this zone for memmap. This affects the watermark
5991 * and per-cpu initialisations
5993 memmap_pages = calc_memmap_size(size, realsize);
5994 if (!is_highmem_idx(j)) {
5995 if (freesize >= memmap_pages) {
5996 freesize -= memmap_pages;
5999 " %s zone: %lu pages used for memmap\n",
6000 zone_names[j], memmap_pages);
6002 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6003 zone_names[j], memmap_pages, freesize);
6006 /* Account for reserved pages */
6007 if (j == 0 && freesize > dma_reserve) {
6008 freesize -= dma_reserve;
6009 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6010 zone_names[0], dma_reserve);
6013 if (!is_highmem_idx(j))
6014 nr_kernel_pages += freesize;
6015 /* Charge for highmem memmap if there are enough kernel pages */
6016 else if (nr_kernel_pages > memmap_pages * 2)
6017 nr_kernel_pages -= memmap_pages;
6018 nr_all_pages += freesize;
6021 * Set an approximate value for lowmem here, it will be adjusted
6022 * when the bootmem allocator frees pages into the buddy system.
6023 * And all highmem pages will be managed by the buddy system.
6025 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6029 zone->name = zone_names[j];
6030 zone->zone_pgdat = pgdat;
6031 spin_lock_init(&zone->lock);
6032 zone_seqlock_init(zone);
6033 zone_pcp_init(zone);
6038 set_pageblock_order();
6039 setup_usemap(pgdat, zone, zone_start_pfn, size);
6040 ret = init_currently_empty_zone(zone, zone_start_pfn, size);
6042 memmap_init(size, nid, j, zone_start_pfn);
6046 static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6048 unsigned long __maybe_unused start = 0;
6049 unsigned long __maybe_unused offset = 0;
6051 /* Skip empty nodes */
6052 if (!pgdat->node_spanned_pages)
6055 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6056 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6057 offset = pgdat->node_start_pfn - start;
6058 /* ia64 gets its own node_mem_map, before this, without bootmem */
6059 if (!pgdat->node_mem_map) {
6060 unsigned long size, end;
6064 * The zone's endpoints aren't required to be MAX_ORDER
6065 * aligned but the node_mem_map endpoints must be in order
6066 * for the buddy allocator to function correctly.
6068 end = pgdat_end_pfn(pgdat);
6069 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6070 size = (end - start) * sizeof(struct page);
6071 map = alloc_remap(pgdat->node_id, size);
6073 map = memblock_virt_alloc_node_nopanic(size,
6075 pgdat->node_mem_map = map + offset;
6077 #ifndef CONFIG_NEED_MULTIPLE_NODES
6079 * With no DISCONTIG, the global mem_map is just set as node 0's
6081 if (pgdat == NODE_DATA(0)) {
6082 mem_map = NODE_DATA(0)->node_mem_map;
6083 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6084 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6086 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6089 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
6092 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6093 unsigned long node_start_pfn, unsigned long *zholes_size)
6095 pg_data_t *pgdat = NODE_DATA(nid);
6096 unsigned long start_pfn = 0;
6097 unsigned long end_pfn = 0;
6099 /* pg_data_t should be reset to zero when it's allocated */
6100 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6102 reset_deferred_meminit(pgdat);
6103 pgdat->node_id = nid;
6104 pgdat->node_start_pfn = node_start_pfn;
6105 pgdat->per_cpu_nodestats = NULL;
6106 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6107 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6108 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6109 (u64)start_pfn << PAGE_SHIFT,
6110 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6112 start_pfn = node_start_pfn;
6114 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6115 zones_size, zholes_size);
6117 alloc_node_mem_map(pgdat);
6118 #ifdef CONFIG_FLAT_NODE_MEM_MAP
6119 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
6120 nid, (unsigned long)pgdat,
6121 (unsigned long)pgdat->node_mem_map);
6124 free_area_init_core(pgdat);
6127 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6129 #if MAX_NUMNODES > 1
6131 * Figure out the number of possible node ids.
6133 void __init setup_nr_node_ids(void)
6135 unsigned int highest;
6137 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6138 nr_node_ids = highest + 1;
6143 * node_map_pfn_alignment - determine the maximum internode alignment
6145 * This function should be called after node map is populated and sorted.
6146 * It calculates the maximum power of two alignment which can distinguish
6149 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6150 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6151 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6152 * shifted, 1GiB is enough and this function will indicate so.
6154 * This is used to test whether pfn -> nid mapping of the chosen memory
6155 * model has fine enough granularity to avoid incorrect mapping for the
6156 * populated node map.
6158 * Returns the determined alignment in pfn's. 0 if there is no alignment
6159 * requirement (single node).
6161 unsigned long __init node_map_pfn_alignment(void)
6163 unsigned long accl_mask = 0, last_end = 0;
6164 unsigned long start, end, mask;
6168 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6169 if (!start || last_nid < 0 || last_nid == nid) {
6176 * Start with a mask granular enough to pin-point to the
6177 * start pfn and tick off bits one-by-one until it becomes
6178 * too coarse to separate the current node from the last.
6180 mask = ~((1 << __ffs(start)) - 1);
6181 while (mask && last_end <= (start & (mask << 1)))
6184 /* accumulate all internode masks */
6188 /* convert mask to number of pages */
6189 return ~accl_mask + 1;
6192 /* Find the lowest pfn for a node */
6193 static unsigned long __init find_min_pfn_for_node(int nid)
6195 unsigned long min_pfn = ULONG_MAX;
6196 unsigned long start_pfn;
6199 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6200 min_pfn = min(min_pfn, start_pfn);
6202 if (min_pfn == ULONG_MAX) {
6203 pr_warn("Could not find start_pfn for node %d\n", nid);
6211 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6213 * It returns the minimum PFN based on information provided via
6214 * memblock_set_node().
6216 unsigned long __init find_min_pfn_with_active_regions(void)
6218 return find_min_pfn_for_node(MAX_NUMNODES);
6222 * early_calculate_totalpages()
6223 * Sum pages in active regions for movable zone.
6224 * Populate N_MEMORY for calculating usable_nodes.
6226 static unsigned long __init early_calculate_totalpages(void)
6228 unsigned long totalpages = 0;
6229 unsigned long start_pfn, end_pfn;
6232 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6233 unsigned long pages = end_pfn - start_pfn;
6235 totalpages += pages;
6237 node_set_state(nid, N_MEMORY);
6243 * Find the PFN the Movable zone begins in each node. Kernel memory
6244 * is spread evenly between nodes as long as the nodes have enough
6245 * memory. When they don't, some nodes will have more kernelcore than
6248 static void __init find_zone_movable_pfns_for_nodes(void)
6251 unsigned long usable_startpfn;
6252 unsigned long kernelcore_node, kernelcore_remaining;
6253 /* save the state before borrow the nodemask */
6254 nodemask_t saved_node_state = node_states[N_MEMORY];
6255 unsigned long totalpages = early_calculate_totalpages();
6256 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6257 struct memblock_region *r;
6259 /* Need to find movable_zone earlier when movable_node is specified. */
6260 find_usable_zone_for_movable();
6263 * If movable_node is specified, ignore kernelcore and movablecore
6266 if (movable_node_is_enabled()) {
6267 for_each_memblock(memory, r) {
6268 if (!memblock_is_hotpluggable(r))
6273 usable_startpfn = PFN_DOWN(r->base);
6274 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6275 min(usable_startpfn, zone_movable_pfn[nid]) :
6283 * If kernelcore=mirror is specified, ignore movablecore option
6285 if (mirrored_kernelcore) {
6286 bool mem_below_4gb_not_mirrored = false;
6288 for_each_memblock(memory, r) {
6289 if (memblock_is_mirror(r))
6294 usable_startpfn = memblock_region_memory_base_pfn(r);
6296 if (usable_startpfn < 0x100000) {
6297 mem_below_4gb_not_mirrored = true;
6301 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6302 min(usable_startpfn, zone_movable_pfn[nid]) :
6306 if (mem_below_4gb_not_mirrored)
6307 pr_warn("This configuration results in unmirrored kernel memory.");
6313 * If movablecore=nn[KMG] was specified, calculate what size of
6314 * kernelcore that corresponds so that memory usable for
6315 * any allocation type is evenly spread. If both kernelcore
6316 * and movablecore are specified, then the value of kernelcore
6317 * will be used for required_kernelcore if it's greater than
6318 * what movablecore would have allowed.
6320 if (required_movablecore) {
6321 unsigned long corepages;
6324 * Round-up so that ZONE_MOVABLE is at least as large as what
6325 * was requested by the user
6327 required_movablecore =
6328 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6329 required_movablecore = min(totalpages, required_movablecore);
6330 corepages = totalpages - required_movablecore;
6332 required_kernelcore = max(required_kernelcore, corepages);
6336 * If kernelcore was not specified or kernelcore size is larger
6337 * than totalpages, there is no ZONE_MOVABLE.
6339 if (!required_kernelcore || required_kernelcore >= totalpages)
6342 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6343 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6346 /* Spread kernelcore memory as evenly as possible throughout nodes */
6347 kernelcore_node = required_kernelcore / usable_nodes;
6348 for_each_node_state(nid, N_MEMORY) {
6349 unsigned long start_pfn, end_pfn;
6352 * Recalculate kernelcore_node if the division per node
6353 * now exceeds what is necessary to satisfy the requested
6354 * amount of memory for the kernel
6356 if (required_kernelcore < kernelcore_node)
6357 kernelcore_node = required_kernelcore / usable_nodes;
6360 * As the map is walked, we track how much memory is usable
6361 * by the kernel using kernelcore_remaining. When it is
6362 * 0, the rest of the node is usable by ZONE_MOVABLE
6364 kernelcore_remaining = kernelcore_node;
6366 /* Go through each range of PFNs within this node */
6367 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6368 unsigned long size_pages;
6370 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6371 if (start_pfn >= end_pfn)
6374 /* Account for what is only usable for kernelcore */
6375 if (start_pfn < usable_startpfn) {
6376 unsigned long kernel_pages;
6377 kernel_pages = min(end_pfn, usable_startpfn)
6380 kernelcore_remaining -= min(kernel_pages,
6381 kernelcore_remaining);
6382 required_kernelcore -= min(kernel_pages,
6383 required_kernelcore);
6385 /* Continue if range is now fully accounted */
6386 if (end_pfn <= usable_startpfn) {
6389 * Push zone_movable_pfn to the end so
6390 * that if we have to rebalance
6391 * kernelcore across nodes, we will
6392 * not double account here
6394 zone_movable_pfn[nid] = end_pfn;
6397 start_pfn = usable_startpfn;
6401 * The usable PFN range for ZONE_MOVABLE is from
6402 * start_pfn->end_pfn. Calculate size_pages as the
6403 * number of pages used as kernelcore
6405 size_pages = end_pfn - start_pfn;
6406 if (size_pages > kernelcore_remaining)
6407 size_pages = kernelcore_remaining;
6408 zone_movable_pfn[nid] = start_pfn + size_pages;
6411 * Some kernelcore has been met, update counts and
6412 * break if the kernelcore for this node has been
6415 required_kernelcore -= min(required_kernelcore,
6417 kernelcore_remaining -= size_pages;
6418 if (!kernelcore_remaining)
6424 * If there is still required_kernelcore, we do another pass with one
6425 * less node in the count. This will push zone_movable_pfn[nid] further
6426 * along on the nodes that still have memory until kernelcore is
6430 if (usable_nodes && required_kernelcore > usable_nodes)
6434 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6435 for (nid = 0; nid < MAX_NUMNODES; nid++)
6436 zone_movable_pfn[nid] =
6437 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6440 /* restore the node_state */
6441 node_states[N_MEMORY] = saved_node_state;
6444 /* Any regular or high memory on that node ? */
6445 static void check_for_memory(pg_data_t *pgdat, int nid)
6447 enum zone_type zone_type;
6449 if (N_MEMORY == N_NORMAL_MEMORY)
6452 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6453 struct zone *zone = &pgdat->node_zones[zone_type];
6454 if (populated_zone(zone)) {
6455 node_set_state(nid, N_HIGH_MEMORY);
6456 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6457 zone_type <= ZONE_NORMAL)
6458 node_set_state(nid, N_NORMAL_MEMORY);
6465 * free_area_init_nodes - Initialise all pg_data_t and zone data
6466 * @max_zone_pfn: an array of max PFNs for each zone
6468 * This will call free_area_init_node() for each active node in the system.
6469 * Using the page ranges provided by memblock_set_node(), the size of each
6470 * zone in each node and their holes is calculated. If the maximum PFN
6471 * between two adjacent zones match, it is assumed that the zone is empty.
6472 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6473 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6474 * starts where the previous one ended. For example, ZONE_DMA32 starts
6475 * at arch_max_dma_pfn.
6477 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6479 unsigned long start_pfn, end_pfn;
6482 /* Record where the zone boundaries are */
6483 memset(arch_zone_lowest_possible_pfn, 0,
6484 sizeof(arch_zone_lowest_possible_pfn));
6485 memset(arch_zone_highest_possible_pfn, 0,
6486 sizeof(arch_zone_highest_possible_pfn));
6488 start_pfn = find_min_pfn_with_active_regions();
6490 for (i = 0; i < MAX_NR_ZONES; i++) {
6491 if (i == ZONE_MOVABLE)
6494 end_pfn = max(max_zone_pfn[i], start_pfn);
6495 arch_zone_lowest_possible_pfn[i] = start_pfn;
6496 arch_zone_highest_possible_pfn[i] = end_pfn;
6498 start_pfn = end_pfn;
6500 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
6501 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
6503 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6504 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6505 find_zone_movable_pfns_for_nodes();
6507 /* Print out the zone ranges */
6508 pr_info("Zone ranges:\n");
6509 for (i = 0; i < MAX_NR_ZONES; i++) {
6510 if (i == ZONE_MOVABLE)
6512 pr_info(" %-8s ", zone_names[i]);
6513 if (arch_zone_lowest_possible_pfn[i] ==
6514 arch_zone_highest_possible_pfn[i])
6517 pr_cont("[mem %#018Lx-%#018Lx]\n",
6518 (u64)arch_zone_lowest_possible_pfn[i]
6520 ((u64)arch_zone_highest_possible_pfn[i]
6521 << PAGE_SHIFT) - 1);
6524 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6525 pr_info("Movable zone start for each node\n");
6526 for (i = 0; i < MAX_NUMNODES; i++) {
6527 if (zone_movable_pfn[i])
6528 pr_info(" Node %d: %#018Lx\n", i,
6529 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6532 /* Print out the early node map */
6533 pr_info("Early memory node ranges\n");
6534 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6535 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6536 (u64)start_pfn << PAGE_SHIFT,
6537 ((u64)end_pfn << PAGE_SHIFT) - 1);
6539 /* Initialise every node */
6540 mminit_verify_pageflags_layout();
6541 setup_nr_node_ids();
6542 for_each_online_node(nid) {
6543 pg_data_t *pgdat = NODE_DATA(nid);
6544 free_area_init_node(nid, NULL,
6545 find_min_pfn_for_node(nid), NULL);
6547 /* Any memory on that node */
6548 if (pgdat->node_present_pages)
6549 node_set_state(nid, N_MEMORY);
6550 check_for_memory(pgdat, nid);
6554 static int __init cmdline_parse_core(char *p, unsigned long *core)
6556 unsigned long long coremem;
6560 coremem = memparse(p, &p);
6561 *core = coremem >> PAGE_SHIFT;
6563 /* Paranoid check that UL is enough for the coremem value */
6564 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6570 * kernelcore=size sets the amount of memory for use for allocations that
6571 * cannot be reclaimed or migrated.
6573 static int __init cmdline_parse_kernelcore(char *p)
6575 /* parse kernelcore=mirror */
6576 if (parse_option_str(p, "mirror")) {
6577 mirrored_kernelcore = true;
6581 return cmdline_parse_core(p, &required_kernelcore);
6585 * movablecore=size sets the amount of memory for use for allocations that
6586 * can be reclaimed or migrated.
6588 static int __init cmdline_parse_movablecore(char *p)
6590 return cmdline_parse_core(p, &required_movablecore);
6593 early_param("kernelcore", cmdline_parse_kernelcore);
6594 early_param("movablecore", cmdline_parse_movablecore);
6596 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6598 void adjust_managed_page_count(struct page *page, long count)
6600 spin_lock(&managed_page_count_lock);
6601 page_zone(page)->managed_pages += count;
6602 totalram_pages += count;
6603 #ifdef CONFIG_HIGHMEM
6604 if (PageHighMem(page))
6605 totalhigh_pages += count;
6607 spin_unlock(&managed_page_count_lock);
6609 EXPORT_SYMBOL(adjust_managed_page_count);
6611 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6614 unsigned long pages = 0;
6616 start = (void *)PAGE_ALIGN((unsigned long)start);
6617 end = (void *)((unsigned long)end & PAGE_MASK);
6618 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6619 if ((unsigned int)poison <= 0xFF)
6620 memset(pos, poison, PAGE_SIZE);
6621 free_reserved_page(virt_to_page(pos));
6625 pr_info("Freeing %s memory: %ldK\n",
6626 s, pages << (PAGE_SHIFT - 10));
6630 EXPORT_SYMBOL(free_reserved_area);
6632 #ifdef CONFIG_HIGHMEM
6633 void free_highmem_page(struct page *page)
6635 __free_reserved_page(page);
6637 page_zone(page)->managed_pages++;
6643 void __init mem_init_print_info(const char *str)
6645 unsigned long physpages, codesize, datasize, rosize, bss_size;
6646 unsigned long init_code_size, init_data_size;
6648 physpages = get_num_physpages();
6649 codesize = _etext - _stext;
6650 datasize = _edata - _sdata;
6651 rosize = __end_rodata - __start_rodata;
6652 bss_size = __bss_stop - __bss_start;
6653 init_data_size = __init_end - __init_begin;
6654 init_code_size = _einittext - _sinittext;
6657 * Detect special cases and adjust section sizes accordingly:
6658 * 1) .init.* may be embedded into .data sections
6659 * 2) .init.text.* may be out of [__init_begin, __init_end],
6660 * please refer to arch/tile/kernel/vmlinux.lds.S.
6661 * 3) .rodata.* may be embedded into .text or .data sections.
6663 #define adj_init_size(start, end, size, pos, adj) \
6665 if (start <= pos && pos < end && size > adj) \
6669 adj_init_size(__init_begin, __init_end, init_data_size,
6670 _sinittext, init_code_size);
6671 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6672 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6673 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6674 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6676 #undef adj_init_size
6678 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6679 #ifdef CONFIG_HIGHMEM
6683 nr_free_pages() << (PAGE_SHIFT - 10),
6684 physpages << (PAGE_SHIFT - 10),
6685 codesize >> 10, datasize >> 10, rosize >> 10,
6686 (init_data_size + init_code_size) >> 10, bss_size >> 10,
6687 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
6688 totalcma_pages << (PAGE_SHIFT - 10),
6689 #ifdef CONFIG_HIGHMEM
6690 totalhigh_pages << (PAGE_SHIFT - 10),
6692 str ? ", " : "", str ? str : "");
6696 * set_dma_reserve - set the specified number of pages reserved in the first zone
6697 * @new_dma_reserve: The number of pages to mark reserved
6699 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
6700 * In the DMA zone, a significant percentage may be consumed by kernel image
6701 * and other unfreeable allocations which can skew the watermarks badly. This
6702 * function may optionally be used to account for unfreeable pages in the
6703 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6704 * smaller per-cpu batchsize.
6706 void __init set_dma_reserve(unsigned long new_dma_reserve)
6708 dma_reserve = new_dma_reserve;
6711 void __init free_area_init(unsigned long *zones_size)
6713 free_area_init_node(0, zones_size,
6714 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
6717 static int page_alloc_cpu_dead(unsigned int cpu)
6720 lru_add_drain_cpu(cpu);
6724 * Spill the event counters of the dead processor
6725 * into the current processors event counters.
6726 * This artificially elevates the count of the current
6729 vm_events_fold_cpu(cpu);
6732 * Zero the differential counters of the dead processor
6733 * so that the vm statistics are consistent.
6735 * This is only okay since the processor is dead and cannot
6736 * race with what we are doing.
6738 cpu_vm_stats_fold(cpu);
6742 void __init page_alloc_init(void)
6746 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
6747 "mm/page_alloc:dead", NULL,
6748 page_alloc_cpu_dead);
6753 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
6754 * or min_free_kbytes changes.
6756 static void calculate_totalreserve_pages(void)
6758 struct pglist_data *pgdat;
6759 unsigned long reserve_pages = 0;
6760 enum zone_type i, j;
6762 for_each_online_pgdat(pgdat) {
6764 pgdat->totalreserve_pages = 0;
6766 for (i = 0; i < MAX_NR_ZONES; i++) {
6767 struct zone *zone = pgdat->node_zones + i;
6770 /* Find valid and maximum lowmem_reserve in the zone */
6771 for (j = i; j < MAX_NR_ZONES; j++) {
6772 if (zone->lowmem_reserve[j] > max)
6773 max = zone->lowmem_reserve[j];
6776 /* we treat the high watermark as reserved pages. */
6777 max += high_wmark_pages(zone);
6779 if (max > zone->managed_pages)
6780 max = zone->managed_pages;
6782 pgdat->totalreserve_pages += max;
6784 reserve_pages += max;
6787 totalreserve_pages = reserve_pages;
6791 * setup_per_zone_lowmem_reserve - called whenever
6792 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
6793 * has a correct pages reserved value, so an adequate number of
6794 * pages are left in the zone after a successful __alloc_pages().
6796 static void setup_per_zone_lowmem_reserve(void)
6798 struct pglist_data *pgdat;
6799 enum zone_type j, idx;
6801 for_each_online_pgdat(pgdat) {
6802 for (j = 0; j < MAX_NR_ZONES; j++) {
6803 struct zone *zone = pgdat->node_zones + j;
6804 unsigned long managed_pages = zone->managed_pages;
6806 zone->lowmem_reserve[j] = 0;
6810 struct zone *lower_zone;
6814 if (sysctl_lowmem_reserve_ratio[idx] < 1)
6815 sysctl_lowmem_reserve_ratio[idx] = 1;
6817 lower_zone = pgdat->node_zones + idx;
6818 lower_zone->lowmem_reserve[j] = managed_pages /
6819 sysctl_lowmem_reserve_ratio[idx];
6820 managed_pages += lower_zone->managed_pages;
6825 /* update totalreserve_pages */
6826 calculate_totalreserve_pages();
6829 static void __setup_per_zone_wmarks(void)
6831 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
6832 unsigned long lowmem_pages = 0;
6834 unsigned long flags;
6836 /* Calculate total number of !ZONE_HIGHMEM pages */
6837 for_each_zone(zone) {
6838 if (!is_highmem(zone))
6839 lowmem_pages += zone->managed_pages;
6842 for_each_zone(zone) {
6845 spin_lock_irqsave(&zone->lock, flags);
6846 tmp = (u64)pages_min * zone->managed_pages;
6847 do_div(tmp, lowmem_pages);
6848 if (is_highmem(zone)) {
6850 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6851 * need highmem pages, so cap pages_min to a small
6854 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6855 * deltas control asynch page reclaim, and so should
6856 * not be capped for highmem.
6858 unsigned long min_pages;
6860 min_pages = zone->managed_pages / 1024;
6861 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
6862 zone->watermark[WMARK_MIN] = min_pages;
6865 * If it's a lowmem zone, reserve a number of pages
6866 * proportionate to the zone's size.
6868 zone->watermark[WMARK_MIN] = tmp;
6872 * Set the kswapd watermarks distance according to the
6873 * scale factor in proportion to available memory, but
6874 * ensure a minimum size on small systems.
6876 tmp = max_t(u64, tmp >> 2,
6877 mult_frac(zone->managed_pages,
6878 watermark_scale_factor, 10000));
6880 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
6881 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
6883 spin_unlock_irqrestore(&zone->lock, flags);
6886 /* update totalreserve_pages */
6887 calculate_totalreserve_pages();
6891 * setup_per_zone_wmarks - called when min_free_kbytes changes
6892 * or when memory is hot-{added|removed}
6894 * Ensures that the watermark[min,low,high] values for each zone are set
6895 * correctly with respect to min_free_kbytes.
6897 void setup_per_zone_wmarks(void)
6899 mutex_lock(&zonelists_mutex);
6900 __setup_per_zone_wmarks();
6901 mutex_unlock(&zonelists_mutex);
6905 * Initialise min_free_kbytes.
6907 * For small machines we want it small (128k min). For large machines
6908 * we want it large (64MB max). But it is not linear, because network
6909 * bandwidth does not increase linearly with machine size. We use
6911 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6912 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6928 int __meminit init_per_zone_wmark_min(void)
6930 unsigned long lowmem_kbytes;
6931 int new_min_free_kbytes;
6933 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
6934 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
6936 if (new_min_free_kbytes > user_min_free_kbytes) {
6937 min_free_kbytes = new_min_free_kbytes;
6938 if (min_free_kbytes < 128)
6939 min_free_kbytes = 128;
6940 if (min_free_kbytes > 65536)
6941 min_free_kbytes = 65536;
6943 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6944 new_min_free_kbytes, user_min_free_kbytes);
6946 setup_per_zone_wmarks();
6947 refresh_zone_stat_thresholds();
6948 setup_per_zone_lowmem_reserve();
6951 setup_min_unmapped_ratio();
6952 setup_min_slab_ratio();
6957 core_initcall(init_per_zone_wmark_min)
6960 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6961 * that we can call two helper functions whenever min_free_kbytes
6964 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
6965 void __user *buffer, size_t *length, loff_t *ppos)
6969 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6974 user_min_free_kbytes = min_free_kbytes;
6975 setup_per_zone_wmarks();
6980 int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
6981 void __user *buffer, size_t *length, loff_t *ppos)
6985 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6990 setup_per_zone_wmarks();
6996 static void setup_min_unmapped_ratio(void)
7001 for_each_online_pgdat(pgdat)
7002 pgdat->min_unmapped_pages = 0;
7005 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7006 sysctl_min_unmapped_ratio) / 100;
7010 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7011 void __user *buffer, size_t *length, loff_t *ppos)
7015 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7019 setup_min_unmapped_ratio();
7024 static void setup_min_slab_ratio(void)
7029 for_each_online_pgdat(pgdat)
7030 pgdat->min_slab_pages = 0;
7033 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7034 sysctl_min_slab_ratio) / 100;
7037 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7038 void __user *buffer, size_t *length, loff_t *ppos)
7042 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7046 setup_min_slab_ratio();
7053 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7054 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7055 * whenever sysctl_lowmem_reserve_ratio changes.
7057 * The reserve ratio obviously has absolutely no relation with the
7058 * minimum watermarks. The lowmem reserve ratio can only make sense
7059 * if in function of the boot time zone sizes.
7061 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7062 void __user *buffer, size_t *length, loff_t *ppos)
7064 proc_dointvec_minmax(table, write, buffer, length, ppos);
7065 setup_per_zone_lowmem_reserve();
7070 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7071 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7072 * pagelist can have before it gets flushed back to buddy allocator.
7074 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7075 void __user *buffer, size_t *length, loff_t *ppos)
7078 int old_percpu_pagelist_fraction;
7081 mutex_lock(&pcp_batch_high_lock);
7082 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7084 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7085 if (!write || ret < 0)
7088 /* Sanity checking to avoid pcp imbalance */
7089 if (percpu_pagelist_fraction &&
7090 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7091 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7097 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7100 for_each_populated_zone(zone) {
7103 for_each_possible_cpu(cpu)
7104 pageset_set_high_and_batch(zone,
7105 per_cpu_ptr(zone->pageset, cpu));
7108 mutex_unlock(&pcp_batch_high_lock);
7113 int hashdist = HASHDIST_DEFAULT;
7115 static int __init set_hashdist(char *str)
7119 hashdist = simple_strtoul(str, &str, 0);
7122 __setup("hashdist=", set_hashdist);
7125 #ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7127 * Returns the number of pages that arch has reserved but
7128 * is not known to alloc_large_system_hash().
7130 static unsigned long __init arch_reserved_kernel_pages(void)
7137 * allocate a large system hash table from bootmem
7138 * - it is assumed that the hash table must contain an exact power-of-2
7139 * quantity of entries
7140 * - limit is the number of hash buckets, not the total allocation size
7142 void *__init alloc_large_system_hash(const char *tablename,
7143 unsigned long bucketsize,
7144 unsigned long numentries,
7147 unsigned int *_hash_shift,
7148 unsigned int *_hash_mask,
7149 unsigned long low_limit,
7150 unsigned long high_limit)
7152 unsigned long long max = high_limit;
7153 unsigned long log2qty, size;
7156 /* allow the kernel cmdline to have a say */
7158 /* round applicable memory size up to nearest megabyte */
7159 numentries = nr_kernel_pages;
7160 numentries -= arch_reserved_kernel_pages();
7162 /* It isn't necessary when PAGE_SIZE >= 1MB */
7163 if (PAGE_SHIFT < 20)
7164 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7166 /* limit to 1 bucket per 2^scale bytes of low memory */
7167 if (scale > PAGE_SHIFT)
7168 numentries >>= (scale - PAGE_SHIFT);
7170 numentries <<= (PAGE_SHIFT - scale);
7172 /* Make sure we've got at least a 0-order allocation.. */
7173 if (unlikely(flags & HASH_SMALL)) {
7174 /* Makes no sense without HASH_EARLY */
7175 WARN_ON(!(flags & HASH_EARLY));
7176 if (!(numentries >> *_hash_shift)) {
7177 numentries = 1UL << *_hash_shift;
7178 BUG_ON(!numentries);
7180 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7181 numentries = PAGE_SIZE / bucketsize;
7183 numentries = roundup_pow_of_two(numentries);
7185 /* limit allocation size to 1/16 total memory by default */
7187 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7188 do_div(max, bucketsize);
7190 max = min(max, 0x80000000ULL);
7192 if (numentries < low_limit)
7193 numentries = low_limit;
7194 if (numentries > max)
7197 log2qty = ilog2(numentries);
7200 size = bucketsize << log2qty;
7201 if (flags & HASH_EARLY)
7202 table = memblock_virt_alloc_nopanic(size, 0);
7204 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
7207 * If bucketsize is not a power-of-two, we may free
7208 * some pages at the end of hash table which
7209 * alloc_pages_exact() automatically does
7211 if (get_order(size) < MAX_ORDER) {
7212 table = alloc_pages_exact(size, GFP_ATOMIC);
7213 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
7216 } while (!table && size > PAGE_SIZE && --log2qty);
7219 panic("Failed to allocate %s hash table\n", tablename);
7221 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7222 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7225 *_hash_shift = log2qty;
7227 *_hash_mask = (1 << log2qty) - 1;
7233 * This function checks whether pageblock includes unmovable pages or not.
7234 * If @count is not zero, it is okay to include less @count unmovable pages
7236 * PageLRU check without isolation or lru_lock could race so that
7237 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
7238 * expect this function should be exact.
7240 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7241 bool skip_hwpoisoned_pages)
7243 unsigned long pfn, iter, found;
7247 * For avoiding noise data, lru_add_drain_all() should be called
7248 * If ZONE_MOVABLE, the zone never contains unmovable pages
7250 if (zone_idx(zone) == ZONE_MOVABLE)
7252 mt = get_pageblock_migratetype(page);
7253 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
7256 pfn = page_to_pfn(page);
7257 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7258 unsigned long check = pfn + iter;
7260 if (!pfn_valid_within(check))
7263 page = pfn_to_page(check);
7266 * Hugepages are not in LRU lists, but they're movable.
7267 * We need not scan over tail pages bacause we don't
7268 * handle each tail page individually in migration.
7270 if (PageHuge(page)) {
7271 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7276 * We can't use page_count without pin a page
7277 * because another CPU can free compound page.
7278 * This check already skips compound tails of THP
7279 * because their page->_refcount is zero at all time.
7281 if (!page_ref_count(page)) {
7282 if (PageBuddy(page))
7283 iter += (1 << page_order(page)) - 1;
7288 * The HWPoisoned page may be not in buddy system, and
7289 * page_count() is not 0.
7291 if (skip_hwpoisoned_pages && PageHWPoison(page))
7297 * If there are RECLAIMABLE pages, we need to check
7298 * it. But now, memory offline itself doesn't call
7299 * shrink_node_slabs() and it still to be fixed.
7302 * If the page is not RAM, page_count()should be 0.
7303 * we don't need more check. This is an _used_ not-movable page.
7305 * The problematic thing here is PG_reserved pages. PG_reserved
7306 * is set to both of a memory hole page and a _used_ kernel
7315 bool is_pageblock_removable_nolock(struct page *page)
7321 * We have to be careful here because we are iterating over memory
7322 * sections which are not zone aware so we might end up outside of
7323 * the zone but still within the section.
7324 * We have to take care about the node as well. If the node is offline
7325 * its NODE_DATA will be NULL - see page_zone.
7327 if (!node_online(page_to_nid(page)))
7330 zone = page_zone(page);
7331 pfn = page_to_pfn(page);
7332 if (!zone_spans_pfn(zone, pfn))
7335 return !has_unmovable_pages(zone, page, 0, true);
7338 #if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7340 static unsigned long pfn_max_align_down(unsigned long pfn)
7342 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7343 pageblock_nr_pages) - 1);
7346 static unsigned long pfn_max_align_up(unsigned long pfn)
7348 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7349 pageblock_nr_pages));
7352 /* [start, end) must belong to a single zone. */
7353 static int __alloc_contig_migrate_range(struct compact_control *cc,
7354 unsigned long start, unsigned long end)
7356 /* This function is based on compact_zone() from compaction.c. */
7357 unsigned long nr_reclaimed;
7358 unsigned long pfn = start;
7359 unsigned int tries = 0;
7364 while (pfn < end || !list_empty(&cc->migratepages)) {
7365 if (fatal_signal_pending(current)) {
7370 if (list_empty(&cc->migratepages)) {
7371 cc->nr_migratepages = 0;
7372 pfn = isolate_migratepages_range(cc, pfn, end);
7378 } else if (++tries == 5) {
7379 ret = ret < 0 ? ret : -EBUSY;
7383 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7385 cc->nr_migratepages -= nr_reclaimed;
7387 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7388 NULL, 0, cc->mode, MR_CMA);
7391 putback_movable_pages(&cc->migratepages);
7398 * alloc_contig_range() -- tries to allocate given range of pages
7399 * @start: start PFN to allocate
7400 * @end: one-past-the-last PFN to allocate
7401 * @migratetype: migratetype of the underlaying pageblocks (either
7402 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7403 * in range must have the same migratetype and it must
7404 * be either of the two.
7406 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7407 * aligned, however it's the caller's responsibility to guarantee that
7408 * we are the only thread that changes migrate type of pageblocks the
7411 * The PFN range must belong to a single zone.
7413 * Returns zero on success or negative error code. On success all
7414 * pages which PFN is in [start, end) are allocated for the caller and
7415 * need to be freed with free_contig_range().
7417 int alloc_contig_range(unsigned long start, unsigned long end,
7418 unsigned migratetype)
7420 unsigned long outer_start, outer_end;
7424 struct compact_control cc = {
7425 .nr_migratepages = 0,
7427 .zone = page_zone(pfn_to_page(start)),
7428 .mode = MIGRATE_SYNC,
7429 .ignore_skip_hint = true,
7430 .gfp_mask = GFP_KERNEL,
7432 INIT_LIST_HEAD(&cc.migratepages);
7435 * What we do here is we mark all pageblocks in range as
7436 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7437 * have different sizes, and due to the way page allocator
7438 * work, we align the range to biggest of the two pages so
7439 * that page allocator won't try to merge buddies from
7440 * different pageblocks and change MIGRATE_ISOLATE to some
7441 * other migration type.
7443 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7444 * migrate the pages from an unaligned range (ie. pages that
7445 * we are interested in). This will put all the pages in
7446 * range back to page allocator as MIGRATE_ISOLATE.
7448 * When this is done, we take the pages in range from page
7449 * allocator removing them from the buddy system. This way
7450 * page allocator will never consider using them.
7452 * This lets us mark the pageblocks back as
7453 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7454 * aligned range but not in the unaligned, original range are
7455 * put back to page allocator so that buddy can use them.
7458 ret = start_isolate_page_range(pfn_max_align_down(start),
7459 pfn_max_align_up(end), migratetype,
7465 * In case of -EBUSY, we'd like to know which page causes problem.
7466 * So, just fall through. We will check it in test_pages_isolated().
7468 ret = __alloc_contig_migrate_range(&cc, start, end);
7469 if (ret && ret != -EBUSY)
7473 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7474 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7475 * more, all pages in [start, end) are free in page allocator.
7476 * What we are going to do is to allocate all pages from
7477 * [start, end) (that is remove them from page allocator).
7479 * The only problem is that pages at the beginning and at the
7480 * end of interesting range may be not aligned with pages that
7481 * page allocator holds, ie. they can be part of higher order
7482 * pages. Because of this, we reserve the bigger range and
7483 * once this is done free the pages we are not interested in.
7485 * We don't have to hold zone->lock here because the pages are
7486 * isolated thus they won't get removed from buddy.
7489 lru_add_drain_all();
7490 drain_all_pages(cc.zone);
7493 outer_start = start;
7494 while (!PageBuddy(pfn_to_page(outer_start))) {
7495 if (++order >= MAX_ORDER) {
7496 outer_start = start;
7499 outer_start &= ~0UL << order;
7502 if (outer_start != start) {
7503 order = page_order(pfn_to_page(outer_start));
7506 * outer_start page could be small order buddy page and
7507 * it doesn't include start page. Adjust outer_start
7508 * in this case to report failed page properly
7509 * on tracepoint in test_pages_isolated()
7511 if (outer_start + (1UL << order) <= start)
7512 outer_start = start;
7515 /* Make sure the range is really isolated. */
7516 if (test_pages_isolated(outer_start, end, false)) {
7517 pr_info("%s: [%lx, %lx) PFNs busy\n",
7518 __func__, outer_start, end);
7523 /* Grab isolated pages from freelists. */
7524 outer_end = isolate_freepages_range(&cc, outer_start, end);
7530 /* Free head and tail (if any) */
7531 if (start != outer_start)
7532 free_contig_range(outer_start, start - outer_start);
7533 if (end != outer_end)
7534 free_contig_range(end, outer_end - end);
7537 undo_isolate_page_range(pfn_max_align_down(start),
7538 pfn_max_align_up(end), migratetype);
7542 void free_contig_range(unsigned long pfn, unsigned nr_pages)
7544 unsigned int count = 0;
7546 for (; nr_pages--; pfn++) {
7547 struct page *page = pfn_to_page(pfn);
7549 count += page_count(page) != 1;
7552 WARN(count != 0, "%d pages are still in use!\n", count);
7556 #ifdef CONFIG_MEMORY_HOTPLUG
7558 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7559 * page high values need to be recalulated.
7561 void __meminit zone_pcp_update(struct zone *zone)
7564 mutex_lock(&pcp_batch_high_lock);
7565 for_each_possible_cpu(cpu)
7566 pageset_set_high_and_batch(zone,
7567 per_cpu_ptr(zone->pageset, cpu));
7568 mutex_unlock(&pcp_batch_high_lock);
7572 void zone_pcp_reset(struct zone *zone)
7574 unsigned long flags;
7576 struct per_cpu_pageset *pset;
7578 /* avoid races with drain_pages() */
7579 local_irq_save(flags);
7580 if (zone->pageset != &boot_pageset) {
7581 for_each_online_cpu(cpu) {
7582 pset = per_cpu_ptr(zone->pageset, cpu);
7583 drain_zonestat(zone, pset);
7585 free_percpu(zone->pageset);
7586 zone->pageset = &boot_pageset;
7588 local_irq_restore(flags);
7591 #ifdef CONFIG_MEMORY_HOTREMOVE
7593 * All pages in the range must be in a single zone and isolated
7594 * before calling this.
7597 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7601 unsigned int order, i;
7603 unsigned long flags;
7604 /* find the first valid pfn */
7605 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7610 zone = page_zone(pfn_to_page(pfn));
7611 spin_lock_irqsave(&zone->lock, flags);
7613 while (pfn < end_pfn) {
7614 if (!pfn_valid(pfn)) {
7618 page = pfn_to_page(pfn);
7620 * The HWPoisoned page may be not in buddy system, and
7621 * page_count() is not 0.
7623 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
7625 SetPageReserved(page);
7629 BUG_ON(page_count(page));
7630 BUG_ON(!PageBuddy(page));
7631 order = page_order(page);
7632 #ifdef CONFIG_DEBUG_VM
7633 pr_info("remove from free list %lx %d %lx\n",
7634 pfn, 1 << order, end_pfn);
7636 list_del(&page->lru);
7637 rmv_page_order(page);
7638 zone->free_area[order].nr_free--;
7639 for (i = 0; i < (1 << order); i++)
7640 SetPageReserved((page+i));
7641 pfn += (1 << order);
7643 spin_unlock_irqrestore(&zone->lock, flags);
7647 bool is_free_buddy_page(struct page *page)
7649 struct zone *zone = page_zone(page);
7650 unsigned long pfn = page_to_pfn(page);
7651 unsigned long flags;
7654 spin_lock_irqsave(&zone->lock, flags);
7655 for (order = 0; order < MAX_ORDER; order++) {
7656 struct page *page_head = page - (pfn & ((1 << order) - 1));
7658 if (PageBuddy(page_head) && page_order(page_head) >= order)
7661 spin_unlock_irqrestore(&zone->lock, flags);
7663 return order < MAX_ORDER;